1 | /* $Id: PGMR0DynMap.cpp 14484 2008-11-21 19:23:59Z vboxsync $ */
|
---|
2 | /** @file
|
---|
3 | * PGM - Page Manager and Monitor, ring-0 dynamic mapping cache.
|
---|
4 | */
|
---|
5 |
|
---|
6 | /*
|
---|
7 | * Copyright (C) 2008 Sun Microsystems, Inc.
|
---|
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 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
|
---|
18 | * Clara, CA 95054 USA or visit http://www.sun.com if you need
|
---|
19 | * additional information or have any questions.
|
---|
20 | */
|
---|
21 |
|
---|
22 | /*******************************************************************************
|
---|
23 | * Internal Functions *
|
---|
24 | *******************************************************************************/
|
---|
25 | #include <VBox/pgm.h>
|
---|
26 | #include "../PGMInternal.h"
|
---|
27 | #include <VBox/vm.h>
|
---|
28 | #include <VBox/sup.h>
|
---|
29 | #include <VBox/err.h>
|
---|
30 | #include <iprt/asm.h>
|
---|
31 | #include <iprt/alloc.h>
|
---|
32 | #include <iprt/assert.h>
|
---|
33 | #include <iprt/cpuset.h>
|
---|
34 | #include <iprt/memobj.h>
|
---|
35 | #include <iprt/mp.h>
|
---|
36 | #include <iprt/spinlock.h>
|
---|
37 | #include <iprt/semaphore.h>
|
---|
38 |
|
---|
39 |
|
---|
40 | /*******************************************************************************
|
---|
41 | * Defined Constants And Macros *
|
---|
42 | *******************************************************************************/
|
---|
43 | /** The max size of the mapping cache (in pages). */
|
---|
44 | #define PGMR0DYNMAP_MAX_PAGES ((8*_1M) >> PAGE_SHIFT)
|
---|
45 | /* * The max segment size. */
|
---|
46 | /** @todo #define PGMR0DYNMAP_SEG_MAX_PAGES (_1M >> PAGE_SHIFT) */
|
---|
47 | /** The number of pages we reserve per CPU. */
|
---|
48 | #define PGMR0DYNMAP_PAGES_PER_CPU 64
|
---|
49 | /** Calcs the overload threshold. Current set at 50%. */
|
---|
50 | #define PGMR0DYNMAP_CALC_OVERLOAD(cPages) ((cPages) / 2)
|
---|
51 |
|
---|
52 |
|
---|
53 | /*******************************************************************************
|
---|
54 | * Structures and Typedefs *
|
---|
55 | *******************************************************************************/
|
---|
56 | /**
|
---|
57 | * Ring-0 dynamic mapping cache segment.
|
---|
58 | *
|
---|
59 | * The dynamic mapping cache can be extended with additional segments if the
|
---|
60 | * load is found to be too high. This done the next time a VM is created, under
|
---|
61 | * the protection of the init mutex. The arrays is reallocated and the new
|
---|
62 | * segment is added to the end of these. Nothing is rehashed of course, as the
|
---|
63 | * indexes / addresses must remain unchanged.
|
---|
64 | *
|
---|
65 | * This structure is only modified while owning the init mutex or during module
|
---|
66 | * init / term.
|
---|
67 | */
|
---|
68 | typedef struct PGMR0DYNMAPSEG
|
---|
69 | {
|
---|
70 | /** Pointer to the next segment. */
|
---|
71 | struct PGMR0DYNMAPSEG *pNext;
|
---|
72 | /** The memory object for the virtual address range that we're abusing. */
|
---|
73 | RTR0MEMOBJ hMemObj;
|
---|
74 | /** The start page in the cache. (I.e. index into the arrays.) */
|
---|
75 | uint16_t iPage;
|
---|
76 | /** The number of pages this segment contributes. */
|
---|
77 | uint16_t cPages;
|
---|
78 | /** The number of page tables. */
|
---|
79 | uint16_t cPTs;
|
---|
80 | /** The memory objects for the page tables. */
|
---|
81 | RTR0MEMOBJ ahMemObjPT[1];
|
---|
82 | } PGMR0DYNMAPSEG;
|
---|
83 | /** Pointer to a ring-0 dynamic mapping cache segment. */
|
---|
84 | typedef PGMR0DYNMAPSEG *PPGMR0DYNMAPSEG;
|
---|
85 |
|
---|
86 |
|
---|
87 | /**
|
---|
88 | * Ring-0 dynamic mapping cache entry.
|
---|
89 | *
|
---|
90 | * This structure tracks
|
---|
91 | */
|
---|
92 | typedef struct PGMR0DYNMAPENTRY
|
---|
93 | {
|
---|
94 | /** The physical address of the currently mapped page.
|
---|
95 | * This is duplicate for three reasons: cache locality, cache policy of the PT
|
---|
96 | * mappings and sanity checks. */
|
---|
97 | RTHCPHYS HCPhys;
|
---|
98 | /** Pointer to the page. */
|
---|
99 | void *pvPage;
|
---|
100 | /** The number of references. */
|
---|
101 | int32_t volatile cRefs;
|
---|
102 | /** PTE pointer union. */
|
---|
103 | union PGMR0DYNMAPENTRY_PPTE
|
---|
104 | {
|
---|
105 | /** PTE pointer, 32-bit legacy version. */
|
---|
106 | PX86PTE pLegacy;
|
---|
107 | /** PTE pointer, PAE version. */
|
---|
108 | PX86PTEPAE pPae;
|
---|
109 | } uPte;
|
---|
110 | /** CPUs that haven't invalidated this entry after it's last update. */
|
---|
111 | RTCPUSET PendingSet;
|
---|
112 | } PGMR0DYNMAPENTRY;
|
---|
113 | /** Pointer to a ring-0 dynamic mapping cache entry. */
|
---|
114 | typedef PGMR0DYNMAPENTRY *PPGMR0DYNMAPENTRY;
|
---|
115 |
|
---|
116 |
|
---|
117 | /**
|
---|
118 | * Ring-0 dynamic mapping cache.
|
---|
119 | *
|
---|
120 | * This is initialized during VMMR0 module init but no segments are allocated at
|
---|
121 | * that time. Segments will be added when the first VM is started and removed
|
---|
122 | * again when the last VM shuts down, thus avoid consuming memory while dormant.
|
---|
123 | * At module termination, the remaining bits will be freed up.
|
---|
124 | */
|
---|
125 | typedef struct PGMR0DYNMAP
|
---|
126 | {
|
---|
127 | /** The usual magic number / eye catcher (PGMR0DYNMAP_MAGIC). */
|
---|
128 | uint32_t u32Magic;
|
---|
129 | /** Spinlock serializing the normal operation of the cache. */
|
---|
130 | RTSPINLOCK hSpinlock;
|
---|
131 | /** Array for tracking and managing the pages. */
|
---|
132 | PPGMR0DYNMAPENTRY paPages;
|
---|
133 | /** The cache size given as a number of pages. */
|
---|
134 | uint32_t cPages;
|
---|
135 | /** Whether it's 32-bit legacy or PAE/AMD64 paging mode. */
|
---|
136 | bool fLegacyMode;
|
---|
137 | /** The current load. */
|
---|
138 | uint32_t cLoad;
|
---|
139 | /** The max load.
|
---|
140 | * This is maintained to get trigger adding of more mapping space. */
|
---|
141 | uint32_t cMaxLoad;
|
---|
142 | /** Initialization / termination lock. */
|
---|
143 | RTSEMFASTMUTEX hInitLock;
|
---|
144 | /** The number of users (protected by hInitLock). */
|
---|
145 | uint32_t cUsers;
|
---|
146 | /** Array containing a copy of the original page tables.
|
---|
147 | * The entries are either X86PTE or X86PTEPAE according to fLegacyMode. */
|
---|
148 | void *pvSavedPTEs;
|
---|
149 | /** List of segments. */
|
---|
150 | PPGMR0DYNMAPSEG pSegHead;
|
---|
151 | /** The paging mode. */
|
---|
152 | SUPPAGINGMODE enmPgMode;
|
---|
153 | } PGMR0DYNMAP;
|
---|
154 | /** Pointer to the ring-0 dynamic mapping cache */
|
---|
155 | typedef PGMR0DYNMAP *PPGMR0DYNMAP;
|
---|
156 |
|
---|
157 | /** PGMR0DYNMAP::u32Magic. (Jens Christian Bugge Wesseltoft) */
|
---|
158 | #define PGMR0DYNMAP_MAGIC 0x19640201
|
---|
159 |
|
---|
160 |
|
---|
161 | /*******************************************************************************
|
---|
162 | * Global Variables *
|
---|
163 | *******************************************************************************/
|
---|
164 | /** Pointer to the ring-0 dynamic mapping cache. */
|
---|
165 | static PPGMR0DYNMAP g_pPGMR0DynMap;
|
---|
166 |
|
---|
167 |
|
---|
168 | /*******************************************************************************
|
---|
169 | * Internal Functions *
|
---|
170 | *******************************************************************************/
|
---|
171 | static void pgmR0DynMapReleasePage(PPGMR0DYNMAP pThis, uint32_t iPage, uint32_t cRefs);
|
---|
172 | static int pgmR0DynMapSetup(PPGMR0DYNMAP pThis);
|
---|
173 | static int pgmR0DynMapGrow(PPGMR0DYNMAP pThis);
|
---|
174 | static void pgmR0DynMapTearDown(PPGMR0DYNMAP pThis);
|
---|
175 |
|
---|
176 |
|
---|
177 | /**
|
---|
178 | * Initializes the ring-0 dynamic mapping cache.
|
---|
179 | *
|
---|
180 | * @returns VBox status code.
|
---|
181 | */
|
---|
182 | VMMR0DECL(int) PGMR0DynMapInit(void)
|
---|
183 | {
|
---|
184 | #ifndef DEBUG_bird
|
---|
185 | return VINF_SUCCESS;
|
---|
186 | #else
|
---|
187 | Assert(!g_pPGMR0DynMap);
|
---|
188 |
|
---|
189 | /*
|
---|
190 | * Create and initialize the cache instance.
|
---|
191 | */
|
---|
192 | PPGMR0DYNMAP pThis = (PPGMR0DYNMAP)RTMemAllocZ(sizeof(*pThis));
|
---|
193 | AssertLogRelReturn(pThis, VERR_NO_MEMORY);
|
---|
194 | int rc = VINF_SUCCESS;
|
---|
195 | pThis->enmPgMode = SUPR0GetPagingMode();
|
---|
196 | switch (pThis->enmPgMode)
|
---|
197 | {
|
---|
198 | case SUPPAGINGMODE_32_BIT:
|
---|
199 | case SUPPAGINGMODE_32_BIT_GLOBAL:
|
---|
200 | pThis->fLegacyMode = false;
|
---|
201 | break;
|
---|
202 | case SUPPAGINGMODE_PAE:
|
---|
203 | case SUPPAGINGMODE_PAE_GLOBAL:
|
---|
204 | case SUPPAGINGMODE_PAE_NX:
|
---|
205 | case SUPPAGINGMODE_PAE_GLOBAL_NX:
|
---|
206 | case SUPPAGINGMODE_AMD64:
|
---|
207 | case SUPPAGINGMODE_AMD64_GLOBAL:
|
---|
208 | case SUPPAGINGMODE_AMD64_NX:
|
---|
209 | case SUPPAGINGMODE_AMD64_GLOBAL_NX:
|
---|
210 | pThis->fLegacyMode = false;
|
---|
211 | break;
|
---|
212 | default:
|
---|
213 | rc = VERR_INTERNAL_ERROR;
|
---|
214 | break;
|
---|
215 | }
|
---|
216 | if (RT_SUCCESS(rc))
|
---|
217 | {
|
---|
218 | rc = RTSemFastMutexCreate(&pThis->hInitLock);
|
---|
219 | if (RT_SUCCESS(rc))
|
---|
220 | {
|
---|
221 | rc = RTSpinlockCreate(&pThis->hSpinlock);
|
---|
222 | if (RT_SUCCESS(rc))
|
---|
223 | {
|
---|
224 | pThis->u32Magic = PGMR0DYNMAP_MAGIC;
|
---|
225 | g_pPGMR0DynMap = pThis;
|
---|
226 | return VINF_SUCCESS;
|
---|
227 | }
|
---|
228 | RTSemFastMutexDestroy(pThis->hInitLock);
|
---|
229 | }
|
---|
230 | }
|
---|
231 | RTMemFree(pThis);
|
---|
232 | return rc;
|
---|
233 | #endif
|
---|
234 | }
|
---|
235 |
|
---|
236 |
|
---|
237 | /**
|
---|
238 | * Terminates the ring-0 dynamic mapping cache.
|
---|
239 | */
|
---|
240 | VMMR0DECL(void) PGMR0DynMapTerm(void)
|
---|
241 | {
|
---|
242 | #ifdef DEBUG_bird
|
---|
243 | /*
|
---|
244 | * Destroy the cache.
|
---|
245 | *
|
---|
246 | * There is not supposed to be any races here, the loader should
|
---|
247 | * make sure about that. So, don't bother locking anything.
|
---|
248 | *
|
---|
249 | * The VM objects should all be destroyed by now, so there is no
|
---|
250 | * dangling users or anything like that to clean up. This routine
|
---|
251 | * is just a mirror image of PGMR0DynMapInit.
|
---|
252 | */
|
---|
253 | PPGMR0DYNMAP pThis = g_pPGMR0DynMap;
|
---|
254 | if (pThis)
|
---|
255 | {
|
---|
256 | AssertPtr(pThis);
|
---|
257 | g_pPGMR0DynMap = NULL;
|
---|
258 |
|
---|
259 | AssertLogRelMsg(!pThis->cUsers && !pThis->paPages && !pThis->cPages,
|
---|
260 | ("cUsers=%d paPages=%p cPages=%#x\n",
|
---|
261 | pThis->cUsers, pThis->paPages, pThis->cPages));
|
---|
262 |
|
---|
263 | /* Free the associated resources. */
|
---|
264 | RTSemFastMutexDestroy(pThis->hInitLock);
|
---|
265 | pThis->hInitLock = NIL_RTSEMFASTMUTEX;
|
---|
266 | RTSpinlockDestroy(pThis->hSpinlock);
|
---|
267 | pThis->hSpinlock = NIL_RTSPINLOCK;
|
---|
268 | pThis->u32Magic = UINT32_MAX;
|
---|
269 | RTMemFree(pThis);
|
---|
270 | }
|
---|
271 | #endif
|
---|
272 | }
|
---|
273 |
|
---|
274 |
|
---|
275 | /**
|
---|
276 | * Initializes the dynamic mapping cache for a new VM.
|
---|
277 | *
|
---|
278 | * @returns VBox status code.
|
---|
279 | * @param pVM Pointer to the shared VM structure.
|
---|
280 | */
|
---|
281 | VMMR0DECL(int) PGMR0DynMapInitVM(PVM pVM)
|
---|
282 | {
|
---|
283 | #ifndef DEBUG_bird
|
---|
284 | return VINF_SUCCESS;
|
---|
285 | #else
|
---|
286 | /*
|
---|
287 | * Initialize the auto sets.
|
---|
288 | */
|
---|
289 | VMCPUID idCpu = pVM->cCPUs;
|
---|
290 | while (idCpu-- > 0)
|
---|
291 | {
|
---|
292 | PPGMMAPSET pSet = &pVM->aCpus[idCpu].pgm.s.AutoSet;
|
---|
293 | uint32_t j = RT_ELEMENTS(pSet->aEntries);
|
---|
294 | while (j-- > 0)
|
---|
295 | {
|
---|
296 | pSet->aEntries[j].iPage = UINT16_MAX;
|
---|
297 | pSet->aEntries[j].cRefs = 0;
|
---|
298 | }
|
---|
299 | pSet->cEntries = PGMMAPSET_CLOSED;
|
---|
300 | }
|
---|
301 |
|
---|
302 | /*
|
---|
303 | * Do we need the cache? Skip the last bit if we don't.
|
---|
304 | */
|
---|
305 | Assert(!pVM->pgm.s.pvR0DynMapUsed);
|
---|
306 | pVM->pgm.s.pvR0DynMapUsed = NULL;
|
---|
307 | if (!HWACCMIsEnabled(pVM))
|
---|
308 | return VINF_SUCCESS;
|
---|
309 |
|
---|
310 | /*
|
---|
311 | * Reference and if necessary setup or grow the cache.
|
---|
312 | */
|
---|
313 | PPGMR0DYNMAP pThis = g_pPGMR0DynMap;
|
---|
314 | AssertPtrReturn(pThis, VERR_INTERNAL_ERROR);
|
---|
315 | int rc = RTSemFastMutexRequest(pThis->hInitLock);
|
---|
316 | AssertLogRelRCReturn(rc, rc);
|
---|
317 |
|
---|
318 | pThis->cUsers++;
|
---|
319 | if (pThis->cUsers == 1)
|
---|
320 | rc = pgmR0DynMapSetup(pThis);
|
---|
321 | else if (pThis->cMaxLoad > PGMR0DYNMAP_CALC_OVERLOAD(pThis->cPages))
|
---|
322 | rc = pgmR0DynMapGrow(pThis);
|
---|
323 | if (RT_FAILURE(rc))
|
---|
324 | pThis->cUsers--;
|
---|
325 |
|
---|
326 | RTSemFastMutexRelease(pThis->hInitLock);
|
---|
327 |
|
---|
328 | return rc;
|
---|
329 | #endif
|
---|
330 | }
|
---|
331 |
|
---|
332 |
|
---|
333 | /**
|
---|
334 | * Terminates the dynamic mapping cache usage for a VM.
|
---|
335 | *
|
---|
336 | * @param pVM Pointer to the shared VM structure.
|
---|
337 | */
|
---|
338 | VMMR0DECL(void) PGMR0DynMapTermVM(PVM pVM)
|
---|
339 | {
|
---|
340 | #ifdef DEBUG_bird
|
---|
341 | /*
|
---|
342 | * Return immediately if we're not using the cache.
|
---|
343 | */
|
---|
344 | if (!pVM->pgm.s.pvR0DynMapUsed)
|
---|
345 | return;
|
---|
346 |
|
---|
347 | PPGMR0DYNMAP pThis = g_pPGMR0DynMap;
|
---|
348 | AssertPtrReturnVoid(pThis);
|
---|
349 |
|
---|
350 | int rc = RTSemFastMutexRequest(pThis->hInitLock);
|
---|
351 | AssertLogRelRCReturnVoid(rc);
|
---|
352 |
|
---|
353 | if (pVM->pgm.s.pvR0DynMapUsed == pThis)
|
---|
354 | {
|
---|
355 | pVM->pgm.s.pvR0DynMapUsed = NULL;
|
---|
356 |
|
---|
357 | /*
|
---|
358 | * Clean up and check the auto sets.
|
---|
359 | */
|
---|
360 | VMCPUID idCpu = pVM->cCPUs;
|
---|
361 | while (idCpu-- > 0)
|
---|
362 | {
|
---|
363 | PPGMMAPSET pSet = &pVM->aCpus[idCpu].pgm.s.AutoSet;
|
---|
364 | uint32_t j = pSet->cEntries;
|
---|
365 | if (j <= RT_ELEMENTS(pSet->aEntries))
|
---|
366 | {
|
---|
367 | /*
|
---|
368 | * The set is open, close it.
|
---|
369 | */
|
---|
370 | while (j-- > 0)
|
---|
371 | {
|
---|
372 | int32_t cRefs = pSet->aEntries[j].cRefs;
|
---|
373 | uint32_t iPage = pSet->aEntries[j].iPage;
|
---|
374 | LogRel(("PGMR0DynMapTermVM: %d dangling refs to %#x\n", cRefs, iPage));
|
---|
375 | if (iPage < pThis->cPages && cRefs > 0)
|
---|
376 | pgmR0DynMapReleasePage(pThis, iPage, cRefs);
|
---|
377 | else
|
---|
378 | AssertMsgFailed(("cRefs=%d iPage=%#x cPages=%u\n", cRefs, iPage, pThis->cPages));
|
---|
379 |
|
---|
380 | pSet->aEntries[j].iPage = UINT16_MAX;
|
---|
381 | pSet->aEntries[j].cRefs = 0;
|
---|
382 | }
|
---|
383 | pSet->cEntries = PGMMAPSET_CLOSED;
|
---|
384 | }
|
---|
385 |
|
---|
386 | j = RT_ELEMENTS(pSet->aEntries);
|
---|
387 | while (j-- > 0)
|
---|
388 | {
|
---|
389 | Assert(pSet->aEntries[j].iPage == UINT16_MAX);
|
---|
390 | Assert(!pSet->aEntries[j].cRefs);
|
---|
391 | }
|
---|
392 | }
|
---|
393 |
|
---|
394 | /*
|
---|
395 | * Release our reference to the mapping cache.
|
---|
396 | */
|
---|
397 | Assert(pThis->cUsers > 0);
|
---|
398 | pThis->cUsers--;
|
---|
399 | if (!pThis->cUsers)
|
---|
400 | pgmR0DynMapTearDown(pThis);
|
---|
401 | }
|
---|
402 | else
|
---|
403 | AssertMsgFailed(("pvR0DynMapUsed=%p pThis=%p\n", pVM->pgm.s.pvR0DynMapUsed, pThis));
|
---|
404 |
|
---|
405 | RTSemFastMutexRelease(pThis->hInitLock);
|
---|
406 | #endif
|
---|
407 | }
|
---|
408 |
|
---|
409 |
|
---|
410 | /**
|
---|
411 | * Calculate the new cache size based on cMaxLoad statistics.
|
---|
412 | *
|
---|
413 | * @returns Number of pages.
|
---|
414 | * @param pThis The dynamic mapping cache instance.
|
---|
415 | */
|
---|
416 | static uint32_t pgmR0DynMapCalcNewSize(PPGMR0DYNMAP pThis)
|
---|
417 | {
|
---|
418 | /*
|
---|
419 | * cCpus * PGMR0DYNMAP_PAGES_PER_CPU.
|
---|
420 | */
|
---|
421 | RTCPUID cCpus = RTMpGetCount();
|
---|
422 | uint32_t cPages = cCpus * PGMR0DYNMAP_PAGES_PER_CPU;
|
---|
423 |
|
---|
424 | /* adjust against cMaxLoad. */
|
---|
425 | AssertMsg(pThis->cMaxLoad <= PGMR0DYNMAP_MAX_PAGES, ("%#x\n", pThis->cMaxLoad));
|
---|
426 | if (pThis->cMaxLoad > PGMR0DYNMAP_MAX_PAGES)
|
---|
427 | pThis->cMaxLoad = 0;
|
---|
428 |
|
---|
429 | while (pThis->cMaxLoad < PGMR0DYNMAP_CALC_OVERLOAD(cPages))
|
---|
430 | cPages += PGMR0DYNMAP_PAGES_PER_CPU;
|
---|
431 |
|
---|
432 | /* adjust against max size. */
|
---|
433 | if (cPages > PGMR0DYNMAP_MAX_PAGES)
|
---|
434 | cPages = PGMR0DYNMAP_MAX_PAGES;
|
---|
435 |
|
---|
436 | return cPages;
|
---|
437 | }
|
---|
438 |
|
---|
439 |
|
---|
440 | /**
|
---|
441 | * Adds a new segment of the specified size.
|
---|
442 | *
|
---|
443 | * @returns VBox status code.
|
---|
444 | * @param pThis The dynamic mapping cache instance.
|
---|
445 | * @param cPages The size of the new segment, give as a page count.
|
---|
446 | */
|
---|
447 | static int pgmR0DynMapAddSeg(PPGMR0DYNMAP pThis, uint32_t cPages)
|
---|
448 | {
|
---|
449 | #if 0
|
---|
450 | int rc2;
|
---|
451 |
|
---|
452 | /*
|
---|
453 | * Do the array rellocation first.
|
---|
454 | * (Too lazy to clean these up on failure.)
|
---|
455 | */
|
---|
456 | void *pv = RTMemRealloc(pThis->paPages, sizeof(pThis->paPages[0]) * (pThis->cPages + cPages));
|
---|
457 | if (!pv)
|
---|
458 | return VERR_NO_MEMORY;
|
---|
459 | pThis->paPages = (PPGMR0DYNMAPENTRY)pv;
|
---|
460 |
|
---|
461 | pv = RTMemRealloc(pThis->pvSavedPTEs, (pThis->fLegacyMode ? sizeof(X86PGUINT) : sizeof(X86PGPAEUINT)) * (pThis->cPages + cPages));
|
---|
462 | if (!pv)
|
---|
463 | return VERR_NO_MEMORY;
|
---|
464 | pThis->pvSavedPTEs = pv;
|
---|
465 |
|
---|
466 | /*
|
---|
467 | * Allocate the segment structure and pages memory.
|
---|
468 | */
|
---|
469 | uint32_t cPTs = cPages / (pThis->fLegacyMode ? X86_PG_ENTRIES : X86_PG_PAE_ENTRIES) + 2;
|
---|
470 | PPGMR0DYNMAPSEG pSeg = RTMemAllocZ(RT_UOFFSETOF(PGMR0DYNMAPSEG, ahMemObjPTs[cPTs]));
|
---|
471 | if (!pSeg)
|
---|
472 | return VERR_NO_MEMORY;
|
---|
473 | pSeg->pNext = NULL;
|
---|
474 | pSeg->cPages = cPages;
|
---|
475 | pSeg->iPage = pThis->cPages;
|
---|
476 | pSeg->cPTs = 0;
|
---|
477 | int rc = RTR0MemObjAllocPage(&pSeg->hMemObj, cPages << PAGE_SHIFT, false);
|
---|
478 | if (RT_SUCCESS(rc))
|
---|
479 | {
|
---|
480 | /*
|
---|
481 | * Walk the paging hierarchy and map the relevant page tables.
|
---|
482 | */
|
---|
483 | uint8_t *pbPage = RTR0MemObjAddress(pSeg->hMemObj);
|
---|
484 | AssertMsg(VALID_PTR(pbPage) && !((uintptr_t)pbPage & PAGE_OFFSET_MASK), ("%p\n", pbPage));
|
---|
485 | uint32_t iPage = pThis->cPages;
|
---|
486 | uint32_t iEndPage = iPage + cPages;
|
---|
487 | struct
|
---|
488 | {
|
---|
489 | RTHCPHYS HCPhys; /**< The entry that's currently mapped */
|
---|
490 | RTHCPHYS fPhysMask; /**< Mask for extracting HCPhys from uEntry. */
|
---|
491 | RTR0MEMOBJ hMemObj;
|
---|
492 | RTR0MEMOBJ hMapObj;
|
---|
493 | uint64_t fPtrMask;
|
---|
494 | uint32_t fPtrShift;
|
---|
495 | uint64_t fAndMask;
|
---|
496 | uint64_t fResMask;
|
---|
497 | union
|
---|
498 | {
|
---|
499 | void *pv;
|
---|
500 | } u;
|
---|
501 | } a[4];
|
---|
502 | RTCCUINTREG cr4 = ASMGetCR4();
|
---|
503 | uint32_t cLevels;
|
---|
504 | switch (pThis->enmPgMode)
|
---|
505 | {
|
---|
506 | case SUPPAGINGMODE_32_BIT:
|
---|
507 | case SUPPAGINGMODE_32_BIT_GLOBAL:
|
---|
508 | cLevels = 2;
|
---|
509 | a[0].fAndMask = X86_PDE_P | X86_PDE_RW | (cr4 & X86_CR4_PSE ? X86_PDE_PS : 0);
|
---|
510 | a[0].fResMask = X86_PDE_P | X86_PDE_RW;
|
---|
511 | a[0].fPtrMask = X86_PD_MASK;
|
---|
512 | a[0].fPtrShift = X86_PD_SHIFT;
|
---|
513 | a[1].fAndMask = X86_PTE_P | X86_PTE_RW;
|
---|
514 | a[1].fResMask = X86_PTE_P | X86_PTE_RW;
|
---|
515 | a[1].fPtrMask = X86_PT_MASK;
|
---|
516 | a[1].fPtrShift = X86_PT_SHIFT;
|
---|
517 | break;
|
---|
518 |
|
---|
519 | case SUPPAGINGMODE_PAE:
|
---|
520 | case SUPPAGINGMODE_PAE_GLOBAL:
|
---|
521 | case SUPPAGINGMODE_PAE_NX:
|
---|
522 | case SUPPAGINGMODE_PAE_GLOBAL_NX:
|
---|
523 | cLevels = 3;
|
---|
524 | a[0].fAndMask = X86_PDPE_P;
|
---|
525 | a[0].fResMask = X86_PDPE_P;
|
---|
526 | a[0].fPtrMask = X86_PDPT_MASK_PAE;
|
---|
527 | a[0].fPtrShift = X86_PDPT_SHIFT;
|
---|
528 | a[1].fAndMask = X86_PDE_P | X86_PDE_RW | (cr4 & X86_CR4_PSE ? X86_PDE_PS : 0);
|
---|
529 | a[1].fResMask = X86_PDE_P | X86_PDE_RW;
|
---|
530 | a[1].fPtrMask = X86_PD_MASK;
|
---|
531 | a[1].fPtrShift = X86_PD_SHIFT;
|
---|
532 | a[2].fAndMask = X86_PTE_P | X86_PTE_RW;
|
---|
533 | a[2].fResMask = X86_PTE_P | X86_PTE_RW;
|
---|
534 | a[2].fPtrMask = X86_PT_MASK;
|
---|
535 | a[2].fPtrShift = X86_PT_SHIFT;
|
---|
536 | break;
|
---|
537 |
|
---|
538 | case SUPPAGINGMODE_AMD64:
|
---|
539 | case SUPPAGINGMODE_AMD64_GLOBAL:
|
---|
540 | case SUPPAGINGMODE_AMD64_NX:
|
---|
541 | case SUPPAGINGMODE_AMD64_GLOBAL_NX:
|
---|
542 | cLevels = 3;
|
---|
543 | a[0].fAndMask = X86_PML4E_P | X86_PML4E_RW;
|
---|
544 | a[0].fResMask = X86_PML4E_P | X86_PML4E_RW;
|
---|
545 | a[0].fPtrMask = X86_PML4_MASK;
|
---|
546 | a[0].fPtrShift = X86_PML4_SHIFT;
|
---|
547 | a[1].fAndMask = X86_PDPE_P | X86_PDPE_RW /** @todo check for X86_PDPT_PS support. */;
|
---|
548 | a[1].fResMask = X86_PDPE_P | X86_PDPE_RW;
|
---|
549 | a[1].fPtrMask = X86_PDPT_MASK_AMD64;
|
---|
550 | a[1].fPtrShift = X86_PDPT_SHIFT;
|
---|
551 | a[2].fAndMask = X86_PDE_P | X86_PDE_RW | (cr4 & X86_CR4_PSE ? X86_PDE_PS : 0);
|
---|
552 | a[2].fResMask = X86_PDE_P | X86_PDE_RW;
|
---|
553 | a[2].fPtrMask = X86_PD_MASK;
|
---|
554 | a[2].fPtrShift = X86_PD_SHIFT;
|
---|
555 | a[3].fAndMask = X86_PTE_P | X86_PTE_RW;
|
---|
556 | a[3].fResMask = X86_PTE_P | X86_PTE_RW;
|
---|
557 | a[3].fPtrMask = X86_PT_MASK;
|
---|
558 | a[3].fPtrShift = X86_PT_SHIFT;
|
---|
559 | break;
|
---|
560 | default:
|
---|
561 | cLevels = 0;
|
---|
562 | break;
|
---|
563 | }
|
---|
564 | for (uint32_t i = 0; i < RT_ELEMENTS(a); i++)
|
---|
565 | {
|
---|
566 | a[i].HCPhys = NIL_RTHCPHYS;
|
---|
567 | a[i].hMapObj = a[i].hMemObj = NIL_RTR0MEMOBJ;
|
---|
568 | a[i].u.pv = NULL;
|
---|
569 | }
|
---|
570 |
|
---|
571 | for (; iPage < iEndPage && RT_SUCCESS(rc); iPage++, pbPage += PAGE_SIZE)
|
---|
572 | {
|
---|
573 | /* Initialize it */
|
---|
574 | pThis->paPages[iPage].HCPhys = NIL_RTHCPHYS;
|
---|
575 | pThis->paPages[iPage].pvPage = pbPage;
|
---|
576 | pThis->paPages[iPage].cRefs = 0;
|
---|
577 | pThis->paPages[iPage].uPte.pPae = NULL;
|
---|
578 | RTCpuSetFill(&pThis->paPages[iPage].PendingSet);
|
---|
579 |
|
---|
580 | /*
|
---|
581 | * Map its page table.
|
---|
582 | *
|
---|
583 | * This is a bit ASSUMPTIVE, it should really do a clean run thru
|
---|
584 | * the tables everything something was mapped and disable preemption
|
---|
585 | * or/and interrupts.
|
---|
586 | */
|
---|
587 | X86PGPAEUINT uEntry = ASMGetCR3();
|
---|
588 | for (unsigned i = 0; i < cLevels && RT_SUCCESS(rc); i++)
|
---|
589 | {
|
---|
590 | RTHCPHYS HCPhys = uEntry & a[i].fPhysMask;
|
---|
591 | if (a[i].HCPhys != HCPhys)
|
---|
592 | {
|
---|
593 | if (i + 1 != cLevels)
|
---|
594 | {
|
---|
595 | RTR0MemObjFree(a[i].hMemObj, true /* fFreeMappings */);
|
---|
596 | a[i].hMemObj = a[i].hMapObj = NIL_RTR0MEMOBJ;
|
---|
597 | }
|
---|
598 | rc = RTR0MemObjEnterPhys(&a[i].hMemObj, HCPhys, PAGE_SIZE);
|
---|
599 | if (RT_SUCCESS(rc))
|
---|
600 | rc = RTR0MemObjMapKernel(&a[i].hMapObj, a[i].hMemObj, &a[i].u.pv, 0, RTMEM_PROT_WRITE | RTMEM_PROT_READ);
|
---|
601 | if (RT_FAILURE(rc))
|
---|
602 | break;
|
---|
603 | }
|
---|
604 |
|
---|
605 | }
|
---|
606 |
|
---|
607 |
|
---|
608 | } /* for each page */
|
---|
609 |
|
---|
610 | for (iPage = 0; i < cLevels; )
|
---|
611 |
|
---|
612 | rc2 = RTR0MemObjFree(hMemObjCR3, true /* fFreeMappings */); AssertRC(rc2);
|
---|
613 |
|
---|
614 | rc2 = RTR0MemObjFree(pSeg->hMemObj, true /* fFreeMappings */); AssertRC(rc2);
|
---|
615 | }
|
---|
616 | RTMemFree(pSeg);
|
---|
617 | return rc;
|
---|
618 | #else
|
---|
619 | return VERR_NOT_IMPLEMENTED;
|
---|
620 | #endif
|
---|
621 | }
|
---|
622 |
|
---|
623 |
|
---|
624 | /**
|
---|
625 | * Called by PGMR0DynMapInitVM under the init lock.
|
---|
626 | *
|
---|
627 | * @returns VBox status code.
|
---|
628 | * @param pThis The dynamic mapping cache instance.
|
---|
629 | */
|
---|
630 | static int pgmR0DynMapSetup(PPGMR0DYNMAP pThis)
|
---|
631 | {
|
---|
632 | /*
|
---|
633 | * Calc the size and add a segment of that size.
|
---|
634 | */
|
---|
635 | uint32_t cPages = pgmR0DynMapCalcNewSize(pThis);
|
---|
636 | AssertReturn(cPages, VERR_INTERNAL_ERROR);
|
---|
637 | return pgmR0DynMapAddSeg(pThis, cPages);
|
---|
638 | }
|
---|
639 |
|
---|
640 |
|
---|
641 | /**
|
---|
642 | * Called by PGMR0DynMapInitVM under the init lock.
|
---|
643 | *
|
---|
644 | * @returns VBox status code.
|
---|
645 | * @param pThis The dynamic mapping cache instance.
|
---|
646 | */
|
---|
647 | static int pgmR0DynMapGrow(PPGMR0DYNMAP pThis)
|
---|
648 | {
|
---|
649 | /*
|
---|
650 | * Calc the new target size and add a segment of the appropriate size.
|
---|
651 | */
|
---|
652 | uint32_t cPages = pgmR0DynMapCalcNewSize(pThis);
|
---|
653 | if (pThis->cPages >= cPages)
|
---|
654 | return VINF_SUCCESS;
|
---|
655 |
|
---|
656 | uint32_t cAdd = cPages - pThis->cPages;
|
---|
657 | return pgmR0DynMapAddSeg(pThis, cAdd);
|
---|
658 | }
|
---|
659 |
|
---|
660 |
|
---|
661 | /**
|
---|
662 | * Shoots down the TLBs for all the cache pages, pgmR0DynMapTearDown helper.
|
---|
663 | *
|
---|
664 | * @param idCpu The current CPU.
|
---|
665 | * @param pvUser1 The dynamic mapping cache instance.
|
---|
666 | * @param pvUser2 Unused, NULL.
|
---|
667 | */
|
---|
668 | static DECLCALLBACK(void) pgmR0DynMapShootDownTlbs(RTCPUID idCpu, void *pvUser1, void *pvUser2)
|
---|
669 | {
|
---|
670 | Assert(!pvUser2);
|
---|
671 | PPGMR0DYNMAP pThis = (PPGMR0DYNMAP)pvUser1;
|
---|
672 | AssertPtr(pThis == g_pPGMR0DynMap);
|
---|
673 | PPGMR0DYNMAPENTRY paPages = pThis->paPages;
|
---|
674 | uint32_t iPage = pThis->cPages;
|
---|
675 | while (iPage-- > 0)
|
---|
676 | ASMInvalidatePage(paPages[iPage].pvPage);
|
---|
677 | }
|
---|
678 |
|
---|
679 |
|
---|
680 | /**
|
---|
681 | * Called by PGMR0DynMapTermVM under the init lock.
|
---|
682 | *
|
---|
683 | * @returns VBox status code.
|
---|
684 | * @param pThis The dynamic mapping cache instance.
|
---|
685 | */
|
---|
686 | static void pgmR0DynMapTearDown(PPGMR0DYNMAP pThis)
|
---|
687 | {
|
---|
688 | /*
|
---|
689 | * Restore the original page table entries
|
---|
690 | */
|
---|
691 | PPGMR0DYNMAPENTRY paPages = pThis->paPages;
|
---|
692 | uint32_t iPage = pThis->cPages;
|
---|
693 | if (pThis->fLegacyMode)
|
---|
694 | {
|
---|
695 | X86PGUINT const *paSavedPTEs = (X86PGUINT const *)pThis->pvSavedPTEs;
|
---|
696 | while (iPage-- > 0)
|
---|
697 | {
|
---|
698 | X86PGUINT uOld = paPages[iPage].uPte.pLegacy->u;
|
---|
699 | X86PGUINT uOld2 = uOld; NOREF(uOld2);
|
---|
700 | X86PGUINT uNew = paSavedPTEs[iPage];
|
---|
701 | while (!ASMAtomicCmpXchgExU32(&paPages[iPage].uPte.pLegacy->u, uNew, uOld, &uOld))
|
---|
702 | AssertMsgFailed(("uOld=%#x uOld2=%#x uNew=%#x\n", uOld, uOld2, uNew));
|
---|
703 | }
|
---|
704 | }
|
---|
705 | else
|
---|
706 | {
|
---|
707 | X86PGPAEUINT const *paSavedPTEs = (X86PGPAEUINT const *)pThis->pvSavedPTEs;
|
---|
708 | while (iPage-- > 0)
|
---|
709 | {
|
---|
710 | X86PGPAEUINT uOld = paPages[iPage].uPte.pPae->u;
|
---|
711 | X86PGPAEUINT uOld2 = uOld; NOREF(uOld2);
|
---|
712 | X86PGPAEUINT uNew = paSavedPTEs[iPage];
|
---|
713 | while (!ASMAtomicCmpXchgExU64(&paPages[iPage].uPte.pPae->u, uNew, uOld, &uOld))
|
---|
714 | AssertMsgFailed(("uOld=%#llx uOld2=%#llx uNew=%#llx\n", uOld, uOld2, uNew));
|
---|
715 | }
|
---|
716 | }
|
---|
717 |
|
---|
718 | /*
|
---|
719 | * Shoot down the TLBs on all CPUs before freeing them.
|
---|
720 | * If RTMpOnAll fails, make sure the TLBs are invalidated on the current CPU at least.
|
---|
721 | */
|
---|
722 | int rc = RTMpOnAll(pgmR0DynMapShootDownTlbs, pThis, NULL);
|
---|
723 | AssertRC(rc);
|
---|
724 | if (RT_FAILURE(rc))
|
---|
725 | {
|
---|
726 | iPage = pThis->cPages;
|
---|
727 | while (iPage-- > 0)
|
---|
728 | ASMInvalidatePage(paPages[iPage].pvPage);
|
---|
729 | }
|
---|
730 |
|
---|
731 | /*
|
---|
732 | * Free the segments.
|
---|
733 | */
|
---|
734 | while (pThis->pSegHead)
|
---|
735 | {
|
---|
736 | PPGMR0DYNMAPSEG pSeg = pThis->pSegHead;
|
---|
737 | pThis->pSegHead = pSeg->pNext;
|
---|
738 |
|
---|
739 | uint32_t iPT = pSeg->cPTs;
|
---|
740 | while (iPT-- > 0)
|
---|
741 | {
|
---|
742 | rc = RTR0MemObjFree(pSeg->ahMemObjPT[iPT], true /* fFreeMappings */); AssertRC(rc);
|
---|
743 | pSeg->ahMemObjPT[iPT] = NIL_RTR0MEMOBJ;
|
---|
744 | }
|
---|
745 | rc = RTR0MemObjFree(pSeg->hMemObj, true /* fFreeMappings */); AssertRC(rc);
|
---|
746 | pSeg->hMemObj = NIL_RTR0MEMOBJ;
|
---|
747 | pSeg->pNext = NULL;
|
---|
748 | pSeg->iPage = UINT16_MAX;
|
---|
749 | pSeg->cPages = 0;
|
---|
750 | pSeg->cPTs = 0;
|
---|
751 | RTMemFree(pSeg);
|
---|
752 | }
|
---|
753 |
|
---|
754 | /*
|
---|
755 | * Free the arrays and restore the initial state.
|
---|
756 | * The cLoadMax value is left behind for the next setup.
|
---|
757 | */
|
---|
758 | RTMemFree(pThis->paPages);
|
---|
759 | pThis->paPages = NULL;
|
---|
760 | RTMemFree(pThis->pvSavedPTEs);
|
---|
761 | pThis->pvSavedPTEs = NULL;
|
---|
762 | pThis->cPages = 0;
|
---|
763 | pThis->cLoad = 0;
|
---|
764 | }
|
---|
765 |
|
---|
766 |
|
---|
767 | /**
|
---|
768 | * Release references to a page, caller owns the spin lock.
|
---|
769 | *
|
---|
770 | * @param pThis The dynamic mapping cache instance.
|
---|
771 | * @param iPage The page.
|
---|
772 | * @param cRefs The number of references to release.
|
---|
773 | */
|
---|
774 | DECLINLINE(void) pgmR0DynMapReleasePageLocked(PPGMR0DYNMAP pThis, uint32_t iPage, int32_t cRefs)
|
---|
775 | {
|
---|
776 | cRefs = ASMAtomicSubS32(&pThis->paPages[iPage].cRefs, cRefs);
|
---|
777 | AssertMsg(cRefs >= 0, ("%d\n", cRefs));
|
---|
778 | if (!cRefs)
|
---|
779 | pThis->cLoad--;
|
---|
780 | }
|
---|
781 |
|
---|
782 |
|
---|
783 | /**
|
---|
784 | * Release references to a page, caller does not own the spin lock.
|
---|
785 | *
|
---|
786 | * @param pThis The dynamic mapping cache instance.
|
---|
787 | * @param iPage The page.
|
---|
788 | * @param cRefs The number of references to release.
|
---|
789 | */
|
---|
790 | static void pgmR0DynMapReleasePage(PPGMR0DYNMAP pThis, uint32_t iPage, uint32_t cRefs)
|
---|
791 | {
|
---|
792 | RTSPINLOCKTMP Tmp = RTSPINLOCKTMP_INITIALIZER;
|
---|
793 | RTSpinlockAcquire(pThis->hSpinlock, &Tmp);
|
---|
794 | pgmR0DynMapReleasePageLocked(pThis, iPage, cRefs);
|
---|
795 | RTSpinlockRelease(pThis->hSpinlock, &Tmp);
|
---|
796 | }
|
---|
797 |
|
---|
798 |
|
---|
799 | /**
|
---|
800 | * pgmR0DynMapPage worker that deals with the tedious bits.
|
---|
801 | *
|
---|
802 | * @returns The page index on success, UINT32_MAX on failure.
|
---|
803 | * @param pThis The dynamic mapping cache instance.
|
---|
804 | * @param HCPhys The address of the page to be mapped.
|
---|
805 | * @param iPage The page index pgmR0DynMapPage hashed HCPhys to.
|
---|
806 | */
|
---|
807 | static uint32_t pgmR0DynMapPageSlow(PPGMR0DYNMAP pThis, RTHCPHYS HCPhys, uint32_t iPage)
|
---|
808 | {
|
---|
809 | /*
|
---|
810 | * Check if any of the first 5 pages are unreferenced since the caller
|
---|
811 | * already has made sure they aren't matching.
|
---|
812 | */
|
---|
813 | uint32_t const cPages = cPages;
|
---|
814 | PPGMR0DYNMAPENTRY paPages = pThis->paPages;
|
---|
815 | uint32_t iFreePage;
|
---|
816 | if (!paPages[iPage].cRefs)
|
---|
817 | iFreePage = iPage;
|
---|
818 | else if (!paPages[(iPage + 1) % cPages].cRefs)
|
---|
819 | iFreePage = iPage;
|
---|
820 | else if (!paPages[(iPage + 2) % cPages].cRefs)
|
---|
821 | iFreePage = iPage;
|
---|
822 | else if (!paPages[(iPage + 3) % cPages].cRefs)
|
---|
823 | iFreePage = iPage;
|
---|
824 | else if (!paPages[(iPage + 4) % cPages].cRefs)
|
---|
825 | iFreePage = iPage;
|
---|
826 | else
|
---|
827 | {
|
---|
828 | /*
|
---|
829 | * Search for an unused or matching entry.
|
---|
830 | */
|
---|
831 | iFreePage = (iPage + 5) % pThis->cPages;
|
---|
832 | for (;;)
|
---|
833 | {
|
---|
834 | if (paPages[iFreePage].HCPhys == HCPhys)
|
---|
835 | return iFreePage;
|
---|
836 | if (!paPages[iFreePage].cRefs)
|
---|
837 | break;
|
---|
838 |
|
---|
839 | /* advance */
|
---|
840 | iFreePage = (iFreePage + 1) % cPages;
|
---|
841 | if (RT_UNLIKELY(iFreePage != iPage))
|
---|
842 | return UINT32_MAX;
|
---|
843 | }
|
---|
844 | }
|
---|
845 |
|
---|
846 | /*
|
---|
847 | * Setup the new entry.
|
---|
848 | */
|
---|
849 | paPages[iFreePage].HCPhys = HCPhys;
|
---|
850 | RTCpuSetFill(&paPages[iFreePage].PendingSet);
|
---|
851 | if (pThis->fLegacyMode)
|
---|
852 | {
|
---|
853 | X86PGUINT uOld = paPages[iFreePage].uPte.pLegacy->u;
|
---|
854 | X86PGUINT uOld2 = uOld; NOREF(uOld2);
|
---|
855 | X86PGUINT uNew = (uOld & X86_PTE_G | X86_PTE_PAT | X86_PTE_PCD | X86_PTE_PWT)
|
---|
856 | | X86_PTE_P | X86_PTE_A | X86_PTE_D
|
---|
857 | | (HCPhys & X86_PTE_PG_MASK);
|
---|
858 | while (!ASMAtomicCmpXchgExU32(&paPages[iFreePage].uPte.pLegacy->u, uNew, uOld, &uOld))
|
---|
859 | AssertMsgFailed(("uOld=%#x uOld2=%#x uNew=%#x\n", uOld, uOld2, uNew));
|
---|
860 | }
|
---|
861 | else
|
---|
862 | {
|
---|
863 | X86PGPAEUINT uOld = paPages[iFreePage].uPte.pPae->u;
|
---|
864 | X86PGPAEUINT uOld2 = uOld; NOREF(uOld2);
|
---|
865 | X86PGPAEUINT uNew = (uOld & X86_PTE_G | X86_PTE_PAT | X86_PTE_PCD | X86_PTE_PWT)
|
---|
866 | | X86_PTE_P | X86_PTE_A | X86_PTE_D
|
---|
867 | | (HCPhys & X86_PTE_PAE_PG_MASK);
|
---|
868 | while (!ASMAtomicCmpXchgExU64(&paPages[iFreePage].uPte.pPae->u, uNew, uOld, &uOld))
|
---|
869 | AssertMsgFailed(("uOld=%#llx uOld2=%#llx uNew=%#llx\n", uOld, uOld2, uNew));
|
---|
870 | }
|
---|
871 | return iFreePage;
|
---|
872 | }
|
---|
873 |
|
---|
874 |
|
---|
875 | /**
|
---|
876 | * Maps a page into the pool.
|
---|
877 | *
|
---|
878 | * @returns Pointer to the mapping.
|
---|
879 | * @param pThis The dynamic mapping cache instance.
|
---|
880 | * @param HCPhys The address of the page to be mapped.
|
---|
881 | * @param piPage Where to store the page index.
|
---|
882 | */
|
---|
883 | DECLINLINE(void *) pgmR0DynMapPage(PPGMR0DYNMAP pThis, RTHCPHYS HCPhys, uint32_t *piPage)
|
---|
884 | {
|
---|
885 | RTSPINLOCKTMP Tmp = RTSPINLOCKTMP_INITIALIZER;
|
---|
886 | RTSpinlockAcquire(pThis->hSpinlock, &Tmp);
|
---|
887 | AssertMsg(!(HCPhys & PAGE_OFFSET_MASK), ("HCPhys=%RHp\n", HCPhys));
|
---|
888 |
|
---|
889 | /*
|
---|
890 | * Find an entry, if possible a matching one. The HCPhys address is hashed
|
---|
891 | * down to a page index, collisions are handled by linear searching. Optimize
|
---|
892 | * for a hit in the first 5 pages.
|
---|
893 | *
|
---|
894 | * To the cheap hits here and defer the tedious searching and inserting
|
---|
895 | * to a helper function.
|
---|
896 | */
|
---|
897 | uint32_t const cPages = cPages;
|
---|
898 | uint32_t iPage = (HCPhys >> PAGE_SHIFT) % cPages;
|
---|
899 | PPGMR0DYNMAPENTRY paPages = pThis->paPages;
|
---|
900 | if (paPages[iPage].HCPhys != HCPhys)
|
---|
901 | {
|
---|
902 | uint32_t iPage2 = (iPage + 1) % cPages;
|
---|
903 | if (paPages[iPage2].HCPhys != HCPhys)
|
---|
904 | {
|
---|
905 | iPage2 = (iPage + 2) % cPages;
|
---|
906 | if (paPages[iPage2].HCPhys != HCPhys)
|
---|
907 | {
|
---|
908 | iPage2 = (iPage + 3) % cPages;
|
---|
909 | if (paPages[iPage2].HCPhys != HCPhys)
|
---|
910 | {
|
---|
911 | iPage2 = (iPage + 4) % cPages;
|
---|
912 | if (paPages[iPage2].HCPhys != HCPhys)
|
---|
913 | {
|
---|
914 | iPage = pgmR0DynMapPageSlow(pThis, HCPhys, iPage);
|
---|
915 | if (RT_UNLIKELY(iPage == UINT32_MAX))
|
---|
916 | {
|
---|
917 | RTSpinlockRelease(pThis->hSpinlock, &Tmp);
|
---|
918 | return NULL;
|
---|
919 | }
|
---|
920 | }
|
---|
921 | else
|
---|
922 | iPage = iPage2;
|
---|
923 | }
|
---|
924 | else
|
---|
925 | iPage = iPage2;
|
---|
926 | }
|
---|
927 | else
|
---|
928 | iPage = iPage2;
|
---|
929 | }
|
---|
930 | else
|
---|
931 | iPage = iPage2;
|
---|
932 | }
|
---|
933 |
|
---|
934 | /*
|
---|
935 | * Reference it, update statistics and get the return address.
|
---|
936 | */
|
---|
937 | if (ASMAtomicIncS32(&paPages[iPage].cRefs) == 1)
|
---|
938 | {
|
---|
939 | pThis->cLoad++;
|
---|
940 | if (pThis->cLoad > pThis->cMaxLoad)
|
---|
941 | pThis->cMaxLoad = pThis->cLoad;
|
---|
942 | Assert(pThis->cLoad <= pThis->cPages);
|
---|
943 | }
|
---|
944 | void *pvPage = paPages[iPage].pvPage;
|
---|
945 |
|
---|
946 | /*
|
---|
947 | * Invalidate the entry?
|
---|
948 | */
|
---|
949 | RTCPUID idRealCpu = RTMpCpuId();
|
---|
950 | bool fInvalidateIt = RTCpuSetIsMember(&paPages[iPage].PendingSet, idRealCpu);
|
---|
951 | if (fInvalidateIt)
|
---|
952 | RTCpuSetDel(&paPages[iPage].PendingSet, idRealCpu);
|
---|
953 |
|
---|
954 | RTSpinlockRelease(pThis->hSpinlock, &Tmp);
|
---|
955 |
|
---|
956 | /*
|
---|
957 | * Do the actual invalidation outside the spinlock.
|
---|
958 | */
|
---|
959 | ASMInvalidatePage(pvPage);
|
---|
960 |
|
---|
961 | *piPage = iPage;
|
---|
962 | return pvPage;
|
---|
963 | }
|
---|
964 |
|
---|
965 |
|
---|
966 | /**
|
---|
967 | * Signals the start of a new set of mappings.
|
---|
968 | *
|
---|
969 | * Mostly for strictness. PGMDynMapHCPage won't work unless this
|
---|
970 | * API is called.
|
---|
971 | *
|
---|
972 | * @param pVCpu The shared data for the current virtual CPU.
|
---|
973 | */
|
---|
974 | VMMDECL(void) PGMDynMapStartAutoSet(PVMCPU pVCpu)
|
---|
975 | {
|
---|
976 | Assert(pVCpu->pgm.s.AutoSet.cEntries == PGMMAPSET_CLOSED);
|
---|
977 | pVCpu->pgm.s.AutoSet.cEntries = 0;
|
---|
978 | }
|
---|
979 |
|
---|
980 |
|
---|
981 | /**
|
---|
982 | * Releases the dynamic memory mappings made by PGMDynMapHCPage and associates
|
---|
983 | * since the PGMDynMapStartAutoSet call.
|
---|
984 | *
|
---|
985 | * @param pVCpu The shared data for the current virtual CPU.
|
---|
986 | */
|
---|
987 | VMMDECL(void) PGMDynMapReleaseAutoSet(PVMCPU pVCpu)
|
---|
988 | {
|
---|
989 | PPGMMAPSET pSet = &pVCpu->pgm.s.AutoSet;
|
---|
990 |
|
---|
991 | /* close the set */
|
---|
992 | uint32_t i = pVCpu->pgm.s.AutoSet.cEntries;
|
---|
993 | AssertMsg(i <= RT_ELEMENTS(pVCpu->pgm.s.AutoSet.aEntries), ("%#x (%u)\n", i, i));
|
---|
994 | pVCpu->pgm.s.AutoSet.cEntries = PGMMAPSET_CLOSED;
|
---|
995 |
|
---|
996 | /* release any pages we're referencing. */
|
---|
997 | if (i != 0 && RT_LIKELY(i <= RT_ELEMENTS(pVCpu->pgm.s.AutoSet.aEntries)))
|
---|
998 | {
|
---|
999 | PPGMR0DYNMAP pThis = g_pPGMR0DynMap;
|
---|
1000 | RTSPINLOCKTMP Tmp = RTSPINLOCKTMP_INITIALIZER;
|
---|
1001 | RTSpinlockAcquire(pThis->hSpinlock, &Tmp);
|
---|
1002 |
|
---|
1003 | while (i-- > 0)
|
---|
1004 | {
|
---|
1005 | uint32_t iPage = pSet->aEntries[i].iPage;
|
---|
1006 | Assert(iPage < pThis->cPages);
|
---|
1007 | int32_t cRefs = pSet->aEntries[i].cRefs;
|
---|
1008 | Assert(cRefs > 0);
|
---|
1009 | pgmR0DynMapReleasePageLocked(pThis, iPage, cRefs);
|
---|
1010 |
|
---|
1011 | pSet->aEntries[i].iPage = UINT16_MAX;
|
---|
1012 | pSet->aEntries[i].cRefs = 0;
|
---|
1013 | }
|
---|
1014 |
|
---|
1015 | Assert(pThis->cLoad <= pThis->cPages);
|
---|
1016 | RTSpinlockRelease(pThis->hSpinlock, &Tmp);
|
---|
1017 | }
|
---|
1018 | }
|
---|
1019 |
|
---|
1020 |
|
---|
1021 | /**
|
---|
1022 | * Migrates the automatic mapping set of the current vCPU if necessary.
|
---|
1023 | *
|
---|
1024 | * This is called when re-entering the hardware assisted execution mode after a
|
---|
1025 | * nip down to ring-3. We run the risk that the CPU might have change and we
|
---|
1026 | * will therefore make sure all the cache entries currently in the auto set will
|
---|
1027 | * be valid on the new CPU. If the cpu didn't change nothing will happen as all
|
---|
1028 | * the entries will have been flagged as invalidated.
|
---|
1029 | *
|
---|
1030 | * @param pVCpu The shared data for the current virtual CPU.
|
---|
1031 | * @thread EMT
|
---|
1032 | */
|
---|
1033 | VMMDECL(void) PGMDynMapMigrateAutoSet(PVMCPU pVCpu)
|
---|
1034 | {
|
---|
1035 | PPGMMAPSET pSet = &pVCpu->pgm.s.AutoSet;
|
---|
1036 | uint32_t i = pVCpu->pgm.s.AutoSet.cEntries;
|
---|
1037 | AssertMsg(i <= RT_ELEMENTS(pVCpu->pgm.s.AutoSet.aEntries), ("%#x (%u)\n", i, i));
|
---|
1038 | if (i != 0 && RT_LIKELY(i <= RT_ELEMENTS(pVCpu->pgm.s.AutoSet.aEntries)))
|
---|
1039 | {
|
---|
1040 | PPGMR0DYNMAP pThis = g_pPGMR0DynMap;
|
---|
1041 | RTCPUID idRealCpu = RTMpCpuId();
|
---|
1042 |
|
---|
1043 | while (i-- > 0)
|
---|
1044 | {
|
---|
1045 | Assert(pSet->aEntries[i].cRefs > 0);
|
---|
1046 | uint32_t iPage = pSet->aEntries[i].iPage;
|
---|
1047 | Assert(iPage < pThis->cPages);
|
---|
1048 | if (RTCpuSetIsMember(&pThis->paPages[iPage].PendingSet, idRealCpu))
|
---|
1049 | {
|
---|
1050 | RTCpuSetDel(&pThis->paPages[iPage].PendingSet, idRealCpu);
|
---|
1051 | ASMInvalidatePage(pThis->paPages[iPage].pvPage);
|
---|
1052 | }
|
---|
1053 | }
|
---|
1054 | }
|
---|
1055 | }
|
---|
1056 |
|
---|
1057 |
|
---|
1058 | /**
|
---|
1059 | * As a final resort for a full auto set, try merge duplicate entries.
|
---|
1060 | *
|
---|
1061 | * @param pSet The set.
|
---|
1062 | */
|
---|
1063 | static void pgmDynMapOptimizeAutoSet(PPGMMAPSET pSet)
|
---|
1064 | {
|
---|
1065 | for (uint32_t i = 0 ; i < pSet->cEntries; i++)
|
---|
1066 | {
|
---|
1067 | uint16_t const iPage = pSet->aEntries[i].iPage;
|
---|
1068 | uint32_t j = i + 1;
|
---|
1069 | while (j < pSet->cEntries)
|
---|
1070 | {
|
---|
1071 | if (pSet->aEntries[j].iPage != iPage)
|
---|
1072 | j++;
|
---|
1073 | else
|
---|
1074 | {
|
---|
1075 | /* merge j with i removing j. */
|
---|
1076 | pSet->aEntries[i].cRefs += pSet->aEntries[j].cRefs;
|
---|
1077 | pSet->cEntries--;
|
---|
1078 | if (j < pSet->cEntries)
|
---|
1079 | {
|
---|
1080 | pSet->aEntries[j] = pSet->aEntries[pSet->cEntries];
|
---|
1081 | pSet->aEntries[pSet->cEntries].iPage = UINT16_MAX;
|
---|
1082 | pSet->aEntries[pSet->cEntries].cRefs = 0;
|
---|
1083 | }
|
---|
1084 | else
|
---|
1085 | {
|
---|
1086 | pSet->aEntries[j].iPage = UINT16_MAX;
|
---|
1087 | pSet->aEntries[j].cRefs = 0;
|
---|
1088 | }
|
---|
1089 | }
|
---|
1090 | }
|
---|
1091 | }
|
---|
1092 | }
|
---|
1093 |
|
---|
1094 |
|
---|
1095 | /* documented elsewhere - a bit of a mess. */
|
---|
1096 | VMMDECL(int) PGMDynMapHCPage(PVM pVM, RTHCPHYS HCPhys, void **ppv)
|
---|
1097 | {
|
---|
1098 | /*
|
---|
1099 | * Validate state.
|
---|
1100 | */
|
---|
1101 | AssertMsgReturn(pVM->pgm.s.pvR0DynMapUsed == g_pPGMR0DynMap,
|
---|
1102 | ("%p != %p\n", pVM->pgm.s.pvR0DynMapUsed, g_pPGMR0DynMap),
|
---|
1103 | VERR_ACCESS_DENIED);
|
---|
1104 | AssertMsg(!(HCPhys & PAGE_OFFSET_MASK), ("HCPhys=%RHp\n", HCPhys));
|
---|
1105 | PVMCPU pVCpu = VMMGetCpu(pVM);
|
---|
1106 | PPGMMAPSET pSet = &pVCpu->pgm.s.AutoSet;
|
---|
1107 | AssertPtrReturn(pVCpu, VERR_INTERNAL_ERROR);
|
---|
1108 | AssertMsgReturn(pSet->cEntries > RT_ELEMENTS(pSet->aEntries),
|
---|
1109 | ("%#x (%u)\n", pSet->cEntries, pSet->cEntries), VERR_WRONG_ORDER);
|
---|
1110 |
|
---|
1111 | /*
|
---|
1112 | * Map it.
|
---|
1113 | */
|
---|
1114 | uint32_t iPage;
|
---|
1115 | void *pvPage = pgmR0DynMapPage(g_pPGMR0DynMap, HCPhys, &iPage);
|
---|
1116 | if (RT_UNLIKELY(!pvPage))
|
---|
1117 | {
|
---|
1118 | static uint32_t s_cBitched = 0;
|
---|
1119 | if (++s_cBitched < 10)
|
---|
1120 | LogRel(("PGMDynMapHCPage: cLoad=%u/%u cPages=%u\n",
|
---|
1121 | g_pPGMR0DynMap->cLoad, g_pPGMR0DynMap->cMaxLoad, g_pPGMR0DynMap->cPages));
|
---|
1122 | return VERR_PGM_DYNMAP_FAILED;
|
---|
1123 | }
|
---|
1124 |
|
---|
1125 | /*
|
---|
1126 | * Add the page to the auto reference set.
|
---|
1127 | * If it's less than half full, don't bother looking for duplicates.
|
---|
1128 | */
|
---|
1129 | if (pSet->cEntries < RT_ELEMENTS(pSet->aEntries) / 2)
|
---|
1130 | {
|
---|
1131 | pSet->aEntries[pSet->cEntries].cRefs = 1;
|
---|
1132 | pSet->aEntries[pSet->cEntries].iPage = iPage;
|
---|
1133 | }
|
---|
1134 | else
|
---|
1135 | {
|
---|
1136 | Assert(pSet->cEntries <= RT_ELEMENTS(pSet->aEntries));
|
---|
1137 | int32_t i = pSet->cEntries;
|
---|
1138 | while (i-- > 0)
|
---|
1139 | if (pSet->aEntries[i].iPage)
|
---|
1140 | {
|
---|
1141 | pSet->aEntries[i].cRefs++;
|
---|
1142 | break;
|
---|
1143 | }
|
---|
1144 | if (i < 0)
|
---|
1145 | {
|
---|
1146 | if (RT_UNLIKELY(pSet->cEntries >= RT_ELEMENTS(pSet->aEntries)))
|
---|
1147 | pgmDynMapOptimizeAutoSet(pSet);
|
---|
1148 | if (RT_LIKELY(pSet->cEntries < RT_ELEMENTS(pSet->aEntries)))
|
---|
1149 | {
|
---|
1150 | pSet->aEntries[pSet->cEntries].cRefs = 1;
|
---|
1151 | pSet->aEntries[pSet->cEntries].iPage = iPage;
|
---|
1152 | }
|
---|
1153 | else
|
---|
1154 | {
|
---|
1155 | /* We're screwed. */
|
---|
1156 | pgmR0DynMapReleasePage(g_pPGMR0DynMap, iPage, 1);
|
---|
1157 |
|
---|
1158 | static uint32_t s_cBitched = 0;
|
---|
1159 | if (++s_cBitched < 10)
|
---|
1160 | LogRel(("PGMDynMapHCPage: set is full!\n"));
|
---|
1161 | return VERR_PGM_DYNMAP_FULL_SET;
|
---|
1162 | }
|
---|
1163 | }
|
---|
1164 | }
|
---|
1165 |
|
---|
1166 | return VINF_SUCCESS;
|
---|
1167 | }
|
---|
1168 |
|
---|