memobj-r0drv-linux.c@ 90249

Last change on this file since 90249 was 90249, checked in by vboxsync, 3 years ago

Linux: fix build when -Wimplicit-fallthrough=5 is enabled, bugref:10066.

Compiler option -Wimplicit-fallthrough=5 was added to KBUILD_CFLAGS
in kernel 5.14. This commit replaces /* fall thru */ comment with
the macro RT_FALL_THRU() in a couple of remaining places, so the warning
can be suppressed.

Property svn:eol-style set to native
Property svn:keywords set to Id Rev Revision

File size: 65.7 KB

Line
1	/* $Id: memobj-r0drv-linux.c 90249 2021-07-20 07:59:02Z vboxsync $ */
2	/** @file
3	* IPRT - Ring-0 Memory Objects, Linux.
4	*/
5
6	/*
7	* Copyright (C) 2006-2020 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/assert.h>
35	#include <iprt/err.h>
36	#include <iprt/log.h>
37	#include <iprt/mem.h>
38	#include <iprt/process.h>
39	#include <iprt/string.h>
40	#include "internal/memobj.h"
41	#include "internal/iprt.h"
42
43
44	/*********************************************************************************************************************************
45	* Defined Constants And Macros *
46	*********************************************************************************************************************************/
47	/* early 2.6 kernels */
48	#ifndef PAGE_SHARED_EXEC
49	# define PAGE_SHARED_EXEC PAGE_SHARED
50	#endif
51	#ifndef PAGE_READONLY_EXEC
52	# define PAGE_READONLY_EXEC PAGE_READONLY
53	#endif
54
55	/** @def IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
56	* Whether we use alloc_vm_area (3.2+) for executable memory.
57	* This is a must for 5.8+, but we enable it all the way back to 3.2.x for
58	* better W^R compliance (fExecutable flag). */
59	#if RTLNX_VER_RANGE(3,2,0, 5,10,0) \|\| defined(DOXYGEN_RUNNING)
60	# define IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
61	#endif
62	/** @def IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
63	* alloc_vm_area was removed with 5.10 so we have to resort to a different way
64	* to allocate executable memory.
65	* It would be possible to remove IPRT_USE_ALLOC_VM_AREA_FOR_EXEC and use
66	* this path execlusively for 3.2+ but no time to test it really works on every
67	* supported kernel, so better play safe for now.
68	*/
69	#if RTLNX_VER_MIN(5,10,0) \|\| defined(DOXYGEN_RUNNING)
70	# define IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
71	#endif
72
73	/*
74	* 2.6.29+ kernels don't work with remap_pfn_range() anymore because
75	* track_pfn_vma_new() is apparently not defined for non-RAM pages.
76	* It should be safe to use vm_insert_page() older kernels as well.
77	*/
78	#if RTLNX_VER_MIN(2,6,23)
79	# define VBOX_USE_INSERT_PAGE
80	#endif
81	#if defined(CONFIG_X86_PAE) \
82	&& ( defined(HAVE_26_STYLE_REMAP_PAGE_RANGE) \
83	\|\| RTLNX_VER_RANGE(2,6,0, 2,6,11) )
84	# define VBOX_USE_PAE_HACK
85	#endif
86
87	/* gfp_t was introduced in 2.6.14, define it for earlier. */
88	#if RTLNX_VER_MAX(2,6,14)
89	# define gfp_t unsigned
90	#endif
91
92	/*
93	* Wrappers around mmap_lock/mmap_sem difference.
94	*/
95	#if RTLNX_VER_MIN(5,8,0)
96	# define LNX_MM_DOWN_READ(a_pMm) down_read(&(a_pMm)->mmap_lock)
97	# define LNX_MM_UP_READ(a_pMm) up_read(&(a_pMm)->mmap_lock)
98	# define LNX_MM_DOWN_WRITE(a_pMm) down_write(&(a_pMm)->mmap_lock)
99	# define LNX_MM_UP_WRITE(a_pMm) up_write(&(a_pMm)->mmap_lock)
100	#else
101	# define LNX_MM_DOWN_READ(a_pMm) down_read(&(a_pMm)->mmap_sem)
102	# define LNX_MM_UP_READ(a_pMm) up_read(&(a_pMm)->mmap_sem)
103	# define LNX_MM_DOWN_WRITE(a_pMm) down_write(&(a_pMm)->mmap_sem)
104	# define LNX_MM_UP_WRITE(a_pMm) up_write(&(a_pMm)->mmap_sem)
105	#endif
106
107
108	/*********************************************************************************************************************************
109	* Structures and Typedefs *
110	*********************************************************************************************************************************/
111	/**
112	* The Linux version of the memory object structure.
113	*/
114	typedef struct RTR0MEMOBJLNX
115	{
116	/** The core structure. */
117	RTR0MEMOBJINTERNAL Core;
118	/** Set if the allocation is contiguous.
119	* This means it has to be given back as one chunk. */
120	bool fContiguous;
121	/** Set if executable allocation. */
122	bool fExecutable;
123	/** Set if we've vmap'ed the memory into ring-0. */
124	bool fMappedToRing0;
125	#ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
126	/** Return from alloc_vm_area() that we now need to use for executable
127	* memory. */
128	struct vm_struct *pArea;
129	/** PTE array that goes along with pArea (must be freed). */
130	pte_t **papPtesForArea;
131	#endif
132	/** The pages in the apPages array. */
133	size_t cPages;
134	/** Array of struct page pointers. (variable size) */
135	struct page *apPages[1];
136	} RTR0MEMOBJLNX;
137	/** Pointer to the linux memory object. */
138	typedef RTR0MEMOBJLNX *PRTR0MEMOBJLNX;
139
140
141	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx);
142
143
144	/**
145	* Helper that converts from a RTR0PROCESS handle to a linux task.
146	*
147	* @returns The corresponding Linux task.
148	* @param R0Process IPRT ring-0 process handle.
149	*/
150	static struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
151	{
152	/** @todo fix rtR0ProcessToLinuxTask!! */
153	/** @todo many (all?) callers currently assume that we return 'current'! */
154	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
155	}
156
157
158	/**
159	* Compute order. Some functions allocate 2^order pages.
160	*
161	* @returns order.
162	* @param cPages Number of pages.
163	*/
164	static int rtR0MemObjLinuxOrder(size_t cPages)
165	{
166	int iOrder;
167	size_t cTmp;
168
169	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
170	;
171	if (cPages & ~((size_t)1 << iOrder))
172	++iOrder;
173
174	return iOrder;
175	}
176
177
178	/**
179	* Converts from RTMEM_PROT_* to Linux PAGE_*.
180	*
181	* @returns Linux page protection constant.
182	* @param fProt The IPRT protection mask.
183	* @param fKernel Whether it applies to kernel or user space.
184	*/
185	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
186	{
187	switch (fProt)
188	{
189	default:
190	AssertMsgFailed(("%#x %d\n", fProt, fKernel)); RT_FALL_THRU();
191	case RTMEM_PROT_NONE:
192	return PAGE_NONE;
193
194	case RTMEM_PROT_READ:
195	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
196
197	case RTMEM_PROT_WRITE:
198	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
199	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
200
201	case RTMEM_PROT_EXEC:
202	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
203	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
204	if (fKernel)
205	{
206	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
207	pgprot_val(fPg) &= ~_PAGE_RW;
208	return fPg;
209	}
210	return PAGE_READONLY_EXEC;
211	#else
212	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
213	#endif
214
215	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
216	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
217	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
218	}
219	}
220
221
222	/**
223	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
224	* an empty user space mapping.
225	*
226	* We acquire the mmap_sem/mmap_lock of the task!
227	*
228	* @returns Pointer to the mapping.
229	* (void *)-1 on failure.
230	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
231	* @param cb The size of the mapping.
232	* @param uAlignment The alignment of the mapping.
233	* @param pTask The Linux task to create this mapping in.
234	* @param fProt The RTMEM_PROT_* mask.
235	*/
236	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
237	{
238	unsigned fLnxProt;
239	unsigned long ulAddr;
240
241	Assert(pTask == current); /* do_mmap */
242	RT_NOREF_PV(pTask);
243
244	/*
245	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
246	*/
247	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
248	if (fProt == RTMEM_PROT_NONE)
249	fLnxProt = PROT_NONE;
250	else
251	{
252	fLnxProt = 0;
253	if (fProt & RTMEM_PROT_READ)
254	fLnxProt \|= PROT_READ;
255	if (fProt & RTMEM_PROT_WRITE)
256	fLnxProt \|= PROT_WRITE;
257	if (fProt & RTMEM_PROT_EXEC)
258	fLnxProt \|= PROT_EXEC;
259	}
260
261	if (R3PtrFixed != (RTR3PTR)-1)
262	{
263	#if RTLNX_VER_MIN(3,5,0)
264	ulAddr = vm_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
265	#else
266	LNX_MM_DOWN_WRITE(pTask->mm);
267	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
268	LNX_MM_UP_WRITE(pTask->mm);
269	#endif
270	}
271	else
272	{
273	#if RTLNX_VER_MIN(3,5,0)
274	ulAddr = vm_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
275	#else
276	LNX_MM_DOWN_WRITE(pTask->mm);
277	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
278	LNX_MM_UP_WRITE(pTask->mm);
279	#endif
280	if ( !(ulAddr & ~PAGE_MASK)
281	&& (ulAddr & (uAlignment - 1)))
282	{
283	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
284	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
285	* ourselves) and further by there begin two mmap strategies (top / bottom). */
286	/* For now, just ignore uAlignment requirements... */
287	}
288	}
289
290
291	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
292	return (void *)-1;
293	return (void *)ulAddr;
294	}
295
296
297	/**
298	* Worker that destroys a user space mapping.
299	* Undoes what rtR0MemObjLinuxDoMmap did.
300	*
301	* We acquire the mmap_sem/mmap_lock of the task!
302	*
303	* @param pv The ring-3 mapping.
304	* @param cb The size of the mapping.
305	* @param pTask The Linux task to destroy this mapping in.
306	*/
307	static void rtR0MemObjLinuxDoMunmap(void pv, size_t cb, struct task_struct pTask)
308	{
309	#if RTLNX_VER_MIN(3,5,0)
310	Assert(pTask == current); RT_NOREF_PV(pTask);
311	vm_munmap((unsigned long)pv, cb);
312	#elif defined(USE_RHEL4_MUNMAP)
313	LNX_MM_DOWN_WRITE(pTask->mm);
314	do_munmap(pTask->mm, (unsigned long)pv, cb, 0); /* should it be 1 or 0? */
315	LNX_MM_UP_WRITE(pTask->mm);
316	#else
317	LNX_MM_DOWN_WRITE(pTask->mm);
318	do_munmap(pTask->mm, (unsigned long)pv, cb);
319	LNX_MM_UP_WRITE(pTask->mm);
320	#endif
321	}
322
323
324	/**
325	* Internal worker that allocates physical pages and creates the memory object for them.
326	*
327	* @returns IPRT status code.
328	* @param ppMemLnx Where to store the memory object pointer.
329	* @param enmType The object type.
330	* @param cb The number of bytes to allocate.
331	* @param uAlignment The alignment of the physical memory.
332	* Only valid if fContiguous == true, ignored otherwise.
333	* @param fFlagsLnx The page allocation flags (GPFs).
334	* @param fContiguous Whether the allocation must be contiguous.
335	* @param fExecutable Whether the memory must be executable.
336	* @param rcNoMem What to return when we're out of pages.
337	*/
338	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
339	size_t uAlignment, gfp_t fFlagsLnx, bool fContiguous, bool fExecutable, int rcNoMem)
340	{
341	size_t iPage;
342	size_t const cPages = cb >> PAGE_SHIFT;
343	struct page *paPages;
344
345	/*
346	* Allocate a memory object structure that's large enough to contain
347	* the page pointer array.
348	*/
349	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
350	if (!pMemLnx)
351	return VERR_NO_MEMORY;
352	pMemLnx->cPages = cPages;
353
354	if (cPages > 255)
355	{
356	# ifdef __GFP_REPEAT
357	/* Try hard to allocate the memory, but the allocation attempt might fail. */
358	fFlagsLnx \|= __GFP_REPEAT;
359	# endif
360	# ifdef __GFP_NOMEMALLOC
361	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
362	fFlagsLnx \|= __GFP_NOMEMALLOC;
363	# endif
364	}
365
366	/*
367	* Allocate the pages.
368	* For small allocations we'll try contiguous first and then fall back on page by page.
369	*/
370	#if RTLNX_VER_MIN(2,4,22)
371	if ( fContiguous
372	\|\| cb <= PAGE_SIZE * 2)
373	{
374	# ifdef VBOX_USE_INSERT_PAGE
375	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
376	# else
377	paPages = alloc_pages(fFlagsLnx \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
378	# endif
379	if (paPages)
380	{
381	fContiguous = true;
382	for (iPage = 0; iPage < cPages; iPage++)
383	pMemLnx->apPages[iPage] = &paPages[iPage];
384	}
385	else if (fContiguous)
386	{
387	rtR0MemObjDelete(&pMemLnx->Core);
388	return rcNoMem;
389	}
390	}
391
392	if (!fContiguous)
393	{
394	for (iPage = 0; iPage < cPages; iPage++)
395	{
396	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
397	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
398	{
399	while (iPage-- > 0)
400	__free_page(pMemLnx->apPages[iPage]);
401	rtR0MemObjDelete(&pMemLnx->Core);
402	return rcNoMem;
403	}
404	}
405	}
406
407	#else /* < 2.4.22 */
408	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
409	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
410	if (!paPages)
411	{
412	rtR0MemObjDelete(&pMemLnx->Core);
413	return rcNoMem;
414	}
415	for (iPage = 0; iPage < cPages; iPage++)
416	{
417	pMemLnx->apPages[iPage] = &paPages[iPage];
418	if (fExecutable)
419	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
420	if (PageHighMem(pMemLnx->apPages[iPage]))
421	BUG();
422	}
423
424	fContiguous = true;
425	#endif /* < 2.4.22 */
426	pMemLnx->fContiguous = fContiguous;
427	pMemLnx->fExecutable = fExecutable;
428
429	#if RTLNX_VER_MAX(4,5,0)
430	/*
431	* Reserve the pages.
432	*
433	* Linux >= 4.5 with CONFIG_DEBUG_VM panics when setting PG_reserved on compound
434	* pages. According to Michal Hocko this shouldn't be necessary anyway because
435	* as pages which are not on the LRU list are never evictable.
436	*/
437	for (iPage = 0; iPage < cPages; iPage++)
438	SetPageReserved(pMemLnx->apPages[iPage]);
439	#endif
440
441	/*
442	* Note that the physical address of memory allocated with alloc_pages(flags, order)
443	* is always 2^(PAGE_SHIFT+order)-aligned.
444	*/
445	if ( fContiguous
446	&& uAlignment > PAGE_SIZE)
447	{
448	/*
449	* Check for alignment constraints. The physical address of memory allocated with
450	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
451	*/
452	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
453	{
454	/*
455	* This should never happen!
456	*/
457	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
458	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
459	rtR0MemObjLinuxFreePages(pMemLnx);
460	return rcNoMem;
461	}
462	}
463
464	*ppMemLnx = pMemLnx;
465	return VINF_SUCCESS;
466	}
467
468
469	/**
470	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
471	*
472	* This method does NOT free the object.
473	*
474	* @param pMemLnx The object which physical pages should be freed.
475	*/
476	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
477	{
478	size_t iPage = pMemLnx->cPages;
479	if (iPage > 0)
480	{
481	/*
482	* Restore the page flags.
483	*/
484	while (iPage-- > 0)
485	{
486	#if RTLNX_VER_MAX(4,5,0)
487	/* See SetPageReserved() in rtR0MemObjLinuxAllocPages() */
488	ClearPageReserved(pMemLnx->apPages[iPage]);
489	#endif
490	#if RTLNX_VER_MAX(2,4,22)
491	if (pMemLnx->fExecutable)
492	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
493	#endif
494	}
495
496	/*
497	* Free the pages.
498	*/
499	#if RTLNX_VER_MIN(2,4,22)
500	if (!pMemLnx->fContiguous)
501	{
502	iPage = pMemLnx->cPages;
503	while (iPage-- > 0)
504	__free_page(pMemLnx->apPages[iPage]);
505	}
506	else
507	#endif
508	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
509
510	pMemLnx->cPages = 0;
511	}
512	}
513
514
515	#ifdef IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
516	/**
517	* User data passed to the apply_to_page_range() callback.
518	*/
519	typedef struct LNXAPPLYPGRANGE
520	{
521	/** Pointer to the memory object. */
522	PRTR0MEMOBJLNX pMemLnx;
523	/** The page protection flags to apply. */
524	pgprot_t fPg;
525	} LNXAPPLYPGRANGE;
526	/** Pointer to the user data. */
527	typedef LNXAPPLYPGRANGE *PLNXAPPLYPGRANGE;
528	/** Pointer to the const user data. */
529	typedef const LNXAPPLYPGRANGE *PCLNXAPPLYPGRANGE;
530
531	/**
532	* Callback called in apply_to_page_range().
533	*
534	* @returns Linux status code.
535	* @param pPte Pointer to the page table entry for the given address.
536	* @param uAddr The address to apply the new protection to.
537	* @param pvUser The opaque user data.
538	*/
539	static int rtR0MemObjLinuxApplyPageRange(pte_t pPte, unsigned long uAddr, void pvUser)
540	{
541	PCLNXAPPLYPGRANGE pArgs = (PCLNXAPPLYPGRANGE)pvUser;
542	PRTR0MEMOBJLNX pMemLnx = pArgs->pMemLnx;
543	size_t idxPg = (uAddr - (unsigned long)pMemLnx->Core.pv) >> PAGE_SHIFT;
544
545	set_pte(pPte, mk_pte(pMemLnx->apPages[idxPg], pArgs->fPg));
546	return 0;
547	}
548	#endif
549
550
551	/**
552	* Maps the allocation into ring-0.
553	*
554	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
555	*
556	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
557	* space, so we'll use that mapping if possible. If execute access is required, we'll
558	* play safe and do our own mapping.
559	*
560	* @returns IPRT status code.
561	* @param pMemLnx The linux memory object to map.
562	* @param fExecutable Whether execute access is required.
563	*/
564	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
565	{
566	int rc = VINF_SUCCESS;
567
568	/*
569	* Choose mapping strategy.
570	*/
571	bool fMustMap = fExecutable
572	\|\| !pMemLnx->fContiguous;
573	if (!fMustMap)
574	{
575	size_t iPage = pMemLnx->cPages;
576	while (iPage-- > 0)
577	if (PageHighMem(pMemLnx->apPages[iPage]))
578	{
579	fMustMap = true;
580	break;
581	}
582	}
583
584	Assert(!pMemLnx->Core.pv);
585	Assert(!pMemLnx->fMappedToRing0);
586
587	if (fMustMap)
588	{
589	/*
590	* Use vmap - 2.4.22 and later.
591	*/
592	#if RTLNX_VER_MIN(2,4,22)
593	pgprot_t fPg;
594	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
595	# ifdef _PAGE_NX
596	if (!fExecutable)
597	pgprot_val(fPg) \|= _PAGE_NX;
598	# endif
599
600	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
601	if (fExecutable)
602	{
603	# if RTLNX_VER_MIN(3,2,51)
604	pte_t papPtes = (pte_t )kmalloc_array(pMemLnx->cPages, sizeof(papPtes[0]), GFP_KERNEL);
605	# else
606	pte_t papPtes = (pte_t )kmalloc(pMemLnx->cPages * sizeof(papPtes[0]), GFP_KERNEL);
607	# endif
608	if (papPtes)
609	{
610	pMemLnx->pArea = alloc_vm_area(pMemLnx->Core.cb, papPtes); /* Note! pArea->nr_pages is not set. */
611	if (pMemLnx->pArea)
612	{
613	size_t i;
614	Assert(pMemLnx->pArea->size >= pMemLnx->Core.cb); /* Note! includes guard page. */
615	Assert(pMemLnx->pArea->addr);
616	# ifdef _PAGE_NX
617	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
618	# endif
619	pMemLnx->papPtesForArea = papPtes;
620	for (i = 0; i < pMemLnx->cPages; i++)
621	*papPtes[i] = mk_pte(pMemLnx->apPages[i], fPg);
622	pMemLnx->Core.pv = pMemLnx->pArea->addr;
623	pMemLnx->fMappedToRing0 = true;
624	}
625	else
626	{
627	kfree(papPtes);
628	rc = VERR_MAP_FAILED;
629	}
630	}
631	else
632	rc = VERR_MAP_FAILED;
633	}
634	else
635	# endif
636	{
637	# if defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
638	if (fExecutable)
639	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
640	# endif
641
642	# ifdef VM_MAP
643	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
644	# else
645	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
646	# endif
647	if (pMemLnx->Core.pv)
648	pMemLnx->fMappedToRing0 = true;
649	else
650	rc = VERR_MAP_FAILED;
651	}
652	#else /* < 2.4.22 */
653	rc = VERR_NOT_SUPPORTED;
654	#endif
655	}
656	else
657	{
658	/*
659	* Use the kernel RAM mapping.
660	*/
661	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
662	Assert(pMemLnx->Core.pv);
663	}
664
665	return rc;
666	}
667
668
669	/**
670	* Undoes what rtR0MemObjLinuxVMap() did.
671	*
672	* @param pMemLnx The linux memory object.
673	*/
674	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
675	{
676	#if RTLNX_VER_MIN(2,4,22)
677	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
678	if (pMemLnx->pArea)
679	{
680	# if 0
681	pte_t **papPtes = pMemLnx->papPtesForArea;
682	size_t i;
683	for (i = 0; i < pMemLnx->cPages; i++)
684	*papPtes[i] = 0;
685	# endif
686	free_vm_area(pMemLnx->pArea);
687	kfree(pMemLnx->papPtesForArea);
688	pMemLnx->pArea = NULL;
689	pMemLnx->papPtesForArea = NULL;
690	}
691	else
692	# endif
693	if (pMemLnx->fMappedToRing0)
694	{
695	Assert(pMemLnx->Core.pv);
696	vunmap(pMemLnx->Core.pv);
697	pMemLnx->fMappedToRing0 = false;
698	}
699	#else /* < 2.4.22 */
700	Assert(!pMemLnx->fMappedToRing0);
701	#endif
702	pMemLnx->Core.pv = NULL;
703	}
704
705
706	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
707	{
708	IPRT_LINUX_SAVE_EFL_AC();
709	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
710
711	/*
712	* Release any memory that we've allocated or locked.
713	*/
714	switch (pMemLnx->Core.enmType)
715	{
716	case RTR0MEMOBJTYPE_LOW:
717	case RTR0MEMOBJTYPE_PAGE:
718	case RTR0MEMOBJTYPE_CONT:
719	case RTR0MEMOBJTYPE_PHYS:
720	case RTR0MEMOBJTYPE_PHYS_NC:
721	rtR0MemObjLinuxVUnmap(pMemLnx);
722	rtR0MemObjLinuxFreePages(pMemLnx);
723	break;
724
725	case RTR0MEMOBJTYPE_LOCK:
726	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
727	{
728	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
729	size_t iPage;
730	Assert(pTask);
731	if (pTask && pTask->mm)
732	LNX_MM_DOWN_READ(pTask->mm);
733
734	iPage = pMemLnx->cPages;
735	while (iPage-- > 0)
736	{
737	if (!PageReserved(pMemLnx->apPages[iPage]))
738	SetPageDirty(pMemLnx->apPages[iPage]);
739	#if RTLNX_VER_MIN(4,6,0)
740	put_page(pMemLnx->apPages[iPage]);
741	#else
742	page_cache_release(pMemLnx->apPages[iPage]);
743	#endif
744	}
745
746	if (pTask && pTask->mm)
747	LNX_MM_UP_READ(pTask->mm);
748	}
749	/* else: kernel memory - nothing to do here. */
750	break;
751
752	case RTR0MEMOBJTYPE_RES_VIRT:
753	Assert(pMemLnx->Core.pv);
754	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
755	{
756	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
757	Assert(pTask);
758	if (pTask && pTask->mm)
759	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
760	}
761	else
762	{
763	vunmap(pMemLnx->Core.pv);
764
765	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
766	__free_page(pMemLnx->apPages[0]);
767	pMemLnx->apPages[0] = NULL;
768	pMemLnx->cPages = 0;
769	}
770	pMemLnx->Core.pv = NULL;
771	break;
772
773	case RTR0MEMOBJTYPE_MAPPING:
774	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
775	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
776	{
777	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
778	Assert(pTask);
779	if (pTask && pTask->mm)
780	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
781	}
782	else
783	vunmap(pMemLnx->Core.pv);
784	pMemLnx->Core.pv = NULL;
785	break;
786
787	default:
788	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
789	return VERR_INTERNAL_ERROR;
790	}
791	IPRT_LINUX_RESTORE_EFL_ONLY_AC();
792	return VINF_SUCCESS;
793	}
794
795
796	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
797	{
798	IPRT_LINUX_SAVE_EFL_AC();
799	PRTR0MEMOBJLNX pMemLnx;
800	int rc;
801
802	#if RTLNX_VER_MIN(2,4,22)
803	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
804	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY);
805	#else
806	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
807	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY);
808	#endif
809	if (RT_SUCCESS(rc))
810	{
811	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
812	if (RT_SUCCESS(rc))
813	{
814	*ppMem = &pMemLnx->Core;
815	IPRT_LINUX_RESTORE_EFL_AC();
816	return rc;
817	}
818
819	rtR0MemObjLinuxFreePages(pMemLnx);
820	rtR0MemObjDelete(&pMemLnx->Core);
821	}
822
823	IPRT_LINUX_RESTORE_EFL_AC();
824	return rc;
825	}
826
827
828	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
829	{
830	IPRT_LINUX_SAVE_EFL_AC();
831	PRTR0MEMOBJLNX pMemLnx;
832	int rc;
833
834	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
835	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
836	/* ZONE_DMA32: 0-4GB */
837	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
838	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY);
839	if (RT_FAILURE(rc))
840	#endif
841	#ifdef RT_ARCH_AMD64
842	/* ZONE_DMA: 0-16MB */
843	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
844	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY);
845	#else
846	# ifdef CONFIG_X86_PAE
847	# endif
848	/* ZONE_NORMAL: 0-896MB */
849	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
850	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY);
851	#endif
852	if (RT_SUCCESS(rc))
853	{
854	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
855	if (RT_SUCCESS(rc))
856	{
857	*ppMem = &pMemLnx->Core;
858	IPRT_LINUX_RESTORE_EFL_AC();
859	return rc;
860	}
861
862	rtR0MemObjLinuxFreePages(pMemLnx);
863	rtR0MemObjDelete(&pMemLnx->Core);
864	}
865
866	IPRT_LINUX_RESTORE_EFL_AC();
867	return rc;
868	}
869
870
871	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
872	{
873	IPRT_LINUX_SAVE_EFL_AC();
874	PRTR0MEMOBJLNX pMemLnx;
875	int rc;
876
877	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
878	/* ZONE_DMA32: 0-4GB */
879	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
880	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY);
881	if (RT_FAILURE(rc))
882	#endif
883	#ifdef RT_ARCH_AMD64
884	/* ZONE_DMA: 0-16MB */
885	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
886	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY);
887	#else
888	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
889	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
890	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY);
891	#endif
892	if (RT_SUCCESS(rc))
893	{
894	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
895	if (RT_SUCCESS(rc))
896	{
897	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
898	size_t iPage = pMemLnx->cPages;
899	while (iPage-- > 0)
900	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
901	#endif
902	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
903	*ppMem = &pMemLnx->Core;
904	IPRT_LINUX_RESTORE_EFL_AC();
905	return rc;
906	}
907
908	rtR0MemObjLinuxFreePages(pMemLnx);
909	rtR0MemObjDelete(&pMemLnx->Core);
910	}
911
912	IPRT_LINUX_RESTORE_EFL_AC();
913	return rc;
914	}
915
916
917	/**
918	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
919	*
920	* @returns IPRT status code.
921	* @param ppMemLnx Where to
922	* @param enmType The object type.
923	* @param cb The size of the allocation.
924	* @param uAlignment The alignment of the physical memory.
925	* Only valid for fContiguous == true, ignored otherwise.
926	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
927	* @param fGfp The Linux GFP flags to use for the allocation.
928	*/
929	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
930	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, gfp_t fGfp)
931	{
932	PRTR0MEMOBJLNX pMemLnx;
933	int rc;
934
935	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
936	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
937	false /fExecutable/, VERR_NO_PHYS_MEMORY);
938	if (RT_FAILURE(rc))
939	return rc;
940
941	/*
942	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
943	*/
944	if (PhysHighest != NIL_RTHCPHYS)
945	{
946	size_t iPage = pMemLnx->cPages;
947	while (iPage-- > 0)
948	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
949	{
950	rtR0MemObjLinuxFreePages(pMemLnx);
951	rtR0MemObjDelete(&pMemLnx->Core);
952	return VERR_NO_MEMORY;
953	}
954	}
955
956	/*
957	* Complete the object.
958	*/
959	if (enmType == RTR0MEMOBJTYPE_PHYS)
960	{
961	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
962	pMemLnx->Core.u.Phys.fAllocated = true;
963	}
964	*ppMem = &pMemLnx->Core;
965	return rc;
966	}
967
968
969	/**
970	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
971	*
972	* @returns IPRT status code.
973	* @param ppMem Where to store the memory object pointer on success.
974	* @param enmType The object type.
975	* @param cb The size of the allocation.
976	* @param uAlignment The alignment of the physical memory.
977	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
978	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
979	*/
980	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
981	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
982	{
983	int rc;
984	IPRT_LINUX_SAVE_EFL_AC();
985
986	/*
987	* There are two clear cases and that's the <=16MB and anything-goes ones.
988	* When the physical address limit is somewhere in-between those two we'll
989	* just have to try, starting with HIGHUSER and working our way thru the
990	* different types, hoping we'll get lucky.
991	*
992	* We should probably move this physical address restriction logic up to
993	* the page alloc function as it would be more efficient there. But since
994	* we don't expect this to be a performance issue just yet it can wait.
995	*/
996	if (PhysHighest == NIL_RTHCPHYS)
997	/* ZONE_HIGHMEM: the whole physical memory */
998	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
999	else if (PhysHighest <= _1M * 16)
1000	/* ZONE_DMA: 0-16MB */
1001	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
1002	else
1003	{
1004	rc = VERR_NO_MEMORY;
1005	if (RT_FAILURE(rc))
1006	/* ZONE_HIGHMEM: the whole physical memory */
1007	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
1008	if (RT_FAILURE(rc))
1009	/* ZONE_NORMAL: 0-896MB */
1010	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
1011	#ifdef GFP_DMA32
1012	if (RT_FAILURE(rc))
1013	/* ZONE_DMA32: 0-4GB */
1014	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
1015	#endif
1016	if (RT_FAILURE(rc))
1017	/* ZONE_DMA: 0-16MB */
1018	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
1019	}
1020	IPRT_LINUX_RESTORE_EFL_AC();
1021	return rc;
1022	}
1023
1024
1025	/**
1026	* Translates a kernel virtual address to a linux page structure by walking the
1027	* page tables.
1028	*
1029	* @note We do assume that the page tables will not change as we are walking
1030	* them. This assumption is rather forced by the fact that I could not
1031	* immediately see any way of preventing this from happening. So, we
1032	* take some extra care when accessing them.
1033	*
1034	* Because of this, we don't want to use this function on memory where
1035	* attribute changes to nearby pages is likely to cause large pages to
1036	* be used or split up. So, don't use this for the linear mapping of
1037	* physical memory.
1038	*
1039	* @returns Pointer to the page structur or NULL if it could not be found.
1040	* @param pv The kernel virtual address.
1041	*/
1042	RTDECL(struct page ) rtR0MemObjLinuxVirtToPage(void pv)
1043	{
1044	unsigned long ulAddr = (unsigned long)pv;
1045	unsigned long pfn;
1046	struct page *pPage;
1047	pte_t *pEntry;
1048	union
1049	{
1050	pgd_t Global;
1051	#if RTLNX_VER_MIN(4,12,0)
1052	p4d_t Four;
1053	#endif
1054	#if RTLNX_VER_MIN(2,6,11)
1055	pud_t Upper;
1056	#endif
1057	pmd_t Middle;
1058	pte_t Entry;
1059	} u;
1060
1061	/* Should this happen in a situation this code will be called in? And if
1062	* so, can it change under our feet? See also
1063	* "Documentation/vm/active_mm.txt" in the kernel sources. */
1064	if (RT_UNLIKELY(!current->active_mm))
1065	return NULL;
1066	u.Global = *pgd_offset(current->active_mm, ulAddr);
1067	if (RT_UNLIKELY(pgd_none(u.Global)))
1068	return NULL;
1069	#if RTLNX_VER_MIN(2,6,11)
1070	# if RTLNX_VER_MIN(4,12,0)
1071	u.Four = *p4d_offset(&u.Global, ulAddr);
1072	if (RT_UNLIKELY(p4d_none(u.Four)))
1073	return NULL;
1074	if (p4d_large(u.Four))
1075	{
1076	pPage = p4d_page(u.Four);
1077	AssertReturn(pPage, NULL);
1078	pfn = page_to_pfn(pPage); /* doing the safe way... */
1079	AssertCompile(P4D_SHIFT - PAGE_SHIFT < 31);
1080	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (P4D_SHIFT - PAGE_SHIFT)) - 1);
1081	return pfn_to_page(pfn);
1082	}
1083	u.Upper = *pud_offset(&u.Four, ulAddr);
1084	# else /* < 4.12 */
1085	u.Upper = *pud_offset(&u.Global, ulAddr);
1086	# endif /* < 4.12 */
1087	if (RT_UNLIKELY(pud_none(u.Upper)))
1088	return NULL;
1089	# if RTLNX_VER_MIN(2,6,25)
1090	if (pud_large(u.Upper))
1091	{
1092	pPage = pud_page(u.Upper);
1093	AssertReturn(pPage, NULL);
1094	pfn = page_to_pfn(pPage); /* doing the safe way... */
1095	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
1096	return pfn_to_page(pfn);
1097	}
1098	# endif
1099	u.Middle = *pmd_offset(&u.Upper, ulAddr);
1100	#else /* < 2.6.11 */
1101	u.Middle = *pmd_offset(&u.Global, ulAddr);
1102	#endif /* < 2.6.11 */
1103	if (RT_UNLIKELY(pmd_none(u.Middle)))
1104	return NULL;
1105	#if RTLNX_VER_MIN(2,6,0)
1106	if (pmd_large(u.Middle))
1107	{
1108	pPage = pmd_page(u.Middle);
1109	AssertReturn(pPage, NULL);
1110	pfn = page_to_pfn(pPage); /* doing the safe way... */
1111	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
1112	return pfn_to_page(pfn);
1113	}
1114	#endif
1115
1116	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
1117	pEntry = pte_offset_map(&u.Middle, ulAddr);
1118	#else
1119	pEntry = pte_offset(&u.Middle, ulAddr);
1120	#endif
1121	if (RT_UNLIKELY(!pEntry))
1122	return NULL;
1123	u.Entry = *pEntry;
1124	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map)
1125	pte_unmap(pEntry);
1126	#endif
1127
1128	if (RT_UNLIKELY(!pte_present(u.Entry)))
1129	return NULL;
1130	return pte_page(u.Entry);
1131	}
1132	RT_EXPORT_SYMBOL(rtR0MemObjLinuxVirtToPage);
1133
1134
1135	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
1136	{
1137	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
1138	}
1139
1140
1141	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
1142	{
1143	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
1144	}
1145
1146
1147	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
1148	{
1149	IPRT_LINUX_SAVE_EFL_AC();
1150
1151	/*
1152	* All we need to do here is to validate that we can use
1153	* ioremap on the specified address (32/64-bit dma_addr_t).
1154	*/
1155	PRTR0MEMOBJLNX pMemLnx;
1156	dma_addr_t PhysAddr = Phys;
1157	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
1158
1159	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
1160	if (!pMemLnx)
1161	{
1162	IPRT_LINUX_RESTORE_EFL_AC();
1163	return VERR_NO_MEMORY;
1164	}
1165
1166	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
1167	pMemLnx->Core.u.Phys.fAllocated = false;
1168	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
1169	Assert(!pMemLnx->cPages);
1170	*ppMem = &pMemLnx->Core;
1171	IPRT_LINUX_RESTORE_EFL_AC();
1172	return VINF_SUCCESS;
1173	}
1174
1175	/* openSUSE Leap 42.3 detection :-/ */
1176	#if RTLNX_VER_RANGE(4,4,0, 4,6,0) && defined(FAULT_FLAG_REMOTE)
1177	# define GET_USER_PAGES_API KERNEL_VERSION(4, 10, 0) /* no typo! */
1178	#else
1179	# define GET_USER_PAGES_API LINUX_VERSION_CODE
1180	#endif
1181
1182	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
1183	{
1184	IPRT_LINUX_SAVE_EFL_AC();
1185	const int cPages = cb >> PAGE_SHIFT;
1186	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1187	struct vm_area_struct **papVMAs;
1188	PRTR0MEMOBJLNX pMemLnx;
1189	int rc = VERR_NO_MEMORY;
1190	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1191
1192	/*
1193	* Check for valid task and size overflows.
1194	*/
1195	if (!pTask)
1196	return VERR_NOT_SUPPORTED;
1197	if (((size_t)cPages << PAGE_SHIFT) != cb)
1198	return VERR_OUT_OF_RANGE;
1199
1200	/*
1201	* Allocate the memory object and a temporary buffer for the VMAs.
1202	*/
1203	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
1204	if (!pMemLnx)
1205	{
1206	IPRT_LINUX_RESTORE_EFL_AC();
1207	return VERR_NO_MEMORY;
1208	}
1209
1210	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1211	if (papVMAs)
1212	{
1213	LNX_MM_DOWN_READ(pTask->mm);
1214
1215	/*
1216	* Get user pages.
1217	*/
1218	/** @todo r=bird: Should we not force read access too? */
1219	#if GET_USER_PAGES_API >= KERNEL_VERSION(4, 6, 0)
1220	if (R0Process == RTR0ProcHandleSelf())
1221	rc = get_user_pages(R3Ptr, /* Where from. */
1222	cPages, /* How many pages. */
1223	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1224	fWrite ? FOLL_WRITE \| /* Write to memory. */
1225	FOLL_FORCE /* force write access. */
1226	: 0, /* Write to memory. */
1227	# else
1228	fWrite, /* Write to memory. */
1229	fWrite, /* force write access. */
1230	# endif
1231	&pMemLnx->apPages[0], /* Page array. */
1232	papVMAs); /* vmas */
1233	/*
1234	* Actually this should not happen at the moment as call this function
1235	* only for our own process.
1236	*/
1237	else
1238	rc = get_user_pages_remote(
1239	# if GET_USER_PAGES_API < KERNEL_VERSION(5, 9, 0)
1240	pTask, /* Task for fault accounting. */
1241	# endif
1242	pTask->mm, /* Whose pages. */
1243	R3Ptr, /* Where from. */
1244	cPages, /* How many pages. */
1245	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1246	fWrite ? FOLL_WRITE \| /* Write to memory. */
1247	FOLL_FORCE /* force write access. */
1248	: 0, /* Write to memory. */
1249	# else
1250	fWrite, /* Write to memory. */
1251	fWrite, /* force write access. */
1252	# endif
1253	&pMemLnx->apPages[0], /* Page array. */
1254	papVMAs /* vmas */
1255	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 10, 0)
1256	, NULL /* locked */
1257	# endif
1258	);
1259	#else /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1260	rc = get_user_pages(pTask, /* Task for fault accounting. */
1261	pTask->mm, /* Whose pages. */
1262	R3Ptr, /* Where from. */
1263	cPages, /* How many pages. */
1264	/* The get_user_pages API change was back-ported to 4.4.168. */
1265	# if RTLNX_VER_RANGE(4,4,168, 4,5,0)
1266	fWrite ? FOLL_WRITE \| /* Write to memory. */
1267	FOLL_FORCE /* force write access. */
1268	: 0, /* Write to memory. */
1269	# else
1270	fWrite, /* Write to memory. */
1271	fWrite, /* force write access. */
1272	# endif
1273	&pMemLnx->apPages[0], /* Page array. */
1274	papVMAs); /* vmas */
1275	#endif /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1276	if (rc == cPages)
1277	{
1278	/*
1279	* Flush dcache (required?), protect against fork and _really_ pin the page
1280	* table entries. get_user_pages() will protect against swapping out the
1281	* pages but it will NOT protect against removing page table entries. This
1282	* can be achieved with
1283	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1284	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1285	* Usual Linux distributions support only a limited size of locked pages
1286	* (e.g. 32KB).
1287	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1288	* or by
1289	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1290	* a range check.
1291	*/
1292	/** @todo The Linux fork() protection will require more work if this API
1293	* is to be used for anything but locking VM pages. */
1294	while (rc-- > 0)
1295	{
1296	flush_dcache_page(pMemLnx->apPages[rc]);
1297	papVMAs[rc]->vm_flags \|= VM_DONTCOPY \| VM_LOCKED;
1298	}
1299
1300	LNX_MM_UP_READ(pTask->mm);
1301
1302	RTMemFree(papVMAs);
1303
1304	pMemLnx->Core.u.Lock.R0Process = R0Process;
1305	pMemLnx->cPages = cPages;
1306	Assert(!pMemLnx->fMappedToRing0);
1307	*ppMem = &pMemLnx->Core;
1308
1309	IPRT_LINUX_RESTORE_EFL_AC();
1310	return VINF_SUCCESS;
1311	}
1312
1313	/*
1314	* Failed - we need to unlock any pages that we succeeded to lock.
1315	*/
1316	while (rc-- > 0)
1317	{
1318	if (!PageReserved(pMemLnx->apPages[rc]))
1319	SetPageDirty(pMemLnx->apPages[rc]);
1320	#if RTLNX_VER_MIN(4,6,0)
1321	put_page(pMemLnx->apPages[rc]);
1322	#else
1323	page_cache_release(pMemLnx->apPages[rc]);
1324	#endif
1325	}
1326
1327	LNX_MM_UP_READ(pTask->mm);
1328
1329	RTMemFree(papVMAs);
1330	rc = VERR_LOCK_FAILED;
1331	}
1332
1333	rtR0MemObjDelete(&pMemLnx->Core);
1334	IPRT_LINUX_RESTORE_EFL_AC();
1335	return rc;
1336	}
1337
1338
1339	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
1340	{
1341	IPRT_LINUX_SAVE_EFL_AC();
1342	void pvLast = (uint8_t )pv + cb - 1;
1343	size_t const cPages = cb >> PAGE_SHIFT;
1344	PRTR0MEMOBJLNX pMemLnx;
1345	bool fLinearMapping;
1346	int rc;
1347	uint8_t *pbPage;
1348	size_t iPage;
1349	NOREF(fAccess);
1350
1351	if ( !RTR0MemKernelIsValidAddr(pv)
1352	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1353	return VERR_INVALID_PARAMETER;
1354
1355	/*
1356	* The lower part of the kernel memory has a linear mapping between
1357	* physical and virtual addresses. So we take a short cut here. This is
1358	* assumed to be the cleanest way to handle those addresses (and the code
1359	* is well tested, though the test for determining it is not very nice).
1360	* If we ever decide it isn't we can still remove it.
1361	*/
1362	#if 0
1363	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1364	#else
1365	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1366	&& (unsigned long)pvLast < (unsigned long)high_memory;
1367	#endif
1368
1369	/*
1370	* Allocate the memory object.
1371	*/
1372	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
1373	if (!pMemLnx)
1374	{
1375	IPRT_LINUX_RESTORE_EFL_AC();
1376	return VERR_NO_MEMORY;
1377	}
1378
1379	/*
1380	* Gather the pages.
1381	* We ASSUME all kernel pages are non-swappable and non-movable.
1382	*/
1383	rc = VINF_SUCCESS;
1384	pbPage = (uint8_t *)pvLast;
1385	iPage = cPages;
1386	if (!fLinearMapping)
1387	{
1388	while (iPage-- > 0)
1389	{
1390	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1391	if (RT_UNLIKELY(!pPage))
1392	{
1393	rc = VERR_LOCK_FAILED;
1394	break;
1395	}
1396	pMemLnx->apPages[iPage] = pPage;
1397	pbPage -= PAGE_SIZE;
1398	}
1399	}
1400	else
1401	{
1402	while (iPage-- > 0)
1403	{
1404	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1405	pbPage -= PAGE_SIZE;
1406	}
1407	}
1408	if (RT_SUCCESS(rc))
1409	{
1410	/*
1411	* Complete the memory object and return.
1412	*/
1413	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1414	pMemLnx->cPages = cPages;
1415	Assert(!pMemLnx->fMappedToRing0);
1416	*ppMem = &pMemLnx->Core;
1417
1418	IPRT_LINUX_RESTORE_EFL_AC();
1419	return VINF_SUCCESS;
1420	}
1421
1422	rtR0MemObjDelete(&pMemLnx->Core);
1423	IPRT_LINUX_RESTORE_EFL_AC();
1424	return rc;
1425	}
1426
1427
1428	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
1429	{
1430	#if RTLNX_VER_MIN(2,4,22)
1431	IPRT_LINUX_SAVE_EFL_AC();
1432	const size_t cPages = cb >> PAGE_SHIFT;
1433	struct page *pDummyPage;
1434	struct page **papPages;
1435
1436	/* check for unsupported stuff. */
1437	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1438	if (uAlignment > PAGE_SIZE)
1439	return VERR_NOT_SUPPORTED;
1440
1441	/*
1442	* Allocate a dummy page and create a page pointer array for vmap such that
1443	* the dummy page is mapped all over the reserved area.
1444	*/
1445	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1446	if (pDummyPage)
1447	{
1448	papPages = RTMemAlloc(sizeof(papPages) cPages);
1449	if (papPages)
1450	{
1451	void *pv;
1452	size_t iPage = cPages;
1453	while (iPage-- > 0)
1454	papPages[iPage] = pDummyPage;
1455	# ifdef VM_MAP
1456	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1457	# else
1458	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1459	# endif
1460	RTMemFree(papPages);
1461	if (pv)
1462	{
1463	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1464	if (pMemLnx)
1465	{
1466	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1467	pMemLnx->cPages = 1;
1468	pMemLnx->apPages[0] = pDummyPage;
1469	*ppMem = &pMemLnx->Core;
1470	IPRT_LINUX_RESTORE_EFL_AC();
1471	return VINF_SUCCESS;
1472	}
1473	vunmap(pv);
1474	}
1475	}
1476	__free_page(pDummyPage);
1477	}
1478	IPRT_LINUX_RESTORE_EFL_AC();
1479	return VERR_NO_MEMORY;
1480
1481	#else /* < 2.4.22 */
1482	/*
1483	* Could probably use ioremap here, but the caller is in a better position than us
1484	* to select some safe physical memory.
1485	*/
1486	return VERR_NOT_SUPPORTED;
1487	#endif
1488	}
1489
1490
1491	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1492	{
1493	IPRT_LINUX_SAVE_EFL_AC();
1494	PRTR0MEMOBJLNX pMemLnx;
1495	void *pv;
1496	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1497	if (!pTask)
1498	return VERR_NOT_SUPPORTED;
1499
1500	/*
1501	* Check that the specified alignment is supported.
1502	*/
1503	if (uAlignment > PAGE_SIZE)
1504	return VERR_NOT_SUPPORTED;
1505
1506	/*
1507	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1508	*/
1509	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1510	if (pv == (void *)-1)
1511	{
1512	IPRT_LINUX_RESTORE_EFL_AC();
1513	return VERR_NO_MEMORY;
1514	}
1515
1516	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1517	if (!pMemLnx)
1518	{
1519	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1520	IPRT_LINUX_RESTORE_EFL_AC();
1521	return VERR_NO_MEMORY;
1522	}
1523
1524	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1525	*ppMem = &pMemLnx->Core;
1526	IPRT_LINUX_RESTORE_EFL_AC();
1527	return VINF_SUCCESS;
1528	}
1529
1530
1531	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap,
1532	void *pvFixed, size_t uAlignment,
1533	unsigned fProt, size_t offSub, size_t cbSub)
1534	{
1535	int rc = VERR_NO_MEMORY;
1536	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1537	PRTR0MEMOBJLNX pMemLnx;
1538	IPRT_LINUX_SAVE_EFL_AC();
1539
1540	/* Fail if requested to do something we can't. */
1541	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1542	if (uAlignment > PAGE_SIZE)
1543	return VERR_NOT_SUPPORTED;
1544
1545	/*
1546	* Create the IPRT memory object.
1547	*/
1548	if (!cbSub)
1549	cbSub = pMemLnxToMap->Core.cb - offSub;
1550	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub);
1551	if (pMemLnx)
1552	{
1553	if (pMemLnxToMap->cPages)
1554	{
1555	#if RTLNX_VER_MIN(2,4,22)
1556	/*
1557	* Use vmap - 2.4.22 and later.
1558	*/
1559	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1560	/** @todo We don't really care too much for EXEC here... 5.8 always adds NX. */
1561	Assert(((offSub + cbSub) >> PAGE_SHIFT) <= pMemLnxToMap->cPages);
1562	# ifdef VM_MAP
1563	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_MAP, fPg);
1564	# else
1565	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_ALLOC, fPg);
1566	# endif
1567	if (pMemLnx->Core.pv)
1568	{
1569	pMemLnx->fMappedToRing0 = true;
1570	rc = VINF_SUCCESS;
1571	}
1572	else
1573	rc = VERR_MAP_FAILED;
1574
1575	#else /* < 2.4.22 */
1576	/*
1577	* Only option here is to share mappings if possible and forget about fProt.
1578	*/
1579	if (rtR0MemObjIsRing3(pMemToMap))
1580	rc = VERR_NOT_SUPPORTED;
1581	else
1582	{
1583	rc = VINF_SUCCESS;
1584	if (!pMemLnxToMap->Core.pv)
1585	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1586	if (RT_SUCCESS(rc))
1587	{
1588	Assert(pMemLnxToMap->Core.pv);
1589	pMemLnx->Core.pv = (uint8_t *)pMemLnxToMap->Core.pv + offSub;
1590	}
1591	}
1592	#endif
1593	}
1594	else
1595	{
1596	/*
1597	* MMIO / physical memory.
1598	*/
1599	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1600	#if RTLNX_VER_MIN(2,6,25)
1601	/*
1602	* ioremap() defaults to no caching since the 2.6 kernels.
1603	* ioremap_nocache() has been removed finally in 5.6-rc1.
1604	*/
1605	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1606	? ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1607	: ioremap_cache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1608	#else /* KERNEL_VERSION < 2.6.25 */
1609	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1610	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1611	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1612	#endif /* KERNEL_VERSION < 2.6.25 */
1613	if (pMemLnx->Core.pv)
1614	{
1615	/** @todo fix protection. */
1616	rc = VINF_SUCCESS;
1617	}
1618	}
1619	if (RT_SUCCESS(rc))
1620	{
1621	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1622	*ppMem = &pMemLnx->Core;
1623	IPRT_LINUX_RESTORE_EFL_AC();
1624	return VINF_SUCCESS;
1625	}
1626	rtR0MemObjDelete(&pMemLnx->Core);
1627	}
1628
1629	IPRT_LINUX_RESTORE_EFL_AC();
1630	return rc;
1631	}
1632
1633
1634	#ifdef VBOX_USE_PAE_HACK
1635	/**
1636	* Replace the PFN of a PTE with the address of the actual page.
1637	*
1638	* The caller maps a reserved dummy page at the address with the desired access
1639	* and flags.
1640	*
1641	* This hack is required for older Linux kernels which don't provide
1642	* remap_pfn_range().
1643	*
1644	* @returns 0 on success, -ENOMEM on failure.
1645	* @param mm The memory context.
1646	* @param ulAddr The mapping address.
1647	* @param Phys The physical address of the page to map.
1648	*/
1649	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1650	{
1651	int rc = -ENOMEM;
1652	pgd_t *pgd;
1653
1654	spin_lock(&mm->page_table_lock);
1655
1656	pgd = pgd_offset(mm, ulAddr);
1657	if (!pgd_none(pgd) && !pgd_bad(pgd))
1658	{
1659	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1660	if (!pmd_none(*pmd))
1661	{
1662	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1663	if (ptep)
1664	{
1665	pte_t pte = *ptep;
1666	pte.pte_high &= 0xfff00000;
1667	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1668	pte.pte_low &= 0x00000fff;
1669	pte.pte_low \|= (Phys & 0xfffff000);
1670	set_pte(ptep, pte);
1671	pte_unmap(ptep);
1672	rc = 0;
1673	}
1674	}
1675	}
1676
1677	spin_unlock(&mm->page_table_lock);
1678	return rc;
1679	}
1680	#endif /* VBOX_USE_PAE_HACK */
1681
1682
1683	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment,
1684	unsigned fProt, RTR0PROCESS R0Process, size_t offSub, size_t cbSub)
1685	{
1686	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1687	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1688	int rc = VERR_NO_MEMORY;
1689	PRTR0MEMOBJLNX pMemLnx;
1690	#ifdef VBOX_USE_PAE_HACK
1691	struct page *pDummyPage;
1692	RTHCPHYS DummyPhys;
1693	#endif
1694	IPRT_LINUX_SAVE_EFL_AC();
1695
1696	/*
1697	* Check for restrictions.
1698	*/
1699	if (!pTask)
1700	return VERR_NOT_SUPPORTED;
1701	if (uAlignment > PAGE_SIZE)
1702	return VERR_NOT_SUPPORTED;
1703
1704	#ifdef VBOX_USE_PAE_HACK
1705	/*
1706	* Allocate a dummy page for use when mapping the memory.
1707	*/
1708	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1709	if (!pDummyPage)
1710	{
1711	IPRT_LINUX_RESTORE_EFL_AC();
1712	return VERR_NO_MEMORY;
1713	}
1714	SetPageReserved(pDummyPage);
1715	DummyPhys = page_to_phys(pDummyPage);
1716	#endif
1717
1718	/*
1719	* Create the IPRT memory object.
1720	*/
1721	Assert(!offSub \|\| cbSub);
1722	if (cbSub == 0)
1723	cbSub = pMemLnxToMap->Core.cb;
1724	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub);
1725	if (pMemLnx)
1726	{
1727	/*
1728	* Allocate user space mapping.
1729	*/
1730	void *pv;
1731	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cbSub, uAlignment, pTask, fProt);
1732	if (pv != (void *)-1)
1733	{
1734	/*
1735	* Map page by page into the mmap area.
1736	* This is generic, paranoid and not very efficient.
1737	*/
1738	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1739	unsigned long ulAddrCur = (unsigned long)pv;
1740	const size_t cPages = (offSub + cbSub) >> PAGE_SHIFT;
1741	size_t iPage;
1742
1743	LNX_MM_DOWN_WRITE(pTask->mm);
1744
1745	rc = VINF_SUCCESS;
1746	if (pMemLnxToMap->cPages)
1747	{
1748	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1749	{
1750	#if RTLNX_VER_MAX(2,6,11)
1751	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1752	#endif
1753	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1754	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1755	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1756	#endif
1757	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1758	/* remap_page_range() limitation on x86 */
1759	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1760	#endif
1761
1762	#if defined(VBOX_USE_INSERT_PAGE) && RTLNX_VER_MIN(2,6,22)
1763	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1764	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1765	* See remap_pfn_range() in mm/memory.c */
1766	#if RTLNX_VER_MIN(3,7,0)
1767	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1768	#else
1769	vma->vm_flags \|= VM_RESERVED;
1770	#endif
1771	#elif RTLNX_VER_MIN(2,6,11)
1772	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1773	#elif defined(VBOX_USE_PAE_HACK)
1774	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1775	if (!rc)
1776	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1777	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1778	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1779	#else /* 2.4 */
1780	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1781	#endif
1782	if (rc)
1783	{
1784	rc = VERR_NO_MEMORY;
1785	break;
1786	}
1787	}
1788	}
1789	else
1790	{
1791	RTHCPHYS Phys;
1792	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1793	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1794	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1795	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1796	else
1797	{
1798	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1799	Phys = NIL_RTHCPHYS;
1800	}
1801	if (Phys != NIL_RTHCPHYS)
1802	{
1803	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1804	{
1805	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1806	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1807	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1808	#endif
1809	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1810	/* remap_page_range() limitation on x86 */
1811	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1812	#endif
1813
1814	#if RTLNX_VER_MIN(2,6,11)
1815	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1816	#elif defined(VBOX_USE_PAE_HACK)
1817	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1818	if (!rc)
1819	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1820	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1821	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1822	#else /* 2.4 */
1823	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1824	#endif
1825	if (rc)
1826	{
1827	rc = VERR_NO_MEMORY;
1828	break;
1829	}
1830	}
1831	}
1832	}
1833
1834	#ifdef CONFIG_NUMA_BALANCING
1835	# if RTLNX_VER_MAX(3,13,0) && RTLNX_RHEL_MAX(7,0)
1836	# define VBOX_NUMA_HACK_OLD
1837	# endif
1838	if (RT_SUCCESS(rc))
1839	{
1840	/** @todo Ugly hack! But right now we have no other means to
1841	* disable automatic NUMA page balancing. */
1842	# ifdef RT_OS_X86
1843	# ifdef VBOX_NUMA_HACK_OLD
1844	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1845	# endif
1846	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1847	# else
1848	# ifdef VBOX_NUMA_HACK_OLD
1849	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1850	# endif
1851	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1852	# endif
1853	}
1854	#endif /* CONFIG_NUMA_BALANCING */
1855
1856	LNX_MM_UP_WRITE(pTask->mm);
1857
1858	if (RT_SUCCESS(rc))
1859	{
1860	#ifdef VBOX_USE_PAE_HACK
1861	__free_page(pDummyPage);
1862	#endif
1863	pMemLnx->Core.pv = pv;
1864	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1865	*ppMem = &pMemLnx->Core;
1866	IPRT_LINUX_RESTORE_EFL_AC();
1867	return VINF_SUCCESS;
1868	}
1869
1870	/*
1871	* Bail out.
1872	*/
1873	rtR0MemObjLinuxDoMunmap(pv, cbSub, pTask);
1874	}
1875	rtR0MemObjDelete(&pMemLnx->Core);
1876	}
1877	#ifdef VBOX_USE_PAE_HACK
1878	__free_page(pDummyPage);
1879	#endif
1880
1881	IPRT_LINUX_RESTORE_EFL_AC();
1882	return rc;
1883	}
1884
1885
1886	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1887	{
1888	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
1889	/*
1890	* Currently only supported when we've got addresses PTEs from the kernel.
1891	*/
1892	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1893	if (pMemLnx->pArea && pMemLnx->papPtesForArea)
1894	{
1895	pgprot_t const fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1896	size_t const cPages = (offSub + cbSub) >> PAGE_SHIFT;
1897	pte_t **papPtes = pMemLnx->papPtesForArea;
1898	size_t i;
1899
1900	for (i = offSub >> PAGE_SHIFT; i < cPages; i++)
1901	{
1902	set_pte(papPtes[i], mk_pte(pMemLnx->apPages[i], fPg));
1903	}
1904	preempt_disable();
1905	__flush_tlb_all();
1906	preempt_enable();
1907	return VINF_SUCCESS;
1908	}
1909	# elif defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
1910	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1911	if ( pMemLnx->fExecutable
1912	&& pMemLnx->fMappedToRing0)
1913	{
1914	LNXAPPLYPGRANGE Args;
1915	Args.pMemLnx = pMemLnx;
1916	Args.fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1917	int rcLnx = apply_to_page_range(current->active_mm, (unsigned long)pMemLnx->Core.pv + offSub, cbSub,
1918	rtR0MemObjLinuxApplyPageRange, (void *)&Args);
1919	if (rcLnx)
1920	return VERR_NOT_SUPPORTED;
1921
1922	return VINF_SUCCESS;
1923	}
1924	# endif
1925
1926	NOREF(pMem);
1927	NOREF(offSub);
1928	NOREF(cbSub);
1929	NOREF(fProt);
1930	return VERR_NOT_SUPPORTED;
1931	}
1932
1933
1934	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1935	{
1936	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1937
1938	if (pMemLnx->cPages)
1939	return page_to_phys(pMemLnx->apPages[iPage]);
1940
1941	switch (pMemLnx->Core.enmType)
1942	{
1943	case RTR0MEMOBJTYPE_CONT:
1944	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1945
1946	case RTR0MEMOBJTYPE_PHYS:
1947	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1948
1949	/* the parent knows */
1950	case RTR0MEMOBJTYPE_MAPPING:
1951	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1952
1953	/* cPages > 0 */
1954	case RTR0MEMOBJTYPE_LOW:
1955	case RTR0MEMOBJTYPE_LOCK:
1956	case RTR0MEMOBJTYPE_PHYS_NC:
1957	case RTR0MEMOBJTYPE_PAGE:
1958	default:
1959	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1960	RT_FALL_THROUGH();
1961
1962	case RTR0MEMOBJTYPE_RES_VIRT:
1963	return NIL_RTHCPHYS;
1964	}
1965	}
1966

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 90249

Download in other formats: