VirtualBox

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 36360

Last change on this file since 36360 was 36360, checked in by vboxsync, 14 years ago

Runtime/r0drv/linux: fail to lock kernel memory for kmap()'ed memory
  • Property svn:eol-style set to native
  • Property svn:keywords set to Id Rev
File size: 47.7 KB
Line 
1/* $Revision: 36360 $ */
2/** @file
3 * IPRT - Ring-0 Memory Objects, Linux.
4 */
5
6/*
7 * Copyright (C) 2006-2007 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 * The contents of this file may alternatively be used under the terms
18 * of the Common Development and Distribution License Version 1.0
19 * (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20 * VirtualBox OSE distribution, in which case the provisions of the
21 * CDDL are applicable instead of those of the GPL.
22 *
23 * You may elect to license modified versions of this file under the
24 * terms and conditions of either the GPL or the CDDL or both.
25 */
26
27
28/*******************************************************************************
29* Header Files *
30*******************************************************************************/
31#include "the-linux-kernel.h"
32
33#include <iprt/memobj.h>
34#include <iprt/alloc.h>
35#include <iprt/assert.h>
36#include <iprt/log.h>
37#include <iprt/process.h>
38#include <iprt/string.h>
39#include "internal/memobj.h"
40
41
42/*******************************************************************************
43* Defined Constants And Macros *
44*******************************************************************************/
45/* early 2.6 kernels */
46#ifndef PAGE_SHARED_EXEC
47# define PAGE_SHARED_EXEC PAGE_SHARED
48#endif
49#ifndef PAGE_READONLY_EXEC
50# define PAGE_READONLY_EXEC PAGE_READONLY
51#endif
52
53/*
54 * 2.6.29+ kernels don't work with remap_pfn_range() anymore because
55 * track_pfn_vma_new() is apparently not defined for non-RAM pages.
56 * It should be safe to use vm_insert_page() older kernels as well.
57 */
58#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 23)
59# define VBOX_USE_INSERT_PAGE
60#endif
61#if defined(CONFIG_X86_PAE) \
62 && ( defined(HAVE_26_STYLE_REMAP_PAGE_RANGE) \
63 || ( LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \
64 && LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)))
65# define VBOX_USE_PAE_HACK
66#endif
67
68
69/*******************************************************************************
70* Structures and Typedefs *
71*******************************************************************************/
72/**
73 * The Darwin version of the memory object structure.
74 */
75typedef struct RTR0MEMOBJLNX
76{
77 /** The core structure. */
78 RTR0MEMOBJINTERNAL Core;
79 /** Set if the allocation is contiguous.
80 * This means it has to be given back as one chunk. */
81 bool fContiguous;
82 /** Set if we've vmap'ed the memory into ring-0. */
83 bool fMappedToRing0;
84 /** The pages in the apPages array. */
85 size_t cPages;
86 /** Array of struct page pointers. (variable size) */
87 struct page *apPages[1];
88} RTR0MEMOBJLNX, *PRTR0MEMOBJLNX;
89
90
91static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx);
92
93
94/**
95 * Helper that converts from a RTR0PROCESS handle to a linux task.
96 *
97 * @returns The corresponding Linux task.
98 * @param R0Process IPRT ring-0 process handle.
99 */
100struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
101{
102 /** @todo fix rtR0ProcessToLinuxTask!! */
103 return R0Process == RTR0ProcHandleSelf() ? current : NULL;
104}
105
106
107/**
108 * Compute order. Some functions allocate 2^order pages.
109 *
110 * @returns order.
111 * @param cPages Number of pages.
112 */
113static int rtR0MemObjLinuxOrder(size_t cPages)
114{
115 int iOrder;
116 size_t cTmp;
117
118 for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
119 ;
120 if (cPages & ~((size_t)1 << iOrder))
121 ++iOrder;
122
123 return iOrder;
124}
125
126
127/**
128 * Converts from RTMEM_PROT_* to Linux PAGE_*.
129 *
130 * @returns Linux page protection constant.
131 * @param fProt The IPRT protection mask.
132 * @param fKernel Whether it applies to kernel or user space.
133 */
134static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
135{
136 switch (fProt)
137 {
138 default:
139 AssertMsgFailed(("%#x %d\n", fProt, fKernel));
140 case RTMEM_PROT_NONE:
141 return PAGE_NONE;
142
143 case RTMEM_PROT_READ:
144 return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
145
146 case RTMEM_PROT_WRITE:
147 case RTMEM_PROT_WRITE | RTMEM_PROT_READ:
148 return fKernel ? PAGE_KERNEL : PAGE_SHARED;
149
150 case RTMEM_PROT_EXEC:
151 case RTMEM_PROT_EXEC | RTMEM_PROT_READ:
152#if defined(RT_ARCH_X86) || defined(RT_ARCH_AMD64)
153 if (fKernel)
154 {
155 pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
156 pgprot_val(fPg) &= ~_PAGE_RW;
157 return fPg;
158 }
159 return PAGE_READONLY_EXEC;
160#else
161 return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
162#endif
163
164 case RTMEM_PROT_WRITE | RTMEM_PROT_EXEC:
165 case RTMEM_PROT_WRITE | RTMEM_PROT_EXEC | RTMEM_PROT_READ:
166 return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
167 }
168}
169
170
171/**
172 * Internal worker that allocates physical pages and creates the memory object for them.
173 *
174 * @returns IPRT status code.
175 * @param ppMemLnx Where to store the memory object pointer.
176 * @param enmType The object type.
177 * @param cb The number of bytes to allocate.
178 * @param uAlignment The alignment of the physical memory.
179 * Only valid if fContiguous == true, ignored otherwise.
180 * @param fFlagsLnx The page allocation flags (GPFs).
181 * @param fContiguous Whether the allocation must be contiguous.
182 */
183static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
184 size_t uAlignment, unsigned fFlagsLnx, bool fContiguous)
185{
186 size_t iPage;
187 size_t const cPages = cb >> PAGE_SHIFT;
188 struct page *paPages;
189
190 /*
191 * Allocate a memory object structure that's large enough to contain
192 * the page pointer array.
193 */
194 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
195 if (!pMemLnx)
196 return VERR_NO_MEMORY;
197 pMemLnx->cPages = cPages;
198
199 if (cPages > 255)
200 {
201# ifdef __GFP_REPEAT
202 /* Try hard to allocate the memory, but the allocation attempt might fail. */
203 fFlagsLnx |= __GFP_REPEAT;
204# endif
205# ifdef __GFP_NOMEMALLOC
206 /* Introduced with Linux 2.6.12: Don't use emergency reserves */
207 fFlagsLnx |= __GFP_NOMEMALLOC;
208# endif
209 }
210
211 /*
212 * Allocate the pages.
213 * For small allocations we'll try contiguous first and then fall back on page by page.
214 */
215#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
216 if ( fContiguous
217 || cb <= PAGE_SIZE * 2)
218 {
219# ifdef VBOX_USE_INSERT_PAGE
220 paPages = alloc_pages(fFlagsLnx | __GFP_COMP, rtR0MemObjLinuxOrder(cPages));
221# else
222 paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
223# endif
224 if (paPages)
225 {
226 fContiguous = true;
227 for (iPage = 0; iPage < cPages; iPage++)
228 pMemLnx->apPages[iPage] = &paPages[iPage];
229 }
230 else if (fContiguous)
231 {
232 rtR0MemObjDelete(&pMemLnx->Core);
233 return VERR_NO_MEMORY;
234 }
235 }
236
237 if (!fContiguous)
238 {
239 for (iPage = 0; iPage < cPages; iPage++)
240 {
241 pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx);
242 if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
243 {
244 while (iPage-- > 0)
245 __free_page(pMemLnx->apPages[iPage]);
246 rtR0MemObjDelete(&pMemLnx->Core);
247 return VERR_NO_MEMORY;
248 }
249 }
250 }
251
252#else /* < 2.4.22 */
253 /** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
254 paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
255 if (!paPages)
256 {
257 rtR0MemObjDelete(&pMemLnx->Core);
258 return VERR_NO_MEMORY;
259 }
260 for (iPage = 0; iPage < cPages; iPage++)
261 {
262 pMemLnx->apPages[iPage] = &paPages[iPage];
263 MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
264 if (PageHighMem(pMemLnx->apPages[iPage]))
265 BUG();
266 }
267
268 fContiguous = true;
269#endif /* < 2.4.22 */
270 pMemLnx->fContiguous = fContiguous;
271
272 /*
273 * Reserve the pages.
274 */
275 for (iPage = 0; iPage < cPages; iPage++)
276 SetPageReserved(pMemLnx->apPages[iPage]);
277
278 /*
279 * Note that the physical address of memory allocated with alloc_pages(flags, order)
280 * is always 2^(PAGE_SHIFT+order)-aligned.
281 */
282 if ( fContiguous
283 && uAlignment > PAGE_SIZE)
284 {
285 /*
286 * Check for alignment constraints. The physical address of memory allocated with
287 * alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
288 */
289 if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
290 {
291 /*
292 * This should never happen!
293 */
294 printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
295 (unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
296 rtR0MemObjLinuxFreePages(pMemLnx);
297 return VERR_NO_MEMORY;
298 }
299 }
300
301 *ppMemLnx = pMemLnx;
302 return VINF_SUCCESS;
303}
304
305
306/**
307 * Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
308 *
309 * This method does NOT free the object.
310 *
311 * @param pMemLnx The object which physical pages should be freed.
312 */
313static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
314{
315 size_t iPage = pMemLnx->cPages;
316 if (iPage > 0)
317 {
318 /*
319 * Restore the page flags.
320 */
321 while (iPage-- > 0)
322 {
323 ClearPageReserved(pMemLnx->apPages[iPage]);
324#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
325#else
326 MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
327#endif
328 }
329
330 /*
331 * Free the pages.
332 */
333#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
334 if (!pMemLnx->fContiguous)
335 {
336 iPage = pMemLnx->cPages;
337 while (iPage-- > 0)
338 __free_page(pMemLnx->apPages[iPage]);
339 }
340 else
341#endif
342 __free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
343
344 pMemLnx->cPages = 0;
345 }
346}
347
348
349/**
350 * Maps the allocation into ring-0.
351 *
352 * This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
353 *
354 * Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
355 * space, so we'll use that mapping if possible. If execute access is required, we'll
356 * play safe and do our own mapping.
357 *
358 * @returns IPRT status code.
359 * @param pMemLnx The linux memory object to map.
360 * @param fExecutable Whether execute access is required.
361 */
362static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
363{
364 int rc = VINF_SUCCESS;
365
366 /*
367 * Choose mapping strategy.
368 */
369 bool fMustMap = fExecutable
370 || !pMemLnx->fContiguous;
371 if (!fMustMap)
372 {
373 size_t iPage = pMemLnx->cPages;
374 while (iPage-- > 0)
375 if (PageHighMem(pMemLnx->apPages[iPage]))
376 {
377 fMustMap = true;
378 break;
379 }
380 }
381
382 Assert(!pMemLnx->Core.pv);
383 Assert(!pMemLnx->fMappedToRing0);
384
385 if (fMustMap)
386 {
387 /*
388 * Use vmap - 2.4.22 and later.
389 */
390#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
391 pgprot_t fPg;
392 pgprot_val(fPg) = _PAGE_PRESENT | _PAGE_RW;
393# ifdef _PAGE_NX
394 if (!fExecutable)
395 pgprot_val(fPg) |= _PAGE_NX;
396# endif
397
398# ifdef VM_MAP
399 pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
400# else
401 pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
402# endif
403 if (pMemLnx->Core.pv)
404 pMemLnx->fMappedToRing0 = true;
405 else
406 rc = VERR_MAP_FAILED;
407#else /* < 2.4.22 */
408 rc = VERR_NOT_SUPPORTED;
409#endif
410 }
411 else
412 {
413 /*
414 * Use the kernel RAM mapping.
415 */
416 pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
417 Assert(pMemLnx->Core.pv);
418 }
419
420 return rc;
421}
422
423
424/**
425 * Undos what rtR0MemObjLinuxVMap() did.
426 *
427 * @param pMemLnx The linux memory object.
428 */
429static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
430{
431#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
432 if (pMemLnx->fMappedToRing0)
433 {
434 Assert(pMemLnx->Core.pv);
435 vunmap(pMemLnx->Core.pv);
436 pMemLnx->fMappedToRing0 = false;
437 }
438#else /* < 2.4.22 */
439 Assert(!pMemLnx->fMappedToRing0);
440#endif
441 pMemLnx->Core.pv = NULL;
442}
443
444
445int rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
446{
447 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
448
449 /*
450 * Release any memory that we've allocated or locked.
451 */
452 switch (pMemLnx->Core.enmType)
453 {
454 case RTR0MEMOBJTYPE_LOW:
455 case RTR0MEMOBJTYPE_PAGE:
456 case RTR0MEMOBJTYPE_CONT:
457 case RTR0MEMOBJTYPE_PHYS:
458 case RTR0MEMOBJTYPE_PHYS_NC:
459 rtR0MemObjLinuxVUnmap(pMemLnx);
460 rtR0MemObjLinuxFreePages(pMemLnx);
461 break;
462
463 case RTR0MEMOBJTYPE_LOCK:
464 if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
465 {
466 struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
467 size_t iPage;
468 Assert(pTask);
469 if (pTask && pTask->mm)
470 down_read(&pTask->mm->mmap_sem);
471
472 iPage = pMemLnx->cPages;
473 while (iPage-- > 0)
474 {
475 if (!PageReserved(pMemLnx->apPages[iPage]))
476 SetPageDirty(pMemLnx->apPages[iPage]);
477 page_cache_release(pMemLnx->apPages[iPage]);
478 }
479
480 if (pTask && pTask->mm)
481 up_read(&pTask->mm->mmap_sem);
482 }
483 /* else: kernel memory - nothing to do here. */
484 break;
485
486 case RTR0MEMOBJTYPE_RES_VIRT:
487 Assert(pMemLnx->Core.pv);
488 if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
489 {
490 struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
491 Assert(pTask);
492 if (pTask && pTask->mm)
493 {
494 down_write(&pTask->mm->mmap_sem);
495 MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
496 up_write(&pTask->mm->mmap_sem);
497 }
498 }
499 else
500 {
501 vunmap(pMemLnx->Core.pv);
502
503 Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
504 __free_page(pMemLnx->apPages[0]);
505 pMemLnx->apPages[0] = NULL;
506 pMemLnx->cPages = 0;
507 }
508 pMemLnx->Core.pv = NULL;
509 break;
510
511 case RTR0MEMOBJTYPE_MAPPING:
512 Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
513 if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
514 {
515 struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
516 Assert(pTask);
517 if (pTask && pTask->mm)
518 {
519 down_write(&pTask->mm->mmap_sem);
520 MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
521 up_write(&pTask->mm->mmap_sem);
522 }
523 }
524 else
525 vunmap(pMemLnx->Core.pv);
526 pMemLnx->Core.pv = NULL;
527 break;
528
529 default:
530 AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
531 return VERR_INTERNAL_ERROR;
532 }
533 return VINF_SUCCESS;
534}
535
536
537int rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
538{
539 PRTR0MEMOBJLNX pMemLnx;
540 int rc;
541
542#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
543 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER, false /* non-contiguous */);
544#else
545 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER, false /* non-contiguous */);
546#endif
547 if (RT_SUCCESS(rc))
548 {
549 rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
550 if (RT_SUCCESS(rc))
551 {
552 *ppMem = &pMemLnx->Core;
553 return rc;
554 }
555
556 rtR0MemObjLinuxFreePages(pMemLnx);
557 rtR0MemObjDelete(&pMemLnx->Core);
558 }
559
560 return rc;
561}
562
563
564int rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
565{
566 PRTR0MEMOBJLNX pMemLnx;
567 int rc;
568
569 /* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
570#if (defined(RT_ARCH_AMD64) || defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
571 /* ZONE_DMA32: 0-4GB */
572 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32, false /* non-contiguous */);
573 if (RT_FAILURE(rc))
574#endif
575#ifdef RT_ARCH_AMD64
576 /* ZONE_DMA: 0-16MB */
577 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA, false /* non-contiguous */);
578#else
579# ifdef CONFIG_X86_PAE
580# endif
581 /* ZONE_NORMAL: 0-896MB */
582 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER, false /* non-contiguous */);
583#endif
584 if (RT_SUCCESS(rc))
585 {
586 rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
587 if (RT_SUCCESS(rc))
588 {
589 *ppMem = &pMemLnx->Core;
590 return rc;
591 }
592
593 rtR0MemObjLinuxFreePages(pMemLnx);
594 rtR0MemObjDelete(&pMemLnx->Core);
595 }
596
597 return rc;
598}
599
600
601int rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
602{
603 PRTR0MEMOBJLNX pMemLnx;
604 int rc;
605
606#if (defined(RT_ARCH_AMD64) || defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
607 /* ZONE_DMA32: 0-4GB */
608 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32, true /* contiguous */);
609 if (RT_FAILURE(rc))
610#endif
611#ifdef RT_ARCH_AMD64
612 /* ZONE_DMA: 0-16MB */
613 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA, true /* contiguous */);
614#else
615 /* ZONE_NORMAL (32-bit hosts): 0-896MB */
616 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER, true /* contiguous */);
617#endif
618 if (RT_SUCCESS(rc))
619 {
620 rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
621 if (RT_SUCCESS(rc))
622 {
623#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) || defined(CONFIG_HIGHMEM64G))
624 size_t iPage = pMemLnx->cPages;
625 while (iPage-- > 0)
626 Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
627#endif
628 pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
629 *ppMem = &pMemLnx->Core;
630 return rc;
631 }
632
633 rtR0MemObjLinuxFreePages(pMemLnx);
634 rtR0MemObjDelete(&pMemLnx->Core);
635 }
636
637 return rc;
638}
639
640
641/**
642 * Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
643 *
644 * @returns IPRT status.
645 * @param ppMemLnx Where to
646 * @param enmType The object type.
647 * @param cb The size of the allocation.
648 * @param uAlignment The alignment of the physical memory.
649 * Only valid for fContiguous == true, ignored otherwise.
650 * @param PhysHighest See rtR0MemObjNativeAllocPhys.
651 * @param fGfp The Linux GFP flags to use for the allocation.
652 */
653static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
654 size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, unsigned fGfp)
655{
656 PRTR0MEMOBJLNX pMemLnx;
657 int rc;
658
659 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
660 enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */);
661 if (RT_FAILURE(rc))
662 return rc;
663
664 /*
665 * Check the addresses if necessary. (Can be optimized a bit for PHYS.)
666 */
667 if (PhysHighest != NIL_RTHCPHYS)
668 {
669 size_t iPage = pMemLnx->cPages;
670 while (iPage-- > 0)
671 if (page_to_phys(pMemLnx->apPages[iPage]) >= PhysHighest)
672 {
673 rtR0MemObjLinuxFreePages(pMemLnx);
674 rtR0MemObjDelete(&pMemLnx->Core);
675 return VERR_NO_MEMORY;
676 }
677 }
678
679 /*
680 * Complete the object.
681 */
682 if (enmType == RTR0MEMOBJTYPE_PHYS)
683 {
684 pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
685 pMemLnx->Core.u.Phys.fAllocated = true;
686 }
687 *ppMem = &pMemLnx->Core;
688 return rc;
689}
690
691
692/**
693 * Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
694 *
695 * @returns IPRT status.
696 * @param ppMem Where to store the memory object pointer on success.
697 * @param enmType The object type.
698 * @param cb The size of the allocation.
699 * @param uAlignment The alignment of the physical memory.
700 * Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
701 * @param PhysHighest See rtR0MemObjNativeAllocPhys.
702 */
703static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
704 size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
705{
706 int rc;
707
708 /*
709 * There are two clear cases and that's the <=16MB and anything-goes ones.
710 * When the physical address limit is somewhere in-between those two we'll
711 * just have to try, starting with HIGHUSER and working our way thru the
712 * different types, hoping we'll get lucky.
713 *
714 * We should probably move this physical address restriction logic up to
715 * the page alloc function as it would be more efficient there. But since
716 * we don't expect this to be a performance issue just yet it can wait.
717 */
718 if (PhysHighest == NIL_RTHCPHYS)
719 /* ZONE_HIGHMEM: the whole physical memory */
720 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
721 else if (PhysHighest <= _1M * 16)
722 /* ZONE_DMA: 0-16MB */
723 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
724 else
725 {
726 rc = VERR_NO_MEMORY;
727 if (RT_FAILURE(rc))
728 /* ZONE_HIGHMEM: the whole physical memory */
729 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
730 if (RT_FAILURE(rc))
731 /* ZONE_NORMAL: 0-896MB */
732 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
733#ifdef GFP_DMA32
734 if (RT_FAILURE(rc))
735 /* ZONE_DMA32: 0-4GB */
736 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
737#endif
738 if (RT_FAILURE(rc))
739 /* ZONE_DMA: 0-16MB */
740 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
741 }
742 return rc;
743}
744
745
746int rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
747{
748 return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
749}
750
751
752int rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
753{
754 return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
755}
756
757
758int rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
759{
760 /*
761 * All we need to do here is to validate that we can use
762 * ioremap on the specified address (32/64-bit dma_addr_t).
763 */
764 PRTR0MEMOBJLNX pMemLnx;
765 dma_addr_t PhysAddr = Phys;
766 AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
767
768 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
769 if (!pMemLnx)
770 return VERR_NO_MEMORY;
771
772 pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
773 pMemLnx->Core.u.Phys.fAllocated = false;
774 pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
775 Assert(!pMemLnx->cPages);
776 *ppMem = &pMemLnx->Core;
777 return VINF_SUCCESS;
778}
779
780
781int rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
782{
783 const int cPages = cb >> PAGE_SHIFT;
784 struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
785 struct vm_area_struct **papVMAs;
786 PRTR0MEMOBJLNX pMemLnx;
787 int rc = VERR_NO_MEMORY;
788 NOREF(fAccess);
789
790 /*
791 * Check for valid task and size overflows.
792 */
793 if (!pTask)
794 return VERR_NOT_SUPPORTED;
795 if (((size_t)cPages << PAGE_SHIFT) != cb)
796 return VERR_OUT_OF_RANGE;
797
798 /*
799 * Allocate the memory object and a temporary buffer for the VMAs.
800 */
801 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
802 if (!pMemLnx)
803 return VERR_NO_MEMORY;
804
805 papVMAs = (struct vm_area_struct **)RTMemAlloc(sizeof(*papVMAs) * cPages);
806 if (papVMAs)
807 {
808 down_read(&pTask->mm->mmap_sem);
809
810 /*
811 * Get user pages.
812 */
813 rc = get_user_pages(pTask, /* Task for fault accounting. */
814 pTask->mm, /* Whose pages. */
815 R3Ptr, /* Where from. */
816 cPages, /* How many pages. */
817 1, /* Write to memory. */
818 0, /* force. */
819 &pMemLnx->apPages[0], /* Page array. */
820 papVMAs); /* vmas */
821 if (rc == cPages)
822 {
823 /*
824 * Flush dcache (required?), protect against fork and _really_ pin the page
825 * table entries. get_user_pages() will protect against swapping out the
826 * pages but it will NOT protect against removing page table entries. This
827 * can be achieved with
828 * - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
829 * an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
830 * Usual Linux distributions support only a limited size of locked pages
831 * (e.g. 32KB).
832 * - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
833 * or by
834 * - setting the VM_LOCKED flag. This is the same as doing mlock() without
835 * a range check.
836 */
837 /** @todo The Linux fork() protection will require more work if this API
838 * is to be used for anything but locking VM pages. */
839 while (rc-- > 0)
840 {
841 flush_dcache_page(pMemLnx->apPages[rc]);
842 papVMAs[rc]->vm_flags |= (VM_DONTCOPY | VM_LOCKED);
843 }
844
845 up_read(&pTask->mm->mmap_sem);
846
847 RTMemFree(papVMAs);
848
849 pMemLnx->Core.u.Lock.R0Process = R0Process;
850 pMemLnx->cPages = cPages;
851 Assert(!pMemLnx->fMappedToRing0);
852 *ppMem = &pMemLnx->Core;
853
854 return VINF_SUCCESS;
855 }
856
857 /*
858 * Failed - we need to unlock any pages that we succeeded to lock.
859 */
860 while (rc-- > 0)
861 {
862 if (!PageReserved(pMemLnx->apPages[rc]))
863 SetPageDirty(pMemLnx->apPages[rc]);
864 page_cache_release(pMemLnx->apPages[rc]);
865 }
866
867 up_read(&pTask->mm->mmap_sem);
868
869 RTMemFree(papVMAs);
870 rc = VERR_LOCK_FAILED;
871 }
872
873 rtR0MemObjDelete(&pMemLnx->Core);
874 return rc;
875}
876
877
878int rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
879{
880 void *pvLast = (uint8_t *)pv + cb - 1;
881 size_t const cPages = cb >> PAGE_SHIFT;
882 PRTR0MEMOBJLNX pMemLnx;
883 bool fLinearMapping;
884 int rc;
885 uint8_t *pbPage;
886 size_t iPage;
887 NOREF(fAccess);
888
889 /*
890 * Classify the memory and check that we can deal with it.
891 */
892#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0)
893 fLinearMapping = virt_addr_valid(pvLast) && virt_addr_valid(pv);
894#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 0)
895 fLinearMapping = VALID_PAGE(virt_to_page(pvLast)) && VALID_PAGE(virt_to_page(pv));
896#else
897# error "not supported"
898#endif
899 /*
900 * kmap()'ed memory. Only relevant for 32-bit Linux kernels with HIGHMEM
901 * enabled. Unfortunately there is no easy way to retrieve the page object
902 * for such temporarily mapped memory, virt_to_page() does not work here.
903 * There is even no function to check if a virtual address is inside the
904 * kmap() area or not :-( kmap_atomic_to_page() looks promising but the test
905 * 'if (vaddr < FIXADDR_START)' if wrong -- the kmap() area is located
906 * below the fixmap area. vmalloc_to_page() would work but is only allowed
907 * for vmalloc'ed memory.
908 */
909#ifdef CONFIG_HIGHMEM
910 if (pv < PKMAP_BASE + LAST_PKMAP*PAGE_SIZE && pvLast >= PKMAP_BASE)
911 return VERR_INVALID_PARAMETER;
912#endif
913 if (!fLinearMapping)
914 {
915#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 19)
916 if ( !RTR0MemKernelIsValidAddr(pv)
917 || !RTR0MemKernelIsValidAddr(pv + cb))
918#endif
919 return VERR_INVALID_PARAMETER;
920 }
921
922 /*
923 * Allocate the memory object.
924 */
925 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
926 if (!pMemLnx)
927 return VERR_NO_MEMORY;
928
929 /*
930 * Gather the pages.
931 * We ASSUME all kernel pages are non-swappable.
932 */
933 rc = VINF_SUCCESS;
934 pbPage = (uint8_t *)pvLast;
935 iPage = cPages;
936#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 19)
937 if (!fLinearMapping)
938 {
939 while (iPage-- > 0)
940 {
941 struct page *pPage = vmalloc_to_page(pbPage);
942 if (RT_UNLIKELY(!pPage))
943 {
944 rc = VERR_LOCK_FAILED;
945 break;
946 }
947 pMemLnx->apPages[iPage] = pPage;
948 pbPage -= PAGE_SIZE;
949 }
950 }
951 else
952#endif
953 {
954 while (iPage-- > 0)
955 {
956 pMemLnx->apPages[iPage] = virt_to_page(pbPage);
957 pbPage -= PAGE_SIZE;
958 }
959 }
960 if (RT_SUCCESS(rc))
961 {
962 /*
963 * Complete the memory object and return.
964 */
965 pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
966 pMemLnx->cPages = cPages;
967 Assert(!pMemLnx->fMappedToRing0);
968 *ppMem = &pMemLnx->Core;
969
970 return VINF_SUCCESS;
971 }
972
973 rtR0MemObjDelete(&pMemLnx->Core);
974 return rc;
975}
976
977
978int rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
979{
980#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
981 const size_t cPages = cb >> PAGE_SHIFT;
982 struct page *pDummyPage;
983 struct page **papPages;
984
985 /* check for unsupported stuff. */
986 AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
987 if (uAlignment > PAGE_SIZE)
988 return VERR_NOT_SUPPORTED;
989
990 /*
991 * Allocate a dummy page and create a page pointer array for vmap such that
992 * the dummy page is mapped all over the reserved area.
993 */
994 pDummyPage = alloc_page(GFP_HIGHUSER);
995 if (!pDummyPage)
996 return VERR_NO_MEMORY;
997 papPages = RTMemAlloc(sizeof(*papPages) * cPages);
998 if (papPages)
999 {
1000 void *pv;
1001 size_t iPage = cPages;
1002 while (iPage-- > 0)
1003 papPages[iPage] = pDummyPage;
1004# ifdef VM_MAP
1005 pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1006# else
1007 pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1008# endif
1009 RTMemFree(papPages);
1010 if (pv)
1011 {
1012 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1013 if (pMemLnx)
1014 {
1015 pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1016 pMemLnx->cPages = 1;
1017 pMemLnx->apPages[0] = pDummyPage;
1018 *ppMem = &pMemLnx->Core;
1019 return VINF_SUCCESS;
1020 }
1021 vunmap(pv);
1022 }
1023 }
1024 __free_page(pDummyPage);
1025 return VERR_NO_MEMORY;
1026
1027#else /* < 2.4.22 */
1028 /*
1029 * Could probably use ioremap here, but the caller is in a better position than us
1030 * to select some safe physical memory.
1031 */
1032 return VERR_NOT_SUPPORTED;
1033#endif
1034}
1035
1036
1037/**
1038 * Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
1039 * an empty user space mapping.
1040 *
1041 * The caller takes care of acquiring the mmap_sem of the task.
1042 *
1043 * @returns Pointer to the mapping.
1044 * (void *)-1 on failure.
1045 * @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
1046 * @param cb The size of the mapping.
1047 * @param uAlignment The alignment of the mapping.
1048 * @param pTask The Linux task to create this mapping in.
1049 * @param fProt The RTMEM_PROT_* mask.
1050 */
1051static void *rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct *pTask, unsigned fProt)
1052{
1053 unsigned fLnxProt;
1054 unsigned long ulAddr;
1055
1056 /*
1057 * Convert from IPRT protection to mman.h PROT_ and call do_mmap.
1058 */
1059 fProt &= (RTMEM_PROT_NONE | RTMEM_PROT_READ | RTMEM_PROT_WRITE | RTMEM_PROT_EXEC);
1060 if (fProt == RTMEM_PROT_NONE)
1061 fLnxProt = PROT_NONE;
1062 else
1063 {
1064 fLnxProt = 0;
1065 if (fProt & RTMEM_PROT_READ)
1066 fLnxProt |= PROT_READ;
1067 if (fProt & RTMEM_PROT_WRITE)
1068 fLnxProt |= PROT_WRITE;
1069 if (fProt & RTMEM_PROT_EXEC)
1070 fLnxProt |= PROT_EXEC;
1071 }
1072
1073 if (R3PtrFixed != (RTR3PTR)-1)
1074 ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED | MAP_ANONYMOUS | MAP_FIXED, 0);
1075 else
1076 {
1077 ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED | MAP_ANONYMOUS, 0);
1078 if ( !(ulAddr & ~PAGE_MASK)
1079 && (ulAddr & (uAlignment - 1)))
1080 {
1081 /** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
1082 * up alignment gaps. This is of course complicated by fragmentation (which we might have cause
1083 * ourselves) and further by there begin two mmap strategies (top / bottom). */
1084 /* For now, just ignore uAlignment requirements... */
1085 }
1086 }
1087 if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
1088 return (void *)-1;
1089 return (void *)ulAddr;
1090}
1091
1092
1093int rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1094{
1095 PRTR0MEMOBJLNX pMemLnx;
1096 void *pv;
1097 struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1098 if (!pTask)
1099 return VERR_NOT_SUPPORTED;
1100
1101 /*
1102 * Check that the specified alignment is supported.
1103 */
1104 if (uAlignment > PAGE_SIZE)
1105 return VERR_NOT_SUPPORTED;
1106
1107 /*
1108 * Let rtR0MemObjLinuxDoMmap do the difficult bits.
1109 */
1110 down_write(&pTask->mm->mmap_sem);
1111 pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1112 up_write(&pTask->mm->mmap_sem);
1113 if (pv == (void *)-1)
1114 return VERR_NO_MEMORY;
1115
1116 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1117 if (!pMemLnx)
1118 {
1119 down_write(&pTask->mm->mmap_sem);
1120 MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, cb);
1121 up_write(&pTask->mm->mmap_sem);
1122 return VERR_NO_MEMORY;
1123 }
1124
1125 pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1126 *ppMem = &pMemLnx->Core;
1127 return VINF_SUCCESS;
1128}
1129
1130
1131int rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment,
1132 unsigned fProt, size_t offSub, size_t cbSub)
1133{
1134 int rc = VERR_NO_MEMORY;
1135 PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1136 PRTR0MEMOBJLNX pMemLnx;
1137
1138 /* Fail if requested to do something we can't. */
1139 AssertMsgReturn(!offSub && !cbSub, ("%#x %#x\n", offSub, cbSub), VERR_NOT_SUPPORTED);
1140 AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1141 if (uAlignment > PAGE_SIZE)
1142 return VERR_NOT_SUPPORTED;
1143
1144 /*
1145 * Create the IPRT memory object.
1146 */
1147 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1148 if (pMemLnx)
1149 {
1150 if (pMemLnxToMap->cPages)
1151 {
1152#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1153 /*
1154 * Use vmap - 2.4.22 and later.
1155 */
1156 pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1157# ifdef VM_MAP
1158 pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
1159# else
1160 pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
1161# endif
1162 if (pMemLnx->Core.pv)
1163 {
1164 pMemLnx->fMappedToRing0 = true;
1165 rc = VINF_SUCCESS;
1166 }
1167 else
1168 rc = VERR_MAP_FAILED;
1169
1170#else /* < 2.4.22 */
1171 /*
1172 * Only option here is to share mappings if possible and forget about fProt.
1173 */
1174 if (rtR0MemObjIsRing3(pMemToMap))
1175 rc = VERR_NOT_SUPPORTED;
1176 else
1177 {
1178 rc = VINF_SUCCESS;
1179 if (!pMemLnxToMap->Core.pv)
1180 rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1181 if (RT_SUCCESS(rc))
1182 {
1183 Assert(pMemLnxToMap->Core.pv);
1184 pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
1185 }
1186 }
1187#endif
1188 }
1189 else
1190 {
1191 /*
1192 * MMIO / physical memory.
1193 */
1194 Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1195 pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1196 ? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb)
1197 : ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
1198 if (pMemLnx->Core.pv)
1199 {
1200 /** @todo fix protection. */
1201 rc = VINF_SUCCESS;
1202 }
1203 }
1204 if (RT_SUCCESS(rc))
1205 {
1206 pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1207 *ppMem = &pMemLnx->Core;
1208 return VINF_SUCCESS;
1209 }
1210 rtR0MemObjDelete(&pMemLnx->Core);
1211 }
1212
1213 return rc;
1214}
1215
1216
1217#ifdef VBOX_USE_PAE_HACK
1218/**
1219 * Replace the PFN of a PTE with the address of the actual page.
1220 *
1221 * The caller maps a reserved dummy page at the address with the desired access
1222 * and flags.
1223 *
1224 * This hack is required for older Linux kernels which don't provide
1225 * remap_pfn_range().
1226 *
1227 * @returns 0 on success, -ENOMEM on failure.
1228 * @param mm The memory context.
1229 * @param ulAddr The mapping address.
1230 * @param Phys The physical address of the page to map.
1231 */
1232static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1233{
1234 int rc = -ENOMEM;
1235 pgd_t *pgd;
1236
1237 spin_lock(&mm->page_table_lock);
1238
1239 pgd = pgd_offset(mm, ulAddr);
1240 if (!pgd_none(*pgd) && !pgd_bad(*pgd))
1241 {
1242 pmd_t *pmd = pmd_offset(pgd, ulAddr);
1243 if (!pmd_none(*pmd))
1244 {
1245 pte_t *ptep = pte_offset_map(pmd, ulAddr);
1246 if (ptep)
1247 {
1248 pte_t pte = *ptep;
1249 pte.pte_high &= 0xfff00000;
1250 pte.pte_high |= ((Phys >> 32) & 0x000fffff);
1251 pte.pte_low &= 0x00000fff;
1252 pte.pte_low |= (Phys & 0xfffff000);
1253 set_pte(ptep, pte);
1254 pte_unmap(ptep);
1255 rc = 0;
1256 }
1257 }
1258 }
1259
1260 spin_unlock(&mm->page_table_lock);
1261 return rc;
1262}
1263#endif /* VBOX_USE_PAE_HACK */
1264
1265
1266int rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1267{
1268 struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1269 PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1270 int rc = VERR_NO_MEMORY;
1271 PRTR0MEMOBJLNX pMemLnx;
1272#ifdef VBOX_USE_PAE_HACK
1273 struct page *pDummyPage;
1274 RTHCPHYS DummyPhys;
1275#endif
1276
1277 /*
1278 * Check for restrictions.
1279 */
1280 if (!pTask)
1281 return VERR_NOT_SUPPORTED;
1282 if (uAlignment > PAGE_SIZE)
1283 return VERR_NOT_SUPPORTED;
1284
1285#ifdef VBOX_USE_PAE_HACK
1286 /*
1287 * Allocate a dummy page for use when mapping the memory.
1288 */
1289 pDummyPage = alloc_page(GFP_USER);
1290 if (!pDummyPage)
1291 return VERR_NO_MEMORY;
1292 SetPageReserved(pDummyPage);
1293 DummyPhys = page_to_phys(pDummyPage);
1294#endif
1295
1296 /*
1297 * Create the IPRT memory object.
1298 */
1299 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1300 if (pMemLnx)
1301 {
1302 /*
1303 * Allocate user space mapping.
1304 */
1305 void *pv;
1306 down_write(&pTask->mm->mmap_sem);
1307 pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1308 if (pv != (void *)-1)
1309 {
1310 /*
1311 * Map page by page into the mmap area.
1312 * This is generic, paranoid and not very efficient.
1313 */
1314 pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1315 unsigned long ulAddrCur = (unsigned long)pv;
1316 const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1317 size_t iPage;
1318
1319 rc = 0;
1320 if (pMemLnxToMap->cPages)
1321 {
1322 for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1323 {
1324#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)
1325 RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1326#endif
1327#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1328 struct vm_area_struct *vma = find_vma(pTask->mm, ulAddrCur); /* this is probably the same for all the pages... */
1329 AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1330#endif
1331#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1332 /* remap_page_range() limitation on x86 */
1333 AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1334#endif
1335
1336#if defined(VBOX_USE_INSERT_PAGE) && LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
1337 rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1338 vma->vm_flags |= VM_RESERVED; /* This flag helps making 100% sure some bad stuff wont happen (swap, core, ++). */
1339#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1340 rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1341#elif defined(VBOX_USE_PAE_HACK)
1342 rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1343 if (!rc)
1344 rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1345#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1346 rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1347#else /* 2.4 */
1348 rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1349#endif
1350 if (rc)
1351 {
1352 rc = VERR_NO_MEMORY;
1353 break;
1354 }
1355 }
1356 }
1357 else
1358 {
1359 RTHCPHYS Phys;
1360 if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1361 Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1362 else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1363 Phys = pMemLnxToMap->Core.u.Cont.Phys;
1364 else
1365 {
1366 AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1367 Phys = NIL_RTHCPHYS;
1368 }
1369 if (Phys != NIL_RTHCPHYS)
1370 {
1371 for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1372 {
1373#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1374 struct vm_area_struct *vma = find_vma(pTask->mm, ulAddrCur); /* this is probably the same for all the pages... */
1375 AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1376#endif
1377#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1378 /* remap_page_range() limitation on x86 */
1379 AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1380#endif
1381
1382#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1383 rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1384#elif defined(VBOX_USE_PAE_HACK)
1385 rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1386 if (!rc)
1387 rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1388#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1389 rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1390#else /* 2.4 */
1391 rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1392#endif
1393 if (rc)
1394 {
1395 rc = VERR_NO_MEMORY;
1396 break;
1397 }
1398 }
1399 }
1400 }
1401 if (!rc)
1402 {
1403 up_write(&pTask->mm->mmap_sem);
1404#ifdef VBOX_USE_PAE_HACK
1405 __free_page(pDummyPage);
1406#endif
1407
1408 pMemLnx->Core.pv = pv;
1409 pMemLnx->Core.u.Mapping.R0Process = R0Process;
1410 *ppMem = &pMemLnx->Core;
1411 return VINF_SUCCESS;
1412 }
1413
1414 /*
1415 * Bail out.
1416 */
1417 MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, pMemLnxToMap->Core.cb);
1418 }
1419 up_write(&pTask->mm->mmap_sem);
1420 rtR0MemObjDelete(&pMemLnx->Core);
1421 }
1422#ifdef VBOX_USE_PAE_HACK
1423 __free_page(pDummyPage);
1424#endif
1425
1426 return rc;
1427}
1428
1429
1430int rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1431{
1432 NOREF(pMem);
1433 NOREF(offSub);
1434 NOREF(cbSub);
1435 NOREF(fProt);
1436 return VERR_NOT_SUPPORTED;
1437}
1438
1439
1440RTHCPHYS rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1441{
1442 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1443
1444 if (pMemLnx->cPages)
1445 return page_to_phys(pMemLnx->apPages[iPage]);
1446
1447 switch (pMemLnx->Core.enmType)
1448 {
1449 case RTR0MEMOBJTYPE_CONT:
1450 return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1451
1452 case RTR0MEMOBJTYPE_PHYS:
1453 return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1454
1455 /* the parent knows */
1456 case RTR0MEMOBJTYPE_MAPPING:
1457 return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1458
1459 /* cPages > 0 */
1460 case RTR0MEMOBJTYPE_LOW:
1461 case RTR0MEMOBJTYPE_LOCK:
1462 case RTR0MEMOBJTYPE_PHYS_NC:
1463 case RTR0MEMOBJTYPE_PAGE:
1464 default:
1465 AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1466 /* fall thru */
1467
1468 case RTR0MEMOBJTYPE_RES_VIRT:
1469 return NIL_RTHCPHYS;
1470 }
1471}
1472
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