VirtualBox

source: vbox/trunk/src/VBox/Devices/Storage/DrvVD.cpp@ 64209

Last change on this file since 64209 was 64093, checked in by vboxsync, 8 years ago

Storage: Move the NonRotationalMedium property from the devices and SCSI driver down to the media access driver (VD, HostDVD or HostFloppy) and add callback to PDMIMEDIA to query the flag
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 194.4 KB
Line 
1/* $Id: DrvVD.cpp 64093 2016-09-29 15:39:23Z vboxsync $ */
2/** @file
3 * DrvVD - Generic VBox disk media driver.
4 */
5
6/*
7 * Copyright (C) 2006-2016 Oracle Corporation
8 *
9 * This file is part of VirtualBox Open Source Edition (OSE), as
10 * available from http://www.virtualbox.org. This file is free software;
11 * you can redistribute it and/or modify it under the terms of the GNU
12 * General Public License (GPL) as published by the Free Software
13 * Foundation, in version 2 as it comes in the "COPYING" file of the
14 * VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15 * hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16 */
17
18
19/*********************************************************************************************************************************
20* Header files *
21*********************************************************************************************************************************/
22#define LOG_GROUP LOG_GROUP_DRV_VD
23#include <VBox/vd.h>
24#include <VBox/vmm/pdmdrv.h>
25#include <VBox/vmm/pdmstorageifs.h>
26#include <VBox/vmm/pdmasynccompletion.h>
27#include <VBox/vmm/pdmblkcache.h>
28#include <VBox/vmm/ssm.h>
29#include <iprt/asm.h>
30#include <iprt/alloc.h>
31#include <iprt/assert.h>
32#include <iprt/uuid.h>
33#include <iprt/file.h>
34#include <iprt/string.h>
35#include <iprt/tcp.h>
36#include <iprt/semaphore.h>
37#include <iprt/sg.h>
38#include <iprt/poll.h>
39#include <iprt/pipe.h>
40#include <iprt/system.h>
41#include <iprt/memsafer.h>
42#include <iprt/memcache.h>
43#include <iprt/list.h>
44
45#ifdef VBOX_WITH_INIP
46/* All lwip header files are not C++ safe. So hack around this. */
47RT_C_DECLS_BEGIN
48#include <lwip/opt.h>
49#include <lwip/inet.h>
50#include <lwip/tcp.h>
51#include <lwip/sockets.h>
52# if LWIP_IPV6
53# include <lwip/inet6.h>
54# endif
55RT_C_DECLS_END
56#endif /* VBOX_WITH_INIP */
57
58#include "HBDMgmt.h"
59#include "IOBufMgmt.h"
60
61#include "VBoxDD.h"
62
63#ifdef VBOX_WITH_INIP
64/* Small hack to get at lwIP initialized status */
65extern bool DevINIPConfigured(void);
66#endif /* VBOX_WITH_INIP */
67
68
69/** @def VBOX_PERIODIC_FLUSH
70 * Enable support for periodically flushing the VDI to disk. This may prove
71 * useful for those nasty problems with the ultra-slow host filesystems.
72 * If this is enabled, it can be configured via the CFGM key
73 * "VBoxInternal/Devices/piix3ide/0/LUN#<x>/Config/FlushInterval". <x>
74 * must be replaced with the correct LUN number of the disk that should
75 * do the periodic flushes. The value of the key is the number of bytes
76 * written between flushes. A value of 0 (the default) denotes no flushes. */
77#define VBOX_PERIODIC_FLUSH
78
79/** @def VBOX_IGNORE_FLUSH
80 * Enable support for ignoring VDI flush requests. This can be useful for
81 * filesystems that show bad guest IDE write performance (especially with
82 * Windows guests). NOTE that this does not disable the flushes caused by
83 * the periodic flush cache feature above.
84 * If this feature is enabled, it can be configured via the CFGM key
85 * "VBoxInternal/Devices/piix3ide/0/LUN#<x>/Config/IgnoreFlush". <x>
86 * must be replaced with the correct LUN number of the disk that should
87 * ignore flush requests. The value of the key is a boolean. The default
88 * is to ignore flushes, i.e. true. */
89#define VBOX_IGNORE_FLUSH
90
91
92/*********************************************************************************************************************************
93* Defined types, constants and macros *
94*********************************************************************************************************************************/
95
96/** Converts a pointer to VBOXDISK::IMedia to a PVBOXDISK. */
97#define PDMIMEDIA_2_VBOXDISK(pInterface) \
98 ( (PVBOXDISK)((uintptr_t)pInterface - RT_OFFSETOF(VBOXDISK, IMedia)) )
99
100/** Saved state version of an I/O request .*/
101#define DRVVD_IOREQ_SAVED_STATE_VERSION UINT32_C(1)
102/** Maximum number of request errors in the release log before muting. */
103#define DRVVD_MAX_LOG_REL_ERRORS 100
104
105/** Forward declaration for the dis kcontainer. */
106typedef struct VBOXDISK *PVBOXDISK;
107
108/**
109 * VBox disk container, image information, private part.
110 */
111
112typedef struct VBOXIMAGE
113{
114 /** Pointer to next image. */
115 struct VBOXIMAGE *pNext;
116 /** Pointer to list of VD interfaces. Per-image. */
117 PVDINTERFACE pVDIfsImage;
118 /** Configuration information interface. */
119 VDINTERFACECONFIG VDIfConfig;
120 /** TCP network stack interface. */
121 VDINTERFACETCPNET VDIfTcpNet;
122 /** I/O interface. */
123 VDINTERFACEIO VDIfIo;
124} VBOXIMAGE, *PVBOXIMAGE;
125
126/**
127 * Storage backend data.
128 */
129typedef struct DRVVDSTORAGEBACKEND
130{
131 /** PDM async completion end point. */
132 PPDMASYNCCOMPLETIONENDPOINT pEndpoint;
133 /** The template. */
134 PPDMASYNCCOMPLETIONTEMPLATE pTemplate;
135 /** Event semaphore for synchronous operations. */
136 RTSEMEVENT EventSem;
137 /** Flag whether a synchronous operation is currently pending. */
138 volatile bool fSyncIoPending;
139 /** Return code of the last completed request. */
140 int rcReqLast;
141 /** Callback routine */
142 PFNVDCOMPLETED pfnCompleted;
143} DRVVDSTORAGEBACKEND, *PDRVVDSTORAGEBACKEND;
144
145/**
146 * VD I/O request state.
147 */
148typedef enum VDIOREQSTATE
149{
150 /** Invalid. */
151 VDIOREQSTATE_INVALID = 0,
152 /** The request is not in use and resides on the free list. */
153 VDIOREQSTATE_FREE,
154 /** The request was just allocated and is not active. */
155 VDIOREQSTATE_ALLOCATED,
156 /** The request was allocated and is in use. */
157 VDIOREQSTATE_ACTIVE,
158 /** The request was suspended and is not actively processed. */
159 VDIOREQSTATE_SUSPENDED,
160 /** The request is in the last step of completion and syncs memory. */
161 VDIOREQSTATE_COMPLETING,
162 /** The request completed. */
163 VDIOREQSTATE_COMPLETED,
164 /** The request was aborted but wasn't returned as complete from the storage
165 * layer below us. */
166 VDIOREQSTATE_CANCELED,
167 /** 32bit hack. */
168 VDIOREQSTATE_32BIT_HACK = 0x7fffffff
169} VDIOREQSTATE;
170
171/**
172 * VD I/O Request.
173 */
174typedef struct PDMMEDIAEXIOREQINT
175{
176 /** List node for the list of allocated requests. */
177 RTLISTNODE NdAllocatedList;
178 /** List for requests waiting for I/O memory or on the redo list. */
179 RTLISTNODE NdLstWait;
180 /** I/O request type. */
181 PDMMEDIAEXIOREQTYPE enmType;
182 /** Request state. */
183 volatile VDIOREQSTATE enmState;
184 /** I/O request ID. */
185 PDMMEDIAEXIOREQID uIoReqId;
186 /** Pointer to the disk container. */
187 PVBOXDISK pDisk;
188 /** Flags. */
189 uint32_t fFlags;
190 /** Timestamp when the request was submitted. */
191 uint64_t tsSubmit;
192 /** Type dependent data. */
193 union
194 {
195 /** Read/Write request sepcific data. */
196 struct
197 {
198 /** Start offset of the request. */
199 uint64_t offStart;
200 /** Size of the request. */
201 size_t cbReq;
202 /** Size left for this request. */
203 size_t cbReqLeft;
204 /** Size of the allocated I/O buffer. */
205 size_t cbIoBuf;
206 /** I/O buffer descriptor. */
207 IOBUFDESC IoBuf;
208 } ReadWrite;
209 /** Discard specific data. */
210 struct
211 {
212 /** Pointer to array of ranges to discard. */
213 PRTRANGE paRanges;
214 /** Number of ranges to discard. */
215 unsigned cRanges;
216 } Discard;
217 };
218 /** Allocator specific memory - variable size. */
219 uint8_t abAlloc[1];
220} PDMMEDIAEXIOREQINT;
221/** Pointer to a VD I/O request. */
222typedef PDMMEDIAEXIOREQINT *PPDMMEDIAEXIOREQINT;
223
224/**
225 * Structure for holding a list of allocated requests.
226 */
227typedef struct VDLSTIOREQALLOC
228{
229 /** Mutex protecting the table of allocated requests. */
230 RTSEMFASTMUTEX hMtxLstIoReqAlloc;
231 /** List anchor. */
232 RTLISTANCHOR LstIoReqAlloc;
233} VDLSTIOREQALLOC;
234typedef VDLSTIOREQALLOC *PVDLSTIOREQALLOC;
235
236/** Number of bins for allocated requests. */
237#define DRVVD_VDIOREQ_ALLOC_BINS 8
238
239/**
240 * VBox disk container media main structure, private part.
241 *
242 * @implements PDMIMEDIA
243 * @implements PDMIMEDIAEX
244 * @implements PDMIMOUNT
245 * @implements VDINTERFACEERROR
246 * @implements VDINTERFACETCPNET
247 * @implements VDINTERFACEASYNCIO
248 * @implements VDINTERFACECONFIG
249 */
250typedef struct VBOXDISK
251{
252 /** The VBox disk container. */
253 PVBOXHDD pDisk;
254 /** The media interface. */
255 PDMIMEDIA IMedia;
256 /** Media port. */
257 PPDMIMEDIAPORT pDrvMediaPort;
258 /** Pointer to the driver instance. */
259 PPDMDRVINS pDrvIns;
260 /** Flag whether suspend has changed image open mode to read only. */
261 bool fTempReadOnly;
262 /** Flag whether to use the runtime (true) or startup error facility. */
263 bool fErrorUseRuntime;
264 /** Pointer to list of VD interfaces. Per-disk. */
265 PVDINTERFACE pVDIfsDisk;
266 /** Error interface. */
267 VDINTERFACEERROR VDIfError;
268 /** Thread synchronization interface. */
269 VDINTERFACETHREADSYNC VDIfThreadSync;
270
271 /** Flag whether opened disk supports async I/O operations. */
272 bool fAsyncIOSupported;
273 /** Pointer to the list of data we need to keep per image. */
274 PVBOXIMAGE pImages;
275 /** Flag whether the media should allow concurrent open for writing. */
276 bool fShareable;
277 /** Flag whether a merge operation has been set up. */
278 bool fMergePending;
279 /** Synchronization to prevent destruction before merge finishes. */
280 RTSEMFASTMUTEX MergeCompleteMutex;
281 /** Synchronization between merge and other image accesses. */
282 RTSEMRW MergeLock;
283 /** Source image index for merging. */
284 unsigned uMergeSource;
285 /** Target image index for merging. */
286 unsigned uMergeTarget;
287
288 /** Flag whether boot acceleration is enabled. */
289 bool fBootAccelEnabled;
290 /** Flag whether boot acceleration is currently active. */
291 bool fBootAccelActive;
292 /** Size of the disk, used for read truncation. */
293 uint64_t cbDisk;
294 /** Size of the configured buffer. */
295 size_t cbBootAccelBuffer;
296 /** Start offset for which the buffer holds data. */
297 uint64_t offDisk;
298 /** Number of valid bytes in the buffer. */
299 size_t cbDataValid;
300 /** The disk buffer. */
301 uint8_t *pbData;
302 /** Bandwidth group the disk is assigned to. */
303 char *pszBwGroup;
304 /** Flag whether async I/O using the host cache is enabled. */
305 bool fAsyncIoWithHostCache;
306
307 /** I/O interface for a cache image. */
308 VDINTERFACEIO VDIfIoCache;
309 /** Interface list for the cache image. */
310 PVDINTERFACE pVDIfsCache;
311
312 /** The block cache handle if configured. */
313 PPDMBLKCACHE pBlkCache;
314 /** Host block device manager. */
315 HBDMGR hHbdMgr;
316
317 /** Drive type. */
318 PDMMEDIATYPE enmType;
319 /** Locked indicator. */
320 bool fLocked;
321 /** Mountable indicator. */
322 bool fMountable;
323 /** Visible to the BIOS. */
324 bool fBiosVisible;
325 /** Flag whether this medium should be presented as non rotational. */
326 bool fNonRotational;
327#ifdef VBOX_PERIODIC_FLUSH
328 /** HACK: Configuration value for number of bytes written after which to flush. */
329 uint32_t cbFlushInterval;
330 /** HACK: Current count for the number of bytes written since the last flush. */
331 uint32_t cbDataWritten;
332#endif /* VBOX_PERIODIC_FLUSH */
333#ifdef VBOX_IGNORE_FLUSH
334 /** HACK: Disable flushes for this drive. */
335 bool fIgnoreFlush;
336 /** Disable async flushes for this drive. */
337 bool fIgnoreFlushAsync;
338#endif /* VBOX_IGNORE_FLUSH */
339 /** Our mountable interface. */
340 PDMIMOUNT IMount;
341 /** Pointer to the mount notify interface above us. */
342 PPDMIMOUNTNOTIFY pDrvMountNotify;
343 /** Uuid of the drive. */
344 RTUUID Uuid;
345 /** BIOS PCHS Geometry. */
346 PDMMEDIAGEOMETRY PCHSGeometry;
347 /** BIOS LCHS Geometry. */
348 PDMMEDIAGEOMETRY LCHSGeometry;
349
350 /** Cryptographic support
351 * @{ */
352 /** Pointer to the CFGM node containing the config of the crypto filter
353 * if enable. */
354 PCFGMNODE pCfgCrypto;
355 /** Config interface for the encryption filter. */
356 VDINTERFACECONFIG VDIfCfg;
357 /** Crypto interface for the encryption filter. */
358 VDINTERFACECRYPTO VDIfCrypto;
359 /** The secret key interface used to retrieve keys. */
360 PPDMISECKEY pIfSecKey;
361 /** The secret key helper interface used to notify about missing keys. */
362 PPDMISECKEYHLP pIfSecKeyHlp;
363 /** @} */
364
365 /** @name IMEDIAEX interface support specific members.
366 * @{ */
367 /** Pointer to the IMEDIAEXPORT interface above us. */
368 PPDMIMEDIAEXPORT pDrvMediaExPort;
369 /** Our extended media interface. */
370 PDMIMEDIAEX IMediaEx;
371 /** Memory cache for the I/O requests. */
372 RTMEMCACHE hIoReqCache;
373 /** I/O buffer manager. */
374 IOBUFMGR hIoBufMgr;
375 /** Active request counter. */
376 volatile uint32_t cIoReqsActive;
377 /** Bins for allocated requests. */
378 VDLSTIOREQALLOC aIoReqAllocBins[DRVVD_VDIOREQ_ALLOC_BINS];
379 /** List of requests for I/O memory to be available - VDIOREQ::NdLstWait. */
380 RTLISTANCHOR LstIoReqIoBufWait;
381 /** Critical section protecting the list of requests waiting for I/O memory. */
382 RTCRITSECT CritSectIoReqsIoBufWait;
383 /** Number of requests waiting for a I/O buffer. */
384 volatile uint32_t cIoReqsWaiting;
385 /** Flag whether we have to resubmit requests on resume because the
386 * VM was suspended due to a recoverable I/O error.
387 */
388 volatile bool fRedo;
389 /** List of requests we have to redo. */
390 RTLISTANCHOR LstIoReqRedo;
391 /** Criticial section protecting the list of waiting requests. */
392 RTCRITSECT CritSectIoReqRedo;
393 /** Number of errors logged so far. */
394 unsigned cErrors;
395 /** @} */
396} VBOXDISK;
397
398
399/*********************************************************************************************************************************
400* Internal Functions *
401*********************************************************************************************************************************/
402
403static DECLCALLBACK(void) drvvdMediaExIoReqComplete(void *pvUser1, void *pvUser2, int rcReq);
404static void drvvdPowerOffOrDestructOrUnmount(PPDMDRVINS pDrvIns);
405DECLINLINE(void) drvvdMediaExIoReqBufFree(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq);
406static int drvvdMediaExIoReqCompleteWorker(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, int rcReq, bool fUpNotify);
407static int drvvdMediaExIoReqReadWriteProcess(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, bool fUpNotify);
408
409/**
410 * Internal: allocate new image descriptor and put it in the list
411 */
412static PVBOXIMAGE drvvdNewImage(PVBOXDISK pThis)
413{
414 AssertPtr(pThis);
415 PVBOXIMAGE pImage = (PVBOXIMAGE)RTMemAllocZ(sizeof(VBOXIMAGE));
416 if (pImage)
417 {
418 pImage->pVDIfsImage = NULL;
419 PVBOXIMAGE *pp = &pThis->pImages;
420 while (*pp != NULL)
421 pp = &(*pp)->pNext;
422 *pp = pImage;
423 pImage->pNext = NULL;
424 }
425
426 return pImage;
427}
428
429/**
430 * Internal: free the list of images descriptors.
431 */
432static void drvvdFreeImages(PVBOXDISK pThis)
433{
434 while (pThis->pImages != NULL)
435 {
436 PVBOXIMAGE p = pThis->pImages;
437 pThis->pImages = pThis->pImages->pNext;
438 RTMemFree(p);
439 }
440}
441
442
443/**
444 * Make the image temporarily read-only.
445 *
446 * @returns VBox status code.
447 * @param pThis The driver instance data.
448 */
449static int drvvdSetReadonly(PVBOXDISK pThis)
450{
451 int rc = VINF_SUCCESS;
452 if ( pThis->pDisk
453 && !VDIsReadOnly(pThis->pDisk))
454 {
455 unsigned uOpenFlags;
456 rc = VDGetOpenFlags(pThis->pDisk, VD_LAST_IMAGE, &uOpenFlags);
457 AssertRC(rc);
458 uOpenFlags |= VD_OPEN_FLAGS_READONLY;
459 rc = VDSetOpenFlags(pThis->pDisk, VD_LAST_IMAGE, uOpenFlags);
460 AssertRC(rc);
461 pThis->fTempReadOnly = true;
462 }
463 return rc;
464}
465
466
467/**
468 * Undo the temporary read-only status of the image.
469 *
470 * @returns VBox status code.
471 * @param pThis The driver instance data.
472 */
473static int drvvdSetWritable(PVBOXDISK pThis)
474{
475 int rc = VINF_SUCCESS;
476 if (pThis->fTempReadOnly)
477 {
478 unsigned uOpenFlags;
479 rc = VDGetOpenFlags(pThis->pDisk, VD_LAST_IMAGE, &uOpenFlags);
480 AssertRC(rc);
481 uOpenFlags &= ~VD_OPEN_FLAGS_READONLY;
482 rc = VDSetOpenFlags(pThis->pDisk, VD_LAST_IMAGE, uOpenFlags);
483 if (RT_SUCCESS(rc))
484 pThis->fTempReadOnly = false;
485 else
486 AssertRC(rc);
487 }
488 return rc;
489}
490
491
492/*********************************************************************************************************************************
493* Error reporting callback *
494*********************************************************************************************************************************/
495
496static DECLCALLBACK(void) drvvdErrorCallback(void *pvUser, int rc, RT_SRC_POS_DECL,
497 const char *pszFormat, va_list va)
498{
499 PPDMDRVINS pDrvIns = (PPDMDRVINS)pvUser;
500 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
501 if (pThis->fErrorUseRuntime)
502 /* We must not pass VMSETRTERR_FLAGS_FATAL as it could lead to a
503 * deadlock: We are probably executed in a thread context != EMT
504 * and the EM thread would wait until every thread is suspended
505 * but we would wait for the EM thread ... */
506
507 PDMDrvHlpVMSetRuntimeErrorV(pDrvIns, /* fFlags=*/ 0, "DrvVD", pszFormat, va);
508 else
509 PDMDrvHlpVMSetErrorV(pDrvIns, rc, RT_SRC_POS_ARGS, pszFormat, va);
510}
511
512
513/*********************************************************************************************************************************
514* VD Async I/O interface implementation *
515*********************************************************************************************************************************/
516
517#ifdef VBOX_WITH_PDM_ASYNC_COMPLETION
518
519static DECLCALLBACK(void) drvvdAsyncTaskCompleted(PPDMDRVINS pDrvIns, void *pvTemplateUser, void *pvUser, int rcReq)
520{
521 RT_NOREF(pDrvIns);
522 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pvTemplateUser;
523
524 LogFlowFunc(("pDrvIns=%#p pvTemplateUser=%#p pvUser=%#p rcReq=%d\n",
525 pDrvIns, pvTemplateUser, pvUser, rcReq));
526
527 if (pStorageBackend->fSyncIoPending)
528 {
529 Assert(!pvUser);
530 pStorageBackend->rcReqLast = rcReq;
531 ASMAtomicWriteBool(&pStorageBackend->fSyncIoPending, false);
532 RTSemEventSignal(pStorageBackend->EventSem);
533 }
534 else
535 {
536 int rc;
537
538 AssertPtr(pvUser);
539
540 AssertPtr(pStorageBackend->pfnCompleted);
541 rc = pStorageBackend->pfnCompleted(pvUser, rcReq);
542 AssertRC(rc);
543 }
544}
545
546static DECLCALLBACK(int) drvvdAsyncIOOpen(void *pvUser, const char *pszLocation,
547 uint32_t fOpen,
548 PFNVDCOMPLETED pfnCompleted,
549 void **ppStorage)
550{
551 PVBOXDISK pThis = (PVBOXDISK)pvUser;
552 PDRVVDSTORAGEBACKEND pStorageBackend = NULL;
553 int rc = VINF_SUCCESS;
554
555 /*
556 * Check whether the backend wants to open a block device and try to prepare it
557 * if we didn't claim it yet.
558 *
559 * We only create a block device manager on demand to not waste any resources.
560 */
561 if (HBDMgrIsBlockDevice(pszLocation))
562 {
563 if (pThis->hHbdMgr == NIL_HBDMGR)
564 rc = HBDMgrCreate(&pThis->hHbdMgr);
565
566 if ( RT_SUCCESS(rc)
567 && !HBDMgrIsBlockDeviceClaimed(pThis->hHbdMgr, pszLocation))
568 rc = HBDMgrClaimBlockDevice(pThis->hHbdMgr, pszLocation);
569
570 if (RT_FAILURE(rc))
571 return rc;
572 }
573
574 pStorageBackend = (PDRVVDSTORAGEBACKEND)RTMemAllocZ(sizeof(DRVVDSTORAGEBACKEND));
575 if (pStorageBackend)
576 {
577 pStorageBackend->fSyncIoPending = false;
578 pStorageBackend->rcReqLast = VINF_SUCCESS;
579 pStorageBackend->pfnCompleted = pfnCompleted;
580
581 rc = RTSemEventCreate(&pStorageBackend->EventSem);
582 if (RT_SUCCESS(rc))
583 {
584 rc = PDMDrvHlpAsyncCompletionTemplateCreate(pThis->pDrvIns, &pStorageBackend->pTemplate,
585 drvvdAsyncTaskCompleted, pStorageBackend, "AsyncTaskCompleted");
586 if (RT_SUCCESS(rc))
587 {
588 uint32_t fFlags = (fOpen & RTFILE_O_ACCESS_MASK) == RTFILE_O_READ
589 ? PDMACEP_FILE_FLAGS_READ_ONLY
590 : 0;
591 if (pThis->fShareable)
592 {
593 Assert((fOpen & RTFILE_O_DENY_MASK) == RTFILE_O_DENY_NONE);
594
595 fFlags |= PDMACEP_FILE_FLAGS_DONT_LOCK;
596 }
597 if (pThis->fAsyncIoWithHostCache)
598 fFlags |= PDMACEP_FILE_FLAGS_HOST_CACHE_ENABLED;
599
600 rc = PDMR3AsyncCompletionEpCreateForFile(&pStorageBackend->pEndpoint,
601 pszLocation, fFlags,
602 pStorageBackend->pTemplate);
603
604 if (RT_SUCCESS(rc))
605 {
606 if (pThis->pszBwGroup)
607 rc = PDMR3AsyncCompletionEpSetBwMgr(pStorageBackend->pEndpoint, pThis->pszBwGroup);
608
609 if (RT_SUCCESS(rc))
610 {
611 LogFlow(("drvvdAsyncIOOpen: Successfully opened '%s'; fOpen=%#x pStorage=%p\n",
612 pszLocation, fOpen, pStorageBackend));
613 *ppStorage = pStorageBackend;
614 return VINF_SUCCESS;
615 }
616
617 PDMR3AsyncCompletionEpClose(pStorageBackend->pEndpoint);
618 }
619
620 PDMR3AsyncCompletionTemplateDestroy(pStorageBackend->pTemplate);
621 }
622 RTSemEventDestroy(pStorageBackend->EventSem);
623 }
624 RTMemFree(pStorageBackend);
625 }
626 else
627 rc = VERR_NO_MEMORY;
628
629 return rc;
630}
631
632static DECLCALLBACK(int) drvvdAsyncIOClose(void *pvUser, void *pStorage)
633{
634 RT_NOREF(pvUser);
635 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
636
637 /*
638 * We don't unclaim any block devices on purpose here because they
639 * might get reopened shortly (switching to readonly during suspend)
640 *
641 * Block devices will get unclaimed during destruction of the driver.
642 */
643
644 PDMR3AsyncCompletionEpClose(pStorageBackend->pEndpoint);
645 PDMR3AsyncCompletionTemplateDestroy(pStorageBackend->pTemplate);
646 RTSemEventDestroy(pStorageBackend->EventSem);
647 RTMemFree(pStorageBackend);
648 return VINF_SUCCESS;;
649}
650
651static DECLCALLBACK(int) drvvdAsyncIOReadSync(void *pvUser, void *pStorage, uint64_t uOffset,
652 void *pvBuf, size_t cbRead, size_t *pcbRead)
653{
654 RT_NOREF(pvUser);
655 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
656 RTSGSEG DataSeg;
657 PPDMASYNCCOMPLETIONTASK pTask;
658
659 bool fOld = ASMAtomicXchgBool(&pStorageBackend->fSyncIoPending, true);
660 Assert(!fOld); NOREF(fOld);
661 DataSeg.cbSeg = cbRead;
662 DataSeg.pvSeg = pvBuf;
663
664 int rc = PDMR3AsyncCompletionEpRead(pStorageBackend->pEndpoint, uOffset, &DataSeg, 1, cbRead, NULL, &pTask);
665 if (RT_FAILURE(rc))
666 return rc;
667
668 if (rc == VINF_AIO_TASK_PENDING)
669 {
670 /* Wait */
671 rc = RTSemEventWait(pStorageBackend->EventSem, RT_INDEFINITE_WAIT);
672 AssertRC(rc);
673 }
674 else
675 ASMAtomicXchgBool(&pStorageBackend->fSyncIoPending, false);
676
677 if (pcbRead)
678 *pcbRead = cbRead;
679
680 return pStorageBackend->rcReqLast;
681}
682
683static DECLCALLBACK(int) drvvdAsyncIOWriteSync(void *pvUser, void *pStorage, uint64_t uOffset,
684 const void *pvBuf, size_t cbWrite, size_t *pcbWritten)
685{
686 RT_NOREF(pvUser);
687 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
688 RTSGSEG DataSeg;
689 PPDMASYNCCOMPLETIONTASK pTask;
690
691 bool fOld = ASMAtomicXchgBool(&pStorageBackend->fSyncIoPending, true);
692 Assert(!fOld); NOREF(fOld);
693 DataSeg.cbSeg = cbWrite;
694 DataSeg.pvSeg = (void *)pvBuf;
695
696 int rc = PDMR3AsyncCompletionEpWrite(pStorageBackend->pEndpoint, uOffset, &DataSeg, 1, cbWrite, NULL, &pTask);
697 if (RT_FAILURE(rc))
698 return rc;
699
700 if (rc == VINF_AIO_TASK_PENDING)
701 {
702 /* Wait */
703 rc = RTSemEventWait(pStorageBackend->EventSem, RT_INDEFINITE_WAIT);
704 AssertRC(rc);
705 }
706 else
707 ASMAtomicXchgBool(&pStorageBackend->fSyncIoPending, false);
708
709 if (pcbWritten)
710 *pcbWritten = cbWrite;
711
712 return pStorageBackend->rcReqLast;
713}
714
715static DECLCALLBACK(int) drvvdAsyncIOFlushSync(void *pvUser, void *pStorage)
716{
717 RT_NOREF(pvUser);
718 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
719 PPDMASYNCCOMPLETIONTASK pTask;
720
721 LogFlowFunc(("pvUser=%#p pStorage=%#p\n", pvUser, pStorage));
722
723 bool fOld = ASMAtomicXchgBool(&pStorageBackend->fSyncIoPending, true);
724 Assert(!fOld); NOREF(fOld);
725
726 int rc = PDMR3AsyncCompletionEpFlush(pStorageBackend->pEndpoint, NULL, &pTask);
727 if (RT_FAILURE(rc))
728 return rc;
729
730 if (rc == VINF_AIO_TASK_PENDING)
731 {
732 /* Wait */
733 LogFlowFunc(("Waiting for flush to complete\n"));
734 rc = RTSemEventWait(pStorageBackend->EventSem, RT_INDEFINITE_WAIT);
735 AssertRC(rc);
736 }
737 else
738 ASMAtomicXchgBool(&pStorageBackend->fSyncIoPending, false);
739
740 return pStorageBackend->rcReqLast;
741}
742
743static DECLCALLBACK(int) drvvdAsyncIOReadAsync(void *pvUser, void *pStorage, uint64_t uOffset,
744 PCRTSGSEG paSegments, size_t cSegments,
745 size_t cbRead, void *pvCompletion,
746 void **ppTask)
747{
748 RT_NOREF(pvUser);
749 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
750
751 int rc = PDMR3AsyncCompletionEpRead(pStorageBackend->pEndpoint, uOffset, paSegments, (unsigned)cSegments, cbRead,
752 pvCompletion, (PPPDMASYNCCOMPLETIONTASK)ppTask);
753 if (rc == VINF_AIO_TASK_PENDING)
754 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
755
756 return rc;
757}
758
759static DECLCALLBACK(int) drvvdAsyncIOWriteAsync(void *pvUser, void *pStorage, uint64_t uOffset,
760 PCRTSGSEG paSegments, size_t cSegments,
761 size_t cbWrite, void *pvCompletion,
762 void **ppTask)
763{
764 RT_NOREF(pvUser);
765 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
766
767 int rc = PDMR3AsyncCompletionEpWrite(pStorageBackend->pEndpoint, uOffset, paSegments, (unsigned)cSegments, cbWrite,
768 pvCompletion, (PPPDMASYNCCOMPLETIONTASK)ppTask);
769 if (rc == VINF_AIO_TASK_PENDING)
770 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
771
772 return rc;
773}
774
775static DECLCALLBACK(int) drvvdAsyncIOFlushAsync(void *pvUser, void *pStorage,
776 void *pvCompletion, void **ppTask)
777{
778 RT_NOREF(pvUser);
779 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
780
781 int rc = PDMR3AsyncCompletionEpFlush(pStorageBackend->pEndpoint, pvCompletion,
782 (PPPDMASYNCCOMPLETIONTASK)ppTask);
783 if (rc == VINF_AIO_TASK_PENDING)
784 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
785
786 return rc;
787}
788
789static DECLCALLBACK(int) drvvdAsyncIOGetSize(void *pvUser, void *pStorage, uint64_t *pcbSize)
790{
791 RT_NOREF(pvUser);
792 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
793
794 return PDMR3AsyncCompletionEpGetSize(pStorageBackend->pEndpoint, pcbSize);
795}
796
797static DECLCALLBACK(int) drvvdAsyncIOSetSize(void *pvUser, void *pStorage, uint64_t cbSize)
798{
799 RT_NOREF(pvUser);
800 PDRVVDSTORAGEBACKEND pStorageBackend = (PDRVVDSTORAGEBACKEND)pStorage;
801
802 return PDMR3AsyncCompletionEpSetSize(pStorageBackend->pEndpoint, cbSize);
803}
804
805static DECLCALLBACK(int) drvvdAsyncIOSetAllocationSize(void *pvUser, void *pvStorage, uint64_t cbSize, uint32_t fFlags)
806{
807 RT_NOREF(pvUser, pvStorage, cbSize, fFlags);
808 return VERR_NOT_SUPPORTED;
809}
810
811#endif /* VBOX_WITH_PDM_ASYNC_COMPLETION */
812
813
814/*********************************************************************************************************************************
815* VD Thread Synchronization interface implementation *
816*********************************************************************************************************************************/
817
818static DECLCALLBACK(int) drvvdThreadStartRead(void *pvUser)
819{
820 PVBOXDISK pThis = (PVBOXDISK)pvUser;
821
822 return RTSemRWRequestRead(pThis->MergeLock, RT_INDEFINITE_WAIT);
823}
824
825static DECLCALLBACK(int) drvvdThreadFinishRead(void *pvUser)
826{
827 PVBOXDISK pThis = (PVBOXDISK)pvUser;
828
829 return RTSemRWReleaseRead(pThis->MergeLock);
830}
831
832static DECLCALLBACK(int) drvvdThreadStartWrite(void *pvUser)
833{
834 PVBOXDISK pThis = (PVBOXDISK)pvUser;
835
836 return RTSemRWRequestWrite(pThis->MergeLock, RT_INDEFINITE_WAIT);
837}
838
839static DECLCALLBACK(int) drvvdThreadFinishWrite(void *pvUser)
840{
841 PVBOXDISK pThis = (PVBOXDISK)pvUser;
842
843 return RTSemRWReleaseWrite(pThis->MergeLock);
844}
845
846
847/*********************************************************************************************************************************
848* VD Configuration interface implementation *
849*********************************************************************************************************************************/
850
851static DECLCALLBACK(bool) drvvdCfgAreKeysValid(void *pvUser, const char *pszzValid)
852{
853 return CFGMR3AreValuesValid((PCFGMNODE)pvUser, pszzValid);
854}
855
856static DECLCALLBACK(int) drvvdCfgQuerySize(void *pvUser, const char *pszName, size_t *pcb)
857{
858 return CFGMR3QuerySize((PCFGMNODE)pvUser, pszName, pcb);
859}
860
861static DECLCALLBACK(int) drvvdCfgQuery(void *pvUser, const char *pszName, char *pszString, size_t cchString)
862{
863 return CFGMR3QueryString((PCFGMNODE)pvUser, pszName, pszString, cchString);
864}
865
866static DECLCALLBACK(int) drvvdCfgQueryBytes(void *pvUser, const char *pszName, void *ppvData, size_t cbData)
867{
868 return CFGMR3QueryBytes((PCFGMNODE)pvUser, pszName, ppvData, cbData);
869}
870
871
872/*******************************************************************************
873* VD Crypto interface implementation for the encryption support *
874*******************************************************************************/
875
876static DECLCALLBACK(int) drvvdCryptoKeyRetain(void *pvUser, const char *pszId, const uint8_t **ppbKey, size_t *pcbKey)
877{
878 PVBOXDISK pThis = (PVBOXDISK)pvUser;
879 int rc = VINF_SUCCESS;
880
881 AssertPtr(pThis->pIfSecKey);
882 if (pThis->pIfSecKey)
883 rc = pThis->pIfSecKey->pfnKeyRetain(pThis->pIfSecKey, pszId, ppbKey, pcbKey);
884 else
885 rc = VERR_NOT_SUPPORTED;
886
887 return rc;
888}
889
890static DECLCALLBACK(int) drvvdCryptoKeyRelease(void *pvUser, const char *pszId)
891{
892 PVBOXDISK pThis = (PVBOXDISK)pvUser;
893 int rc = VINF_SUCCESS;
894
895 AssertPtr(pThis->pIfSecKey);
896 if (pThis->pIfSecKey)
897 rc = pThis->pIfSecKey->pfnKeyRelease(pThis->pIfSecKey, pszId);
898 else
899 rc = VERR_NOT_SUPPORTED;
900
901 return rc;
902}
903
904static DECLCALLBACK(int) drvvdCryptoKeyStorePasswordRetain(void *pvUser, const char *pszId, const char **ppszPassword)
905{
906 PVBOXDISK pThis = (PVBOXDISK)pvUser;
907 int rc = VINF_SUCCESS;
908
909 AssertPtr(pThis->pIfSecKey);
910 if (pThis->pIfSecKey)
911 rc = pThis->pIfSecKey->pfnPasswordRetain(pThis->pIfSecKey, pszId, ppszPassword);
912 else
913 rc = VERR_NOT_SUPPORTED;
914
915 return rc;
916}
917
918static DECLCALLBACK(int) drvvdCryptoKeyStorePasswordRelease(void *pvUser, const char *pszId)
919{
920 PVBOXDISK pThis = (PVBOXDISK)pvUser;
921 int rc = VINF_SUCCESS;
922
923 AssertPtr(pThis->pIfSecKey);
924 if (pThis->pIfSecKey)
925 rc = pThis->pIfSecKey->pfnPasswordRelease(pThis->pIfSecKey, pszId);
926 else
927 rc = VERR_NOT_SUPPORTED;
928
929 return rc;
930}
931
932#ifdef VBOX_WITH_INIP
933
934
935/*********************************************************************************************************************************
936* VD TCP network stack interface implementation - INIP case *
937*********************************************************************************************************************************/
938
939/**
940 * vvl: this structure duplicate meaning of sockaddr,
941 * perhaps it'd be better to get rid of it.
942 */
943typedef union INIPSOCKADDRUNION
944{
945 struct sockaddr Addr;
946 struct sockaddr_in Ipv4;
947#if LWIP_IPV6
948 struct sockaddr_in6 Ipv6;
949#endif
950} INIPSOCKADDRUNION;
951
952typedef struct INIPSOCKET
953{
954 int hSock;
955} INIPSOCKET, *PINIPSOCKET;
956
957static DECLCALLBACK(int) drvvdINIPFlush(VDSOCKET Sock);
958
959/** @interface_method_impl{VDINTERFACETCPNET,pfnSocketCreate} */
960static DECLCALLBACK(int) drvvdINIPSocketCreate(uint32_t fFlags, PVDSOCKET pSock)
961{
962 PINIPSOCKET pSocketInt = NULL;
963
964 /*
965 * The extended select method is not supported because it is impossible to wakeup
966 * the thread.
967 */
968 if (fFlags & VD_INTERFACETCPNET_CONNECT_EXTENDED_SELECT)
969 return VERR_NOT_SUPPORTED;
970
971 pSocketInt = (PINIPSOCKET)RTMemAllocZ(sizeof(INIPSOCKET));
972 if (pSocketInt)
973 {
974 pSocketInt->hSock = INT32_MAX;
975 *pSock = (VDSOCKET)pSocketInt;
976 return VINF_SUCCESS;
977 }
978
979 return VERR_NO_MEMORY;
980}
981
982/** @interface_method_impl{VDINTERFACETCPNET,pfnSocketCreate} */
983static DECLCALLBACK(int) drvvdINIPSocketDestroy(VDSOCKET Sock)
984{
985 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
986
987 RTMemFree(pSocketInt);
988 return VINF_SUCCESS;
989}
990
991/** @interface_method_impl{VDINTERFACETCPNET,pfnClientConnect} */
992static DECLCALLBACK(int) drvvdINIPClientConnect(VDSOCKET Sock, const char *pszAddress, uint32_t uPort,
993 RTMSINTERVAL cMillies)
994{
995 int rc = VINF_SUCCESS;
996 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
997 int iInetFamily = PF_INET;
998 struct in_addr ip;
999#if LWIP_IPV6
1000 ip6_addr_t ip6;
1001 RT_ZERO(ip6);
1002#endif
1003
1004 NOREF(cMillies); /* LwIP doesn't support connect timeout. */
1005 RT_ZERO(ip); /* Shut up MSC. */
1006
1007 /* Check whether lwIP is set up in this VM instance. */
1008 if (!DevINIPConfigured())
1009 {
1010 LogRelFunc(("no IP stack\n"));
1011 return VERR_NET_HOST_UNREACHABLE;
1012 }
1013 /* Resolve hostname. As there is no standard resolver for lwIP yet,
1014 * just accept numeric IP addresses for now. */
1015#if LWIP_IPV6
1016 if (inet6_aton(pszAddress, &ip6))
1017 iInetFamily = PF_INET6;
1018 else /* concatination with if */
1019#endif
1020 if (!lwip_inet_aton(pszAddress, &ip))
1021 {
1022 LogRelFunc(("cannot resolve IP %s\n", pszAddress));
1023 return VERR_NET_HOST_UNREACHABLE;
1024 }
1025 /* Create socket and connect. */
1026 int iSock = lwip_socket(iInetFamily, SOCK_STREAM, 0);
1027 if (iSock != -1)
1028 {
1029 struct sockaddr *pSockAddr = NULL;
1030 struct sockaddr_in InAddr = {0};
1031#if LWIP_IPV6
1032 struct sockaddr_in6 In6Addr = {0};
1033#endif
1034 if (iInetFamily == PF_INET)
1035 {
1036 InAddr.sin_family = AF_INET;
1037 InAddr.sin_port = htons(uPort);
1038 InAddr.sin_addr = ip;
1039 InAddr.sin_len = sizeof(InAddr);
1040 pSockAddr = (struct sockaddr *)&InAddr;
1041 }
1042#if LWIP_IPV6
1043 else
1044 {
1045 In6Addr.sin6_family = AF_INET6;
1046 In6Addr.sin6_port = htons(uPort);
1047 memcpy(&In6Addr.sin6_addr, &ip6, sizeof(ip6));
1048 In6Addr.sin6_len = sizeof(In6Addr);
1049 pSockAddr = (struct sockaddr *)&In6Addr;
1050 }
1051#endif
1052 if ( pSockAddr
1053 && !lwip_connect(iSock, pSockAddr, pSockAddr->sa_len))
1054 {
1055 pSocketInt->hSock = iSock;
1056 return VINF_SUCCESS;
1057 }
1058 rc = VERR_NET_CONNECTION_REFUSED; /** @todo real solution needed */
1059 lwip_close(iSock);
1060 }
1061 else
1062 rc = VERR_NET_CONNECTION_REFUSED; /** @todo real solution needed */
1063 return rc;
1064}
1065
1066/** @interface_method_impl{VDINTERFACETCPNET,pfnClientClose} */
1067static DECLCALLBACK(int) drvvdINIPClientClose(VDSOCKET Sock)
1068{
1069 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1070
1071 lwip_close(pSocketInt->hSock);
1072 pSocketInt->hSock = INT32_MAX;
1073 return VINF_SUCCESS; /** @todo real solution needed */
1074}
1075
1076/** @interface_method_impl{VDINTERFACETCPNET,pfnIsClientConnected} */
1077static DECLCALLBACK(bool) drvvdINIPIsClientConnected(VDSOCKET Sock)
1078{
1079 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1080
1081 return pSocketInt->hSock != INT32_MAX;
1082}
1083
1084/** @interface_method_impl{VDINTERFACETCPNET,pfnSelectOne} */
1085static DECLCALLBACK(int) drvvdINIPSelectOne(VDSOCKET Sock, RTMSINTERVAL cMillies)
1086{
1087 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1088 fd_set fdsetR;
1089 FD_ZERO(&fdsetR);
1090 FD_SET((uintptr_t)pSocketInt->hSock, &fdsetR);
1091 fd_set fdsetE = fdsetR;
1092
1093 int rc;
1094 if (cMillies == RT_INDEFINITE_WAIT)
1095 rc = lwip_select(pSocketInt->hSock + 1, &fdsetR, NULL, &fdsetE, NULL);
1096 else
1097 {
1098 struct timeval timeout;
1099 timeout.tv_sec = cMillies / 1000;
1100 timeout.tv_usec = (cMillies % 1000) * 1000;
1101 rc = lwip_select(pSocketInt->hSock + 1, &fdsetR, NULL, &fdsetE, &timeout);
1102 }
1103 if (rc > 0)
1104 return VINF_SUCCESS;
1105 if (rc == 0)
1106 return VERR_TIMEOUT;
1107 return VERR_NET_CONNECTION_REFUSED; /** @todo real solution needed */
1108}
1109
1110/** @interface_method_impl{VDINTERFACETCPNET,pfnRead} */
1111static DECLCALLBACK(int) drvvdINIPRead(VDSOCKET Sock, void *pvBuffer, size_t cbBuffer, size_t *pcbRead)
1112{
1113 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1114
1115 /* Do params checking */
1116 if (!pvBuffer || !cbBuffer)
1117 {
1118 AssertMsgFailed(("Invalid params\n"));
1119 return VERR_INVALID_PARAMETER;
1120 }
1121
1122 /*
1123 * Read loop.
1124 * If pcbRead is NULL we have to fill the entire buffer!
1125 */
1126 size_t cbRead = 0;
1127 size_t cbToRead = cbBuffer;
1128 for (;;)
1129 {
1130 /** @todo this clipping here is just in case (the send function
1131 * needed it, so I added it here, too). Didn't investigate if this
1132 * really has issues. Better be safe than sorry. */
1133 ssize_t cbBytesRead = lwip_recv(pSocketInt->hSock, (char *)pvBuffer + cbRead,
1134 RT_MIN(cbToRead, 32768), 0);
1135 if (cbBytesRead < 0)
1136 return VERR_NET_CONNECTION_REFUSED; /** @todo real solution */
1137 if (cbBytesRead == 0 && errno) /** @todo r=bird: lwip_recv will not touch errno on Windows. This may apply to other hosts as well */
1138 return VERR_NET_CONNECTION_REFUSED; /** @todo real solution */
1139 if (pcbRead)
1140 {
1141 /* return partial data */
1142 *pcbRead = cbBytesRead;
1143 break;
1144 }
1145
1146 /* read more? */
1147 cbRead += cbBytesRead;
1148 if (cbRead == cbBuffer)
1149 break;
1150
1151 /* next */
1152 cbToRead = cbBuffer - cbRead;
1153 }
1154
1155 return VINF_SUCCESS;
1156}
1157
1158/** @interface_method_impl{VDINTERFACETCPNET,pfnWrite} */
1159static DECLCALLBACK(int) drvvdINIPWrite(VDSOCKET Sock, const void *pvBuffer, size_t cbBuffer)
1160{
1161 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1162
1163 do
1164 {
1165 /** @todo lwip send only supports up to 65535 bytes in a single
1166 * send (stupid limitation buried in the code), so make sure we
1167 * don't get any wraparounds. This should be moved to DevINIP
1168 * stack interface once that's implemented. */
1169 ssize_t cbWritten = lwip_send(pSocketInt->hSock, (void *)pvBuffer,
1170 RT_MIN(cbBuffer, 32768), 0);
1171 if (cbWritten < 0)
1172 return VERR_NET_CONNECTION_REFUSED; /** @todo real solution needed */
1173 AssertMsg(cbBuffer >= (size_t)cbWritten, ("Wrote more than we requested!!! cbWritten=%d cbBuffer=%d\n",
1174 cbWritten, cbBuffer));
1175 cbBuffer -= cbWritten;
1176 pvBuffer = (const char *)pvBuffer + cbWritten;
1177 } while (cbBuffer);
1178
1179 return VINF_SUCCESS;
1180}
1181
1182/** @interface_method_impl{VDINTERFACETCPNET,pfnSgWrite} */
1183static DECLCALLBACK(int) drvvdINIPSgWrite(VDSOCKET Sock, PCRTSGBUF pSgBuf)
1184{
1185 int rc = VINF_SUCCESS;
1186
1187 /* This is an extremely crude emulation, however it's good enough
1188 * for our iSCSI code. INIP has no sendmsg(). */
1189 for (unsigned i = 0; i < pSgBuf->cSegs; i++)
1190 {
1191 rc = drvvdINIPWrite(Sock, pSgBuf->paSegs[i].pvSeg,
1192 pSgBuf->paSegs[i].cbSeg);
1193 if (RT_FAILURE(rc))
1194 break;
1195 }
1196 if (RT_SUCCESS(rc))
1197 drvvdINIPFlush(Sock);
1198
1199 return rc;
1200}
1201
1202/** @interface_method_impl{VDINTERFACETCPNET,pfnFlush} */
1203static DECLCALLBACK(int) drvvdINIPFlush(VDSOCKET Sock)
1204{
1205 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1206
1207 int fFlag = 1;
1208 lwip_setsockopt(pSocketInt->hSock, IPPROTO_TCP, TCP_NODELAY,
1209 (const char *)&fFlag, sizeof(fFlag));
1210 fFlag = 0;
1211 lwip_setsockopt(pSocketInt->hSock, IPPROTO_TCP, TCP_NODELAY,
1212 (const char *)&fFlag, sizeof(fFlag));
1213 return VINF_SUCCESS;
1214}
1215
1216/** @interface_method_impl{VDINTERFACETCPNET,pfnSetSendCoalescing} */
1217static DECLCALLBACK(int) drvvdINIPSetSendCoalescing(VDSOCKET Sock, bool fEnable)
1218{
1219 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1220
1221 int fFlag = fEnable ? 0 : 1;
1222 lwip_setsockopt(pSocketInt->hSock, IPPROTO_TCP, TCP_NODELAY,
1223 (const char *)&fFlag, sizeof(fFlag));
1224 return VINF_SUCCESS;
1225}
1226
1227/** @interface_method_impl{VDINTERFACETCPNET,pfnGetLocalAddress} */
1228static DECLCALLBACK(int) drvvdINIPGetLocalAddress(VDSOCKET Sock, PRTNETADDR pAddr)
1229{
1230 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1231 INIPSOCKADDRUNION u;
1232 socklen_t cbAddr = sizeof(u);
1233 RT_ZERO(u);
1234 if (!lwip_getsockname(pSocketInt->hSock, &u.Addr, &cbAddr))
1235 {
1236 /*
1237 * Convert the address.
1238 */
1239 if ( cbAddr == sizeof(struct sockaddr_in)
1240 && u.Addr.sa_family == AF_INET)
1241 {
1242 RT_ZERO(*pAddr);
1243 pAddr->enmType = RTNETADDRTYPE_IPV4;
1244 pAddr->uPort = RT_N2H_U16(u.Ipv4.sin_port);
1245 pAddr->uAddr.IPv4.u = u.Ipv4.sin_addr.s_addr;
1246 }
1247#if LWIP_IPV6
1248 else if ( cbAddr == sizeof(struct sockaddr_in6)
1249 && u.Addr.sa_family == AF_INET6)
1250 {
1251 RT_ZERO(*pAddr);
1252 pAddr->enmType = RTNETADDRTYPE_IPV6;
1253 pAddr->uPort = RT_N2H_U16(u.Ipv6.sin6_port);
1254 memcpy(&pAddr->uAddr.IPv6, &u.Ipv6.sin6_addr, sizeof(RTNETADDRIPV6));
1255 }
1256#endif
1257 else
1258 return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
1259 return VINF_SUCCESS;
1260 }
1261 return VERR_NET_OPERATION_NOT_SUPPORTED;
1262}
1263
1264/** @interface_method_impl{VDINTERFACETCPNET,pfnGetPeerAddress} */
1265static DECLCALLBACK(int) drvvdINIPGetPeerAddress(VDSOCKET Sock, PRTNETADDR pAddr)
1266{
1267 PINIPSOCKET pSocketInt = (PINIPSOCKET)Sock;
1268 INIPSOCKADDRUNION u;
1269 socklen_t cbAddr = sizeof(u);
1270 RT_ZERO(u);
1271 if (!lwip_getpeername(pSocketInt->hSock, &u.Addr, &cbAddr))
1272 {
1273 /*
1274 * Convert the address.
1275 */
1276 if ( cbAddr == sizeof(struct sockaddr_in)
1277 && u.Addr.sa_family == AF_INET)
1278 {
1279 RT_ZERO(*pAddr);
1280 pAddr->enmType = RTNETADDRTYPE_IPV4;
1281 pAddr->uPort = RT_N2H_U16(u.Ipv4.sin_port);
1282 pAddr->uAddr.IPv4.u = u.Ipv4.sin_addr.s_addr;
1283 }
1284#if LWIP_IPV6
1285 else if ( cbAddr == sizeof(struct sockaddr_in6)
1286 && u.Addr.sa_family == AF_INET6)
1287 {
1288 RT_ZERO(*pAddr);
1289 pAddr->enmType = RTNETADDRTYPE_IPV6;
1290 pAddr->uPort = RT_N2H_U16(u.Ipv6.sin6_port);
1291 memcpy(&pAddr->uAddr.IPv6, &u.Ipv6.sin6_addr, sizeof(RTNETADDRIPV6));
1292 }
1293#endif
1294 else
1295 return VERR_NET_ADDRESS_FAMILY_NOT_SUPPORTED;
1296 return VINF_SUCCESS;
1297 }
1298 return VERR_NET_OPERATION_NOT_SUPPORTED;
1299}
1300
1301/** @interface_method_impl{VDINTERFACETCPNET,pfnSelectOneEx} */
1302static DECLCALLBACK(int) drvvdINIPSelectOneEx(VDSOCKET Sock, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
1303{
1304 RT_NOREF(Sock, fEvents, pfEvents, cMillies);
1305 AssertMsgFailed(("Not supported!\n"));
1306 return VERR_NOT_SUPPORTED;
1307}
1308
1309/** @interface_method_impl{VDINTERFACETCPNET,pfnPoke} */
1310static DECLCALLBACK(int) drvvdINIPPoke(VDSOCKET Sock)
1311{
1312 RT_NOREF(Sock);
1313 AssertMsgFailed(("Not supported!\n"));
1314 return VERR_NOT_SUPPORTED;
1315}
1316
1317#endif /* VBOX_WITH_INIP */
1318
1319
1320/*********************************************************************************************************************************
1321* VD TCP network stack interface implementation - Host TCP case *
1322*********************************************************************************************************************************/
1323
1324/**
1325 * Socket data.
1326 */
1327typedef struct VDSOCKETINT
1328{
1329 /** IPRT socket handle. */
1330 RTSOCKET hSocket;
1331 /** Pollset with the wakeup pipe and socket. */
1332 RTPOLLSET hPollSet;
1333 /** Pipe endpoint - read (in the pollset). */
1334 RTPIPE hPipeR;
1335 /** Pipe endpoint - write. */
1336 RTPIPE hPipeW;
1337 /** Flag whether the thread was woken up. */
1338 volatile bool fWokenUp;
1339 /** Flag whether the thread is waiting in the select call. */
1340 volatile bool fWaiting;
1341 /** Old event mask. */
1342 uint32_t fEventsOld;
1343} VDSOCKETINT, *PVDSOCKETINT;
1344
1345/** Pollset id of the socket. */
1346#define VDSOCKET_POLL_ID_SOCKET 0
1347/** Pollset id of the pipe. */
1348#define VDSOCKET_POLL_ID_PIPE 1
1349
1350/** @interface_method_impl{VDINTERFACETCPNET,pfnSocketCreate} */
1351static DECLCALLBACK(int) drvvdTcpSocketCreate(uint32_t fFlags, PVDSOCKET pSock)
1352{
1353 int rc = VINF_SUCCESS;
1354 int rc2 = VINF_SUCCESS;
1355 PVDSOCKETINT pSockInt = NULL;
1356
1357 pSockInt = (PVDSOCKETINT)RTMemAllocZ(sizeof(VDSOCKETINT));
1358 if (!pSockInt)
1359 return VERR_NO_MEMORY;
1360
1361 pSockInt->hSocket = NIL_RTSOCKET;
1362 pSockInt->hPollSet = NIL_RTPOLLSET;
1363 pSockInt->hPipeR = NIL_RTPIPE;
1364 pSockInt->hPipeW = NIL_RTPIPE;
1365 pSockInt->fWokenUp = false;
1366 pSockInt->fWaiting = false;
1367
1368 if (fFlags & VD_INTERFACETCPNET_CONNECT_EXTENDED_SELECT)
1369 {
1370 /* Init pipe and pollset. */
1371 rc = RTPipeCreate(&pSockInt->hPipeR, &pSockInt->hPipeW, 0);
1372 if (RT_SUCCESS(rc))
1373 {
1374 rc = RTPollSetCreate(&pSockInt->hPollSet);
1375 if (RT_SUCCESS(rc))
1376 {
1377 rc = RTPollSetAddPipe(pSockInt->hPollSet, pSockInt->hPipeR,
1378 RTPOLL_EVT_READ, VDSOCKET_POLL_ID_PIPE);
1379 if (RT_SUCCESS(rc))
1380 {
1381 *pSock = pSockInt;
1382 return VINF_SUCCESS;
1383 }
1384
1385 RTPollSetRemove(pSockInt->hPollSet, VDSOCKET_POLL_ID_PIPE);
1386 rc2 = RTPollSetDestroy(pSockInt->hPollSet);
1387 AssertRC(rc2);
1388 }
1389
1390 rc2 = RTPipeClose(pSockInt->hPipeR);
1391 AssertRC(rc2);
1392 rc2 = RTPipeClose(pSockInt->hPipeW);
1393 AssertRC(rc2);
1394 }
1395 }
1396 else
1397 {
1398 *pSock = pSockInt;
1399 return VINF_SUCCESS;
1400 }
1401
1402 RTMemFree(pSockInt);
1403
1404 return rc;
1405}
1406
1407/** @interface_method_impl{VDINTERFACETCPNET,pfnSocketDestroy} */
1408static DECLCALLBACK(int) drvvdTcpSocketDestroy(VDSOCKET Sock)
1409{
1410 int rc = VINF_SUCCESS;
1411 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1412
1413 /* Destroy the pipe and pollset if necessary. */
1414 if (pSockInt->hPollSet != NIL_RTPOLLSET)
1415 {
1416 if (pSockInt->hSocket != NIL_RTSOCKET)
1417 {
1418 rc = RTPollSetRemove(pSockInt->hPollSet, VDSOCKET_POLL_ID_SOCKET);
1419 Assert(RT_SUCCESS(rc) || rc == VERR_POLL_HANDLE_ID_NOT_FOUND);
1420 }
1421 rc = RTPollSetRemove(pSockInt->hPollSet, VDSOCKET_POLL_ID_PIPE);
1422 AssertRC(rc);
1423 rc = RTPollSetDestroy(pSockInt->hPollSet);
1424 AssertRC(rc);
1425 rc = RTPipeClose(pSockInt->hPipeR);
1426 AssertRC(rc);
1427 rc = RTPipeClose(pSockInt->hPipeW);
1428 AssertRC(rc);
1429 }
1430
1431 if (pSockInt->hSocket != NIL_RTSOCKET)
1432 rc = RTTcpClientCloseEx(pSockInt->hSocket, false /*fGracefulShutdown*/);
1433
1434 RTMemFree(pSockInt);
1435
1436 return rc;
1437}
1438
1439/** @interface_method_impl{VDINTERFACETCPNET,pfnClientConnect} */
1440static DECLCALLBACK(int) drvvdTcpClientConnect(VDSOCKET Sock, const char *pszAddress, uint32_t uPort,
1441 RTMSINTERVAL cMillies)
1442{
1443 int rc = VINF_SUCCESS;
1444 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1445
1446 rc = RTTcpClientConnectEx(pszAddress, uPort, &pSockInt->hSocket, cMillies, NULL);
1447 if (RT_SUCCESS(rc))
1448 {
1449 /* Add to the pollset if required. */
1450 if (pSockInt->hPollSet != NIL_RTPOLLSET)
1451 {
1452 pSockInt->fEventsOld = RTPOLL_EVT_READ | RTPOLL_EVT_WRITE | RTPOLL_EVT_ERROR;
1453
1454 rc = RTPollSetAddSocket(pSockInt->hPollSet, pSockInt->hSocket,
1455 pSockInt->fEventsOld, VDSOCKET_POLL_ID_SOCKET);
1456 }
1457
1458 if (RT_SUCCESS(rc))
1459 return VINF_SUCCESS;
1460
1461 rc = RTTcpClientCloseEx(pSockInt->hSocket, false /*fGracefulShutdown*/);
1462 }
1463
1464 return rc;
1465}
1466
1467/** @interface_method_impl{VDINTERFACETCPNET,pfnClientClose} */
1468static DECLCALLBACK(int) drvvdTcpClientClose(VDSOCKET Sock)
1469{
1470 int rc = VINF_SUCCESS;
1471 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1472
1473 if (pSockInt->hPollSet != NIL_RTPOLLSET)
1474 {
1475 rc = RTPollSetRemove(pSockInt->hPollSet, VDSOCKET_POLL_ID_SOCKET);
1476 AssertRC(rc);
1477 }
1478
1479 rc = RTTcpClientCloseEx(pSockInt->hSocket, false /*fGracefulShutdown*/);
1480 pSockInt->hSocket = NIL_RTSOCKET;
1481
1482 return rc;
1483}
1484
1485/** @interface_method_impl{VDINTERFACETCPNET,pfnIsClientConnected} */
1486static DECLCALLBACK(bool) drvvdTcpIsClientConnected(VDSOCKET Sock)
1487{
1488 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1489
1490 return pSockInt->hSocket != NIL_RTSOCKET;
1491}
1492
1493/** @interface_method_impl{VDINTERFACETCPNET,pfnSelectOne} */
1494static DECLCALLBACK(int) drvvdTcpSelectOne(VDSOCKET Sock, RTMSINTERVAL cMillies)
1495{
1496 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1497
1498 return RTTcpSelectOne(pSockInt->hSocket, cMillies);
1499}
1500
1501/** @interface_method_impl{VDINTERFACETCPNET,pfnRead} */
1502static DECLCALLBACK(int) drvvdTcpRead(VDSOCKET Sock, void *pvBuffer, size_t cbBuffer, size_t *pcbRead)
1503{
1504 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1505
1506 return RTTcpRead(pSockInt->hSocket, pvBuffer, cbBuffer, pcbRead);
1507}
1508
1509/** @interface_method_impl{VDINTERFACETCPNET,pfnWrite} */
1510static DECLCALLBACK(int) drvvdTcpWrite(VDSOCKET Sock, const void *pvBuffer, size_t cbBuffer)
1511{
1512 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1513
1514 return RTTcpWrite(pSockInt->hSocket, pvBuffer, cbBuffer);
1515}
1516
1517/** @interface_method_impl{VDINTERFACETCPNET,pfnSgWrite} */
1518static DECLCALLBACK(int) drvvdTcpSgWrite(VDSOCKET Sock, PCRTSGBUF pSgBuf)
1519{
1520 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1521
1522 return RTTcpSgWrite(pSockInt->hSocket, pSgBuf);
1523}
1524
1525/** @interface_method_impl{VDINTERFACETCPNET,pfnReadNB} */
1526static DECLCALLBACK(int) drvvdTcpReadNB(VDSOCKET Sock, void *pvBuffer, size_t cbBuffer, size_t *pcbRead)
1527{
1528 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1529
1530 return RTTcpReadNB(pSockInt->hSocket, pvBuffer, cbBuffer, pcbRead);
1531}
1532
1533/** @interface_method_impl{VDINTERFACETCPNET,pfnWriteNB} */
1534static DECLCALLBACK(int) drvvdTcpWriteNB(VDSOCKET Sock, const void *pvBuffer, size_t cbBuffer, size_t *pcbWritten)
1535{
1536 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1537
1538 return RTTcpWriteNB(pSockInt->hSocket, pvBuffer, cbBuffer, pcbWritten);
1539}
1540
1541/** @interface_method_impl{VDINTERFACETCPNET,pfnSgWriteNB} */
1542static DECLCALLBACK(int) drvvdTcpSgWriteNB(VDSOCKET Sock, PRTSGBUF pSgBuf, size_t *pcbWritten)
1543{
1544 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1545
1546 return RTTcpSgWriteNB(pSockInt->hSocket, pSgBuf, pcbWritten);
1547}
1548
1549/** @interface_method_impl{VDINTERFACETCPNET,pfnFlush} */
1550static DECLCALLBACK(int) drvvdTcpFlush(VDSOCKET Sock)
1551{
1552 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1553
1554 return RTTcpFlush(pSockInt->hSocket);
1555}
1556
1557/** @interface_method_impl{VDINTERFACETCPNET,pfnSetSendCoalescing} */
1558static DECLCALLBACK(int) drvvdTcpSetSendCoalescing(VDSOCKET Sock, bool fEnable)
1559{
1560 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1561
1562 return RTTcpSetSendCoalescing(pSockInt->hSocket, fEnable);
1563}
1564
1565/** @interface_method_impl{VDINTERFACETCPNET,pfnGetLocalAddress} */
1566static DECLCALLBACK(int) drvvdTcpGetLocalAddress(VDSOCKET Sock, PRTNETADDR pAddr)
1567{
1568 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1569
1570 return RTTcpGetLocalAddress(pSockInt->hSocket, pAddr);
1571}
1572
1573/** @interface_method_impl{VDINTERFACETCPNET,pfnGetPeerAddress} */
1574static DECLCALLBACK(int) drvvdTcpGetPeerAddress(VDSOCKET Sock, PRTNETADDR pAddr)
1575{
1576 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1577
1578 return RTTcpGetPeerAddress(pSockInt->hSocket, pAddr);
1579}
1580
1581static DECLCALLBACK(int) drvvdTcpSelectOneExPoll(VDSOCKET Sock, uint32_t fEvents,
1582 uint32_t *pfEvents, RTMSINTERVAL cMillies)
1583{
1584 int rc = VINF_SUCCESS;
1585 uint32_t id = 0;
1586 uint32_t fEventsRecv = 0;
1587 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1588
1589 *pfEvents = 0;
1590
1591 if ( pSockInt->fEventsOld != fEvents
1592 && pSockInt->hSocket != NIL_RTSOCKET)
1593 {
1594 uint32_t fPollEvents = 0;
1595
1596 if (fEvents & VD_INTERFACETCPNET_EVT_READ)
1597 fPollEvents |= RTPOLL_EVT_READ;
1598 if (fEvents & VD_INTERFACETCPNET_EVT_WRITE)
1599 fPollEvents |= RTPOLL_EVT_WRITE;
1600 if (fEvents & VD_INTERFACETCPNET_EVT_ERROR)
1601 fPollEvents |= RTPOLL_EVT_ERROR;
1602
1603 rc = RTPollSetEventsChange(pSockInt->hPollSet, VDSOCKET_POLL_ID_SOCKET, fPollEvents);
1604 if (RT_FAILURE(rc))
1605 return rc;
1606
1607 pSockInt->fEventsOld = fEvents;
1608 }
1609
1610 ASMAtomicXchgBool(&pSockInt->fWaiting, true);
1611 if (ASMAtomicXchgBool(&pSockInt->fWokenUp, false))
1612 {
1613 ASMAtomicXchgBool(&pSockInt->fWaiting, false);
1614 return VERR_INTERRUPTED;
1615 }
1616
1617 rc = RTPoll(pSockInt->hPollSet, cMillies, &fEventsRecv, &id);
1618 Assert(RT_SUCCESS(rc) || rc == VERR_TIMEOUT);
1619
1620 ASMAtomicXchgBool(&pSockInt->fWaiting, false);
1621
1622 if (RT_SUCCESS(rc))
1623 {
1624 if (id == VDSOCKET_POLL_ID_SOCKET)
1625 {
1626 fEventsRecv &= RTPOLL_EVT_VALID_MASK;
1627
1628 if (fEventsRecv & RTPOLL_EVT_READ)
1629 *pfEvents |= VD_INTERFACETCPNET_EVT_READ;
1630 if (fEventsRecv & RTPOLL_EVT_WRITE)
1631 *pfEvents |= VD_INTERFACETCPNET_EVT_WRITE;
1632 if (fEventsRecv & RTPOLL_EVT_ERROR)
1633 *pfEvents |= VD_INTERFACETCPNET_EVT_ERROR;
1634 }
1635 else
1636 {
1637 size_t cbRead = 0;
1638 uint8_t abBuf[10];
1639 Assert(id == VDSOCKET_POLL_ID_PIPE);
1640 Assert((fEventsRecv & RTPOLL_EVT_VALID_MASK) == RTPOLL_EVT_READ);
1641
1642 /* We got interrupted, drain the pipe. */
1643 rc = RTPipeRead(pSockInt->hPipeR, abBuf, sizeof(abBuf), &cbRead);
1644 AssertRC(rc);
1645
1646 ASMAtomicXchgBool(&pSockInt->fWokenUp, false);
1647
1648 rc = VERR_INTERRUPTED;
1649 }
1650 }
1651
1652 return rc;
1653}
1654
1655/** @interface_method_impl{VDINTERFACETCPNET,pfnSelectOneEx} */
1656static DECLCALLBACK(int) drvvdTcpSelectOneExNoPoll(VDSOCKET Sock, uint32_t fEvents, uint32_t *pfEvents, RTMSINTERVAL cMillies)
1657{
1658 RT_NOREF(cMillies); /** @todo timeouts */
1659 int rc = VINF_SUCCESS;
1660 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1661
1662 *pfEvents = 0;
1663
1664 ASMAtomicXchgBool(&pSockInt->fWaiting, true);
1665 if (ASMAtomicXchgBool(&pSockInt->fWokenUp, false))
1666 {
1667 ASMAtomicXchgBool(&pSockInt->fWaiting, false);
1668 return VERR_INTERRUPTED;
1669 }
1670
1671 if ( pSockInt->hSocket == NIL_RTSOCKET
1672 || !fEvents)
1673 {
1674 /*
1675 * Only the pipe is configured or the caller doesn't wait for a socket event,
1676 * wait until there is something to read from the pipe.
1677 */
1678 size_t cbRead = 0;
1679 char ch = 0;
1680 rc = RTPipeReadBlocking(pSockInt->hPipeR, &ch, 1, &cbRead);
1681 if (RT_SUCCESS(rc))
1682 {
1683 Assert(cbRead == 1);
1684 rc = VERR_INTERRUPTED;
1685 ASMAtomicXchgBool(&pSockInt->fWokenUp, false);
1686 }
1687 }
1688 else
1689 {
1690 uint32_t fSelectEvents = 0;
1691
1692 if (fEvents & VD_INTERFACETCPNET_EVT_READ)
1693 fSelectEvents |= RTSOCKET_EVT_READ;
1694 if (fEvents & VD_INTERFACETCPNET_EVT_WRITE)
1695 fSelectEvents |= RTSOCKET_EVT_WRITE;
1696 if (fEvents & VD_INTERFACETCPNET_EVT_ERROR)
1697 fSelectEvents |= RTSOCKET_EVT_ERROR;
1698
1699 if (fEvents & VD_INTERFACETCPNET_HINT_INTERRUPT)
1700 {
1701 uint32_t fEventsRecv = 0;
1702
1703 /* Make sure the socket is not in the pollset. */
1704 rc = RTPollSetRemove(pSockInt->hPollSet, VDSOCKET_POLL_ID_SOCKET);
1705 Assert(RT_SUCCESS(rc) || rc == VERR_POLL_HANDLE_ID_NOT_FOUND);
1706
1707 for (;;)
1708 {
1709 uint32_t id = 0;
1710 rc = RTPoll(pSockInt->hPollSet, 5, &fEvents, &id);
1711 if (rc == VERR_TIMEOUT)
1712 {
1713 /* Check the socket. */
1714 rc = RTTcpSelectOneEx(pSockInt->hSocket, fSelectEvents, &fEventsRecv, 0);
1715 if (RT_SUCCESS(rc))
1716 {
1717 if (fEventsRecv & RTSOCKET_EVT_READ)
1718 *pfEvents |= VD_INTERFACETCPNET_EVT_READ;
1719 if (fEventsRecv & RTSOCKET_EVT_WRITE)
1720 *pfEvents |= VD_INTERFACETCPNET_EVT_WRITE;
1721 if (fEventsRecv & RTSOCKET_EVT_ERROR)
1722 *pfEvents |= VD_INTERFACETCPNET_EVT_ERROR;
1723 break; /* Quit */
1724 }
1725 else if (rc != VERR_TIMEOUT)
1726 break;
1727 }
1728 else if (RT_SUCCESS(rc))
1729 {
1730 size_t cbRead = 0;
1731 uint8_t abBuf[10];
1732 Assert(id == VDSOCKET_POLL_ID_PIPE);
1733 Assert((fEventsRecv & RTPOLL_EVT_VALID_MASK) == RTPOLL_EVT_READ);
1734
1735 /* We got interrupted, drain the pipe. */
1736 rc = RTPipeRead(pSockInt->hPipeR, abBuf, sizeof(abBuf), &cbRead);
1737 AssertRC(rc);
1738
1739 ASMAtomicXchgBool(&pSockInt->fWokenUp, false);
1740
1741 rc = VERR_INTERRUPTED;
1742 break;
1743 }
1744 else
1745 break;
1746 }
1747 }
1748 else /* The caller waits for a socket event. */
1749 {
1750 uint32_t fEventsRecv = 0;
1751
1752 /* Loop until we got woken up or a socket event occurred. */
1753 for (;;)
1754 {
1755 /** @todo find an adaptive wait algorithm based on the
1756 * number of wakeups in the past. */
1757 rc = RTTcpSelectOneEx(pSockInt->hSocket, fSelectEvents, &fEventsRecv, 5);
1758 if (rc == VERR_TIMEOUT)
1759 {
1760 /* Check if there is an event pending. */
1761 size_t cbRead = 0;
1762 char ch = 0;
1763 rc = RTPipeRead(pSockInt->hPipeR, &ch, 1, &cbRead);
1764 if (RT_SUCCESS(rc) && rc != VINF_TRY_AGAIN)
1765 {
1766 Assert(cbRead == 1);
1767 rc = VERR_INTERRUPTED;
1768 ASMAtomicXchgBool(&pSockInt->fWokenUp, false);
1769 break; /* Quit */
1770 }
1771 else
1772 Assert(rc == VINF_TRY_AGAIN);
1773 }
1774 else if (RT_SUCCESS(rc))
1775 {
1776 if (fEventsRecv & RTSOCKET_EVT_READ)
1777 *pfEvents |= VD_INTERFACETCPNET_EVT_READ;
1778 if (fEventsRecv & RTSOCKET_EVT_WRITE)
1779 *pfEvents |= VD_INTERFACETCPNET_EVT_WRITE;
1780 if (fEventsRecv & RTSOCKET_EVT_ERROR)
1781 *pfEvents |= VD_INTERFACETCPNET_EVT_ERROR;
1782 break; /* Quit */
1783 }
1784 else
1785 break;
1786 }
1787 }
1788 }
1789
1790 ASMAtomicXchgBool(&pSockInt->fWaiting, false);
1791
1792 return rc;
1793}
1794
1795/** @interface_method_impl{VDINTERFACETCPNET,pfnPoke} */
1796static DECLCALLBACK(int) drvvdTcpPoke(VDSOCKET Sock)
1797{
1798 int rc = VINF_SUCCESS;
1799 size_t cbWritten = 0;
1800 PVDSOCKETINT pSockInt = (PVDSOCKETINT)Sock;
1801
1802 ASMAtomicXchgBool(&pSockInt->fWokenUp, true);
1803
1804 if (ASMAtomicReadBool(&pSockInt->fWaiting))
1805 {
1806 rc = RTPipeWrite(pSockInt->hPipeW, "", 1, &cbWritten);
1807 Assert(RT_SUCCESS(rc) || cbWritten == 0);
1808 }
1809
1810 return VINF_SUCCESS;
1811}
1812
1813/**
1814 * Checks the prerequisites for encrypted I/O.
1815 *
1816 * @returns VBox status code.
1817 * @param pThis The VD driver instance data.
1818 * @param fSetError Flag whether to set a runtime error.
1819 */
1820static int drvvdKeyCheckPrereqs(PVBOXDISK pThis, bool fSetError)
1821{
1822 if ( pThis->pCfgCrypto
1823 && !pThis->pIfSecKey)
1824 {
1825 AssertPtr(pThis->pIfSecKeyHlp);
1826 pThis->pIfSecKeyHlp->pfnKeyMissingNotify(pThis->pIfSecKeyHlp);
1827
1828 if (fSetError)
1829 {
1830 int rc = PDMDrvHlpVMSetRuntimeError(pThis->pDrvIns, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "DrvVD_DEKMISSING",
1831 N_("VD: The DEK for this disk is missing"));
1832 AssertRC(rc);
1833 }
1834 return VERR_VD_DEK_MISSING;
1835 }
1836
1837 return VINF_SUCCESS;
1838}
1839
1840
1841/*********************************************************************************************************************************
1842* Media interface methods *
1843*********************************************************************************************************************************/
1844
1845/** @interface_method_impl{PDMIMEDIA,pfnRead} */
1846static DECLCALLBACK(int) drvvdRead(PPDMIMEDIA pInterface,
1847 uint64_t off, void *pvBuf, size_t cbRead)
1848{
1849 int rc = VINF_SUCCESS;
1850
1851 LogFlowFunc(("off=%#llx pvBuf=%p cbRead=%d\n", off, pvBuf, cbRead));
1852 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
1853
1854 /*
1855 * Check the state.
1856 */
1857 if (!pThis->pDisk)
1858 {
1859 AssertMsgFailed(("Invalid state! Not mounted!\n"));
1860 return VERR_PDM_MEDIA_NOT_MOUNTED;
1861 }
1862
1863 rc = drvvdKeyCheckPrereqs(pThis, true /* fSetError */);
1864 if (RT_FAILURE(rc))
1865 return rc;
1866
1867 if (!pThis->fBootAccelActive)
1868 rc = VDRead(pThis->pDisk, off, pvBuf, cbRead);
1869 else
1870 {
1871 /* Can we serve the request from the buffer? */
1872 if ( off >= pThis->offDisk
1873 && off - pThis->offDisk < pThis->cbDataValid)
1874 {
1875 size_t cbToCopy = RT_MIN(cbRead, pThis->offDisk + pThis->cbDataValid - off);
1876
1877 memcpy(pvBuf, pThis->pbData + (off - pThis->offDisk), cbToCopy);
1878 cbRead -= cbToCopy;
1879 off += cbToCopy;
1880 pvBuf = (char *)pvBuf + cbToCopy;
1881 }
1882
1883 if ( cbRead > 0
1884 && cbRead < pThis->cbBootAccelBuffer)
1885 {
1886 /* Increase request to the buffer size and read. */
1887 pThis->cbDataValid = RT_MIN(pThis->cbDisk - off, pThis->cbBootAccelBuffer);
1888 pThis->offDisk = off;
1889 rc = VDRead(pThis->pDisk, off, pThis->pbData, pThis->cbDataValid);
1890 if (RT_FAILURE(rc))
1891 pThis->cbDataValid = 0;
1892 else
1893 memcpy(pvBuf, pThis->pbData, cbRead);
1894 }
1895 else if (cbRead >= pThis->cbBootAccelBuffer)
1896 {
1897 pThis->fBootAccelActive = false; /* Deactiviate */
1898 }
1899 }
1900
1901 if (RT_SUCCESS(rc))
1902 Log2(("%s: off=%#llx pvBuf=%p cbRead=%d\n%.*Rhxd\n", __FUNCTION__,
1903 off, pvBuf, cbRead, cbRead, pvBuf));
1904 LogFlowFunc(("returns %Rrc\n", rc));
1905 return rc;
1906}
1907
1908/** @interface_method_impl{PDMIMEDIA,pfnRead} */
1909static DECLCALLBACK(int) drvvdReadPcBios(PPDMIMEDIA pInterface,
1910 uint64_t off, void *pvBuf, size_t cbRead)
1911{
1912 int rc = VINF_SUCCESS;
1913
1914 LogFlowFunc(("off=%#llx pvBuf=%p cbRead=%d\n", off, pvBuf, cbRead));
1915 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
1916
1917 /*
1918 * Check the state.
1919 */
1920 if (!pThis->pDisk)
1921 {
1922 AssertMsgFailed(("Invalid state! Not mounted!\n"));
1923 return VERR_PDM_MEDIA_NOT_MOUNTED;
1924 }
1925
1926 if ( pThis->pCfgCrypto
1927 && !pThis->pIfSecKey)
1928 return VERR_VD_DEK_MISSING;
1929
1930 if (!pThis->fBootAccelActive)
1931 rc = VDRead(pThis->pDisk, off, pvBuf, cbRead);
1932 else
1933 {
1934 /* Can we serve the request from the buffer? */
1935 if ( off >= pThis->offDisk
1936 && off - pThis->offDisk < pThis->cbDataValid)
1937 {
1938 size_t cbToCopy = RT_MIN(cbRead, pThis->offDisk + pThis->cbDataValid - off);
1939
1940 memcpy(pvBuf, pThis->pbData + (off - pThis->offDisk), cbToCopy);
1941 cbRead -= cbToCopy;
1942 off += cbToCopy;
1943 pvBuf = (char *)pvBuf + cbToCopy;
1944 }
1945
1946 if ( cbRead > 0
1947 && cbRead < pThis->cbBootAccelBuffer)
1948 {
1949 /* Increase request to the buffer size and read. */
1950 pThis->cbDataValid = RT_MIN(pThis->cbDisk - off, pThis->cbBootAccelBuffer);
1951 pThis->offDisk = off;
1952 rc = VDRead(pThis->pDisk, off, pThis->pbData, pThis->cbDataValid);
1953 if (RT_FAILURE(rc))
1954 pThis->cbDataValid = 0;
1955 else
1956 memcpy(pvBuf, pThis->pbData, cbRead);
1957 }
1958 else if (cbRead >= pThis->cbBootAccelBuffer)
1959 {
1960 pThis->fBootAccelActive = false; /* Deactiviate */
1961 }
1962 }
1963
1964 if (RT_SUCCESS(rc))
1965 Log2(("%s: off=%#llx pvBuf=%p cbRead=%d\n%.*Rhxd\n", __FUNCTION__,
1966 off, pvBuf, cbRead, cbRead, pvBuf));
1967 LogFlowFunc(("returns %Rrc\n", rc));
1968 return rc;
1969}
1970
1971
1972/** @interface_method_impl{PDMIMEDIA,pfnWrite} */
1973static DECLCALLBACK(int) drvvdWrite(PPDMIMEDIA pInterface,
1974 uint64_t off, const void *pvBuf,
1975 size_t cbWrite)
1976{
1977 LogFlowFunc(("off=%#llx pvBuf=%p cbWrite=%d\n", off, pvBuf, cbWrite));
1978 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
1979 Log2(("%s: off=%#llx pvBuf=%p cbWrite=%d\n%.*Rhxd\n", __FUNCTION__,
1980 off, pvBuf, cbWrite, cbWrite, pvBuf));
1981
1982 /*
1983 * Check the state.
1984 */
1985 if (!pThis->pDisk)
1986 {
1987 AssertMsgFailed(("Invalid state! Not mounted!\n"));
1988 return VERR_PDM_MEDIA_NOT_MOUNTED;
1989 }
1990
1991 /* Set an FTM checkpoint as this operation changes the state permanently. */
1992 PDMDrvHlpFTSetCheckpoint(pThis->pDrvIns, FTMCHECKPOINTTYPE_STORAGE);
1993
1994 int rc = drvvdKeyCheckPrereqs(pThis, true /* fSetError */);
1995 if (RT_FAILURE(rc))
1996 return rc;
1997
1998 /* Invalidate any buffer if boot acceleration is enabled. */
1999 if (pThis->fBootAccelActive)
2000 {
2001 pThis->cbDataValid = 0;
2002 pThis->offDisk = 0;
2003 }
2004
2005 rc = VDWrite(pThis->pDisk, off, pvBuf, cbWrite);
2006#ifdef VBOX_PERIODIC_FLUSH
2007 if (pThis->cbFlushInterval)
2008 {
2009 pThis->cbDataWritten += (uint32_t)cbWrite;
2010 if (pThis->cbDataWritten > pThis->cbFlushInterval)
2011 {
2012 pThis->cbDataWritten = 0;
2013 VDFlush(pThis->pDisk);
2014 }
2015 }
2016#endif /* VBOX_PERIODIC_FLUSH */
2017
2018 LogFlowFunc(("returns %Rrc\n", rc));
2019 return rc;
2020}
2021
2022/** @interface_method_impl{PDMIMEDIA,pfnFlush} */
2023static DECLCALLBACK(int) drvvdFlush(PPDMIMEDIA pInterface)
2024{
2025 LogFlowFunc(("\n"));
2026 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2027
2028 /*
2029 * Check the state.
2030 */
2031 if (!pThis->pDisk)
2032 {
2033 AssertMsgFailed(("Invalid state! Not mounted!\n"));
2034 return VERR_PDM_MEDIA_NOT_MOUNTED;
2035 }
2036
2037#ifdef VBOX_IGNORE_FLUSH
2038 if (pThis->fIgnoreFlush)
2039 return VINF_SUCCESS;
2040#endif /* VBOX_IGNORE_FLUSH */
2041
2042 int rc = VDFlush(pThis->pDisk);
2043 LogFlowFunc(("returns %Rrc\n", rc));
2044 return rc;
2045}
2046
2047/** @interface_method_impl{PDMIMEDIA,pfnMerge} */
2048static DECLCALLBACK(int) drvvdMerge(PPDMIMEDIA pInterface,
2049 PFNSIMPLEPROGRESS pfnProgress,
2050 void *pvUser)
2051{
2052 LogFlowFunc(("\n"));
2053 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2054 int rc = VINF_SUCCESS;
2055
2056 /*
2057 * Check the state.
2058 */
2059 if (!pThis->pDisk)
2060 {
2061 AssertMsgFailed(("Invalid state! Not mounted!\n"));
2062 return VERR_PDM_MEDIA_NOT_MOUNTED;
2063 }
2064
2065 /* Note: There is an unavoidable race between destruction and another
2066 * thread invoking this function. This is handled safely and gracefully by
2067 * atomically invalidating the lock handle in drvvdDestruct. */
2068 int rc2 = RTSemFastMutexRequest(pThis->MergeCompleteMutex);
2069 AssertRC(rc2);
2070 if (RT_SUCCESS(rc2) && pThis->fMergePending)
2071 {
2072 /* Take shortcut: PFNSIMPLEPROGRESS is exactly the same type as
2073 * PFNVDPROGRESS, so there's no need for a conversion function. */
2074 /** @todo maybe introduce a conversion which limits update frequency. */
2075 PVDINTERFACE pVDIfsOperation = NULL;
2076 VDINTERFACEPROGRESS VDIfProgress;
2077 VDIfProgress.pfnProgress = pfnProgress;
2078 rc2 = VDInterfaceAdd(&VDIfProgress.Core, "DrvVD_VDIProgress", VDINTERFACETYPE_PROGRESS,
2079 pvUser, sizeof(VDINTERFACEPROGRESS), &pVDIfsOperation);
2080 AssertRC(rc2);
2081 pThis->fMergePending = false;
2082 rc = VDMerge(pThis->pDisk, pThis->uMergeSource,
2083 pThis->uMergeTarget, pVDIfsOperation);
2084 }
2085 rc2 = RTSemFastMutexRelease(pThis->MergeCompleteMutex);
2086 AssertRC(rc2);
2087 LogFlowFunc(("returns %Rrc\n", rc));
2088 return rc;
2089}
2090
2091/** @interface_method_impl{PDMIMEDIA,pfnSetSecKeyIf} */
2092static DECLCALLBACK(int) drvvdSetSecKeyIf(PPDMIMEDIA pInterface, PPDMISECKEY pIfSecKey, PPDMISECKEYHLP pIfSecKeyHlp)
2093{
2094 LogFlowFunc(("\n"));
2095 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2096 int rc = VINF_SUCCESS;
2097
2098 if (pThis->pCfgCrypto)
2099 {
2100 PVDINTERFACE pVDIfFilter = NULL;
2101
2102 pThis->pIfSecKeyHlp = pIfSecKeyHlp;
2103
2104 if ( pThis->pIfSecKey
2105 && !pIfSecKey)
2106 {
2107 /* Unload the crypto filter first to make sure it doesn't access the keys anymore. */
2108 rc = VDFilterRemove(pThis->pDisk, VD_FILTER_FLAGS_DEFAULT);
2109 AssertRC(rc);
2110
2111 pThis->pIfSecKey = NULL;
2112 }
2113
2114 if ( pIfSecKey
2115 && RT_SUCCESS(rc))
2116 {
2117 pThis->pIfSecKey = pIfSecKey;
2118
2119 rc = VDInterfaceAdd(&pThis->VDIfCfg.Core, "DrvVD_Config", VDINTERFACETYPE_CONFIG,
2120 pThis->pCfgCrypto, sizeof(VDINTERFACECONFIG), &pVDIfFilter);
2121 AssertRC(rc);
2122
2123 rc = VDInterfaceAdd(&pThis->VDIfCrypto.Core, "DrvVD_Crypto", VDINTERFACETYPE_CRYPTO,
2124 pThis, sizeof(VDINTERFACECRYPTO), &pVDIfFilter);
2125 AssertRC(rc);
2126
2127 /* Load the crypt filter plugin. */
2128 rc = VDFilterAdd(pThis->pDisk, "CRYPT", VD_FILTER_FLAGS_DEFAULT, pVDIfFilter);
2129 if (RT_FAILURE(rc))
2130 pThis->pIfSecKey = NULL;
2131 }
2132 }
2133 else
2134 rc = VERR_NOT_SUPPORTED;
2135
2136 LogFlowFunc(("returns %Rrc\n", rc));
2137 return rc;
2138}
2139
2140/** @interface_method_impl{PDMIMEDIA,pfnGetSize} */
2141static DECLCALLBACK(uint64_t) drvvdGetSize(PPDMIMEDIA pInterface)
2142{
2143 LogFlowFunc(("\n"));
2144 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2145
2146 /*
2147 * Check the state.
2148 */
2149 if (!pThis->pDisk)
2150 return 0;
2151
2152 uint64_t cb = VDGetSize(pThis->pDisk, VD_LAST_IMAGE);
2153 LogFlowFunc(("returns %#llx (%llu)\n", cb, cb));
2154 return cb;
2155}
2156
2157/** @interface_method_impl{PDMIMEDIA,pfnGetSectorSize} */
2158static DECLCALLBACK(uint32_t) drvvdGetSectorSize(PPDMIMEDIA pInterface)
2159{
2160 LogFlowFunc(("\n"));
2161 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2162
2163 /*
2164 * Check the state.
2165 */
2166 if (!pThis->pDisk)
2167 return 0;
2168
2169 uint32_t cb = VDGetSectorSize(pThis->pDisk, VD_LAST_IMAGE);
2170 LogFlowFunc(("returns %u\n", cb));
2171 return cb;
2172}
2173
2174/** @interface_method_impl{PDMIMEDIA,pfnIsReadOnly} */
2175static DECLCALLBACK(bool) drvvdIsReadOnly(PPDMIMEDIA pInterface)
2176{
2177 LogFlowFunc(("\n"));
2178 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2179
2180 /*
2181 * Check the state.
2182 */
2183 if (!pThis->pDisk)
2184 return false;
2185
2186 bool f = VDIsReadOnly(pThis->pDisk);
2187 LogFlowFunc(("returns %d\n", f));
2188 return f;
2189}
2190
2191/** @interface_method_impl{PDMIMEDIA,pfnIsNonRotational} */
2192static DECLCALLBACK(bool) drvvdIsNonRotational(PPDMIMEDIA pInterface)
2193{
2194 LogFlowFunc(("\n"));
2195 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2196
2197 return pThis->fNonRotational;
2198}
2199
2200/** @interface_method_impl{PDMIMEDIA,pfnBiosGetPCHSGeometry} */
2201static DECLCALLBACK(int) drvvdBiosGetPCHSGeometry(PPDMIMEDIA pInterface,
2202 PPDMMEDIAGEOMETRY pPCHSGeometry)
2203{
2204 LogFlowFunc(("\n"));
2205 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2206 VDGEOMETRY geo;
2207
2208 /*
2209 * Check the state.
2210 */
2211 if (!pThis->pDisk)
2212 return VERR_PDM_MEDIA_NOT_MOUNTED;
2213
2214 /*
2215 * Use configured/cached values if present.
2216 */
2217 if ( pThis->PCHSGeometry.cCylinders > 0
2218 && pThis->PCHSGeometry.cHeads > 0
2219 && pThis->PCHSGeometry.cSectors > 0)
2220 {
2221 *pPCHSGeometry = pThis->PCHSGeometry;
2222 LogFlow(("%s: returns VINF_SUCCESS {%d,%d,%d}\n", __FUNCTION__, pThis->PCHSGeometry.cCylinders, pThis->PCHSGeometry.cHeads, pThis->PCHSGeometry.cSectors));
2223 return VINF_SUCCESS;
2224 }
2225
2226 int rc = VDGetPCHSGeometry(pThis->pDisk, VD_LAST_IMAGE, &geo);
2227 if (RT_SUCCESS(rc))
2228 {
2229 pPCHSGeometry->cCylinders = geo.cCylinders;
2230 pPCHSGeometry->cHeads = geo.cHeads;
2231 pPCHSGeometry->cSectors = geo.cSectors;
2232 pThis->PCHSGeometry = *pPCHSGeometry;
2233 }
2234 else
2235 {
2236 LogFunc(("geometry not available.\n"));
2237 rc = VERR_PDM_GEOMETRY_NOT_SET;
2238 }
2239 LogFlowFunc(("returns %Rrc (CHS=%d/%d/%d)\n",
2240 rc, pPCHSGeometry->cCylinders, pPCHSGeometry->cHeads, pPCHSGeometry->cSectors));
2241 return rc;
2242}
2243
2244/** @interface_method_impl{PDMIMEDIA,pfnBiosSetPCHSGeometry} */
2245static DECLCALLBACK(int) drvvdBiosSetPCHSGeometry(PPDMIMEDIA pInterface,
2246 PCPDMMEDIAGEOMETRY pPCHSGeometry)
2247{
2248 LogFlowFunc(("CHS=%d/%d/%d\n",
2249 pPCHSGeometry->cCylinders, pPCHSGeometry->cHeads, pPCHSGeometry->cSectors));
2250 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2251 VDGEOMETRY geo;
2252
2253 /*
2254 * Check the state.
2255 */
2256 if (!pThis->pDisk)
2257 {
2258 AssertMsgFailed(("Invalid state! Not mounted!\n"));
2259 return VERR_PDM_MEDIA_NOT_MOUNTED;
2260 }
2261
2262 geo.cCylinders = pPCHSGeometry->cCylinders;
2263 geo.cHeads = pPCHSGeometry->cHeads;
2264 geo.cSectors = pPCHSGeometry->cSectors;
2265 int rc = VDSetPCHSGeometry(pThis->pDisk, VD_LAST_IMAGE, &geo);
2266 if (rc == VERR_VD_GEOMETRY_NOT_SET)
2267 rc = VERR_PDM_GEOMETRY_NOT_SET;
2268 if (RT_SUCCESS(rc))
2269 pThis->PCHSGeometry = *pPCHSGeometry;
2270 LogFlowFunc(("returns %Rrc\n", rc));
2271 return rc;
2272}
2273
2274/** @interface_method_impl{PDMIMEDIA,pfnBiosGetLCHSGeometry} */
2275static DECLCALLBACK(int) drvvdBiosGetLCHSGeometry(PPDMIMEDIA pInterface,
2276 PPDMMEDIAGEOMETRY pLCHSGeometry)
2277{
2278 LogFlowFunc(("\n"));
2279 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2280 VDGEOMETRY geo;
2281
2282 /*
2283 * Check the state.
2284 */
2285 if (!pThis->pDisk)
2286 return VERR_PDM_MEDIA_NOT_MOUNTED;
2287
2288 /*
2289 * Use configured/cached values if present.
2290 */
2291 if ( pThis->LCHSGeometry.cCylinders > 0
2292 && pThis->LCHSGeometry.cHeads > 0
2293 && pThis->LCHSGeometry.cSectors > 0)
2294 {
2295 *pLCHSGeometry = pThis->LCHSGeometry;
2296 LogFlow(("%s: returns VINF_SUCCESS {%d,%d,%d}\n", __FUNCTION__, pThis->LCHSGeometry.cCylinders, pThis->LCHSGeometry.cHeads, pThis->LCHSGeometry.cSectors));
2297 return VINF_SUCCESS;
2298 }
2299
2300 int rc = VDGetLCHSGeometry(pThis->pDisk, VD_LAST_IMAGE, &geo);
2301 if (RT_SUCCESS(rc))
2302 {
2303 pLCHSGeometry->cCylinders = geo.cCylinders;
2304 pLCHSGeometry->cHeads = geo.cHeads;
2305 pLCHSGeometry->cSectors = geo.cSectors;
2306 pThis->LCHSGeometry = *pLCHSGeometry;
2307 }
2308 else
2309 {
2310 LogFunc(("geometry not available.\n"));
2311 rc = VERR_PDM_GEOMETRY_NOT_SET;
2312 }
2313 LogFlowFunc(("returns %Rrc (CHS=%d/%d/%d)\n",
2314 rc, pLCHSGeometry->cCylinders, pLCHSGeometry->cHeads, pLCHSGeometry->cSectors));
2315 return rc;
2316}
2317
2318/** @interface_method_impl{PDMIMEDIA,pfnBiosSetLCHSGeometry} */
2319static DECLCALLBACK(int) drvvdBiosSetLCHSGeometry(PPDMIMEDIA pInterface,
2320 PCPDMMEDIAGEOMETRY pLCHSGeometry)
2321{
2322 LogFlowFunc(("CHS=%d/%d/%d\n",
2323 pLCHSGeometry->cCylinders, pLCHSGeometry->cHeads, pLCHSGeometry->cSectors));
2324 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2325 VDGEOMETRY geo;
2326
2327 /*
2328 * Check the state.
2329 */
2330 if (!pThis->pDisk)
2331 {
2332 AssertMsgFailed(("Invalid state! Not mounted!\n"));
2333 return VERR_PDM_MEDIA_NOT_MOUNTED;
2334 }
2335
2336 geo.cCylinders = pLCHSGeometry->cCylinders;
2337 geo.cHeads = pLCHSGeometry->cHeads;
2338 geo.cSectors = pLCHSGeometry->cSectors;
2339 int rc = VDSetLCHSGeometry(pThis->pDisk, VD_LAST_IMAGE, &geo);
2340 if (rc == VERR_VD_GEOMETRY_NOT_SET)
2341 rc = VERR_PDM_GEOMETRY_NOT_SET;
2342 if (RT_SUCCESS(rc))
2343 pThis->LCHSGeometry = *pLCHSGeometry;
2344 LogFlowFunc(("returns %Rrc\n", rc));
2345 return rc;
2346}
2347
2348/** @interface_method_impl{PDMIMEDIA,pfnBiosIsVisible} */
2349static DECLCALLBACK(bool) drvvdBiosIsVisible(PPDMIMEDIA pInterface)
2350{
2351 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2352 LogFlow(("drvvdBiosIsVisible: returns %d\n", pThis->fBiosVisible));
2353 return pThis->fBiosVisible;
2354}
2355
2356/** @interface_method_impl{PDMIMEDIA,pfnGetType} */
2357static DECLCALLBACK(PDMMEDIATYPE) drvvdGetType(PPDMIMEDIA pInterface)
2358{
2359 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2360 LogFlow(("drvvdBiosIsVisible: returns %d\n", pThis->fBiosVisible));
2361 return pThis->enmType;
2362}
2363
2364/** @interface_method_impl{PDMIMEDIA,pfnGetUuid} */
2365static DECLCALLBACK(int) drvvdGetUuid(PPDMIMEDIA pInterface, PRTUUID pUuid)
2366{
2367 LogFlowFunc(("\n"));
2368 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2369
2370 /*
2371 * Copy the uuid.
2372 */
2373 *pUuid = pThis->Uuid;
2374 LogFlowFunc(("returns {%RTuuid}\n", pUuid));
2375 return VINF_SUCCESS;
2376}
2377
2378static DECLCALLBACK(int) drvvdDiscard(PPDMIMEDIA pInterface, PCRTRANGE paRanges, unsigned cRanges)
2379{
2380 LogFlowFunc(("\n"));
2381 PVBOXDISK pThis = PDMIMEDIA_2_VBOXDISK(pInterface);
2382
2383 int rc = VDDiscardRanges(pThis->pDisk, paRanges, cRanges);
2384 LogFlowFunc(("returns %Rrc\n", rc));
2385 return rc;
2386}
2387
2388/* -=-=-=-=- IMount -=-=-=-=- */
2389
2390/** @interface_method_impl{PDMIMOUNT,pfnUnmount} */
2391static DECLCALLBACK(int) drvvdUnmount(PPDMIMOUNT pInterface, bool fForce, bool fEject)
2392{
2393 RT_NOREF(fEject);
2394 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMount);
2395
2396 /*
2397 * Validate state.
2398 */
2399 if (!pThis->pDisk)
2400 {
2401 Log(("drvvdUnmount: Not mounted\n"));
2402 return VERR_PDM_MEDIA_NOT_MOUNTED;
2403 }
2404 if (pThis->fLocked && !fForce)
2405 {
2406 Log(("drvvdUnmount: Locked\n"));
2407 return VERR_PDM_MEDIA_LOCKED;
2408 }
2409
2410 /* Media is no longer locked even if it was previously. */
2411 pThis->fLocked = false;
2412 drvvdPowerOffOrDestructOrUnmount(pThis->pDrvIns);
2413
2414 /*
2415 * Notify driver/device above us.
2416 */
2417 if (pThis->pDrvMountNotify)
2418 pThis->pDrvMountNotify->pfnUnmountNotify(pThis->pDrvMountNotify);
2419 Log(("drvblockUnmount: success\n"));
2420 return VINF_SUCCESS;
2421}
2422
2423
2424/** @interface_method_impl{PDMIMOUNT,pfnIsMounted} */
2425static DECLCALLBACK(bool) drvvdIsMounted(PPDMIMOUNT pInterface)
2426{
2427 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMount);
2428 return pThis->pDisk != NULL;
2429}
2430
2431/** @interface_method_impl{PDMIMOUNT,pfnLock} */
2432static DECLCALLBACK(int) drvvdLock(PPDMIMOUNT pInterface)
2433{
2434 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMount);
2435 Log(("drvblockLock: %d -> %d\n", pThis->fLocked, true));
2436 pThis->fLocked = true;
2437 return VINF_SUCCESS;
2438}
2439
2440/** @interface_method_impl{PDMIMOUNT,pfnUnlock} */
2441static DECLCALLBACK(int) drvvdUnlock(PPDMIMOUNT pInterface)
2442{
2443 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMount);
2444 Log(("drvblockUnlock: %d -> %d\n", pThis->fLocked, false));
2445 pThis->fLocked = false;
2446 return VINF_SUCCESS;
2447}
2448
2449/** @interface_method_impl{PDMIMOUNT,pfnIsLocked} */
2450static DECLCALLBACK(bool) drvvdIsLocked(PPDMIMOUNT pInterface)
2451{
2452 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMount);
2453 return pThis->fLocked;
2454}
2455
2456
2457static DECLCALLBACK(void) drvvdBlkCacheReqComplete(void *pvUser1, void *pvUser2, int rcReq)
2458{
2459 PVBOXDISK pThis = (PVBOXDISK)pvUser1;
2460
2461 AssertPtr(pThis->pBlkCache);
2462 PDMR3BlkCacheIoXferComplete(pThis->pBlkCache, (PPDMBLKCACHEIOXFER)pvUser2, rcReq);
2463}
2464
2465
2466/** @copydoc FNPDMBLKCACHEXFERCOMPLETEDRV */
2467static DECLCALLBACK(void) drvvdBlkCacheXferCompleteIoReq(PPDMDRVINS pDrvIns, void *pvUser, int rcReq)
2468{
2469 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
2470
2471 drvvdMediaExIoReqCompleteWorker(pThis, (PPDMMEDIAEXIOREQINT)pvUser, rcReq, true /* fUpNotify */);
2472}
2473
2474/** @copydoc FNPDMBLKCACHEXFERENQUEUEDRV */
2475static DECLCALLBACK(int) drvvdBlkCacheXferEnqueue(PPDMDRVINS pDrvIns,
2476 PDMBLKCACHEXFERDIR enmXferDir,
2477 uint64_t off, size_t cbXfer,
2478 PCRTSGBUF pcSgBuf, PPDMBLKCACHEIOXFER hIoXfer)
2479{
2480 int rc = VINF_SUCCESS;
2481 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
2482
2483 Assert (!pThis->pCfgCrypto);
2484
2485 switch (enmXferDir)
2486 {
2487 case PDMBLKCACHEXFERDIR_READ:
2488 rc = VDAsyncRead(pThis->pDisk, off, cbXfer, pcSgBuf, drvvdBlkCacheReqComplete,
2489 pThis, hIoXfer);
2490 break;
2491 case PDMBLKCACHEXFERDIR_WRITE:
2492 rc = VDAsyncWrite(pThis->pDisk, off, cbXfer, pcSgBuf, drvvdBlkCacheReqComplete,
2493 pThis, hIoXfer);
2494 break;
2495 case PDMBLKCACHEXFERDIR_FLUSH:
2496 rc = VDAsyncFlush(pThis->pDisk, drvvdBlkCacheReqComplete, pThis, hIoXfer);
2497 break;
2498 default:
2499 AssertMsgFailed(("Invalid transfer type %d\n", enmXferDir));
2500 rc = VERR_INVALID_PARAMETER;
2501 }
2502
2503 if (rc == VINF_VD_ASYNC_IO_FINISHED)
2504 PDMR3BlkCacheIoXferComplete(pThis->pBlkCache, hIoXfer, VINF_SUCCESS);
2505 else if (RT_FAILURE(rc) && rc != VERR_VD_ASYNC_IO_IN_PROGRESS)
2506 PDMR3BlkCacheIoXferComplete(pThis->pBlkCache, hIoXfer, rc);
2507
2508 return VINF_SUCCESS;
2509}
2510
2511/** @copydoc FNPDMBLKCACHEXFERENQUEUEDISCARDDRV */
2512static DECLCALLBACK(int) drvvdBlkCacheXferEnqueueDiscard(PPDMDRVINS pDrvIns, PCRTRANGE paRanges,
2513 unsigned cRanges, PPDMBLKCACHEIOXFER hIoXfer)
2514{
2515 int rc = VINF_SUCCESS;
2516 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
2517
2518 rc = VDAsyncDiscardRanges(pThis->pDisk, paRanges, cRanges,
2519 drvvdBlkCacheReqComplete, pThis, hIoXfer);
2520
2521 if (rc == VINF_VD_ASYNC_IO_FINISHED)
2522 PDMR3BlkCacheIoXferComplete(pThis->pBlkCache, hIoXfer, VINF_SUCCESS);
2523 else if (RT_FAILURE(rc) && rc != VERR_VD_ASYNC_IO_IN_PROGRESS)
2524 PDMR3BlkCacheIoXferComplete(pThis->pBlkCache, hIoXfer, rc);
2525
2526 return VINF_SUCCESS;
2527}
2528
2529
2530/*********************************************************************************************************************************
2531* Extended media interface methods *
2532*********************************************************************************************************************************/
2533
2534static void drvvdMediaExIoReqWarningDiskFull(PPDMDRVINS pDrvIns)
2535{
2536 int rc;
2537 LogRel(("VD#%u: Host disk full\n", pDrvIns->iInstance));
2538 rc = PDMDrvHlpVMSetRuntimeError(pDrvIns, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "DrvVD_DISKFULL",
2539 N_("Host system reported disk full. VM execution is suspended. You can resume after freeing some space"));
2540 AssertRC(rc);
2541}
2542
2543static void drvvdMediaExIoReqWarningFileTooBig(PPDMDRVINS pDrvIns)
2544{
2545 int rc;
2546 LogRel(("VD#%u: File too big\n", pDrvIns->iInstance));
2547 rc = PDMDrvHlpVMSetRuntimeError(pDrvIns, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "DrvVD_FILETOOBIG",
2548 N_("Host system reported that the file size limit of the host file system has been exceeded. VM execution is suspended. You need to move your virtual hard disk to a filesystem which allows bigger files"));
2549 AssertRC(rc);
2550}
2551
2552static void drvvdMediaExIoReqWarningISCSI(PPDMDRVINS pDrvIns)
2553{
2554 int rc;
2555 LogRel(("VD#%u: iSCSI target unavailable\n", pDrvIns->iInstance));
2556 rc = PDMDrvHlpVMSetRuntimeError(pDrvIns, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "DrvVD_ISCSIDOWN",
2557 N_("The iSCSI target has stopped responding. VM execution is suspended. You can resume when it is available again"));
2558 AssertRC(rc);
2559}
2560
2561static void drvvdMediaExIoReqWarningDekMissing(PPDMDRVINS pDrvIns)
2562{
2563 LogRel(("VD#%u: DEK is missing\n", pDrvIns->iInstance));
2564 int rc = PDMDrvHlpVMSetRuntimeError(pDrvIns, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "DrvVD_DEKMISSING",
2565 N_("VD: The DEK for this disk is missing"));
2566 AssertRC(rc);
2567}
2568
2569/**
2570 * Checks whether a given status code indicates a recoverable error
2571 * suspending the VM if it is.
2572 *
2573 * @returns Flag indicating whether the status code is a recoverable error
2574 * (full disk, broken network connection).
2575 * @param pThis VBox disk container instance data.
2576 * @param rc Status code to check.
2577 */
2578bool drvvdMediaExIoReqIsRedoSetWarning(PVBOXDISK pThis, int rc)
2579{
2580 if (rc == VERR_DISK_FULL)
2581 {
2582 if (ASMAtomicCmpXchgBool(&pThis->fRedo, true, false))
2583 drvvdMediaExIoReqWarningDiskFull(pThis->pDrvIns);
2584 return true;
2585 }
2586 if (rc == VERR_FILE_TOO_BIG)
2587 {
2588 if (ASMAtomicCmpXchgBool(&pThis->fRedo, true, false))
2589 drvvdMediaExIoReqWarningFileTooBig(pThis->pDrvIns);
2590 return true;
2591 }
2592 if (rc == VERR_BROKEN_PIPE || rc == VERR_NET_CONNECTION_REFUSED)
2593 {
2594 /* iSCSI connection abort (first error) or failure to reestablish
2595 * connection (second error). Pause VM. On resume we'll retry. */
2596 if (ASMAtomicCmpXchgBool(&pThis->fRedo, true, false))
2597 drvvdMediaExIoReqWarningISCSI(pThis->pDrvIns);
2598 return true;
2599 }
2600 if (rc == VERR_VD_DEK_MISSING)
2601 {
2602 /* Error message already set. */
2603 if (ASMAtomicCmpXchgBool(&pThis->fRedo, true, false))
2604 drvvdMediaExIoReqWarningDekMissing(pThis->pDrvIns);
2605 return true;
2606 }
2607
2608 return false;
2609}
2610
2611/**
2612 * Syncs the memory buffers between the I/O request allocator and the internal buffer.
2613 *
2614 * @returns VBox status code.
2615 * @param pThis VBox disk container instance data.
2616 * @param pIoReq I/O request to sync.
2617 * @param fToIoBuf Flag indicating the sync direction.
2618 * true to copy data from the allocators buffer to our internal buffer.
2619 * false for the other direction.
2620 */
2621DECLINLINE(int) drvvdMediaExIoReqBufSync(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, bool fToIoBuf)
2622{
2623 int rc = VINF_SUCCESS;
2624
2625 Assert(pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_WRITE);
2626 Assert(pIoReq->ReadWrite.cbIoBuf > 0);
2627
2628 /* Make sure the buffer is reset. */
2629 RTSgBufReset(&pIoReq->ReadWrite.IoBuf.SgBuf);
2630
2631 size_t const offSrc = pIoReq->ReadWrite.cbReq - pIoReq->ReadWrite.cbReqLeft;
2632 Assert((uint32_t)offSrc == offSrc);
2633 if (fToIoBuf)
2634 rc = pThis->pDrvMediaExPort->pfnIoReqCopyToBuf(pThis->pDrvMediaExPort, pIoReq, &pIoReq->abAlloc[0], (uint32_t)offSrc,
2635 &pIoReq->ReadWrite.IoBuf.SgBuf,
2636 RT_MIN(pIoReq->ReadWrite.cbIoBuf, pIoReq->ReadWrite.cbReqLeft));
2637 else
2638 rc = pThis->pDrvMediaExPort->pfnIoReqCopyFromBuf(pThis->pDrvMediaExPort, pIoReq, &pIoReq->abAlloc[0], (uint32_t)offSrc,
2639 &pIoReq->ReadWrite.IoBuf.SgBuf,
2640 (uint32_t)RT_MIN(pIoReq->ReadWrite.cbIoBuf, pIoReq->ReadWrite.cbReqLeft));
2641
2642 RTSgBufReset(&pIoReq->ReadWrite.IoBuf.SgBuf);
2643 return rc;
2644}
2645
2646/**
2647 * Hashes the I/O request ID to an index for the allocated I/O request bin.
2648 */
2649DECLINLINE(unsigned) drvvdMediaExIoReqIdHash(PDMMEDIAEXIOREQID uIoReqId)
2650{
2651 return uIoReqId % DRVVD_VDIOREQ_ALLOC_BINS; /** @todo Find something better? */
2652}
2653
2654/**
2655 * Inserts the given I/O request in to the list of allocated I/O requests.
2656 *
2657 * @returns VBox status code.
2658 * @param pThis VBox disk container instance data.
2659 * @param pIoReq I/O request to insert.
2660 */
2661static int drvvdMediaExIoReqInsert(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq)
2662{
2663 int rc = VINF_SUCCESS;
2664 unsigned idxBin = drvvdMediaExIoReqIdHash(pIoReq->uIoReqId);
2665
2666 rc = RTSemFastMutexRequest(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
2667 if (RT_SUCCESS(rc))
2668 {
2669 /* Search for conflicting I/O request ID. */
2670 PPDMMEDIAEXIOREQINT pIt;
2671 RTListForEach(&pThis->aIoReqAllocBins[idxBin].LstIoReqAlloc, pIt, PDMMEDIAEXIOREQINT, NdAllocatedList)
2672 {
2673 if (RT_UNLIKELY(pIt->uIoReqId == pIoReq->uIoReqId))
2674 {
2675 rc = VERR_PDM_MEDIAEX_IOREQID_CONFLICT;
2676 break;
2677 }
2678 }
2679 if (RT_SUCCESS(rc))
2680 RTListAppend(&pThis->aIoReqAllocBins[idxBin].LstIoReqAlloc, &pIoReq->NdAllocatedList);
2681 RTSemFastMutexRelease(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
2682 }
2683
2684 return rc;
2685}
2686
2687/**
2688 * Removes the given I/O request from the list of allocated I/O requests.
2689 *
2690 * @returns VBox status code.
2691 * @param pThis VBox disk container instance data.
2692 * @param pIoReq I/O request to insert.
2693 */
2694static int drvvdMediaExIoReqRemove(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq)
2695{
2696 int rc = VINF_SUCCESS;
2697 unsigned idxBin = drvvdMediaExIoReqIdHash(pIoReq->uIoReqId);
2698
2699 rc = RTSemFastMutexRequest(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
2700 if (RT_SUCCESS(rc))
2701 {
2702 RTListNodeRemove(&pIoReq->NdAllocatedList);
2703 RTSemFastMutexRelease(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
2704 }
2705
2706 return rc;
2707}
2708
2709/**
2710 * Retires a given I/O request marking it as complete and notiyfing the
2711 * device/driver above about the completion if requested.
2712 *
2713 * @returns VBox status code.
2714 * @param pThis VBox disk container instance data.
2715 * @param pIoReq I/O request to complete.
2716 * @param rcReq The status code the request completed with.
2717 * @param fUpNotify Flag whether to notify the driver/device above us about the completion.
2718 */
2719static void drvvdMediaExIoReqRetire(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, int rcReq, bool fUpNotify)
2720{
2721 LogFlowFunc(("pThis=%#p pIoReq=%#p rcReq=%Rrc fUpNotify=%RTbool\n",
2722 pThis, pIoReq, rcReq, fUpNotify));
2723
2724 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_COMPLETING, VDIOREQSTATE_ACTIVE);
2725 if (fXchg)
2726 ASMAtomicDecU32(&pThis->cIoReqsActive);
2727 else
2728 {
2729 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
2730 rcReq = VERR_PDM_MEDIAEX_IOREQ_CANCELED;
2731 }
2732
2733 ASMAtomicXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_COMPLETED);
2734 drvvdMediaExIoReqBufFree(pThis, pIoReq);
2735
2736 /*
2737 * Leave a release log entry if the request was active for more than 25 seconds
2738 * (30 seconds is the timeout of the guest).
2739 */
2740 uint64_t tsNow = RTTimeMilliTS();
2741 if (tsNow - pIoReq->tsSubmit >= 25 * 1000)
2742 {
2743 const char *pcszReq = NULL;
2744
2745 switch (pIoReq->enmType)
2746 {
2747 case PDMMEDIAEXIOREQTYPE_READ:
2748 pcszReq = "Read";
2749 break;
2750 case PDMMEDIAEXIOREQTYPE_WRITE:
2751 pcszReq = "Write";
2752 break;
2753 case PDMMEDIAEXIOREQTYPE_FLUSH:
2754 pcszReq = "Flush";
2755 break;
2756 case PDMMEDIAEXIOREQTYPE_DISCARD:
2757 pcszReq = "Discard";
2758 break;
2759 default:
2760 pcszReq = "<Invalid>";
2761 }
2762
2763 LogRel(("VD#%u: %s request was active for %llu seconds\n",
2764 pThis->pDrvIns->iInstance, pcszReq, (tsNow - pIoReq->tsSubmit) / 1000));
2765 }
2766
2767 if (RT_FAILURE(rcReq))
2768 {
2769 /* Log the error. */
2770 if (pThis->cErrors++ < DRVVD_MAX_LOG_REL_ERRORS)
2771 {
2772 if (rcReq == VERR_PDM_MEDIAEX_IOREQ_CANCELED)
2773 {
2774 if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_FLUSH)
2775 LogRel(("VD#%u: Aborted flush returned rc=%Rrc\n",
2776 pThis->pDrvIns->iInstance, rcReq));
2777 else if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD)
2778 LogRel(("VD#%u: Aborted discard returned rc=%Rrc\n",
2779 pThis->pDrvIns->iInstance, rcReq));
2780 else
2781 LogRel(("VD#%u: Aborted %s (%u bytes left) returned rc=%Rrc\n",
2782 pThis->pDrvIns->iInstance,
2783 pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ
2784 ? "read"
2785 : "write",
2786 pIoReq->ReadWrite.cbReqLeft, rcReq));
2787 }
2788 else
2789 {
2790 if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_FLUSH)
2791 LogRel(("VD#%u: Flush returned rc=%Rrc\n",
2792 pThis->pDrvIns->iInstance, rcReq));
2793 else if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD)
2794 LogRel(("VD#%u: Discard returned rc=%Rrc\n",
2795 pThis->pDrvIns->iInstance, rcReq));
2796 else
2797 LogRel(("VD#%u: %s (%u bytes left) returned rc=%Rrc\n",
2798 pThis->pDrvIns->iInstance,
2799 pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ
2800 ? "Read"
2801 : "Write",
2802 pIoReq->ReadWrite.cbReqLeft, rcReq));
2803 }
2804 }
2805 }
2806
2807 if (fUpNotify)
2808 {
2809 int rc = pThis->pDrvMediaExPort->pfnIoReqCompleteNotify(pThis->pDrvMediaExPort,
2810 pIoReq, &pIoReq->abAlloc[0], rcReq);
2811 AssertRC(rc);
2812 }
2813
2814 LogFlowFunc(("returns\n"));
2815}
2816
2817/**
2818 * I/O request completion worker.
2819 *
2820 * @returns VBox status code.
2821 * @param pThis VBox disk container instance data.
2822 * @param pIoReq I/O request to complete.
2823 * @param rcReq The status code the request completed with.
2824 * @param fUpNotify Flag whether to notify the driver/device above us about the completion.
2825 */
2826static int drvvdMediaExIoReqCompleteWorker(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, int rcReq, bool fUpNotify)
2827{
2828 LogFlowFunc(("pThis=%#p pIoReq=%#p rcReq=%Rrc fUpNotify=%RTbool\n",
2829 pThis, pIoReq, rcReq, fUpNotify));
2830
2831 /*
2832 * For a read we need to sync the memory before continuing to process
2833 * the request further.
2834 */
2835 if ( RT_SUCCESS(rcReq)
2836 && pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ)
2837 rcReq = drvvdMediaExIoReqBufSync(pThis, pIoReq, false /* fToIoBuf */);
2838
2839 /*
2840 * When the request owner instructs us to handle recoverable errors like full disks
2841 * do it. Mark the request as suspended, notify the owner and put the request on the
2842 * redo list.
2843 */
2844 if ( RT_FAILURE(rcReq)
2845 && (pIoReq->fFlags & PDMIMEDIAEX_F_SUSPEND_ON_RECOVERABLE_ERR)
2846 && drvvdMediaExIoReqIsRedoSetWarning(pThis, rcReq))
2847 {
2848 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_SUSPENDED, VDIOREQSTATE_ACTIVE);
2849 if (fXchg)
2850 {
2851 /* Put on redo list and adjust active request counter. */
2852 RTCritSectEnter(&pThis->CritSectIoReqRedo);
2853 RTListAppend(&pThis->LstIoReqRedo, &pIoReq->NdLstWait);
2854 RTCritSectLeave(&pThis->CritSectIoReqRedo);
2855 ASMAtomicDecU32(&pThis->cIoReqsActive);
2856 pThis->pDrvMediaExPort->pfnIoReqStateChanged(pThis->pDrvMediaExPort, pIoReq, &pIoReq->abAlloc[0],
2857 PDMMEDIAEXIOREQSTATE_SUSPENDED);
2858 rcReq = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
2859 }
2860 else
2861 {
2862 /* Request was canceled inbetween, so don't care and notify the owner about the completed request. */
2863 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
2864 drvvdMediaExIoReqRetire(pThis, pIoReq, rcReq, fUpNotify);
2865 }
2866 }
2867 else
2868 {
2869 /* Adjust the remaining amount to transfer. */
2870 Assert(pIoReq->ReadWrite.cbIoBuf > 0);
2871
2872 size_t cbReqIo = RT_MIN(pIoReq->ReadWrite.cbReqLeft, pIoReq->ReadWrite.cbIoBuf);
2873 pIoReq->ReadWrite.offStart += cbReqIo;
2874 pIoReq->ReadWrite.cbReqLeft -= cbReqIo;
2875
2876 if ( RT_FAILURE(rcReq)
2877 || !pIoReq->ReadWrite.cbReqLeft
2878 || ( pIoReq->enmType != PDMMEDIAEXIOREQTYPE_READ
2879 && pIoReq->enmType != PDMMEDIAEXIOREQTYPE_WRITE))
2880 drvvdMediaExIoReqRetire(pThis, pIoReq, rcReq, fUpNotify);
2881 else
2882 drvvdMediaExIoReqReadWriteProcess(pThis, pIoReq, fUpNotify);
2883 }
2884
2885 LogFlowFunc(("returns %Rrc\n", rcReq));
2886 return rcReq;
2887}
2888
2889
2890/**
2891 * Allocates a memory buffer suitable for I/O for the given request.
2892 *
2893 * @returns VBox status code.
2894 * @param VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS if there is no I/O memory available to allocate and
2895 * the request was placed on a waiting list.
2896 * @param pThis VBox disk container instance data.
2897 * @param pIoReq I/O request to allocate memory for.
2898 * @param cb Size of the buffer.
2899 */
2900DECLINLINE(int) drvvdMediaExIoReqBufAlloc(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, size_t cb)
2901{
2902 LogFlowFunc(("pThis=%#p pIoReq=%#p cb=%zu\n", pThis, pIoReq, cb));
2903
2904 int rc = IOBUFMgrAllocBuf(pThis->hIoBufMgr, &pIoReq->ReadWrite.IoBuf, cb, &pIoReq->ReadWrite.cbIoBuf);
2905 if (rc == VERR_NO_MEMORY)
2906 {
2907 LogFlowFunc(("Could not allocate memory for request, deferring\n"));
2908 RTCritSectEnter(&pThis->CritSectIoReqsIoBufWait);
2909 RTListAppend(&pThis->LstIoReqIoBufWait, &pIoReq->NdLstWait);
2910 RTCritSectLeave(&pThis->CritSectIoReqsIoBufWait);
2911 ASMAtomicIncU32(&pThis->cIoReqsWaiting);
2912 rc = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
2913 }
2914 else
2915 {
2916 LogFlowFunc(("Allocated %zu bytes of memory\n", pIoReq->ReadWrite.cbIoBuf));
2917 Assert(pIoReq->ReadWrite.cbIoBuf > 0);
2918 }
2919
2920 LogFlowFunc(("returns %Rrc\n", rc));
2921 return rc;
2922}
2923
2924/**
2925 * Wrapper around the various ways to read from the underlying medium (cache, async vs. sync).
2926 *
2927 * @returns VBox status code.
2928 * @param pThis VBox disk container instance data.
2929 * @param pIoReq I/O request to process.
2930 * @param cbReqIo Transfer size.
2931 * @param pcbReqIo Where to store the amount of transferred data.
2932 */
2933static int drvvdMediaExIoReqReadWrapper(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, size_t cbReqIo, size_t *pcbReqIo)
2934{
2935 int rc = VINF_SUCCESS;
2936
2937 LogFlowFunc(("pThis=%#p pIoReq=%#p cbReqIo=%zu pcbReqIo=%#p\n", pThis, pIoReq, cbReqIo, pcbReqIo));
2938
2939 Assert(cbReqIo > 0);
2940
2941 if ( pThis->fAsyncIOSupported
2942 && !(pIoReq->fFlags & PDMIMEDIAEX_F_SYNC))
2943 {
2944 if (pThis->pBlkCache)
2945 {
2946 rc = PDMR3BlkCacheRead(pThis->pBlkCache, pIoReq->ReadWrite.offStart,
2947 &pIoReq->ReadWrite.IoBuf.SgBuf, cbReqIo, pIoReq);
2948 if (rc == VINF_SUCCESS)
2949 rc = VINF_VD_ASYNC_IO_FINISHED;
2950 else if (rc == VINF_AIO_TASK_PENDING)
2951 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
2952 }
2953 else
2954 rc = VDAsyncRead(pThis->pDisk, pIoReq->ReadWrite.offStart, cbReqIo, &pIoReq->ReadWrite.IoBuf.SgBuf,
2955 drvvdMediaExIoReqComplete, pThis, pIoReq);
2956 }
2957 else
2958 {
2959 void *pvBuf = RTSgBufGetNextSegment(&pIoReq->ReadWrite.IoBuf.SgBuf, &cbReqIo);
2960
2961 Assert(cbReqIo > 0 && VALID_PTR(pvBuf));
2962 rc = VDRead(pThis->pDisk, pIoReq->ReadWrite.offStart, pvBuf, cbReqIo);
2963 if (RT_SUCCESS(rc))
2964 rc = VINF_VD_ASYNC_IO_FINISHED;
2965 }
2966
2967 *pcbReqIo = cbReqIo;
2968
2969 LogFlowFunc(("returns %Rrc *pcbReqIo=%zu\n", rc, *pcbReqIo));
2970 return rc;
2971}
2972
2973/**
2974 * Wrapper around the various ways to write to the underlying medium (cache, async vs. sync).
2975 *
2976 * @returns VBox status code.
2977 * @param pThis VBox disk container instance data.
2978 * @param pIoReq I/O request to process.
2979 * @param cbReqIo Transfer size.
2980 * @param pcbReqIo Where to store the amount of transferred data.
2981 */
2982static int drvvdMediaExIoReqWriteWrapper(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, size_t cbReqIo, size_t *pcbReqIo)
2983{
2984 int rc = VINF_SUCCESS;
2985
2986 Assert(cbReqIo > 0);
2987
2988 LogFlowFunc(("pThis=%#p pIoReq=%#p cbReqIo=%zu pcbReqIo=%#p\n", pThis, pIoReq, cbReqIo, pcbReqIo));
2989
2990 if ( pThis->fAsyncIOSupported
2991 && !(pIoReq->fFlags & PDMIMEDIAEX_F_SYNC))
2992 {
2993 if (pThis->pBlkCache)
2994 {
2995 rc = PDMR3BlkCacheWrite(pThis->pBlkCache, pIoReq->ReadWrite.offStart,
2996 &pIoReq->ReadWrite.IoBuf.SgBuf, cbReqIo, pIoReq);
2997 if (rc == VINF_SUCCESS)
2998 rc = VINF_VD_ASYNC_IO_FINISHED;
2999 else if (rc == VINF_AIO_TASK_PENDING)
3000 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
3001 }
3002 else
3003 rc = VDAsyncWrite(pThis->pDisk, pIoReq->ReadWrite.offStart, cbReqIo, &pIoReq->ReadWrite.IoBuf.SgBuf,
3004 drvvdMediaExIoReqComplete, pThis, pIoReq);
3005 }
3006 else
3007 {
3008 void *pvBuf = RTSgBufGetNextSegment(&pIoReq->ReadWrite.IoBuf.SgBuf, &cbReqIo);
3009
3010 Assert(cbReqIo > 0 && VALID_PTR(pvBuf));
3011 rc = VDWrite(pThis->pDisk, pIoReq->ReadWrite.offStart, pvBuf, cbReqIo);
3012 if (RT_SUCCESS(rc))
3013 rc = VINF_VD_ASYNC_IO_FINISHED;
3014 }
3015
3016 *pcbReqIo = cbReqIo;
3017
3018 LogFlowFunc(("returns %Rrc *pcbReqIo=%zu\n", rc, *pcbReqIo));
3019 return rc;
3020}
3021
3022/**
3023 * Wrapper around the various ways to flush all data to the underlying medium (cache, async vs. sync).
3024 *
3025 * @returns VBox status code.
3026 * @param pThis VBox disk container instance data.
3027 * @param pIoReq I/O request to process.
3028 */
3029static int drvvdMediaExIoReqFlushWrapper(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq)
3030{
3031 int rc = VINF_SUCCESS;
3032
3033 LogFlowFunc(("pThis=%#p pIoReq=%#p\n", pThis, pIoReq));
3034
3035 if ( pThis->fAsyncIOSupported
3036 && !(pIoReq->fFlags & PDMIMEDIAEX_F_SYNC))
3037 {
3038 if (pThis->pBlkCache)
3039 {
3040 rc = PDMR3BlkCacheFlush(pThis->pBlkCache, pIoReq);
3041 if (rc == VINF_SUCCESS)
3042 rc = VINF_VD_ASYNC_IO_FINISHED;
3043 else if (rc == VINF_AIO_TASK_PENDING)
3044 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
3045 }
3046 else
3047 rc = VDAsyncFlush(pThis->pDisk, drvvdMediaExIoReqComplete, pThis, pIoReq);
3048 }
3049 else
3050 {
3051 rc = VDFlush(pThis->pDisk);
3052 if (RT_SUCCESS(rc))
3053 rc = VINF_VD_ASYNC_IO_FINISHED;
3054 }
3055
3056 LogFlowFunc(("returns %Rrc\n", rc));
3057 return rc;
3058}
3059
3060/**
3061 * Wrapper around the various ways to discard data blocks on the underlying medium (cache, async vs. sync).
3062 *
3063 * @returns VBox status code.
3064 * @param pThis VBox disk container instance data.
3065 * @param pIoReq I/O request to process.
3066 */
3067static int drvvdMediaExIoReqDiscardWrapper(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq)
3068{
3069 int rc = VINF_SUCCESS;
3070
3071 LogFlowFunc(("pThis=%#p pIoReq=%#p\n", pThis, pIoReq));
3072
3073 if ( pThis->fAsyncIOSupported
3074 && !(pIoReq->fFlags & PDMIMEDIAEX_F_SYNC))
3075 {
3076 if (pThis->pBlkCache)
3077 {
3078 rc = PDMR3BlkCacheDiscard(pThis->pBlkCache, pIoReq->Discard.paRanges, pIoReq->Discard.cRanges, pIoReq);
3079 if (rc == VINF_SUCCESS)
3080 rc = VINF_VD_ASYNC_IO_FINISHED;
3081 else if (rc == VINF_AIO_TASK_PENDING)
3082 rc = VERR_VD_ASYNC_IO_IN_PROGRESS;
3083 }
3084 else
3085 rc = VDAsyncDiscardRanges(pThis->pDisk, pIoReq->Discard.paRanges, pIoReq->Discard.cRanges,
3086 drvvdMediaExIoReqComplete, pThis, pIoReq);
3087 }
3088 else
3089 {
3090 rc = VDDiscardRanges(pThis->pDisk, pIoReq->Discard.paRanges, pIoReq->Discard.cRanges);
3091 if (RT_SUCCESS(rc))
3092 rc = VINF_VD_ASYNC_IO_FINISHED;
3093 }
3094
3095 LogFlowFunc(("returns %Rrc\n", rc));
3096 return rc;
3097}
3098
3099/**
3100 * Processes a read/write request.
3101 *
3102 * @returns VBox status code.
3103 * @param pThis VBox disk container instance data.
3104 * @param pIoReq I/O request to process.
3105 * @param fUpNotify Flag whether to notify the driver/device above us about the completion.
3106 */
3107static int drvvdMediaExIoReqReadWriteProcess(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq, bool fUpNotify)
3108{
3109 int rc = VINF_SUCCESS;
3110
3111 LogFlowFunc(("pThis=%#p pIoReq=%#p fUpNotify=%RTbool\n", pThis, pIoReq, fUpNotify));
3112
3113 Assert(pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_WRITE);
3114
3115 rc = drvvdKeyCheckPrereqs(pThis, false /* fSetError */);
3116
3117 while ( pIoReq->ReadWrite.cbReqLeft
3118 && rc == VINF_SUCCESS)
3119 {
3120 Assert(pIoReq->ReadWrite.cbIoBuf > 0);
3121
3122 size_t cbReqIo = RT_MIN(pIoReq->ReadWrite.cbReqLeft, pIoReq->ReadWrite.cbIoBuf);
3123
3124 if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ)
3125 rc = drvvdMediaExIoReqReadWrapper(pThis, pIoReq, cbReqIo, &cbReqIo);
3126 else
3127 {
3128 /* Sync memory buffer from the request initiator. */
3129 rc = drvvdMediaExIoReqBufSync(pThis, pIoReq, true /* fToIoBuf */);
3130 if (RT_SUCCESS(rc))
3131 rc = drvvdMediaExIoReqWriteWrapper(pThis, pIoReq, cbReqIo, &cbReqIo);
3132 }
3133
3134 if (rc == VERR_VD_ASYNC_IO_IN_PROGRESS)
3135 rc = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
3136 else if (rc == VINF_VD_ASYNC_IO_FINISHED)
3137 {
3138 if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ)
3139 rc = drvvdMediaExIoReqBufSync(pThis, pIoReq, false /* fToIoBuf */);
3140 else
3141 rc = VINF_SUCCESS;
3142 pIoReq->ReadWrite.offStart += cbReqIo;
3143 pIoReq->ReadWrite.cbReqLeft -= cbReqIo;
3144 }
3145 }
3146
3147 if (rc != VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS)
3148 {
3149 Assert(!pIoReq->ReadWrite.cbReqLeft || RT_FAILURE(rc));
3150 rc = drvvdMediaExIoReqCompleteWorker(pThis, pIoReq, rc, fUpNotify);
3151 }
3152
3153 LogFlowFunc(("returns %Rrc\n", rc));
3154 return rc;
3155}
3156
3157
3158/**
3159 * Frees a I/O memory buffer allocated previously.
3160 *
3161 * @returns nothing.
3162 * @param pThis VBox disk container instance data.
3163 * @param pIoReq I/O request for which to free memory.
3164 */
3165DECLINLINE(void) drvvdMediaExIoReqBufFree(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq)
3166{
3167 LogFlowFunc(("pThis=%#p pIoReq=%#p{.cbIoBuf=%zu}\n", pThis, pIoReq, pIoReq->ReadWrite.cbIoBuf));
3168
3169 if ( pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ
3170 || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_WRITE)
3171 {
3172 IOBUFMgrFreeBuf(&pIoReq->ReadWrite.IoBuf);
3173
3174 uint32_t cIoReqsWaiting = ASMAtomicXchgU32(&pThis->cIoReqsWaiting, 0);
3175 if (cIoReqsWaiting > 0)
3176 {
3177 /* Try to process as many requests as possible. */
3178 RTCritSectEnter(&pThis->CritSectIoReqsIoBufWait);
3179 PPDMMEDIAEXIOREQINT pIoReqCur, pIoReqNext;
3180
3181 RTListForEachSafe(&pThis->LstIoReqIoBufWait, pIoReqCur, pIoReqNext, PDMMEDIAEXIOREQINT, NdLstWait)
3182 {
3183 LogFlowFunc(("Found I/O request %#p on waiting list, trying to allocate buffer of size %zu bytes\n",
3184 pIoReqCur, pIoReqCur->ReadWrite.cbReq));
3185
3186 /* Allocate a suitable I/O buffer for this request. */
3187 int rc = IOBUFMgrAllocBuf(pThis->hIoBufMgr, &pIoReqCur->ReadWrite.IoBuf, pIoReqCur->ReadWrite.cbReq,
3188 &pIoReqCur->ReadWrite.cbIoBuf);
3189 if (rc == VINF_SUCCESS)
3190 {
3191 Assert(pIoReq->ReadWrite.cbIoBuf > 0);
3192
3193 cIoReqsWaiting--;
3194 RTListNodeRemove(&pIoReqCur->NdLstWait);
3195
3196 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReqCur->enmState, VDIOREQSTATE_ACTIVE, VDIOREQSTATE_ALLOCATED);
3197 if (RT_UNLIKELY(!fXchg))
3198 {
3199 /* Must have been canceled inbetween. */
3200 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
3201 drvvdMediaExIoReqCompleteWorker(pThis, pIoReqCur, VERR_PDM_MEDIAEX_IOREQ_CANCELED, true /* fUpNotify */);
3202 }
3203 ASMAtomicIncU32(&pThis->cIoReqsActive);
3204 rc = drvvdMediaExIoReqReadWriteProcess(pThis, pIoReqCur, true /* fUpNotify */);
3205 }
3206 else
3207 {
3208 Assert(rc == VERR_NO_MEMORY);
3209 break;
3210 }
3211 }
3212 RTCritSectLeave(&pThis->CritSectIoReqsIoBufWait);
3213
3214 ASMAtomicAddU32(&pThis->cIoReqsWaiting, cIoReqsWaiting);
3215 }
3216 }
3217
3218 LogFlowFunc(("returns\n"));
3219}
3220
3221
3222/**
3223 * Returns whether the VM is in a running state.
3224 *
3225 * @returns Flag indicating whether the VM is currently in a running state.
3226 * @param pThis VBox disk container instance data.
3227 */
3228DECLINLINE(bool) drvvdMediaExIoReqIsVmRunning(PVBOXDISK pThis)
3229{
3230 VMSTATE enmVmState = PDMDrvHlpVMState(pThis->pDrvIns);
3231 if ( enmVmState == VMSTATE_RESUMING
3232 || enmVmState == VMSTATE_RUNNING
3233 || enmVmState == VMSTATE_RUNNING_LS
3234 || enmVmState == VMSTATE_RUNNING_FT
3235 || enmVmState == VMSTATE_RESETTING
3236 || enmVmState == VMSTATE_RESETTING_LS
3237 || enmVmState == VMSTATE_SOFT_RESETTING
3238 || enmVmState == VMSTATE_SOFT_RESETTING_LS
3239 || enmVmState == VMSTATE_SUSPENDING
3240 || enmVmState == VMSTATE_SUSPENDING_LS
3241 || enmVmState == VMSTATE_SUSPENDING_EXT_LS)
3242 return true;
3243
3244 return false;
3245}
3246
3247/**
3248 * @copydoc FNVDASYNCTRANSFERCOMPLETE
3249 */
3250static DECLCALLBACK(void) drvvdMediaExIoReqComplete(void *pvUser1, void *pvUser2, int rcReq)
3251{
3252 PVBOXDISK pThis = (PVBOXDISK)pvUser1;
3253 PPDMMEDIAEXIOREQINT pIoReq = (PPDMMEDIAEXIOREQINT)pvUser2;
3254
3255 drvvdMediaExIoReqCompleteWorker(pThis, pIoReq, rcReq, true /* fUpNotify */);
3256}
3257
3258/**
3259 * Tries to cancel the given I/O request returning the result.
3260 *
3261 * @returns Flag whether the request was successfully canceled or whether it
3262 * already complete inbetween.
3263 * @param pThis VBox disk container instance data.
3264 * @param pIoReq The I/O request to cancel.
3265 */
3266static bool drvvdMediaExIoReqCancel(PVBOXDISK pThis, PPDMMEDIAEXIOREQINT pIoReq)
3267{
3268 bool fXchg = true;
3269 VDIOREQSTATE enmStateOld = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3270
3271 /*
3272 * We might have to try canceling the request multiple times if it transitioned from
3273 * ALLOCATED to ACTIVE or to SUSPENDED between reading the state and trying to change it.
3274 */
3275 while ( ( enmStateOld == VDIOREQSTATE_ALLOCATED
3276 || enmStateOld == VDIOREQSTATE_ACTIVE
3277 || enmStateOld == VDIOREQSTATE_SUSPENDED)
3278 && !fXchg)
3279 {
3280 fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_CANCELED, enmStateOld);
3281 if (!fXchg)
3282 enmStateOld = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3283 }
3284
3285 if (fXchg)
3286 ASMAtomicDecU32(&pThis->cIoReqsActive);
3287
3288 return fXchg;
3289}
3290
3291/**
3292 * @interface_method_impl{PDMIMEDIAEX,pfnQueryFeatures}
3293 */
3294static DECLCALLBACK(int) drvvdQueryFeatures(PPDMIMEDIAEX pInterface, uint32_t *pfFeatures)
3295{
3296 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3297
3298 AssertPtrReturn(pfFeatures, VERR_INVALID_POINTER);
3299
3300 uint32_t fFeatures = 0;
3301 if (pThis->fAsyncIOSupported)
3302 fFeatures |= PDMIMEDIAEX_FEATURE_F_ASYNC;
3303 if (pThis->IMedia.pfnDiscard)
3304 fFeatures |= PDMIMEDIAEX_FEATURE_F_DISCARD;
3305
3306 *pfFeatures = fFeatures;
3307
3308 return VINF_SUCCESS;
3309}
3310
3311
3312/**
3313 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqAllocSizeSet}
3314 */
3315static DECLCALLBACK(int) drvvdIoReqAllocSizeSet(PPDMIMEDIAEX pInterface, size_t cbIoReqAlloc)
3316{
3317 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3318 if (RT_UNLIKELY(pThis->hIoReqCache != NIL_RTMEMCACHE))
3319 return VERR_INVALID_STATE;
3320
3321 return RTMemCacheCreate(&pThis->hIoReqCache, sizeof(PDMMEDIAEXIOREQINT) + cbIoReqAlloc, 0, UINT32_MAX,
3322 NULL, NULL, NULL, 0);
3323}
3324
3325/**
3326 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqAlloc}
3327 */
3328static DECLCALLBACK(int) drvvdIoReqAlloc(PPDMIMEDIAEX pInterface, PPDMMEDIAEXIOREQ phIoReq, void **ppvIoReqAlloc,
3329 PDMMEDIAEXIOREQID uIoReqId, uint32_t fFlags)
3330{
3331 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3332
3333 AssertReturn(!(fFlags & ~PDMIMEDIAEX_F_VALID), VERR_INVALID_PARAMETER);
3334
3335 PPDMMEDIAEXIOREQINT pIoReq = (PPDMMEDIAEXIOREQINT)RTMemCacheAlloc(pThis->hIoReqCache);
3336
3337 if (RT_UNLIKELY(!pIoReq))
3338 return VERR_NO_MEMORY;
3339
3340 pIoReq->uIoReqId = uIoReqId;
3341 pIoReq->fFlags = fFlags;
3342 pIoReq->pDisk = pThis;
3343 pIoReq->enmState = VDIOREQSTATE_ALLOCATED;
3344 pIoReq->enmType = PDMMEDIAEXIOREQTYPE_INVALID;
3345
3346 int rc = drvvdMediaExIoReqInsert(pThis, pIoReq);
3347 if (RT_SUCCESS(rc))
3348 {
3349 *phIoReq = pIoReq;
3350 *ppvIoReqAlloc = &pIoReq->abAlloc[0];
3351 }
3352 else
3353 RTMemCacheFree(pThis->hIoReqCache, pIoReq);
3354
3355 return rc;
3356}
3357
3358/**
3359 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqFree}
3360 */
3361static DECLCALLBACK(int) drvvdIoReqFree(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq)
3362{
3363 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3364 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3365
3366 if ( pIoReq->enmState != VDIOREQSTATE_COMPLETED
3367 && pIoReq->enmState != VDIOREQSTATE_ALLOCATED)
3368 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3369
3370 /* Remove from allocated list. */
3371 int rc = drvvdMediaExIoReqRemove(pThis, pIoReq);
3372 if (RT_FAILURE(rc))
3373 return rc;
3374
3375 /* Free any associated I/O memory. */
3376 drvvdMediaExIoReqBufFree(pThis, pIoReq);
3377
3378 /* For discard request discard the range array. */
3379 if ( pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD
3380 && pIoReq->Discard.paRanges)
3381 {
3382 RTMemFree(pIoReq->Discard.paRanges);
3383 pIoReq->Discard.paRanges = NULL;
3384 }
3385
3386 pIoReq->enmState = VDIOREQSTATE_FREE;
3387 RTMemCacheFree(pThis->hIoReqCache, pIoReq);
3388 return VINF_SUCCESS;
3389}
3390
3391/**
3392 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqQueryResidual}
3393 */
3394static DECLCALLBACK(int) drvvdIoReqQueryResidual(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq, size_t *pcbResidual)
3395{
3396 RT_NOREF1(pInterface);
3397
3398 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3399
3400 if (pIoReq->enmState != VDIOREQSTATE_COMPLETED)
3401 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3402
3403 if ( pIoReq->enmType != PDMMEDIAEXIOREQTYPE_READ
3404 && pIoReq->enmType != PDMMEDIAEXIOREQTYPE_WRITE)
3405 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3406
3407 *pcbResidual = 0; /* No data left to transfer always. */
3408 return VINF_SUCCESS;
3409}
3410
3411/**
3412 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqCancelAll}
3413 */
3414static DECLCALLBACK(int) drvvdIoReqCancelAll(PPDMIMEDIAEX pInterface)
3415{
3416 int rc = VINF_SUCCESS;
3417 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3418
3419 for (unsigned idxBin = 0; idxBin < RT_ELEMENTS(pThis->aIoReqAllocBins); idxBin++)
3420 {
3421 rc = RTSemFastMutexRequest(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
3422 if (RT_SUCCESS(rc))
3423 {
3424 /* Search for I/O request with ID. */
3425 PPDMMEDIAEXIOREQINT pIt;
3426
3427 RTListForEach(&pThis->aIoReqAllocBins[idxBin].LstIoReqAlloc, pIt, PDMMEDIAEXIOREQINT, NdAllocatedList)
3428 {
3429 drvvdMediaExIoReqCancel(pThis, pIt);
3430 }
3431 RTSemFastMutexRelease(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
3432 }
3433 }
3434
3435 return rc;
3436}
3437
3438/**
3439 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqCancel}
3440 */
3441static DECLCALLBACK(int) drvvdIoReqCancel(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQID uIoReqId)
3442{
3443 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3444 unsigned idxBin = drvvdMediaExIoReqIdHash(uIoReqId);
3445
3446 int rc = RTSemFastMutexRequest(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
3447 if (RT_SUCCESS(rc))
3448 {
3449 /* Search for I/O request with ID. */
3450 PPDMMEDIAEXIOREQINT pIt;
3451 rc = VERR_PDM_MEDIAEX_IOREQID_NOT_FOUND;
3452
3453 RTListForEach(&pThis->aIoReqAllocBins[idxBin].LstIoReqAlloc, pIt, PDMMEDIAEXIOREQINT, NdAllocatedList)
3454 {
3455 if (pIt->uIoReqId == uIoReqId)
3456 {
3457 if (drvvdMediaExIoReqCancel(pThis, pIt))
3458 rc = VINF_SUCCESS;
3459
3460 break;
3461 }
3462 }
3463 RTSemFastMutexRelease(pThis->aIoReqAllocBins[idxBin].hMtxLstIoReqAlloc);
3464 }
3465
3466 return rc;
3467}
3468
3469/**
3470 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqRead}
3471 */
3472static DECLCALLBACK(int) drvvdIoReqRead(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq, uint64_t off, size_t cbRead)
3473{
3474 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3475 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3476 VDIOREQSTATE enmState = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3477
3478 if (RT_UNLIKELY(enmState == VDIOREQSTATE_CANCELED))
3479 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3480
3481 if (RT_UNLIKELY(enmState != VDIOREQSTATE_ALLOCATED))
3482 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3483
3484 pIoReq->enmType = PDMMEDIAEXIOREQTYPE_READ;
3485 pIoReq->tsSubmit = RTTimeMilliTS();
3486 pIoReq->ReadWrite.offStart = off;
3487 pIoReq->ReadWrite.cbReq = cbRead;
3488 pIoReq->ReadWrite.cbReqLeft = cbRead;
3489 /* Allocate a suitable I/O buffer for this request. */
3490 int rc = drvvdMediaExIoReqBufAlloc(pThis, pIoReq, cbRead);
3491 if (rc == VINF_SUCCESS)
3492 {
3493 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_ACTIVE, VDIOREQSTATE_ALLOCATED);
3494 if (RT_UNLIKELY(!fXchg))
3495 {
3496 /* Must have been canceled inbetween. */
3497 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
3498 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3499 }
3500 ASMAtomicIncU32(&pThis->cIoReqsActive);
3501
3502 rc = drvvdMediaExIoReqReadWriteProcess(pThis, pIoReq, false /* fUpNotify */);
3503 }
3504
3505 return rc;
3506}
3507
3508/**
3509 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqWrite}
3510 */
3511static DECLCALLBACK(int) drvvdIoReqWrite(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq, uint64_t off, size_t cbWrite)
3512{
3513 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3514 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3515 VDIOREQSTATE enmState = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3516
3517 if (RT_UNLIKELY(enmState == VDIOREQSTATE_CANCELED))
3518 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3519
3520 if (RT_UNLIKELY(enmState != VDIOREQSTATE_ALLOCATED))
3521 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3522
3523 pIoReq->enmType = PDMMEDIAEXIOREQTYPE_WRITE;
3524 pIoReq->tsSubmit = RTTimeMilliTS();
3525 pIoReq->ReadWrite.offStart = off;
3526 pIoReq->ReadWrite.cbReq = cbWrite;
3527 pIoReq->ReadWrite.cbReqLeft = cbWrite;
3528 /* Allocate a suitable I/O buffer for this request. */
3529 int rc = drvvdMediaExIoReqBufAlloc(pThis, pIoReq, cbWrite);
3530 if (rc == VINF_SUCCESS)
3531 {
3532 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_ACTIVE, VDIOREQSTATE_ALLOCATED);
3533 if (RT_UNLIKELY(!fXchg))
3534 {
3535 /* Must have been canceled inbetween. */
3536 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
3537 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3538 }
3539 ASMAtomicIncU32(&pThis->cIoReqsActive);
3540
3541 rc = drvvdMediaExIoReqReadWriteProcess(pThis, pIoReq, false /* fUpNotify */);
3542 }
3543
3544 return rc;
3545}
3546
3547/**
3548 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqFlush}
3549 */
3550static DECLCALLBACK(int) drvvdIoReqFlush(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq)
3551{
3552 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3553 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3554 VDIOREQSTATE enmState = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3555
3556 if (RT_UNLIKELY(enmState == VDIOREQSTATE_CANCELED))
3557 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3558
3559 if (RT_UNLIKELY(enmState != VDIOREQSTATE_ALLOCATED))
3560 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3561
3562 pIoReq->enmType = PDMMEDIAEXIOREQTYPE_FLUSH;
3563 pIoReq->tsSubmit = RTTimeMilliTS();
3564 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_ACTIVE, VDIOREQSTATE_ALLOCATED);
3565 if (RT_UNLIKELY(!fXchg))
3566 {
3567 /* Must have been canceled inbetween. */
3568 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
3569 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3570 }
3571
3572 ASMAtomicIncU32(&pThis->cIoReqsActive);
3573 int rc = drvvdMediaExIoReqFlushWrapper(pThis, pIoReq);
3574 if (rc == VERR_VD_ASYNC_IO_IN_PROGRESS)
3575 rc = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
3576 else if (rc == VINF_VD_ASYNC_IO_FINISHED)
3577 rc = VINF_SUCCESS;
3578
3579 if (rc != VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS)
3580 rc = drvvdMediaExIoReqCompleteWorker(pThis, pIoReq, rc, false /* fUpNotify */);
3581
3582 return rc;
3583}
3584
3585/**
3586 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqDiscard}
3587 */
3588static DECLCALLBACK(int) drvvdIoReqDiscard(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq, unsigned cRangesMax)
3589{
3590 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3591 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3592 VDIOREQSTATE enmState = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3593
3594 if (RT_UNLIKELY(enmState == VDIOREQSTATE_CANCELED))
3595 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3596
3597 if (RT_UNLIKELY(enmState != VDIOREQSTATE_ALLOCATED))
3598 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3599
3600 /* Copy the ranges over now, this can be optimized in the future. */
3601 pIoReq->Discard.paRanges = (PRTRANGE)RTMemAllocZ(cRangesMax * sizeof(RTRANGE));
3602 if (RT_UNLIKELY(!pIoReq->Discard.paRanges))
3603 return VERR_NO_MEMORY;
3604
3605 int rc = pThis->pDrvMediaExPort->pfnIoReqQueryDiscardRanges(pThis->pDrvMediaExPort, pIoReq, &pIoReq->abAlloc[0],
3606 0, cRangesMax, pIoReq->Discard.paRanges,
3607 &pIoReq->Discard.cRanges);
3608 if (RT_SUCCESS(rc))
3609 {
3610 pIoReq->enmType = PDMMEDIAEXIOREQTYPE_DISCARD;
3611 pIoReq->tsSubmit = RTTimeMilliTS();
3612 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_ACTIVE, VDIOREQSTATE_ALLOCATED);
3613 if (RT_UNLIKELY(!fXchg))
3614 {
3615 /* Must have been canceled inbetween. */
3616 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
3617 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3618 }
3619
3620 ASMAtomicIncU32(&pThis->cIoReqsActive);
3621 rc = drvvdMediaExIoReqDiscardWrapper(pThis, pIoReq);
3622 if (rc == VERR_VD_ASYNC_IO_IN_PROGRESS)
3623 rc = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
3624 else if (rc == VINF_VD_ASYNC_IO_FINISHED)
3625 rc = VINF_SUCCESS;
3626
3627 if (rc != VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS)
3628 rc = drvvdMediaExIoReqCompleteWorker(pThis, pIoReq, rc, false /* fUpNotify */);
3629 }
3630
3631 return rc;
3632}
3633
3634/**
3635 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqSendScsiCmd}
3636 */
3637static DECLCALLBACK(int) drvvdIoReqSendScsiCmd(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq, uint32_t uLun,
3638 const uint8_t *pbCdb, size_t cbCdb, PDMMEDIAEXIOREQSCSITXDIR enmTxDir,
3639 size_t cbBuf, uint8_t *pabSense, size_t cbSense, uint8_t *pu8ScsiSts,
3640 uint32_t cTimeoutMillies)
3641{
3642 RT_NOREF10(pInterface, uLun, pbCdb, cbCdb, enmTxDir, cbBuf, pabSense, cbSense, pu8ScsiSts, cTimeoutMillies);
3643 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3644 VDIOREQSTATE enmState = (VDIOREQSTATE)ASMAtomicReadU32((volatile uint32_t *)&pIoReq->enmState);
3645
3646 if (RT_UNLIKELY(enmState == VDIOREQSTATE_CANCELED))
3647 return VERR_PDM_MEDIAEX_IOREQ_CANCELED;
3648
3649 if (RT_UNLIKELY(enmState != VDIOREQSTATE_ALLOCATED))
3650 return VERR_PDM_MEDIAEX_IOREQ_INVALID_STATE;
3651
3652 return VERR_NOT_SUPPORTED;
3653}
3654
3655/**
3656 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqGetActiveCount}
3657 */
3658static DECLCALLBACK(uint32_t) drvvdIoReqGetActiveCount(PPDMIMEDIAEX pInterface)
3659{
3660 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3661 return ASMAtomicReadU32(&pThis->cIoReqsActive);
3662}
3663
3664/**
3665 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqGetSuspendedCount}
3666 */
3667static DECLCALLBACK(uint32_t) drvvdIoReqGetSuspendedCount(PPDMIMEDIAEX pInterface)
3668{
3669 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3670
3671 AssertReturn(!drvvdMediaExIoReqIsVmRunning(pThis), 0);
3672
3673 uint32_t cIoReqSuspended = 0;
3674 PPDMMEDIAEXIOREQINT pIoReq;
3675 RTCritSectEnter(&pThis->CritSectIoReqRedo);
3676 RTListForEach(&pThis->LstIoReqRedo, pIoReq, PDMMEDIAEXIOREQINT, NdLstWait)
3677 {
3678 cIoReqSuspended++;
3679 }
3680 RTCritSectLeave(&pThis->CritSectIoReqRedo);
3681
3682 return cIoReqSuspended;
3683}
3684
3685/**
3686 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqQuerySuspendedFirst}
3687 */
3688static DECLCALLBACK(int) drvvdIoReqQuerySuspendedStart(PPDMIMEDIAEX pInterface, PPDMMEDIAEXIOREQ phIoReq,
3689 void **ppvIoReqAlloc)
3690{
3691 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3692
3693 AssertReturn(!drvvdMediaExIoReqIsVmRunning(pThis), VERR_INVALID_STATE);
3694 AssertReturn(!RTListIsEmpty(&pThis->LstIoReqRedo), VERR_NOT_FOUND);
3695
3696 RTCritSectEnter(&pThis->CritSectIoReqRedo);
3697 PPDMMEDIAEXIOREQINT pIoReq = RTListGetFirst(&pThis->LstIoReqRedo, PDMMEDIAEXIOREQINT, NdLstWait);
3698 *phIoReq = pIoReq;
3699 *ppvIoReqAlloc = &pIoReq->abAlloc[0];
3700 RTCritSectLeave(&pThis->CritSectIoReqRedo);
3701
3702 return VINF_SUCCESS;
3703}
3704
3705/**
3706 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqQuerySuspendedNext}
3707 */
3708static DECLCALLBACK(int) drvvdIoReqQuerySuspendedNext(PPDMIMEDIAEX pInterface, PDMMEDIAEXIOREQ hIoReq,
3709 PPDMMEDIAEXIOREQ phIoReqNext, void **ppvIoReqAllocNext)
3710{
3711 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3712 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3713
3714 AssertReturn(!drvvdMediaExIoReqIsVmRunning(pThis), VERR_INVALID_STATE);
3715 AssertPtrReturn(pIoReq, VERR_INVALID_HANDLE);
3716 AssertReturn(!RTListNodeIsLast(&pThis->LstIoReqRedo, &pIoReq->NdLstWait), VERR_NOT_FOUND);
3717
3718 RTCritSectEnter(&pThis->CritSectIoReqRedo);
3719 PPDMMEDIAEXIOREQINT pIoReqNext = RTListNodeGetNext(&pIoReq->NdLstWait, PDMMEDIAEXIOREQINT, NdLstWait);
3720 *phIoReqNext = pIoReqNext;
3721 *ppvIoReqAllocNext = &pIoReqNext->abAlloc[0];
3722 RTCritSectLeave(&pThis->CritSectIoReqRedo);
3723
3724 return VINF_SUCCESS;
3725}
3726
3727/**
3728 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqSuspendedSave}
3729 */
3730static DECLCALLBACK(int) drvvdIoReqSuspendedSave(PPDMIMEDIAEX pInterface, PSSMHANDLE pSSM, PDMMEDIAEXIOREQ hIoReq)
3731{
3732 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3733 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3734
3735 AssertReturn(!drvvdMediaExIoReqIsVmRunning(pThis), VERR_INVALID_STATE);
3736 AssertPtrReturn(pIoReq, VERR_INVALID_HANDLE);
3737 AssertReturn(pIoReq->enmState == VDIOREQSTATE_SUSPENDED, VERR_INVALID_STATE);
3738
3739 SSMR3PutU32(pSSM, DRVVD_IOREQ_SAVED_STATE_VERSION);
3740 SSMR3PutU32(pSSM, (uint32_t)pIoReq->enmType);
3741 SSMR3PutU32(pSSM, pIoReq->uIoReqId);
3742 SSMR3PutU32(pSSM, pIoReq->fFlags);
3743 if ( pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ
3744 || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_WRITE)
3745 {
3746 SSMR3PutU64(pSSM, pIoReq->ReadWrite.offStart);
3747 SSMR3PutU64(pSSM, pIoReq->ReadWrite.cbReq);
3748 SSMR3PutU64(pSSM, pIoReq->ReadWrite.cbReqLeft);
3749 }
3750 else if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD)
3751 {
3752 SSMR3PutU32(pSSM, pIoReq->Discard.cRanges);
3753 for (unsigned i = 0; i < pIoReq->Discard.cRanges; i++)
3754 {
3755 SSMR3PutU64(pSSM, pIoReq->Discard.paRanges[i].offStart);
3756 SSMR3PutU64(pSSM, pIoReq->Discard.paRanges[i].cbRange);
3757 }
3758 }
3759
3760 return SSMR3PutU32(pSSM, UINT32_MAX); /* sanity/terminator */
3761}
3762
3763/**
3764 * @interface_method_impl{PDMIMEDIAEX,pfnIoReqSuspendedLoad}
3765 */
3766static DECLCALLBACK(int) drvvdIoReqSuspendedLoad(PPDMIMEDIAEX pInterface, PSSMHANDLE pSSM, PDMMEDIAEXIOREQ hIoReq)
3767{
3768 PVBOXDISK pThis = RT_FROM_MEMBER(pInterface, VBOXDISK, IMediaEx);
3769 PPDMMEDIAEXIOREQINT pIoReq = hIoReq;
3770
3771 AssertReturn(!drvvdMediaExIoReqIsVmRunning(pThis), VERR_INVALID_STATE);
3772 AssertPtrReturn(pIoReq, VERR_INVALID_HANDLE);
3773 AssertReturn(pIoReq->enmState == VDIOREQSTATE_ALLOCATED, VERR_INVALID_STATE);
3774
3775 uint32_t u32;
3776 uint64_t u64;
3777 int rc = VINF_SUCCESS;
3778 bool fPlaceOnRedoList = true;
3779
3780 SSMR3GetU32(pSSM, &u32);
3781 if (u32 <= DRVVD_IOREQ_SAVED_STATE_VERSION)
3782 {
3783 SSMR3GetU32(pSSM, &u32);
3784 AssertReturn( u32 == PDMMEDIAEXIOREQTYPE_WRITE
3785 || u32 == PDMMEDIAEXIOREQTYPE_READ
3786 || u32 == PDMMEDIAEXIOREQTYPE_DISCARD
3787 || u32 == PDMMEDIAEXIOREQTYPE_FLUSH,
3788 VERR_SSM_DATA_UNIT_FORMAT_CHANGED);
3789 pIoReq->enmType = (PDMMEDIAEXIOREQTYPE)u32;
3790
3791 SSMR3GetU32(pSSM, &u32);
3792 AssertReturn(u32 == pIoReq->uIoReqId, VERR_SSM_DATA_UNIT_FORMAT_CHANGED);
3793
3794 SSMR3GetU32(pSSM, &u32);
3795 AssertReturn(u32 == pIoReq->fFlags, VERR_SSM_DATA_UNIT_FORMAT_CHANGED);
3796
3797 if ( pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ
3798 || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_WRITE)
3799 {
3800 SSMR3GetU64(pSSM, &pIoReq->ReadWrite.offStart);
3801 SSMR3GetU64(pSSM, &u64);
3802 pIoReq->ReadWrite.cbReq = (size_t)u64;
3803 SSMR3GetU64(pSSM, &u64);
3804 pIoReq->ReadWrite.cbReqLeft = (size_t)u64;
3805
3806 /*
3807 * Try to allocate enough I/O buffer, if this fails for some reason put it onto the
3808 * waitign list instead of the redo list.
3809 */
3810 pIoReq->ReadWrite.cbIoBuf = 0;
3811 rc = IOBUFMgrAllocBuf(pThis->hIoBufMgr, &pIoReq->ReadWrite.IoBuf, pIoReq->ReadWrite.cbReqLeft,
3812 &pIoReq->ReadWrite.cbIoBuf);
3813 if (rc == VERR_NO_MEMORY)
3814 {
3815 pIoReq->enmState = VDIOREQSTATE_ALLOCATED;
3816 ASMAtomicIncU32(&pThis->cIoReqsWaiting);
3817 RTListAppend(&pThis->LstIoReqIoBufWait, &pIoReq->NdLstWait);
3818 fPlaceOnRedoList = false;
3819 rc = VINF_SUCCESS;
3820 }
3821 }
3822 else if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD)
3823 {
3824 rc = SSMR3GetU32(pSSM, &pIoReq->Discard.cRanges);
3825 if (RT_SUCCESS(rc))
3826 {
3827 pIoReq->Discard.paRanges = (PRTRANGE)RTMemAllocZ(pIoReq->Discard.cRanges * sizeof(RTRANGE));
3828 if (RT_LIKELY(pIoReq->Discard.paRanges))
3829 {
3830 for (unsigned i = 0; i < pIoReq->Discard.cRanges; i++)
3831 {
3832 SSMR3GetU64(pSSM, &pIoReq->Discard.paRanges[i].offStart);
3833 SSMR3GetU64(pSSM, &u64);
3834 pIoReq->Discard.paRanges[i].cbRange = (size_t)u64;
3835 }
3836 }
3837 else
3838 rc = VERR_NO_MEMORY;
3839 }
3840 }
3841
3842 if (RT_SUCCESS(rc))
3843 rc = SSMR3GetU32(pSSM, &u32); /* sanity/terminator */
3844 if (RT_SUCCESS(rc))
3845 AssertReturn(u32 == UINT32_MAX, VERR_SSM_DATA_UNIT_FORMAT_CHANGED);
3846 if ( RT_SUCCESS(rc)
3847 && fPlaceOnRedoList)
3848 {
3849 /* Mark as suspended */
3850 pIoReq->enmState = VDIOREQSTATE_SUSPENDED;
3851
3852 /* Link into suspended list so it gets kicked off again when we resume. */
3853 RTCritSectEnter(&pThis->CritSectIoReqRedo);
3854 RTListAppend(&pThis->LstIoReqRedo, &pIoReq->NdLstWait);
3855 RTCritSectLeave(&pThis->CritSectIoReqRedo);
3856 }
3857 }
3858
3859 return rc;
3860}
3861
3862/**
3863 * Loads all configured plugins.
3864 *
3865 * @returns VBox status code.
3866 * @param pCfg CFGM node holding plugin list.
3867 */
3868static int drvvdLoadPlugins(PCFGMNODE pCfg)
3869{
3870 PCFGMNODE pCfgPlugins = CFGMR3GetChild(pCfg, "Plugins");
3871
3872 if (pCfgPlugins)
3873 {
3874 PCFGMNODE pPluginCur = CFGMR3GetFirstChild(pCfgPlugins);
3875 while (pPluginCur)
3876 {
3877 int rc = VINF_SUCCESS;
3878 char *pszPluginFilename = NULL;
3879 rc = CFGMR3QueryStringAlloc(pPluginCur, "Path", &pszPluginFilename);
3880 if (RT_SUCCESS(rc))
3881 rc = VDPluginLoadFromFilename(pszPluginFilename);
3882
3883 if (RT_FAILURE(rc))
3884 LogRel(("VD: Failed to load plugin '%s' with %Rrc, continuing\n", pszPluginFilename, rc));
3885
3886 pPluginCur = CFGMR3GetNextChild(pPluginCur);
3887 }
3888 }
3889
3890 return VINF_SUCCESS;
3891}
3892
3893
3894/**
3895 * Sets up the disk filter chain.
3896 *
3897 * @returns VBox status code.
3898 * @param pThis The disk instance.
3899 * @param pCfg CFGM node holding the filter parameters.
3900 */
3901static int drvvdSetupFilters(PVBOXDISK pThis, PCFGMNODE pCfg)
3902{
3903 int rc = VINF_SUCCESS;
3904 PCFGMNODE pCfgFilter = CFGMR3GetChild(pCfg, "Filters");
3905
3906 if (pCfgFilter)
3907 {
3908 PCFGMNODE pCfgFilterConfig = CFGMR3GetChild(pCfgFilter, "VDConfig");
3909 char *pszFilterName = NULL;
3910 VDINTERFACECONFIG VDIfConfig;
3911 PVDINTERFACE pVDIfsFilter = NULL;
3912
3913 rc = CFGMR3QueryStringAlloc(pCfgFilter, "FilterName", &pszFilterName);
3914 if (RT_SUCCESS(rc))
3915 {
3916 VDIfConfig.pfnAreKeysValid = drvvdCfgAreKeysValid;
3917 VDIfConfig.pfnQuerySize = drvvdCfgQuerySize;
3918 VDIfConfig.pfnQuery = drvvdCfgQuery;
3919 VDIfConfig.pfnQueryBytes = drvvdCfgQueryBytes;
3920 rc = VDInterfaceAdd(&VDIfConfig.Core, "DrvVD_Config", VDINTERFACETYPE_CONFIG,
3921 pCfgFilterConfig, sizeof(VDINTERFACECONFIG), &pVDIfsFilter);
3922 AssertRC(rc);
3923
3924 rc = VDFilterAdd(pThis->pDisk, pszFilterName, VD_FILTER_FLAGS_DEFAULT, pVDIfsFilter);
3925
3926 MMR3HeapFree(pszFilterName);
3927 }
3928 }
3929
3930 return rc;
3931}
3932
3933
3934/**
3935 * Translates a PDMMEDIATYPE value into a string.
3936 *
3937 * @returns Read only string.
3938 * @param enmType The type value.
3939 */
3940static const char *drvvdGetTypeName(PDMMEDIATYPE enmType)
3941{
3942 switch (enmType)
3943 {
3944 case PDMMEDIATYPE_ERROR: return "ERROR";
3945 case PDMMEDIATYPE_FLOPPY_360: return "FLOPPY_360";
3946 case PDMMEDIATYPE_FLOPPY_720: return "FLOPPY_720";
3947 case PDMMEDIATYPE_FLOPPY_1_20: return "FLOPPY_1_20";
3948 case PDMMEDIATYPE_FLOPPY_1_44: return "FLOPPY_1_44";
3949 case PDMMEDIATYPE_FLOPPY_2_88: return "FLOPPY_2_88";
3950 case PDMMEDIATYPE_FLOPPY_FAKE_15_6: return "FLOPPY_FAKE_15_6";
3951 case PDMMEDIATYPE_FLOPPY_FAKE_63_5: return "FLOPPY_FAKE_63_5";
3952 case PDMMEDIATYPE_CDROM: return "CDROM";
3953 case PDMMEDIATYPE_DVD: return "DVD";
3954 case PDMMEDIATYPE_HARD_DISK: return "HARD_DISK";
3955 default: return "Unknown";
3956 }
3957}
3958
3959/**
3960 * Returns the appropriate PDMMEDIATYPE for t he given string.
3961 *
3962 * @returns PDMMEDIATYPE
3963 * @param pszType The string representation of the media type.
3964 */
3965static PDMMEDIATYPE drvvdGetMediaTypeFromString(const char *pszType)
3966{
3967 PDMMEDIATYPE enmType = PDMMEDIATYPE_ERROR;
3968
3969 if (!strcmp(pszType, "HardDisk"))
3970 enmType = PDMMEDIATYPE_HARD_DISK;
3971 else if (!strcmp(pszType, "DVD"))
3972 enmType = PDMMEDIATYPE_DVD;
3973 else if (!strcmp(pszType, "CDROM"))
3974 enmType = PDMMEDIATYPE_CDROM;
3975 else if (!strcmp(pszType, "Floppy 2.88"))
3976 enmType = PDMMEDIATYPE_FLOPPY_2_88;
3977 else if (!strcmp(pszType, "Floppy 1.44"))
3978 enmType = PDMMEDIATYPE_FLOPPY_1_44;
3979 else if (!strcmp(pszType, "Floppy 1.20"))
3980 enmType = PDMMEDIATYPE_FLOPPY_1_20;
3981 else if (!strcmp(pszType, "Floppy 720"))
3982 enmType = PDMMEDIATYPE_FLOPPY_720;
3983 else if (!strcmp(pszType, "Floppy 360"))
3984 enmType = PDMMEDIATYPE_FLOPPY_360;
3985 else if (!strcmp(pszType, "Floppy 15.6"))
3986 enmType = PDMMEDIATYPE_FLOPPY_FAKE_15_6;
3987 else if (!strcmp(pszType, "Floppy 63.5"))
3988 enmType = PDMMEDIATYPE_FLOPPY_FAKE_63_5;
3989
3990 return enmType;
3991}
3992
3993/**
3994 * Converts PDMMEDIATYPE to the appropriate VDTYPE.
3995 *
3996 * @returns The VDTYPE.
3997 * @param enmType The PDMMEDIATYPE to convert from.
3998 */
3999static VDTYPE drvvdGetVDFromMediaType(PDMMEDIATYPE enmType)
4000{
4001 if (PDMMEDIATYPE_IS_FLOPPY(enmType))
4002 return VDTYPE_FLOPPY;
4003 else if (enmType == PDMMEDIATYPE_DVD || enmType == PDMMEDIATYPE_CDROM)
4004 return VDTYPE_DVD;
4005 else if (enmType == PDMMEDIATYPE_HARD_DISK)
4006 return VDTYPE_HDD;
4007
4008 AssertMsgFailed(("Invalid media type %d{%s} given!\n", enmType, drvvdGetTypeName(enmType)));
4009 return VDTYPE_HDD;
4010}
4011
4012
4013/*********************************************************************************************************************************
4014* Base interface methods *
4015*********************************************************************************************************************************/
4016
4017/**
4018 * @interface_method_impl{PDMIBASE,pfnQueryInterface}
4019 */
4020static DECLCALLBACK(void *) drvvdQueryInterface(PPDMIBASE pInterface, const char *pszIID)
4021{
4022 PPDMDRVINS pDrvIns = PDMIBASE_2_PDMDRV(pInterface);
4023 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4024
4025 PDMIBASE_RETURN_INTERFACE(pszIID, PDMIBASE, &pDrvIns->IBase);
4026 PDMIBASE_RETURN_INTERFACE(pszIID, PDMIMEDIA, &pThis->IMedia);
4027 PDMIBASE_RETURN_INTERFACE(pszIID, PDMIMOUNT, pThis->fMountable ? &pThis->IMount : NULL);
4028 PDMIBASE_RETURN_INTERFACE(pszIID, PDMIMEDIAEX, pThis->pDrvMediaExPort ? &pThis->IMediaEx : NULL);
4029 return NULL;
4030}
4031
4032
4033/*********************************************************************************************************************************
4034* Saved state notification methods *
4035*********************************************************************************************************************************/
4036
4037/**
4038 * Load done callback for re-opening the image writable during teleportation.
4039 *
4040 * This is called both for successful and failed load runs, we only care about
4041 * successful ones.
4042 *
4043 * @returns VBox status code.
4044 * @param pDrvIns The driver instance.
4045 * @param pSSM The saved state handle.
4046 */
4047static DECLCALLBACK(int) drvvdLoadDone(PPDMDRVINS pDrvIns, PSSMHANDLE pSSM)
4048{
4049 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4050 Assert(!pThis->fErrorUseRuntime);
4051
4052 /* Drop out if we don't have any work to do or if it's a failed load. */
4053 if ( !pThis->fTempReadOnly
4054 || RT_FAILURE(SSMR3HandleGetStatus(pSSM)))
4055 return VINF_SUCCESS;
4056
4057 int rc = drvvdSetWritable(pThis);
4058 if (RT_FAILURE(rc)) /** @todo does the bugger set any errors? */
4059 return SSMR3SetLoadError(pSSM, rc, RT_SRC_POS,
4060 N_("Failed to write lock the images"));
4061 return VINF_SUCCESS;
4062}
4063
4064
4065/*********************************************************************************************************************************
4066* Driver methods *
4067*********************************************************************************************************************************/
4068
4069/**
4070 * Worker for the power off or destruct callback.
4071 *
4072 * @returns nothing.
4073 * @param pDrvIns The driver instance.
4074 */
4075static void drvvdPowerOffOrDestructOrUnmount(PPDMDRVINS pDrvIns)
4076{
4077 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4078 LogFlowFunc(("\n"));
4079
4080 RTSEMFASTMUTEX mutex;
4081 ASMAtomicXchgHandle(&pThis->MergeCompleteMutex, NIL_RTSEMFASTMUTEX, &mutex);
4082 if (mutex != NIL_RTSEMFASTMUTEX)
4083 {
4084 /* Request the semaphore to wait until a potentially running merge
4085 * operation has been finished. */
4086 int rc = RTSemFastMutexRequest(mutex);
4087 AssertRC(rc);
4088 pThis->fMergePending = false;
4089 rc = RTSemFastMutexRelease(mutex);
4090 AssertRC(rc);
4091 rc = RTSemFastMutexDestroy(mutex);
4092 AssertRC(rc);
4093 }
4094
4095 if (RT_VALID_PTR(pThis->pBlkCache))
4096 {
4097 PDMR3BlkCacheRelease(pThis->pBlkCache);
4098 pThis->pBlkCache = NULL;
4099 }
4100
4101 if (RT_VALID_PTR(pThis->pDisk))
4102 {
4103 VDDestroy(pThis->pDisk);
4104 pThis->pDisk = NULL;
4105 }
4106 drvvdFreeImages(pThis);
4107}
4108
4109/**
4110 * @copydoc FNPDMDRVPOWEROFF
4111 */
4112static DECLCALLBACK(void) drvvdPowerOff(PPDMDRVINS pDrvIns)
4113{
4114 PDMDRV_CHECK_VERSIONS_RETURN_VOID(pDrvIns);
4115 drvvdPowerOffOrDestructOrUnmount(pDrvIns);
4116}
4117
4118/**
4119 * @callback_method_impl{FNPDMDRVRESUME}
4120 *
4121 * VM resume notification that we use to undo what the temporary read-only image
4122 * mode set by drvvdSuspend.
4123 *
4124 * Also switch to runtime error mode if we're resuming after a state load
4125 * without having been powered on first.
4126 *
4127 * @todo The VMSetError vs VMSetRuntimeError mess must be fixed elsewhere,
4128 * we're making assumptions about Main behavior here!
4129 */
4130static DECLCALLBACK(void) drvvdResume(PPDMDRVINS pDrvIns)
4131{
4132 LogFlowFunc(("\n"));
4133 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4134
4135 drvvdSetWritable(pThis);
4136 pThis->fErrorUseRuntime = true;
4137
4138 if (pThis->pBlkCache)
4139 {
4140 int rc = PDMR3BlkCacheResume(pThis->pBlkCache);
4141 AssertRC(rc);
4142 }
4143
4144 if (pThis->pDrvMediaExPort)
4145 {
4146 /* Kick of any request we have to redo. */
4147 PPDMMEDIAEXIOREQINT pIoReq, pIoReqNext;
4148 RTCritSectEnter(&pThis->CritSectIoReqRedo);
4149 RTListForEachSafe(&pThis->LstIoReqRedo, pIoReq, pIoReqNext, PDMMEDIAEXIOREQINT, NdLstWait)
4150 {
4151 int rc = VINF_SUCCESS;
4152 bool fXchg = ASMAtomicCmpXchgU32((volatile uint32_t *)&pIoReq->enmState, VDIOREQSTATE_ACTIVE, VDIOREQSTATE_SUSPENDED);
4153
4154 RTListNodeRemove(&pIoReq->NdLstWait);
4155 ASMAtomicIncU32(&pThis->cIoReqsActive);
4156
4157 if (fXchg)
4158 {
4159 pThis->pDrvMediaExPort->pfnIoReqStateChanged(pThis->pDrvMediaExPort, pIoReq, &pIoReq->abAlloc[0],
4160 PDMMEDIAEXIOREQSTATE_ACTIVE);
4161 if ( pIoReq->enmType == PDMMEDIAEXIOREQTYPE_READ
4162 || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_WRITE)
4163 rc = drvvdMediaExIoReqReadWriteProcess(pThis, pIoReq, true /* fUpNotify */);
4164 else if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_FLUSH)
4165 {
4166 rc = drvvdMediaExIoReqFlushWrapper(pThis, pIoReq);
4167 if (rc == VERR_VD_ASYNC_IO_IN_PROGRESS)
4168 rc = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
4169 else if (rc == VINF_VD_ASYNC_IO_FINISHED)
4170 rc = VINF_SUCCESS;
4171 }
4172 else if (pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD)
4173 {
4174 rc = drvvdMediaExIoReqDiscardWrapper(pThis, pIoReq);
4175 if (rc == VERR_VD_ASYNC_IO_IN_PROGRESS)
4176 rc = VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS;
4177 else if (rc == VINF_VD_ASYNC_IO_FINISHED)
4178 rc = VINF_SUCCESS;
4179 }
4180 else
4181 AssertMsgFailed(("Invalid request type %u\n", pIoReq->enmType));
4182
4183 /* The read write process will call the completion callback on its own. */
4184 if ( rc != VINF_PDM_MEDIAEX_IOREQ_IN_PROGRESS
4185 && ( pIoReq->enmType == PDMMEDIAEXIOREQTYPE_DISCARD
4186 || pIoReq->enmType == PDMMEDIAEXIOREQTYPE_FLUSH))
4187 {
4188 Assert( ( pIoReq->enmType != PDMMEDIAEXIOREQTYPE_WRITE
4189 && pIoReq->enmType != PDMMEDIAEXIOREQTYPE_READ)
4190 || !pIoReq->ReadWrite.cbReqLeft
4191 || RT_FAILURE(rc));
4192 drvvdMediaExIoReqCompleteWorker(pThis, pIoReq, rc, true /* fUpNotify */);
4193 }
4194
4195 }
4196 else
4197 {
4198 /* Request was canceled inbetween, so don't care and notify the owner about the completed request. */
4199 Assert(pIoReq->enmState == VDIOREQSTATE_CANCELED);
4200 drvvdMediaExIoReqCompleteWorker(pThis, pIoReq, VERR_PDM_MEDIAEX_IOREQ_CANCELED, true /* fUpNotify */);
4201 }
4202 }
4203 Assert(RTListIsEmpty(&pThis->LstIoReqRedo));
4204 RTCritSectLeave(&pThis->CritSectIoReqRedo);
4205 }
4206}
4207
4208/**
4209 * @callback_method_impl{FNPDMDRVSUSPEND}
4210 *
4211 * When the VM is being suspended, temporarily change to read-only image mode.
4212 *
4213 * This is important for several reasons:
4214 * -# It makes sure that there are no pending writes to the image. Most
4215 * backends implements this by closing and reopening the image in read-only
4216 * mode.
4217 * -# It allows Main to read the images during snapshotting without having
4218 * to account for concurrent writes.
4219 * -# This is essential for making teleportation targets sharing images work
4220 * right. Both with regards to caching and with regards to file sharing
4221 * locks (RTFILE_O_DENY_*). (See also drvvdLoadDone.)
4222 */
4223static DECLCALLBACK(void) drvvdSuspend(PPDMDRVINS pDrvIns)
4224{
4225 LogFlowFunc(("\n"));
4226 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4227
4228 if (pThis->pBlkCache)
4229 {
4230 int rc = PDMR3BlkCacheSuspend(pThis->pBlkCache);
4231 AssertRC(rc);
4232 }
4233
4234 drvvdSetReadonly(pThis);
4235}
4236
4237/**
4238 * @callback_method_impl{FNPDMDRVPOWERON}
4239 */
4240static DECLCALLBACK(void) drvvdPowerOn(PPDMDRVINS pDrvIns)
4241{
4242 LogFlowFunc(("\n"));
4243 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4244 drvvdSetWritable(pThis);
4245 pThis->fErrorUseRuntime = true;
4246}
4247
4248/**
4249 * @callback_method_impl{FNPDMDRVRESET}
4250 */
4251static DECLCALLBACK(void) drvvdReset(PPDMDRVINS pDrvIns)
4252{
4253 LogFlowFunc(("\n"));
4254 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4255
4256 if (pThis->pBlkCache)
4257 {
4258 int rc = PDMR3BlkCacheClear(pThis->pBlkCache);
4259 AssertRC(rc);
4260 }
4261
4262 if (pThis->fBootAccelEnabled)
4263 {
4264 pThis->fBootAccelActive = true;
4265 pThis->cbDataValid = 0;
4266 pThis->offDisk = 0;
4267 }
4268}
4269
4270/**
4271 * @callback_method_impl{FNPDMDRVDESTRUCT}
4272 */
4273static DECLCALLBACK(void) drvvdDestruct(PPDMDRVINS pDrvIns)
4274{
4275 PDMDRV_CHECK_VERSIONS_RETURN_VOID(pDrvIns);
4276 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4277 LogFlowFunc(("\n"));
4278
4279 /*
4280 * Make sure the block cache and disks are closed when this driver is
4281 * destroyed. This method will get called without calling the power off
4282 * callback first when we reconfigure the driver chain after a snapshot.
4283 */
4284 drvvdPowerOffOrDestructOrUnmount(pDrvIns);
4285 if (pThis->MergeLock != NIL_RTSEMRW)
4286 {
4287 int rc = RTSemRWDestroy(pThis->MergeLock);
4288 AssertRC(rc);
4289 pThis->MergeLock = NIL_RTSEMRW;
4290 }
4291 if (pThis->pbData)
4292 {
4293 RTMemFree(pThis->pbData);
4294 pThis->pbData = NULL;
4295 }
4296 if (pThis->pszBwGroup)
4297 {
4298 MMR3HeapFree(pThis->pszBwGroup);
4299 pThis->pszBwGroup = NULL;
4300 }
4301 if (pThis->hHbdMgr != NIL_HBDMGR)
4302 HBDMgrDestroy(pThis->hHbdMgr);
4303 if (pThis->hIoReqCache != NIL_RTMEMCACHE)
4304 RTMemCacheDestroy(pThis->hIoReqCache);
4305 if (pThis->hIoBufMgr != NIL_IOBUFMGR)
4306 IOBUFMgrDestroy(pThis->hIoBufMgr);
4307 if (RTCritSectIsInitialized(&pThis->CritSectIoReqsIoBufWait))
4308 RTCritSectDelete(&pThis->CritSectIoReqsIoBufWait);
4309 if (RTCritSectIsInitialized(&pThis->CritSectIoReqRedo))
4310 RTCritSectDelete(&pThis->CritSectIoReqRedo);
4311 for (unsigned i = 0; i < RT_ELEMENTS(pThis->aIoReqAllocBins); i++)
4312 if (pThis->aIoReqAllocBins[i].hMtxLstIoReqAlloc != NIL_RTSEMFASTMUTEX)
4313 RTSemFastMutexDestroy(pThis->aIoReqAllocBins[i].hMtxLstIoReqAlloc);
4314}
4315
4316/**
4317 * @callback_method_impl{FNPDMDRVCONSTRUCT,
4318 * Construct a VBox disk media driver instance.}
4319 */
4320static DECLCALLBACK(int) drvvdConstruct(PPDMDRVINS pDrvIns, PCFGMNODE pCfg, uint32_t fFlags)
4321{
4322 RT_NOREF(fFlags);
4323 LogFlowFunc(("\n"));
4324 PDMDRV_CHECK_VERSIONS_RETURN(pDrvIns);
4325 PVBOXDISK pThis = PDMINS_2_DATA(pDrvIns, PVBOXDISK);
4326 int rc = VINF_SUCCESS;
4327 char *pszName = NULL; /* The path of the disk image file. */
4328 char *pszFormat = NULL; /* The format backed to use for this image. */
4329 char *pszCachePath = NULL; /* The path to the cache image. */
4330 char *pszCacheFormat = NULL; /* The format backend to use for the cache image. */
4331 bool fReadOnly = false; /* True if the media is read-only. */
4332 bool fMaybeReadOnly = false; /* True if the media may or may not be read-only. */
4333 bool fHonorZeroWrites = false; /* True if zero blocks should be written. */
4334
4335 /*
4336 * Init the static parts.
4337 */
4338 pDrvIns->IBase.pfnQueryInterface = drvvdQueryInterface;
4339 pThis->pDrvIns = pDrvIns;
4340 pThis->fTempReadOnly = false;
4341 pThis->pDisk = NULL;
4342 pThis->fAsyncIOSupported = false;
4343 pThis->fShareable = false;
4344 pThis->fMergePending = false;
4345 pThis->MergeCompleteMutex = NIL_RTSEMFASTMUTEX;
4346 pThis->MergeLock = NIL_RTSEMRW;
4347 pThis->uMergeSource = VD_LAST_IMAGE;
4348 pThis->uMergeTarget = VD_LAST_IMAGE;
4349 pThis->pCfgCrypto = NULL;
4350 pThis->pIfSecKey = NULL;
4351 pThis->hIoReqCache = NIL_RTMEMCACHE;
4352 pThis->hIoBufMgr = NIL_IOBUFMGR;
4353
4354 for (unsigned i = 0; i < RT_ELEMENTS(pThis->aIoReqAllocBins); i++)
4355 pThis->aIoReqAllocBins[i].hMtxLstIoReqAlloc = NIL_RTSEMFASTMUTEX;
4356
4357 /* IMedia */
4358 pThis->IMedia.pfnRead = drvvdRead;
4359 pThis->IMedia.pfnReadPcBios = drvvdReadPcBios;
4360 pThis->IMedia.pfnWrite = drvvdWrite;
4361 pThis->IMedia.pfnFlush = drvvdFlush;
4362 pThis->IMedia.pfnMerge = drvvdMerge;
4363 pThis->IMedia.pfnSetSecKeyIf = drvvdSetSecKeyIf;
4364 pThis->IMedia.pfnGetSize = drvvdGetSize;
4365 pThis->IMedia.pfnGetSectorSize = drvvdGetSectorSize;
4366 pThis->IMedia.pfnIsReadOnly = drvvdIsReadOnly;
4367 pThis->IMedia.pfnIsNonRotational = drvvdIsNonRotational;
4368 pThis->IMedia.pfnBiosGetPCHSGeometry = drvvdBiosGetPCHSGeometry;
4369 pThis->IMedia.pfnBiosSetPCHSGeometry = drvvdBiosSetPCHSGeometry;
4370 pThis->IMedia.pfnBiosGetLCHSGeometry = drvvdBiosGetLCHSGeometry;
4371 pThis->IMedia.pfnBiosSetLCHSGeometry = drvvdBiosSetLCHSGeometry;
4372 pThis->IMedia.pfnBiosIsVisible = drvvdBiosIsVisible;
4373 pThis->IMedia.pfnGetType = drvvdGetType;
4374 pThis->IMedia.pfnGetUuid = drvvdGetUuid;
4375 pThis->IMedia.pfnDiscard = drvvdDiscard;
4376 pThis->IMedia.pfnSendCmd = NULL;
4377
4378 /* IMount */
4379 pThis->IMount.pfnUnmount = drvvdUnmount;
4380 pThis->IMount.pfnIsMounted = drvvdIsMounted;
4381 pThis->IMount.pfnLock = drvvdLock;
4382 pThis->IMount.pfnUnlock = drvvdUnlock;
4383 pThis->IMount.pfnIsLocked = drvvdIsLocked;
4384
4385 /* IMediaEx */
4386 pThis->IMediaEx.pfnQueryFeatures = drvvdQueryFeatures;
4387 pThis->IMediaEx.pfnIoReqAllocSizeSet = drvvdIoReqAllocSizeSet;
4388 pThis->IMediaEx.pfnIoReqAlloc = drvvdIoReqAlloc;
4389 pThis->IMediaEx.pfnIoReqFree = drvvdIoReqFree;
4390 pThis->IMediaEx.pfnIoReqQueryResidual = drvvdIoReqQueryResidual;
4391 pThis->IMediaEx.pfnIoReqCancelAll = drvvdIoReqCancelAll;
4392 pThis->IMediaEx.pfnIoReqCancel = drvvdIoReqCancel;
4393 pThis->IMediaEx.pfnIoReqRead = drvvdIoReqRead;
4394 pThis->IMediaEx.pfnIoReqWrite = drvvdIoReqWrite;
4395 pThis->IMediaEx.pfnIoReqFlush = drvvdIoReqFlush;
4396 pThis->IMediaEx.pfnIoReqDiscard = drvvdIoReqDiscard;
4397 pThis->IMediaEx.pfnIoReqSendScsiCmd = drvvdIoReqSendScsiCmd;
4398 pThis->IMediaEx.pfnIoReqGetActiveCount = drvvdIoReqGetActiveCount;
4399 pThis->IMediaEx.pfnIoReqGetSuspendedCount = drvvdIoReqGetSuspendedCount;
4400 pThis->IMediaEx.pfnIoReqQuerySuspendedStart = drvvdIoReqQuerySuspendedStart;
4401 pThis->IMediaEx.pfnIoReqQuerySuspendedNext = drvvdIoReqQuerySuspendedNext;
4402 pThis->IMediaEx.pfnIoReqSuspendedSave = drvvdIoReqSuspendedSave;
4403 pThis->IMediaEx.pfnIoReqSuspendedLoad = drvvdIoReqSuspendedLoad;
4404
4405 /* Initialize supported VD interfaces. */
4406 pThis->pVDIfsDisk = NULL;
4407
4408 pThis->VDIfError.pfnError = drvvdErrorCallback;
4409 pThis->VDIfError.pfnMessage = NULL;
4410 rc = VDInterfaceAdd(&pThis->VDIfError.Core, "DrvVD_VDIError", VDINTERFACETYPE_ERROR,
4411 pDrvIns, sizeof(VDINTERFACEERROR), &pThis->pVDIfsDisk);
4412 AssertRC(rc);
4413
4414 /* List of images is empty now. */
4415 pThis->pImages = NULL;
4416
4417 pThis->pDrvMediaPort = PDMIBASE_QUERY_INTERFACE(pDrvIns->pUpBase, PDMIMEDIAPORT);
4418 if (!pThis->pDrvMediaPort)
4419 return PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_MISSING_INTERFACE_ABOVE,
4420 N_("No media port interface above"));
4421
4422 pThis->pDrvMountNotify = PDMIBASE_QUERY_INTERFACE(pDrvIns->pUpBase, PDMIMOUNTNOTIFY);
4423
4424 /*
4425 * Try to attach the optional extended media interface port above and initialize associated
4426 * structures if available.
4427 */
4428 pThis->pDrvMediaExPort = PDMIBASE_QUERY_INTERFACE(pDrvIns->pUpBase, PDMIMEDIAEXPORT);
4429 if (pThis->pDrvMediaExPort)
4430 {
4431 for (unsigned i = 0; i < RT_ELEMENTS(pThis->aIoReqAllocBins); i++)
4432 {
4433 rc = RTSemFastMutexCreate(&pThis->aIoReqAllocBins[i].hMtxLstIoReqAlloc);
4434 if (RT_FAILURE(rc))
4435 break;
4436 RTListInit(&pThis->aIoReqAllocBins[i].LstIoReqAlloc);
4437 }
4438
4439 if (RT_SUCCESS(rc))
4440 rc = RTCritSectInit(&pThis->CritSectIoReqsIoBufWait);
4441
4442 if (RT_SUCCESS(rc))
4443 rc = RTCritSectInit(&pThis->CritSectIoReqRedo);
4444
4445 if (RT_FAILURE(rc))
4446 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Creating Mutex failed"));
4447
4448 RTListInit(&pThis->LstIoReqIoBufWait);
4449 RTListInit(&pThis->LstIoReqRedo);
4450 }
4451
4452 /* Before we access any VD API load all given plugins. */
4453 rc = drvvdLoadPlugins(pCfg);
4454 if (RT_FAILURE(rc))
4455 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Loading VD plugins failed"));
4456
4457 /*
4458 * Validate configuration and find all parent images.
4459 * It's sort of up side down from the image dependency tree.
4460 */
4461 bool fHostIP = false;
4462 bool fUseNewIo = false;
4463 bool fUseBlockCache = false;
4464 bool fDiscard = false;
4465 bool fInformAboutZeroBlocks = false;
4466 bool fSkipConsistencyChecks = false;
4467 bool fEmptyDrive = false;
4468 unsigned iLevel = 0;
4469 PCFGMNODE pCurNode = pCfg;
4470 uint32_t cbIoBufMax = 0;
4471
4472 for (;;)
4473 {
4474 bool fValid;
4475
4476 if (pCurNode == pCfg)
4477 {
4478 /* Toplevel configuration additionally contains the global image
4479 * open flags. Some might be converted to per-image flags later. */
4480 fValid = CFGMR3AreValuesValid(pCurNode,
4481 "Format\0Path\0"
4482 "ReadOnly\0MaybeReadOnly\0TempReadOnly\0Shareable\0HonorZeroWrites\0"
4483 "HostIPStack\0UseNewIo\0BootAcceleration\0BootAccelerationBuffer\0"
4484 "SetupMerge\0MergeSource\0MergeTarget\0BwGroup\0Type\0BlockCache\0"
4485 "CachePath\0CacheFormat\0Discard\0InformAboutZeroBlocks\0"
4486 "SkipConsistencyChecks\0"
4487 "Locked\0BIOSVisible\0Cylinders\0Heads\0Sectors\0Mountable\0"
4488 "EmptyDrive\0IoBufMax\0NonRotationalMedium\0"
4489#if defined(VBOX_PERIODIC_FLUSH) || defined(VBOX_IGNORE_FLUSH)
4490 "FlushInterval\0IgnoreFlush\0IgnoreFlushAsync\0"
4491#endif /* !(VBOX_PERIODIC_FLUSH || VBOX_IGNORE_FLUSH) */
4492 );
4493 }
4494 else
4495 {
4496 /* All other image configurations only contain image name and
4497 * the format information. */
4498 fValid = CFGMR3AreValuesValid(pCurNode, "Format\0Path\0"
4499 "MergeSource\0MergeTarget\0");
4500 }
4501 if (!fValid)
4502 {
4503 rc = PDMDrvHlpVMSetError(pDrvIns, VERR_PDM_DRVINS_UNKNOWN_CFG_VALUES,
4504 RT_SRC_POS, N_("DrvVD: Configuration error: keys incorrect at level %d"), iLevel);
4505 break;
4506 }
4507
4508 if (pCurNode == pCfg)
4509 {
4510 rc = CFGMR3QueryBoolDef(pCurNode, "HostIPStack", &fHostIP, true);
4511 if (RT_FAILURE(rc))
4512 {
4513 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4514 N_("DrvVD: Configuration error: Querying \"HostIPStack\" as boolean failed"));
4515 break;
4516 }
4517
4518 rc = CFGMR3QueryBoolDef(pCurNode, "HonorZeroWrites", &fHonorZeroWrites, false);
4519 if (RT_FAILURE(rc))
4520 {
4521 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4522 N_("DrvVD: Configuration error: Querying \"HonorZeroWrites\" as boolean failed"));
4523 break;
4524 }
4525
4526 rc = CFGMR3QueryBoolDef(pCurNode, "ReadOnly", &fReadOnly, false);
4527 if (RT_FAILURE(rc))
4528 {
4529 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4530 N_("DrvVD: Configuration error: Querying \"ReadOnly\" as boolean failed"));
4531 break;
4532 }
4533
4534 rc = CFGMR3QueryBoolDef(pCurNode, "MaybeReadOnly", &fMaybeReadOnly, false);
4535 if (RT_FAILURE(rc))
4536 {
4537 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4538 N_("DrvVD: Configuration error: Querying \"MaybeReadOnly\" as boolean failed"));
4539 break;
4540 }
4541
4542 rc = CFGMR3QueryBoolDef(pCurNode, "TempReadOnly", &pThis->fTempReadOnly, false);
4543 if (RT_FAILURE(rc))
4544 {
4545 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4546 N_("DrvVD: Configuration error: Querying \"TempReadOnly\" as boolean failed"));
4547 break;
4548 }
4549 if (fReadOnly && pThis->fTempReadOnly)
4550 {
4551 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
4552 N_("DrvVD: Configuration error: Both \"ReadOnly\" and \"TempReadOnly\" are set"));
4553 break;
4554 }
4555
4556 rc = CFGMR3QueryBoolDef(pCurNode, "Shareable", &pThis->fShareable, false);
4557 if (RT_FAILURE(rc))
4558 {
4559 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4560 N_("DrvVD: Configuration error: Querying \"Shareable\" as boolean failed"));
4561 break;
4562 }
4563
4564 rc = CFGMR3QueryBoolDef(pCurNode, "UseNewIo", &fUseNewIo, false);
4565 if (RT_FAILURE(rc))
4566 {
4567 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4568 N_("DrvVD: Configuration error: Querying \"UseNewIo\" as boolean failed"));
4569 break;
4570 }
4571 rc = CFGMR3QueryBoolDef(pCurNode, "SetupMerge", &pThis->fMergePending, false);
4572 if (RT_FAILURE(rc))
4573 {
4574 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4575 N_("DrvVD: Configuration error: Querying \"SetupMerge\" as boolean failed"));
4576 break;
4577 }
4578 if (fReadOnly && pThis->fMergePending)
4579 {
4580 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
4581 N_("DrvVD: Configuration error: Both \"ReadOnly\" and \"MergePending\" are set"));
4582 break;
4583 }
4584 rc = CFGMR3QueryBoolDef(pCurNode, "BootAcceleration", &pThis->fBootAccelEnabled, false);
4585 if (RT_FAILURE(rc))
4586 {
4587 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4588 N_("DrvVD: Configuration error: Querying \"BootAcceleration\" as boolean failed"));
4589 break;
4590 }
4591 rc = CFGMR3QueryU32Def(pCurNode, "BootAccelerationBuffer", (uint32_t *)&pThis->cbBootAccelBuffer, 16 * _1K);
4592 if (RT_FAILURE(rc))
4593 {
4594 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4595 N_("DrvVD: Configuration error: Querying \"BootAccelerationBuffer\" as integer failed"));
4596 break;
4597 }
4598 rc = CFGMR3QueryBoolDef(pCurNode, "BlockCache", &fUseBlockCache, false);
4599 if (RT_FAILURE(rc))
4600 {
4601 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4602 N_("DrvVD: Configuration error: Querying \"BlockCache\" as boolean failed"));
4603 break;
4604 }
4605 rc = CFGMR3QueryStringAlloc(pCurNode, "BwGroup", &pThis->pszBwGroup);
4606 if (RT_FAILURE(rc) && rc != VERR_CFGM_VALUE_NOT_FOUND)
4607 {
4608 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4609 N_("DrvVD: Configuration error: Querying \"BwGroup\" as string failed"));
4610 break;
4611 }
4612 else
4613 rc = VINF_SUCCESS;
4614 rc = CFGMR3QueryBoolDef(pCurNode, "Discard", &fDiscard, false);
4615 if (RT_FAILURE(rc))
4616 {
4617 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4618 N_("DrvVD: Configuration error: Querying \"Discard\" as boolean failed"));
4619 break;
4620 }
4621 if (fReadOnly && fDiscard)
4622 {
4623 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
4624 N_("DrvVD: Configuration error: Both \"ReadOnly\" and \"Discard\" are set"));
4625 break;
4626 }
4627 rc = CFGMR3QueryBoolDef(pCurNode, "InformAboutZeroBlocks", &fInformAboutZeroBlocks, false);
4628 if (RT_FAILURE(rc))
4629 {
4630 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4631 N_("DrvVD: Configuration error: Querying \"InformAboutZeroBlocks\" as boolean failed"));
4632 break;
4633 }
4634 rc = CFGMR3QueryBoolDef(pCurNode, "SkipConsistencyChecks", &fSkipConsistencyChecks, true);
4635 if (RT_FAILURE(rc))
4636 {
4637 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4638 N_("DrvVD: Configuration error: Querying \"SKipConsistencyChecks\" as boolean failed"));
4639 break;
4640 }
4641
4642 char *psz = NULL;
4643 rc = CFGMR3QueryStringAlloc(pCfg, "Type", &psz);
4644 if (RT_FAILURE(rc))
4645 return PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_BLOCK_NO_TYPE, N_("Failed to obtain the sub type"));
4646 pThis->enmType = drvvdGetMediaTypeFromString(psz);
4647 if (pThis->enmType == PDMMEDIATYPE_ERROR)
4648 {
4649 PDMDrvHlpVMSetError(pDrvIns, VERR_PDM_BLOCK_UNKNOWN_TYPE, RT_SRC_POS,
4650 N_("Unknown type \"%s\""), psz);
4651 MMR3HeapFree(psz);
4652 return VERR_PDM_BLOCK_UNKNOWN_TYPE;
4653 }
4654 MMR3HeapFree(psz); psz = NULL;
4655
4656 rc = CFGMR3QueryStringAlloc(pCurNode, "CachePath", &pszCachePath);
4657 if (RT_FAILURE(rc) && rc != VERR_CFGM_VALUE_NOT_FOUND)
4658 {
4659 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4660 N_("DrvVD: Configuration error: Querying \"CachePath\" as string failed"));
4661 break;
4662 }
4663 else
4664 rc = VINF_SUCCESS;
4665
4666 if (pszCachePath)
4667 {
4668 rc = CFGMR3QueryStringAlloc(pCurNode, "CacheFormat", &pszCacheFormat);
4669 if (RT_FAILURE(rc))
4670 {
4671 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4672 N_("DrvVD: Configuration error: Querying \"CacheFormat\" as string failed"));
4673 break;
4674 }
4675 }
4676
4677 /* Mountable */
4678 rc = CFGMR3QueryBoolDef(pCfg, "Mountable", &pThis->fMountable, false);
4679 if (RT_FAILURE(rc))
4680 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"Mountable\" from the config"));
4681
4682 /* Locked */
4683 rc = CFGMR3QueryBoolDef(pCfg, "Locked", &pThis->fLocked, false);
4684 if (RT_FAILURE(rc))
4685 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"Locked\" from the config"));
4686
4687 /* BIOS visible */
4688 rc = CFGMR3QueryBoolDef(pCfg, "BIOSVisible", &pThis->fBiosVisible, true);
4689 if (RT_FAILURE(rc))
4690 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"BIOSVisible\" from the config"));
4691
4692 /* Cylinders */
4693 rc = CFGMR3QueryU32Def(pCfg, "Cylinders", &pThis->LCHSGeometry.cCylinders, 0);
4694 if (RT_FAILURE(rc))
4695 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"Cylinders\" from the config"));
4696
4697 /* Heads */
4698 rc = CFGMR3QueryU32Def(pCfg, "Heads", &pThis->LCHSGeometry.cHeads, 0);
4699 if (RT_FAILURE(rc))
4700 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"Heads\" from the config"));
4701
4702 /* Sectors */
4703 rc = CFGMR3QueryU32Def(pCfg, "Sectors", &pThis->LCHSGeometry.cSectors, 0);
4704 if (RT_FAILURE(rc))
4705 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"Sectors\" from the config"));
4706
4707 /* Uuid */
4708 rc = CFGMR3QueryStringAlloc(pCfg, "Uuid", &psz);
4709 if (rc == VERR_CFGM_VALUE_NOT_FOUND)
4710 RTUuidClear(&pThis->Uuid);
4711 else if (RT_SUCCESS(rc))
4712 {
4713 rc = RTUuidFromStr(&pThis->Uuid, psz);
4714 if (RT_FAILURE(rc))
4715 {
4716 PDMDrvHlpVMSetError(pDrvIns, rc, RT_SRC_POS, N_("Uuid from string failed on \"%s\""), psz);
4717 MMR3HeapFree(psz);
4718 return rc;
4719 }
4720 MMR3HeapFree(psz); psz = NULL;
4721 }
4722 else
4723 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"Uuid\" from the config"));
4724
4725#ifdef VBOX_PERIODIC_FLUSH
4726 rc = CFGMR3QueryU32Def(pCfg, "FlushInterval", &pThis->cbFlushInterval, 0);
4727 if (RT_FAILURE(rc))
4728 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"FlushInterval\" from the config"));
4729#endif /* VBOX_PERIODIC_FLUSH */
4730
4731#ifdef VBOX_IGNORE_FLUSH
4732 rc = CFGMR3QueryBoolDef(pCfg, "IgnoreFlush", &pThis->fIgnoreFlush, true);
4733 if (RT_FAILURE(rc))
4734 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"IgnoreFlush\" from the config"));
4735
4736 if (pThis->fIgnoreFlush)
4737 LogRel(("DrvVD: Flushes will be ignored\n"));
4738 else
4739 LogRel(("DrvVD: Flushes will be passed to the disk\n"));
4740
4741 rc = CFGMR3QueryBoolDef(pCfg, "IgnoreFlushAsync", &pThis->fIgnoreFlushAsync, false);
4742 if (RT_FAILURE(rc))
4743 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"IgnoreFlushAsync\" from the config"));
4744
4745 if (pThis->fIgnoreFlushAsync)
4746 LogRel(("DrvVD: Async flushes will be ignored\n"));
4747 else
4748 LogRel(("DrvVD: Async flushes will be passed to the disk\n"));
4749#endif /* VBOX_IGNORE_FLUSH */
4750
4751 rc = CFGMR3QueryBoolDef(pCurNode, "EmptyDrive", &fEmptyDrive, false);
4752 if (RT_FAILURE(rc))
4753 {
4754 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4755 N_("DrvVD: Configuration error: Querying \"EmptyDrive\" as boolean failed"));
4756 break;
4757 }
4758
4759 rc = CFGMR3QueryU32Def(pCfg, "IoBufMax", &cbIoBufMax, 5 * _1M);
4760 if (RT_FAILURE(rc))
4761 return PDMDRV_SET_ERROR(pDrvIns, rc, N_("Failed to query \"IoBufMax\" from the config"));
4762
4763 rc = CFGMR3QueryBoolDef(pCfg, "NonRotationalMedium", &pThis->fNonRotational, false);
4764 if (RT_FAILURE(rc))
4765 return PDMDRV_SET_ERROR(pDrvIns, rc,
4766 N_("DrvVD configuration error: Querying \"NonRotationalMedium\" as boolean failed"));
4767 }
4768
4769 PCFGMNODE pParent = CFGMR3GetChild(pCurNode, "Parent");
4770 if (!pParent)
4771 break;
4772 pCurNode = pParent;
4773 iLevel++;
4774 }
4775
4776 if (pThis->pDrvMediaExPort)
4777 rc = IOBUFMgrCreate(&pThis->hIoBufMgr, cbIoBufMax, pThis->pCfgCrypto ? IOBUFMGR_F_REQUIRE_NOT_PAGABLE : IOBUFMGR_F_DEFAULT);
4778
4779 if ( !fEmptyDrive
4780 && RT_SUCCESS(rc))
4781 {
4782 /*
4783 * Create the image container and the necessary interfaces.
4784 */
4785 if (RT_SUCCESS(rc))
4786 {
4787 /*
4788 * The image has a bandwidth group but the host cache is enabled.
4789 * Use the async I/O framework but tell it to enable the host cache.
4790 */
4791 if (!fUseNewIo && pThis->pszBwGroup)
4792 {
4793 pThis->fAsyncIoWithHostCache = true;
4794 fUseNewIo = true;
4795 }
4796
4797 /** @todo quick hack to work around problems in the async I/O
4798 * implementation (rw semaphore thread ownership problem)
4799 * while a merge is running. Remove once this is fixed. */
4800 if (pThis->fMergePending)
4801 fUseNewIo = false;
4802
4803 if (RT_SUCCESS(rc) && pThis->fMergePending)
4804 {
4805 rc = RTSemFastMutexCreate(&pThis->MergeCompleteMutex);
4806 if (RT_SUCCESS(rc))
4807 rc = RTSemRWCreate(&pThis->MergeLock);
4808 if (RT_SUCCESS(rc))
4809 {
4810 pThis->VDIfThreadSync.pfnStartRead = drvvdThreadStartRead;
4811 pThis->VDIfThreadSync.pfnFinishRead = drvvdThreadFinishRead;
4812 pThis->VDIfThreadSync.pfnStartWrite = drvvdThreadStartWrite;
4813 pThis->VDIfThreadSync.pfnFinishWrite = drvvdThreadFinishWrite;
4814
4815 rc = VDInterfaceAdd(&pThis->VDIfThreadSync.Core, "DrvVD_ThreadSync", VDINTERFACETYPE_THREADSYNC,
4816 pThis, sizeof(VDINTERFACETHREADSYNC), &pThis->pVDIfsDisk);
4817 }
4818 else
4819 {
4820 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4821 N_("DrvVD: Failed to create semaphores for \"MergePending\""));
4822 }
4823 }
4824
4825 if (RT_SUCCESS(rc))
4826 {
4827 rc = VDCreate(pThis->pVDIfsDisk, drvvdGetVDFromMediaType(pThis->enmType), &pThis->pDisk);
4828 /* Error message is already set correctly. */
4829 }
4830 }
4831
4832 if (pThis->pDrvMediaExPort && fUseNewIo)
4833 pThis->fAsyncIOSupported = true;
4834
4835 uint64_t tsStart = RTTimeNanoTS();
4836
4837 unsigned iImageIdx = 0;
4838 while (pCurNode && RT_SUCCESS(rc))
4839 {
4840 /* Allocate per-image data. */
4841 PVBOXIMAGE pImage = drvvdNewImage(pThis);
4842 if (!pImage)
4843 {
4844 rc = VERR_NO_MEMORY;
4845 break;
4846 }
4847
4848 /*
4849 * Read the image configuration.
4850 */
4851 rc = CFGMR3QueryStringAlloc(pCurNode, "Path", &pszName);
4852 if (RT_FAILURE(rc))
4853 {
4854 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4855 N_("DrvVD: Configuration error: Querying \"Path\" as string failed"));
4856 break;
4857 }
4858
4859 rc = CFGMR3QueryStringAlloc(pCurNode, "Format", &pszFormat);
4860 if (RT_FAILURE(rc))
4861 {
4862 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4863 N_("DrvVD: Configuration error: Querying \"Format\" as string failed"));
4864 break;
4865 }
4866
4867 bool fMergeSource;
4868 rc = CFGMR3QueryBoolDef(pCurNode, "MergeSource", &fMergeSource, false);
4869 if (RT_FAILURE(rc))
4870 {
4871 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4872 N_("DrvVD: Configuration error: Querying \"MergeSource\" as boolean failed"));
4873 break;
4874 }
4875 if (fMergeSource)
4876 {
4877 if (pThis->uMergeSource == VD_LAST_IMAGE)
4878 pThis->uMergeSource = iImageIdx;
4879 else
4880 {
4881 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
4882 N_("DrvVD: Configuration error: Multiple \"MergeSource\" occurrences"));
4883 break;
4884 }
4885 }
4886
4887 bool fMergeTarget;
4888 rc = CFGMR3QueryBoolDef(pCurNode, "MergeTarget", &fMergeTarget, false);
4889 if (RT_FAILURE(rc))
4890 {
4891 rc = PDMDRV_SET_ERROR(pDrvIns, rc,
4892 N_("DrvVD: Configuration error: Querying \"MergeTarget\" as boolean failed"));
4893 break;
4894 }
4895 if (fMergeTarget)
4896 {
4897 if (pThis->uMergeTarget == VD_LAST_IMAGE)
4898 pThis->uMergeTarget = iImageIdx;
4899 else
4900 {
4901 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
4902 N_("DrvVD: Configuration error: Multiple \"MergeTarget\" occurrences"));
4903 break;
4904 }
4905 }
4906
4907 PCFGMNODE pCfgVDConfig = CFGMR3GetChild(pCurNode, "VDConfig");
4908 pImage->VDIfConfig.pfnAreKeysValid = drvvdCfgAreKeysValid;
4909 pImage->VDIfConfig.pfnQuerySize = drvvdCfgQuerySize;
4910 pImage->VDIfConfig.pfnQuery = drvvdCfgQuery;
4911 pImage->VDIfConfig.pfnQueryBytes = NULL;
4912 rc = VDInterfaceAdd(&pImage->VDIfConfig.Core, "DrvVD_Config", VDINTERFACETYPE_CONFIG,
4913 pCfgVDConfig, sizeof(VDINTERFACECONFIG), &pImage->pVDIfsImage);
4914 AssertRC(rc);
4915
4916 /* Check VDConfig for encryption config. */
4917 if (pCfgVDConfig)
4918 pThis->pCfgCrypto = CFGMR3GetChild(pCfgVDConfig, "CRYPT");
4919
4920 if (pThis->pCfgCrypto)
4921 {
4922 /* Setup VDConfig interface for disk encryption support. */
4923 pThis->VDIfCfg.pfnAreKeysValid = drvvdCfgAreKeysValid;
4924 pThis->VDIfCfg.pfnQuerySize = drvvdCfgQuerySize;
4925 pThis->VDIfCfg.pfnQuery = drvvdCfgQuery;
4926 pThis->VDIfCfg.pfnQueryBytes = NULL;
4927
4928 pThis->VDIfCrypto.pfnKeyRetain = drvvdCryptoKeyRetain;
4929 pThis->VDIfCrypto.pfnKeyRelease = drvvdCryptoKeyRelease;
4930 pThis->VDIfCrypto.pfnKeyStorePasswordRetain = drvvdCryptoKeyStorePasswordRetain;
4931 pThis->VDIfCrypto.pfnKeyStorePasswordRelease = drvvdCryptoKeyStorePasswordRelease;
4932 }
4933
4934 /* Unconditionally insert the TCPNET interface, don't bother to check
4935 * if an image really needs it. Will be ignored. Since the TCPNET
4936 * interface is per image we could make this more flexible in the
4937 * future if we want to. */
4938 /* Construct TCPNET callback table depending on the config. This is
4939 * done unconditionally, as uninterested backends will ignore it. */
4940 if (fHostIP)
4941 {
4942 pImage->VDIfTcpNet.pfnSocketCreate = drvvdTcpSocketCreate;
4943 pImage->VDIfTcpNet.pfnSocketDestroy = drvvdTcpSocketDestroy;
4944 pImage->VDIfTcpNet.pfnClientConnect = drvvdTcpClientConnect;
4945 pImage->VDIfTcpNet.pfnIsClientConnected = drvvdTcpIsClientConnected;
4946 pImage->VDIfTcpNet.pfnClientClose = drvvdTcpClientClose;
4947 pImage->VDIfTcpNet.pfnSelectOne = drvvdTcpSelectOne;
4948 pImage->VDIfTcpNet.pfnRead = drvvdTcpRead;
4949 pImage->VDIfTcpNet.pfnWrite = drvvdTcpWrite;
4950 pImage->VDIfTcpNet.pfnSgWrite = drvvdTcpSgWrite;
4951 pImage->VDIfTcpNet.pfnReadNB = drvvdTcpReadNB;
4952 pImage->VDIfTcpNet.pfnWriteNB = drvvdTcpWriteNB;
4953 pImage->VDIfTcpNet.pfnSgWriteNB = drvvdTcpSgWriteNB;
4954 pImage->VDIfTcpNet.pfnFlush = drvvdTcpFlush;
4955 pImage->VDIfTcpNet.pfnSetSendCoalescing = drvvdTcpSetSendCoalescing;
4956 pImage->VDIfTcpNet.pfnGetLocalAddress = drvvdTcpGetLocalAddress;
4957 pImage->VDIfTcpNet.pfnGetPeerAddress = drvvdTcpGetPeerAddress;
4958
4959 /*
4960 * There is a 15ms delay between receiving the data and marking the socket
4961 * as readable on Windows XP which hurts async I/O performance of
4962 * TCP backends badly. Provide a different select method without
4963 * using poll on XP.
4964 * This is only used on XP because it is not as efficient as the one using poll
4965 * and all other Windows versions are working fine.
4966 */
4967 char szOS[64];
4968 memset(szOS, 0, sizeof(szOS));
4969 rc = RTSystemQueryOSInfo(RTSYSOSINFO_PRODUCT, &szOS[0], sizeof(szOS));
4970
4971 if (RT_SUCCESS(rc) && !strncmp(szOS, "Windows XP", 10))
4972 {
4973 LogRel(("VD: Detected Windows XP, disabled poll based waiting for TCP\n"));
4974 pImage->VDIfTcpNet.pfnSelectOneEx = drvvdTcpSelectOneExNoPoll;
4975 }
4976 else
4977 pImage->VDIfTcpNet.pfnSelectOneEx = drvvdTcpSelectOneExPoll;
4978
4979 pImage->VDIfTcpNet.pfnPoke = drvvdTcpPoke;
4980 }
4981 else
4982 {
4983#ifndef VBOX_WITH_INIP
4984 rc = PDMDrvHlpVMSetError(pDrvIns, VERR_PDM_DRVINS_UNKNOWN_CFG_VALUES,
4985 RT_SRC_POS, N_("DrvVD: Configuration error: TCP over Internal Networking not compiled in"));
4986#else /* VBOX_WITH_INIP */
4987 pImage->VDIfTcpNet.pfnSocketCreate = drvvdINIPSocketCreate;
4988 pImage->VDIfTcpNet.pfnSocketDestroy = drvvdINIPSocketDestroy;
4989 pImage->VDIfTcpNet.pfnClientConnect = drvvdINIPClientConnect;
4990 pImage->VDIfTcpNet.pfnClientClose = drvvdINIPClientClose;
4991 pImage->VDIfTcpNet.pfnIsClientConnected = drvvdINIPIsClientConnected;
4992 pImage->VDIfTcpNet.pfnSelectOne = drvvdINIPSelectOne;
4993 pImage->VDIfTcpNet.pfnRead = drvvdINIPRead;
4994 pImage->VDIfTcpNet.pfnWrite = drvvdINIPWrite;
4995 pImage->VDIfTcpNet.pfnSgWrite = drvvdINIPSgWrite;
4996 pImage->VDIfTcpNet.pfnFlush = drvvdINIPFlush;
4997 pImage->VDIfTcpNet.pfnSetSendCoalescing = drvvdINIPSetSendCoalescing;
4998 pImage->VDIfTcpNet.pfnGetLocalAddress = drvvdINIPGetLocalAddress;
4999 pImage->VDIfTcpNet.pfnGetPeerAddress = drvvdINIPGetPeerAddress;
5000 pImage->VDIfTcpNet.pfnSelectOneEx = drvvdINIPSelectOneEx;
5001 pImage->VDIfTcpNet.pfnPoke = drvvdINIPPoke;
5002#endif /* VBOX_WITH_INIP */
5003 }
5004 rc = VDInterfaceAdd(&pImage->VDIfTcpNet.Core, "DrvVD_TCPNET",
5005 VDINTERFACETYPE_TCPNET, NULL,
5006 sizeof(VDINTERFACETCPNET), &pImage->pVDIfsImage);
5007 AssertRC(rc);
5008
5009 /* Insert the custom I/O interface only if we're told to use new IO.
5010 * Since the I/O interface is per image we could make this more
5011 * flexible in the future if we want to. */
5012 if (fUseNewIo)
5013 {
5014#ifdef VBOX_WITH_PDM_ASYNC_COMPLETION
5015 pImage->VDIfIo.pfnOpen = drvvdAsyncIOOpen;
5016 pImage->VDIfIo.pfnClose = drvvdAsyncIOClose;
5017 pImage->VDIfIo.pfnGetSize = drvvdAsyncIOGetSize;
5018 pImage->VDIfIo.pfnSetSize = drvvdAsyncIOSetSize;
5019 pImage->VDIfIo.pfnSetAllocationSize = drvvdAsyncIOSetAllocationSize;
5020 pImage->VDIfIo.pfnReadSync = drvvdAsyncIOReadSync;
5021 pImage->VDIfIo.pfnWriteSync = drvvdAsyncIOWriteSync;
5022 pImage->VDIfIo.pfnFlushSync = drvvdAsyncIOFlushSync;
5023 pImage->VDIfIo.pfnReadAsync = drvvdAsyncIOReadAsync;
5024 pImage->VDIfIo.pfnWriteAsync = drvvdAsyncIOWriteAsync;
5025 pImage->VDIfIo.pfnFlushAsync = drvvdAsyncIOFlushAsync;
5026#else /* !VBOX_WITH_PDM_ASYNC_COMPLETION */
5027 rc = PDMDrvHlpVMSetError(pDrvIns, VERR_PDM_DRVINS_UNKNOWN_CFG_VALUES,
5028 RT_SRC_POS, N_("DrvVD: Configuration error: Async Completion Framework not compiled in"));
5029#endif /* !VBOX_WITH_PDM_ASYNC_COMPLETION */
5030 if (RT_SUCCESS(rc))
5031 rc = VDInterfaceAdd(&pImage->VDIfIo.Core, "DrvVD_IO", VDINTERFACETYPE_IO,
5032 pThis, sizeof(VDINTERFACEIO), &pImage->pVDIfsImage);
5033 AssertRC(rc);
5034 }
5035
5036 /*
5037 * Open the image.
5038 */
5039 unsigned uOpenFlags;
5040 if (fReadOnly || pThis->fTempReadOnly || iLevel != 0)
5041 uOpenFlags = VD_OPEN_FLAGS_READONLY;
5042 else
5043 uOpenFlags = VD_OPEN_FLAGS_NORMAL;
5044 if (fHonorZeroWrites)
5045 uOpenFlags |= VD_OPEN_FLAGS_HONOR_ZEROES;
5046 if (pThis->fAsyncIOSupported)
5047 uOpenFlags |= VD_OPEN_FLAGS_ASYNC_IO;
5048 if (pThis->fShareable)
5049 uOpenFlags |= VD_OPEN_FLAGS_SHAREABLE;
5050 if (fDiscard && iLevel == 0)
5051 uOpenFlags |= VD_OPEN_FLAGS_DISCARD;
5052 if (fInformAboutZeroBlocks)
5053 uOpenFlags |= VD_OPEN_FLAGS_INFORM_ABOUT_ZERO_BLOCKS;
5054 if ( (uOpenFlags & VD_OPEN_FLAGS_READONLY)
5055 && fSkipConsistencyChecks)
5056 uOpenFlags |= VD_OPEN_FLAGS_SKIP_CONSISTENCY_CHECKS;
5057
5058 /* Try to open backend in async I/O mode first. */
5059 rc = VDOpen(pThis->pDisk, pszFormat, pszName, uOpenFlags, pImage->pVDIfsImage);
5060 if (rc == VERR_NOT_SUPPORTED)
5061 {
5062 pThis->fAsyncIOSupported = false;
5063 uOpenFlags &= ~VD_OPEN_FLAGS_ASYNC_IO;
5064 rc = VDOpen(pThis->pDisk, pszFormat, pszName, uOpenFlags, pImage->pVDIfsImage);
5065 }
5066
5067 if (rc == VERR_VD_DISCARD_NOT_SUPPORTED)
5068 {
5069 fDiscard = false;
5070 uOpenFlags &= ~VD_OPEN_FLAGS_DISCARD;
5071 rc = VDOpen(pThis->pDisk, pszFormat, pszName, uOpenFlags, pImage->pVDIfsImage);
5072 }
5073
5074 if (!fDiscard)
5075 {
5076 pThis->IMedia.pfnDiscard = NULL;
5077 pThis->IMediaEx.pfnIoReqDiscard = NULL;
5078 }
5079
5080 if (RT_SUCCESS(rc))
5081 {
5082 LogFunc(("%d - Opened '%s' in %s mode\n",
5083 iLevel, pszName,
5084 VDIsReadOnly(pThis->pDisk) ? "read-only" : "read-write"));
5085 if ( VDIsReadOnly(pThis->pDisk)
5086 && !fReadOnly
5087 && !fMaybeReadOnly
5088 && !pThis->fTempReadOnly
5089 && iLevel == 0)
5090 {
5091 rc = PDMDrvHlpVMSetError(pDrvIns, VERR_VD_IMAGE_READ_ONLY, RT_SRC_POS,
5092 N_("Failed to open image '%s' for writing due to wrong permissions"),
5093 pszName);
5094 break;
5095 }
5096 }
5097 else
5098 {
5099 rc = PDMDrvHlpVMSetError(pDrvIns, rc, RT_SRC_POS,
5100 N_("Failed to open image '%s' in %s mode"), pszName,
5101 (uOpenFlags & VD_OPEN_FLAGS_READONLY) ? "read-only" : "read-write");
5102 break;
5103 }
5104
5105 MMR3HeapFree(pszName);
5106 pszName = NULL;
5107 MMR3HeapFree(pszFormat);
5108 pszFormat = NULL;
5109
5110 /* next */
5111 iLevel--;
5112 iImageIdx++;
5113 pCurNode = CFGMR3GetParent(pCurNode);
5114 }
5115
5116 LogRel(("VD: Opening the disk took %lld ns\n", RTTimeNanoTS() - tsStart));
5117
5118 /* Open the cache image if set. */
5119 if ( RT_SUCCESS(rc)
5120 && RT_VALID_PTR(pszCachePath))
5121 {
5122 /* Insert the custom I/O interface only if we're told to use new IO.
5123 * Since the I/O interface is per image we could make this more
5124 * flexible in the future if we want to. */
5125 if (fUseNewIo)
5126 {
5127#ifdef VBOX_WITH_PDM_ASYNC_COMPLETION
5128 pThis->VDIfIoCache.pfnOpen = drvvdAsyncIOOpen;
5129 pThis->VDIfIoCache.pfnClose = drvvdAsyncIOClose;
5130 pThis->VDIfIoCache.pfnGetSize = drvvdAsyncIOGetSize;
5131 pThis->VDIfIoCache.pfnSetSize = drvvdAsyncIOSetSize;
5132 pThis->VDIfIoCache.pfnReadSync = drvvdAsyncIOReadSync;
5133 pThis->VDIfIoCache.pfnWriteSync = drvvdAsyncIOWriteSync;
5134 pThis->VDIfIoCache.pfnFlushSync = drvvdAsyncIOFlushSync;
5135 pThis->VDIfIoCache.pfnReadAsync = drvvdAsyncIOReadAsync;
5136 pThis->VDIfIoCache.pfnWriteAsync = drvvdAsyncIOWriteAsync;
5137 pThis->VDIfIoCache.pfnFlushAsync = drvvdAsyncIOFlushAsync;
5138#else /* !VBOX_WITH_PDM_ASYNC_COMPLETION */
5139 rc = PDMDrvHlpVMSetError(pDrvIns, VERR_PDM_DRVINS_UNKNOWN_CFG_VALUES,
5140 RT_SRC_POS, N_("DrvVD: Configuration error: Async Completion Framework not compiled in"));
5141#endif /* !VBOX_WITH_PDM_ASYNC_COMPLETION */
5142 if (RT_SUCCESS(rc))
5143 rc = VDInterfaceAdd(&pThis->VDIfIoCache.Core, "DrvVD_IO", VDINTERFACETYPE_IO,
5144 pThis, sizeof(VDINTERFACEIO), &pThis->pVDIfsCache);
5145 AssertRC(rc);
5146 }
5147
5148 rc = VDCacheOpen(pThis->pDisk, pszCacheFormat, pszCachePath, VD_OPEN_FLAGS_NORMAL, pThis->pVDIfsCache);
5149 if (RT_FAILURE(rc))
5150 rc = PDMDRV_SET_ERROR(pDrvIns, rc, N_("DrvVD: Could not open cache image"));
5151 }
5152
5153 if (RT_VALID_PTR(pszCachePath))
5154 MMR3HeapFree(pszCachePath);
5155 if (RT_VALID_PTR(pszCacheFormat))
5156 MMR3HeapFree(pszCacheFormat);
5157
5158 if ( RT_SUCCESS(rc)
5159 && pThis->fMergePending
5160 && ( pThis->uMergeSource == VD_LAST_IMAGE
5161 || pThis->uMergeTarget == VD_LAST_IMAGE))
5162 {
5163 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
5164 N_("DrvVD: Configuration error: Inconsistent image merge data"));
5165 }
5166
5167 /* Create the block cache if enabled. */
5168 if ( fUseBlockCache
5169 && !pThis->fShareable
5170 && !fDiscard
5171 && !pThis->pCfgCrypto /* Disk encryption disables the block cache for security reasons */
5172 && RT_SUCCESS(rc))
5173 {
5174 /*
5175 * We need a unique ID for the block cache (to identify the owner of data
5176 * blocks in a saved state). UUIDs are not really suitable because
5177 * there are image formats which don't support them. Furthermore it is
5178 * possible that a new diff image was attached after a saved state
5179 * which changes the UUID.
5180 * However the device "name + device instance + LUN" triple the disk is
5181 * attached to is always constant for saved states.
5182 */
5183 char *pszId = NULL;
5184 uint32_t iInstance, iLUN;
5185 const char *pcszController;
5186
5187 rc = pThis->pDrvMediaPort->pfnQueryDeviceLocation(pThis->pDrvMediaPort, &pcszController,
5188 &iInstance, &iLUN);
5189 if (RT_FAILURE(rc))
5190 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
5191 N_("DrvVD: Configuration error: Could not query device data"));
5192 else
5193 {
5194 int cbStr = RTStrAPrintf(&pszId, "%s-%d-%d", pcszController, iInstance, iLUN);
5195
5196 if (cbStr > 0)
5197 {
5198 rc = PDMDrvHlpBlkCacheRetain(pDrvIns, &pThis->pBlkCache,
5199 drvvdBlkCacheXferCompleteIoReq,
5200 drvvdBlkCacheXferEnqueue,
5201 drvvdBlkCacheXferEnqueueDiscard,
5202 pszId);
5203 if (rc == VERR_NOT_SUPPORTED)
5204 {
5205 LogRel(("VD: Block cache is not supported\n"));
5206 rc = VINF_SUCCESS;
5207 }
5208 else
5209 AssertRC(rc);
5210
5211 RTStrFree(pszId);
5212 }
5213 else
5214 rc = PDMDRV_SET_ERROR(pDrvIns, VERR_PDM_DRIVER_INVALID_PROPERTIES,
5215 N_("DrvVD: Out of memory when creating block cache"));
5216 }
5217 }
5218
5219 if (RT_SUCCESS(rc))
5220 rc = drvvdSetupFilters(pThis, pCfg);
5221
5222 /*
5223 * Register a load-done callback so we can undo TempReadOnly config before
5224 * we get to drvvdResume. Automatically deregistered upon destruction.
5225 */
5226 if (RT_SUCCESS(rc))
5227 rc = PDMDrvHlpSSMRegisterEx(pDrvIns, 0 /* version */, 0 /* cbGuess */,
5228 NULL /*pfnLivePrep*/, NULL /*pfnLiveExec*/, NULL /*pfnLiveVote*/,
5229 NULL /*pfnSavePrep*/, NULL /*pfnSaveExec*/, NULL /*pfnSaveDone*/,
5230 NULL /*pfnDonePrep*/, NULL /*pfnLoadExec*/, drvvdLoadDone);
5231
5232 /* Setup the boot acceleration stuff if enabled. */
5233 if (RT_SUCCESS(rc) && pThis->fBootAccelEnabled)
5234 {
5235 pThis->cbDisk = VDGetSize(pThis->pDisk, VD_LAST_IMAGE);
5236 Assert(pThis->cbDisk > 0);
5237 pThis->pbData = (uint8_t *)RTMemAllocZ(pThis->cbBootAccelBuffer);
5238 if (pThis->pbData)
5239 {
5240 pThis->fBootAccelActive = true;
5241 pThis->offDisk = 0;
5242 pThis->cbDataValid = 0;
5243 LogRel(("VD: Boot acceleration enabled\n"));
5244 }
5245 else
5246 LogRel(("VD: Boot acceleration, out of memory, disabled\n"));
5247 }
5248
5249 if ( RTUuidIsNull(&pThis->Uuid)
5250 && pThis->enmType == PDMMEDIATYPE_HARD_DISK)
5251 VDGetUuid(pThis->pDisk, 0, &pThis->Uuid);
5252
5253 /*
5254 * Automatically upgrade the floppy drive if the specified one is too
5255 * small to represent the whole boot time image. (We cannot do this later
5256 * since the BIOS (and others) gets the info via CMOS.)
5257 *
5258 * This trick should make 2.88 images as well as the fake 15.6 and 63.5 MB
5259 * images despite the hardcoded default 1.44 drive.
5260 */
5261 if ( PDMMEDIATYPE_IS_FLOPPY(pThis->enmType)
5262 && pThis->pDisk)
5263 {
5264 uint64_t const cbFloppyImg = VDGetSize(pThis->pDisk, VD_LAST_IMAGE);
5265 PDMMEDIATYPE const enmCfgType = pThis->enmType;
5266 switch (enmCfgType)
5267 {
5268 default:
5269 AssertFailed();
5270 case PDMMEDIATYPE_FLOPPY_360:
5271 if (cbFloppyImg > 40 * 2 * 9 * 512)
5272 pThis->enmType = PDMMEDIATYPE_FLOPPY_720;
5273 /* fall thru */
5274 case PDMMEDIATYPE_FLOPPY_720:
5275 if (cbFloppyImg > 80 * 2 * 14 * 512)
5276 pThis->enmType = PDMMEDIATYPE_FLOPPY_1_20;
5277 /* fall thru */
5278 case PDMMEDIATYPE_FLOPPY_1_20:
5279 if (cbFloppyImg > 80 * 2 * 20 * 512)
5280 pThis->enmType = PDMMEDIATYPE_FLOPPY_1_44;
5281 /* fall thru */
5282 case PDMMEDIATYPE_FLOPPY_1_44:
5283 if (cbFloppyImg > 80 * 2 * 24 * 512)
5284 pThis->enmType = PDMMEDIATYPE_FLOPPY_2_88;
5285 /* fall thru */
5286 case PDMMEDIATYPE_FLOPPY_2_88:
5287 if (cbFloppyImg > 80 * 2 * 48 * 512)
5288 pThis->enmType = PDMMEDIATYPE_FLOPPY_FAKE_15_6;
5289 /* fall thru */
5290 case PDMMEDIATYPE_FLOPPY_FAKE_15_6:
5291 if (cbFloppyImg > 255 * 2 * 63 * 512)
5292 pThis->enmType = PDMMEDIATYPE_FLOPPY_FAKE_63_5;
5293 case PDMMEDIATYPE_FLOPPY_FAKE_63_5:
5294 if (cbFloppyImg > 255 * 2 * 255 * 512)
5295 LogRel(("Warning: Floppy image is larger that 63.5 MB! (%llu bytes)\n", cbFloppyImg));
5296 break;
5297 }
5298 if (pThis->enmType != enmCfgType)
5299 LogRel(("DrvVD: Automatically upgraded floppy drive from %s to %s to better support the %u byte image\n",
5300 drvvdGetTypeName(enmCfgType), drvvdGetTypeName(pThis->enmType), cbFloppyImg));
5301 }
5302 } /* !fEmptyDrive */
5303
5304 if (RT_FAILURE(rc))
5305 {
5306 if (RT_VALID_PTR(pszName))
5307 MMR3HeapFree(pszName);
5308 if (RT_VALID_PTR(pszFormat))
5309 MMR3HeapFree(pszFormat);
5310 /* drvvdDestruct does the rest. */
5311 }
5312
5313 LogFlowFunc(("returns %Rrc\n", rc));
5314 return rc;
5315}
5316
5317/**
5318 * VBox disk container media driver registration record.
5319 */
5320const PDMDRVREG g_DrvVD =
5321{
5322 /* u32Version */
5323 PDM_DRVREG_VERSION,
5324 /* szName */
5325 "VD",
5326 /* szRCMod */
5327 "",
5328 /* szR0Mod */
5329 "",
5330 /* pszDescription */
5331 "Generic VBox disk media driver.",
5332 /* fFlags */
5333 PDM_DRVREG_FLAGS_HOST_BITS_DEFAULT,
5334 /* fClass. */
5335 PDM_DRVREG_CLASS_MEDIA,
5336 /* cMaxInstances */
5337 ~0U,
5338 /* cbInstance */
5339 sizeof(VBOXDISK),
5340 /* pfnConstruct */
5341 drvvdConstruct,
5342 /* pfnDestruct */
5343 drvvdDestruct,
5344 /* pfnRelocate */
5345 NULL,
5346 /* pfnIOCtl */
5347 NULL,
5348 /* pfnPowerOn */
5349 drvvdPowerOn,
5350 /* pfnReset */
5351 drvvdReset,
5352 /* pfnSuspend */
5353 drvvdSuspend,
5354 /* pfnResume */
5355 drvvdResume,
5356 /* pfnAttach */
5357 NULL,
5358 /* pfnDetach */
5359 NULL,
5360 /* pfnPowerOff */
5361 drvvdPowerOff,
5362 /* pfnSoftReset */
5363 NULL,
5364 /* u32EndVersion */
5365 PDM_DRVREG_VERSION
5366};
5367
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette