memobj-r0drv-linux.c@ 91477

Last change on this file since 91477 was 91477, checked in by vboxsync, 3 years ago
IPRT/memobj-r0drv-linux.c: todo. bugref:10112
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev Revision`
File size: 66.2 KB

Line
1	/* $Id: memobj-r0drv-linux.c 91477 2021-09-29 23:27:09Z 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	/** @todo Try use alloc_pages_bulk_array when available, it should be faster
395	* than a alloc_page loop. Put it in #ifdefs similar to
396	* IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC. */
397	for (iPage = 0; iPage < cPages; iPage++)
398	{
399	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
400	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
401	{
402	while (iPage-- > 0)
403	__free_page(pMemLnx->apPages[iPage]);
404	rtR0MemObjDelete(&pMemLnx->Core);
405	return rcNoMem;
406	}
407	}
408	}
409
410	#else /* < 2.4.22 */
411	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
412	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
413	if (!paPages)
414	{
415	rtR0MemObjDelete(&pMemLnx->Core);
416	return rcNoMem;
417	}
418	for (iPage = 0; iPage < cPages; iPage++)
419	{
420	pMemLnx->apPages[iPage] = &paPages[iPage];
421	if (fExecutable)
422	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
423	if (PageHighMem(pMemLnx->apPages[iPage]))
424	BUG();
425	}
426
427	fContiguous = true;
428	#endif /* < 2.4.22 */
429	pMemLnx->fContiguous = fContiguous;
430	pMemLnx->fExecutable = fExecutable;
431
432	#if RTLNX_VER_MAX(4,5,0)
433	/*
434	* Reserve the pages.
435	*
436	* Linux >= 4.5 with CONFIG_DEBUG_VM panics when setting PG_reserved on compound
437	* pages. According to Michal Hocko this shouldn't be necessary anyway because
438	* as pages which are not on the LRU list are never evictable.
439	*/
440	for (iPage = 0; iPage < cPages; iPage++)
441	SetPageReserved(pMemLnx->apPages[iPage]);
442	#endif
443
444	/*
445	* Note that the physical address of memory allocated with alloc_pages(flags, order)
446	* is always 2^(PAGE_SHIFT+order)-aligned.
447	*/
448	if ( fContiguous
449	&& uAlignment > PAGE_SIZE)
450	{
451	/*
452	* Check for alignment constraints. The physical address of memory allocated with
453	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
454	*/
455	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
456	{
457	/*
458	* This should never happen!
459	*/
460	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
461	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
462	rtR0MemObjLinuxFreePages(pMemLnx);
463	return rcNoMem;
464	}
465	}
466
467	*ppMemLnx = pMemLnx;
468	return VINF_SUCCESS;
469	}
470
471
472	/**
473	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
474	*
475	* This method does NOT free the object.
476	*
477	* @param pMemLnx The object which physical pages should be freed.
478	*/
479	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
480	{
481	size_t iPage = pMemLnx->cPages;
482	if (iPage > 0)
483	{
484	/*
485	* Restore the page flags.
486	*/
487	while (iPage-- > 0)
488	{
489	#if RTLNX_VER_MAX(4,5,0)
490	/* See SetPageReserved() in rtR0MemObjLinuxAllocPages() */
491	ClearPageReserved(pMemLnx->apPages[iPage]);
492	#endif
493	#if RTLNX_VER_MAX(2,4,22)
494	if (pMemLnx->fExecutable)
495	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
496	#endif
497	}
498
499	/*
500	* Free the pages.
501	*/
502	#if RTLNX_VER_MIN(2,4,22)
503	if (!pMemLnx->fContiguous)
504	{
505	iPage = pMemLnx->cPages;
506	while (iPage-- > 0)
507	__free_page(pMemLnx->apPages[iPage]);
508	}
509	else
510	#endif
511	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
512
513	pMemLnx->cPages = 0;
514	}
515	}
516
517
518	#ifdef IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
519	/**
520	* User data passed to the apply_to_page_range() callback.
521	*/
522	typedef struct LNXAPPLYPGRANGE
523	{
524	/** Pointer to the memory object. */
525	PRTR0MEMOBJLNX pMemLnx;
526	/** The page protection flags to apply. */
527	pgprot_t fPg;
528	} LNXAPPLYPGRANGE;
529	/** Pointer to the user data. */
530	typedef LNXAPPLYPGRANGE *PLNXAPPLYPGRANGE;
531	/** Pointer to the const user data. */
532	typedef const LNXAPPLYPGRANGE *PCLNXAPPLYPGRANGE;
533
534	/**
535	* Callback called in apply_to_page_range().
536	*
537	* @returns Linux status code.
538	* @param pPte Pointer to the page table entry for the given address.
539	* @param uAddr The address to apply the new protection to.
540	* @param pvUser The opaque user data.
541	*/
542	static int rtR0MemObjLinuxApplyPageRange(pte_t pPte, unsigned long uAddr, void pvUser)
543	{
544	PCLNXAPPLYPGRANGE pArgs = (PCLNXAPPLYPGRANGE)pvUser;
545	PRTR0MEMOBJLNX pMemLnx = pArgs->pMemLnx;
546	size_t idxPg = (uAddr - (unsigned long)pMemLnx->Core.pv) >> PAGE_SHIFT;
547
548	set_pte(pPte, mk_pte(pMemLnx->apPages[idxPg], pArgs->fPg));
549	return 0;
550	}
551	#endif
552
553
554	/**
555	* Maps the allocation into ring-0.
556	*
557	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
558	*
559	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
560	* space, so we'll use that mapping if possible. If execute access is required, we'll
561	* play safe and do our own mapping.
562	*
563	* @returns IPRT status code.
564	* @param pMemLnx The linux memory object to map.
565	* @param fExecutable Whether execute access is required.
566	*/
567	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
568	{
569	int rc = VINF_SUCCESS;
570
571	/*
572	* Choose mapping strategy.
573	*/
574	bool fMustMap = fExecutable
575	\|\| !pMemLnx->fContiguous;
576	if (!fMustMap)
577	{
578	size_t iPage = pMemLnx->cPages;
579	while (iPage-- > 0)
580	if (PageHighMem(pMemLnx->apPages[iPage]))
581	{
582	fMustMap = true;
583	break;
584	}
585	}
586
587	Assert(!pMemLnx->Core.pv);
588	Assert(!pMemLnx->fMappedToRing0);
589
590	if (fMustMap)
591	{
592	/*
593	* Use vmap - 2.4.22 and later.
594	*/
595	#if RTLNX_VER_MIN(2,4,22)
596	pgprot_t fPg;
597	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
598	# ifdef _PAGE_NX
599	if (!fExecutable)
600	pgprot_val(fPg) \|= _PAGE_NX;
601	# endif
602
603	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
604	if (fExecutable)
605	{
606	# if RTLNX_VER_MIN(3,2,51)
607	pte_t papPtes = (pte_t )kmalloc_array(pMemLnx->cPages, sizeof(papPtes[0]), GFP_KERNEL);
608	# else
609	pte_t papPtes = (pte_t )kmalloc(pMemLnx->cPages * sizeof(papPtes[0]), GFP_KERNEL);
610	# endif
611	if (papPtes)
612	{
613	pMemLnx->pArea = alloc_vm_area(pMemLnx->Core.cb, papPtes); /* Note! pArea->nr_pages is not set. */
614	if (pMemLnx->pArea)
615	{
616	size_t i;
617	Assert(pMemLnx->pArea->size >= pMemLnx->Core.cb); /* Note! includes guard page. */
618	Assert(pMemLnx->pArea->addr);
619	# ifdef _PAGE_NX
620	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
621	# endif
622	pMemLnx->papPtesForArea = papPtes;
623	for (i = 0; i < pMemLnx->cPages; i++)
624	*papPtes[i] = mk_pte(pMemLnx->apPages[i], fPg);
625	pMemLnx->Core.pv = pMemLnx->pArea->addr;
626	pMemLnx->fMappedToRing0 = true;
627	}
628	else
629	{
630	kfree(papPtes);
631	rc = VERR_MAP_FAILED;
632	}
633	}
634	else
635	rc = VERR_MAP_FAILED;
636	}
637	else
638	# endif
639	{
640	# if defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
641	if (fExecutable)
642	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
643	# endif
644
645	# ifdef VM_MAP
646	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
647	# else
648	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
649	# endif
650	if (pMemLnx->Core.pv)
651	pMemLnx->fMappedToRing0 = true;
652	else
653	rc = VERR_MAP_FAILED;
654	}
655	#else /* < 2.4.22 */
656	rc = VERR_NOT_SUPPORTED;
657	#endif
658	}
659	else
660	{
661	/*
662	* Use the kernel RAM mapping.
663	*/
664	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
665	Assert(pMemLnx->Core.pv);
666	}
667
668	return rc;
669	}
670
671
672	/**
673	* Undoes what rtR0MemObjLinuxVMap() did.
674	*
675	* @param pMemLnx The linux memory object.
676	*/
677	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
678	{
679	#if RTLNX_VER_MIN(2,4,22)
680	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
681	if (pMemLnx->pArea)
682	{
683	# if 0
684	pte_t **papPtes = pMemLnx->papPtesForArea;
685	size_t i;
686	for (i = 0; i < pMemLnx->cPages; i++)
687	*papPtes[i] = 0;
688	# endif
689	free_vm_area(pMemLnx->pArea);
690	kfree(pMemLnx->papPtesForArea);
691	pMemLnx->pArea = NULL;
692	pMemLnx->papPtesForArea = NULL;
693	}
694	else
695	# endif
696	if (pMemLnx->fMappedToRing0)
697	{
698	Assert(pMemLnx->Core.pv);
699	vunmap(pMemLnx->Core.pv);
700	pMemLnx->fMappedToRing0 = false;
701	}
702	#else /* < 2.4.22 */
703	Assert(!pMemLnx->fMappedToRing0);
704	#endif
705	pMemLnx->Core.pv = NULL;
706	}
707
708
709	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
710	{
711	IPRT_LINUX_SAVE_EFL_AC();
712	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
713
714	/*
715	* Release any memory that we've allocated or locked.
716	*/
717	switch (pMemLnx->Core.enmType)
718	{
719	case RTR0MEMOBJTYPE_LOW:
720	case RTR0MEMOBJTYPE_PAGE:
721	case RTR0MEMOBJTYPE_CONT:
722	case RTR0MEMOBJTYPE_PHYS:
723	case RTR0MEMOBJTYPE_PHYS_NC:
724	rtR0MemObjLinuxVUnmap(pMemLnx);
725	rtR0MemObjLinuxFreePages(pMemLnx);
726	break;
727
728	case RTR0MEMOBJTYPE_LOCK:
729	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
730	{
731	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
732	size_t iPage;
733	Assert(pTask);
734	if (pTask && pTask->mm)
735	LNX_MM_DOWN_READ(pTask->mm);
736
737	iPage = pMemLnx->cPages;
738	while (iPage-- > 0)
739	{
740	if (!PageReserved(pMemLnx->apPages[iPage]))
741	SetPageDirty(pMemLnx->apPages[iPage]);
742	#if RTLNX_VER_MIN(4,6,0)
743	put_page(pMemLnx->apPages[iPage]);
744	#else
745	page_cache_release(pMemLnx->apPages[iPage]);
746	#endif
747	}
748
749	if (pTask && pTask->mm)
750	LNX_MM_UP_READ(pTask->mm);
751	}
752	/* else: kernel memory - nothing to do here. */
753	break;
754
755	case RTR0MEMOBJTYPE_RES_VIRT:
756	Assert(pMemLnx->Core.pv);
757	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
758	{
759	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
760	Assert(pTask);
761	if (pTask && pTask->mm)
762	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
763	}
764	else
765	{
766	vunmap(pMemLnx->Core.pv);
767
768	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
769	__free_page(pMemLnx->apPages[0]);
770	pMemLnx->apPages[0] = NULL;
771	pMemLnx->cPages = 0;
772	}
773	pMemLnx->Core.pv = NULL;
774	break;
775
776	case RTR0MEMOBJTYPE_MAPPING:
777	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
778	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
779	{
780	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
781	Assert(pTask);
782	if (pTask && pTask->mm)
783	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
784	}
785	else
786	vunmap(pMemLnx->Core.pv);
787	pMemLnx->Core.pv = NULL;
788	break;
789
790	default:
791	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
792	return VERR_INTERNAL_ERROR;
793	}
794	IPRT_LINUX_RESTORE_EFL_ONLY_AC();
795	return VINF_SUCCESS;
796	}
797
798
799	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
800	{
801	IPRT_LINUX_SAVE_EFL_AC();
802	PRTR0MEMOBJLNX pMemLnx;
803	int rc;
804
805	#if RTLNX_VER_MIN(2,4,22)
806	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
807	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY);
808	#else
809	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
810	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY);
811	#endif
812	if (RT_SUCCESS(rc))
813	{
814	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
815	if (RT_SUCCESS(rc))
816	{
817	*ppMem = &pMemLnx->Core;
818	IPRT_LINUX_RESTORE_EFL_AC();
819	return rc;
820	}
821
822	rtR0MemObjLinuxFreePages(pMemLnx);
823	rtR0MemObjDelete(&pMemLnx->Core);
824	}
825
826	IPRT_LINUX_RESTORE_EFL_AC();
827	return rc;
828	}
829
830
831	DECLHIDDEN(int) rtR0MemObjNativeAllocLarge(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, size_t cbLargePage, uint32_t fFlags,
832	const char *pszTag)
833	{
834	return rtR0MemObjFallbackAllocLarge(ppMem, cb, cbLargePage, fFlags, pszTag);
835	}
836
837
838	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
839	{
840	IPRT_LINUX_SAVE_EFL_AC();
841	PRTR0MEMOBJLNX pMemLnx;
842	int rc;
843
844	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
845	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
846	/* ZONE_DMA32: 0-4GB */
847	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
848	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY);
849	if (RT_FAILURE(rc))
850	#endif
851	#ifdef RT_ARCH_AMD64
852	/* ZONE_DMA: 0-16MB */
853	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
854	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY);
855	#else
856	# ifdef CONFIG_X86_PAE
857	# endif
858	/* ZONE_NORMAL: 0-896MB */
859	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
860	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY);
861	#endif
862	if (RT_SUCCESS(rc))
863	{
864	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
865	if (RT_SUCCESS(rc))
866	{
867	*ppMem = &pMemLnx->Core;
868	IPRT_LINUX_RESTORE_EFL_AC();
869	return rc;
870	}
871
872	rtR0MemObjLinuxFreePages(pMemLnx);
873	rtR0MemObjDelete(&pMemLnx->Core);
874	}
875
876	IPRT_LINUX_RESTORE_EFL_AC();
877	return rc;
878	}
879
880
881	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
882	{
883	IPRT_LINUX_SAVE_EFL_AC();
884	PRTR0MEMOBJLNX pMemLnx;
885	int rc;
886
887	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
888	/* ZONE_DMA32: 0-4GB */
889	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
890	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY);
891	if (RT_FAILURE(rc))
892	#endif
893	#ifdef RT_ARCH_AMD64
894	/* ZONE_DMA: 0-16MB */
895	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
896	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY);
897	#else
898	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
899	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
900	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY);
901	#endif
902	if (RT_SUCCESS(rc))
903	{
904	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
905	if (RT_SUCCESS(rc))
906	{
907	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
908	size_t iPage = pMemLnx->cPages;
909	while (iPage-- > 0)
910	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
911	#endif
912	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
913	*ppMem = &pMemLnx->Core;
914	IPRT_LINUX_RESTORE_EFL_AC();
915	return rc;
916	}
917
918	rtR0MemObjLinuxFreePages(pMemLnx);
919	rtR0MemObjDelete(&pMemLnx->Core);
920	}
921
922	IPRT_LINUX_RESTORE_EFL_AC();
923	return rc;
924	}
925
926
927	/**
928	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
929	*
930	* @returns IPRT status code.
931	* @param ppMemLnx Where to
932	* @param enmType The object type.
933	* @param cb The size of the allocation.
934	* @param uAlignment The alignment of the physical memory.
935	* Only valid for fContiguous == true, ignored otherwise.
936	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
937	* @param fGfp The Linux GFP flags to use for the allocation.
938	*/
939	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
940	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, gfp_t fGfp)
941	{
942	PRTR0MEMOBJLNX pMemLnx;
943	int rc;
944
945	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
946	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
947	false /fExecutable/, VERR_NO_PHYS_MEMORY);
948	if (RT_FAILURE(rc))
949	return rc;
950
951	/*
952	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
953	*/
954	if (PhysHighest != NIL_RTHCPHYS)
955	{
956	size_t iPage = pMemLnx->cPages;
957	while (iPage-- > 0)
958	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
959	{
960	rtR0MemObjLinuxFreePages(pMemLnx);
961	rtR0MemObjDelete(&pMemLnx->Core);
962	return VERR_NO_MEMORY;
963	}
964	}
965
966	/*
967	* Complete the object.
968	*/
969	if (enmType == RTR0MEMOBJTYPE_PHYS)
970	{
971	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
972	pMemLnx->Core.u.Phys.fAllocated = true;
973	}
974	*ppMem = &pMemLnx->Core;
975	return rc;
976	}
977
978
979	/**
980	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
981	*
982	* @returns IPRT status code.
983	* @param ppMem Where to store the memory object pointer on success.
984	* @param enmType The object type.
985	* @param cb The size of the allocation.
986	* @param uAlignment The alignment of the physical memory.
987	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
988	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
989	*/
990	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
991	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest)
992	{
993	int rc;
994	IPRT_LINUX_SAVE_EFL_AC();
995
996	/*
997	* There are two clear cases and that's the <=16MB and anything-goes ones.
998	* When the physical address limit is somewhere in-between those two we'll
999	* just have to try, starting with HIGHUSER and working our way thru the
1000	* different types, hoping we'll get lucky.
1001	*
1002	* We should probably move this physical address restriction logic up to
1003	* the page alloc function as it would be more efficient there. But since
1004	* we don't expect this to be a performance issue just yet it can wait.
1005	*/
1006	if (PhysHighest == NIL_RTHCPHYS)
1007	/* ZONE_HIGHMEM: the whole physical memory */
1008	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
1009	else if (PhysHighest <= _1M * 16)
1010	/* ZONE_DMA: 0-16MB */
1011	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
1012	else
1013	{
1014	rc = VERR_NO_MEMORY;
1015	if (RT_FAILURE(rc))
1016	/* ZONE_HIGHMEM: the whole physical memory */
1017	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_HIGHUSER);
1018	if (RT_FAILURE(rc))
1019	/* ZONE_NORMAL: 0-896MB */
1020	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_USER);
1021	#ifdef GFP_DMA32
1022	if (RT_FAILURE(rc))
1023	/* ZONE_DMA32: 0-4GB */
1024	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA32);
1025	#endif
1026	if (RT_FAILURE(rc))
1027	/* ZONE_DMA: 0-16MB */
1028	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, GFP_DMA);
1029	}
1030	IPRT_LINUX_RESTORE_EFL_AC();
1031	return rc;
1032	}
1033
1034
1035	/**
1036	* Translates a kernel virtual address to a linux page structure by walking the
1037	* page tables.
1038	*
1039	* @note We do assume that the page tables will not change as we are walking
1040	* them. This assumption is rather forced by the fact that I could not
1041	* immediately see any way of preventing this from happening. So, we
1042	* take some extra care when accessing them.
1043	*
1044	* Because of this, we don't want to use this function on memory where
1045	* attribute changes to nearby pages is likely to cause large pages to
1046	* be used or split up. So, don't use this for the linear mapping of
1047	* physical memory.
1048	*
1049	* @returns Pointer to the page structur or NULL if it could not be found.
1050	* @param pv The kernel virtual address.
1051	*/
1052	RTDECL(struct page ) rtR0MemObjLinuxVirtToPage(void pv)
1053	{
1054	unsigned long ulAddr = (unsigned long)pv;
1055	unsigned long pfn;
1056	struct page *pPage;
1057	pte_t *pEntry;
1058	union
1059	{
1060	pgd_t Global;
1061	#if RTLNX_VER_MIN(4,12,0)
1062	p4d_t Four;
1063	#endif
1064	#if RTLNX_VER_MIN(2,6,11)
1065	pud_t Upper;
1066	#endif
1067	pmd_t Middle;
1068	pte_t Entry;
1069	} u;
1070
1071	/* Should this happen in a situation this code will be called in? And if
1072	* so, can it change under our feet? See also
1073	* "Documentation/vm/active_mm.txt" in the kernel sources. */
1074	if (RT_UNLIKELY(!current->active_mm))
1075	return NULL;
1076	u.Global = *pgd_offset(current->active_mm, ulAddr);
1077	if (RT_UNLIKELY(pgd_none(u.Global)))
1078	return NULL;
1079	#if RTLNX_VER_MIN(2,6,11)
1080	# if RTLNX_VER_MIN(4,12,0)
1081	u.Four = *p4d_offset(&u.Global, ulAddr);
1082	if (RT_UNLIKELY(p4d_none(u.Four)))
1083	return NULL;
1084	if (p4d_large(u.Four))
1085	{
1086	pPage = p4d_page(u.Four);
1087	AssertReturn(pPage, NULL);
1088	pfn = page_to_pfn(pPage); /* doing the safe way... */
1089	AssertCompile(P4D_SHIFT - PAGE_SHIFT < 31);
1090	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (P4D_SHIFT - PAGE_SHIFT)) - 1);
1091	return pfn_to_page(pfn);
1092	}
1093	u.Upper = *pud_offset(&u.Four, ulAddr);
1094	# else /* < 4.12 */
1095	u.Upper = *pud_offset(&u.Global, ulAddr);
1096	# endif /* < 4.12 */
1097	if (RT_UNLIKELY(pud_none(u.Upper)))
1098	return NULL;
1099	# if RTLNX_VER_MIN(2,6,25)
1100	if (pud_large(u.Upper))
1101	{
1102	pPage = pud_page(u.Upper);
1103	AssertReturn(pPage, NULL);
1104	pfn = page_to_pfn(pPage); /* doing the safe way... */
1105	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
1106	return pfn_to_page(pfn);
1107	}
1108	# endif
1109	u.Middle = *pmd_offset(&u.Upper, ulAddr);
1110	#else /* < 2.6.11 */
1111	u.Middle = *pmd_offset(&u.Global, ulAddr);
1112	#endif /* < 2.6.11 */
1113	if (RT_UNLIKELY(pmd_none(u.Middle)))
1114	return NULL;
1115	#if RTLNX_VER_MIN(2,6,0)
1116	if (pmd_large(u.Middle))
1117	{
1118	pPage = pmd_page(u.Middle);
1119	AssertReturn(pPage, NULL);
1120	pfn = page_to_pfn(pPage); /* doing the safe way... */
1121	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
1122	return pfn_to_page(pfn);
1123	}
1124	#endif
1125
1126	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
1127	pEntry = pte_offset_map(&u.Middle, ulAddr);
1128	#else
1129	pEntry = pte_offset(&u.Middle, ulAddr);
1130	#endif
1131	if (RT_UNLIKELY(!pEntry))
1132	return NULL;
1133	u.Entry = *pEntry;
1134	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map)
1135	pte_unmap(pEntry);
1136	#endif
1137
1138	if (RT_UNLIKELY(!pte_present(u.Entry)))
1139	return NULL;
1140	return pte_page(u.Entry);
1141	}
1142	RT_EXPORT_SYMBOL(rtR0MemObjLinuxVirtToPage);
1143
1144
1145	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment)
1146	{
1147	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest);
1148	}
1149
1150
1151	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
1152	{
1153	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest);
1154	}
1155
1156
1157	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy)
1158	{
1159	IPRT_LINUX_SAVE_EFL_AC();
1160
1161	/*
1162	* All we need to do here is to validate that we can use
1163	* ioremap on the specified address (32/64-bit dma_addr_t).
1164	*/
1165	PRTR0MEMOBJLNX pMemLnx;
1166	dma_addr_t PhysAddr = Phys;
1167	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
1168
1169	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
1170	if (!pMemLnx)
1171	{
1172	IPRT_LINUX_RESTORE_EFL_AC();
1173	return VERR_NO_MEMORY;
1174	}
1175
1176	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
1177	pMemLnx->Core.u.Phys.fAllocated = false;
1178	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
1179	Assert(!pMemLnx->cPages);
1180	*ppMem = &pMemLnx->Core;
1181	IPRT_LINUX_RESTORE_EFL_AC();
1182	return VINF_SUCCESS;
1183	}
1184
1185	/* openSUSE Leap 42.3 detection :-/ */
1186	#if RTLNX_VER_RANGE(4,4,0, 4,6,0) && defined(FAULT_FLAG_REMOTE)
1187	# define GET_USER_PAGES_API KERNEL_VERSION(4, 10, 0) /* no typo! */
1188	#else
1189	# define GET_USER_PAGES_API LINUX_VERSION_CODE
1190	#endif
1191
1192	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess, RTR0PROCESS R0Process)
1193	{
1194	IPRT_LINUX_SAVE_EFL_AC();
1195	const int cPages = cb >> PAGE_SHIFT;
1196	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1197	struct vm_area_struct **papVMAs;
1198	PRTR0MEMOBJLNX pMemLnx;
1199	int rc = VERR_NO_MEMORY;
1200	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1201
1202	/*
1203	* Check for valid task and size overflows.
1204	*/
1205	if (!pTask)
1206	return VERR_NOT_SUPPORTED;
1207	if (((size_t)cPages << PAGE_SHIFT) != cb)
1208	return VERR_OUT_OF_RANGE;
1209
1210	/*
1211	* Allocate the memory object and a temporary buffer for the VMAs.
1212	*/
1213	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
1214	if (!pMemLnx)
1215	{
1216	IPRT_LINUX_RESTORE_EFL_AC();
1217	return VERR_NO_MEMORY;
1218	}
1219
1220	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1221	if (papVMAs)
1222	{
1223	LNX_MM_DOWN_READ(pTask->mm);
1224
1225	/*
1226	* Get user pages.
1227	*/
1228	/** @todo r=bird: Should we not force read access too? */
1229	#if GET_USER_PAGES_API >= KERNEL_VERSION(4, 6, 0)
1230	if (R0Process == RTR0ProcHandleSelf())
1231	rc = get_user_pages(R3Ptr, /* Where from. */
1232	cPages, /* How many pages. */
1233	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1234	fWrite ? FOLL_WRITE \| /* Write to memory. */
1235	FOLL_FORCE /* force write access. */
1236	: 0, /* Write to memory. */
1237	# else
1238	fWrite, /* Write to memory. */
1239	fWrite, /* force write access. */
1240	# endif
1241	&pMemLnx->apPages[0], /* Page array. */
1242	papVMAs); /* vmas */
1243	/*
1244	* Actually this should not happen at the moment as call this function
1245	* only for our own process.
1246	*/
1247	else
1248	rc = get_user_pages_remote(
1249	# if GET_USER_PAGES_API < KERNEL_VERSION(5, 9, 0)
1250	pTask, /* Task for fault accounting. */
1251	# endif
1252	pTask->mm, /* Whose pages. */
1253	R3Ptr, /* Where from. */
1254	cPages, /* How many pages. */
1255	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1256	fWrite ? FOLL_WRITE \| /* Write to memory. */
1257	FOLL_FORCE /* force write access. */
1258	: 0, /* Write to memory. */
1259	# else
1260	fWrite, /* Write to memory. */
1261	fWrite, /* force write access. */
1262	# endif
1263	&pMemLnx->apPages[0], /* Page array. */
1264	papVMAs /* vmas */
1265	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 10, 0)
1266	, NULL /* locked */
1267	# endif
1268	);
1269	#else /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1270	rc = get_user_pages(pTask, /* Task for fault accounting. */
1271	pTask->mm, /* Whose pages. */
1272	R3Ptr, /* Where from. */
1273	cPages, /* How many pages. */
1274	/* The get_user_pages API change was back-ported to 4.4.168. */
1275	# if RTLNX_VER_RANGE(4,4,168, 4,5,0)
1276	fWrite ? FOLL_WRITE \| /* Write to memory. */
1277	FOLL_FORCE /* force write access. */
1278	: 0, /* Write to memory. */
1279	# else
1280	fWrite, /* Write to memory. */
1281	fWrite, /* force write access. */
1282	# endif
1283	&pMemLnx->apPages[0], /* Page array. */
1284	papVMAs); /* vmas */
1285	#endif /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1286	if (rc == cPages)
1287	{
1288	/*
1289	* Flush dcache (required?), protect against fork and _really_ pin the page
1290	* table entries. get_user_pages() will protect against swapping out the
1291	* pages but it will NOT protect against removing page table entries. This
1292	* can be achieved with
1293	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1294	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1295	* Usual Linux distributions support only a limited size of locked pages
1296	* (e.g. 32KB).
1297	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1298	* or by
1299	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1300	* a range check.
1301	*/
1302	/** @todo The Linux fork() protection will require more work if this API
1303	* is to be used for anything but locking VM pages. */
1304	while (rc-- > 0)
1305	{
1306	flush_dcache_page(pMemLnx->apPages[rc]);
1307	papVMAs[rc]->vm_flags \|= VM_DONTCOPY \| VM_LOCKED;
1308	}
1309
1310	LNX_MM_UP_READ(pTask->mm);
1311
1312	RTMemFree(papVMAs);
1313
1314	pMemLnx->Core.u.Lock.R0Process = R0Process;
1315	pMemLnx->cPages = cPages;
1316	Assert(!pMemLnx->fMappedToRing0);
1317	*ppMem = &pMemLnx->Core;
1318
1319	IPRT_LINUX_RESTORE_EFL_AC();
1320	return VINF_SUCCESS;
1321	}
1322
1323	/*
1324	* Failed - we need to unlock any pages that we succeeded to lock.
1325	*/
1326	while (rc-- > 0)
1327	{
1328	if (!PageReserved(pMemLnx->apPages[rc]))
1329	SetPageDirty(pMemLnx->apPages[rc]);
1330	#if RTLNX_VER_MIN(4,6,0)
1331	put_page(pMemLnx->apPages[rc]);
1332	#else
1333	page_cache_release(pMemLnx->apPages[rc]);
1334	#endif
1335	}
1336
1337	LNX_MM_UP_READ(pTask->mm);
1338
1339	RTMemFree(papVMAs);
1340	rc = VERR_LOCK_FAILED;
1341	}
1342
1343	rtR0MemObjDelete(&pMemLnx->Core);
1344	IPRT_LINUX_RESTORE_EFL_AC();
1345	return rc;
1346	}
1347
1348
1349	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb, uint32_t fAccess)
1350	{
1351	IPRT_LINUX_SAVE_EFL_AC();
1352	void pvLast = (uint8_t )pv + cb - 1;
1353	size_t const cPages = cb >> PAGE_SHIFT;
1354	PRTR0MEMOBJLNX pMemLnx;
1355	bool fLinearMapping;
1356	int rc;
1357	uint8_t *pbPage;
1358	size_t iPage;
1359	NOREF(fAccess);
1360
1361	if ( !RTR0MemKernelIsValidAddr(pv)
1362	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1363	return VERR_INVALID_PARAMETER;
1364
1365	/*
1366	* The lower part of the kernel memory has a linear mapping between
1367	* physical and virtual addresses. So we take a short cut here. This is
1368	* assumed to be the cleanest way to handle those addresses (and the code
1369	* is well tested, though the test for determining it is not very nice).
1370	* If we ever decide it isn't we can still remove it.
1371	*/
1372	#if 0
1373	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1374	#else
1375	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1376	&& (unsigned long)pvLast < (unsigned long)high_memory;
1377	#endif
1378
1379	/*
1380	* Allocate the memory object.
1381	*/
1382	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, pv, cb);
1383	if (!pMemLnx)
1384	{
1385	IPRT_LINUX_RESTORE_EFL_AC();
1386	return VERR_NO_MEMORY;
1387	}
1388
1389	/*
1390	* Gather the pages.
1391	* We ASSUME all kernel pages are non-swappable and non-movable.
1392	*/
1393	rc = VINF_SUCCESS;
1394	pbPage = (uint8_t *)pvLast;
1395	iPage = cPages;
1396	if (!fLinearMapping)
1397	{
1398	while (iPage-- > 0)
1399	{
1400	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1401	if (RT_UNLIKELY(!pPage))
1402	{
1403	rc = VERR_LOCK_FAILED;
1404	break;
1405	}
1406	pMemLnx->apPages[iPage] = pPage;
1407	pbPage -= PAGE_SIZE;
1408	}
1409	}
1410	else
1411	{
1412	while (iPage-- > 0)
1413	{
1414	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1415	pbPage -= PAGE_SIZE;
1416	}
1417	}
1418	if (RT_SUCCESS(rc))
1419	{
1420	/*
1421	* Complete the memory object and return.
1422	*/
1423	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1424	pMemLnx->cPages = cPages;
1425	Assert(!pMemLnx->fMappedToRing0);
1426	*ppMem = &pMemLnx->Core;
1427
1428	IPRT_LINUX_RESTORE_EFL_AC();
1429	return VINF_SUCCESS;
1430	}
1431
1432	rtR0MemObjDelete(&pMemLnx->Core);
1433	IPRT_LINUX_RESTORE_EFL_AC();
1434	return rc;
1435	}
1436
1437
1438	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
1439	{
1440	#if RTLNX_VER_MIN(2,4,22)
1441	IPRT_LINUX_SAVE_EFL_AC();
1442	const size_t cPages = cb >> PAGE_SHIFT;
1443	struct page *pDummyPage;
1444	struct page **papPages;
1445
1446	/* check for unsupported stuff. */
1447	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1448	if (uAlignment > PAGE_SIZE)
1449	return VERR_NOT_SUPPORTED;
1450
1451	/*
1452	* Allocate a dummy page and create a page pointer array for vmap such that
1453	* the dummy page is mapped all over the reserved area.
1454	*/
1455	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1456	if (pDummyPage)
1457	{
1458	papPages = RTMemAlloc(sizeof(papPages) cPages);
1459	if (papPages)
1460	{
1461	void *pv;
1462	size_t iPage = cPages;
1463	while (iPage-- > 0)
1464	papPages[iPage] = pDummyPage;
1465	# ifdef VM_MAP
1466	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1467	# else
1468	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1469	# endif
1470	RTMemFree(papPages);
1471	if (pv)
1472	{
1473	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1474	if (pMemLnx)
1475	{
1476	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1477	pMemLnx->cPages = 1;
1478	pMemLnx->apPages[0] = pDummyPage;
1479	*ppMem = &pMemLnx->Core;
1480	IPRT_LINUX_RESTORE_EFL_AC();
1481	return VINF_SUCCESS;
1482	}
1483	vunmap(pv);
1484	}
1485	}
1486	__free_page(pDummyPage);
1487	}
1488	IPRT_LINUX_RESTORE_EFL_AC();
1489	return VERR_NO_MEMORY;
1490
1491	#else /* < 2.4.22 */
1492	/*
1493	* Could probably use ioremap here, but the caller is in a better position than us
1494	* to select some safe physical memory.
1495	*/
1496	return VERR_NOT_SUPPORTED;
1497	#endif
1498	}
1499
1500
1501	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
1502	{
1503	IPRT_LINUX_SAVE_EFL_AC();
1504	PRTR0MEMOBJLNX pMemLnx;
1505	void *pv;
1506	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1507	if (!pTask)
1508	return VERR_NOT_SUPPORTED;
1509
1510	/*
1511	* Check that the specified alignment is supported.
1512	*/
1513	if (uAlignment > PAGE_SIZE)
1514	return VERR_NOT_SUPPORTED;
1515
1516	/*
1517	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1518	*/
1519	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1520	if (pv == (void *)-1)
1521	{
1522	IPRT_LINUX_RESTORE_EFL_AC();
1523	return VERR_NO_MEMORY;
1524	}
1525
1526	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
1527	if (!pMemLnx)
1528	{
1529	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1530	IPRT_LINUX_RESTORE_EFL_AC();
1531	return VERR_NO_MEMORY;
1532	}
1533
1534	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1535	*ppMem = &pMemLnx->Core;
1536	IPRT_LINUX_RESTORE_EFL_AC();
1537	return VINF_SUCCESS;
1538	}
1539
1540
1541	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap,
1542	void *pvFixed, size_t uAlignment,
1543	unsigned fProt, size_t offSub, size_t cbSub)
1544	{
1545	int rc = VERR_NO_MEMORY;
1546	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1547	PRTR0MEMOBJLNX pMemLnx;
1548	IPRT_LINUX_SAVE_EFL_AC();
1549
1550	/* Fail if requested to do something we can't. */
1551	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1552	if (uAlignment > PAGE_SIZE)
1553	return VERR_NOT_SUPPORTED;
1554
1555	/*
1556	* Create the IPRT memory object.
1557	*/
1558	if (!cbSub)
1559	cbSub = pMemLnxToMap->Core.cb - offSub;
1560	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub);
1561	if (pMemLnx)
1562	{
1563	if (pMemLnxToMap->cPages)
1564	{
1565	#if RTLNX_VER_MIN(2,4,22)
1566	/*
1567	* Use vmap - 2.4.22 and later.
1568	*/
1569	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1570	/** @todo We don't really care too much for EXEC here... 5.8 always adds NX. */
1571	Assert(((offSub + cbSub) >> PAGE_SHIFT) <= pMemLnxToMap->cPages);
1572	# ifdef VM_MAP
1573	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_MAP, fPg);
1574	# else
1575	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_ALLOC, fPg);
1576	# endif
1577	if (pMemLnx->Core.pv)
1578	{
1579	pMemLnx->fMappedToRing0 = true;
1580	rc = VINF_SUCCESS;
1581	}
1582	else
1583	rc = VERR_MAP_FAILED;
1584
1585	#else /* < 2.4.22 */
1586	/*
1587	* Only option here is to share mappings if possible and forget about fProt.
1588	*/
1589	if (rtR0MemObjIsRing3(pMemToMap))
1590	rc = VERR_NOT_SUPPORTED;
1591	else
1592	{
1593	rc = VINF_SUCCESS;
1594	if (!pMemLnxToMap->Core.pv)
1595	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1596	if (RT_SUCCESS(rc))
1597	{
1598	Assert(pMemLnxToMap->Core.pv);
1599	pMemLnx->Core.pv = (uint8_t *)pMemLnxToMap->Core.pv + offSub;
1600	}
1601	}
1602	#endif
1603	}
1604	else
1605	{
1606	/*
1607	* MMIO / physical memory.
1608	*/
1609	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1610	#if RTLNX_VER_MIN(2,6,25)
1611	/*
1612	* ioremap() defaults to no caching since the 2.6 kernels.
1613	* ioremap_nocache() has been removed finally in 5.6-rc1.
1614	*/
1615	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1616	? ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1617	: ioremap_cache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1618	#else /* KERNEL_VERSION < 2.6.25 */
1619	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1620	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1621	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1622	#endif /* KERNEL_VERSION < 2.6.25 */
1623	if (pMemLnx->Core.pv)
1624	{
1625	/** @todo fix protection. */
1626	rc = VINF_SUCCESS;
1627	}
1628	}
1629	if (RT_SUCCESS(rc))
1630	{
1631	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1632	*ppMem = &pMemLnx->Core;
1633	IPRT_LINUX_RESTORE_EFL_AC();
1634	return VINF_SUCCESS;
1635	}
1636	rtR0MemObjDelete(&pMemLnx->Core);
1637	}
1638
1639	IPRT_LINUX_RESTORE_EFL_AC();
1640	return rc;
1641	}
1642
1643
1644	#ifdef VBOX_USE_PAE_HACK
1645	/**
1646	* Replace the PFN of a PTE with the address of the actual page.
1647	*
1648	* The caller maps a reserved dummy page at the address with the desired access
1649	* and flags.
1650	*
1651	* This hack is required for older Linux kernels which don't provide
1652	* remap_pfn_range().
1653	*
1654	* @returns 0 on success, -ENOMEM on failure.
1655	* @param mm The memory context.
1656	* @param ulAddr The mapping address.
1657	* @param Phys The physical address of the page to map.
1658	*/
1659	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1660	{
1661	int rc = -ENOMEM;
1662	pgd_t *pgd;
1663
1664	spin_lock(&mm->page_table_lock);
1665
1666	pgd = pgd_offset(mm, ulAddr);
1667	if (!pgd_none(pgd) && !pgd_bad(pgd))
1668	{
1669	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1670	if (!pmd_none(*pmd))
1671	{
1672	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1673	if (ptep)
1674	{
1675	pte_t pte = *ptep;
1676	pte.pte_high &= 0xfff00000;
1677	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1678	pte.pte_low &= 0x00000fff;
1679	pte.pte_low \|= (Phys & 0xfffff000);
1680	set_pte(ptep, pte);
1681	pte_unmap(ptep);
1682	rc = 0;
1683	}
1684	}
1685	}
1686
1687	spin_unlock(&mm->page_table_lock);
1688	return rc;
1689	}
1690	#endif /* VBOX_USE_PAE_HACK */
1691
1692
1693	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment,
1694	unsigned fProt, RTR0PROCESS R0Process, size_t offSub, size_t cbSub)
1695	{
1696	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1697	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1698	int rc = VERR_NO_MEMORY;
1699	PRTR0MEMOBJLNX pMemLnx;
1700	#ifdef VBOX_USE_PAE_HACK
1701	struct page *pDummyPage;
1702	RTHCPHYS DummyPhys;
1703	#endif
1704	IPRT_LINUX_SAVE_EFL_AC();
1705
1706	/*
1707	* Check for restrictions.
1708	*/
1709	if (!pTask)
1710	return VERR_NOT_SUPPORTED;
1711	if (uAlignment > PAGE_SIZE)
1712	return VERR_NOT_SUPPORTED;
1713
1714	#ifdef VBOX_USE_PAE_HACK
1715	/*
1716	* Allocate a dummy page for use when mapping the memory.
1717	*/
1718	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1719	if (!pDummyPage)
1720	{
1721	IPRT_LINUX_RESTORE_EFL_AC();
1722	return VERR_NO_MEMORY;
1723	}
1724	SetPageReserved(pDummyPage);
1725	DummyPhys = page_to_phys(pDummyPage);
1726	#endif
1727
1728	/*
1729	* Create the IPRT memory object.
1730	*/
1731	Assert(!offSub \|\| cbSub);
1732	if (cbSub == 0)
1733	cbSub = pMemLnxToMap->Core.cb;
1734	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub);
1735	if (pMemLnx)
1736	{
1737	/*
1738	* Allocate user space mapping.
1739	*/
1740	void *pv;
1741	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cbSub, uAlignment, pTask, fProt);
1742	if (pv != (void *)-1)
1743	{
1744	/*
1745	* Map page by page into the mmap area.
1746	* This is generic, paranoid and not very efficient.
1747	*/
1748	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1749	unsigned long ulAddrCur = (unsigned long)pv;
1750	const size_t cPages = (offSub + cbSub) >> PAGE_SHIFT;
1751	size_t iPage;
1752
1753	LNX_MM_DOWN_WRITE(pTask->mm);
1754
1755	rc = VINF_SUCCESS;
1756	if (pMemLnxToMap->cPages)
1757	{
1758	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1759	{
1760	#if RTLNX_VER_MAX(2,6,11)
1761	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1762	#endif
1763	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1764	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1765	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1766	#endif
1767	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1768	/* remap_page_range() limitation on x86 */
1769	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1770	#endif
1771
1772	#if defined(VBOX_USE_INSERT_PAGE) && RTLNX_VER_MIN(2,6,22)
1773	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1774	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1775	* See remap_pfn_range() in mm/memory.c */
1776	#if RTLNX_VER_MIN(3,7,0)
1777	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1778	#else
1779	vma->vm_flags \|= VM_RESERVED;
1780	#endif
1781	#elif RTLNX_VER_MIN(2,6,11)
1782	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1783	#elif defined(VBOX_USE_PAE_HACK)
1784	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1785	if (!rc)
1786	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1787	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1788	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1789	#else /* 2.4 */
1790	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1791	#endif
1792	if (rc)
1793	{
1794	rc = VERR_NO_MEMORY;
1795	break;
1796	}
1797	}
1798	}
1799	else
1800	{
1801	RTHCPHYS Phys;
1802	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1803	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1804	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1805	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1806	else
1807	{
1808	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1809	Phys = NIL_RTHCPHYS;
1810	}
1811	if (Phys != NIL_RTHCPHYS)
1812	{
1813	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1814	{
1815	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1816	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1817	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1818	#endif
1819	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1820	/* remap_page_range() limitation on x86 */
1821	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1822	#endif
1823
1824	#if RTLNX_VER_MIN(2,6,11)
1825	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1826	#elif defined(VBOX_USE_PAE_HACK)
1827	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1828	if (!rc)
1829	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1830	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1831	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1832	#else /* 2.4 */
1833	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1834	#endif
1835	if (rc)
1836	{
1837	rc = VERR_NO_MEMORY;
1838	break;
1839	}
1840	}
1841	}
1842	}
1843
1844	#ifdef CONFIG_NUMA_BALANCING
1845	# if RTLNX_VER_MAX(3,13,0) && RTLNX_RHEL_MAX(7,0)
1846	# define VBOX_NUMA_HACK_OLD
1847	# endif
1848	if (RT_SUCCESS(rc))
1849	{
1850	/** @todo Ugly hack! But right now we have no other means to
1851	* disable automatic NUMA page balancing. */
1852	# ifdef RT_OS_X86
1853	# ifdef VBOX_NUMA_HACK_OLD
1854	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1855	# endif
1856	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1857	# else
1858	# ifdef VBOX_NUMA_HACK_OLD
1859	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1860	# endif
1861	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1862	# endif
1863	}
1864	#endif /* CONFIG_NUMA_BALANCING */
1865
1866	LNX_MM_UP_WRITE(pTask->mm);
1867
1868	if (RT_SUCCESS(rc))
1869	{
1870	#ifdef VBOX_USE_PAE_HACK
1871	__free_page(pDummyPage);
1872	#endif
1873	pMemLnx->Core.pv = pv;
1874	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1875	*ppMem = &pMemLnx->Core;
1876	IPRT_LINUX_RESTORE_EFL_AC();
1877	return VINF_SUCCESS;
1878	}
1879
1880	/*
1881	* Bail out.
1882	*/
1883	rtR0MemObjLinuxDoMunmap(pv, cbSub, pTask);
1884	}
1885	rtR0MemObjDelete(&pMemLnx->Core);
1886	}
1887	#ifdef VBOX_USE_PAE_HACK
1888	__free_page(pDummyPage);
1889	#endif
1890
1891	IPRT_LINUX_RESTORE_EFL_AC();
1892	return rc;
1893	}
1894
1895
1896	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1897	{
1898	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
1899	/*
1900	* Currently only supported when we've got addresses PTEs from the kernel.
1901	*/
1902	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1903	if (pMemLnx->pArea && pMemLnx->papPtesForArea)
1904	{
1905	pgprot_t const fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1906	size_t const cPages = (offSub + cbSub) >> PAGE_SHIFT;
1907	pte_t **papPtes = pMemLnx->papPtesForArea;
1908	size_t i;
1909
1910	for (i = offSub >> PAGE_SHIFT; i < cPages; i++)
1911	{
1912	set_pte(papPtes[i], mk_pte(pMemLnx->apPages[i], fPg));
1913	}
1914	preempt_disable();
1915	__flush_tlb_all();
1916	preempt_enable();
1917	return VINF_SUCCESS;
1918	}
1919	# elif defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
1920	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1921	if ( pMemLnx->fExecutable
1922	&& pMemLnx->fMappedToRing0)
1923	{
1924	LNXAPPLYPGRANGE Args;
1925	Args.pMemLnx = pMemLnx;
1926	Args.fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1927	int rcLnx = apply_to_page_range(current->active_mm, (unsigned long)pMemLnx->Core.pv + offSub, cbSub,
1928	rtR0MemObjLinuxApplyPageRange, (void *)&Args);
1929	if (rcLnx)
1930	return VERR_NOT_SUPPORTED;
1931
1932	return VINF_SUCCESS;
1933	}
1934	# endif
1935
1936	NOREF(pMem);
1937	NOREF(offSub);
1938	NOREF(cbSub);
1939	NOREF(fProt);
1940	return VERR_NOT_SUPPORTED;
1941	}
1942
1943
1944	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1945	{
1946	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1947
1948	if (pMemLnx->cPages)
1949	return page_to_phys(pMemLnx->apPages[iPage]);
1950
1951	switch (pMemLnx->Core.enmType)
1952	{
1953	case RTR0MEMOBJTYPE_CONT:
1954	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1955
1956	case RTR0MEMOBJTYPE_PHYS:
1957	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1958
1959	/* the parent knows */
1960	case RTR0MEMOBJTYPE_MAPPING:
1961	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1962
1963	/* cPages > 0 */
1964	case RTR0MEMOBJTYPE_LOW:
1965	case RTR0MEMOBJTYPE_LOCK:
1966	case RTR0MEMOBJTYPE_PHYS_NC:
1967	case RTR0MEMOBJTYPE_PAGE:
1968	default:
1969	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1970	RT_FALL_THROUGH();
1971
1972	case RTR0MEMOBJTYPE_RES_VIRT:
1973	return NIL_RTHCPHYS;
1974	}
1975	}
1976

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 91477

Download in other formats: