memobj-r0drv-linux.c@ 41116

Last change on this file since 41116 was 41116, checked in by vboxsync, 13 years ago
rtR0MemObjNativeLockUser: Force write access to COW pages.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev`
File size: 51.0 KB

Line
1	/* $Revision: 41116 $ */
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	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
104	}
105
106
107	/**
108	* Compute order. Some functions allocate 2^order pages.
109	*
110	* @returns order.
111	* @param cPages Number of pages.
112	*/
113	static int rtR0MemObjLinuxOrder(size_t cPages)
114	{
115	int iOrder;
116	size_t cTmp;
117
118	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
119	;
120	if (cPages & ~((size_t)1 << iOrder))
121	++iOrder;
122
123	return iOrder;
124	}
125
126
127	/**
128	* Converts from RTMEM_PROT_* to Linux PAGE_*.
129	*
130	* @returns Linux page protection constant.
131	* @param fProt The IPRT protection mask.
132	* @param fKernel Whether it applies to kernel or user space.
133	*/
134	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
135	{
136	switch (fProt)
137	{
138	default:
139	AssertMsgFailed(("%#x %d\n", fProt, fKernel));
140	case RTMEM_PROT_NONE:
141	return PAGE_NONE;
142
143	case RTMEM_PROT_READ:
144	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
145
146	case RTMEM_PROT_WRITE:
147	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
148	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
149
150	case RTMEM_PROT_EXEC:
151	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
152	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
153	if (fKernel)
154	{
155	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
156	pgprot_val(fPg) &= ~_PAGE_RW;
157	return fPg;
158	}
159	return PAGE_READONLY_EXEC;
160	#else
161	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
162	#endif
163
164	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
165	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
166	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
167	}
168	}
169
170
171	/**
172	* Internal worker that allocates physical pages and creates the memory object for them.
173	*
174	* @returns IPRT status code.
175	* @param ppMemLnx Where to store the memory object pointer.
176	* @param enmType The object type.
177	* @param cb The number of bytes to allocate.
178	* @param uAlignment The alignment of the physical memory.
179	* Only valid if fContiguous == true, ignored otherwise.
180	* @param fFlagsLnx The page allocation flags (GPFs).
181	* @param fContiguous Whether the allocation must be contiguous.
182	* @param rcNoMem What to return when we're out of pages.
183	*/
184	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
185	size_t uAlignment, unsigned fFlagsLnx, bool fContiguous, int rcNoMem)
186	{
187	size_t iPage;
188	size_t const cPages = cb >> PAGE_SHIFT;
189	struct page *paPages;
190
191	/*
192	* Allocate a memory object structure that's large enough to contain
193	* the page pointer array.
194	*/
195	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
196	if (!pMemLnx)
197	return VERR_NO_MEMORY;
198	pMemLnx->cPages = cPages;
199
200	if (cPages > 255)
201	{
202	# ifdef __GFP_REPEAT
203	/* Try hard to allocate the memory, but the allocation attempt might fail. */
204	fFlagsLnx \|= __GFP_REPEAT;
205	# endif
206	# ifdef __GFP_NOMEMALLOC
207	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
208	fFlagsLnx \|= __GFP_NOMEMALLOC;
209	# endif
210	}
211
212	/*
213	* Allocate the pages.
214	* For small allocations we'll try contiguous first and then fall back on page by page.
215	*/
216	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
217	if ( fContiguous
218	\|\| cb <= PAGE_SIZE * 2)
219	{
220	# ifdef VBOX_USE_INSERT_PAGE
221	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP, rtR0MemObjLinuxOrder(cPages));
222	# else
223	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
224	# endif
225	if (paPages)
226	{
227	fContiguous = true;
228	for (iPage = 0; iPage < cPages; iPage++)
229	pMemLnx->apPages[iPage] = &paPages[iPage];
230	}
231	else if (fContiguous)
232	{
233	rtR0MemObjDelete(&pMemLnx->Core);
234	return rcNoMem;
235	}
236	}
237
238	if (!fContiguous)
239	{
240	for (iPage = 0; iPage < cPages; iPage++)
241	{
242	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx);
243	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
244	{
245	while (iPage-- > 0)
246	__free_page(pMemLnx->apPages[iPage]);
247	rtR0MemObjDelete(&pMemLnx->Core);
248	return rcNoMem;
249	}
250	}
251	}
252
253	#else /* < 2.4.22 */
254	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
255	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
256	if (!paPages)
257	{
258	rtR0MemObjDelete(&pMemLnx->Core);
259	return rcNoMem;
260	}
261	for (iPage = 0; iPage < cPages; iPage++)
262	{
263	pMemLnx->apPages[iPage] = &paPages[iPage];
264	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
265	if (PageHighMem(pMemLnx->apPages[iPage]))
266	BUG();
267	}
268
269	fContiguous = true;
270	#endif /* < 2.4.22 */
271	pMemLnx->fContiguous = fContiguous;
272
273	/*
274	* Reserve the pages.
275	*/
276	for (iPage = 0; iPage < cPages; iPage++)
277	SetPageReserved(pMemLnx->apPages[iPage]);
278
279	/*
280	* Note that the physical address of memory allocated with alloc_pages(flags, order)
281	* is always 2^(PAGE_SHIFT+order)-aligned.
282	*/
283	if ( fContiguous
284	&& uAlignment > PAGE_SIZE)
285	{
286	/*
287	* Check for alignment constraints. The physical address of memory allocated with
288	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
289	*/
290	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
291	{
292	/*
293	* This should never happen!
294	*/
295	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
296	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
297	rtR0MemObjLinuxFreePages(pMemLnx);
298	return rcNoMem;
299	}
300	}
301
302	*ppMemLnx = pMemLnx;
303	return VINF_SUCCESS;
304	}
305
306
307	/**
308	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
309	*
310	* This method does NOT free the object.
311	*
312	* @param pMemLnx The object which physical pages should be freed.
313	*/
314	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
315	{
316	size_t iPage = pMemLnx->cPages;
317	if (iPage > 0)
318	{
319	/*
320	* Restore the page flags.
321	*/
322	while (iPage-- > 0)
323	{
324	ClearPageReserved(pMemLnx->apPages[iPage]);
325	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
326	#else
327	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
328	#endif
329	}
330
331	/*
332	* Free the pages.
333	*/
334	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
335	if (!pMemLnx->fContiguous)
336	{
337	iPage = pMemLnx->cPages;
338	while (iPage-- > 0)
339	__free_page(pMemLnx->apPages[iPage]);
340	}
341	else
342	#endif
343	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
344
345	pMemLnx->cPages = 0;
346	}
347	}
348
349
350	/**
351	* Maps the allocation into ring-0.
352	*
353	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
354	*
355	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
356	* space, so we'll use that mapping if possible. If execute access is required, we'll
357	* play safe and do our own mapping.
358	*
359	* @returns IPRT status code.
360	* @param pMemLnx The linux memory object to map.
361	* @param fExecutable Whether execute access is required.
362	*/
363	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
364	{
365	int rc = VINF_SUCCESS;
366
367	/*
368	* Choose mapping strategy.
369	*/
370	bool fMustMap = fExecutable
371	\|\| !pMemLnx->fContiguous;
372	if (!fMustMap)
373	{
374	size_t iPage = pMemLnx->cPages;
375	while (iPage-- > 0)
376	if (PageHighMem(pMemLnx->apPages[iPage]))
377	{
378	fMustMap = true;
379	break;
380	}
381	}
382
383	Assert(!pMemLnx->Core.pv);
384	Assert(!pMemLnx->fMappedToRing0);
385
386	if (fMustMap)
387	{
388	/*
389	* Use vmap - 2.4.22 and later.
390	*/
391	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
392	pgprot_t fPg;
393	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
394	# ifdef _PAGE_NX
395	if (!fExecutable)
396	pgprot_val(fPg) \|= _PAGE_NX;
397	# endif
398
399	# ifdef VM_MAP
400	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
401	# else
402	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
403	# endif
404	if (pMemLnx->Core.pv)
405	pMemLnx->fMappedToRing0 = true;
406	else
407	rc = VERR_MAP_FAILED;
408	#else /* < 2.4.22 */
409	rc = VERR_NOT_SUPPORTED;
410	#endif
411	}
412	else
413	{
414	/*
415	* Use the kernel RAM mapping.
416	*/
417	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
418	Assert(pMemLnx->Core.pv);
419	}
420
421	return rc;
422	}
423
424
425	/**
426	* Undos what rtR0MemObjLinuxVMap() did.
427	*
428	* @param pMemLnx The linux memory object.
429	*/
430	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
431	{
432	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
433	if (pMemLnx->fMappedToRing0)
434	{
435	Assert(pMemLnx->Core.pv);
436	vunmap(pMemLnx->Core.pv);
437	pMemLnx->fMappedToRing0 = false;
438	}
439	#else /* < 2.4.22 */
440	Assert(!pMemLnx->fMappedToRing0);
441	#endif
442	pMemLnx->Core.pv = NULL;
443	}
444
445
446	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
447	{
448	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
449
450	/*
451	* Release any memory that we've allocated or locked.
452	*/
453	switch (pMemLnx->Core.enmType)
454	{
455	case RTR0MEMOBJTYPE_LOW:
456	case RTR0MEMOBJTYPE_PAGE:
457	case RTR0MEMOBJTYPE_CONT:
458	case RTR0MEMOBJTYPE_PHYS:
459	case RTR0MEMOBJTYPE_PHYS_NC:
460	rtR0MemObjLinuxVUnmap(pMemLnx);
461	rtR0MemObjLinuxFreePages(pMemLnx);
462	break;
463
464	case RTR0MEMOBJTYPE_LOCK:
465	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
466	{
467	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
468	size_t iPage;
469	Assert(pTask);
470	if (pTask && pTask->mm)
471	down_read(&pTask->mm->mmap_sem);
472
473	iPage = pMemLnx->cPages;
474	while (iPage-- > 0)
475	{
476	if (!PageReserved(pMemLnx->apPages[iPage]))
477	SetPageDirty(pMemLnx->apPages[iPage]);
478	page_cache_release(pMemLnx->apPages[iPage]);
479	}
480
481	if (pTask && pTask->mm)
482	up_read(&pTask->mm->mmap_sem);
483	}
484	/* else: kernel memory - nothing to do here. */
485	break;
486
487	case RTR0MEMOBJTYPE_RES_VIRT:
488	Assert(pMemLnx->Core.pv);
489	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
490	{
491	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
492	Assert(pTask);
493	if (pTask && pTask->mm)
494	{
495	down_write(&pTask->mm->mmap_sem);
496	MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
497	up_write(&pTask->mm->mmap_sem);
498	}
499	}
500	else
501	{
502	vunmap(pMemLnx->Core.pv);
503
504	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
505	__free_page(pMemLnx->apPages[0]);
506	pMemLnx->apPages[0] = NULL;
507	pMemLnx->cPages = 0;
508	}
509	pMemLnx->Core.pv = NULL;
510	break;
511
512	case RTR0MEMOBJTYPE_MAPPING:
513	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
514	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
515	{
516	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
517	Assert(pTask);
518	if (pTask && pTask->mm)
519	{
520	down_write(&pTask->mm->mmap_sem);
521	MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
522	up_write(&pTask->mm->mmap_sem);
523	}
524	}
525	else
526	vunmap(pMemLnx->Core.pv);
527	pMemLnx->Core.pv = NULL;
528	break;
529
530	default:
531	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
532	return VERR_INTERNAL_ERROR;
533	}
534	return VINF_SUCCESS;
535	}
536
537
538	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
539	{
540	PRTR0MEMOBJLNX pMemLnx;
541	int rc;
542
543	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
544	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
545	false /* non-contiguous */, VERR_NO_MEMORY);
546	#else
547	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
548	false /* non-contiguous */, VERR_NO_MEMORY);
549	#endif
550	if (RT_SUCCESS(rc))
551	{
552	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
553	if (RT_SUCCESS(rc))
554	{
555	*ppMem = &pMemLnx->Core;
556	return rc;
557	}
558
559	rtR0MemObjLinuxFreePages(pMemLnx);
560	rtR0MemObjDelete(&pMemLnx->Core);
561	}
562
563	return rc;
564	}
565
566
567	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
568	{
569	PRTR0MEMOBJLNX pMemLnx;
570	int rc;
571
572	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
573	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
574	/* ZONE_DMA32: 0-4GB */
575	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
576	false /* non-contiguous */, VERR_NO_LOW_MEMORY);
577	if (RT_FAILURE(rc))
578	#endif
579	#ifdef RT_ARCH_AMD64
580	/* ZONE_DMA: 0-16MB */
581	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
582	false /* non-contiguous */, VERR_NO_LOW_MEMORY);
583	#else
584	# ifdef CONFIG_X86_PAE
585	# endif
586	/* ZONE_NORMAL: 0-896MB */
587	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
588	false /* non-contiguous */, VERR_NO_LOW_MEMORY);
589	#endif
590	if (RT_SUCCESS(rc))
591	{
592	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
593	if (RT_SUCCESS(rc))
594	{
595	*ppMem = &pMemLnx->Core;
596	return rc;
597	}
598
599	rtR0MemObjLinuxFreePages(pMemLnx);
600	rtR0MemObjDelete(&pMemLnx->Core);
601	}
602
603	return rc;
604	}
605
606
607	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
608	{
609	PRTR0MEMOBJLNX pMemLnx;
610	int rc;
611
612	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
613	/* ZONE_DMA32: 0-4GB */
614	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
615	true /* contiguous */, VERR_NO_CONT_MEMORY);
616	if (RT_FAILURE(rc))
617	#endif
618	#ifdef RT_ARCH_AMD64
619	/* ZONE_DMA: 0-16MB */
620	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
621	true /* contiguous */, VERR_NO_CONT_MEMORY);
622	#else
623	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
624	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
625	true /* contiguous */, VERR_NO_CONT_MEMORY);
626	#endif
627	if (RT_SUCCESS(rc))
628	{
629	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
630	if (RT_SUCCESS(rc))
631	{
632	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
633	size_t iPage = pMemLnx->cPages;
634	while (iPage-- > 0)
635	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
636	#endif
637	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
638	*ppMem = &pMemLnx->Core;
639	return rc;
640	}
641
642	rtR0MemObjLinuxFreePages(pMemLnx);
643	rtR0MemObjDelete(&pMemLnx->Core);
644	}
645
646	return rc;
647	}
648
649
650	/**
651	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
652	*
653	* @returns IPRT status.
654	* @param ppMemLnx Where to
655	* @param enmType The object type.
656	* @param cb The size of the allocation.
657	* @param uAlignment The alignment of the physical memory.
658	* Only valid for fContiguous == true, ignored otherwise.
659	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
660	* @param fGfp The Linux GFP flags to use for the allocation.
661	*/
662	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
663	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, unsigned fGfp)
664	{
665	PRTR0MEMOBJLNX pMemLnx;
666	int rc;
667
668	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
669	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
670	VERR_NO_PHYS_MEMORY);
671	if (RT_FAILURE(rc))
672	return rc;
673
674	/*
675	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
676	*/
677	if (PhysHighest != NIL_RTHCPHYS)
678	{
679	size_t iPage = pMemLnx->cPages;
680	while (iPage-- > 0)
681	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
682	{
683	rtR0MemObjLinuxFreePages(pMemLnx);
684	rtR0MemObjDelete(&pMemLnx->Core);
685	return VERR_NO_MEMORY;
686	}
687	}
688
689	/*
690	* Complete the object.
691	*/
692	if (enmType == RTR0MEMOBJTYPE_PHYS)
693	{
694	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
695	pMemLnx->Core.u.Phys.fAllocated = true;
696	}
697	*ppMem = &pMemLnx->Core;
698	return rc;
699	}
700
701
702	/**
703	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
704	*
705	* @returns IPRT status.
706	* @param ppMem Where to store the memory object pointer on success.
707	* @param enmType The object type.
708	* @param cb The size of the allocation.
709	* @param uAlignment The alignment of the physical memory.
710	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
711	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
712	*/
713	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
714	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
715	{
716	int rc;
717
718	/*
719	* There are two clear cases and that's the <=16MB and anything-goes ones.
720	* When the physical address limit is somewhere in-between those two we'll
721	* just have to try, starting with HIGHUSER and working our way thru the
722	* different types, hoping we'll get lucky.
723	*
724	* We should probably move this physical address restriction logic up to
725	* the page alloc function as it would be more efficient there. But since
726	* we don't expect this to be a performance issue just yet it can wait.
727	*/
728	if (PhysHighest == NIL_RTHCPHYS)
729	/* ZONE_HIGHMEM: the whole physical memory */
730	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
731	else if (PhysHighest <= _1M * 16)
732	/* ZONE_DMA: 0-16MB */
733	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
734	else
735	{
736	rc = VERR_NO_MEMORY;
737	if (RT_FAILURE(rc))
738	/* ZONE_HIGHMEM: the whole physical memory */
739	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
740	if (RT_FAILURE(rc))
741	/* ZONE_NORMAL: 0-896MB */
742	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
743	#ifdef GFP_DMA32
744	if (RT_FAILURE(rc))
745	/* ZONE_DMA32: 0-4GB */
746	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
747	#endif
748	if (RT_FAILURE(rc))
749	/* ZONE_DMA: 0-16MB */
750	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
751	}
752	return rc;
753	}
754
755
756	/**
757	* Translates a kernel virtual address to a linux page structure by walking the
758	* page tables.
759	*
760	* @note We do assume that the page tables will not change as we are walking
761	* them. This assumption is rather forced by the fact that I could not
762	* immediately see any way of preventing this from happening. So, we
763	* take some extra care when accessing them.
764	*
765	* Because of this, we don't want to use this function on memory where
766	* attribute changes to nearby pages is likely to cause large pages to
767	* be used or split up. So, don't use this for the linear mapping of
768	* physical memory.
769	*
770	* @returns Pointer to the page structur or NULL if it could not be found.
771	* @param pv The kernel virtual address.
772	*/
773	static struct page rtR0MemObjLinuxVirtToPage(void pv)
774	{
775	unsigned long ulAddr = (unsigned long)pv;
776	unsigned long pfn;
777	struct page *pPage;
778	pte_t *pEntry;
779	union
780	{
781	pgd_t Global;
782	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
783	pud_t Upper;
784	#endif
785	pmd_t Middle;
786	pte_t Entry;
787	} u;
788
789	/* Should this happen in a situation this code will be called in? And if
790	* so, can it change under our feet? See also
791	* "Documentation/vm/active_mm.txt" in the kernel sources. */
792	if (RT_UNLIKELY(!current->active_mm))
793	return NULL;
794	u.Global = *pgd_offset(current->active_mm, ulAddr);
795	if (RT_UNLIKELY(pgd_none(u.Global)))
796	return NULL;
797
798	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
799	u.Upper = *pud_offset(&u.Global, ulAddr);
800	if (RT_UNLIKELY(pud_none(u.Upper)))
801	return NULL;
802	# if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 25)
803	if (pud_large(u.Upper))
804	{
805	pPage = pud_page(u.Upper);
806	AssertReturn(pPage, NULL);
807	pfn = page_to_pfn(pPage); /* doing the safe way... */
808	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
809	return pfn_to_page(pfn);
810	}
811	# endif
812
813	u.Middle = *pmd_offset(&u.Upper, ulAddr);
814	#else /* < 2.6.11 */
815	u.Middle = *pmd_offset(&u.Global, ulAddr);
816	#endif /* < 2.6.11 */
817	if (RT_UNLIKELY(pmd_none(u.Middle)))
818	return NULL;
819	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0)
820	if (pmd_large(u.Middle))
821	{
822	pPage = pmd_page(u.Middle);
823	AssertReturn(pPage, NULL);
824	pfn = page_to_pfn(pPage); /* doing the safe way... */
825	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
826	return pfn_to_page(pfn);
827	}
828	#endif
829
830	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
831	pEntry = pte_offset_map(&u.Middle, ulAddr);
832	#else
833	pEntry = pte_offset(&u.Middle, ulAddr);
834	#endif
835	if (RT_UNLIKELY(!pEntry))
836	return NULL;
837	u.Entry = *pEntry;
838	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) \|\| defined(pte_offset_map)
839	pte_unmap(pEntry);
840	#endif
841
842	if (RT_UNLIKELY(!pte_present(u.Entry)))
843	return NULL;
844	return pte_page(u.Entry);
845	}
846
847
848	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
849	{
850	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
851	}
852
853
854	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
855	{
856	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
857	}
858
859
860	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
861	{
862	/*
863	* All we need to do here is to validate that we can use
864	* ioremap on the specified address (32/64-bit dma_addr_t).
865	*/
866	PRTR0MEMOBJLNX pMemLnx;
867	dma_addr_t PhysAddr = Phys;
868	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
869
870	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
871	if (!pMemLnx)
872	return VERR_NO_MEMORY;
873
874	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
875	pMemLnx->Core.u.Phys.fAllocated = false;
876	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
877	Assert(!pMemLnx->cPages);
878	*ppMem = &pMemLnx->Core;
879	return VINF_SUCCESS;
880	}
881
882
883	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
884	{
885	const int cPages = cb >> PAGE_SHIFT;
886	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
887	struct vm_area_struct **papVMAs;
888	PRTR0MEMOBJLNX pMemLnx;
889	int rc = VERR_NO_MEMORY;
890	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
891
892	/*
893	* Check for valid task and size overflows.
894	*/
895	if (!pTask)
896	return VERR_NOT_SUPPORTED;
897	if (((size_t)cPages << PAGE_SHIFT) != cb)
898	return VERR_OUT_OF_RANGE;
899
900	/*
901	* Allocate the memory object and a temporary buffer for the VMAs.
902	*/
903	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
904	if (!pMemLnx)
905	return VERR_NO_MEMORY;
906
907	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
908	if (papVMAs)
909	{
910	down_read(&pTask->mm->mmap_sem);
911
912	/*
913	* Get user pages.
914	*/
915	rc = get_user_pages(pTask, /* Task for fault accounting. */
916	pTask->mm, /* Whose pages. */
917	R3Ptr, /* Where from. */
918	cPages, /* How many pages. */
919	fWrite, /* Write to memory. */
920	fWrite, /* force write access. */
921	&pMemLnx->apPages[0], /* Page array. */
922	papVMAs); /* vmas */
923	if (rc == cPages)
924	{
925	/*
926	* Flush dcache (required?), protect against fork and _really_ pin the page
927	* table entries. get_user_pages() will protect against swapping out the
928	* pages but it will NOT protect against removing page table entries. This
929	* can be achieved with
930	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
931	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
932	* Usual Linux distributions support only a limited size of locked pages
933	* (e.g. 32KB).
934	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
935	* or by
936	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
937	* a range check.
938	*/
939	/** @todo The Linux fork() protection will require more work if this API
940	* is to be used for anything but locking VM pages. */
941	while (rc-- > 0)
942	{
943	flush_dcache_page(pMemLnx->apPages[rc]);
944	papVMAs[rc]->vm_flags \|= (VM_DONTCOPY \| VM_LOCKED);
945	}
946
947	up_read(&pTask->mm->mmap_sem);
948
949	RTMemFree(papVMAs);
950
951	pMemLnx->Core.u.Lock.R0Process = R0Process;
952	pMemLnx->cPages = cPages;
953	Assert(!pMemLnx->fMappedToRing0);
954	*ppMem = &pMemLnx->Core;
955
956	return VINF_SUCCESS;
957	}
958
959	/*
960	* Failed - we need to unlock any pages that we succeeded to lock.
961	*/
962	while (rc-- > 0)
963	{
964	if (!PageReserved(pMemLnx->apPages[rc]))
965	SetPageDirty(pMemLnx->apPages[rc]);
966	page_cache_release(pMemLnx->apPages[rc]);
967	}
968
969	up_read(&pTask->mm->mmap_sem);
970
971	RTMemFree(papVMAs);
972	rc = VERR_LOCK_FAILED;
973	}
974
975	rtR0MemObjDelete(&pMemLnx->Core);
976	return rc;
977	}
978
979
980	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
981	{
982	void pvLast = (uint8_t )pv + cb - 1;
983	size_t const cPages = cb >> PAGE_SHIFT;
984	PRTR0MEMOBJLNX pMemLnx;
985	bool fLinearMapping;
986	int rc;
987	uint8_t *pbPage;
988	size_t iPage;
989	NOREF(fAccess);
990
991	if ( !RTR0MemKernelIsValidAddr(pv)
992	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
993	return VERR_INVALID_PARAMETER;
994
995	/*
996	* The lower part of the kernel memory has a linear mapping between
997	* physical and virtual addresses. So we take a short cut here. This is
998	* assumed to be the cleanest way to handle those addresses (and the code
999	* is well tested, though the test for determining it is not very nice).
1000	* If we ever decide it isn't we can still remove it.
1001	*/
1002	#if 0
1003	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1004	#else
1005	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1006	&& (unsigned long)pvLast < (unsigned long)high_memory;
1007	#endif
1008
1009	/*
1010	* Allocate the memory object.
1011	*/
1012	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
1013	if (!pMemLnx)
1014	return VERR_NO_MEMORY;
1015
1016	/*
1017	* Gather the pages.
1018	* We ASSUME all kernel pages are non-swappable and non-movable.
1019	*/
1020	rc = VINF_SUCCESS;
1021	pbPage = (uint8_t *)pvLast;
1022	iPage = cPages;
1023	if (!fLinearMapping)
1024	{
1025	while (iPage-- > 0)
1026	{
1027	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1028	if (RT_UNLIKELY(!pPage))
1029	{
1030	rc = VERR_LOCK_FAILED;
1031	break;
1032	}
1033	pMemLnx->apPages[iPage] = pPage;
1034	pbPage -= PAGE_SIZE;
1035	}
1036	}
1037	else
1038	{
1039	while (iPage-- > 0)
1040	{
1041	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1042	pbPage -= PAGE_SIZE;
1043	}
1044	}
1045	if (RT_SUCCESS(rc))
1046	{
1047	/*
1048	* Complete the memory object and return.
1049	*/
1050	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1051	pMemLnx->cPages = cPages;
1052	Assert(!pMemLnx->fMappedToRing0);
1053	*ppMem = &pMemLnx->Core;
1054
1055	return VINF_SUCCESS;
1056	}
1057
1058	rtR0MemObjDelete(&pMemLnx->Core);
1059	return rc;
1060	}
1061
1062
1063	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
1064	{
1065	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1066	const size_t cPages = cb >> PAGE_SHIFT;
1067	struct page *pDummyPage;
1068	struct page **papPages;
1069
1070	/* check for unsupported stuff. */
1071	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1072	if (uAlignment > PAGE_SIZE)
1073	return VERR_NOT_SUPPORTED;
1074
1075	/*
1076	* Allocate a dummy page and create a page pointer array for vmap such that
1077	* the dummy page is mapped all over the reserved area.
1078	*/
1079	pDummyPage = alloc_page(GFP_HIGHUSER);
1080	if (!pDummyPage)
1081	return VERR_NO_MEMORY;
1082	papPages = RTMemAlloc(sizeof(papPages) cPages);
1083	if (papPages)
1084	{
1085	void *pv;
1086	size_t iPage = cPages;
1087	while (iPage-- > 0)
1088	papPages[iPage] = pDummyPage;
1089	# ifdef VM_MAP
1090	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1091	# else
1092	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1093	# endif
1094	RTMemFree(papPages);
1095	if (pv)
1096	{
1097	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1098	if (pMemLnx)
1099	{
1100	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1101	pMemLnx->cPages = 1;
1102	pMemLnx->apPages[0] = pDummyPage;
1103	*ppMem = &pMemLnx->Core;
1104	return VINF_SUCCESS;
1105	}
1106	vunmap(pv);
1107	}
1108	}
1109	__free_page(pDummyPage);
1110	return VERR_NO_MEMORY;
1111
1112	#else /* < 2.4.22 */
1113	/*
1114	* Could probably use ioremap here, but the caller is in a better position than us
1115	* to select some safe physical memory.
1116	*/
1117	return VERR_NOT_SUPPORTED;
1118	#endif
1119	}
1120
1121
1122	/**
1123	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
1124	* an empty user space mapping.
1125	*
1126	* The caller takes care of acquiring the mmap_sem of the task.
1127	*
1128	* @returns Pointer to the mapping.
1129	* (void *)-1 on failure.
1130	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
1131	* @param cb The size of the mapping.
1132	* @param uAlignment The alignment of the mapping.
1133	* @param pTask The Linux task to create this mapping in.
1134	* @param fProt The RTMEM_PROT_* mask.
1135	*/
1136	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
1137	{
1138	unsigned fLnxProt;
1139	unsigned long ulAddr;
1140
1141	/*
1142	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
1143	*/
1144	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
1145	if (fProt == RTMEM_PROT_NONE)
1146	fLnxProt = PROT_NONE;
1147	else
1148	{
1149	fLnxProt = 0;
1150	if (fProt & RTMEM_PROT_READ)
1151	fLnxProt \|= PROT_READ;
1152	if (fProt & RTMEM_PROT_WRITE)
1153	fLnxProt \|= PROT_WRITE;
1154	if (fProt & RTMEM_PROT_EXEC)
1155	fLnxProt \|= PROT_EXEC;
1156	}
1157
1158	if (R3PtrFixed != (RTR3PTR)-1)
1159	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
1160	else
1161	{
1162	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
1163	if ( !(ulAddr & ~PAGE_MASK)
1164	&& (ulAddr & (uAlignment - 1)))
1165	{
1166	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
1167	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
1168	* ourselves) and further by there begin two mmap strategies (top / bottom). */
1169	/* For now, just ignore uAlignment requirements... */
1170	}
1171	}
1172	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
1173	return (void *)-1;
1174	return (void *)ulAddr;
1175	}
1176
1177
1178	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1179	{
1180	PRTR0MEMOBJLNX pMemLnx;
1181	void *pv;
1182	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1183	if (!pTask)
1184	return VERR_NOT_SUPPORTED;
1185
1186	/*
1187	* Check that the specified alignment is supported.
1188	*/
1189	if (uAlignment > PAGE_SIZE)
1190	return VERR_NOT_SUPPORTED;
1191
1192	/*
1193	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1194	*/
1195	down_write(&pTask->mm->mmap_sem);
1196	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1197	up_write(&pTask->mm->mmap_sem);
1198	if (pv == (void *)-1)
1199	return VERR_NO_MEMORY;
1200
1201	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1202	if (!pMemLnx)
1203	{
1204	down_write(&pTask->mm->mmap_sem);
1205	MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, cb);
1206	up_write(&pTask->mm->mmap_sem);
1207	return VERR_NO_MEMORY;
1208	}
1209
1210	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1211	*ppMem = &pMemLnx->Core;
1212	return VINF_SUCCESS;
1213	}
1214
1215
1216	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap,
1217	void *pvFixed, size_t uAlignment,
1218	unsigned fProt, size_t offSub, size_t cbSub)
1219	{
1220	int rc = VERR_NO_MEMORY;
1221	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1222	PRTR0MEMOBJLNX pMemLnx;
1223
1224	/* Fail if requested to do something we can't. */
1225	AssertMsgReturn(!offSub && !cbSub, ("%#x %#x\n", offSub, cbSub), VERR_NOT_SUPPORTED);
1226	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1227	if (uAlignment > PAGE_SIZE)
1228	return VERR_NOT_SUPPORTED;
1229
1230	/*
1231	* Create the IPRT memory object.
1232	*/
1233	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1234	if (pMemLnx)
1235	{
1236	if (pMemLnxToMap->cPages)
1237	{
1238	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1239	/*
1240	* Use vmap - 2.4.22 and later.
1241	*/
1242	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1243	# ifdef VM_MAP
1244	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
1245	# else
1246	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
1247	# endif
1248	if (pMemLnx->Core.pv)
1249	{
1250	pMemLnx->fMappedToRing0 = true;
1251	rc = VINF_SUCCESS;
1252	}
1253	else
1254	rc = VERR_MAP_FAILED;
1255
1256	#else /* < 2.4.22 */
1257	/*
1258	* Only option here is to share mappings if possible and forget about fProt.
1259	*/
1260	if (rtR0MemObjIsRing3(pMemToMap))
1261	rc = VERR_NOT_SUPPORTED;
1262	else
1263	{
1264	rc = VINF_SUCCESS;
1265	if (!pMemLnxToMap->Core.pv)
1266	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1267	if (RT_SUCCESS(rc))
1268	{
1269	Assert(pMemLnxToMap->Core.pv);
1270	pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
1271	}
1272	}
1273	#endif
1274	}
1275	else
1276	{
1277	/*
1278	* MMIO / physical memory.
1279	*/
1280	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1281	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1282	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb)
1283	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
1284	if (pMemLnx->Core.pv)
1285	{
1286	/** @todo fix protection. */
1287	rc = VINF_SUCCESS;
1288	}
1289	}
1290	if (RT_SUCCESS(rc))
1291	{
1292	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1293	*ppMem = &pMemLnx->Core;
1294	return VINF_SUCCESS;
1295	}
1296	rtR0MemObjDelete(&pMemLnx->Core);
1297	}
1298
1299	return rc;
1300	}
1301
1302
1303	#ifdef VBOX_USE_PAE_HACK
1304	/**
1305	* Replace the PFN of a PTE with the address of the actual page.
1306	*
1307	* The caller maps a reserved dummy page at the address with the desired access
1308	* and flags.
1309	*
1310	* This hack is required for older Linux kernels which don't provide
1311	* remap_pfn_range().
1312	*
1313	* @returns 0 on success, -ENOMEM on failure.
1314	* @param mm The memory context.
1315	* @param ulAddr The mapping address.
1316	* @param Phys The physical address of the page to map.
1317	*/
1318	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1319	{
1320	int rc = -ENOMEM;
1321	pgd_t *pgd;
1322
1323	spin_lock(&mm->page_table_lock);
1324
1325	pgd = pgd_offset(mm, ulAddr);
1326	if (!pgd_none(pgd) && !pgd_bad(pgd))
1327	{
1328	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1329	if (!pmd_none(*pmd))
1330	{
1331	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1332	if (ptep)
1333	{
1334	pte_t pte = *ptep;
1335	pte.pte_high &= 0xfff00000;
1336	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1337	pte.pte_low &= 0x00000fff;
1338	pte.pte_low \|= (Phys & 0xfffff000);
1339	set_pte(ptep, pte);
1340	pte_unmap(ptep);
1341	rc = 0;
1342	}
1343	}
1344	}
1345
1346	spin_unlock(&mm->page_table_lock);
1347	return rc;
1348	}
1349	#endif /* VBOX_USE_PAE_HACK */
1350
1351
1352	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed,
1353	size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1354	{
1355	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1356	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1357	int rc = VERR_NO_MEMORY;
1358	PRTR0MEMOBJLNX pMemLnx;
1359	#ifdef VBOX_USE_PAE_HACK
1360	struct page *pDummyPage;
1361	RTHCPHYS DummyPhys;
1362	#endif
1363
1364	/*
1365	* Check for restrictions.
1366	*/
1367	if (!pTask)
1368	return VERR_NOT_SUPPORTED;
1369	if (uAlignment > PAGE_SIZE)
1370	return VERR_NOT_SUPPORTED;
1371
1372	#ifdef VBOX_USE_PAE_HACK
1373	/*
1374	* Allocate a dummy page for use when mapping the memory.
1375	*/
1376	pDummyPage = alloc_page(GFP_USER);
1377	if (!pDummyPage)
1378	return VERR_NO_MEMORY;
1379	SetPageReserved(pDummyPage);
1380	DummyPhys = page_to_phys(pDummyPage);
1381	#endif
1382
1383	/*
1384	* Create the IPRT memory object.
1385	*/
1386	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1387	if (pMemLnx)
1388	{
1389	/*
1390	* Allocate user space mapping.
1391	*/
1392	void *pv;
1393	down_write(&pTask->mm->mmap_sem);
1394	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1395	if (pv != (void *)-1)
1396	{
1397	/*
1398	* Map page by page into the mmap area.
1399	* This is generic, paranoid and not very efficient.
1400	*/
1401	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1402	unsigned long ulAddrCur = (unsigned long)pv;
1403	const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1404	size_t iPage;
1405
1406	rc = 0;
1407	if (pMemLnxToMap->cPages)
1408	{
1409	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1410	{
1411	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)
1412	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1413	#endif
1414	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1415	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1416	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1417	#endif
1418	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1419	/* remap_page_range() limitation on x86 */
1420	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1421	#endif
1422
1423	#if defined(VBOX_USE_INSERT_PAGE) && LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
1424	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1425	vma->vm_flags \|= VM_RESERVED; /* This flag helps making 100% sure some bad stuff wont happen (swap, core, ++). */
1426	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1427	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1428	#elif defined(VBOX_USE_PAE_HACK)
1429	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1430	if (!rc)
1431	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1432	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1433	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1434	#else /* 2.4 */
1435	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1436	#endif
1437	if (rc)
1438	{
1439	rc = VERR_NO_MEMORY;
1440	break;
1441	}
1442	}
1443	}
1444	else
1445	{
1446	RTHCPHYS Phys;
1447	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1448	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1449	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1450	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1451	else
1452	{
1453	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1454	Phys = NIL_RTHCPHYS;
1455	}
1456	if (Phys != NIL_RTHCPHYS)
1457	{
1458	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1459	{
1460	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1461	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1462	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1463	#endif
1464	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1465	/* remap_page_range() limitation on x86 */
1466	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1467	#endif
1468
1469	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1470	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1471	#elif defined(VBOX_USE_PAE_HACK)
1472	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1473	if (!rc)
1474	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1475	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1476	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1477	#else /* 2.4 */
1478	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1479	#endif
1480	if (rc)
1481	{
1482	rc = VERR_NO_MEMORY;
1483	break;
1484	}
1485	}
1486	}
1487	}
1488	if (!rc)
1489	{
1490	up_write(&pTask->mm->mmap_sem);
1491	#ifdef VBOX_USE_PAE_HACK
1492	__free_page(pDummyPage);
1493	#endif
1494
1495	pMemLnx->Core.pv = pv;
1496	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1497	*ppMem = &pMemLnx->Core;
1498	return VINF_SUCCESS;
1499	}
1500
1501	/*
1502	* Bail out.
1503	*/
1504	MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, pMemLnxToMap->Core.cb);
1505	}
1506	up_write(&pTask->mm->mmap_sem);
1507	rtR0MemObjDelete(&pMemLnx->Core);
1508	}
1509	#ifdef VBOX_USE_PAE_HACK
1510	__free_page(pDummyPage);
1511	#endif
1512
1513	return rc;
1514	}
1515
1516
1517	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1518	{
1519	NOREF(pMem);
1520	NOREF(offSub);
1521	NOREF(cbSub);
1522	NOREF(fProt);
1523	return VERR_NOT_SUPPORTED;
1524	}
1525
1526
1527	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1528	{
1529	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1530
1531	if (pMemLnx->cPages)
1532	return page_to_phys(pMemLnx->apPages[iPage]);
1533
1534	switch (pMemLnx->Core.enmType)
1535	{
1536	case RTR0MEMOBJTYPE_CONT:
1537	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1538
1539	case RTR0MEMOBJTYPE_PHYS:
1540	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1541
1542	/* the parent knows */
1543	case RTR0MEMOBJTYPE_MAPPING:
1544	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1545
1546	/* cPages > 0 */
1547	case RTR0MEMOBJTYPE_LOW:
1548	case RTR0MEMOBJTYPE_LOCK:
1549	case RTR0MEMOBJTYPE_PHYS_NC:
1550	case RTR0MEMOBJTYPE_PAGE:
1551	default:
1552	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1553	/* fall thru */
1554
1555	case RTR0MEMOBJTYPE_RES_VIRT:
1556	return NIL_RTHCPHYS;
1557	}
1558	}
1559

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 41116

Download in other formats: