memobj-r0drv-linux.c@ 53402

Last change on this file since 53402 was 51938, checked in by vboxsync, 10 years ago
Runtime/r0drv: RHEL7 hack
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev Revision`
File size: 53.1 KB

Line
1	/* $Revision: 51938 $ */
2	/** @file
3	* IPRT - Ring-0 Memory Objects, Linux.
4	*/
5
6	/*
7	* Copyright (C) 2006-2012 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*******************************************************************************
29	* Header Files *
30	*******************************************************************************/
31	#include "the-linux-kernel.h"
32
33	#include <iprt/memobj.h>
34	#include <iprt/alloc.h>
35	#include <iprt/assert.h>
36	#include <iprt/log.h>
37	#include <iprt/process.h>
38	#include <iprt/string.h>
39	#include "internal/memobj.h"
40
41
42	/*******************************************************************************
43	* Defined Constants And Macros *
44	*******************************************************************************/
45	/* early 2.6 kernels */
46	#ifndef PAGE_SHARED_EXEC
47	# define PAGE_SHARED_EXEC PAGE_SHARED
48	#endif
49	#ifndef PAGE_READONLY_EXEC
50	# define PAGE_READONLY_EXEC PAGE_READONLY
51	#endif
52
53	/*
54	* 2.6.29+ kernels don't work with remap_pfn_range() anymore because
55	* track_pfn_vma_new() is apparently not defined for non-RAM pages.
56	* It should be safe to use vm_insert_page() older kernels as well.
57	*/
58	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 23)
59	# define VBOX_USE_INSERT_PAGE
60	#endif
61	#if defined(CONFIG_X86_PAE) \
62	&& ( defined(HAVE_26_STYLE_REMAP_PAGE_RANGE) \
63	\|\| ( LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \
64	&& LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)))
65	# define VBOX_USE_PAE_HACK
66	#endif
67
68
69	/*******************************************************************************
70	* Structures and Typedefs *
71	*******************************************************************************/
72	/**
73	* The Darwin version of the memory object structure.
74	*/
75	typedef struct RTR0MEMOBJLNX
76	{
77	/** The core structure. */
78	RTR0MEMOBJINTERNAL Core;
79	/** Set if the allocation is contiguous.
80	* This means it has to be given back as one chunk. */
81	bool fContiguous;
82	/** Set if we've vmap'ed the memory into ring-0. */
83	bool fMappedToRing0;
84	/** The pages in the apPages array. */
85	size_t cPages;
86	/** Array of struct page pointers. (variable size) */
87	struct page *apPages[1];
88	} RTR0MEMOBJLNX, *PRTR0MEMOBJLNX;
89
90
91	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx);
92
93
94	/**
95	* Helper that converts from a RTR0PROCESS handle to a linux task.
96	*
97	* @returns The corresponding Linux task.
98	* @param R0Process IPRT ring-0 process handle.
99	*/
100	static struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
101	{
102	/** @todo fix rtR0ProcessToLinuxTask!! */
103	/** @todo many (all?) callers currently assume that we return 'current'! */
104	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
105	}
106
107
108	/**
109	* Compute order. Some functions allocate 2^order pages.
110	*
111	* @returns order.
112	* @param cPages Number of pages.
113	*/
114	static int rtR0MemObjLinuxOrder(size_t cPages)
115	{
116	int iOrder;
117	size_t cTmp;
118
119	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
120	;
121	if (cPages & ~((size_t)1 << iOrder))
122	++iOrder;
123
124	return iOrder;
125	}
126
127
128	/**
129	* Converts from RTMEM_PROT_* to Linux PAGE_*.
130	*
131	* @returns Linux page protection constant.
132	* @param fProt The IPRT protection mask.
133	* @param fKernel Whether it applies to kernel or user space.
134	*/
135	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
136	{
137	switch (fProt)
138	{
139	default:
140	AssertMsgFailed(("%#x %d\n", fProt, fKernel));
141	case RTMEM_PROT_NONE:
142	return PAGE_NONE;
143
144	case RTMEM_PROT_READ:
145	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
146
147	case RTMEM_PROT_WRITE:
148	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
149	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
150
151	case RTMEM_PROT_EXEC:
152	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
153	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
154	if (fKernel)
155	{
156	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
157	pgprot_val(fPg) &= ~_PAGE_RW;
158	return fPg;
159	}
160	return PAGE_READONLY_EXEC;
161	#else
162	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
163	#endif
164
165	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
166	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
167	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
168	}
169	}
170
171
172	/**
173	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
174	* an empty user space mapping.
175	*
176	* We acquire the mmap_sem of the task!
177	*
178	* @returns Pointer to the mapping.
179	* (void *)-1 on failure.
180	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
181	* @param cb The size of the mapping.
182	* @param uAlignment The alignment of the mapping.
183	* @param pTask The Linux task to create this mapping in.
184	* @param fProt The RTMEM_PROT_* mask.
185	*/
186	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
187	{
188	unsigned fLnxProt;
189	unsigned long ulAddr;
190
191	Assert((pTask == current)); /* do_mmap */
192
193	/*
194	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
195	*/
196	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
197	if (fProt == RTMEM_PROT_NONE)
198	fLnxProt = PROT_NONE;
199	else
200	{
201	fLnxProt = 0;
202	if (fProt & RTMEM_PROT_READ)
203	fLnxProt \|= PROT_READ;
204	if (fProt & RTMEM_PROT_WRITE)
205	fLnxProt \|= PROT_WRITE;
206	if (fProt & RTMEM_PROT_EXEC)
207	fLnxProt \|= PROT_EXEC;
208	}
209
210	if (R3PtrFixed != (RTR3PTR)-1)
211	{
212	#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 5, 0)
213	ulAddr = vm_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
214	#else
215	down_write(&pTask->mm->mmap_sem);
216	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
217	up_write(&pTask->mm->mmap_sem);
218	#endif
219	}
220	else
221	{
222	#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 5, 0)
223	ulAddr = vm_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
224	#else
225	down_write(&pTask->mm->mmap_sem);
226	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
227	up_write(&pTask->mm->mmap_sem);
228	#endif
229	if ( !(ulAddr & ~PAGE_MASK)
230	&& (ulAddr & (uAlignment - 1)))
231	{
232	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
233	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
234	* ourselves) and further by there begin two mmap strategies (top / bottom). */
235	/* For now, just ignore uAlignment requirements... */
236	}
237	}
238
239
240	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
241	return (void *)-1;
242	return (void *)ulAddr;
243	}
244
245
246	/**
247	* Worker that destroys a user space mapping.
248	* Undoes what rtR0MemObjLinuxDoMmap did.
249	*
250	* We acquire the mmap_sem of the task!
251	*
252	* @param pv The ring-3 mapping.
253	* @param cb The size of the mapping.
254	* @param pTask The Linux task to destroy this mapping in.
255	*/
256	static void rtR0MemObjLinuxDoMunmap(void pv, size_t cb, struct task_struct pTask)
257	{
258	#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 5, 0)
259	Assert(pTask == current);
260	vm_munmap((unsigned long)pv, cb);
261	#elif defined(USE_RHEL4_MUNMAP)
262	down_write(&pTask->mm->mmap_sem);
263	do_munmap(pTask->mm, (unsigned long)pv, cb, 0); /* should it be 1 or 0? */
264	up_write(&pTask->mm->mmap_sem);
265	#else
266	down_write(&pTask->mm->mmap_sem);
267	do_munmap(pTask->mm, (unsigned long)pv, cb);
268	up_write(&pTask->mm->mmap_sem);
269	#endif
270	}
271
272
273	/**
274	* Internal worker that allocates physical pages and creates the memory object for them.
275	*
276	* @returns IPRT status code.
277	* @param ppMemLnx Where to store the memory object pointer.
278	* @param enmType The object type.
279	* @param cb The number of bytes to allocate.
280	* @param uAlignment The alignment of the physical memory.
281	* Only valid if fContiguous == true, ignored otherwise.
282	* @param fFlagsLnx The page allocation flags (GPFs).
283	* @param fContiguous Whether the allocation must be contiguous.
284	* @param rcNoMem What to return when we're out of pages.
285	*/
286	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
287	size_t uAlignment, unsigned fFlagsLnx, bool fContiguous, int rcNoMem)
288	{
289	size_t iPage;
290	size_t const cPages = cb >> PAGE_SHIFT;
291	struct page *paPages;
292
293	/*
294	* Allocate a memory object structure that's large enough to contain
295	* the page pointer array.
296	*/
297	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
298	if (!pMemLnx)
299	return VERR_NO_MEMORY;
300	pMemLnx->cPages = cPages;
301
302	if (cPages > 255)
303	{
304	# ifdef __GFP_REPEAT
305	/* Try hard to allocate the memory, but the allocation attempt might fail. */
306	fFlagsLnx \|= __GFP_REPEAT;
307	# endif
308	# ifdef __GFP_NOMEMALLOC
309	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
310	fFlagsLnx \|= __GFP_NOMEMALLOC;
311	# endif
312	}
313
314	/*
315	* Allocate the pages.
316	* For small allocations we'll try contiguous first and then fall back on page by page.
317	*/
318	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
319	if ( fContiguous
320	\|\| cb <= PAGE_SIZE * 2)
321	{
322	# ifdef VBOX_USE_INSERT_PAGE
323	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
324	# else
325	paPages = alloc_pages(fFlagsLnx \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
326	# endif
327	if (paPages)
328	{
329	fContiguous = true;
330	for (iPage = 0; iPage < cPages; iPage++)
331	pMemLnx->apPages[iPage] = &paPages[iPage];
332	}
333	else if (fContiguous)
334	{
335	rtR0MemObjDelete(&pMemLnx->Core);
336	return rcNoMem;
337	}
338	}
339
340	if (!fContiguous)
341	{
342	for (iPage = 0; iPage < cPages; iPage++)
343	{
344	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
345	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
346	{
347	while (iPage-- > 0)
348	__free_page(pMemLnx->apPages[iPage]);
349	rtR0MemObjDelete(&pMemLnx->Core);
350	return rcNoMem;
351	}
352	}
353	}
354
355	#else /* < 2.4.22 */
356	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
357	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
358	if (!paPages)
359	{
360	rtR0MemObjDelete(&pMemLnx->Core);
361	return rcNoMem;
362	}
363	for (iPage = 0; iPage < cPages; iPage++)
364	{
365	pMemLnx->apPages[iPage] = &paPages[iPage];
366	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
367	if (PageHighMem(pMemLnx->apPages[iPage]))
368	BUG();
369	}
370
371	fContiguous = true;
372	#endif /* < 2.4.22 */
373	pMemLnx->fContiguous = fContiguous;
374
375	/*
376	* Reserve the pages.
377	*/
378	for (iPage = 0; iPage < cPages; iPage++)
379	SetPageReserved(pMemLnx->apPages[iPage]);
380
381	/*
382	* Note that the physical address of memory allocated with alloc_pages(flags, order)
383	* is always 2^(PAGE_SHIFT+order)-aligned.
384	*/
385	if ( fContiguous
386	&& uAlignment > PAGE_SIZE)
387	{
388	/*
389	* Check for alignment constraints. The physical address of memory allocated with
390	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
391	*/
392	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
393	{
394	/*
395	* This should never happen!
396	*/
397	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
398	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
399	rtR0MemObjLinuxFreePages(pMemLnx);
400	return rcNoMem;
401	}
402	}
403
404	*ppMemLnx = pMemLnx;
405	return VINF_SUCCESS;
406	}
407
408
409	/**
410	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
411	*
412	* This method does NOT free the object.
413	*
414	* @param pMemLnx The object which physical pages should be freed.
415	*/
416	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
417	{
418	size_t iPage = pMemLnx->cPages;
419	if (iPage > 0)
420	{
421	/*
422	* Restore the page flags.
423	*/
424	while (iPage-- > 0)
425	{
426	ClearPageReserved(pMemLnx->apPages[iPage]);
427	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
428	#else
429	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
430	#endif
431	}
432
433	/*
434	* Free the pages.
435	*/
436	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
437	if (!pMemLnx->fContiguous)
438	{
439	iPage = pMemLnx->cPages;
440	while (iPage-- > 0)
441	__free_page(pMemLnx->apPages[iPage]);
442	}
443	else
444	#endif
445	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
446
447	pMemLnx->cPages = 0;
448	}
449	}
450
451
452	/**
453	* Maps the allocation into ring-0.
454	*
455	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
456	*
457	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
458	* space, so we'll use that mapping if possible. If execute access is required, we'll
459	* play safe and do our own mapping.
460	*
461	* @returns IPRT status code.
462	* @param pMemLnx The linux memory object to map.
463	* @param fExecutable Whether execute access is required.
464	*/
465	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
466	{
467	int rc = VINF_SUCCESS;
468
469	/*
470	* Choose mapping strategy.
471	*/
472	bool fMustMap = fExecutable
473	\|\| !pMemLnx->fContiguous;
474	if (!fMustMap)
475	{
476	size_t iPage = pMemLnx->cPages;
477	while (iPage-- > 0)
478	if (PageHighMem(pMemLnx->apPages[iPage]))
479	{
480	fMustMap = true;
481	break;
482	}
483	}
484
485	Assert(!pMemLnx->Core.pv);
486	Assert(!pMemLnx->fMappedToRing0);
487
488	if (fMustMap)
489	{
490	/*
491	* Use vmap - 2.4.22 and later.
492	*/
493	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
494	pgprot_t fPg;
495	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
496	# ifdef _PAGE_NX
497	if (!fExecutable)
498	pgprot_val(fPg) \|= _PAGE_NX;
499	# endif
500
501	# ifdef VM_MAP
502	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
503	# else
504	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
505	# endif
506	if (pMemLnx->Core.pv)
507	pMemLnx->fMappedToRing0 = true;
508	else
509	rc = VERR_MAP_FAILED;
510	#else /* < 2.4.22 */
511	rc = VERR_NOT_SUPPORTED;
512	#endif
513	}
514	else
515	{
516	/*
517	* Use the kernel RAM mapping.
518	*/
519	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
520	Assert(pMemLnx->Core.pv);
521	}
522
523	return rc;
524	}
525
526
527	/**
528	* Undoes what rtR0MemObjLinuxVMap() did.
529	*
530	* @param pMemLnx The linux memory object.
531	*/
532	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
533	{
534	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
535	if (pMemLnx->fMappedToRing0)
536	{
537	Assert(pMemLnx->Core.pv);
538	vunmap(pMemLnx->Core.pv);
539	pMemLnx->fMappedToRing0 = false;
540	}
541	#else /* < 2.4.22 */
542	Assert(!pMemLnx->fMappedToRing0);
543	#endif
544	pMemLnx->Core.pv = NULL;
545	}
546
547
548	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
549	{
550	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
551
552	/*
553	* Release any memory that we've allocated or locked.
554	*/
555	switch (pMemLnx->Core.enmType)
556	{
557	case RTR0MEMOBJTYPE_LOW:
558	case RTR0MEMOBJTYPE_PAGE:
559	case RTR0MEMOBJTYPE_CONT:
560	case RTR0MEMOBJTYPE_PHYS:
561	case RTR0MEMOBJTYPE_PHYS_NC:
562	rtR0MemObjLinuxVUnmap(pMemLnx);
563	rtR0MemObjLinuxFreePages(pMemLnx);
564	break;
565
566	case RTR0MEMOBJTYPE_LOCK:
567	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
568	{
569	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
570	size_t iPage;
571	Assert(pTask);
572	if (pTask && pTask->mm)
573	down_read(&pTask->mm->mmap_sem);
574
575	iPage = pMemLnx->cPages;
576	while (iPage-- > 0)
577	{
578	if (!PageReserved(pMemLnx->apPages[iPage]))
579	SetPageDirty(pMemLnx->apPages[iPage]);
580	page_cache_release(pMemLnx->apPages[iPage]);
581	}
582
583	if (pTask && pTask->mm)
584	up_read(&pTask->mm->mmap_sem);
585	}
586	/* else: kernel memory - nothing to do here. */
587	break;
588
589	case RTR0MEMOBJTYPE_RES_VIRT:
590	Assert(pMemLnx->Core.pv);
591	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
592	{
593	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
594	Assert(pTask);
595	if (pTask && pTask->mm)
596	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
597	}
598	else
599	{
600	vunmap(pMemLnx->Core.pv);
601
602	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
603	__free_page(pMemLnx->apPages[0]);
604	pMemLnx->apPages[0] = NULL;
605	pMemLnx->cPages = 0;
606	}
607	pMemLnx->Core.pv = NULL;
608	break;
609
610	case RTR0MEMOBJTYPE_MAPPING:
611	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
612	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
613	{
614	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
615	Assert(pTask);
616	if (pTask && pTask->mm)
617	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
618	}
619	else
620	vunmap(pMemLnx->Core.pv);
621	pMemLnx->Core.pv = NULL;
622	break;
623
624	default:
625	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
626	return VERR_INTERNAL_ERROR;
627	}
628	return VINF_SUCCESS;
629	}
630
631
632	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
633	{
634	PRTR0MEMOBJLNX pMemLnx;
635	int rc;
636
637	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
638	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
639	false /* non-contiguous */, VERR_NO_MEMORY);
640	#else
641	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
642	false /* non-contiguous */, VERR_NO_MEMORY);
643	#endif
644	if (RT_SUCCESS(rc))
645	{
646	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
647	if (RT_SUCCESS(rc))
648	{
649	*ppMem = &pMemLnx->Core;
650	return rc;
651	}
652
653	rtR0MemObjLinuxFreePages(pMemLnx);
654	rtR0MemObjDelete(&pMemLnx->Core);
655	}
656
657	return rc;
658	}
659
660
661	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
662	{
663	PRTR0MEMOBJLNX pMemLnx;
664	int rc;
665
666	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
667	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
668	/* ZONE_DMA32: 0-4GB */
669	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
670	false /* non-contiguous */, VERR_NO_LOW_MEMORY);
671	if (RT_FAILURE(rc))
672	#endif
673	#ifdef RT_ARCH_AMD64
674	/* ZONE_DMA: 0-16MB */
675	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
676	false /* non-contiguous */, VERR_NO_LOW_MEMORY);
677	#else
678	# ifdef CONFIG_X86_PAE
679	# endif
680	/* ZONE_NORMAL: 0-896MB */
681	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
682	false /* non-contiguous */, VERR_NO_LOW_MEMORY);
683	#endif
684	if (RT_SUCCESS(rc))
685	{
686	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
687	if (RT_SUCCESS(rc))
688	{
689	*ppMem = &pMemLnx->Core;
690	return rc;
691	}
692
693	rtR0MemObjLinuxFreePages(pMemLnx);
694	rtR0MemObjDelete(&pMemLnx->Core);
695	}
696
697	return rc;
698	}
699
700
701	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
702	{
703	PRTR0MEMOBJLNX pMemLnx;
704	int rc;
705
706	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
707	/* ZONE_DMA32: 0-4GB */
708	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
709	true /* contiguous */, VERR_NO_CONT_MEMORY);
710	if (RT_FAILURE(rc))
711	#endif
712	#ifdef RT_ARCH_AMD64
713	/* ZONE_DMA: 0-16MB */
714	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
715	true /* contiguous */, VERR_NO_CONT_MEMORY);
716	#else
717	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
718	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
719	true /* contiguous */, VERR_NO_CONT_MEMORY);
720	#endif
721	if (RT_SUCCESS(rc))
722	{
723	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
724	if (RT_SUCCESS(rc))
725	{
726	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
727	size_t iPage = pMemLnx->cPages;
728	while (iPage-- > 0)
729	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
730	#endif
731	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
732	*ppMem = &pMemLnx->Core;
733	return rc;
734	}
735
736	rtR0MemObjLinuxFreePages(pMemLnx);
737	rtR0MemObjDelete(&pMemLnx->Core);
738	}
739
740	return rc;
741	}
742
743
744	/**
745	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
746	*
747	* @returns IPRT status.
748	* @param ppMemLnx Where to
749	* @param enmType The object type.
750	* @param cb The size of the allocation.
751	* @param uAlignment The alignment of the physical memory.
752	* Only valid for fContiguous == true, ignored otherwise.
753	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
754	* @param fGfp The Linux GFP flags to use for the allocation.
755	*/
756	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
757	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, unsigned fGfp)
758	{
759	PRTR0MEMOBJLNX pMemLnx;
760	int rc;
761
762	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
763	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
764	VERR_NO_PHYS_MEMORY);
765	if (RT_FAILURE(rc))
766	return rc;
767
768	/*
769	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
770	*/
771	if (PhysHighest != NIL_RTHCPHYS)
772	{
773	size_t iPage = pMemLnx->cPages;
774	while (iPage-- > 0)
775	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
776	{
777	rtR0MemObjLinuxFreePages(pMemLnx);
778	rtR0MemObjDelete(&pMemLnx->Core);
779	return VERR_NO_MEMORY;
780	}
781	}
782
783	/*
784	* Complete the object.
785	*/
786	if (enmType == RTR0MEMOBJTYPE_PHYS)
787	{
788	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
789	pMemLnx->Core.u.Phys.fAllocated = true;
790	}
791	*ppMem = &pMemLnx->Core;
792	return rc;
793	}
794
795
796	/**
797	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
798	*
799	* @returns IPRT status.
800	* @param ppMem Where to store the memory object pointer on success.
801	* @param enmType The object type.
802	* @param cb The size of the allocation.
803	* @param uAlignment The alignment of the physical memory.
804	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
805	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
806	*/
807	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
808	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
809	{
810	int rc;
811
812	/*
813	* There are two clear cases and that's the <=16MB and anything-goes ones.
814	* When the physical address limit is somewhere in-between those two we'll
815	* just have to try, starting with HIGHUSER and working our way thru the
816	* different types, hoping we'll get lucky.
817	*
818	* We should probably move this physical address restriction logic up to
819	* the page alloc function as it would be more efficient there. But since
820	* we don't expect this to be a performance issue just yet it can wait.
821	*/
822	if (PhysHighest == NIL_RTHCPHYS)
823	/* ZONE_HIGHMEM: the whole physical memory */
824	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
825	else if (PhysHighest <= _1M * 16)
826	/* ZONE_DMA: 0-16MB */
827	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
828	else
829	{
830	rc = VERR_NO_MEMORY;
831	if (RT_FAILURE(rc))
832	/* ZONE_HIGHMEM: the whole physical memory */
833	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
834	if (RT_FAILURE(rc))
835	/* ZONE_NORMAL: 0-896MB */
836	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
837	#ifdef GFP_DMA32
838	if (RT_FAILURE(rc))
839	/* ZONE_DMA32: 0-4GB */
840	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
841	#endif
842	if (RT_FAILURE(rc))
843	/* ZONE_DMA: 0-16MB */
844	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
845	}
846	return rc;
847	}
848
849
850	/**
851	* Translates a kernel virtual address to a linux page structure by walking the
852	* page tables.
853	*
854	* @note We do assume that the page tables will not change as we are walking
855	* them. This assumption is rather forced by the fact that I could not
856	* immediately see any way of preventing this from happening. So, we
857	* take some extra care when accessing them.
858	*
859	* Because of this, we don't want to use this function on memory where
860	* attribute changes to nearby pages is likely to cause large pages to
861	* be used or split up. So, don't use this for the linear mapping of
862	* physical memory.
863	*
864	* @returns Pointer to the page structur or NULL if it could not be found.
865	* @param pv The kernel virtual address.
866	*/
867	static struct page rtR0MemObjLinuxVirtToPage(void pv)
868	{
869	unsigned long ulAddr = (unsigned long)pv;
870	unsigned long pfn;
871	struct page *pPage;
872	pte_t *pEntry;
873	union
874	{
875	pgd_t Global;
876	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
877	pud_t Upper;
878	#endif
879	pmd_t Middle;
880	pte_t Entry;
881	} u;
882
883	/* Should this happen in a situation this code will be called in? And if
884	* so, can it change under our feet? See also
885	* "Documentation/vm/active_mm.txt" in the kernel sources. */
886	if (RT_UNLIKELY(!current->active_mm))
887	return NULL;
888	u.Global = *pgd_offset(current->active_mm, ulAddr);
889	if (RT_UNLIKELY(pgd_none(u.Global)))
890	return NULL;
891
892	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
893	u.Upper = *pud_offset(&u.Global, ulAddr);
894	if (RT_UNLIKELY(pud_none(u.Upper)))
895	return NULL;
896	# if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 25)
897	if (pud_large(u.Upper))
898	{
899	pPage = pud_page(u.Upper);
900	AssertReturn(pPage, NULL);
901	pfn = page_to_pfn(pPage); /* doing the safe way... */
902	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
903	return pfn_to_page(pfn);
904	}
905	# endif
906
907	u.Middle = *pmd_offset(&u.Upper, ulAddr);
908	#else /* < 2.6.11 */
909	u.Middle = *pmd_offset(&u.Global, ulAddr);
910	#endif /* < 2.6.11 */
911	if (RT_UNLIKELY(pmd_none(u.Middle)))
912	return NULL;
913	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0)
914	if (pmd_large(u.Middle))
915	{
916	pPage = pmd_page(u.Middle);
917	AssertReturn(pPage, NULL);
918	pfn = page_to_pfn(pPage); /* doing the safe way... */
919	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
920	return pfn_to_page(pfn);
921	}
922	#endif
923
924	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
925	pEntry = pte_offset_map(&u.Middle, ulAddr);
926	#else
927	pEntry = pte_offset(&u.Middle, ulAddr);
928	#endif
929	if (RT_UNLIKELY(!pEntry))
930	return NULL;
931	u.Entry = *pEntry;
932	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 5) \|\| defined(pte_offset_map)
933	pte_unmap(pEntry);
934	#endif
935
936	if (RT_UNLIKELY(!pte_present(u.Entry)))
937	return NULL;
938	return pte_page(u.Entry);
939	}
940
941
942	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
943	{
944	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
945	}
946
947
948	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
949	{
950	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
951	}
952
953
954	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
955	{
956	/*
957	* All we need to do here is to validate that we can use
958	* ioremap on the specified address (32/64-bit dma_addr_t).
959	*/
960	PRTR0MEMOBJLNX pMemLnx;
961	dma_addr_t PhysAddr = Phys;
962	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
963
964	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
965	if (!pMemLnx)
966	return VERR_NO_MEMORY;
967
968	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
969	pMemLnx->Core.u.Phys.fAllocated = false;
970	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
971	Assert(!pMemLnx->cPages);
972	*ppMem = &pMemLnx->Core;
973	return VINF_SUCCESS;
974	}
975
976
977	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
978	{
979	const int cPages = cb >> PAGE_SHIFT;
980	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
981	struct vm_area_struct **papVMAs;
982	PRTR0MEMOBJLNX pMemLnx;
983	int rc = VERR_NO_MEMORY;
984	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
985
986	/*
987	* Check for valid task and size overflows.
988	*/
989	if (!pTask)
990	return VERR_NOT_SUPPORTED;
991	if (((size_t)cPages << PAGE_SHIFT) != cb)
992	return VERR_OUT_OF_RANGE;
993
994	/*
995	* Allocate the memory object and a temporary buffer for the VMAs.
996	*/
997	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
998	if (!pMemLnx)
999	return VERR_NO_MEMORY;
1000
1001	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1002	if (papVMAs)
1003	{
1004	down_read(&pTask->mm->mmap_sem);
1005
1006	/*
1007	* Get user pages.
1008	*/
1009	rc = get_user_pages(pTask, /* Task for fault accounting. */
1010	pTask->mm, /* Whose pages. */
1011	R3Ptr, /* Where from. */
1012	cPages, /* How many pages. */
1013	fWrite, /* Write to memory. */
1014	fWrite, /* force write access. */
1015	&pMemLnx->apPages[0], /* Page array. */
1016	papVMAs); /* vmas */
1017	if (rc == cPages)
1018	{
1019	/*
1020	* Flush dcache (required?), protect against fork and _really_ pin the page
1021	* table entries. get_user_pages() will protect against swapping out the
1022	* pages but it will NOT protect against removing page table entries. This
1023	* can be achieved with
1024	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1025	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1026	* Usual Linux distributions support only a limited size of locked pages
1027	* (e.g. 32KB).
1028	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1029	* or by
1030	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1031	* a range check.
1032	*/
1033	/** @todo The Linux fork() protection will require more work if this API
1034	* is to be used for anything but locking VM pages. */
1035	while (rc-- > 0)
1036	{
1037	flush_dcache_page(pMemLnx->apPages[rc]);
1038	papVMAs[rc]->vm_flags \|= (VM_DONTCOPY \| VM_LOCKED);
1039	}
1040
1041	up_read(&pTask->mm->mmap_sem);
1042
1043	RTMemFree(papVMAs);
1044
1045	pMemLnx->Core.u.Lock.R0Process = R0Process;
1046	pMemLnx->cPages = cPages;
1047	Assert(!pMemLnx->fMappedToRing0);
1048	*ppMem = &pMemLnx->Core;
1049
1050	return VINF_SUCCESS;
1051	}
1052
1053	/*
1054	* Failed - we need to unlock any pages that we succeeded to lock.
1055	*/
1056	while (rc-- > 0)
1057	{
1058	if (!PageReserved(pMemLnx->apPages[rc]))
1059	SetPageDirty(pMemLnx->apPages[rc]);
1060	page_cache_release(pMemLnx->apPages[rc]);
1061	}
1062
1063	up_read(&pTask->mm->mmap_sem);
1064
1065	RTMemFree(papVMAs);
1066	rc = VERR_LOCK_FAILED;
1067	}
1068
1069	rtR0MemObjDelete(&pMemLnx->Core);
1070	return rc;
1071	}
1072
1073
1074	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
1075	{
1076	void pvLast = (uint8_t )pv + cb - 1;
1077	size_t const cPages = cb >> PAGE_SHIFT;
1078	PRTR0MEMOBJLNX pMemLnx;
1079	bool fLinearMapping;
1080	int rc;
1081	uint8_t *pbPage;
1082	size_t iPage;
1083	NOREF(fAccess);
1084
1085	if ( !RTR0MemKernelIsValidAddr(pv)
1086	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1087	return VERR_INVALID_PARAMETER;
1088
1089	/*
1090	* The lower part of the kernel memory has a linear mapping between
1091	* physical and virtual addresses. So we take a short cut here. This is
1092	* assumed to be the cleanest way to handle those addresses (and the code
1093	* is well tested, though the test for determining it is not very nice).
1094	* If we ever decide it isn't we can still remove it.
1095	*/
1096	#if 0
1097	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1098	#else
1099	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1100	&& (unsigned long)pvLast < (unsigned long)high_memory;
1101	#endif
1102
1103	/*
1104	* Allocate the memory object.
1105	*/
1106	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
1107	if (!pMemLnx)
1108	return VERR_NO_MEMORY;
1109
1110	/*
1111	* Gather the pages.
1112	* We ASSUME all kernel pages are non-swappable and non-movable.
1113	*/
1114	rc = VINF_SUCCESS;
1115	pbPage = (uint8_t *)pvLast;
1116	iPage = cPages;
1117	if (!fLinearMapping)
1118	{
1119	while (iPage-- > 0)
1120	{
1121	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1122	if (RT_UNLIKELY(!pPage))
1123	{
1124	rc = VERR_LOCK_FAILED;
1125	break;
1126	}
1127	pMemLnx->apPages[iPage] = pPage;
1128	pbPage -= PAGE_SIZE;
1129	}
1130	}
1131	else
1132	{
1133	while (iPage-- > 0)
1134	{
1135	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1136	pbPage -= PAGE_SIZE;
1137	}
1138	}
1139	if (RT_SUCCESS(rc))
1140	{
1141	/*
1142	* Complete the memory object and return.
1143	*/
1144	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1145	pMemLnx->cPages = cPages;
1146	Assert(!pMemLnx->fMappedToRing0);
1147	*ppMem = &pMemLnx->Core;
1148
1149	return VINF_SUCCESS;
1150	}
1151
1152	rtR0MemObjDelete(&pMemLnx->Core);
1153	return rc;
1154	}
1155
1156
1157	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
1158	{
1159	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1160	const size_t cPages = cb >> PAGE_SHIFT;
1161	struct page *pDummyPage;
1162	struct page **papPages;
1163
1164	/* check for unsupported stuff. */
1165	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1166	if (uAlignment > PAGE_SIZE)
1167	return VERR_NOT_SUPPORTED;
1168
1169	/*
1170	* Allocate a dummy page and create a page pointer array for vmap such that
1171	* the dummy page is mapped all over the reserved area.
1172	*/
1173	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1174	if (!pDummyPage)
1175	return VERR_NO_MEMORY;
1176	papPages = RTMemAlloc(sizeof(papPages) cPages);
1177	if (papPages)
1178	{
1179	void *pv;
1180	size_t iPage = cPages;
1181	while (iPage-- > 0)
1182	papPages[iPage] = pDummyPage;
1183	# ifdef VM_MAP
1184	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1185	# else
1186	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1187	# endif
1188	RTMemFree(papPages);
1189	if (pv)
1190	{
1191	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1192	if (pMemLnx)
1193	{
1194	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1195	pMemLnx->cPages = 1;
1196	pMemLnx->apPages[0] = pDummyPage;
1197	*ppMem = &pMemLnx->Core;
1198	return VINF_SUCCESS;
1199	}
1200	vunmap(pv);
1201	}
1202	}
1203	__free_page(pDummyPage);
1204	return VERR_NO_MEMORY;
1205
1206	#else /* < 2.4.22 */
1207	/*
1208	* Could probably use ioremap here, but the caller is in a better position than us
1209	* to select some safe physical memory.
1210	*/
1211	return VERR_NOT_SUPPORTED;
1212	#endif
1213	}
1214
1215
1216	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1217	{
1218	PRTR0MEMOBJLNX pMemLnx;
1219	void *pv;
1220	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1221	if (!pTask)
1222	return VERR_NOT_SUPPORTED;
1223
1224	/*
1225	* Check that the specified alignment is supported.
1226	*/
1227	if (uAlignment > PAGE_SIZE)
1228	return VERR_NOT_SUPPORTED;
1229
1230	/*
1231	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1232	*/
1233	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1234	if (pv == (void *)-1)
1235	return VERR_NO_MEMORY;
1236
1237	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1238	if (!pMemLnx)
1239	{
1240	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1241	return VERR_NO_MEMORY;
1242	}
1243
1244	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1245	*ppMem = &pMemLnx->Core;
1246	return VINF_SUCCESS;
1247	}
1248
1249
1250	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap,
1251	void *pvFixed, size_t uAlignment,
1252	unsigned fProt, size_t offSub, size_t cbSub)
1253	{
1254	int rc = VERR_NO_MEMORY;
1255	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1256	PRTR0MEMOBJLNX pMemLnx;
1257
1258	/* Fail if requested to do something we can't. */
1259	AssertMsgReturn(!offSub && !cbSub, ("%#x %#x\n", offSub, cbSub), VERR_NOT_SUPPORTED);
1260	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1261	if (uAlignment > PAGE_SIZE)
1262	return VERR_NOT_SUPPORTED;
1263
1264	/*
1265	* Create the IPRT memory object.
1266	*/
1267	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1268	if (pMemLnx)
1269	{
1270	if (pMemLnxToMap->cPages)
1271	{
1272	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
1273	/*
1274	* Use vmap - 2.4.22 and later.
1275	*/
1276	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1277	# ifdef VM_MAP
1278	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
1279	# else
1280	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
1281	# endif
1282	if (pMemLnx->Core.pv)
1283	{
1284	pMemLnx->fMappedToRing0 = true;
1285	rc = VINF_SUCCESS;
1286	}
1287	else
1288	rc = VERR_MAP_FAILED;
1289
1290	#else /* < 2.4.22 */
1291	/*
1292	* Only option here is to share mappings if possible and forget about fProt.
1293	*/
1294	if (rtR0MemObjIsRing3(pMemToMap))
1295	rc = VERR_NOT_SUPPORTED;
1296	else
1297	{
1298	rc = VINF_SUCCESS;
1299	if (!pMemLnxToMap->Core.pv)
1300	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1301	if (RT_SUCCESS(rc))
1302	{
1303	Assert(pMemLnxToMap->Core.pv);
1304	pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
1305	}
1306	}
1307	#endif
1308	}
1309	else
1310	{
1311	/*
1312	* MMIO / physical memory.
1313	*/
1314	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1315	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1316	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb)
1317	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
1318	if (pMemLnx->Core.pv)
1319	{
1320	/** @todo fix protection. */
1321	rc = VINF_SUCCESS;
1322	}
1323	}
1324	if (RT_SUCCESS(rc))
1325	{
1326	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1327	*ppMem = &pMemLnx->Core;
1328	return VINF_SUCCESS;
1329	}
1330	rtR0MemObjDelete(&pMemLnx->Core);
1331	}
1332
1333	return rc;
1334	}
1335
1336
1337	#ifdef VBOX_USE_PAE_HACK
1338	/**
1339	* Replace the PFN of a PTE with the address of the actual page.
1340	*
1341	* The caller maps a reserved dummy page at the address with the desired access
1342	* and flags.
1343	*
1344	* This hack is required for older Linux kernels which don't provide
1345	* remap_pfn_range().
1346	*
1347	* @returns 0 on success, -ENOMEM on failure.
1348	* @param mm The memory context.
1349	* @param ulAddr The mapping address.
1350	* @param Phys The physical address of the page to map.
1351	*/
1352	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1353	{
1354	int rc = -ENOMEM;
1355	pgd_t *pgd;
1356
1357	spin_lock(&mm->page_table_lock);
1358
1359	pgd = pgd_offset(mm, ulAddr);
1360	if (!pgd_none(pgd) && !pgd_bad(pgd))
1361	{
1362	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1363	if (!pmd_none(*pmd))
1364	{
1365	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1366	if (ptep)
1367	{
1368	pte_t pte = *ptep;
1369	pte.pte_high &= 0xfff00000;
1370	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1371	pte.pte_low &= 0x00000fff;
1372	pte.pte_low \|= (Phys & 0xfffff000);
1373	set_pte(ptep, pte);
1374	pte_unmap(ptep);
1375	rc = 0;
1376	}
1377	}
1378	}
1379
1380	spin_unlock(&mm->page_table_lock);
1381	return rc;
1382	}
1383	#endif /* VBOX_USE_PAE_HACK */
1384
1385
1386	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed,
1387	size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1388	{
1389	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1390	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1391	int rc = VERR_NO_MEMORY;
1392	PRTR0MEMOBJLNX pMemLnx;
1393	#ifdef VBOX_USE_PAE_HACK
1394	struct page *pDummyPage;
1395	RTHCPHYS DummyPhys;
1396	#endif
1397
1398	/*
1399	* Check for restrictions.
1400	*/
1401	if (!pTask)
1402	return VERR_NOT_SUPPORTED;
1403	if (uAlignment > PAGE_SIZE)
1404	return VERR_NOT_SUPPORTED;
1405
1406	#ifdef VBOX_USE_PAE_HACK
1407	/*
1408	* Allocate a dummy page for use when mapping the memory.
1409	*/
1410	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1411	if (!pDummyPage)
1412	return VERR_NO_MEMORY;
1413	SetPageReserved(pDummyPage);
1414	DummyPhys = page_to_phys(pDummyPage);
1415	#endif
1416
1417	/*
1418	* Create the IPRT memory object.
1419	*/
1420	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1421	if (pMemLnx)
1422	{
1423	/*
1424	* Allocate user space mapping.
1425	*/
1426	void *pv;
1427	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1428	if (pv != (void *)-1)
1429	{
1430	/*
1431	* Map page by page into the mmap area.
1432	* This is generic, paranoid and not very efficient.
1433	*/
1434	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1435	unsigned long ulAddrCur = (unsigned long)pv;
1436	const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1437	size_t iPage;
1438
1439	down_write(&pTask->mm->mmap_sem);
1440
1441	rc = VINF_SUCCESS;
1442	if (pMemLnxToMap->cPages)
1443	{
1444	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1445	{
1446	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 11)
1447	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1448	#endif
1449	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1450	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1451	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1452	#endif
1453	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1454	/* remap_page_range() limitation on x86 */
1455	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1456	#endif
1457
1458	#if defined(VBOX_USE_INSERT_PAGE) && LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
1459	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1460	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1461	* See remap_pfn_range() in mm/memory.c */
1462	#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 7, 0)
1463	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1464	#else
1465	vma->vm_flags \|= VM_RESERVED;
1466	#endif
1467	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1468	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1469	#elif defined(VBOX_USE_PAE_HACK)
1470	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1471	if (!rc)
1472	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1473	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1474	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1475	#else /* 2.4 */
1476	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1477	#endif
1478	if (rc)
1479	{
1480	rc = VERR_NO_MEMORY;
1481	break;
1482	}
1483	}
1484	}
1485	else
1486	{
1487	RTHCPHYS Phys;
1488	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1489	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1490	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1491	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1492	else
1493	{
1494	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1495	Phys = NIL_RTHCPHYS;
1496	}
1497	if (Phys != NIL_RTHCPHYS)
1498	{
1499	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1500	{
1501	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1502	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1503	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1504	#endif
1505	#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 0) && defined(RT_ARCH_X86)
1506	/* remap_page_range() limitation on x86 */
1507	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1508	#endif
1509
1510	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1511	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1512	#elif defined(VBOX_USE_PAE_HACK)
1513	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1514	if (!rc)
1515	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1516	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1517	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1518	#else /* 2.4 */
1519	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1520	#endif
1521	if (rc)
1522	{
1523	rc = VERR_NO_MEMORY;
1524	break;
1525	}
1526	}
1527	}
1528	}
1529
1530	#ifdef CONFIG_NUMA_BALANCING
1531	# if LINUX_VERSION_CODE < KERNEL_VERSION(3, 13, 0)
1532	# ifdef RHEL_RELEASE_CODE
1533	# if RHEL_RELEASE_CODE < RHEL_RELEASE_VERSION(7, 0)
1534	# define VBOX_NUMA_HACK_OLD
1535	# endif
1536	# endif
1537	# endif
1538	if (RT_SUCCESS(rc))
1539	{
1540	/** @todo Ugly hack! But right now we have no other means to
1541	* disable automatic NUMA page balancing. */
1542	# ifdef RT_OS_X86
1543	# ifdef VBOX_NUMA_HACK_OLD
1544	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1545	# endif
1546	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1547	# else
1548	# ifdef VBOX_NUMA_HACK_OLD
1549	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1550	# endif
1551	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1552	# endif
1553	}
1554	#endif /* CONFIG_NUMA_BALANCING */
1555
1556	up_write(&pTask->mm->mmap_sem);
1557
1558	if (RT_SUCCESS(rc))
1559	{
1560	#ifdef VBOX_USE_PAE_HACK
1561	__free_page(pDummyPage);
1562	#endif
1563	pMemLnx->Core.pv = pv;
1564	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1565	*ppMem = &pMemLnx->Core;
1566	return VINF_SUCCESS;
1567	}
1568
1569	/*
1570	* Bail out.
1571	*/
1572	rtR0MemObjLinuxDoMunmap(pv, pMemLnxToMap->Core.cb, pTask);
1573	}
1574	rtR0MemObjDelete(&pMemLnx->Core);
1575	}
1576	#ifdef VBOX_USE_PAE_HACK
1577	__free_page(pDummyPage);
1578	#endif
1579
1580	return rc;
1581	}
1582
1583
1584	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1585	{
1586	NOREF(pMem);
1587	NOREF(offSub);
1588	NOREF(cbSub);
1589	NOREF(fProt);
1590	return VERR_NOT_SUPPORTED;
1591	}
1592
1593
1594	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1595	{
1596	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1597
1598	if (pMemLnx->cPages)
1599	return page_to_phys(pMemLnx->apPages[iPage]);
1600
1601	switch (pMemLnx->Core.enmType)
1602	{
1603	case RTR0MEMOBJTYPE_CONT:
1604	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1605
1606	case RTR0MEMOBJTYPE_PHYS:
1607	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1608
1609	/* the parent knows */
1610	case RTR0MEMOBJTYPE_MAPPING:
1611	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1612
1613	/* cPages > 0 */
1614	case RTR0MEMOBJTYPE_LOW:
1615	case RTR0MEMOBJTYPE_LOCK:
1616	case RTR0MEMOBJTYPE_PHYS_NC:
1617	case RTR0MEMOBJTYPE_PAGE:
1618	default:
1619	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1620	/* fall thru */
1621
1622	case RTR0MEMOBJTYPE_RES_VIRT:
1623	return NIL_RTHCPHYS;
1624	}
1625	}
1626

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 53402

Download in other formats: