memobj-r0drv-linux.c@ 33640

Last change on this file since 33640 was 33640, checked in by vboxsync, 14 years ago
Linux host page allocator: don't use system reserves for >=1M reqs
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev`
File size: 46.8 KB

Line
1	/* $Revision: 33640 $ */
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	&& ( HAVE_26_STYLE_REMAP_PAGE_RANGE \
63	\|\| (LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) && LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)))
64	# define VBOX_USE_PAE_HACK
65	#endif
66
67
68	/*******************************************************************************
69	* Structures and Typedefs *
70	*******************************************************************************/
71	/**
72	* The Darwin version of the memory object structure.
73	*/
74	typedef struct RTR0MEMOBJLNX
75	{
76	/** The core structure. */
77	RTR0MEMOBJINTERNAL Core;
78	/** Set if the allocation is contiguous.
79	* This means it has to be given back as one chunk. */
80	bool fContiguous;
81	/** Set if we've vmap'ed the memory into ring-0. */
82	bool fMappedToRing0;
83	/** The pages in the apPages array. */
84	size_t cPages;
85	/** Array of struct page pointers. (variable size) */
86	struct page *apPages[1];
87	} RTR0MEMOBJLNX, *PRTR0MEMOBJLNX;
88
89
90	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx);
91
92
93	/**
94	* Helper that converts from a RTR0PROCESS handle to a linux task.
95	*
96	* @returns The corresponding Linux task.
97	* @param R0Process IPRT ring-0 process handle.
98	*/
99	struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
100	{
101	/** @todo fix rtR0ProcessToLinuxTask!! */
102	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
103	}
104
105
106	/**
107	* Compute order. Some functions allocate 2^order pages.
108	*
109	* @returns order.
110	* @param cPages Number of pages.
111	*/
112	static int rtR0MemObjLinuxOrder(size_t cPages)
113	{
114	int iOrder;
115	size_t cTmp;
116
117	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
118	;
119	if (cPages & ~((size_t)1 << iOrder))
120	++iOrder;
121
122	return iOrder;
123	}
124
125
126	/**
127	* Converts from RTMEM_PROT_* to Linux PAGE_*.
128	*
129	* @returns Linux page protection constant.
130	* @param fProt The IPRT protection mask.
131	* @param fKernel Whether it applies to kernel or user space.
132	*/
133	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
134	{
135	switch (fProt)
136	{
137	default:
138	AssertMsgFailed(("%#x %d\n", fProt, fKernel));
139	case RTMEM_PROT_NONE:
140	return PAGE_NONE;
141
142	case RTMEM_PROT_READ:
143	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
144
145	case RTMEM_PROT_WRITE:
146	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
147	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
148
149	case RTMEM_PROT_EXEC:
150	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
151	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
152	if (fKernel)
153	{
154	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
155	pgprot_val(fPg) &= ~_PAGE_RW;
156	return fPg;
157	}
158	return PAGE_READONLY_EXEC;
159	#else
160	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
161	#endif
162
163	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
164	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
165	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
166	}
167	}
168
169
170	/**
171	* Internal worker that allocates physical pages and creates the memory object for them.
172	*
173	* @returns IPRT status code.
174	* @param ppMemLnx Where to store the memory object pointer.
175	* @param enmType The object type.
176	* @param cb The number of bytes to allocate.
177	* @param uAlignment The alignment of the physical memory.
178	* Only valid if fContiguous == true, ignored otherwise.
179	* @param fFlagsLnx The page allocation flags (GPFs).
180	* @param fContiguous Whether the allocation must be contiguous.
181	*/
182	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
183	size_t uAlignment, unsigned fFlagsLnx, bool fContiguous)
184	{
185	size_t iPage;
186	size_t const cPages = cb >> PAGE_SHIFT;
187	struct page *paPages;
188
189	/*
190	* Allocate a memory object structure that's large enough to contain
191	* the page pointer array.
192	*/
193	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
194	if (!pMemLnx)
195	return VERR_NO_MEMORY;
196	pMemLnx->cPages = cPages;
197
198	# ifdef __GFP_NOMEMALLOC
199	if (cPages > 255)
200	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
201	fFlagsLnx \|= __GFP_NOMEMALLOC;
202	# endif
203
204	/*
205	* Allocate the pages.
206	* For small allocations we'll try contiguous first and then fall back on page by page.
207	*/
208	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
209	if ( fContiguous
210	\|\| cb <= PAGE_SIZE * 2)
211	{
212	# ifdef VBOX_USE_INSERT_PAGE
213	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP, rtR0MemObjLinuxOrder(cPages));
214	# else
215	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
216	# endif
217	if (paPages)
218	{
219	fContiguous = true;
220	for (iPage = 0; iPage < cPages; iPage++)
221	pMemLnx->apPages[iPage] = &paPages[iPage];
222	}
223	else if (fContiguous)
224	{
225	rtR0MemObjDelete(&pMemLnx->Core);
226	return VERR_NO_MEMORY;
227	}
228	}
229
230	if (!fContiguous)
231	{
232	for (iPage = 0; iPage < cPages; iPage++)
233	{
234	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx);
235	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
236	{
237	while (iPage-- > 0)
238	__free_page(pMemLnx->apPages[iPage]);
239	rtR0MemObjDelete(&pMemLnx->Core);
240	return VERR_NO_MEMORY;
241	}
242	}
243	}
244
245	#else /* < 2.4.22 */
246	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
247	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
248	if (!paPages)
249	{
250	rtR0MemObjDelete(&pMemLnx->Core);
251	return VERR_NO_MEMORY;
252	}
253	for (iPage = 0; iPage < cPages; iPage++)
254	{
255	pMemLnx->apPages[iPage] = &paPages[iPage];
256	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
257	if (PageHighMem(pMemLnx->apPages[iPage]))
258	BUG();
259	}
260
261	fContiguous = true;
262	#endif /* < 2.4.22 */
263	pMemLnx->fContiguous = fContiguous;
264
265	/*
266	* Reserve the pages.
267	*/
268	for (iPage = 0; iPage < cPages; iPage++)
269	SetPageReserved(pMemLnx->apPages[iPage]);
270
271	/*
272	* Note that the physical address of memory allocated with alloc_pages(flags, order)
273	* is always 2^(PAGE_SHIFT+order)-aligned.
274	*/
275	if ( fContiguous
276	&& uAlignment > PAGE_SIZE)
277	{
278	/*
279	* Check for alignment constraints. The physical address of memory allocated with
280	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
281	*/
282	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
283	{
284	/*
285	* This should never happen!
286	*/
287	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
288	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
289	rtR0MemObjLinuxFreePages(pMemLnx);
290	return VERR_NO_MEMORY;
291	}
292	}
293
294	*ppMemLnx = pMemLnx;
295	return VINF_SUCCESS;
296	}
297
298
299	/**
300	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
301	*
302	* This method does NOT free the object.
303	*
304	* @param pMemLnx The object which physical pages should be freed.
305	*/
306	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
307	{
308	size_t iPage = pMemLnx->cPages;
309	if (iPage > 0)
310	{
311	/*
312	* Restore the page flags.
313	*/
314	while (iPage-- > 0)
315	{
316	ClearPageReserved(pMemLnx->apPages[iPage]);
317	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
318	#else
319	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
320	#endif
321	}
322
323	/*
324	* Free the pages.
325	*/
326	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
327	if (!pMemLnx->fContiguous)
328	{
329	iPage = pMemLnx->cPages;
330	while (iPage-- > 0)
331	__free_page(pMemLnx->apPages[iPage]);
332	}
333	else
334	#endif
335	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
336
337	pMemLnx->cPages = 0;
338	}
339	}
340
341
342	/**
343	* Maps the allocation into ring-0.
344	*
345	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
346	*
347	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
348	* space, so we'll use that mapping if possible. If execute access is required, we'll
349	* play safe and do our own mapping.
350	*
351	* @returns IPRT status code.
352	* @param pMemLnx The linux memory object to map.
353	* @param fExecutable Whether execute access is required.
354	*/
355	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
356	{
357	int rc = VINF_SUCCESS;
358
359	/*
360	* Choose mapping strategy.
361	*/
362	bool fMustMap = fExecutable
363	\|\| !pMemLnx->fContiguous;
364	if (!fMustMap)
365	{
366	size_t iPage = pMemLnx->cPages;
367	while (iPage-- > 0)
368	if (PageHighMem(pMemLnx->apPages[iPage]))
369	{
370	fMustMap = true;
371	break;
372	}
373	}
374
375	Assert(!pMemLnx->Core.pv);
376	Assert(!pMemLnx->fMappedToRing0);
377
378	if (fMustMap)
379	{
380	/*
381	* Use vmap - 2.4.22 and later.
382	*/
383	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
384	pgprot_t fPg;
385	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
386	# ifdef _PAGE_NX
387	if (!fExecutable)
388	pgprot_val(fPg) \|= _PAGE_NX;
389	# endif
390
391	# ifdef VM_MAP
392	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
393	# else
394	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
395	# endif
396	if (pMemLnx->Core.pv)
397	pMemLnx->fMappedToRing0 = true;
398	else
399	rc = VERR_MAP_FAILED;
400	#else /* < 2.4.22 */
401	rc = VERR_NOT_SUPPORTED;
402	#endif
403	}
404	else
405	{
406	/*
407	* Use the kernel RAM mapping.
408	*/
409	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
410	Assert(pMemLnx->Core.pv);
411	}
412
413	return rc;
414	}
415
416
417	/**
418	* Undos what rtR0MemObjLinuxVMap() did.
419	*
420	* @param pMemLnx The linux memory object.
421	*/
422	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
423	{
424	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
425	if (pMemLnx->fMappedToRing0)
426	{
427	Assert(pMemLnx->Core.pv);
428	vunmap(pMemLnx->Core.pv);
429	pMemLnx->fMappedToRing0 = false;
430	}
431	#else /* < 2.4.22 */
432	Assert(!pMemLnx->fMappedToRing0);
433	#endif
434	pMemLnx->Core.pv = NULL;
435	}
436
437
438	int rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
439	{
440	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
441
442	/*
443	* Release any memory that we've allocated or locked.
444	*/
445	switch (pMemLnx->Core.enmType)
446	{
447	case RTR0MEMOBJTYPE_LOW:
448	case RTR0MEMOBJTYPE_PAGE:
449	case RTR0MEMOBJTYPE_CONT:
450	case RTR0MEMOBJTYPE_PHYS:
451	case RTR0MEMOBJTYPE_PHYS_NC:
452	rtR0MemObjLinuxVUnmap(pMemLnx);
453	rtR0MemObjLinuxFreePages(pMemLnx);
454	break;
455
456	case RTR0MEMOBJTYPE_LOCK:
457	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
458	{
459	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
460	size_t iPage;
461	Assert(pTask);
462	if (pTask && pTask->mm)
463	down_read(&pTask->mm->mmap_sem);
464
465	iPage = pMemLnx->cPages;
466	while (iPage-- > 0)
467	{
468	if (!PageReserved(pMemLnx->apPages[iPage]))
469	SetPageDirty(pMemLnx->apPages[iPage]);
470	page_cache_release(pMemLnx->apPages[iPage]);
471	}
472
473	if (pTask && pTask->mm)
474	up_read(&pTask->mm->mmap_sem);
475	}
476	/* else: kernel memory - nothing to do here. */
477	break;
478
479	case RTR0MEMOBJTYPE_RES_VIRT:
480	Assert(pMemLnx->Core.pv);
481	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
482	{
483	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
484	Assert(pTask);
485	if (pTask && pTask->mm)
486	{
487	down_write(&pTask->mm->mmap_sem);
488	MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
489	up_write(&pTask->mm->mmap_sem);
490	}
491	}
492	else
493	{
494	vunmap(pMemLnx->Core.pv);
495
496	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
497	__free_page(pMemLnx->apPages[0]);
498	pMemLnx->apPages[0] = NULL;
499	pMemLnx->cPages = 0;
500	}
501	pMemLnx->Core.pv = NULL;
502	break;
503
504	case RTR0MEMOBJTYPE_MAPPING:
505	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
506	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
507	{
508	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
509	Assert(pTask);
510	if (pTask && pTask->mm)
511	{
512	down_write(&pTask->mm->mmap_sem);
513	MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
514	up_write(&pTask->mm->mmap_sem);
515	}
516	}
517	else
518	vunmap(pMemLnx->Core.pv);
519	pMemLnx->Core.pv = NULL;
520	break;
521
522	default:
523	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
524	return VERR_INTERNAL_ERROR;
525	}
526	return VINF_SUCCESS;
527	}
528
529
530	int rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
531	{
532	PRTR0MEMOBJLNX pMemLnx;
533	int rc;
534
535	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
536	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER, false /* non-contiguous */);
537	#else
538	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER, false /* non-contiguous */);
539	#endif
540	if (RT_SUCCESS(rc))
541	{
542	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
543	if (RT_SUCCESS(rc))
544	{
545	*ppMem = &pMemLnx->Core;
546	return rc;
547	}
548
549	rtR0MemObjLinuxFreePages(pMemLnx);
550	rtR0MemObjDelete(&pMemLnx->Core);
551	}
552
553	return rc;
554	}
555
556
557	int rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
558	{
559	PRTR0MEMOBJLNX pMemLnx;
560	int rc;
561
562	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
563	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
564	/* ZONE_DMA32: 0-4GB */
565	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32, false /* non-contiguous */);
566	if (RT_FAILURE(rc))
567	#endif
568	#ifdef RT_ARCH_AMD64
569	/* ZONE_DMA: 0-16MB */
570	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA, false /* non-contiguous */);
571	#else
572	# ifdef CONFIG_X86_PAE
573	# endif
574	/* ZONE_NORMAL: 0-896MB */
575	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER, false /* non-contiguous */);
576	#endif
577	if (RT_SUCCESS(rc))
578	{
579	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
580	if (RT_SUCCESS(rc))
581	{
582	*ppMem = &pMemLnx->Core;
583	return rc;
584	}
585
586	rtR0MemObjLinuxFreePages(pMemLnx);
587	rtR0MemObjDelete(&pMemLnx->Core);
588	}
589
590	return rc;
591	}
592
593
594	int rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
595	{
596	PRTR0MEMOBJLNX pMemLnx;
597	int rc;
598
599	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
600	/* ZONE_DMA32: 0-4GB */
601	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32, true /* contiguous */);
602	if (RT_FAILURE(rc))
603	#endif
604	#ifdef RT_ARCH_AMD64
605	/* ZONE_DMA: 0-16MB */
606	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA, true /* contiguous */);
607	#else
608	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
609	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER, true /* contiguous */);
610	#endif
611	if (RT_SUCCESS(rc))
612	{
613	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
614	if (RT_SUCCESS(rc))
615	{
616	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
617	size_t iPage = pMemLnx->cPages;
618	while (iPage-- > 0)
619	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
620	#endif
621	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
622	*ppMem = &pMemLnx->Core;
623	return rc;
624	}
625
626	rtR0MemObjLinuxFreePages(pMemLnx);
627	rtR0MemObjDelete(&pMemLnx->Core);
628	}
629
630	return rc;
631	}
632
633
634	/**
635	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
636	*
637	* @returns IPRT status.
638	* @param ppMemLnx Where to
639	* @param enmType The object type.
640	* @param cb The size of the allocation.
641	* @param uAlignment The alignment of the physical memory.
642	* Only valid for fContiguous == true, ignored otherwise.
643	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
644	* @param fGfp The Linux GFP flags to use for the allocation.
645	*/
646	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
647	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, unsigned fGfp)
648	{
649	PRTR0MEMOBJLNX pMemLnx;
650	int rc;
651
652	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
653	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */);
654	if (RT_FAILURE(rc))
655	return rc;
656
657	/*
658	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
659	*/
660	if (PhysHighest != NIL_RTHCPHYS)
661	{
662	size_t iPage = pMemLnx->cPages;
663	while (iPage-- > 0)
664	if (page_to_phys(pMemLnx->apPages[iPage]) >= PhysHighest)
665	{
666	rtR0MemObjLinuxFreePages(pMemLnx);
667	rtR0MemObjDelete(&pMemLnx->Core);
668	return VERR_NO_MEMORY;
669	}
670	}
671
672	/*
673	* Complete the object.
674	*/
675	if (enmType == RTR0MEMOBJTYPE_PHYS)
676	{
677	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
678	pMemLnx->Core.u.Phys.fAllocated = true;
679	}
680	*ppMem = &pMemLnx->Core;
681	return rc;
682	}
683
684
685	/**
686	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
687	*
688	* @returns IPRT status.
689	* @param ppMem Where to store the memory object pointer on success.
690	* @param enmType The object type.
691	* @param cb The size of the allocation.
692	* @param uAlignment The alignment of the physical memory.
693	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
694	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
695	*/
696	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
697	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
698	{
699	int rc;
700
701	/*
702	* There are two clear cases and that's the <=16MB and anything-goes ones.
703	* When the physical address limit is somewhere in-between those two we'll
704	* just have to try, starting with HIGHUSER and working our way thru the
705	* different types, hoping we'll get lucky.
706	*
707	* We should probably move this physical address restriction logic up to
708	* the page alloc function as it would be more efficient there. But since
709	* we don't expect this to be a performance issue just yet it can wait.
710	*/
711	if (PhysHighest == NIL_RTHCPHYS)
712	/* ZONE_HIGHMEM: the whole physical memory */
713	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
714	else if (PhysHighest <= _1M * 16)
715	/* ZONE_DMA: 0-16MB */
716	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
717	else
718	{
719	rc = VERR_NO_MEMORY;
720	if (RT_FAILURE(rc))
721	/* ZONE_HIGHMEM: the whole physical memory */
722	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
723	if (RT_FAILURE(rc))
724	/* ZONE_NORMAL: 0-896MB */
725	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
726	#ifdef GFP_DMA32
727	if (RT_FAILURE(rc))
728	/* ZONE_DMA32: 0-4GB */
729	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
730	#endif
731	if (RT_FAILURE(rc))
732	/* ZONE_DMA: 0-16MB */
733	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
734	}
735	return rc;
736	}
737
738
739	int rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
740	{
741	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
742	}
743
744
745	int rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
746	{
747	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
748	}
749
750
751	int rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
752	{
753	/*
754	* All we need to do here is to validate that we can use
755	* ioremap on the specified address (32/64-bit dma_addr_t).
756	*/
757	PRTR0MEMOBJLNX pMemLnx;
758	dma_addr_t PhysAddr = Phys;
759	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
760
761	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
762	if (!pMemLnx)
763	return VERR_NO_MEMORY;
764
765	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
766	pMemLnx->Core.u.Phys.fAllocated = false;
767	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
768	Assert(!pMemLnx->cPages);
769	*ppMem = &pMemLnx->Core;
770	return VINF_SUCCESS;
771	}
772
773
774	int rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
775	{
776	const int cPages = cb >> PAGE_SHIFT;
777	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
778	struct vm_area_struct **papVMAs;
779	PRTR0MEMOBJLNX pMemLnx;
780	int rc = VERR_NO_MEMORY;
781	NOREF(fAccess);
782
783	/*
784	* Check for valid task and size overflows.
785	*/
786	if (!pTask)
787	return VERR_NOT_SUPPORTED;
788	if (((size_t)cPages << PAGE_SHIFT) != cb)
789	return VERR_OUT_OF_RANGE;
790
791	/*
792	* Allocate the memory object and a temporary buffer for the VMAs.
793	*/
794	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
795	if (!pMemLnx)
796	return VERR_NO_MEMORY;
797
798	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
799	if (papVMAs)
800	{
801	down_read(&pTask->mm->mmap_sem);
802
803	/*
804	* Get user pages.
805	*/
806	rc = get_user_pages(pTask, /* Task for fault accounting. */
807	pTask->mm, /* Whose pages. */
808	R3Ptr, /* Where from. */
809	cPages, /* How many pages. */
810	1, /* Write to memory. */
811	0, /* force. */
812	&pMemLnx->apPages[0], /* Page array. */
813	papVMAs); /* vmas */
814	if (rc == cPages)
815	{
816	/*
817	* Flush dcache (required?), protect against fork and _really_ pin the page
818	* table entries. get_user_pages() will protect against swapping out the
819	* pages but it will NOT protect against removing page table entries. This
820	* can be achieved with
821	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
822	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
823	* Usual Linux distributions support only a limited size of locked pages
824	* (e.g. 32KB).
825	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
826	* or by
827	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
828	* a range check.
829	*/
830	/** @todo The Linux fork() protection will require more work if this API
831	* is to be used for anything but locking VM pages. */
832	while (rc-- > 0)
833	{
834	flush_dcache_page(pMemLnx->apPages[rc]);
835	papVMAs[rc]->vm_flags \|= (VM_DONTCOPY \| VM_LOCKED);
836	}
837
838	up_read(&pTask->mm->mmap_sem);
839
840	RTMemFree(papVMAs);
841
842	pMemLnx->Core.u.Lock.R0Process = R0Process;
843	pMemLnx->cPages = cPages;
844	Assert(!pMemLnx->fMappedToRing0);
845	*ppMem = &pMemLnx->Core;
846
847	return VINF_SUCCESS;
848	}
849
850	/*
851	* Failed - we need to unlock any pages that we succeeded to lock.
852	*/
853	while (rc-- > 0)
854	{
855	if (!PageReserved(pMemLnx->apPages[rc]))
856	SetPageDirty(pMemLnx->apPages[rc]);
857	page_cache_release(pMemLnx->apPages[rc]);
858	}
859
860	up_read(&pTask->mm->mmap_sem);
861
862	RTMemFree(papVMAs);
863	rc = VERR_LOCK_FAILED;
864	}
865
866	rtR0MemObjDelete(&pMemLnx->Core);
867	return rc;
868	}
869
870
871	int rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
872	{
873	void pvLast = (uint8_t )pv + cb - 1;
874	size_t const cPages = cb >> PAGE_SHIFT;
875	PRTR0MEMOBJLNX pMemLnx;
876	bool fLinearMapping;
877	int rc;
878	uint8_t *pbPage;
879	size_t iPage;
880	NOREF(fAccess);
881
882	/*
883	* Classify the memory and check that we can deal with it.
884	*/
885	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0)
886	fLinearMapping = virt_addr_valid(pvLast) && virt_addr_valid(pv);
887	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 0)
888	fLinearMapping = VALID_PAGE(virt_to_page(pvLast)) && VALID_PAGE(virt_to_page(pv));
889	#else
890	# error "not supported"
891	#endif
892	if (!fLinearMapping)
893	{
894	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 19)
895	if ( !RTR0MemKernelIsValidAddr(pv)
896	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
897	#endif
898	return VERR_INVALID_PARAMETER;
899	}
900
901	/*
902	* Allocate the memory object.
903	*/
904	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
905	if (!pMemLnx)
906	return VERR_NO_MEMORY;
907
908	/*
909	* Gather the pages.
910	* We ASSUME all kernel pages are non-swappable.
911	*/
912	rc = VINF_SUCCESS;
913	pbPage = (uint8_t *)pvLast;
914	iPage = cPages;
915	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 19)
916	if (!fLinearMapping)
917	{
918	while (iPage-- > 0)
919	{
920	struct page *pPage = vmalloc_to_page(pbPage);
921	if (RT_UNLIKELY(!pPage))
922	{
923	rc = VERR_LOCK_FAILED;
924	break;
925	}
926	pMemLnx->apPages[iPage] = pPage;
927	pbPage -= PAGE_SIZE;
928	}
929	}
930	else
931	#endif
932	{
933	while (iPage-- > 0)
934	{
935	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
936	pbPage -= PAGE_SIZE;
937	}
938	}
939	if (RT_SUCCESS(rc))
940	{
941	/*
942	* Complete the memory object and return.
943	*/
944	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
945	pMemLnx->cPages = cPages;
946	Assert(!pMemLnx->fMappedToRing0);
947	*ppMem = &pMemLnx->Core;
948
949	return VINF_SUCCESS;
950	}
951
952	rtR0MemObjDelete(&pMemLnx->Core);
953	return rc;
954	}
955
956
957	int rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
958	{
959	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
960	const size_t cPages = cb >> PAGE_SHIFT;
961	struct page *pDummyPage;
962	struct page **papPages;
963
964	/* check for unsupported stuff. */
965	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
966	if (uAlignment > PAGE_SIZE)
967	return VERR_NOT_SUPPORTED;
968
969	/*
970	* Allocate a dummy page and create a page pointer array for vmap such that
971	* the dummy page is mapped all over the reserved area.
972	*/
973	pDummyPage = alloc_page(GFP_HIGHUSER);
974	if (!pDummyPage)
975	return VERR_NO_MEMORY;
976	papPages = RTMemAlloc(sizeof(papPages) cPages);
977	if (papPages)
978	{
979	void *pv;
980	size_t iPage = cPages;
981	while (iPage-- > 0)
982	papPages[iPage] = pDummyPage;
983	# ifdef VM_MAP
984	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
985	# else
986	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
987	# endif
988	RTMemFree(papPages);
989	if (pv)
990	{
991	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
992	if (pMemLnx)
993	{
994	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
995	pMemLnx->cPages = 1;
996	pMemLnx->apPages[0] = pDummyPage;
997	*ppMem = &pMemLnx->Core;
998	return VINF_SUCCESS;
999	}
1000	vunmap(pv);
1001	}
1002	}
1003	__free_page(pDummyPage);
1004	return VERR_NO_MEMORY;
1005
1006	#else /* < 2.4.22 */
1007	/*
1008	* Could probably use ioremap here, but the caller is in a better position than us
1009	* to select some safe physical memory.
1010	*/
1011	return VERR_NOT_SUPPORTED;
1012	#endif
1013	}
1014
1015
1016	/**
1017	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
1018	* an empty user space mapping.
1019	*
1020	* The caller takes care of acquiring the mmap_sem of the task.
1021	*
1022	* @returns Pointer to the mapping.
1023	* (void *)-1 on failure.
1024	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
1025	* @param cb The size of the mapping.
1026	* @param uAlignment The alignment of the mapping.
1027	* @param pTask The Linux task to create this mapping in.
1028	* @param fProt The RTMEM_PROT_* mask.
1029	*/
1030	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
1031	{
1032	unsigned fLnxProt;
1033	unsigned long ulAddr;
1034
1035	/*
1036	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
1037	*/
1038	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
1039	if (fProt == RTMEM_PROT_NONE)
1040	fLnxProt = PROT_NONE;
1041	else
1042	{
1043	fLnxProt = 0;
1044	if (fProt & RTMEM_PROT_READ)
1045	fLnxProt \|= PROT_READ;
1046	if (fProt & RTMEM_PROT_WRITE)
1047	fLnxProt \|= PROT_WRITE;
1048	if (fProt & RTMEM_PROT_EXEC)
1049	fLnxProt \|= PROT_EXEC;
1050	}
1051
1052	if (R3PtrFixed != (RTR3PTR)-1)
1053	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
1054	else
1055	{
1056	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
1057	if ( !(ulAddr & ~PAGE_MASK)
1058	&& (ulAddr & (uAlignment - 1)))
1059	{
1060	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
1061	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
1062	* ourselves) and further by there begin two mmap strategies (top / bottom). */
1063	/* For now, just ignore uAlignment requirements... */
1064	}
1065	}
1066	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
1067	return (void *)-1;
1068	return (void *)ulAddr;
1069	}
1070
1071
1072	int rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1073	{
1074	PRTR0MEMOBJLNX pMemLnx;
1075	void *pv;
1076	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1077	if (!pTask)
1078	return VERR_NOT_SUPPORTED;
1079
1080	/*
1081	* Check that the specified alignment is supported.
1082	*/
1083	if (uAlignment > PAGE_SIZE)
1084	return VERR_NOT_SUPPORTED;
1085
1086	/*
1087	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1088	*/
1089	down_write(&pTask->mm->mmap_sem);
1090	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1091	up_write(&pTask->mm->mmap_sem);
1092	if (pv == (void *)-1)
1093	return VERR_NO_MEMORY;
1094
1095	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1096	if (!pMemLnx)
1097	{
1098	down_write(&pTask->mm->mmap_sem);
1099	MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, cb);
1100	up_write(&pTask->mm->mmap_sem);
1101	return VERR_NO_MEMORY;
1102	}
1103
1104	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1105	*ppMem = &pMemLnx->Core;
1106	return VINF_SUCCESS;
1107	}
1108
1109
1110	int rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment,
1111	unsigned fProt, size_t offSub, size_t cbSub)
1112	{
1113	int rc = VERR_NO_MEMORY;
1114	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1115	PRTR0MEMOBJLNX pMemLnx;
1116
1117	/* Fail if requested to do something we can't. */
1118	AssertMsgReturn(!offSub && !cbSub, ("%#x %#x\n", offSub, cbSub), VERR_NOT_SUPPORTED);
1119	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1120	if (uAlignment > PAGE_SIZE)
1121	return VERR_NOT_SUPPORTED;
1122
1123	/*
1124	* Create the IPRT memory object.
1125	*/
1126	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1127	if (pMemLnx)
1128	{
1129	if (pMemLnxToMap->cPages)
1130	{
1131	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1132	/*
1133	* Use vmap - 2.4.22 and later.
1134	*/
1135	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1136	# ifdef VM_MAP
1137	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
1138	# else
1139	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
1140	# endif
1141	if (pMemLnx->Core.pv)
1142	{
1143	pMemLnx->fMappedToRing0 = true;
1144	rc = VINF_SUCCESS;
1145	}
1146	else
1147	rc = VERR_MAP_FAILED;
1148
1149	#else /* < 2.4.22 */
1150	/*
1151	* Only option here is to share mappings if possible and forget about fProt.
1152	*/
1153	if (rtR0MemObjIsRing3(pMemToMap))
1154	rc = VERR_NOT_SUPPORTED;
1155	else
1156	{
1157	rc = VINF_SUCCESS;
1158	if (!pMemLnxToMap->Core.pv)
1159	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1160	if (RT_SUCCESS(rc))
1161	{
1162	Assert(pMemLnxToMap->Core.pv);
1163	pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
1164	}
1165	}
1166	#endif
1167	}
1168	else
1169	{
1170	/*
1171	* MMIO / physical memory.
1172	*/
1173	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1174	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1175	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb)
1176	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
1177	if (pMemLnx->Core.pv)
1178	{
1179	/** @todo fix protection. */
1180	rc = VINF_SUCCESS;
1181	}
1182	}
1183	if (RT_SUCCESS(rc))
1184	{
1185	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1186	*ppMem = &pMemLnx->Core;
1187	return VINF_SUCCESS;
1188	}
1189	rtR0MemObjDelete(&pMemLnx->Core);
1190	}
1191
1192	return rc;
1193	}
1194
1195
1196	#ifdef VBOX_USE_PAE_HACK
1197	/**
1198	* Replace the PFN of a PTE with the address of the actual page.
1199	*
1200	* The caller maps a reserved dummy page at the address with the desired access
1201	* and flags.
1202	*
1203	* This hack is required for older Linux kernels which don't provide
1204	* remap_pfn_range().
1205	*
1206	* @returns 0 on success, -ENOMEM on failure.
1207	* @param mm The memory context.
1208	* @param ulAddr The mapping address.
1209	* @param Phys The physical address of the page to map.
1210	*/
1211	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1212	{
1213	int rc = -ENOMEM;
1214	pgd_t *pgd;
1215
1216	spin_lock(&mm->page_table_lock);
1217
1218	pgd = pgd_offset(mm, ulAddr);
1219	if (!pgd_none(pgd) && !pgd_bad(pgd))
1220	{
1221	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1222	if (!pmd_none(*pmd))
1223	{
1224	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1225	if (ptep)
1226	{
1227	pte_t pte = *ptep;
1228	pte.pte_high &= 0xfff00000;
1229	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1230	pte.pte_low &= 0x00000fff;
1231	pte.pte_low \|= (Phys & 0xfffff000);
1232	set_pte(ptep, pte);
1233	pte_unmap(ptep);
1234	rc = 0;
1235	}
1236	}
1237	}
1238
1239	spin_unlock(&mm->page_table_lock);
1240	return rc;
1241	}
1242	#endif /* VBOX_USE_PAE_HACK */
1243
1244
1245	int rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1246	{
1247	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1248	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1249	int rc = VERR_NO_MEMORY;
1250	PRTR0MEMOBJLNX pMemLnx;
1251	#ifdef VBOX_USE_PAE_HACK
1252	struct page *pDummyPage;
1253	RTHCPHYS DummyPhys;
1254	#endif
1255
1256	/*
1257	* Check for restrictions.
1258	*/
1259	if (!pTask)
1260	return VERR_NOT_SUPPORTED;
1261	if (uAlignment > PAGE_SIZE)
1262	return VERR_NOT_SUPPORTED;
1263
1264	#ifdef VBOX_USE_PAE_HACK
1265	/*
1266	* Allocate a dummy page for use when mapping the memory.
1267	*/
1268	pDummyPage = alloc_page(GFP_USER);
1269	if (!pDummyPage)
1270	return VERR_NO_MEMORY;
1271	SetPageReserved(pDummyPage);
1272	DummyPhys = page_to_phys(pDummyPage);
1273	#endif
1274
1275	/*
1276	* Create the IPRT memory object.
1277	*/
1278	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1279	if (pMemLnx)
1280	{
1281	/*
1282	* Allocate user space mapping.
1283	*/
1284	void *pv;
1285	down_write(&pTask->mm->mmap_sem);
1286	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1287	if (pv != (void *)-1)
1288	{
1289	/*
1290	* Map page by page into the mmap area.
1291	* This is generic, paranoid and not very efficient.
1292	*/
1293	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1294	unsigned long ulAddrCur = (unsigned long)pv;
1295	const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1296	size_t iPage;
1297
1298	rc = 0;
1299	if (pMemLnxToMap->cPages)
1300	{
1301	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1302	{
1303	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)
1304	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1305	#endif
1306	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1307	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1308	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1309	#endif
1310	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1311	/* remap_page_range() limitation on x86 */
1312	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1313	#endif
1314
1315	#if defined(VBOX_USE_INSERT_PAGE) && LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
1316	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1317	vma->vm_flags \|= VM_RESERVED; /* This flag helps making 100% sure some bad stuff wont happen (swap, core, ++). */
1318	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1319	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1320	#elif defined(VBOX_USE_PAE_HACK)
1321	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1322	if (!rc)
1323	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1324	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1325	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1326	#else /* 2.4 */
1327	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1328	#endif
1329	if (rc)
1330	{
1331	rc = VERR_NO_MEMORY;
1332	break;
1333	}
1334	}
1335	}
1336	else
1337	{
1338	RTHCPHYS Phys;
1339	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1340	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1341	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1342	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1343	else
1344	{
1345	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1346	Phys = NIL_RTHCPHYS;
1347	}
1348	if (Phys != NIL_RTHCPHYS)
1349	{
1350	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1351	{
1352	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1353	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1354	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1355	#endif
1356	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1357	/* remap_page_range() limitation on x86 */
1358	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1359	#endif
1360
1361	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1362	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1363	#elif defined(VBOX_USE_PAE_HACK)
1364	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1365	if (!rc)
1366	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1367	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1368	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1369	#else /* 2.4 */
1370	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1371	#endif
1372	if (rc)
1373	{
1374	rc = VERR_NO_MEMORY;
1375	break;
1376	}
1377	}
1378	}
1379	}
1380	if (!rc)
1381	{
1382	up_write(&pTask->mm->mmap_sem);
1383	#ifdef VBOX_USE_PAE_HACK
1384	__free_page(pDummyPage);
1385	#endif
1386
1387	pMemLnx->Core.pv = pv;
1388	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1389	*ppMem = &pMemLnx->Core;
1390	return VINF_SUCCESS;
1391	}
1392
1393	/*
1394	* Bail out.
1395	*/
1396	MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, pMemLnxToMap->Core.cb);
1397	}
1398	up_write(&pTask->mm->mmap_sem);
1399	rtR0MemObjDelete(&pMemLnx->Core);
1400	}
1401	#ifdef VBOX_USE_PAE_HACK
1402	__free_page(pDummyPage);
1403	#endif
1404
1405	return rc;
1406	}
1407
1408
1409	int rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1410	{
1411	NOREF(pMem);
1412	NOREF(offSub);
1413	NOREF(cbSub);
1414	NOREF(fProt);
1415	return VERR_NOT_SUPPORTED;
1416	}
1417
1418
1419	RTHCPHYS rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1420	{
1421	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1422
1423	if (pMemLnx->cPages)
1424	return page_to_phys(pMemLnx->apPages[iPage]);
1425
1426	switch (pMemLnx->Core.enmType)
1427	{
1428	case RTR0MEMOBJTYPE_CONT:
1429	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1430
1431	case RTR0MEMOBJTYPE_PHYS:
1432	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1433
1434	/* the parent knows */
1435	case RTR0MEMOBJTYPE_MAPPING:
1436	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1437
1438	/* cPages > 0 */
1439	case RTR0MEMOBJTYPE_LOW:
1440	case RTR0MEMOBJTYPE_LOCK:
1441	case RTR0MEMOBJTYPE_PHYS_NC:
1442	case RTR0MEMOBJTYPE_PAGE:
1443	default:
1444	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1445	/* fall thru */
1446
1447	case RTR0MEMOBJTYPE_RES_VIRT:
1448	return NIL_RTHCPHYS;
1449	}
1450	}
1451

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 33640

Download in other formats: