memobj-r0drv-linux.c@ 92374

Last change on this file since 92374 was 92246, checked in by vboxsync, 3 years ago
IPRT/RTR0MemObj: Added RTR0MemObjWasZeroInitialized and a couple of flags with which the backend can feed it the necessary info. It would be good to try avoid zeroing memory twice when we can. bugref:10093
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev Revision`
File size: 67.3 KB

Line
1	/* $Id: memobj-r0drv-linux.c 92246 2021-11-06 03:10:49Z 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	* @param pszTag Allocation tag used for statistics and such.
338	*/
339	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
340	size_t uAlignment, gfp_t fFlagsLnx, bool fContiguous, bool fExecutable, int rcNoMem,
341	const char *pszTag)
342	{
343	size_t iPage;
344	size_t const cPages = cb >> PAGE_SHIFT;
345	struct page *paPages;
346
347	/*
348	* Allocate a memory object structure that's large enough to contain
349	* the page pointer array.
350	*/
351	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), enmType,
352	NULL, cb, pszTag);
353	if (!pMemLnx)
354	return VERR_NO_MEMORY;
355	pMemLnx->Core.fFlags \|= RTR0MEMOBJ_FLAGS_UNINITIALIZED_AT_ALLOC;
356	pMemLnx->cPages = cPages;
357
358	if (cPages > 255)
359	{
360	# ifdef __GFP_REPEAT
361	/* Try hard to allocate the memory, but the allocation attempt might fail. */
362	fFlagsLnx \|= __GFP_REPEAT;
363	# endif
364	# ifdef __GFP_NOMEMALLOC
365	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
366	fFlagsLnx \|= __GFP_NOMEMALLOC;
367	# endif
368	}
369
370	/*
371	* Allocate the pages.
372	* For small allocations we'll try contiguous first and then fall back on page by page.
373	*/
374	#if RTLNX_VER_MIN(2,4,22)
375	if ( fContiguous
376	\|\| cb <= PAGE_SIZE * 2)
377	{
378	# ifdef VBOX_USE_INSERT_PAGE
379	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
380	# else
381	paPages = alloc_pages(fFlagsLnx \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
382	# endif
383	if (paPages)
384	{
385	fContiguous = true;
386	for (iPage = 0; iPage < cPages; iPage++)
387	pMemLnx->apPages[iPage] = &paPages[iPage];
388	}
389	else if (fContiguous)
390	{
391	rtR0MemObjDelete(&pMemLnx->Core);
392	return rcNoMem;
393	}
394	}
395
396	if (!fContiguous)
397	{
398	/** @todo Try use alloc_pages_bulk_array when available, it should be faster
399	* than a alloc_page loop. Put it in #ifdefs similar to
400	* IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC. */
401	for (iPage = 0; iPage < cPages; iPage++)
402	{
403	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
404	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
405	{
406	while (iPage-- > 0)
407	__free_page(pMemLnx->apPages[iPage]);
408	rtR0MemObjDelete(&pMemLnx->Core);
409	return rcNoMem;
410	}
411	}
412	}
413
414	#else /* < 2.4.22 */
415	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
416	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
417	if (!paPages)
418	{
419	rtR0MemObjDelete(&pMemLnx->Core);
420	return rcNoMem;
421	}
422	for (iPage = 0; iPage < cPages; iPage++)
423	{
424	pMemLnx->apPages[iPage] = &paPages[iPage];
425	if (fExecutable)
426	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
427	if (PageHighMem(pMemLnx->apPages[iPage]))
428	BUG();
429	}
430
431	fContiguous = true;
432	#endif /* < 2.4.22 */
433	pMemLnx->fContiguous = fContiguous;
434	pMemLnx->fExecutable = fExecutable;
435
436	#if RTLNX_VER_MAX(4,5,0)
437	/*
438	* Reserve the pages.
439	*
440	* Linux >= 4.5 with CONFIG_DEBUG_VM panics when setting PG_reserved on compound
441	* pages. According to Michal Hocko this shouldn't be necessary anyway because
442	* as pages which are not on the LRU list are never evictable.
443	*/
444	for (iPage = 0; iPage < cPages; iPage++)
445	SetPageReserved(pMemLnx->apPages[iPage]);
446	#endif
447
448	/*
449	* Note that the physical address of memory allocated with alloc_pages(flags, order)
450	* is always 2^(PAGE_SHIFT+order)-aligned.
451	*/
452	if ( fContiguous
453	&& uAlignment > PAGE_SIZE)
454	{
455	/*
456	* Check for alignment constraints. The physical address of memory allocated with
457	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
458	*/
459	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
460	{
461	/*
462	* This should never happen!
463	*/
464	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
465	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
466	rtR0MemObjLinuxFreePages(pMemLnx);
467	return rcNoMem;
468	}
469	}
470
471	*ppMemLnx = pMemLnx;
472	return VINF_SUCCESS;
473	}
474
475
476	/**
477	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
478	*
479	* This method does NOT free the object.
480	*
481	* @param pMemLnx The object which physical pages should be freed.
482	*/
483	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
484	{
485	size_t iPage = pMemLnx->cPages;
486	if (iPage > 0)
487	{
488	/*
489	* Restore the page flags.
490	*/
491	while (iPage-- > 0)
492	{
493	#if RTLNX_VER_MAX(4,5,0)
494	/* See SetPageReserved() in rtR0MemObjLinuxAllocPages() */
495	ClearPageReserved(pMemLnx->apPages[iPage]);
496	#endif
497	#if RTLNX_VER_MAX(2,4,22)
498	if (pMemLnx->fExecutable)
499	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
500	#endif
501	}
502
503	/*
504	* Free the pages.
505	*/
506	#if RTLNX_VER_MIN(2,4,22)
507	if (!pMemLnx->fContiguous)
508	{
509	iPage = pMemLnx->cPages;
510	while (iPage-- > 0)
511	__free_page(pMemLnx->apPages[iPage]);
512	}
513	else
514	#endif
515	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
516
517	pMemLnx->cPages = 0;
518	}
519	}
520
521
522	#ifdef IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
523	/**
524	* User data passed to the apply_to_page_range() callback.
525	*/
526	typedef struct LNXAPPLYPGRANGE
527	{
528	/** Pointer to the memory object. */
529	PRTR0MEMOBJLNX pMemLnx;
530	/** The page protection flags to apply. */
531	pgprot_t fPg;
532	} LNXAPPLYPGRANGE;
533	/** Pointer to the user data. */
534	typedef LNXAPPLYPGRANGE *PLNXAPPLYPGRANGE;
535	/** Pointer to the const user data. */
536	typedef const LNXAPPLYPGRANGE *PCLNXAPPLYPGRANGE;
537
538	/**
539	* Callback called in apply_to_page_range().
540	*
541	* @returns Linux status code.
542	* @param pPte Pointer to the page table entry for the given address.
543	* @param uAddr The address to apply the new protection to.
544	* @param pvUser The opaque user data.
545	*/
546	static int rtR0MemObjLinuxApplyPageRange(pte_t pPte, unsigned long uAddr, void pvUser)
547	{
548	PCLNXAPPLYPGRANGE pArgs = (PCLNXAPPLYPGRANGE)pvUser;
549	PRTR0MEMOBJLNX pMemLnx = pArgs->pMemLnx;
550	size_t idxPg = (uAddr - (unsigned long)pMemLnx->Core.pv) >> PAGE_SHIFT;
551
552	set_pte(pPte, mk_pte(pMemLnx->apPages[idxPg], pArgs->fPg));
553	return 0;
554	}
555	#endif
556
557
558	/**
559	* Maps the allocation into ring-0.
560	*
561	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
562	*
563	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
564	* space, so we'll use that mapping if possible. If execute access is required, we'll
565	* play safe and do our own mapping.
566	*
567	* @returns IPRT status code.
568	* @param pMemLnx The linux memory object to map.
569	* @param fExecutable Whether execute access is required.
570	*/
571	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
572	{
573	int rc = VINF_SUCCESS;
574
575	/*
576	* Choose mapping strategy.
577	*/
578	bool fMustMap = fExecutable
579	\|\| !pMemLnx->fContiguous;
580	if (!fMustMap)
581	{
582	size_t iPage = pMemLnx->cPages;
583	while (iPage-- > 0)
584	if (PageHighMem(pMemLnx->apPages[iPage]))
585	{
586	fMustMap = true;
587	break;
588	}
589	}
590
591	Assert(!pMemLnx->Core.pv);
592	Assert(!pMemLnx->fMappedToRing0);
593
594	if (fMustMap)
595	{
596	/*
597	* Use vmap - 2.4.22 and later.
598	*/
599	#if RTLNX_VER_MIN(2,4,22)
600	pgprot_t fPg;
601	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
602	# ifdef _PAGE_NX
603	if (!fExecutable)
604	pgprot_val(fPg) \|= _PAGE_NX;
605	# endif
606
607	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
608	if (fExecutable)
609	{
610	# if RTLNX_VER_MIN(3,2,51)
611	pte_t papPtes = (pte_t )kmalloc_array(pMemLnx->cPages, sizeof(papPtes[0]), GFP_KERNEL);
612	# else
613	pte_t papPtes = (pte_t )kmalloc(pMemLnx->cPages * sizeof(papPtes[0]), GFP_KERNEL);
614	# endif
615	if (papPtes)
616	{
617	pMemLnx->pArea = alloc_vm_area(pMemLnx->Core.cb, papPtes); /* Note! pArea->nr_pages is not set. */
618	if (pMemLnx->pArea)
619	{
620	size_t i;
621	Assert(pMemLnx->pArea->size >= pMemLnx->Core.cb); /* Note! includes guard page. */
622	Assert(pMemLnx->pArea->addr);
623	# ifdef _PAGE_NX
624	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
625	# endif
626	pMemLnx->papPtesForArea = papPtes;
627	for (i = 0; i < pMemLnx->cPages; i++)
628	*papPtes[i] = mk_pte(pMemLnx->apPages[i], fPg);
629	pMemLnx->Core.pv = pMemLnx->pArea->addr;
630	pMemLnx->fMappedToRing0 = true;
631	}
632	else
633	{
634	kfree(papPtes);
635	rc = VERR_MAP_FAILED;
636	}
637	}
638	else
639	rc = VERR_MAP_FAILED;
640	}
641	else
642	# endif
643	{
644	# if defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
645	if (fExecutable)
646	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
647	# endif
648
649	# ifdef VM_MAP
650	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
651	# else
652	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
653	# endif
654	if (pMemLnx->Core.pv)
655	pMemLnx->fMappedToRing0 = true;
656	else
657	rc = VERR_MAP_FAILED;
658	}
659	#else /* < 2.4.22 */
660	rc = VERR_NOT_SUPPORTED;
661	#endif
662	}
663	else
664	{
665	/*
666	* Use the kernel RAM mapping.
667	*/
668	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
669	Assert(pMemLnx->Core.pv);
670	}
671
672	return rc;
673	}
674
675
676	/**
677	* Undoes what rtR0MemObjLinuxVMap() did.
678	*
679	* @param pMemLnx The linux memory object.
680	*/
681	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
682	{
683	#if RTLNX_VER_MIN(2,4,22)
684	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
685	if (pMemLnx->pArea)
686	{
687	# if 0
688	pte_t **papPtes = pMemLnx->papPtesForArea;
689	size_t i;
690	for (i = 0; i < pMemLnx->cPages; i++)
691	*papPtes[i] = 0;
692	# endif
693	free_vm_area(pMemLnx->pArea);
694	kfree(pMemLnx->papPtesForArea);
695	pMemLnx->pArea = NULL;
696	pMemLnx->papPtesForArea = NULL;
697	}
698	else
699	# endif
700	if (pMemLnx->fMappedToRing0)
701	{
702	Assert(pMemLnx->Core.pv);
703	vunmap(pMemLnx->Core.pv);
704	pMemLnx->fMappedToRing0 = false;
705	}
706	#else /* < 2.4.22 */
707	Assert(!pMemLnx->fMappedToRing0);
708	#endif
709	pMemLnx->Core.pv = NULL;
710	}
711
712
713	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
714	{
715	IPRT_LINUX_SAVE_EFL_AC();
716	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
717
718	/*
719	* Release any memory that we've allocated or locked.
720	*/
721	switch (pMemLnx->Core.enmType)
722	{
723	case RTR0MEMOBJTYPE_LOW:
724	case RTR0MEMOBJTYPE_PAGE:
725	case RTR0MEMOBJTYPE_CONT:
726	case RTR0MEMOBJTYPE_PHYS:
727	case RTR0MEMOBJTYPE_PHYS_NC:
728	rtR0MemObjLinuxVUnmap(pMemLnx);
729	rtR0MemObjLinuxFreePages(pMemLnx);
730	break;
731
732	case RTR0MEMOBJTYPE_LOCK:
733	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
734	{
735	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
736	size_t iPage;
737	Assert(pTask);
738	if (pTask && pTask->mm)
739	LNX_MM_DOWN_READ(pTask->mm);
740
741	iPage = pMemLnx->cPages;
742	while (iPage-- > 0)
743	{
744	if (!PageReserved(pMemLnx->apPages[iPage]))
745	SetPageDirty(pMemLnx->apPages[iPage]);
746	#if RTLNX_VER_MIN(4,6,0)
747	put_page(pMemLnx->apPages[iPage]);
748	#else
749	page_cache_release(pMemLnx->apPages[iPage]);
750	#endif
751	}
752
753	if (pTask && pTask->mm)
754	LNX_MM_UP_READ(pTask->mm);
755	}
756	/* else: kernel memory - nothing to do here. */
757	break;
758
759	case RTR0MEMOBJTYPE_RES_VIRT:
760	Assert(pMemLnx->Core.pv);
761	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
762	{
763	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
764	Assert(pTask);
765	if (pTask && pTask->mm)
766	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
767	}
768	else
769	{
770	vunmap(pMemLnx->Core.pv);
771
772	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
773	__free_page(pMemLnx->apPages[0]);
774	pMemLnx->apPages[0] = NULL;
775	pMemLnx->cPages = 0;
776	}
777	pMemLnx->Core.pv = NULL;
778	break;
779
780	case RTR0MEMOBJTYPE_MAPPING:
781	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
782	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
783	{
784	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
785	Assert(pTask);
786	if (pTask && pTask->mm)
787	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
788	}
789	else
790	vunmap(pMemLnx->Core.pv);
791	pMemLnx->Core.pv = NULL;
792	break;
793
794	default:
795	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
796	return VERR_INTERNAL_ERROR;
797	}
798	IPRT_LINUX_RESTORE_EFL_ONLY_AC();
799	return VINF_SUCCESS;
800	}
801
802
803	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
804	{
805	IPRT_LINUX_SAVE_EFL_AC();
806	PRTR0MEMOBJLNX pMemLnx;
807	int rc;
808
809	#if RTLNX_VER_MIN(2,4,22)
810	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
811	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, pszTag);
812	#else
813	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
814	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, pszTag);
815	#endif
816	if (RT_SUCCESS(rc))
817	{
818	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
819	if (RT_SUCCESS(rc))
820	{
821	*ppMem = &pMemLnx->Core;
822	IPRT_LINUX_RESTORE_EFL_AC();
823	return rc;
824	}
825
826	rtR0MemObjLinuxFreePages(pMemLnx);
827	rtR0MemObjDelete(&pMemLnx->Core);
828	}
829
830	IPRT_LINUX_RESTORE_EFL_AC();
831	return rc;
832	}
833
834
835	DECLHIDDEN(int) rtR0MemObjNativeAllocLarge(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, size_t cbLargePage, uint32_t fFlags,
836	const char *pszTag)
837	{
838	return rtR0MemObjFallbackAllocLarge(ppMem, cb, cbLargePage, fFlags, pszTag);
839	}
840
841
842	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
843	{
844	IPRT_LINUX_SAVE_EFL_AC();
845	PRTR0MEMOBJLNX pMemLnx;
846	int rc;
847
848	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
849	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
850	/* ZONE_DMA32: 0-4GB */
851	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
852	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
853	if (RT_FAILURE(rc))
854	#endif
855	#ifdef RT_ARCH_AMD64
856	/* ZONE_DMA: 0-16MB */
857	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
858	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
859	#else
860	# ifdef CONFIG_X86_PAE
861	# endif
862	/* ZONE_NORMAL: 0-896MB */
863	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
864	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
865	#endif
866	if (RT_SUCCESS(rc))
867	{
868	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
869	if (RT_SUCCESS(rc))
870	{
871	*ppMem = &pMemLnx->Core;
872	IPRT_LINUX_RESTORE_EFL_AC();
873	return rc;
874	}
875
876	rtR0MemObjLinuxFreePages(pMemLnx);
877	rtR0MemObjDelete(&pMemLnx->Core);
878	}
879
880	IPRT_LINUX_RESTORE_EFL_AC();
881	return rc;
882	}
883
884
885	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
886	{
887	IPRT_LINUX_SAVE_EFL_AC();
888	PRTR0MEMOBJLNX pMemLnx;
889	int rc;
890
891	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
892	/* ZONE_DMA32: 0-4GB */
893	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
894	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
895	if (RT_FAILURE(rc))
896	#endif
897	#ifdef RT_ARCH_AMD64
898	/* ZONE_DMA: 0-16MB */
899	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
900	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
901	#else
902	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
903	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
904	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
905	#endif
906	if (RT_SUCCESS(rc))
907	{
908	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
909	if (RT_SUCCESS(rc))
910	{
911	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
912	size_t iPage = pMemLnx->cPages;
913	while (iPage-- > 0)
914	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
915	#endif
916	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
917	*ppMem = &pMemLnx->Core;
918	IPRT_LINUX_RESTORE_EFL_AC();
919	return rc;
920	}
921
922	rtR0MemObjLinuxFreePages(pMemLnx);
923	rtR0MemObjDelete(&pMemLnx->Core);
924	}
925
926	IPRT_LINUX_RESTORE_EFL_AC();
927	return rc;
928	}
929
930
931	/**
932	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
933	*
934	* @returns IPRT status code.
935	* @param ppMemLnx Where to
936	* @param enmType The object type.
937	* @param cb The size of the allocation.
938	* @param uAlignment The alignment of the physical memory.
939	* Only valid for fContiguous == true, ignored otherwise.
940	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
941	* @param pszTag Allocation tag used for statistics and such.
942	* @param fGfp The Linux GFP flags to use for the allocation.
943	*/
944	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
945	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag, gfp_t fGfp)
946	{
947	PRTR0MEMOBJLNX pMemLnx;
948	int rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
949	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
950	false /fExecutable/, VERR_NO_PHYS_MEMORY, pszTag);
951	if (RT_FAILURE(rc))
952	return rc;
953
954	/*
955	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
956	*/
957	if (PhysHighest != NIL_RTHCPHYS)
958	{
959	size_t iPage = pMemLnx->cPages;
960	while (iPage-- > 0)
961	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
962	{
963	rtR0MemObjLinuxFreePages(pMemLnx);
964	rtR0MemObjDelete(&pMemLnx->Core);
965	return VERR_NO_MEMORY;
966	}
967	}
968
969	/*
970	* Complete the object.
971	*/
972	if (enmType == RTR0MEMOBJTYPE_PHYS)
973	{
974	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
975	pMemLnx->Core.u.Phys.fAllocated = true;
976	}
977	*ppMem = &pMemLnx->Core;
978	return rc;
979	}
980
981
982	/**
983	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
984	*
985	* @returns IPRT status code.
986	* @param ppMem Where to store the memory object pointer on success.
987	* @param enmType The object type.
988	* @param cb The size of the allocation.
989	* @param uAlignment The alignment of the physical memory.
990	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
991	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
992	* @param pszTag Allocation tag used for statistics and such.
993	*/
994	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
995	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag)
996	{
997	int rc;
998	IPRT_LINUX_SAVE_EFL_AC();
999
1000	/*
1001	* There are two clear cases and that's the <=16MB and anything-goes ones.
1002	* When the physical address limit is somewhere in-between those two we'll
1003	* just have to try, starting with HIGHUSER and working our way thru the
1004	* different types, hoping we'll get lucky.
1005	*
1006	* We should probably move this physical address restriction logic up to
1007	* the page alloc function as it would be more efficient there. But since
1008	* we don't expect this to be a performance issue just yet it can wait.
1009	*/
1010	if (PhysHighest == NIL_RTHCPHYS)
1011	/* ZONE_HIGHMEM: the whole physical memory */
1012	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1013	else if (PhysHighest <= _1M * 16)
1014	/* ZONE_DMA: 0-16MB */
1015	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1016	else
1017	{
1018	rc = VERR_NO_MEMORY;
1019	if (RT_FAILURE(rc))
1020	/* ZONE_HIGHMEM: the whole physical memory */
1021	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1022	if (RT_FAILURE(rc))
1023	/* ZONE_NORMAL: 0-896MB */
1024	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_USER);
1025	#ifdef GFP_DMA32
1026	if (RT_FAILURE(rc))
1027	/* ZONE_DMA32: 0-4GB */
1028	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA32);
1029	#endif
1030	if (RT_FAILURE(rc))
1031	/* ZONE_DMA: 0-16MB */
1032	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1033	}
1034	IPRT_LINUX_RESTORE_EFL_AC();
1035	return rc;
1036	}
1037
1038
1039	/**
1040	* Translates a kernel virtual address to a linux page structure by walking the
1041	* page tables.
1042	*
1043	* @note We do assume that the page tables will not change as we are walking
1044	* them. This assumption is rather forced by the fact that I could not
1045	* immediately see any way of preventing this from happening. So, we
1046	* take some extra care when accessing them.
1047	*
1048	* Because of this, we don't want to use this function on memory where
1049	* attribute changes to nearby pages is likely to cause large pages to
1050	* be used or split up. So, don't use this for the linear mapping of
1051	* physical memory.
1052	*
1053	* @returns Pointer to the page structur or NULL if it could not be found.
1054	* @param pv The kernel virtual address.
1055	*/
1056	RTDECL(struct page ) rtR0MemObjLinuxVirtToPage(void pv)
1057	{
1058	unsigned long ulAddr = (unsigned long)pv;
1059	unsigned long pfn;
1060	struct page *pPage;
1061	pte_t *pEntry;
1062	union
1063	{
1064	pgd_t Global;
1065	#if RTLNX_VER_MIN(4,12,0)
1066	p4d_t Four;
1067	#endif
1068	#if RTLNX_VER_MIN(2,6,11)
1069	pud_t Upper;
1070	#endif
1071	pmd_t Middle;
1072	pte_t Entry;
1073	} u;
1074
1075	/* Should this happen in a situation this code will be called in? And if
1076	* so, can it change under our feet? See also
1077	* "Documentation/vm/active_mm.txt" in the kernel sources. */
1078	if (RT_UNLIKELY(!current->active_mm))
1079	return NULL;
1080	u.Global = *pgd_offset(current->active_mm, ulAddr);
1081	if (RT_UNLIKELY(pgd_none(u.Global)))
1082	return NULL;
1083	#if RTLNX_VER_MIN(2,6,11)
1084	# if RTLNX_VER_MIN(4,12,0)
1085	u.Four = *p4d_offset(&u.Global, ulAddr);
1086	if (RT_UNLIKELY(p4d_none(u.Four)))
1087	return NULL;
1088	if (p4d_large(u.Four))
1089	{
1090	pPage = p4d_page(u.Four);
1091	AssertReturn(pPage, NULL);
1092	pfn = page_to_pfn(pPage); /* doing the safe way... */
1093	AssertCompile(P4D_SHIFT - PAGE_SHIFT < 31);
1094	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (P4D_SHIFT - PAGE_SHIFT)) - 1);
1095	return pfn_to_page(pfn);
1096	}
1097	u.Upper = *pud_offset(&u.Four, ulAddr);
1098	# else /* < 4.12 */
1099	u.Upper = *pud_offset(&u.Global, ulAddr);
1100	# endif /* < 4.12 */
1101	if (RT_UNLIKELY(pud_none(u.Upper)))
1102	return NULL;
1103	# if RTLNX_VER_MIN(2,6,25)
1104	if (pud_large(u.Upper))
1105	{
1106	pPage = pud_page(u.Upper);
1107	AssertReturn(pPage, NULL);
1108	pfn = page_to_pfn(pPage); /* doing the safe way... */
1109	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
1110	return pfn_to_page(pfn);
1111	}
1112	# endif
1113	u.Middle = *pmd_offset(&u.Upper, ulAddr);
1114	#else /* < 2.6.11 */
1115	u.Middle = *pmd_offset(&u.Global, ulAddr);
1116	#endif /* < 2.6.11 */
1117	if (RT_UNLIKELY(pmd_none(u.Middle)))
1118	return NULL;
1119	#if RTLNX_VER_MIN(2,6,0)
1120	if (pmd_large(u.Middle))
1121	{
1122	pPage = pmd_page(u.Middle);
1123	AssertReturn(pPage, NULL);
1124	pfn = page_to_pfn(pPage); /* doing the safe way... */
1125	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
1126	return pfn_to_page(pfn);
1127	}
1128	#endif
1129
1130	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
1131	pEntry = pte_offset_map(&u.Middle, ulAddr);
1132	#else
1133	pEntry = pte_offset(&u.Middle, ulAddr);
1134	#endif
1135	if (RT_UNLIKELY(!pEntry))
1136	return NULL;
1137	u.Entry = *pEntry;
1138	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map)
1139	pte_unmap(pEntry);
1140	#endif
1141
1142	if (RT_UNLIKELY(!pte_present(u.Entry)))
1143	return NULL;
1144	return pte_page(u.Entry);
1145	}
1146	RT_EXPORT_SYMBOL(rtR0MemObjLinuxVirtToPage);
1147
1148
1149	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment,
1150	const char *pszTag)
1151	{
1152	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest, pszTag);
1153	}
1154
1155
1156	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, const char *pszTag)
1157	{
1158	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest, pszTag);
1159	}
1160
1161
1162	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy,
1163	const char *pszTag)
1164	{
1165	IPRT_LINUX_SAVE_EFL_AC();
1166
1167	/*
1168	* All we need to do here is to validate that we can use
1169	* ioremap on the specified address (32/64-bit dma_addr_t).
1170	*/
1171	PRTR0MEMOBJLNX pMemLnx;
1172	dma_addr_t PhysAddr = Phys;
1173	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
1174
1175	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb, pszTag);
1176	if (!pMemLnx)
1177	{
1178	IPRT_LINUX_RESTORE_EFL_AC();
1179	return VERR_NO_MEMORY;
1180	}
1181
1182	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
1183	pMemLnx->Core.u.Phys.fAllocated = false;
1184	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
1185	Assert(!pMemLnx->cPages);
1186	*ppMem = &pMemLnx->Core;
1187	IPRT_LINUX_RESTORE_EFL_AC();
1188	return VINF_SUCCESS;
1189	}
1190
1191	/* openSUSE Leap 42.3 detection :-/ */
1192	#if RTLNX_VER_RANGE(4,4,0, 4,6,0) && defined(FAULT_FLAG_REMOTE)
1193	# define GET_USER_PAGES_API KERNEL_VERSION(4, 10, 0) /* no typo! */
1194	#else
1195	# define GET_USER_PAGES_API LINUX_VERSION_CODE
1196	#endif
1197
1198	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess,
1199	RTR0PROCESS R0Process, const char *pszTag)
1200	{
1201	IPRT_LINUX_SAVE_EFL_AC();
1202	const int cPages = cb >> PAGE_SHIFT;
1203	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1204	struct vm_area_struct **papVMAs;
1205	PRTR0MEMOBJLNX pMemLnx;
1206	int rc = VERR_NO_MEMORY;
1207	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1208
1209	/*
1210	* Check for valid task and size overflows.
1211	*/
1212	if (!pTask)
1213	return VERR_NOT_SUPPORTED;
1214	if (((size_t)cPages << PAGE_SHIFT) != cb)
1215	return VERR_OUT_OF_RANGE;
1216
1217	/*
1218	* Allocate the memory object and a temporary buffer for the VMAs.
1219	*/
1220	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1221	(void *)R3Ptr, cb, pszTag);
1222	if (!pMemLnx)
1223	{
1224	IPRT_LINUX_RESTORE_EFL_AC();
1225	return VERR_NO_MEMORY;
1226	}
1227
1228	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1229	if (papVMAs)
1230	{
1231	LNX_MM_DOWN_READ(pTask->mm);
1232
1233	/*
1234	* Get user pages.
1235	*/
1236	/** @todo r=bird: Should we not force read access too? */
1237	#if GET_USER_PAGES_API >= KERNEL_VERSION(4, 6, 0)
1238	if (R0Process == RTR0ProcHandleSelf())
1239	rc = get_user_pages(R3Ptr, /* Where from. */
1240	cPages, /* How many pages. */
1241	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1242	fWrite ? FOLL_WRITE \| /* Write to memory. */
1243	FOLL_FORCE /* force write access. */
1244	: 0, /* Write to memory. */
1245	# else
1246	fWrite, /* Write to memory. */
1247	fWrite, /* force write access. */
1248	# endif
1249	&pMemLnx->apPages[0], /* Page array. */
1250	papVMAs); /* vmas */
1251	/*
1252	* Actually this should not happen at the moment as call this function
1253	* only for our own process.
1254	*/
1255	else
1256	rc = get_user_pages_remote(
1257	# if GET_USER_PAGES_API < KERNEL_VERSION(5, 9, 0)
1258	pTask, /* Task for fault accounting. */
1259	# endif
1260	pTask->mm, /* Whose pages. */
1261	R3Ptr, /* Where from. */
1262	cPages, /* How many pages. */
1263	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1264	fWrite ? FOLL_WRITE \| /* Write to memory. */
1265	FOLL_FORCE /* force write access. */
1266	: 0, /* Write to memory. */
1267	# else
1268	fWrite, /* Write to memory. */
1269	fWrite, /* force write access. */
1270	# endif
1271	&pMemLnx->apPages[0], /* Page array. */
1272	papVMAs /* vmas */
1273	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 10, 0)
1274	, NULL /* locked */
1275	# endif
1276	);
1277	#else /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1278	rc = get_user_pages(pTask, /* Task for fault accounting. */
1279	pTask->mm, /* Whose pages. */
1280	R3Ptr, /* Where from. */
1281	cPages, /* How many pages. */
1282	/* The get_user_pages API change was back-ported to 4.4.168. */
1283	# if RTLNX_VER_RANGE(4,4,168, 4,5,0)
1284	fWrite ? FOLL_WRITE \| /* Write to memory. */
1285	FOLL_FORCE /* force write access. */
1286	: 0, /* Write to memory. */
1287	# else
1288	fWrite, /* Write to memory. */
1289	fWrite, /* force write access. */
1290	# endif
1291	&pMemLnx->apPages[0], /* Page array. */
1292	papVMAs); /* vmas */
1293	#endif /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1294	if (rc == cPages)
1295	{
1296	/*
1297	* Flush dcache (required?), protect against fork and _really_ pin the page
1298	* table entries. get_user_pages() will protect against swapping out the
1299	* pages but it will NOT protect against removing page table entries. This
1300	* can be achieved with
1301	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1302	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1303	* Usual Linux distributions support only a limited size of locked pages
1304	* (e.g. 32KB).
1305	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1306	* or by
1307	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1308	* a range check.
1309	*/
1310	/** @todo The Linux fork() protection will require more work if this API
1311	* is to be used for anything but locking VM pages. */
1312	while (rc-- > 0)
1313	{
1314	flush_dcache_page(pMemLnx->apPages[rc]);
1315	papVMAs[rc]->vm_flags \|= VM_DONTCOPY \| VM_LOCKED;
1316	}
1317
1318	LNX_MM_UP_READ(pTask->mm);
1319
1320	RTMemFree(papVMAs);
1321
1322	pMemLnx->Core.u.Lock.R0Process = R0Process;
1323	pMemLnx->cPages = cPages;
1324	Assert(!pMemLnx->fMappedToRing0);
1325	*ppMem = &pMemLnx->Core;
1326
1327	IPRT_LINUX_RESTORE_EFL_AC();
1328	return VINF_SUCCESS;
1329	}
1330
1331	/*
1332	* Failed - we need to unlock any pages that we succeeded to lock.
1333	*/
1334	while (rc-- > 0)
1335	{
1336	if (!PageReserved(pMemLnx->apPages[rc]))
1337	SetPageDirty(pMemLnx->apPages[rc]);
1338	#if RTLNX_VER_MIN(4,6,0)
1339	put_page(pMemLnx->apPages[rc]);
1340	#else
1341	page_cache_release(pMemLnx->apPages[rc]);
1342	#endif
1343	}
1344
1345	LNX_MM_UP_READ(pTask->mm);
1346
1347	RTMemFree(papVMAs);
1348	rc = VERR_LOCK_FAILED;
1349	}
1350
1351	rtR0MemObjDelete(&pMemLnx->Core);
1352	IPRT_LINUX_RESTORE_EFL_AC();
1353	return rc;
1354	}
1355
1356
1357	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void pv, size_t cb, uint32_t fAccess, const char pszTag)
1358	{
1359	IPRT_LINUX_SAVE_EFL_AC();
1360	void pvLast = (uint8_t )pv + cb - 1;
1361	size_t const cPages = cb >> PAGE_SHIFT;
1362	PRTR0MEMOBJLNX pMemLnx;
1363	bool fLinearMapping;
1364	int rc;
1365	uint8_t *pbPage;
1366	size_t iPage;
1367	NOREF(fAccess);
1368
1369	if ( !RTR0MemKernelIsValidAddr(pv)
1370	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1371	return VERR_INVALID_PARAMETER;
1372
1373	/*
1374	* The lower part of the kernel memory has a linear mapping between
1375	* physical and virtual addresses. So we take a short cut here. This is
1376	* assumed to be the cleanest way to handle those addresses (and the code
1377	* is well tested, though the test for determining it is not very nice).
1378	* If we ever decide it isn't we can still remove it.
1379	*/
1380	#if 0
1381	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1382	#else
1383	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1384	&& (unsigned long)pvLast < (unsigned long)high_memory;
1385	#endif
1386
1387	/*
1388	* Allocate the memory object.
1389	*/
1390	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1391	pv, cb, pszTag);
1392	if (!pMemLnx)
1393	{
1394	IPRT_LINUX_RESTORE_EFL_AC();
1395	return VERR_NO_MEMORY;
1396	}
1397
1398	/*
1399	* Gather the pages.
1400	* We ASSUME all kernel pages are non-swappable and non-movable.
1401	*/
1402	rc = VINF_SUCCESS;
1403	pbPage = (uint8_t *)pvLast;
1404	iPage = cPages;
1405	if (!fLinearMapping)
1406	{
1407	while (iPage-- > 0)
1408	{
1409	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1410	if (RT_UNLIKELY(!pPage))
1411	{
1412	rc = VERR_LOCK_FAILED;
1413	break;
1414	}
1415	pMemLnx->apPages[iPage] = pPage;
1416	pbPage -= PAGE_SIZE;
1417	}
1418	}
1419	else
1420	{
1421	while (iPage-- > 0)
1422	{
1423	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1424	pbPage -= PAGE_SIZE;
1425	}
1426	}
1427	if (RT_SUCCESS(rc))
1428	{
1429	/*
1430	* Complete the memory object and return.
1431	*/
1432	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1433	pMemLnx->cPages = cPages;
1434	Assert(!pMemLnx->fMappedToRing0);
1435	*ppMem = &pMemLnx->Core;
1436
1437	IPRT_LINUX_RESTORE_EFL_AC();
1438	return VINF_SUCCESS;
1439	}
1440
1441	rtR0MemObjDelete(&pMemLnx->Core);
1442	IPRT_LINUX_RESTORE_EFL_AC();
1443	return rc;
1444	}
1445
1446
1447	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment,
1448	const char *pszTag)
1449	{
1450	#if RTLNX_VER_MIN(2,4,22)
1451	IPRT_LINUX_SAVE_EFL_AC();
1452	const size_t cPages = cb >> PAGE_SHIFT;
1453	struct page *pDummyPage;
1454	struct page **papPages;
1455
1456	/* check for unsupported stuff. */
1457	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1458	if (uAlignment > PAGE_SIZE)
1459	return VERR_NOT_SUPPORTED;
1460
1461	/*
1462	* Allocate a dummy page and create a page pointer array for vmap such that
1463	* the dummy page is mapped all over the reserved area.
1464	*/
1465	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1466	if (pDummyPage)
1467	{
1468	papPages = RTMemAlloc(sizeof(papPages) cPages);
1469	if (papPages)
1470	{
1471	void *pv;
1472	size_t iPage = cPages;
1473	while (iPage-- > 0)
1474	papPages[iPage] = pDummyPage;
1475	# ifdef VM_MAP
1476	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1477	# else
1478	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1479	# endif
1480	RTMemFree(papPages);
1481	if (pv)
1482	{
1483	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1484	if (pMemLnx)
1485	{
1486	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1487	pMemLnx->cPages = 1;
1488	pMemLnx->apPages[0] = pDummyPage;
1489	*ppMem = &pMemLnx->Core;
1490	IPRT_LINUX_RESTORE_EFL_AC();
1491	return VINF_SUCCESS;
1492	}
1493	vunmap(pv);
1494	}
1495	}
1496	__free_page(pDummyPage);
1497	}
1498	IPRT_LINUX_RESTORE_EFL_AC();
1499	return VERR_NO_MEMORY;
1500
1501	#else /* < 2.4.22 */
1502	/*
1503	* Could probably use ioremap here, but the caller is in a better position than us
1504	* to select some safe physical memory.
1505	*/
1506	return VERR_NOT_SUPPORTED;
1507	#endif
1508	}
1509
1510
1511	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment,
1512	RTR0PROCESS R0Process, const char *pszTag)
1513	{
1514	IPRT_LINUX_SAVE_EFL_AC();
1515	PRTR0MEMOBJLNX pMemLnx;
1516	void *pv;
1517	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1518	if (!pTask)
1519	return VERR_NOT_SUPPORTED;
1520
1521	/*
1522	* Check that the specified alignment is supported.
1523	*/
1524	if (uAlignment > PAGE_SIZE)
1525	return VERR_NOT_SUPPORTED;
1526
1527	/*
1528	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1529	*/
1530	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1531	if (pv == (void *)-1)
1532	{
1533	IPRT_LINUX_RESTORE_EFL_AC();
1534	return VERR_NO_MEMORY;
1535	}
1536
1537	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1538	if (!pMemLnx)
1539	{
1540	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1541	IPRT_LINUX_RESTORE_EFL_AC();
1542	return VERR_NO_MEMORY;
1543	}
1544
1545	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1546	*ppMem = &pMemLnx->Core;
1547	IPRT_LINUX_RESTORE_EFL_AC();
1548	return VINF_SUCCESS;
1549	}
1550
1551
1552	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment,
1553	unsigned fProt, size_t offSub, size_t cbSub, const char *pszTag)
1554	{
1555	int rc = VERR_NO_MEMORY;
1556	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1557	PRTR0MEMOBJLNX pMemLnx;
1558	IPRT_LINUX_SAVE_EFL_AC();
1559
1560	/* Fail if requested to do something we can't. */
1561	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1562	if (uAlignment > PAGE_SIZE)
1563	return VERR_NOT_SUPPORTED;
1564
1565	/*
1566	* Create the IPRT memory object.
1567	*/
1568	if (!cbSub)
1569	cbSub = pMemLnxToMap->Core.cb - offSub;
1570	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1571	if (pMemLnx)
1572	{
1573	if (pMemLnxToMap->cPages)
1574	{
1575	#if RTLNX_VER_MIN(2,4,22)
1576	/*
1577	* Use vmap - 2.4.22 and later.
1578	*/
1579	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1580	/** @todo We don't really care too much for EXEC here... 5.8 always adds NX. */
1581	Assert(((offSub + cbSub) >> PAGE_SHIFT) <= pMemLnxToMap->cPages);
1582	# ifdef VM_MAP
1583	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_MAP, fPg);
1584	# else
1585	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_ALLOC, fPg);
1586	# endif
1587	if (pMemLnx->Core.pv)
1588	{
1589	pMemLnx->fMappedToRing0 = true;
1590	rc = VINF_SUCCESS;
1591	}
1592	else
1593	rc = VERR_MAP_FAILED;
1594
1595	#else /* < 2.4.22 */
1596	/*
1597	* Only option here is to share mappings if possible and forget about fProt.
1598	*/
1599	if (rtR0MemObjIsRing3(pMemToMap))
1600	rc = VERR_NOT_SUPPORTED;
1601	else
1602	{
1603	rc = VINF_SUCCESS;
1604	if (!pMemLnxToMap->Core.pv)
1605	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1606	if (RT_SUCCESS(rc))
1607	{
1608	Assert(pMemLnxToMap->Core.pv);
1609	pMemLnx->Core.pv = (uint8_t *)pMemLnxToMap->Core.pv + offSub;
1610	}
1611	}
1612	#endif
1613	}
1614	else
1615	{
1616	/*
1617	* MMIO / physical memory.
1618	*/
1619	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1620	#if RTLNX_VER_MIN(2,6,25)
1621	/*
1622	* ioremap() defaults to no caching since the 2.6 kernels.
1623	* ioremap_nocache() has been removed finally in 5.6-rc1.
1624	*/
1625	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1626	? ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1627	: ioremap_cache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1628	#else /* KERNEL_VERSION < 2.6.25 */
1629	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1630	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1631	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1632	#endif /* KERNEL_VERSION < 2.6.25 */
1633	if (pMemLnx->Core.pv)
1634	{
1635	/** @todo fix protection. */
1636	rc = VINF_SUCCESS;
1637	}
1638	}
1639	if (RT_SUCCESS(rc))
1640	{
1641	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1642	*ppMem = &pMemLnx->Core;
1643	IPRT_LINUX_RESTORE_EFL_AC();
1644	return VINF_SUCCESS;
1645	}
1646	rtR0MemObjDelete(&pMemLnx->Core);
1647	}
1648
1649	IPRT_LINUX_RESTORE_EFL_AC();
1650	return rc;
1651	}
1652
1653
1654	#ifdef VBOX_USE_PAE_HACK
1655	/**
1656	* Replace the PFN of a PTE with the address of the actual page.
1657	*
1658	* The caller maps a reserved dummy page at the address with the desired access
1659	* and flags.
1660	*
1661	* This hack is required for older Linux kernels which don't provide
1662	* remap_pfn_range().
1663	*
1664	* @returns 0 on success, -ENOMEM on failure.
1665	* @param mm The memory context.
1666	* @param ulAddr The mapping address.
1667	* @param Phys The physical address of the page to map.
1668	*/
1669	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1670	{
1671	int rc = -ENOMEM;
1672	pgd_t *pgd;
1673
1674	spin_lock(&mm->page_table_lock);
1675
1676	pgd = pgd_offset(mm, ulAddr);
1677	if (!pgd_none(pgd) && !pgd_bad(pgd))
1678	{
1679	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1680	if (!pmd_none(*pmd))
1681	{
1682	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1683	if (ptep)
1684	{
1685	pte_t pte = *ptep;
1686	pte.pte_high &= 0xfff00000;
1687	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1688	pte.pte_low &= 0x00000fff;
1689	pte.pte_low \|= (Phys & 0xfffff000);
1690	set_pte(ptep, pte);
1691	pte_unmap(ptep);
1692	rc = 0;
1693	}
1694	}
1695	}
1696
1697	spin_unlock(&mm->page_table_lock);
1698	return rc;
1699	}
1700	#endif /* VBOX_USE_PAE_HACK */
1701
1702
1703	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment,
1704	unsigned fProt, RTR0PROCESS R0Process, size_t offSub, size_t cbSub, const char *pszTag)
1705	{
1706	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1707	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1708	int rc = VERR_NO_MEMORY;
1709	PRTR0MEMOBJLNX pMemLnx;
1710	#ifdef VBOX_USE_PAE_HACK
1711	struct page *pDummyPage;
1712	RTHCPHYS DummyPhys;
1713	#endif
1714	IPRT_LINUX_SAVE_EFL_AC();
1715
1716	/*
1717	* Check for restrictions.
1718	*/
1719	if (!pTask)
1720	return VERR_NOT_SUPPORTED;
1721	if (uAlignment > PAGE_SIZE)
1722	return VERR_NOT_SUPPORTED;
1723
1724	#ifdef VBOX_USE_PAE_HACK
1725	/*
1726	* Allocate a dummy page for use when mapping the memory.
1727	*/
1728	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1729	if (!pDummyPage)
1730	{
1731	IPRT_LINUX_RESTORE_EFL_AC();
1732	return VERR_NO_MEMORY;
1733	}
1734	SetPageReserved(pDummyPage);
1735	DummyPhys = page_to_phys(pDummyPage);
1736	#endif
1737
1738	/*
1739	* Create the IPRT memory object.
1740	*/
1741	Assert(!offSub \|\| cbSub);
1742	if (cbSub == 0)
1743	cbSub = pMemLnxToMap->Core.cb;
1744	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1745	if (pMemLnx)
1746	{
1747	/*
1748	* Allocate user space mapping.
1749	*/
1750	void *pv;
1751	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cbSub, uAlignment, pTask, fProt);
1752	if (pv != (void *)-1)
1753	{
1754	/*
1755	* Map page by page into the mmap area.
1756	* This is generic, paranoid and not very efficient.
1757	*/
1758	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1759	unsigned long ulAddrCur = (unsigned long)pv;
1760	const size_t cPages = (offSub + cbSub) >> PAGE_SHIFT;
1761	size_t iPage;
1762
1763	LNX_MM_DOWN_WRITE(pTask->mm);
1764
1765	rc = VINF_SUCCESS;
1766	if (pMemLnxToMap->cPages)
1767	{
1768	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1769	{
1770	#if RTLNX_VER_MAX(2,6,11)
1771	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1772	#endif
1773	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1774	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1775	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1776	#endif
1777	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1778	/* remap_page_range() limitation on x86 */
1779	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1780	#endif
1781
1782	#if defined(VBOX_USE_INSERT_PAGE) && RTLNX_VER_MIN(2,6,22)
1783	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1784	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1785	* See remap_pfn_range() in mm/memory.c */
1786	#if RTLNX_VER_MIN(3,7,0)
1787	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1788	#else
1789	vma->vm_flags \|= VM_RESERVED;
1790	#endif
1791	#elif RTLNX_VER_MIN(2,6,11)
1792	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1793	#elif defined(VBOX_USE_PAE_HACK)
1794	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1795	if (!rc)
1796	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1797	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1798	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1799	#else /* 2.4 */
1800	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1801	#endif
1802	if (rc)
1803	{
1804	rc = VERR_NO_MEMORY;
1805	break;
1806	}
1807	}
1808	}
1809	else
1810	{
1811	RTHCPHYS Phys;
1812	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1813	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1814	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1815	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1816	else
1817	{
1818	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1819	Phys = NIL_RTHCPHYS;
1820	}
1821	if (Phys != NIL_RTHCPHYS)
1822	{
1823	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1824	{
1825	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1826	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1827	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1828	#endif
1829	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1830	/* remap_page_range() limitation on x86 */
1831	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1832	#endif
1833
1834	#if RTLNX_VER_MIN(2,6,11)
1835	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1836	#elif defined(VBOX_USE_PAE_HACK)
1837	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1838	if (!rc)
1839	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1840	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1841	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1842	#else /* 2.4 */
1843	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1844	#endif
1845	if (rc)
1846	{
1847	rc = VERR_NO_MEMORY;
1848	break;
1849	}
1850	}
1851	}
1852	}
1853
1854	#ifdef CONFIG_NUMA_BALANCING
1855	# if RTLNX_VER_MAX(3,13,0) && RTLNX_RHEL_MAX(7,0)
1856	# define VBOX_NUMA_HACK_OLD
1857	# endif
1858	if (RT_SUCCESS(rc))
1859	{
1860	/** @todo Ugly hack! But right now we have no other means to
1861	* disable automatic NUMA page balancing. */
1862	# ifdef RT_OS_X86
1863	# ifdef VBOX_NUMA_HACK_OLD
1864	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1865	# endif
1866	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1867	# else
1868	# ifdef VBOX_NUMA_HACK_OLD
1869	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1870	# endif
1871	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1872	# endif
1873	}
1874	#endif /* CONFIG_NUMA_BALANCING */
1875
1876	LNX_MM_UP_WRITE(pTask->mm);
1877
1878	if (RT_SUCCESS(rc))
1879	{
1880	#ifdef VBOX_USE_PAE_HACK
1881	__free_page(pDummyPage);
1882	#endif
1883	pMemLnx->Core.pv = pv;
1884	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1885	*ppMem = &pMemLnx->Core;
1886	IPRT_LINUX_RESTORE_EFL_AC();
1887	return VINF_SUCCESS;
1888	}
1889
1890	/*
1891	* Bail out.
1892	*/
1893	rtR0MemObjLinuxDoMunmap(pv, cbSub, pTask);
1894	}
1895	rtR0MemObjDelete(&pMemLnx->Core);
1896	}
1897	#ifdef VBOX_USE_PAE_HACK
1898	__free_page(pDummyPage);
1899	#endif
1900
1901	IPRT_LINUX_RESTORE_EFL_AC();
1902	return rc;
1903	}
1904
1905
1906	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1907	{
1908	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
1909	/*
1910	* Currently only supported when we've got addresses PTEs from the kernel.
1911	*/
1912	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1913	if (pMemLnx->pArea && pMemLnx->papPtesForArea)
1914	{
1915	pgprot_t const fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1916	size_t const cPages = (offSub + cbSub) >> PAGE_SHIFT;
1917	pte_t **papPtes = pMemLnx->papPtesForArea;
1918	size_t i;
1919
1920	for (i = offSub >> PAGE_SHIFT; i < cPages; i++)
1921	{
1922	set_pte(papPtes[i], mk_pte(pMemLnx->apPages[i], fPg));
1923	}
1924	preempt_disable();
1925	__flush_tlb_all();
1926	preempt_enable();
1927	return VINF_SUCCESS;
1928	}
1929	# elif defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
1930	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1931	if ( pMemLnx->fExecutable
1932	&& pMemLnx->fMappedToRing0)
1933	{
1934	LNXAPPLYPGRANGE Args;
1935	Args.pMemLnx = pMemLnx;
1936	Args.fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1937	int rcLnx = apply_to_page_range(current->active_mm, (unsigned long)pMemLnx->Core.pv + offSub, cbSub,
1938	rtR0MemObjLinuxApplyPageRange, (void *)&Args);
1939	if (rcLnx)
1940	return VERR_NOT_SUPPORTED;
1941
1942	return VINF_SUCCESS;
1943	}
1944	# endif
1945
1946	NOREF(pMem);
1947	NOREF(offSub);
1948	NOREF(cbSub);
1949	NOREF(fProt);
1950	return VERR_NOT_SUPPORTED;
1951	}
1952
1953
1954	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1955	{
1956	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1957
1958	if (pMemLnx->cPages)
1959	return page_to_phys(pMemLnx->apPages[iPage]);
1960
1961	switch (pMemLnx->Core.enmType)
1962	{
1963	case RTR0MEMOBJTYPE_CONT:
1964	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1965
1966	case RTR0MEMOBJTYPE_PHYS:
1967	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1968
1969	/* the parent knows */
1970	case RTR0MEMOBJTYPE_MAPPING:
1971	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1972
1973	/* cPages > 0 */
1974	case RTR0MEMOBJTYPE_LOW:
1975	case RTR0MEMOBJTYPE_LOCK:
1976	case RTR0MEMOBJTYPE_PHYS_NC:
1977	case RTR0MEMOBJTYPE_PAGE:
1978	default:
1979	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1980	RT_FALL_THROUGH();
1981
1982	case RTR0MEMOBJTYPE_RES_VIRT:
1983	return NIL_RTHCPHYS;
1984	}
1985	}
1986

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 92374

Download in other formats: