VirtualBox

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 59527

Last change on this file since 59527 was 58171, checked in by vboxsync, 9 years ago

doxygen: fixes.

  • Property svn:eol-style set to native
  • Property svn:keywords set to Id Rev Revision
File size: 54.9 KB
Line 
1/* $Id: memobj-r0drv-linux.c 58171 2015-10-12 09:30:58Z vboxsync $ */
2/** @file
3 * IPRT - Ring-0 Memory Objects, Linux.
4 */
5
6/*
7 * Copyright (C) 2006-2015 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 */
100static struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
101{
102 /** @todo fix rtR0ProcessToLinuxTask!! */
103 /** @todo many (all?) callers currently assume that we return 'current'! */
104 return R0Process == RTR0ProcHandleSelf() ? current : NULL;
105}
106
107
108/**
109 * Compute order. Some functions allocate 2^order pages.
110 *
111 * @returns order.
112 * @param cPages Number of pages.
113 */
114static int rtR0MemObjLinuxOrder(size_t cPages)
115{
116 int iOrder;
117 size_t cTmp;
118
119 for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
120 ;
121 if (cPages & ~((size_t)1 << iOrder))
122 ++iOrder;
123
124 return iOrder;
125}
126
127
128/**
129 * Converts from RTMEM_PROT_* to Linux PAGE_*.
130 *
131 * @returns Linux page protection constant.
132 * @param fProt The IPRT protection mask.
133 * @param fKernel Whether it applies to kernel or user space.
134 */
135static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
136{
137 switch (fProt)
138 {
139 default:
140 AssertMsgFailed(("%#x %d\n", fProt, fKernel));
141 case RTMEM_PROT_NONE:
142 return PAGE_NONE;
143
144 case RTMEM_PROT_READ:
145 return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
146
147 case RTMEM_PROT_WRITE:
148 case RTMEM_PROT_WRITE | RTMEM_PROT_READ:
149 return fKernel ? PAGE_KERNEL : PAGE_SHARED;
150
151 case RTMEM_PROT_EXEC:
152 case RTMEM_PROT_EXEC | RTMEM_PROT_READ:
153#if defined(RT_ARCH_X86) || defined(RT_ARCH_AMD64)
154 if (fKernel)
155 {
156 pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
157 pgprot_val(fPg) &= ~_PAGE_RW;
158 return fPg;
159 }
160 return PAGE_READONLY_EXEC;
161#else
162 return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
163#endif
164
165 case RTMEM_PROT_WRITE | RTMEM_PROT_EXEC:
166 case RTMEM_PROT_WRITE | RTMEM_PROT_EXEC | RTMEM_PROT_READ:
167 return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
168 }
169}
170
171
172/**
173 * Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
174 * an empty user space mapping.
175 *
176 * We acquire the mmap_sem of the task!
177 *
178 * @returns Pointer to the mapping.
179 * (void *)-1 on failure.
180 * @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
181 * @param cb The size of the mapping.
182 * @param uAlignment The alignment of the mapping.
183 * @param pTask The Linux task to create this mapping in.
184 * @param fProt The RTMEM_PROT_* mask.
185 */
186static void *rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct *pTask, unsigned fProt)
187{
188 unsigned fLnxProt;
189 unsigned long ulAddr;
190
191 Assert((pTask == current)); /* do_mmap */
192
193 /*
194 * Convert from IPRT protection to mman.h PROT_ and call do_mmap.
195 */
196 fProt &= (RTMEM_PROT_NONE | RTMEM_PROT_READ | RTMEM_PROT_WRITE | RTMEM_PROT_EXEC);
197 if (fProt == RTMEM_PROT_NONE)
198 fLnxProt = PROT_NONE;
199 else
200 {
201 fLnxProt = 0;
202 if (fProt & RTMEM_PROT_READ)
203 fLnxProt |= PROT_READ;
204 if (fProt & RTMEM_PROT_WRITE)
205 fLnxProt |= PROT_WRITE;
206 if (fProt & RTMEM_PROT_EXEC)
207 fLnxProt |= PROT_EXEC;
208 }
209
210 if (R3PtrFixed != (RTR3PTR)-1)
211 {
212#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 5, 0)
213 ulAddr = vm_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED | MAP_ANONYMOUS | MAP_FIXED, 0);
214#else
215 down_write(&pTask->mm->mmap_sem);
216 ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED | MAP_ANONYMOUS | MAP_FIXED, 0);
217 up_write(&pTask->mm->mmap_sem);
218#endif
219 }
220 else
221 {
222#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 5, 0)
223 ulAddr = vm_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED | MAP_ANONYMOUS, 0);
224#else
225 down_write(&pTask->mm->mmap_sem);
226 ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED | MAP_ANONYMOUS, 0);
227 up_write(&pTask->mm->mmap_sem);
228#endif
229 if ( !(ulAddr & ~PAGE_MASK)
230 && (ulAddr & (uAlignment - 1)))
231 {
232 /** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
233 * up alignment gaps. This is of course complicated by fragmentation (which we might have cause
234 * ourselves) and further by there begin two mmap strategies (top / bottom). */
235 /* For now, just ignore uAlignment requirements... */
236 }
237 }
238
239
240 if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
241 return (void *)-1;
242 return (void *)ulAddr;
243}
244
245
246/**
247 * Worker that destroys a user space mapping.
248 * Undoes what rtR0MemObjLinuxDoMmap did.
249 *
250 * We acquire the mmap_sem of the task!
251 *
252 * @param pv The ring-3 mapping.
253 * @param cb The size of the mapping.
254 * @param pTask The Linux task to destroy this mapping in.
255 */
256static void rtR0MemObjLinuxDoMunmap(void *pv, size_t cb, struct task_struct *pTask)
257{
258#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 5, 0)
259 Assert(pTask == current);
260 vm_munmap((unsigned long)pv, cb);
261#elif defined(USE_RHEL4_MUNMAP)
262 down_write(&pTask->mm->mmap_sem);
263 do_munmap(pTask->mm, (unsigned long)pv, cb, 0); /* should it be 1 or 0? */
264 up_write(&pTask->mm->mmap_sem);
265#else
266 down_write(&pTask->mm->mmap_sem);
267 do_munmap(pTask->mm, (unsigned long)pv, cb);
268 up_write(&pTask->mm->mmap_sem);
269#endif
270}
271
272
273/**
274 * Internal worker that allocates physical pages and creates the memory object for them.
275 *
276 * @returns IPRT status code.
277 * @param ppMemLnx Where to store the memory object pointer.
278 * @param enmType The object type.
279 * @param cb The number of bytes to allocate.
280 * @param uAlignment The alignment of the physical memory.
281 * Only valid if fContiguous == true, ignored otherwise.
282 * @param fFlagsLnx The page allocation flags (GPFs).
283 * @param fContiguous Whether the allocation must be contiguous.
284 * @param rcNoMem What to return when we're out of pages.
285 */
286static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
287 size_t uAlignment, unsigned fFlagsLnx, bool fContiguous, int rcNoMem)
288{
289 size_t iPage;
290 size_t const cPages = cb >> PAGE_SHIFT;
291 struct page *paPages;
292
293 /*
294 * Allocate a memory object structure that's large enough to contain
295 * the page pointer array.
296 */
297 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
298 if (!pMemLnx)
299 return VERR_NO_MEMORY;
300 pMemLnx->cPages = cPages;
301
302 if (cPages > 255)
303 {
304# ifdef __GFP_REPEAT
305 /* Try hard to allocate the memory, but the allocation attempt might fail. */
306 fFlagsLnx |= __GFP_REPEAT;
307# endif
308# ifdef __GFP_NOMEMALLOC
309 /* Introduced with Linux 2.6.12: Don't use emergency reserves */
310 fFlagsLnx |= __GFP_NOMEMALLOC;
311# endif
312 }
313
314 /*
315 * Allocate the pages.
316 * For small allocations we'll try contiguous first and then fall back on page by page.
317 */
318#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
319 if ( fContiguous
320 || cb <= PAGE_SIZE * 2)
321 {
322# ifdef VBOX_USE_INSERT_PAGE
323 paPages = alloc_pages(fFlagsLnx | __GFP_COMP | __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
324# else
325 paPages = alloc_pages(fFlagsLnx | __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
326# endif
327 if (paPages)
328 {
329 fContiguous = true;
330 for (iPage = 0; iPage < cPages; iPage++)
331 pMemLnx->apPages[iPage] = &paPages[iPage];
332 }
333 else if (fContiguous)
334 {
335 rtR0MemObjDelete(&pMemLnx->Core);
336 return rcNoMem;
337 }
338 }
339
340 if (!fContiguous)
341 {
342 for (iPage = 0; iPage < cPages; iPage++)
343 {
344 pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx | __GFP_NOWARN);
345 if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
346 {
347 while (iPage-- > 0)
348 __free_page(pMemLnx->apPages[iPage]);
349 rtR0MemObjDelete(&pMemLnx->Core);
350 return rcNoMem;
351 }
352 }
353 }
354
355#else /* < 2.4.22 */
356 /** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
357 paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
358 if (!paPages)
359 {
360 rtR0MemObjDelete(&pMemLnx->Core);
361 return rcNoMem;
362 }
363 for (iPage = 0; iPage < cPages; iPage++)
364 {
365 pMemLnx->apPages[iPage] = &paPages[iPage];
366 MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
367 if (PageHighMem(pMemLnx->apPages[iPage]))
368 BUG();
369 }
370
371 fContiguous = true;
372#endif /* < 2.4.22 */
373 pMemLnx->fContiguous = fContiguous;
374
375 /*
376 * Reserve the pages.
377 */
378 for (iPage = 0; iPage < cPages; iPage++)
379 SetPageReserved(pMemLnx->apPages[iPage]);
380
381 /*
382 * Note that the physical address of memory allocated with alloc_pages(flags, order)
383 * is always 2^(PAGE_SHIFT+order)-aligned.
384 */
385 if ( fContiguous
386 && uAlignment > PAGE_SIZE)
387 {
388 /*
389 * Check for alignment constraints. The physical address of memory allocated with
390 * alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
391 */
392 if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
393 {
394 /*
395 * This should never happen!
396 */
397 printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
398 (unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
399 rtR0MemObjLinuxFreePages(pMemLnx);
400 return rcNoMem;
401 }
402 }
403
404 *ppMemLnx = pMemLnx;
405 return VINF_SUCCESS;
406}
407
408
409/**
410 * Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
411 *
412 * This method does NOT free the object.
413 *
414 * @param pMemLnx The object which physical pages should be freed.
415 */
416static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
417{
418 size_t iPage = pMemLnx->cPages;
419 if (iPage > 0)
420 {
421 /*
422 * Restore the page flags.
423 */
424 while (iPage-- > 0)
425 {
426 ClearPageReserved(pMemLnx->apPages[iPage]);
427#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
428#else
429 MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
430#endif
431 }
432
433 /*
434 * Free the pages.
435 */
436#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
437 if (!pMemLnx->fContiguous)
438 {
439 iPage = pMemLnx->cPages;
440 while (iPage-- > 0)
441 __free_page(pMemLnx->apPages[iPage]);
442 }
443 else
444#endif
445 __free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
446
447 pMemLnx->cPages = 0;
448 }
449}
450
451
452/**
453 * Maps the allocation into ring-0.
454 *
455 * This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
456 *
457 * Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
458 * space, so we'll use that mapping if possible. If execute access is required, we'll
459 * play safe and do our own mapping.
460 *
461 * @returns IPRT status code.
462 * @param pMemLnx The linux memory object to map.
463 * @param fExecutable Whether execute access is required.
464 */
465static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
466{
467 int rc = VINF_SUCCESS;
468
469 /*
470 * Choose mapping strategy.
471 */
472 bool fMustMap = fExecutable
473 || !pMemLnx->fContiguous;
474 if (!fMustMap)
475 {
476 size_t iPage = pMemLnx->cPages;
477 while (iPage-- > 0)
478 if (PageHighMem(pMemLnx->apPages[iPage]))
479 {
480 fMustMap = true;
481 break;
482 }
483 }
484
485 Assert(!pMemLnx->Core.pv);
486 Assert(!pMemLnx->fMappedToRing0);
487
488 if (fMustMap)
489 {
490 /*
491 * Use vmap - 2.4.22 and later.
492 */
493#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
494 pgprot_t fPg;
495 pgprot_val(fPg) = _PAGE_PRESENT | _PAGE_RW;
496# ifdef _PAGE_NX
497 if (!fExecutable)
498 pgprot_val(fPg) |= _PAGE_NX;
499# endif
500
501# ifdef VM_MAP
502 pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
503# else
504 pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
505# endif
506 if (pMemLnx->Core.pv)
507 pMemLnx->fMappedToRing0 = true;
508 else
509 rc = VERR_MAP_FAILED;
510#else /* < 2.4.22 */
511 rc = VERR_NOT_SUPPORTED;
512#endif
513 }
514 else
515 {
516 /*
517 * Use the kernel RAM mapping.
518 */
519 pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
520 Assert(pMemLnx->Core.pv);
521 }
522
523 return rc;
524}
525
526
527/**
528 * Undoes what rtR0MemObjLinuxVMap() did.
529 *
530 * @param pMemLnx The linux memory object.
531 */
532static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
533{
534#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
535 if (pMemLnx->fMappedToRing0)
536 {
537 Assert(pMemLnx->Core.pv);
538 vunmap(pMemLnx->Core.pv);
539 pMemLnx->fMappedToRing0 = false;
540 }
541#else /* < 2.4.22 */
542 Assert(!pMemLnx->fMappedToRing0);
543#endif
544 pMemLnx->Core.pv = NULL;
545}
546
547
548DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
549{
550 IPRT_LINUX_SAVE_EFL_AC();
551 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
552
553 /*
554 * Release any memory that we've allocated or locked.
555 */
556 switch (pMemLnx->Core.enmType)
557 {
558 case RTR0MEMOBJTYPE_LOW:
559 case RTR0MEMOBJTYPE_PAGE:
560 case RTR0MEMOBJTYPE_CONT:
561 case RTR0MEMOBJTYPE_PHYS:
562 case RTR0MEMOBJTYPE_PHYS_NC:
563 rtR0MemObjLinuxVUnmap(pMemLnx);
564 rtR0MemObjLinuxFreePages(pMemLnx);
565 break;
566
567 case RTR0MEMOBJTYPE_LOCK:
568 if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
569 {
570 struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
571 size_t iPage;
572 Assert(pTask);
573 if (pTask && pTask->mm)
574 down_read(&pTask->mm->mmap_sem);
575
576 iPage = pMemLnx->cPages;
577 while (iPage-- > 0)
578 {
579 if (!PageReserved(pMemLnx->apPages[iPage]))
580 SetPageDirty(pMemLnx->apPages[iPage]);
581 page_cache_release(pMemLnx->apPages[iPage]);
582 }
583
584 if (pTask && pTask->mm)
585 up_read(&pTask->mm->mmap_sem);
586 }
587 /* else: kernel memory - nothing to do here. */
588 break;
589
590 case RTR0MEMOBJTYPE_RES_VIRT:
591 Assert(pMemLnx->Core.pv);
592 if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
593 {
594 struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
595 Assert(pTask);
596 if (pTask && pTask->mm)
597 rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
598 }
599 else
600 {
601 vunmap(pMemLnx->Core.pv);
602
603 Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
604 __free_page(pMemLnx->apPages[0]);
605 pMemLnx->apPages[0] = NULL;
606 pMemLnx->cPages = 0;
607 }
608 pMemLnx->Core.pv = NULL;
609 break;
610
611 case RTR0MEMOBJTYPE_MAPPING:
612 Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
613 if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
614 {
615 struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
616 Assert(pTask);
617 if (pTask && pTask->mm)
618 rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
619 }
620 else
621 vunmap(pMemLnx->Core.pv);
622 pMemLnx->Core.pv = NULL;
623 break;
624
625 default:
626 AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
627 return VERR_INTERNAL_ERROR;
628 }
629 IPRT_LINUX_RESTORE_EFL_ONLY_AC();
630 return VINF_SUCCESS;
631}
632
633
634DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
635{
636 IPRT_LINUX_SAVE_EFL_AC();
637 PRTR0MEMOBJLNX pMemLnx;
638 int rc;
639
640#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
641 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
642 false /* non-contiguous */, VERR_NO_MEMORY);
643#else
644 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
645 false /* non-contiguous */, VERR_NO_MEMORY);
646#endif
647 if (RT_SUCCESS(rc))
648 {
649 rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
650 if (RT_SUCCESS(rc))
651 {
652 *ppMem = &pMemLnx->Core;
653 IPRT_LINUX_RESTORE_EFL_AC();
654 return rc;
655 }
656
657 rtR0MemObjLinuxFreePages(pMemLnx);
658 rtR0MemObjDelete(&pMemLnx->Core);
659 }
660
661 IPRT_LINUX_RESTORE_EFL_AC();
662 return rc;
663}
664
665
666DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
667{
668 IPRT_LINUX_SAVE_EFL_AC();
669 PRTR0MEMOBJLNX pMemLnx;
670 int rc;
671
672 /* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
673#if (defined(RT_ARCH_AMD64) || defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
674 /* ZONE_DMA32: 0-4GB */
675 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
676 false /* non-contiguous */, VERR_NO_LOW_MEMORY);
677 if (RT_FAILURE(rc))
678#endif
679#ifdef RT_ARCH_AMD64
680 /* ZONE_DMA: 0-16MB */
681 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
682 false /* non-contiguous */, VERR_NO_LOW_MEMORY);
683#else
684# ifdef CONFIG_X86_PAE
685# endif
686 /* ZONE_NORMAL: 0-896MB */
687 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
688 false /* non-contiguous */, VERR_NO_LOW_MEMORY);
689#endif
690 if (RT_SUCCESS(rc))
691 {
692 rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
693 if (RT_SUCCESS(rc))
694 {
695 *ppMem = &pMemLnx->Core;
696 IPRT_LINUX_RESTORE_EFL_AC();
697 return rc;
698 }
699
700 rtR0MemObjLinuxFreePages(pMemLnx);
701 rtR0MemObjDelete(&pMemLnx->Core);
702 }
703
704 IPRT_LINUX_RESTORE_EFL_AC();
705 return rc;
706}
707
708
709DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
710{
711 IPRT_LINUX_SAVE_EFL_AC();
712 PRTR0MEMOBJLNX pMemLnx;
713 int rc;
714
715#if (defined(RT_ARCH_AMD64) || defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
716 /* ZONE_DMA32: 0-4GB */
717 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
718 true /* contiguous */, VERR_NO_CONT_MEMORY);
719 if (RT_FAILURE(rc))
720#endif
721#ifdef RT_ARCH_AMD64
722 /* ZONE_DMA: 0-16MB */
723 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
724 true /* contiguous */, VERR_NO_CONT_MEMORY);
725#else
726 /* ZONE_NORMAL (32-bit hosts): 0-896MB */
727 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
728 true /* contiguous */, VERR_NO_CONT_MEMORY);
729#endif
730 if (RT_SUCCESS(rc))
731 {
732 rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
733 if (RT_SUCCESS(rc))
734 {
735#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) || defined(CONFIG_HIGHMEM64G))
736 size_t iPage = pMemLnx->cPages;
737 while (iPage-- > 0)
738 Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
739#endif
740 pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
741 *ppMem = &pMemLnx->Core;
742 IPRT_LINUX_RESTORE_EFL_AC();
743 return rc;
744 }
745
746 rtR0MemObjLinuxFreePages(pMemLnx);
747 rtR0MemObjDelete(&pMemLnx->Core);
748 }
749
750 IPRT_LINUX_RESTORE_EFL_AC();
751 return rc;
752}
753
754
755/**
756 * Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
757 *
758 * @returns IPRT status code.
759 * @param ppMemLnx Where to
760 * @param enmType The object type.
761 * @param cb The size of the allocation.
762 * @param uAlignment The alignment of the physical memory.
763 * Only valid for fContiguous == true, ignored otherwise.
764 * @param PhysHighest See rtR0MemObjNativeAllocPhys.
765 * @param fGfp The Linux GFP flags to use for the allocation.
766 */
767static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
768 size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, unsigned fGfp)
769{
770 PRTR0MEMOBJLNX pMemLnx;
771 int rc;
772
773 rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
774 enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
775 VERR_NO_PHYS_MEMORY);
776 if (RT_FAILURE(rc))
777 return rc;
778
779 /*
780 * Check the addresses if necessary. (Can be optimized a bit for PHYS.)
781 */
782 if (PhysHighest != NIL_RTHCPHYS)
783 {
784 size_t iPage = pMemLnx->cPages;
785 while (iPage-- > 0)
786 if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
787 {
788 rtR0MemObjLinuxFreePages(pMemLnx);
789 rtR0MemObjDelete(&pMemLnx->Core);
790 return VERR_NO_MEMORY;
791 }
792 }
793
794 /*
795 * Complete the object.
796 */
797 if (enmType == RTR0MEMOBJTYPE_PHYS)
798 {
799 pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
800 pMemLnx->Core.u.Phys.fAllocated = true;
801 }
802 *ppMem = &pMemLnx->Core;
803 return rc;
804}
805
806
807/**
808 * Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
809 *
810 * @returns IPRT status code.
811 * @param ppMem Where to store the memory object pointer on success.
812 * @param enmType The object type.
813 * @param cb The size of the allocation.
814 * @param uAlignment The alignment of the physical memory.
815 * Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
816 * @param PhysHighest See rtR0MemObjNativeAllocPhys.
817 */
818static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
819 size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
820{
821 int rc;
822 IPRT_LINUX_SAVE_EFL_AC();
823
824 /*
825 * There are two clear cases and that's the <=16MB and anything-goes ones.
826 * When the physical address limit is somewhere in-between those two we'll
827 * just have to try, starting with HIGHUSER and working our way thru the
828 * different types, hoping we'll get lucky.
829 *
830 * We should probably move this physical address restriction logic up to
831 * the page alloc function as it would be more efficient there. But since
832 * we don't expect this to be a performance issue just yet it can wait.
833 */
834 if (PhysHighest == NIL_RTHCPHYS)
835 /* ZONE_HIGHMEM: the whole physical memory */
836 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
837 else if (PhysHighest <= _1M * 16)
838 /* ZONE_DMA: 0-16MB */
839 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
840 else
841 {
842 rc = VERR_NO_MEMORY;
843 if (RT_FAILURE(rc))
844 /* ZONE_HIGHMEM: the whole physical memory */
845 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
846 if (RT_FAILURE(rc))
847 /* ZONE_NORMAL: 0-896MB */
848 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
849#ifdef GFP_DMA32
850 if (RT_FAILURE(rc))
851 /* ZONE_DMA32: 0-4GB */
852 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
853#endif
854 if (RT_FAILURE(rc))
855 /* ZONE_DMA: 0-16MB */
856 rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
857 }
858 IPRT_LINUX_RESTORE_EFL_AC();
859 return rc;
860}
861
862
863/**
864 * Translates a kernel virtual address to a linux page structure by walking the
865 * page tables.
866 *
867 * @note We do assume that the page tables will not change as we are walking
868 * them. This assumption is rather forced by the fact that I could not
869 * immediately see any way of preventing this from happening. So, we
870 * take some extra care when accessing them.
871 *
872 * Because of this, we don't want to use this function on memory where
873 * attribute changes to nearby pages is likely to cause large pages to
874 * be used or split up. So, don't use this for the linear mapping of
875 * physical memory.
876 *
877 * @returns Pointer to the page structur or NULL if it could not be found.
878 * @param pv The kernel virtual address.
879 */
880static struct page *rtR0MemObjLinuxVirtToPage(void *pv)
881{
882 unsigned long ulAddr = (unsigned long)pv;
883 unsigned long pfn;
884 struct page *pPage;
885 pte_t *pEntry;
886 union
887 {
888 pgd_t Global;
889#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
890 pud_t Upper;
891#endif
892 pmd_t Middle;
893 pte_t Entry;
894 } u;
895
896 /* Should this happen in a situation this code will be called in? And if
897 * so, can it change under our feet? See also
898 * "Documentation/vm/active_mm.txt" in the kernel sources. */
899 if (RT_UNLIKELY(!current->active_mm))
900 return NULL;
901 u.Global = *pgd_offset(current->active_mm, ulAddr);
902 if (RT_UNLIKELY(pgd_none(u.Global)))
903 return NULL;
904
905#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
906 u.Upper = *pud_offset(&u.Global, ulAddr);
907 if (RT_UNLIKELY(pud_none(u.Upper)))
908 return NULL;
909# if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 25)
910 if (pud_large(u.Upper))
911 {
912 pPage = pud_page(u.Upper);
913 AssertReturn(pPage, NULL);
914 pfn = page_to_pfn(pPage); /* doing the safe way... */
915 pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
916 return pfn_to_page(pfn);
917 }
918# endif
919
920 u.Middle = *pmd_offset(&u.Upper, ulAddr);
921#else /* < 2.6.11 */
922 u.Middle = *pmd_offset(&u.Global, ulAddr);
923#endif /* < 2.6.11 */
924 if (RT_UNLIKELY(pmd_none(u.Middle)))
925 return NULL;
926#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0)
927 if (pmd_large(u.Middle))
928 {
929 pPage = pmd_page(u.Middle);
930 AssertReturn(pPage, NULL);
931 pfn = page_to_pfn(pPage); /* doing the safe way... */
932 pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
933 return pfn_to_page(pfn);
934 }
935#endif
936
937#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) || defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
938 pEntry = pte_offset_map(&u.Middle, ulAddr);
939#else
940 pEntry = pte_offset(&u.Middle, ulAddr);
941#endif
942 if (RT_UNLIKELY(!pEntry))
943 return NULL;
944 u.Entry = *pEntry;
945#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) || defined(pte_offset_map)
946 pte_unmap(pEntry);
947#endif
948
949 if (RT_UNLIKELY(!pte_present(u.Entry)))
950 return NULL;
951 return pte_page(u.Entry);
952}
953
954
955DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
956{
957 return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
958}
959
960
961DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
962{
963 return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
964}
965
966
967DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
968{
969 IPRT_LINUX_SAVE_EFL_AC();
970
971 /*
972 * All we need to do here is to validate that we can use
973 * ioremap on the specified address (32/64-bit dma_addr_t).
974 */
975 PRTR0MEMOBJLNX pMemLnx;
976 dma_addr_t PhysAddr = Phys;
977 AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
978
979 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
980 if (!pMemLnx)
981 {
982 IPRT_LINUX_RESTORE_EFL_AC();
983 return VERR_NO_MEMORY;
984 }
985
986 pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
987 pMemLnx->Core.u.Phys.fAllocated = false;
988 pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
989 Assert(!pMemLnx->cPages);
990 *ppMem = &pMemLnx->Core;
991 IPRT_LINUX_RESTORE_EFL_AC();
992 return VINF_SUCCESS;
993}
994
995
996DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
997{
998 IPRT_LINUX_SAVE_EFL_AC();
999 const int cPages = cb >> PAGE_SHIFT;
1000 struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1001 struct vm_area_struct **papVMAs;
1002 PRTR0MEMOBJLNX pMemLnx;
1003 int rc = VERR_NO_MEMORY;
1004 int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1005
1006 /*
1007 * Check for valid task and size overflows.
1008 */
1009 if (!pTask)
1010 return VERR_NOT_SUPPORTED;
1011 if (((size_t)cPages << PAGE_SHIFT) != cb)
1012 return VERR_OUT_OF_RANGE;
1013
1014 /*
1015 * Allocate the memory object and a temporary buffer for the VMAs.
1016 */
1017 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
1018 if (!pMemLnx)
1019 {
1020 IPRT_LINUX_RESTORE_EFL_AC();
1021 return VERR_NO_MEMORY;
1022 }
1023
1024 papVMAs = (struct vm_area_struct **)RTMemAlloc(sizeof(*papVMAs) * cPages);
1025 if (papVMAs)
1026 {
1027 down_read(&pTask->mm->mmap_sem);
1028
1029 /*
1030 * Get user pages.
1031 */
1032 rc = get_user_pages(pTask, /* Task for fault accounting. */
1033 pTask->mm, /* Whose pages. */
1034 R3Ptr, /* Where from. */
1035 cPages, /* How many pages. */
1036 fWrite, /* Write to memory. */
1037 fWrite, /* force write access. */
1038 &pMemLnx->apPages[0], /* Page array. */
1039 papVMAs); /* vmas */
1040 if (rc == cPages)
1041 {
1042 /*
1043 * Flush dcache (required?), protect against fork and _really_ pin the page
1044 * table entries. get_user_pages() will protect against swapping out the
1045 * pages but it will NOT protect against removing page table entries. This
1046 * can be achieved with
1047 * - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1048 * an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1049 * Usual Linux distributions support only a limited size of locked pages
1050 * (e.g. 32KB).
1051 * - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1052 * or by
1053 * - setting the VM_LOCKED flag. This is the same as doing mlock() without
1054 * a range check.
1055 */
1056 /** @todo The Linux fork() protection will require more work if this API
1057 * is to be used for anything but locking VM pages. */
1058 while (rc-- > 0)
1059 {
1060 flush_dcache_page(pMemLnx->apPages[rc]);
1061 papVMAs[rc]->vm_flags |= (VM_DONTCOPY | VM_LOCKED);
1062 }
1063
1064 up_read(&pTask->mm->mmap_sem);
1065
1066 RTMemFree(papVMAs);
1067
1068 pMemLnx->Core.u.Lock.R0Process = R0Process;
1069 pMemLnx->cPages = cPages;
1070 Assert(!pMemLnx->fMappedToRing0);
1071 *ppMem = &pMemLnx->Core;
1072
1073 IPRT_LINUX_RESTORE_EFL_AC();
1074 return VINF_SUCCESS;
1075 }
1076
1077 /*
1078 * Failed - we need to unlock any pages that we succeeded to lock.
1079 */
1080 while (rc-- > 0)
1081 {
1082 if (!PageReserved(pMemLnx->apPages[rc]))
1083 SetPageDirty(pMemLnx->apPages[rc]);
1084 page_cache_release(pMemLnx->apPages[rc]);
1085 }
1086
1087 up_read(&pTask->mm->mmap_sem);
1088
1089 RTMemFree(papVMAs);
1090 rc = VERR_LOCK_FAILED;
1091 }
1092
1093 rtR0MemObjDelete(&pMemLnx->Core);
1094 IPRT_LINUX_RESTORE_EFL_AC();
1095 return rc;
1096}
1097
1098
1099DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
1100{
1101 IPRT_LINUX_SAVE_EFL_AC();
1102 void *pvLast = (uint8_t *)pv + cb - 1;
1103 size_t const cPages = cb >> PAGE_SHIFT;
1104 PRTR0MEMOBJLNX pMemLnx;
1105 bool fLinearMapping;
1106 int rc;
1107 uint8_t *pbPage;
1108 size_t iPage;
1109 NOREF(fAccess);
1110
1111 if ( !RTR0MemKernelIsValidAddr(pv)
1112 || !RTR0MemKernelIsValidAddr(pv + cb))
1113 return VERR_INVALID_PARAMETER;
1114
1115 /*
1116 * The lower part of the kernel memory has a linear mapping between
1117 * physical and virtual addresses. So we take a short cut here. This is
1118 * assumed to be the cleanest way to handle those addresses (and the code
1119 * is well tested, though the test for determining it is not very nice).
1120 * If we ever decide it isn't we can still remove it.
1121 */
1122#if 0
1123 fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1124#else
1125 fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1126 && (unsigned long)pvLast < (unsigned long)high_memory;
1127#endif
1128
1129 /*
1130 * Allocate the memory object.
1131 */
1132 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
1133 if (!pMemLnx)
1134 {
1135 IPRT_LINUX_RESTORE_EFL_AC();
1136 return VERR_NO_MEMORY;
1137 }
1138
1139 /*
1140 * Gather the pages.
1141 * We ASSUME all kernel pages are non-swappable and non-movable.
1142 */
1143 rc = VINF_SUCCESS;
1144 pbPage = (uint8_t *)pvLast;
1145 iPage = cPages;
1146 if (!fLinearMapping)
1147 {
1148 while (iPage-- > 0)
1149 {
1150 struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1151 if (RT_UNLIKELY(!pPage))
1152 {
1153 rc = VERR_LOCK_FAILED;
1154 break;
1155 }
1156 pMemLnx->apPages[iPage] = pPage;
1157 pbPage -= PAGE_SIZE;
1158 }
1159 }
1160 else
1161 {
1162 while (iPage-- > 0)
1163 {
1164 pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1165 pbPage -= PAGE_SIZE;
1166 }
1167 }
1168 if (RT_SUCCESS(rc))
1169 {
1170 /*
1171 * Complete the memory object and return.
1172 */
1173 pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1174 pMemLnx->cPages = cPages;
1175 Assert(!pMemLnx->fMappedToRing0);
1176 *ppMem = &pMemLnx->Core;
1177
1178 IPRT_LINUX_RESTORE_EFL_AC();
1179 return VINF_SUCCESS;
1180 }
1181
1182 rtR0MemObjDelete(&pMemLnx->Core);
1183 IPRT_LINUX_RESTORE_EFL_AC();
1184 return rc;
1185}
1186
1187
1188DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
1189{
1190#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1191 IPRT_LINUX_SAVE_EFL_AC();
1192 const size_t cPages = cb >> PAGE_SHIFT;
1193 struct page *pDummyPage;
1194 struct page **papPages;
1195
1196 /* check for unsupported stuff. */
1197 AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1198 if (uAlignment > PAGE_SIZE)
1199 return VERR_NOT_SUPPORTED;
1200
1201 /*
1202 * Allocate a dummy page and create a page pointer array for vmap such that
1203 * the dummy page is mapped all over the reserved area.
1204 */
1205 pDummyPage = alloc_page(GFP_HIGHUSER | __GFP_NOWARN);
1206 if (pDummyPage)
1207 {
1208 papPages = RTMemAlloc(sizeof(*papPages) * cPages);
1209 if (papPages)
1210 {
1211 void *pv;
1212 size_t iPage = cPages;
1213 while (iPage-- > 0)
1214 papPages[iPage] = pDummyPage;
1215# ifdef VM_MAP
1216 pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1217# else
1218 pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1219# endif
1220 RTMemFree(papPages);
1221 if (pv)
1222 {
1223 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1224 if (pMemLnx)
1225 {
1226 pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1227 pMemLnx->cPages = 1;
1228 pMemLnx->apPages[0] = pDummyPage;
1229 *ppMem = &pMemLnx->Core;
1230 IPRT_LINUX_RESTORE_EFL_AC();
1231 return VINF_SUCCESS;
1232 }
1233 vunmap(pv);
1234 }
1235 }
1236 __free_page(pDummyPage);
1237 }
1238 IPRT_LINUX_RESTORE_EFL_AC();
1239 return VERR_NO_MEMORY;
1240
1241#else /* < 2.4.22 */
1242 /*
1243 * Could probably use ioremap here, but the caller is in a better position than us
1244 * to select some safe physical memory.
1245 */
1246 return VERR_NOT_SUPPORTED;
1247#endif
1248}
1249
1250
1251DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1252{
1253 IPRT_LINUX_SAVE_EFL_AC();
1254 PRTR0MEMOBJLNX pMemLnx;
1255 void *pv;
1256 struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1257 if (!pTask)
1258 return VERR_NOT_SUPPORTED;
1259
1260 /*
1261 * Check that the specified alignment is supported.
1262 */
1263 if (uAlignment > PAGE_SIZE)
1264 return VERR_NOT_SUPPORTED;
1265
1266 /*
1267 * Let rtR0MemObjLinuxDoMmap do the difficult bits.
1268 */
1269 pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1270 if (pv == (void *)-1)
1271 {
1272 IPRT_LINUX_RESTORE_EFL_AC();
1273 return VERR_NO_MEMORY;
1274 }
1275
1276 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1277 if (!pMemLnx)
1278 {
1279 rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1280 IPRT_LINUX_RESTORE_EFL_AC();
1281 return VERR_NO_MEMORY;
1282 }
1283
1284 pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1285 *ppMem = &pMemLnx->Core;
1286 IPRT_LINUX_RESTORE_EFL_AC();
1287 return VINF_SUCCESS;
1288}
1289
1290
1291DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap,
1292 void *pvFixed, size_t uAlignment,
1293 unsigned fProt, size_t offSub, size_t cbSub)
1294{
1295 int rc = VERR_NO_MEMORY;
1296 PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1297 PRTR0MEMOBJLNX pMemLnx;
1298 IPRT_LINUX_SAVE_EFL_AC();
1299
1300 /* Fail if requested to do something we can't. */
1301 AssertMsgReturn(!offSub && !cbSub, ("%#x %#x\n", offSub, cbSub), VERR_NOT_SUPPORTED);
1302 AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1303 if (uAlignment > PAGE_SIZE)
1304 return VERR_NOT_SUPPORTED;
1305
1306 /*
1307 * Create the IPRT memory object.
1308 */
1309 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1310 if (pMemLnx)
1311 {
1312 if (pMemLnxToMap->cPages)
1313 {
1314#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1315 /*
1316 * Use vmap - 2.4.22 and later.
1317 */
1318 pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1319# ifdef VM_MAP
1320 pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
1321# else
1322 pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
1323# endif
1324 if (pMemLnx->Core.pv)
1325 {
1326 pMemLnx->fMappedToRing0 = true;
1327 rc = VINF_SUCCESS;
1328 }
1329 else
1330 rc = VERR_MAP_FAILED;
1331
1332#else /* < 2.4.22 */
1333 /*
1334 * Only option here is to share mappings if possible and forget about fProt.
1335 */
1336 if (rtR0MemObjIsRing3(pMemToMap))
1337 rc = VERR_NOT_SUPPORTED;
1338 else
1339 {
1340 rc = VINF_SUCCESS;
1341 if (!pMemLnxToMap->Core.pv)
1342 rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1343 if (RT_SUCCESS(rc))
1344 {
1345 Assert(pMemLnxToMap->Core.pv);
1346 pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
1347 }
1348 }
1349#endif
1350 }
1351 else
1352 {
1353 /*
1354 * MMIO / physical memory.
1355 */
1356 Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1357 pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1358 ? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb)
1359 : ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
1360 if (pMemLnx->Core.pv)
1361 {
1362 /** @todo fix protection. */
1363 rc = VINF_SUCCESS;
1364 }
1365 }
1366 if (RT_SUCCESS(rc))
1367 {
1368 pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1369 *ppMem = &pMemLnx->Core;
1370 IPRT_LINUX_RESTORE_EFL_AC();
1371 return VINF_SUCCESS;
1372 }
1373 rtR0MemObjDelete(&pMemLnx->Core);
1374 }
1375
1376 IPRT_LINUX_RESTORE_EFL_AC();
1377 return rc;
1378}
1379
1380
1381#ifdef VBOX_USE_PAE_HACK
1382/**
1383 * Replace the PFN of a PTE with the address of the actual page.
1384 *
1385 * The caller maps a reserved dummy page at the address with the desired access
1386 * and flags.
1387 *
1388 * This hack is required for older Linux kernels which don't provide
1389 * remap_pfn_range().
1390 *
1391 * @returns 0 on success, -ENOMEM on failure.
1392 * @param mm The memory context.
1393 * @param ulAddr The mapping address.
1394 * @param Phys The physical address of the page to map.
1395 */
1396static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1397{
1398 int rc = -ENOMEM;
1399 pgd_t *pgd;
1400
1401 spin_lock(&mm->page_table_lock);
1402
1403 pgd = pgd_offset(mm, ulAddr);
1404 if (!pgd_none(*pgd) && !pgd_bad(*pgd))
1405 {
1406 pmd_t *pmd = pmd_offset(pgd, ulAddr);
1407 if (!pmd_none(*pmd))
1408 {
1409 pte_t *ptep = pte_offset_map(pmd, ulAddr);
1410 if (ptep)
1411 {
1412 pte_t pte = *ptep;
1413 pte.pte_high &= 0xfff00000;
1414 pte.pte_high |= ((Phys >> 32) & 0x000fffff);
1415 pte.pte_low &= 0x00000fff;
1416 pte.pte_low |= (Phys & 0xfffff000);
1417 set_pte(ptep, pte);
1418 pte_unmap(ptep);
1419 rc = 0;
1420 }
1421 }
1422 }
1423
1424 spin_unlock(&mm->page_table_lock);
1425 return rc;
1426}
1427#endif /* VBOX_USE_PAE_HACK */
1428
1429
1430DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed,
1431 size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1432{
1433 struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1434 PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1435 int rc = VERR_NO_MEMORY;
1436 PRTR0MEMOBJLNX pMemLnx;
1437#ifdef VBOX_USE_PAE_HACK
1438 struct page *pDummyPage;
1439 RTHCPHYS DummyPhys;
1440#endif
1441 IPRT_LINUX_SAVE_EFL_AC();
1442
1443 /*
1444 * Check for restrictions.
1445 */
1446 if (!pTask)
1447 return VERR_NOT_SUPPORTED;
1448 if (uAlignment > PAGE_SIZE)
1449 return VERR_NOT_SUPPORTED;
1450
1451#ifdef VBOX_USE_PAE_HACK
1452 /*
1453 * Allocate a dummy page for use when mapping the memory.
1454 */
1455 pDummyPage = alloc_page(GFP_USER | __GFP_NOWARN);
1456 if (!pDummyPage)
1457 {
1458 IPRT_LINUX_RESTORE_EFL_AC();
1459 return VERR_NO_MEMORY;
1460 }
1461 SetPageReserved(pDummyPage);
1462 DummyPhys = page_to_phys(pDummyPage);
1463#endif
1464
1465 /*
1466 * Create the IPRT memory object.
1467 */
1468 pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1469 if (pMemLnx)
1470 {
1471 /*
1472 * Allocate user space mapping.
1473 */
1474 void *pv;
1475 pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1476 if (pv != (void *)-1)
1477 {
1478 /*
1479 * Map page by page into the mmap area.
1480 * This is generic, paranoid and not very efficient.
1481 */
1482 pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1483 unsigned long ulAddrCur = (unsigned long)pv;
1484 const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1485 size_t iPage;
1486
1487 down_write(&pTask->mm->mmap_sem);
1488
1489 rc = VINF_SUCCESS;
1490 if (pMemLnxToMap->cPages)
1491 {
1492 for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1493 {
1494#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)
1495 RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1496#endif
1497#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1498 struct vm_area_struct *vma = find_vma(pTask->mm, ulAddrCur); /* this is probably the same for all the pages... */
1499 AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1500#endif
1501#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1502 /* remap_page_range() limitation on x86 */
1503 AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1504#endif
1505
1506#if defined(VBOX_USE_INSERT_PAGE) && LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
1507 rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1508 /* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1509 * See remap_pfn_range() in mm/memory.c */
1510#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 7, 0)
1511 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
1512#else
1513 vma->vm_flags |= VM_RESERVED;
1514#endif
1515#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1516 rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1517#elif defined(VBOX_USE_PAE_HACK)
1518 rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1519 if (!rc)
1520 rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1521#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1522 rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1523#else /* 2.4 */
1524 rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1525#endif
1526 if (rc)
1527 {
1528 rc = VERR_NO_MEMORY;
1529 break;
1530 }
1531 }
1532 }
1533 else
1534 {
1535 RTHCPHYS Phys;
1536 if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1537 Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1538 else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1539 Phys = pMemLnxToMap->Core.u.Cont.Phys;
1540 else
1541 {
1542 AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1543 Phys = NIL_RTHCPHYS;
1544 }
1545 if (Phys != NIL_RTHCPHYS)
1546 {
1547 for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1548 {
1549#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1550 struct vm_area_struct *vma = find_vma(pTask->mm, ulAddrCur); /* this is probably the same for all the pages... */
1551 AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1552#endif
1553#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1554 /* remap_page_range() limitation on x86 */
1555 AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1556#endif
1557
1558#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1559 rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1560#elif defined(VBOX_USE_PAE_HACK)
1561 rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1562 if (!rc)
1563 rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1564#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) || defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1565 rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1566#else /* 2.4 */
1567 rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1568#endif
1569 if (rc)
1570 {
1571 rc = VERR_NO_MEMORY;
1572 break;
1573 }
1574 }
1575 }
1576 }
1577
1578#ifdef CONFIG_NUMA_BALANCING
1579# if LINUX_VERSION_CODE < KERNEL_VERSION(3, 13, 0)
1580# ifdef RHEL_RELEASE_CODE
1581# if RHEL_RELEASE_CODE < RHEL_RELEASE_VERSION(7, 0)
1582# define VBOX_NUMA_HACK_OLD
1583# endif
1584# endif
1585# endif
1586 if (RT_SUCCESS(rc))
1587 {
1588 /** @todo Ugly hack! But right now we have no other means to
1589 * disable automatic NUMA page balancing. */
1590# ifdef RT_OS_X86
1591# ifdef VBOX_NUMA_HACK_OLD
1592 pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1593# endif
1594 pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1595# else
1596# ifdef VBOX_NUMA_HACK_OLD
1597 pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1598# endif
1599 pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1600# endif
1601 }
1602#endif /* CONFIG_NUMA_BALANCING */
1603
1604 up_write(&pTask->mm->mmap_sem);
1605
1606 if (RT_SUCCESS(rc))
1607 {
1608#ifdef VBOX_USE_PAE_HACK
1609 __free_page(pDummyPage);
1610#endif
1611 pMemLnx->Core.pv = pv;
1612 pMemLnx->Core.u.Mapping.R0Process = R0Process;
1613 *ppMem = &pMemLnx->Core;
1614 IPRT_LINUX_RESTORE_EFL_AC();
1615 return VINF_SUCCESS;
1616 }
1617
1618 /*
1619 * Bail out.
1620 */
1621 rtR0MemObjLinuxDoMunmap(pv, pMemLnxToMap->Core.cb, pTask);
1622 }
1623 rtR0MemObjDelete(&pMemLnx->Core);
1624 }
1625#ifdef VBOX_USE_PAE_HACK
1626 __free_page(pDummyPage);
1627#endif
1628
1629 IPRT_LINUX_RESTORE_EFL_AC();
1630 return rc;
1631}
1632
1633
1634DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1635{
1636 NOREF(pMem);
1637 NOREF(offSub);
1638 NOREF(cbSub);
1639 NOREF(fProt);
1640 return VERR_NOT_SUPPORTED;
1641}
1642
1643
1644DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1645{
1646 PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1647
1648 if (pMemLnx->cPages)
1649 return page_to_phys(pMemLnx->apPages[iPage]);
1650
1651 switch (pMemLnx->Core.enmType)
1652 {
1653 case RTR0MEMOBJTYPE_CONT:
1654 return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1655
1656 case RTR0MEMOBJTYPE_PHYS:
1657 return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1658
1659 /* the parent knows */
1660 case RTR0MEMOBJTYPE_MAPPING:
1661 return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1662
1663 /* cPages > 0 */
1664 case RTR0MEMOBJTYPE_LOW:
1665 case RTR0MEMOBJTYPE_LOCK:
1666 case RTR0MEMOBJTYPE_PHYS_NC:
1667 case RTR0MEMOBJTYPE_PAGE:
1668 default:
1669 AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1670 /* fall thru */
1671
1672 case RTR0MEMOBJTYPE_RES_VIRT:
1673 return NIL_RTHCPHYS;
1674 }
1675}
1676
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