Ticket #10709: memobj-r0drv-linux.c

File memobj-r0drv-linux.c, 51.3 KB (added by Andre Robatino, 12 years ago)
/usr/src/vboxhost-4.1.18/vboxdrv/r0drv/linux/memobj-r0drv-linux.c from F17 x86_64 host

Line
1	/* $Revision: 75790 $ */
2	/** @file
3	* IPRT - Ring-0 Memory Objects, Linux.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*******************************************************************************
29	* Header Files *
30	*******************************************************************************/
31	#include "the-linux-kernel.h"
32
33	#include <iprt/memobj.h>
34	#include <iprt/alloc.h>
35	#include <iprt/assert.h>
36	#include <iprt/log.h>
37	#include <iprt/process.h>
38	#include <iprt/string.h>
39	#include "internal/memobj.h"
40
41
42	/*******************************************************************************
43	* Defined Constants And Macros *
44	*******************************************************************************/
45	/* early 2.6 kernels */
46	#ifndef PAGE_SHARED_EXEC
47	# define PAGE_SHARED_EXEC PAGE_SHARED
48	#endif
49	#ifndef PAGE_READONLY_EXEC
50	# define PAGE_READONLY_EXEC PAGE_READONLY
51	#endif
52
53	/*
54	* 2.6.29+ kernels don't work with remap_pfn_range() anymore because
55	* track_pfn_vma_new() is apparently not defined for non-RAM pages.
56	* It should be safe to use vm_insert_page() older kernels as well.
57	*/
58	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 23)
59	# define VBOX_USE_INSERT_PAGE
60	#endif
61	#if defined(CONFIG_X86_PAE) \
62	&& ( defined(HAVE_26_STYLE_REMAP_PAGE_RANGE) \
63	\|\| ( LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \
64	&& LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)))
65	# define VBOX_USE_PAE_HACK
66	#endif
67
68
69	/*******************************************************************************
70	* Structures and Typedefs *
71	*******************************************************************************/
72	/**
73	* The Darwin version of the memory object structure.
74	*/
75	typedef 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
91	static 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	*/
100	static 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	*/
114	static 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	*/
135	static 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	*/
186	static 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	*/
256	static 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	*/
285	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
286	size_t uAlignment, unsigned fFlagsLnx, bool fContiguous)
287	{
288	size_t iPage;
289	size_t const cPages = cb >> PAGE_SHIFT;
290	struct page *paPages;
291
292	/*
293	* Allocate a memory object structure that's large enough to contain
294	* the page pointer array.
295	*/
296	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
297	if (!pMemLnx)
298	return VERR_NO_MEMORY;
299	pMemLnx->cPages = cPages;
300
301	if (cPages > 255)
302	{
303	# ifdef __GFP_REPEAT
304	/* Try hard to allocate the memory, but the allocation attempt might fail. */
305	fFlagsLnx \|= __GFP_REPEAT;
306	# endif
307	# ifdef __GFP_NOMEMALLOC
308	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
309	fFlagsLnx \|= __GFP_NOMEMALLOC;
310	# endif
311	}
312
313	/*
314	* Allocate the pages.
315	* For small allocations we'll try contiguous first and then fall back on page by page.
316	*/
317	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
318	if ( fContiguous
319	\|\| cb <= PAGE_SIZE * 2)
320	{
321	# ifdef VBOX_USE_INSERT_PAGE
322	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP, rtR0MemObjLinuxOrder(cPages));
323	# else
324	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
325	# endif
326	if (paPages)
327	{
328	fContiguous = true;
329	for (iPage = 0; iPage < cPages; iPage++)
330	pMemLnx->apPages[iPage] = &paPages[iPage];
331	}
332	else if (fContiguous)
333	{
334	rtR0MemObjDelete(&pMemLnx->Core);
335	return VERR_NO_MEMORY;
336	}
337	}
338
339	if (!fContiguous)
340	{
341	for (iPage = 0; iPage < cPages; iPage++)
342	{
343	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx);
344	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
345	{
346	while (iPage-- > 0)
347	__free_page(pMemLnx->apPages[iPage]);
348	rtR0MemObjDelete(&pMemLnx->Core);
349	return VERR_NO_MEMORY;
350	}
351	}
352	}
353
354	#else /* < 2.4.22 */
355	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
356	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
357	if (!paPages)
358	{
359	rtR0MemObjDelete(&pMemLnx->Core);
360	return VERR_NO_MEMORY;
361	}
362	for (iPage = 0; iPage < cPages; iPage++)
363	{
364	pMemLnx->apPages[iPage] = &paPages[iPage];
365	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
366	if (PageHighMem(pMemLnx->apPages[iPage]))
367	BUG();
368	}
369
370	fContiguous = true;
371	#endif /* < 2.4.22 */
372	pMemLnx->fContiguous = fContiguous;
373
374	/*
375	* Reserve the pages.
376	*/
377	for (iPage = 0; iPage < cPages; iPage++)
378	SetPageReserved(pMemLnx->apPages[iPage]);
379
380	/*
381	* Note that the physical address of memory allocated with alloc_pages(flags, order)
382	* is always 2^(PAGE_SHIFT+order)-aligned.
383	*/
384	if ( fContiguous
385	&& uAlignment > PAGE_SIZE)
386	{
387	/*
388	* Check for alignment constraints. The physical address of memory allocated with
389	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
390	*/
391	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
392	{
393	/*
394	* This should never happen!
395	*/
396	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
397	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
398	rtR0MemObjLinuxFreePages(pMemLnx);
399	return VERR_NO_MEMORY;
400	}
401	}
402
403	*ppMemLnx = pMemLnx;
404	return VINF_SUCCESS;
405	}
406
407
408	/**
409	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
410	*
411	* This method does NOT free the object.
412	*
413	* @param pMemLnx The object which physical pages should be freed.
414	*/
415	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
416	{
417	size_t iPage = pMemLnx->cPages;
418	if (iPage > 0)
419	{
420	/*
421	* Restore the page flags.
422	*/
423	while (iPage-- > 0)
424	{
425	ClearPageReserved(pMemLnx->apPages[iPage]);
426	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
427	#else
428	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
429	#endif
430	}
431
432	/*
433	* Free the pages.
434	*/
435	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
436	if (!pMemLnx->fContiguous)
437	{
438	iPage = pMemLnx->cPages;
439	while (iPage-- > 0)
440	__free_page(pMemLnx->apPages[iPage]);
441	}
442	else
443	#endif
444	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
445
446	pMemLnx->cPages = 0;
447	}
448	}
449
450
451	/**
452	* Maps the allocation into ring-0.
453	*
454	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
455	*
456	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
457	* space, so we'll use that mapping if possible. If execute access is required, we'll
458	* play safe and do our own mapping.
459	*
460	* @returns IPRT status code.
461	* @param pMemLnx The linux memory object to map.
462	* @param fExecutable Whether execute access is required.
463	*/
464	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
465	{
466	int rc = VINF_SUCCESS;
467
468	/*
469	* Choose mapping strategy.
470	*/
471	bool fMustMap = fExecutable
472	\|\| !pMemLnx->fContiguous;
473	if (!fMustMap)
474	{
475	size_t iPage = pMemLnx->cPages;
476	while (iPage-- > 0)
477	if (PageHighMem(pMemLnx->apPages[iPage]))
478	{
479	fMustMap = true;
480	break;
481	}
482	}
483
484	Assert(!pMemLnx->Core.pv);
485	Assert(!pMemLnx->fMappedToRing0);
486
487	if (fMustMap)
488	{
489	/*
490	* Use vmap - 2.4.22 and later.
491	*/
492	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
493	pgprot_t fPg;
494	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
495	# ifdef _PAGE_NX
496	if (!fExecutable)
497	pgprot_val(fPg) \|= _PAGE_NX;
498	# endif
499
500	# ifdef VM_MAP
501	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
502	# else
503	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
504	# endif
505	if (pMemLnx->Core.pv)
506	pMemLnx->fMappedToRing0 = true;
507	else
508	rc = VERR_MAP_FAILED;
509	#else /* < 2.4.22 */
510	rc = VERR_NOT_SUPPORTED;
511	#endif
512	}
513	else
514	{
515	/*
516	* Use the kernel RAM mapping.
517	*/
518	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
519	Assert(pMemLnx->Core.pv);
520	}
521
522	return rc;
523	}
524
525
526	/**
527	* Undoes what rtR0MemObjLinuxVMap() did.
528	*
529	* @param pMemLnx The linux memory object.
530	*/
531	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
532	{
533	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
534	if (pMemLnx->fMappedToRing0)
535	{
536	Assert(pMemLnx->Core.pv);
537	vunmap(pMemLnx->Core.pv);
538	pMemLnx->fMappedToRing0 = false;
539	}
540	#else /* < 2.4.22 */
541	Assert(!pMemLnx->fMappedToRing0);
542	#endif
543	pMemLnx->Core.pv = NULL;
544	}
545
546
547	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
548	{
549	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
550
551	/*
552	* Release any memory that we've allocated or locked.
553	*/
554	switch (pMemLnx->Core.enmType)
555	{
556	case RTR0MEMOBJTYPE_LOW:
557	case RTR0MEMOBJTYPE_PAGE:
558	case RTR0MEMOBJTYPE_CONT:
559	case RTR0MEMOBJTYPE_PHYS:
560	case RTR0MEMOBJTYPE_PHYS_NC:
561	rtR0MemObjLinuxVUnmap(pMemLnx);
562	rtR0MemObjLinuxFreePages(pMemLnx);
563	break;
564
565	case RTR0MEMOBJTYPE_LOCK:
566	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
567	{
568	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
569	size_t iPage;
570	Assert(pTask);
571	if (pTask && pTask->mm)
572	down_read(&pTask->mm->mmap_sem);
573
574	iPage = pMemLnx->cPages;
575	while (iPage-- > 0)
576	{
577	if (!PageReserved(pMemLnx->apPages[iPage]))
578	SetPageDirty(pMemLnx->apPages[iPage]);
579	page_cache_release(pMemLnx->apPages[iPage]);
580	}
581
582	if (pTask && pTask->mm)
583	up_read(&pTask->mm->mmap_sem);
584	}
585	/* else: kernel memory - nothing to do here. */
586	break;
587
588	case RTR0MEMOBJTYPE_RES_VIRT:
589	Assert(pMemLnx->Core.pv);
590	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
591	{
592	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
593	Assert(pTask);
594	if (pTask && pTask->mm)
595	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
596	}
597	else
598	{
599	vunmap(pMemLnx->Core.pv);
600
601	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
602	__free_page(pMemLnx->apPages[0]);
603	pMemLnx->apPages[0] = NULL;
604	pMemLnx->cPages = 0;
605	}
606	pMemLnx->Core.pv = NULL;
607	break;
608
609	case RTR0MEMOBJTYPE_MAPPING:
610	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
611	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
612	{
613	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
614	Assert(pTask);
615	if (pTask && pTask->mm)
616	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
617	}
618	else
619	vunmap(pMemLnx->Core.pv);
620	pMemLnx->Core.pv = NULL;
621	break;
622
623	default:
624	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
625	return VERR_INTERNAL_ERROR;
626	}
627	return VINF_SUCCESS;
628	}
629
630
631	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
632	{
633	PRTR0MEMOBJLNX pMemLnx;
634	int rc;
635
636	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
637	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER, false /* non-contiguous */);
638	#else
639	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER, false /* non-contiguous */);
640	#endif
641	if (RT_SUCCESS(rc))
642	{
643	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
644	if (RT_SUCCESS(rc))
645	{
646	*ppMem = &pMemLnx->Core;
647	return rc;
648	}
649
650	rtR0MemObjLinuxFreePages(pMemLnx);
651	rtR0MemObjDelete(&pMemLnx->Core);
652	}
653
654	return rc;
655	}
656
657
658	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
659	{
660	PRTR0MEMOBJLNX pMemLnx;
661	int rc;
662
663	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
664	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
665	/* ZONE_DMA32: 0-4GB */
666	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32, false /* non-contiguous */);
667	if (RT_FAILURE(rc))
668	#endif
669	#ifdef RT_ARCH_AMD64
670	/* ZONE_DMA: 0-16MB */
671	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA, false /* non-contiguous */);
672	#else
673	# ifdef CONFIG_X86_PAE
674	# endif
675	/* ZONE_NORMAL: 0-896MB */
676	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER, false /* non-contiguous */);
677	#endif
678	if (RT_SUCCESS(rc))
679	{
680	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
681	if (RT_SUCCESS(rc))
682	{
683	*ppMem = &pMemLnx->Core;
684	return rc;
685	}
686
687	rtR0MemObjLinuxFreePages(pMemLnx);
688	rtR0MemObjDelete(&pMemLnx->Core);
689	}
690
691	return rc;
692	}
693
694
695	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
696	{
697	PRTR0MEMOBJLNX pMemLnx;
698	int rc;
699
700	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
701	/* ZONE_DMA32: 0-4GB */
702	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32, true /* contiguous */);
703	if (RT_FAILURE(rc))
704	#endif
705	#ifdef RT_ARCH_AMD64
706	/* ZONE_DMA: 0-16MB */
707	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA, true /* contiguous */);
708	#else
709	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
710	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER, true /* contiguous */);
711	#endif
712	if (RT_SUCCESS(rc))
713	{
714	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
715	if (RT_SUCCESS(rc))
716	{
717	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
718	size_t iPage = pMemLnx->cPages;
719	while (iPage-- > 0)
720	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
721	#endif
722	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
723	*ppMem = &pMemLnx->Core;
724	return rc;
725	}
726
727	rtR0MemObjLinuxFreePages(pMemLnx);
728	rtR0MemObjDelete(&pMemLnx->Core);
729	}
730
731	return rc;
732	}
733
734
735	/**
736	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
737	*
738	* @returns IPRT status.
739	* @param ppMemLnx Where to
740	* @param enmType The object type.
741	* @param cb The size of the allocation.
742	* @param uAlignment The alignment of the physical memory.
743	* Only valid for fContiguous == true, ignored otherwise.
744	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
745	* @param fGfp The Linux GFP flags to use for the allocation.
746	*/
747	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
748	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, unsigned fGfp)
749	{
750	PRTR0MEMOBJLNX pMemLnx;
751	int rc;
752
753	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
754	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */);
755	if (RT_FAILURE(rc))
756	return rc;
757
758	/*
759	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
760	*/
761	if (PhysHighest != NIL_RTHCPHYS)
762	{
763	size_t iPage = pMemLnx->cPages;
764	while (iPage-- > 0)
765	if (page_to_phys(pMemLnx->apPages[iPage]) >= PhysHighest)
766	{
767	rtR0MemObjLinuxFreePages(pMemLnx);
768	rtR0MemObjDelete(&pMemLnx->Core);
769	return VERR_NO_MEMORY;
770	}
771	}
772
773	/*
774	* Complete the object.
775	*/
776	if (enmType == RTR0MEMOBJTYPE_PHYS)
777	{
778	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
779	pMemLnx->Core.u.Phys.fAllocated = true;
780	}
781	*ppMem = &pMemLnx->Core;
782	return rc;
783	}
784
785
786	/**
787	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
788	*
789	* @returns IPRT status.
790	* @param ppMem Where to store the memory object pointer on success.
791	* @param enmType The object type.
792	* @param cb The size of the allocation.
793	* @param uAlignment The alignment of the physical memory.
794	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
795	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
796	*/
797	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
798	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
799	{
800	int rc;
801
802	/*
803	* There are two clear cases and that's the <=16MB and anything-goes ones.
804	* When the physical address limit is somewhere in-between those two we'll
805	* just have to try, starting with HIGHUSER and working our way thru the
806	* different types, hoping we'll get lucky.
807	*
808	* We should probably move this physical address restriction logic up to
809	* the page alloc function as it would be more efficient there. But since
810	* we don't expect this to be a performance issue just yet it can wait.
811	*/
812	if (PhysHighest == NIL_RTHCPHYS)
813	/* ZONE_HIGHMEM: the whole physical memory */
814	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
815	else if (PhysHighest <= _1M * 16)
816	/* ZONE_DMA: 0-16MB */
817	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
818	else
819	{
820	rc = VERR_NO_MEMORY;
821	if (RT_FAILURE(rc))
822	/* ZONE_HIGHMEM: the whole physical memory */
823	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
824	if (RT_FAILURE(rc))
825	/* ZONE_NORMAL: 0-896MB */
826	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
827	#ifdef GFP_DMA32
828	if (RT_FAILURE(rc))
829	/* ZONE_DMA32: 0-4GB */
830	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
831	#endif
832	if (RT_FAILURE(rc))
833	/* ZONE_DMA: 0-16MB */
834	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
835	}
836	return rc;
837	}
838
839
840	/**
841	* Translates a kernel virtual address to a linux page structure by walking the
842	* page tables.
843	*
844	* @note We do assume that the page tables will not change as we are walking
845	* them. This assumption is rather forced by the fact that I could not
846	* immediately see any way of preventing this from happening. So, we
847	* take some extra care when accessing them.
848	*
849	* Because of this, we don't want to use this function on memory where
850	* attribute changes to nearby pages is likely to cause large pages to
851	* be used or split up. So, don't use this for the linear mapping of
852	* physical memory.
853	*
854	* @returns Pointer to the page structur or NULL if it could not be found.
855	* @param pv The kernel virtual address.
856	*/
857	static struct page rtR0MemObjLinuxVirtToPage(void pv)
858	{
859	unsigned long ulAddr = (unsigned long)pv;
860	unsigned long pfn;
861	struct page *pPage;
862	pte_t *pEntry;
863	union
864	{
865	pgd_t Global;
866	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
867	pud_t Upper;
868	#endif
869	pmd_t Middle;
870	pte_t Entry;
871	} u;
872
873	/* Should this happen in a situation this code will be called in? And if
874	* so, can it change under our feet? See also
875	* "Documentation/vm/active_mm.txt" in the kernel sources. */
876	if (RT_UNLIKELY(!current->active_mm))
877	return NULL;
878	u.Global = *pgd_offset(current->active_mm, ulAddr);
879	if (RT_UNLIKELY(pgd_none(u.Global)))
880	return NULL;
881
882	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
883	u.Upper = *pud_offset(&u.Global, ulAddr);
884	if (RT_UNLIKELY(pud_none(u.Upper)))
885	return NULL;
886	# if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 25)
887	if (pud_large(u.Upper))
888	{
889	pPage = pud_page(u.Upper);
890	AssertReturn(pPage, NULL);
891	pfn = page_to_pfn(pPage); /* doing the safe way... */
892	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
893	return pfn_to_page(pfn);
894	}
895	# endif
896
897	u.Middle = *pmd_offset(&u.Upper, ulAddr);
898	#else /* < 2.6.11 */
899	u.Middle = *pmd_offset(&u.Global, ulAddr);
900	#endif /* < 2.6.11 */
901	if (RT_UNLIKELY(pmd_none(u.Middle)))
902	return NULL;
903	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0)
904	if (pmd_large(u.Middle))
905	{
906	pPage = pmd_page(u.Middle);
907	AssertReturn(pPage, NULL);
908	pfn = page_to_pfn(pPage); /* doing the safe way... */
909	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
910	return pfn_to_page(pfn);
911	}
912	#endif
913
914	/* As usual, RHEL 3 had pte_offset_map earlier. */
915	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) \|\| defined(pte_offset_map)
916	pEntry = pte_offset_map(&u.Middle, ulAddr);
917	#else
918	pEntry = pte_offset(&u.Middle, ulAddr);
919	#endif
920	if (RT_UNLIKELY(!pEntry))
921	return NULL;
922	u.Entry = *pEntry;
923	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) \|\| defined(pte_offset_map)
924	pte_unmap(pEntry);
925	#endif
926
927	if (RT_UNLIKELY(!pte_present(u.Entry)))
928	return NULL;
929	return pte_page(u.Entry);
930	}
931
932
933	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
934	{
935	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
936	}
937
938
939	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
940	{
941	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
942	}
943
944
945	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
946	{
947	/*
948	* All we need to do here is to validate that we can use
949	* ioremap on the specified address (32/64-bit dma_addr_t).
950	*/
951	PRTR0MEMOBJLNX pMemLnx;
952	dma_addr_t PhysAddr = Phys;
953	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
954
955	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
956	if (!pMemLnx)
957	return VERR_NO_MEMORY;
958
959	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
960	pMemLnx->Core.u.Phys.fAllocated = false;
961	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
962	Assert(!pMemLnx->cPages);
963	*ppMem = &pMemLnx->Core;
964	return VINF_SUCCESS;
965	}
966
967
968	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
969	{
970	const int cPages = cb >> PAGE_SHIFT;
971	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
972	struct vm_area_struct **papVMAs;
973	PRTR0MEMOBJLNX pMemLnx;
974	int rc = VERR_NO_MEMORY;
975	NOREF(fAccess);
976
977	/*
978	* Check for valid task and size overflows.
979	*/
980	if (!pTask)
981	return VERR_NOT_SUPPORTED;
982	if (((size_t)cPages << PAGE_SHIFT) != cb)
983	return VERR_OUT_OF_RANGE;
984
985	/*
986	* Allocate the memory object and a temporary buffer for the VMAs.
987	*/
988	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
989	if (!pMemLnx)
990	return VERR_NO_MEMORY;
991
992	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
993	if (papVMAs)
994	{
995	down_read(&pTask->mm->mmap_sem);
996
997	/*
998	* Get user pages.
999	*/
1000	rc = get_user_pages(pTask, /* Task for fault accounting. */
1001	pTask->mm, /* Whose pages. */
1002	R3Ptr, /* Where from. */
1003	cPages, /* How many pages. */
1004	1, /* Write to memory. */
1005	0, /* force. */
1006	&pMemLnx->apPages[0], /* Page array. */
1007	papVMAs); /* vmas */
1008	if (rc == cPages)
1009	{
1010	/*
1011	* Flush dcache (required?), protect against fork and _really_ pin the page
1012	* table entries. get_user_pages() will protect against swapping out the
1013	* pages but it will NOT protect against removing page table entries. This
1014	* can be achieved with
1015	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1016	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1017	* Usual Linux distributions support only a limited size of locked pages
1018	* (e.g. 32KB).
1019	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1020	* or by
1021	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1022	* a range check.
1023	*/
1024	/** @todo The Linux fork() protection will require more work if this API
1025	* is to be used for anything but locking VM pages. */
1026	while (rc-- > 0)
1027	{
1028	flush_dcache_page(pMemLnx->apPages[rc]);
1029	papVMAs[rc]->vm_flags \|= (VM_DONTCOPY \| VM_LOCKED);
1030	}
1031
1032	up_read(&pTask->mm->mmap_sem);
1033
1034	RTMemFree(papVMAs);
1035
1036	pMemLnx->Core.u.Lock.R0Process = R0Process;
1037	pMemLnx->cPages = cPages;
1038	Assert(!pMemLnx->fMappedToRing0);
1039	*ppMem = &pMemLnx->Core;
1040
1041	return VINF_SUCCESS;
1042	}
1043
1044	/*
1045	* Failed - we need to unlock any pages that we succeeded to lock.
1046	*/
1047	while (rc-- > 0)
1048	{
1049	if (!PageReserved(pMemLnx->apPages[rc]))
1050	SetPageDirty(pMemLnx->apPages[rc]);
1051	page_cache_release(pMemLnx->apPages[rc]);
1052	}
1053
1054	up_read(&pTask->mm->mmap_sem);
1055
1056	RTMemFree(papVMAs);
1057	rc = VERR_LOCK_FAILED;
1058	}
1059
1060	rtR0MemObjDelete(&pMemLnx->Core);
1061	return rc;
1062	}
1063
1064
1065	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
1066	{
1067	void pvLast = (uint8_t )pv + cb - 1;
1068	size_t const cPages = cb >> PAGE_SHIFT;
1069	PRTR0MEMOBJLNX pMemLnx;
1070	bool fLinearMapping;
1071	int rc;
1072	uint8_t *pbPage;
1073	size_t iPage;
1074	NOREF(fAccess);
1075
1076	if ( !RTR0MemKernelIsValidAddr(pv)
1077	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1078	return VERR_INVALID_PARAMETER;
1079
1080	/*
1081	* The lower part of the kernel memory has a linear mapping between
1082	* physical and virtual addresses. So we take a short cut here. This is
1083	* assumed to be the cleanest way to handle those addresses (and the code
1084	* is well tested, though the test for determining it is not very nice).
1085	* If we ever decide it isn't we can still remove it.
1086	*/
1087	#if 0
1088	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1089	#else
1090	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1091	&& (unsigned long)pvLast < (unsigned long)high_memory;
1092	#endif
1093
1094	/*
1095	* Allocate the memory object.
1096	*/
1097	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
1098	if (!pMemLnx)
1099	return VERR_NO_MEMORY;
1100
1101	/*
1102	* Gather the pages.
1103	* We ASSUME all kernel pages are non-swappable and non-movable.
1104	*/
1105	rc = VINF_SUCCESS;
1106	pbPage = (uint8_t *)pvLast;
1107	iPage = cPages;
1108	if (!fLinearMapping)
1109	{
1110	while (iPage-- > 0)
1111	{
1112	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1113	if (RT_UNLIKELY(!pPage))
1114	{
1115	rc = VERR_LOCK_FAILED;
1116	break;
1117	}
1118	pMemLnx->apPages[iPage] = pPage;
1119	pbPage -= PAGE_SIZE;
1120	}
1121	}
1122	else
1123	{
1124	while (iPage-- > 0)
1125	{
1126	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1127	pbPage -= PAGE_SIZE;
1128	}
1129	}
1130	if (RT_SUCCESS(rc))
1131	{
1132	/*
1133	* Complete the memory object and return.
1134	*/
1135	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1136	pMemLnx->cPages = cPages;
1137	Assert(!pMemLnx->fMappedToRing0);
1138	*ppMem = &pMemLnx->Core;
1139
1140	return VINF_SUCCESS;
1141	}
1142
1143	rtR0MemObjDelete(&pMemLnx->Core);
1144	return rc;
1145	}
1146
1147
1148	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
1149	{
1150	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1151	const size_t cPages = cb >> PAGE_SHIFT;
1152	struct page *pDummyPage;
1153	struct page **papPages;
1154
1155	/* check for unsupported stuff. */
1156	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1157	if (uAlignment > PAGE_SIZE)
1158	return VERR_NOT_SUPPORTED;
1159
1160	/*
1161	* Allocate a dummy page and create a page pointer array for vmap such that
1162	* the dummy page is mapped all over the reserved area.
1163	*/
1164	pDummyPage = alloc_page(GFP_HIGHUSER);
1165	if (!pDummyPage)
1166	return VERR_NO_MEMORY;
1167	papPages = RTMemAlloc(sizeof(papPages) cPages);
1168	if (papPages)
1169	{
1170	void *pv;
1171	size_t iPage = cPages;
1172	while (iPage-- > 0)
1173	papPages[iPage] = pDummyPage;
1174	# ifdef VM_MAP
1175	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1176	# else
1177	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1178	# endif
1179	RTMemFree(papPages);
1180	if (pv)
1181	{
1182	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1183	if (pMemLnx)
1184	{
1185	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1186	pMemLnx->cPages = 1;
1187	pMemLnx->apPages[0] = pDummyPage;
1188	*ppMem = &pMemLnx->Core;
1189	return VINF_SUCCESS;
1190	}
1191	vunmap(pv);
1192	}
1193	}
1194	__free_page(pDummyPage);
1195	return VERR_NO_MEMORY;
1196
1197	#else /* < 2.4.22 */
1198	/*
1199	* Could probably use ioremap here, but the caller is in a better position than us
1200	* to select some safe physical memory.
1201	*/
1202	return VERR_NOT_SUPPORTED;
1203	#endif
1204	}
1205
1206
1207	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1208	{
1209	PRTR0MEMOBJLNX pMemLnx;
1210	void *pv;
1211	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1212	if (!pTask)
1213	return VERR_NOT_SUPPORTED;
1214
1215	/*
1216	* Check that the specified alignment is supported.
1217	*/
1218	if (uAlignment > PAGE_SIZE)
1219	return VERR_NOT_SUPPORTED;
1220
1221	/*
1222	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1223	*/
1224	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1225	if (pv == (void *)-1)
1226	return VERR_NO_MEMORY;
1227
1228	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1229	if (!pMemLnx)
1230	{
1231	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1232	return VERR_NO_MEMORY;
1233	}
1234
1235	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1236	*ppMem = &pMemLnx->Core;
1237	return VINF_SUCCESS;
1238	}
1239
1240
1241	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap,
1242	void *pvFixed, size_t uAlignment,
1243	unsigned fProt, size_t offSub, size_t cbSub)
1244	{
1245	int rc = VERR_NO_MEMORY;
1246	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1247	PRTR0MEMOBJLNX pMemLnx;
1248
1249	/* Fail if requested to do something we can't. */
1250	AssertMsgReturn(!offSub && !cbSub, ("%#x %#x\n", offSub, cbSub), VERR_NOT_SUPPORTED);
1251	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1252	if (uAlignment > PAGE_SIZE)
1253	return VERR_NOT_SUPPORTED;
1254
1255	/*
1256	* Create the IPRT memory object.
1257	*/
1258	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1259	if (pMemLnx)
1260	{
1261	if (pMemLnxToMap->cPages)
1262	{
1263	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1264	/*
1265	* Use vmap - 2.4.22 and later.
1266	*/
1267	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1268	# ifdef VM_MAP
1269	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
1270	# else
1271	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
1272	# endif
1273	if (pMemLnx->Core.pv)
1274	{
1275	pMemLnx->fMappedToRing0 = true;
1276	rc = VINF_SUCCESS;
1277	}
1278	else
1279	rc = VERR_MAP_FAILED;
1280
1281	#else /* < 2.4.22 */
1282	/*
1283	* Only option here is to share mappings if possible and forget about fProt.
1284	*/
1285	if (rtR0MemObjIsRing3(pMemToMap))
1286	rc = VERR_NOT_SUPPORTED;
1287	else
1288	{
1289	rc = VINF_SUCCESS;
1290	if (!pMemLnxToMap->Core.pv)
1291	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1292	if (RT_SUCCESS(rc))
1293	{
1294	Assert(pMemLnxToMap->Core.pv);
1295	pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
1296	}
1297	}
1298	#endif
1299	}
1300	else
1301	{
1302	/*
1303	* MMIO / physical memory.
1304	*/
1305	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1306	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1307	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb)
1308	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
1309	if (pMemLnx->Core.pv)
1310	{
1311	/** @todo fix protection. */
1312	rc = VINF_SUCCESS;
1313	}
1314	}
1315	if (RT_SUCCESS(rc))
1316	{
1317	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1318	*ppMem = &pMemLnx->Core;
1319	return VINF_SUCCESS;
1320	}
1321	rtR0MemObjDelete(&pMemLnx->Core);
1322	}
1323
1324	return rc;
1325	}
1326
1327
1328	#ifdef VBOX_USE_PAE_HACK
1329	/**
1330	* Replace the PFN of a PTE with the address of the actual page.
1331	*
1332	* The caller maps a reserved dummy page at the address with the desired access
1333	* and flags.
1334	*
1335	* This hack is required for older Linux kernels which don't provide
1336	* remap_pfn_range().
1337	*
1338	* @returns 0 on success, -ENOMEM on failure.
1339	* @param mm The memory context.
1340	* @param ulAddr The mapping address.
1341	* @param Phys The physical address of the page to map.
1342	*/
1343	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1344	{
1345	int rc = -ENOMEM;
1346	pgd_t *pgd;
1347
1348	spin_lock(&mm->page_table_lock);
1349
1350	pgd = pgd_offset(mm, ulAddr);
1351	if (!pgd_none(pgd) && !pgd_bad(pgd))
1352	{
1353	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1354	if (!pmd_none(*pmd))
1355	{
1356	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1357	if (ptep)
1358	{
1359	pte_t pte = *ptep;
1360	pte.pte_high &= 0xfff00000;
1361	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1362	pte.pte_low &= 0x00000fff;
1363	pte.pte_low \|= (Phys & 0xfffff000);
1364	set_pte(ptep, pte);
1365	pte_unmap(ptep);
1366	rc = 0;
1367	}
1368	}
1369	}
1370
1371	spin_unlock(&mm->page_table_lock);
1372	return rc;
1373	}
1374	#endif /* VBOX_USE_PAE_HACK */
1375
1376
1377	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed,
1378	size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1379	{
1380	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1381	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1382	int rc = VERR_NO_MEMORY;
1383	PRTR0MEMOBJLNX pMemLnx;
1384	#ifdef VBOX_USE_PAE_HACK
1385	struct page *pDummyPage;
1386	RTHCPHYS DummyPhys;
1387	#endif
1388
1389	/*
1390	* Check for restrictions.
1391	*/
1392	if (!pTask)
1393	return VERR_NOT_SUPPORTED;
1394	if (uAlignment > PAGE_SIZE)
1395	return VERR_NOT_SUPPORTED;
1396
1397	#ifdef VBOX_USE_PAE_HACK
1398	/*
1399	* Allocate a dummy page for use when mapping the memory.
1400	*/
1401	pDummyPage = alloc_page(GFP_USER);
1402	if (!pDummyPage)
1403	return VERR_NO_MEMORY;
1404	SetPageReserved(pDummyPage);
1405	DummyPhys = page_to_phys(pDummyPage);
1406	#endif
1407
1408	/*
1409	* Create the IPRT memory object.
1410	*/
1411	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1412	if (pMemLnx)
1413	{
1414	/*
1415	* Allocate user space mapping.
1416	*/
1417	void *pv;
1418	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1419	if (pv != (void *)-1)
1420	{
1421	/*
1422	* Map page by page into the mmap area.
1423	* This is generic, paranoid and not very efficient.
1424	*/
1425	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1426	unsigned long ulAddrCur = (unsigned long)pv;
1427	const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1428	size_t iPage;
1429
1430	down_write(&pTask->mm->mmap_sem);
1431
1432	rc = VINF_SUCCESS;
1433	if (pMemLnxToMap->cPages)
1434	{
1435	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1436	{
1437	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)
1438	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1439	#endif
1440	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1441	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1442	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1443	#endif
1444	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1445	/* remap_page_range() limitation on x86 */
1446	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1447	#endif
1448
1449	#if defined(VBOX_USE_INSERT_PAGE) && LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
1450	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1451	vma->vm_flags \|= VM_RESERVED; /* This flag helps making 100% sure some bad stuff wont happen (swap, core, ++). */
1452	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1453	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1454	#elif defined(VBOX_USE_PAE_HACK)
1455	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1456	if (!rc)
1457	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1458	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1459	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1460	#else /* 2.4 */
1461	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1462	#endif
1463	if (rc)
1464	{
1465	rc = VERR_NO_MEMORY;
1466	break;
1467	}
1468	}
1469	}
1470	else
1471	{
1472	RTHCPHYS Phys;
1473	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1474	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1475	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1476	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1477	else
1478	{
1479	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1480	Phys = NIL_RTHCPHYS;
1481	}
1482	if (Phys != NIL_RTHCPHYS)
1483	{
1484	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1485	{
1486	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1487	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1488	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1489	#endif
1490	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1491	/* remap_page_range() limitation on x86 */
1492	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1493	#endif
1494
1495	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1496	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1497	#elif defined(VBOX_USE_PAE_HACK)
1498	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1499	if (!rc)
1500	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1501	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1502	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1503	#else /* 2.4 */
1504	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1505	#endif
1506	if (rc)
1507	{
1508	rc = VERR_NO_MEMORY;
1509	break;
1510	}
1511	}
1512	}
1513	}
1514
1515	up_write(&pTask->mm->mmap_sem);
1516
1517	if (RT_SUCCESS(rc))
1518	{
1519	#ifdef VBOX_USE_PAE_HACK
1520	__free_page(pDummyPage);
1521	#endif
1522	pMemLnx->Core.pv = pv;
1523	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1524	*ppMem = &pMemLnx->Core;
1525	return VINF_SUCCESS;
1526	}
1527
1528	/*
1529	* Bail out.
1530	*/
1531	rtR0MemObjLinuxDoMunmap(pv, pMemLnxToMap->Core.cb, pTask);
1532	}
1533	rtR0MemObjDelete(&pMemLnx->Core);
1534	}
1535	#ifdef VBOX_USE_PAE_HACK
1536	__free_page(pDummyPage);
1537	#endif
1538
1539	return rc;
1540	}
1541
1542
1543	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1544	{
1545	NOREF(pMem);
1546	NOREF(offSub);
1547	NOREF(cbSub);
1548	NOREF(fProt);
1549	return VERR_NOT_SUPPORTED;
1550	}
1551
1552
1553	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1554	{
1555	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1556
1557	if (pMemLnx->cPages)
1558	return page_to_phys(pMemLnx->apPages[iPage]);
1559
1560	switch (pMemLnx->Core.enmType)
1561	{
1562	case RTR0MEMOBJTYPE_CONT:
1563	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1564
1565	case RTR0MEMOBJTYPE_PHYS:
1566	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1567
1568	/* the parent knows */
1569	case RTR0MEMOBJTYPE_MAPPING:
1570	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1571
1572	/* cPages > 0 */
1573	case RTR0MEMOBJTYPE_LOW:
1574	case RTR0MEMOBJTYPE_LOCK:
1575	case RTR0MEMOBJTYPE_PHYS_NC:
1576	case RTR0MEMOBJTYPE_PAGE:
1577	default:
1578	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1579	/* fall thru */
1580
1581	case RTR0MEMOBJTYPE_RES_VIRT:
1582	return NIL_RTHCPHYS;
1583	}
1584	}
1585

Download in other formats:

Original Format