memobj-r0drv-linux.c@ 91482

Last change on this file since 91482 was 91482, checked in by vboxsync, 3 years ago
IPRT/memobj: Passing pszTag around...
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev Revision`
File size: 67.2 KB

Line
1	/* $Id: memobj-r0drv-linux.c 91482 2021-09-30 00:12:26Z 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->cPages = cPages;
356
357	if (cPages > 255)
358	{
359	# ifdef __GFP_REPEAT
360	/* Try hard to allocate the memory, but the allocation attempt might fail. */
361	fFlagsLnx \|= __GFP_REPEAT;
362	# endif
363	# ifdef __GFP_NOMEMALLOC
364	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
365	fFlagsLnx \|= __GFP_NOMEMALLOC;
366	# endif
367	}
368
369	/*
370	* Allocate the pages.
371	* For small allocations we'll try contiguous first and then fall back on page by page.
372	*/
373	#if RTLNX_VER_MIN(2,4,22)
374	if ( fContiguous
375	\|\| cb <= PAGE_SIZE * 2)
376	{
377	# ifdef VBOX_USE_INSERT_PAGE
378	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
379	# else
380	paPages = alloc_pages(fFlagsLnx \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
381	# endif
382	if (paPages)
383	{
384	fContiguous = true;
385	for (iPage = 0; iPage < cPages; iPage++)
386	pMemLnx->apPages[iPage] = &paPages[iPage];
387	}
388	else if (fContiguous)
389	{
390	rtR0MemObjDelete(&pMemLnx->Core);
391	return rcNoMem;
392	}
393	}
394
395	if (!fContiguous)
396	{
397	/** @todo Try use alloc_pages_bulk_array when available, it should be faster
398	* than a alloc_page loop. Put it in #ifdefs similar to
399	* IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC. */
400	for (iPage = 0; iPage < cPages; iPage++)
401	{
402	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
403	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
404	{
405	while (iPage-- > 0)
406	__free_page(pMemLnx->apPages[iPage]);
407	rtR0MemObjDelete(&pMemLnx->Core);
408	return rcNoMem;
409	}
410	}
411	}
412
413	#else /* < 2.4.22 */
414	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
415	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
416	if (!paPages)
417	{
418	rtR0MemObjDelete(&pMemLnx->Core);
419	return rcNoMem;
420	}
421	for (iPage = 0; iPage < cPages; iPage++)
422	{
423	pMemLnx->apPages[iPage] = &paPages[iPage];
424	if (fExecutable)
425	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
426	if (PageHighMem(pMemLnx->apPages[iPage]))
427	BUG();
428	}
429
430	fContiguous = true;
431	#endif /* < 2.4.22 */
432	pMemLnx->fContiguous = fContiguous;
433	pMemLnx->fExecutable = fExecutable;
434
435	#if RTLNX_VER_MAX(4,5,0)
436	/*
437	* Reserve the pages.
438	*
439	* Linux >= 4.5 with CONFIG_DEBUG_VM panics when setting PG_reserved on compound
440	* pages. According to Michal Hocko this shouldn't be necessary anyway because
441	* as pages which are not on the LRU list are never evictable.
442	*/
443	for (iPage = 0; iPage < cPages; iPage++)
444	SetPageReserved(pMemLnx->apPages[iPage]);
445	#endif
446
447	/*
448	* Note that the physical address of memory allocated with alloc_pages(flags, order)
449	* is always 2^(PAGE_SHIFT+order)-aligned.
450	*/
451	if ( fContiguous
452	&& uAlignment > PAGE_SIZE)
453	{
454	/*
455	* Check for alignment constraints. The physical address of memory allocated with
456	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
457	*/
458	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
459	{
460	/*
461	* This should never happen!
462	*/
463	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
464	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
465	rtR0MemObjLinuxFreePages(pMemLnx);
466	return rcNoMem;
467	}
468	}
469
470	*ppMemLnx = pMemLnx;
471	return VINF_SUCCESS;
472	}
473
474
475	/**
476	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
477	*
478	* This method does NOT free the object.
479	*
480	* @param pMemLnx The object which physical pages should be freed.
481	*/
482	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
483	{
484	size_t iPage = pMemLnx->cPages;
485	if (iPage > 0)
486	{
487	/*
488	* Restore the page flags.
489	*/
490	while (iPage-- > 0)
491	{
492	#if RTLNX_VER_MAX(4,5,0)
493	/* See SetPageReserved() in rtR0MemObjLinuxAllocPages() */
494	ClearPageReserved(pMemLnx->apPages[iPage]);
495	#endif
496	#if RTLNX_VER_MAX(2,4,22)
497	if (pMemLnx->fExecutable)
498	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
499	#endif
500	}
501
502	/*
503	* Free the pages.
504	*/
505	#if RTLNX_VER_MIN(2,4,22)
506	if (!pMemLnx->fContiguous)
507	{
508	iPage = pMemLnx->cPages;
509	while (iPage-- > 0)
510	__free_page(pMemLnx->apPages[iPage]);
511	}
512	else
513	#endif
514	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
515
516	pMemLnx->cPages = 0;
517	}
518	}
519
520
521	#ifdef IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
522	/**
523	* User data passed to the apply_to_page_range() callback.
524	*/
525	typedef struct LNXAPPLYPGRANGE
526	{
527	/** Pointer to the memory object. */
528	PRTR0MEMOBJLNX pMemLnx;
529	/** The page protection flags to apply. */
530	pgprot_t fPg;
531	} LNXAPPLYPGRANGE;
532	/** Pointer to the user data. */
533	typedef LNXAPPLYPGRANGE *PLNXAPPLYPGRANGE;
534	/** Pointer to the const user data. */
535	typedef const LNXAPPLYPGRANGE *PCLNXAPPLYPGRANGE;
536
537	/**
538	* Callback called in apply_to_page_range().
539	*
540	* @returns Linux status code.
541	* @param pPte Pointer to the page table entry for the given address.
542	* @param uAddr The address to apply the new protection to.
543	* @param pvUser The opaque user data.
544	*/
545	static int rtR0MemObjLinuxApplyPageRange(pte_t pPte, unsigned long uAddr, void pvUser)
546	{
547	PCLNXAPPLYPGRANGE pArgs = (PCLNXAPPLYPGRANGE)pvUser;
548	PRTR0MEMOBJLNX pMemLnx = pArgs->pMemLnx;
549	size_t idxPg = (uAddr - (unsigned long)pMemLnx->Core.pv) >> PAGE_SHIFT;
550
551	set_pte(pPte, mk_pte(pMemLnx->apPages[idxPg], pArgs->fPg));
552	return 0;
553	}
554	#endif
555
556
557	/**
558	* Maps the allocation into ring-0.
559	*
560	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
561	*
562	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
563	* space, so we'll use that mapping if possible. If execute access is required, we'll
564	* play safe and do our own mapping.
565	*
566	* @returns IPRT status code.
567	* @param pMemLnx The linux memory object to map.
568	* @param fExecutable Whether execute access is required.
569	*/
570	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
571	{
572	int rc = VINF_SUCCESS;
573
574	/*
575	* Choose mapping strategy.
576	*/
577	bool fMustMap = fExecutable
578	\|\| !pMemLnx->fContiguous;
579	if (!fMustMap)
580	{
581	size_t iPage = pMemLnx->cPages;
582	while (iPage-- > 0)
583	if (PageHighMem(pMemLnx->apPages[iPage]))
584	{
585	fMustMap = true;
586	break;
587	}
588	}
589
590	Assert(!pMemLnx->Core.pv);
591	Assert(!pMemLnx->fMappedToRing0);
592
593	if (fMustMap)
594	{
595	/*
596	* Use vmap - 2.4.22 and later.
597	*/
598	#if RTLNX_VER_MIN(2,4,22)
599	pgprot_t fPg;
600	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
601	# ifdef _PAGE_NX
602	if (!fExecutable)
603	pgprot_val(fPg) \|= _PAGE_NX;
604	# endif
605
606	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
607	if (fExecutable)
608	{
609	# if RTLNX_VER_MIN(3,2,51)
610	pte_t papPtes = (pte_t )kmalloc_array(pMemLnx->cPages, sizeof(papPtes[0]), GFP_KERNEL);
611	# else
612	pte_t papPtes = (pte_t )kmalloc(pMemLnx->cPages * sizeof(papPtes[0]), GFP_KERNEL);
613	# endif
614	if (papPtes)
615	{
616	pMemLnx->pArea = alloc_vm_area(pMemLnx->Core.cb, papPtes); /* Note! pArea->nr_pages is not set. */
617	if (pMemLnx->pArea)
618	{
619	size_t i;
620	Assert(pMemLnx->pArea->size >= pMemLnx->Core.cb); /* Note! includes guard page. */
621	Assert(pMemLnx->pArea->addr);
622	# ifdef _PAGE_NX
623	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
624	# endif
625	pMemLnx->papPtesForArea = papPtes;
626	for (i = 0; i < pMemLnx->cPages; i++)
627	*papPtes[i] = mk_pte(pMemLnx->apPages[i], fPg);
628	pMemLnx->Core.pv = pMemLnx->pArea->addr;
629	pMemLnx->fMappedToRing0 = true;
630	}
631	else
632	{
633	kfree(papPtes);
634	rc = VERR_MAP_FAILED;
635	}
636	}
637	else
638	rc = VERR_MAP_FAILED;
639	}
640	else
641	# endif
642	{
643	# if defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
644	if (fExecutable)
645	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
646	# endif
647
648	# ifdef VM_MAP
649	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
650	# else
651	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
652	# endif
653	if (pMemLnx->Core.pv)
654	pMemLnx->fMappedToRing0 = true;
655	else
656	rc = VERR_MAP_FAILED;
657	}
658	#else /* < 2.4.22 */
659	rc = VERR_NOT_SUPPORTED;
660	#endif
661	}
662	else
663	{
664	/*
665	* Use the kernel RAM mapping.
666	*/
667	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
668	Assert(pMemLnx->Core.pv);
669	}
670
671	return rc;
672	}
673
674
675	/**
676	* Undoes what rtR0MemObjLinuxVMap() did.
677	*
678	* @param pMemLnx The linux memory object.
679	*/
680	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
681	{
682	#if RTLNX_VER_MIN(2,4,22)
683	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
684	if (pMemLnx->pArea)
685	{
686	# if 0
687	pte_t **papPtes = pMemLnx->papPtesForArea;
688	size_t i;
689	for (i = 0; i < pMemLnx->cPages; i++)
690	*papPtes[i] = 0;
691	# endif
692	free_vm_area(pMemLnx->pArea);
693	kfree(pMemLnx->papPtesForArea);
694	pMemLnx->pArea = NULL;
695	pMemLnx->papPtesForArea = NULL;
696	}
697	else
698	# endif
699	if (pMemLnx->fMappedToRing0)
700	{
701	Assert(pMemLnx->Core.pv);
702	vunmap(pMemLnx->Core.pv);
703	pMemLnx->fMappedToRing0 = false;
704	}
705	#else /* < 2.4.22 */
706	Assert(!pMemLnx->fMappedToRing0);
707	#endif
708	pMemLnx->Core.pv = NULL;
709	}
710
711
712	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
713	{
714	IPRT_LINUX_SAVE_EFL_AC();
715	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
716
717	/*
718	* Release any memory that we've allocated or locked.
719	*/
720	switch (pMemLnx->Core.enmType)
721	{
722	case RTR0MEMOBJTYPE_LOW:
723	case RTR0MEMOBJTYPE_PAGE:
724	case RTR0MEMOBJTYPE_CONT:
725	case RTR0MEMOBJTYPE_PHYS:
726	case RTR0MEMOBJTYPE_PHYS_NC:
727	rtR0MemObjLinuxVUnmap(pMemLnx);
728	rtR0MemObjLinuxFreePages(pMemLnx);
729	break;
730
731	case RTR0MEMOBJTYPE_LOCK:
732	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
733	{
734	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
735	size_t iPage;
736	Assert(pTask);
737	if (pTask && pTask->mm)
738	LNX_MM_DOWN_READ(pTask->mm);
739
740	iPage = pMemLnx->cPages;
741	while (iPage-- > 0)
742	{
743	if (!PageReserved(pMemLnx->apPages[iPage]))
744	SetPageDirty(pMemLnx->apPages[iPage]);
745	#if RTLNX_VER_MIN(4,6,0)
746	put_page(pMemLnx->apPages[iPage]);
747	#else
748	page_cache_release(pMemLnx->apPages[iPage]);
749	#endif
750	}
751
752	if (pTask && pTask->mm)
753	LNX_MM_UP_READ(pTask->mm);
754	}
755	/* else: kernel memory - nothing to do here. */
756	break;
757
758	case RTR0MEMOBJTYPE_RES_VIRT:
759	Assert(pMemLnx->Core.pv);
760	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
761	{
762	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
763	Assert(pTask);
764	if (pTask && pTask->mm)
765	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
766	}
767	else
768	{
769	vunmap(pMemLnx->Core.pv);
770
771	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
772	__free_page(pMemLnx->apPages[0]);
773	pMemLnx->apPages[0] = NULL;
774	pMemLnx->cPages = 0;
775	}
776	pMemLnx->Core.pv = NULL;
777	break;
778
779	case RTR0MEMOBJTYPE_MAPPING:
780	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
781	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
782	{
783	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
784	Assert(pTask);
785	if (pTask && pTask->mm)
786	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
787	}
788	else
789	vunmap(pMemLnx->Core.pv);
790	pMemLnx->Core.pv = NULL;
791	break;
792
793	default:
794	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
795	return VERR_INTERNAL_ERROR;
796	}
797	IPRT_LINUX_RESTORE_EFL_ONLY_AC();
798	return VINF_SUCCESS;
799	}
800
801
802	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
803	{
804	IPRT_LINUX_SAVE_EFL_AC();
805	PRTR0MEMOBJLNX pMemLnx;
806	int rc;
807
808	#if RTLNX_VER_MIN(2,4,22)
809	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
810	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, NULL);
811	#else
812	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
813	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, NULL);
814	#endif
815	if (RT_SUCCESS(rc))
816	{
817	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
818	if (RT_SUCCESS(rc))
819	{
820	*ppMem = &pMemLnx->Core;
821	IPRT_LINUX_RESTORE_EFL_AC();
822	return rc;
823	}
824
825	rtR0MemObjLinuxFreePages(pMemLnx);
826	rtR0MemObjDelete(&pMemLnx->Core);
827	}
828
829	IPRT_LINUX_RESTORE_EFL_AC();
830	return rc;
831	}
832
833
834	DECLHIDDEN(int) rtR0MemObjNativeAllocLarge(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, size_t cbLargePage, uint32_t fFlags,
835	const char *pszTag)
836	{
837	return rtR0MemObjFallbackAllocLarge(ppMem, cb, cbLargePage, fFlags, pszTag);
838	}
839
840
841	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
842	{
843	IPRT_LINUX_SAVE_EFL_AC();
844	PRTR0MEMOBJLNX pMemLnx;
845	int rc;
846
847	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
848	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
849	/* ZONE_DMA32: 0-4GB */
850	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
851	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, NULL);
852	if (RT_FAILURE(rc))
853	#endif
854	#ifdef RT_ARCH_AMD64
855	/* ZONE_DMA: 0-16MB */
856	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
857	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, NULL);
858	#else
859	# ifdef CONFIG_X86_PAE
860	# endif
861	/* ZONE_NORMAL: 0-896MB */
862	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
863	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, NULL);
864	#endif
865	if (RT_SUCCESS(rc))
866	{
867	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
868	if (RT_SUCCESS(rc))
869	{
870	*ppMem = &pMemLnx->Core;
871	IPRT_LINUX_RESTORE_EFL_AC();
872	return rc;
873	}
874
875	rtR0MemObjLinuxFreePages(pMemLnx);
876	rtR0MemObjDelete(&pMemLnx->Core);
877	}
878
879	IPRT_LINUX_RESTORE_EFL_AC();
880	return rc;
881	}
882
883
884	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
885	{
886	IPRT_LINUX_SAVE_EFL_AC();
887	PRTR0MEMOBJLNX pMemLnx;
888	int rc;
889
890	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
891	/* ZONE_DMA32: 0-4GB */
892	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
893	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, NULL);
894	if (RT_FAILURE(rc))
895	#endif
896	#ifdef RT_ARCH_AMD64
897	/* ZONE_DMA: 0-16MB */
898	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
899	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, NULL);
900	#else
901	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
902	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
903	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, NULL);
904	#endif
905	if (RT_SUCCESS(rc))
906	{
907	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
908	if (RT_SUCCESS(rc))
909	{
910	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
911	size_t iPage = pMemLnx->cPages;
912	while (iPage-- > 0)
913	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
914	#endif
915	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
916	*ppMem = &pMemLnx->Core;
917	IPRT_LINUX_RESTORE_EFL_AC();
918	return rc;
919	}
920
921	rtR0MemObjLinuxFreePages(pMemLnx);
922	rtR0MemObjDelete(&pMemLnx->Core);
923	}
924
925	IPRT_LINUX_RESTORE_EFL_AC();
926	return rc;
927	}
928
929
930	/**
931	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
932	*
933	* @returns IPRT status code.
934	* @param ppMemLnx Where to
935	* @param enmType The object type.
936	* @param cb The size of the allocation.
937	* @param uAlignment The alignment of the physical memory.
938	* Only valid for fContiguous == true, ignored otherwise.
939	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
940	* @param pszTag Allocation tag used for statistics and such.
941	* @param fGfp The Linux GFP flags to use for the allocation.
942	*/
943	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
944	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag, gfp_t fGfp)
945	{
946	PRTR0MEMOBJLNX pMemLnx;
947	int rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
948	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
949	false /fExecutable/, VERR_NO_PHYS_MEMORY, pszTag);
950	if (RT_FAILURE(rc))
951	return rc;
952
953	/*
954	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
955	*/
956	if (PhysHighest != NIL_RTHCPHYS)
957	{
958	size_t iPage = pMemLnx->cPages;
959	while (iPage-- > 0)
960	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
961	{
962	rtR0MemObjLinuxFreePages(pMemLnx);
963	rtR0MemObjDelete(&pMemLnx->Core);
964	return VERR_NO_MEMORY;
965	}
966	}
967
968	/*
969	* Complete the object.
970	*/
971	if (enmType == RTR0MEMOBJTYPE_PHYS)
972	{
973	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
974	pMemLnx->Core.u.Phys.fAllocated = true;
975	}
976	*ppMem = &pMemLnx->Core;
977	return rc;
978	}
979
980
981	/**
982	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
983	*
984	* @returns IPRT status code.
985	* @param ppMem Where to store the memory object pointer on success.
986	* @param enmType The object type.
987	* @param cb The size of the allocation.
988	* @param uAlignment The alignment of the physical memory.
989	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
990	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
991	* @param pszTag Allocation tag used for statistics and such.
992	*/
993	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
994	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag)
995	{
996	int rc;
997	IPRT_LINUX_SAVE_EFL_AC();
998
999	/*
1000	* There are two clear cases and that's the <=16MB and anything-goes ones.
1001	* When the physical address limit is somewhere in-between those two we'll
1002	* just have to try, starting with HIGHUSER and working our way thru the
1003	* different types, hoping we'll get lucky.
1004	*
1005	* We should probably move this physical address restriction logic up to
1006	* the page alloc function as it would be more efficient there. But since
1007	* we don't expect this to be a performance issue just yet it can wait.
1008	*/
1009	if (PhysHighest == NIL_RTHCPHYS)
1010	/* ZONE_HIGHMEM: the whole physical memory */
1011	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1012	else if (PhysHighest <= _1M * 16)
1013	/* ZONE_DMA: 0-16MB */
1014	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1015	else
1016	{
1017	rc = VERR_NO_MEMORY;
1018	if (RT_FAILURE(rc))
1019	/* ZONE_HIGHMEM: the whole physical memory */
1020	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1021	if (RT_FAILURE(rc))
1022	/* ZONE_NORMAL: 0-896MB */
1023	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_USER);
1024	#ifdef GFP_DMA32
1025	if (RT_FAILURE(rc))
1026	/* ZONE_DMA32: 0-4GB */
1027	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA32);
1028	#endif
1029	if (RT_FAILURE(rc))
1030	/* ZONE_DMA: 0-16MB */
1031	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1032	}
1033	IPRT_LINUX_RESTORE_EFL_AC();
1034	return rc;
1035	}
1036
1037
1038	/**
1039	* Translates a kernel virtual address to a linux page structure by walking the
1040	* page tables.
1041	*
1042	* @note We do assume that the page tables will not change as we are walking
1043	* them. This assumption is rather forced by the fact that I could not
1044	* immediately see any way of preventing this from happening. So, we
1045	* take some extra care when accessing them.
1046	*
1047	* Because of this, we don't want to use this function on memory where
1048	* attribute changes to nearby pages is likely to cause large pages to
1049	* be used or split up. So, don't use this for the linear mapping of
1050	* physical memory.
1051	*
1052	* @returns Pointer to the page structur or NULL if it could not be found.
1053	* @param pv The kernel virtual address.
1054	*/
1055	RTDECL(struct page ) rtR0MemObjLinuxVirtToPage(void pv)
1056	{
1057	unsigned long ulAddr = (unsigned long)pv;
1058	unsigned long pfn;
1059	struct page *pPage;
1060	pte_t *pEntry;
1061	union
1062	{
1063	pgd_t Global;
1064	#if RTLNX_VER_MIN(4,12,0)
1065	p4d_t Four;
1066	#endif
1067	#if RTLNX_VER_MIN(2,6,11)
1068	pud_t Upper;
1069	#endif
1070	pmd_t Middle;
1071	pte_t Entry;
1072	} u;
1073
1074	/* Should this happen in a situation this code will be called in? And if
1075	* so, can it change under our feet? See also
1076	* "Documentation/vm/active_mm.txt" in the kernel sources. */
1077	if (RT_UNLIKELY(!current->active_mm))
1078	return NULL;
1079	u.Global = *pgd_offset(current->active_mm, ulAddr);
1080	if (RT_UNLIKELY(pgd_none(u.Global)))
1081	return NULL;
1082	#if RTLNX_VER_MIN(2,6,11)
1083	# if RTLNX_VER_MIN(4,12,0)
1084	u.Four = *p4d_offset(&u.Global, ulAddr);
1085	if (RT_UNLIKELY(p4d_none(u.Four)))
1086	return NULL;
1087	if (p4d_large(u.Four))
1088	{
1089	pPage = p4d_page(u.Four);
1090	AssertReturn(pPage, NULL);
1091	pfn = page_to_pfn(pPage); /* doing the safe way... */
1092	AssertCompile(P4D_SHIFT - PAGE_SHIFT < 31);
1093	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (P4D_SHIFT - PAGE_SHIFT)) - 1);
1094	return pfn_to_page(pfn);
1095	}
1096	u.Upper = *pud_offset(&u.Four, ulAddr);
1097	# else /* < 4.12 */
1098	u.Upper = *pud_offset(&u.Global, ulAddr);
1099	# endif /* < 4.12 */
1100	if (RT_UNLIKELY(pud_none(u.Upper)))
1101	return NULL;
1102	# if RTLNX_VER_MIN(2,6,25)
1103	if (pud_large(u.Upper))
1104	{
1105	pPage = pud_page(u.Upper);
1106	AssertReturn(pPage, NULL);
1107	pfn = page_to_pfn(pPage); /* doing the safe way... */
1108	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
1109	return pfn_to_page(pfn);
1110	}
1111	# endif
1112	u.Middle = *pmd_offset(&u.Upper, ulAddr);
1113	#else /* < 2.6.11 */
1114	u.Middle = *pmd_offset(&u.Global, ulAddr);
1115	#endif /* < 2.6.11 */
1116	if (RT_UNLIKELY(pmd_none(u.Middle)))
1117	return NULL;
1118	#if RTLNX_VER_MIN(2,6,0)
1119	if (pmd_large(u.Middle))
1120	{
1121	pPage = pmd_page(u.Middle);
1122	AssertReturn(pPage, NULL);
1123	pfn = page_to_pfn(pPage); /* doing the safe way... */
1124	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
1125	return pfn_to_page(pfn);
1126	}
1127	#endif
1128
1129	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
1130	pEntry = pte_offset_map(&u.Middle, ulAddr);
1131	#else
1132	pEntry = pte_offset(&u.Middle, ulAddr);
1133	#endif
1134	if (RT_UNLIKELY(!pEntry))
1135	return NULL;
1136	u.Entry = *pEntry;
1137	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map)
1138	pte_unmap(pEntry);
1139	#endif
1140
1141	if (RT_UNLIKELY(!pte_present(u.Entry)))
1142	return NULL;
1143	return pte_page(u.Entry);
1144	}
1145	RT_EXPORT_SYMBOL(rtR0MemObjLinuxVirtToPage);
1146
1147
1148	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment,
1149	const char *pszTag)
1150	{
1151	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest, pszTag);
1152	}
1153
1154
1155	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, const char *pszTag)
1156	{
1157	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest, pszTag);
1158	}
1159
1160
1161	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy,
1162	const char *pszTag)
1163	{
1164	IPRT_LINUX_SAVE_EFL_AC();
1165
1166	/*
1167	* All we need to do here is to validate that we can use
1168	* ioremap on the specified address (32/64-bit dma_addr_t).
1169	*/
1170	PRTR0MEMOBJLNX pMemLnx;
1171	dma_addr_t PhysAddr = Phys;
1172	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
1173
1174	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb, pszTag);
1175	if (!pMemLnx)
1176	{
1177	IPRT_LINUX_RESTORE_EFL_AC();
1178	return VERR_NO_MEMORY;
1179	}
1180
1181	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
1182	pMemLnx->Core.u.Phys.fAllocated = false;
1183	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
1184	Assert(!pMemLnx->cPages);
1185	*ppMem = &pMemLnx->Core;
1186	IPRT_LINUX_RESTORE_EFL_AC();
1187	return VINF_SUCCESS;
1188	}
1189
1190	/* openSUSE Leap 42.3 detection :-/ */
1191	#if RTLNX_VER_RANGE(4,4,0, 4,6,0) && defined(FAULT_FLAG_REMOTE)
1192	# define GET_USER_PAGES_API KERNEL_VERSION(4, 10, 0) /* no typo! */
1193	#else
1194	# define GET_USER_PAGES_API LINUX_VERSION_CODE
1195	#endif
1196
1197	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess,
1198	RTR0PROCESS R0Process, const char *pszTag)
1199	{
1200	IPRT_LINUX_SAVE_EFL_AC();
1201	const int cPages = cb >> PAGE_SHIFT;
1202	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1203	struct vm_area_struct **papVMAs;
1204	PRTR0MEMOBJLNX pMemLnx;
1205	int rc = VERR_NO_MEMORY;
1206	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1207
1208	/*
1209	* Check for valid task and size overflows.
1210	*/
1211	if (!pTask)
1212	return VERR_NOT_SUPPORTED;
1213	if (((size_t)cPages << PAGE_SHIFT) != cb)
1214	return VERR_OUT_OF_RANGE;
1215
1216	/*
1217	* Allocate the memory object and a temporary buffer for the VMAs.
1218	*/
1219	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1220	(void *)R3Ptr, cb, pszTag);
1221	if (!pMemLnx)
1222	{
1223	IPRT_LINUX_RESTORE_EFL_AC();
1224	return VERR_NO_MEMORY;
1225	}
1226
1227	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1228	if (papVMAs)
1229	{
1230	LNX_MM_DOWN_READ(pTask->mm);
1231
1232	/*
1233	* Get user pages.
1234	*/
1235	/** @todo r=bird: Should we not force read access too? */
1236	#if GET_USER_PAGES_API >= KERNEL_VERSION(4, 6, 0)
1237	if (R0Process == RTR0ProcHandleSelf())
1238	rc = get_user_pages(R3Ptr, /* Where from. */
1239	cPages, /* How many pages. */
1240	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1241	fWrite ? FOLL_WRITE \| /* Write to memory. */
1242	FOLL_FORCE /* force write access. */
1243	: 0, /* Write to memory. */
1244	# else
1245	fWrite, /* Write to memory. */
1246	fWrite, /* force write access. */
1247	# endif
1248	&pMemLnx->apPages[0], /* Page array. */
1249	papVMAs); /* vmas */
1250	/*
1251	* Actually this should not happen at the moment as call this function
1252	* only for our own process.
1253	*/
1254	else
1255	rc = get_user_pages_remote(
1256	# if GET_USER_PAGES_API < KERNEL_VERSION(5, 9, 0)
1257	pTask, /* Task for fault accounting. */
1258	# endif
1259	pTask->mm, /* Whose pages. */
1260	R3Ptr, /* Where from. */
1261	cPages, /* How many pages. */
1262	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1263	fWrite ? FOLL_WRITE \| /* Write to memory. */
1264	FOLL_FORCE /* force write access. */
1265	: 0, /* Write to memory. */
1266	# else
1267	fWrite, /* Write to memory. */
1268	fWrite, /* force write access. */
1269	# endif
1270	&pMemLnx->apPages[0], /* Page array. */
1271	papVMAs /* vmas */
1272	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 10, 0)
1273	, NULL /* locked */
1274	# endif
1275	);
1276	#else /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1277	rc = get_user_pages(pTask, /* Task for fault accounting. */
1278	pTask->mm, /* Whose pages. */
1279	R3Ptr, /* Where from. */
1280	cPages, /* How many pages. */
1281	/* The get_user_pages API change was back-ported to 4.4.168. */
1282	# if RTLNX_VER_RANGE(4,4,168, 4,5,0)
1283	fWrite ? FOLL_WRITE \| /* Write to memory. */
1284	FOLL_FORCE /* force write access. */
1285	: 0, /* Write to memory. */
1286	# else
1287	fWrite, /* Write to memory. */
1288	fWrite, /* force write access. */
1289	# endif
1290	&pMemLnx->apPages[0], /* Page array. */
1291	papVMAs); /* vmas */
1292	#endif /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1293	if (rc == cPages)
1294	{
1295	/*
1296	* Flush dcache (required?), protect against fork and _really_ pin the page
1297	* table entries. get_user_pages() will protect against swapping out the
1298	* pages but it will NOT protect against removing page table entries. This
1299	* can be achieved with
1300	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1301	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1302	* Usual Linux distributions support only a limited size of locked pages
1303	* (e.g. 32KB).
1304	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1305	* or by
1306	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1307	* a range check.
1308	*/
1309	/** @todo The Linux fork() protection will require more work if this API
1310	* is to be used for anything but locking VM pages. */
1311	while (rc-- > 0)
1312	{
1313	flush_dcache_page(pMemLnx->apPages[rc]);
1314	papVMAs[rc]->vm_flags \|= VM_DONTCOPY \| VM_LOCKED;
1315	}
1316
1317	LNX_MM_UP_READ(pTask->mm);
1318
1319	RTMemFree(papVMAs);
1320
1321	pMemLnx->Core.u.Lock.R0Process = R0Process;
1322	pMemLnx->cPages = cPages;
1323	Assert(!pMemLnx->fMappedToRing0);
1324	*ppMem = &pMemLnx->Core;
1325
1326	IPRT_LINUX_RESTORE_EFL_AC();
1327	return VINF_SUCCESS;
1328	}
1329
1330	/*
1331	* Failed - we need to unlock any pages that we succeeded to lock.
1332	*/
1333	while (rc-- > 0)
1334	{
1335	if (!PageReserved(pMemLnx->apPages[rc]))
1336	SetPageDirty(pMemLnx->apPages[rc]);
1337	#if RTLNX_VER_MIN(4,6,0)
1338	put_page(pMemLnx->apPages[rc]);
1339	#else
1340	page_cache_release(pMemLnx->apPages[rc]);
1341	#endif
1342	}
1343
1344	LNX_MM_UP_READ(pTask->mm);
1345
1346	RTMemFree(papVMAs);
1347	rc = VERR_LOCK_FAILED;
1348	}
1349
1350	rtR0MemObjDelete(&pMemLnx->Core);
1351	IPRT_LINUX_RESTORE_EFL_AC();
1352	return rc;
1353	}
1354
1355
1356	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void pv, size_t cb, uint32_t fAccess, const char pszTag)
1357	{
1358	IPRT_LINUX_SAVE_EFL_AC();
1359	void pvLast = (uint8_t )pv + cb - 1;
1360	size_t const cPages = cb >> PAGE_SHIFT;
1361	PRTR0MEMOBJLNX pMemLnx;
1362	bool fLinearMapping;
1363	int rc;
1364	uint8_t *pbPage;
1365	size_t iPage;
1366	NOREF(fAccess);
1367
1368	if ( !RTR0MemKernelIsValidAddr(pv)
1369	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1370	return VERR_INVALID_PARAMETER;
1371
1372	/*
1373	* The lower part of the kernel memory has a linear mapping between
1374	* physical and virtual addresses. So we take a short cut here. This is
1375	* assumed to be the cleanest way to handle those addresses (and the code
1376	* is well tested, though the test for determining it is not very nice).
1377	* If we ever decide it isn't we can still remove it.
1378	*/
1379	#if 0
1380	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1381	#else
1382	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1383	&& (unsigned long)pvLast < (unsigned long)high_memory;
1384	#endif
1385
1386	/*
1387	* Allocate the memory object.
1388	*/
1389	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1390	pv, cb, pszTag);
1391	if (!pMemLnx)
1392	{
1393	IPRT_LINUX_RESTORE_EFL_AC();
1394	return VERR_NO_MEMORY;
1395	}
1396
1397	/*
1398	* Gather the pages.
1399	* We ASSUME all kernel pages are non-swappable and non-movable.
1400	*/
1401	rc = VINF_SUCCESS;
1402	pbPage = (uint8_t *)pvLast;
1403	iPage = cPages;
1404	if (!fLinearMapping)
1405	{
1406	while (iPage-- > 0)
1407	{
1408	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1409	if (RT_UNLIKELY(!pPage))
1410	{
1411	rc = VERR_LOCK_FAILED;
1412	break;
1413	}
1414	pMemLnx->apPages[iPage] = pPage;
1415	pbPage -= PAGE_SIZE;
1416	}
1417	}
1418	else
1419	{
1420	while (iPage-- > 0)
1421	{
1422	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1423	pbPage -= PAGE_SIZE;
1424	}
1425	}
1426	if (RT_SUCCESS(rc))
1427	{
1428	/*
1429	* Complete the memory object and return.
1430	*/
1431	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1432	pMemLnx->cPages = cPages;
1433	Assert(!pMemLnx->fMappedToRing0);
1434	*ppMem = &pMemLnx->Core;
1435
1436	IPRT_LINUX_RESTORE_EFL_AC();
1437	return VINF_SUCCESS;
1438	}
1439
1440	rtR0MemObjDelete(&pMemLnx->Core);
1441	IPRT_LINUX_RESTORE_EFL_AC();
1442	return rc;
1443	}
1444
1445
1446	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment,
1447	const char *pszTag)
1448	{
1449	#if RTLNX_VER_MIN(2,4,22)
1450	IPRT_LINUX_SAVE_EFL_AC();
1451	const size_t cPages = cb >> PAGE_SHIFT;
1452	struct page *pDummyPage;
1453	struct page **papPages;
1454
1455	/* check for unsupported stuff. */
1456	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1457	if (uAlignment > PAGE_SIZE)
1458	return VERR_NOT_SUPPORTED;
1459
1460	/*
1461	* Allocate a dummy page and create a page pointer array for vmap such that
1462	* the dummy page is mapped all over the reserved area.
1463	*/
1464	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1465	if (pDummyPage)
1466	{
1467	papPages = RTMemAlloc(sizeof(papPages) cPages);
1468	if (papPages)
1469	{
1470	void *pv;
1471	size_t iPage = cPages;
1472	while (iPage-- > 0)
1473	papPages[iPage] = pDummyPage;
1474	# ifdef VM_MAP
1475	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1476	# else
1477	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1478	# endif
1479	RTMemFree(papPages);
1480	if (pv)
1481	{
1482	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1483	if (pMemLnx)
1484	{
1485	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1486	pMemLnx->cPages = 1;
1487	pMemLnx->apPages[0] = pDummyPage;
1488	*ppMem = &pMemLnx->Core;
1489	IPRT_LINUX_RESTORE_EFL_AC();
1490	return VINF_SUCCESS;
1491	}
1492	vunmap(pv);
1493	}
1494	}
1495	__free_page(pDummyPage);
1496	}
1497	IPRT_LINUX_RESTORE_EFL_AC();
1498	return VERR_NO_MEMORY;
1499
1500	#else /* < 2.4.22 */
1501	/*
1502	* Could probably use ioremap here, but the caller is in a better position than us
1503	* to select some safe physical memory.
1504	*/
1505	return VERR_NOT_SUPPORTED;
1506	#endif
1507	}
1508
1509
1510	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment,
1511	RTR0PROCESS R0Process, const char *pszTag)
1512	{
1513	IPRT_LINUX_SAVE_EFL_AC();
1514	PRTR0MEMOBJLNX pMemLnx;
1515	void *pv;
1516	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1517	if (!pTask)
1518	return VERR_NOT_SUPPORTED;
1519
1520	/*
1521	* Check that the specified alignment is supported.
1522	*/
1523	if (uAlignment > PAGE_SIZE)
1524	return VERR_NOT_SUPPORTED;
1525
1526	/*
1527	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1528	*/
1529	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1530	if (pv == (void *)-1)
1531	{
1532	IPRT_LINUX_RESTORE_EFL_AC();
1533	return VERR_NO_MEMORY;
1534	}
1535
1536	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1537	if (!pMemLnx)
1538	{
1539	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1540	IPRT_LINUX_RESTORE_EFL_AC();
1541	return VERR_NO_MEMORY;
1542	}
1543
1544	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1545	*ppMem = &pMemLnx->Core;
1546	IPRT_LINUX_RESTORE_EFL_AC();
1547	return VINF_SUCCESS;
1548	}
1549
1550
1551	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment,
1552	unsigned fProt, size_t offSub, size_t cbSub, const char *pszTag)
1553	{
1554	int rc = VERR_NO_MEMORY;
1555	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1556	PRTR0MEMOBJLNX pMemLnx;
1557	IPRT_LINUX_SAVE_EFL_AC();
1558
1559	/* Fail if requested to do something we can't. */
1560	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1561	if (uAlignment > PAGE_SIZE)
1562	return VERR_NOT_SUPPORTED;
1563
1564	/*
1565	* Create the IPRT memory object.
1566	*/
1567	if (!cbSub)
1568	cbSub = pMemLnxToMap->Core.cb - offSub;
1569	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1570	if (pMemLnx)
1571	{
1572	if (pMemLnxToMap->cPages)
1573	{
1574	#if RTLNX_VER_MIN(2,4,22)
1575	/*
1576	* Use vmap - 2.4.22 and later.
1577	*/
1578	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1579	/** @todo We don't really care too much for EXEC here... 5.8 always adds NX. */
1580	Assert(((offSub + cbSub) >> PAGE_SHIFT) <= pMemLnxToMap->cPages);
1581	# ifdef VM_MAP
1582	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_MAP, fPg);
1583	# else
1584	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_ALLOC, fPg);
1585	# endif
1586	if (pMemLnx->Core.pv)
1587	{
1588	pMemLnx->fMappedToRing0 = true;
1589	rc = VINF_SUCCESS;
1590	}
1591	else
1592	rc = VERR_MAP_FAILED;
1593
1594	#else /* < 2.4.22 */
1595	/*
1596	* Only option here is to share mappings if possible and forget about fProt.
1597	*/
1598	if (rtR0MemObjIsRing3(pMemToMap))
1599	rc = VERR_NOT_SUPPORTED;
1600	else
1601	{
1602	rc = VINF_SUCCESS;
1603	if (!pMemLnxToMap->Core.pv)
1604	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1605	if (RT_SUCCESS(rc))
1606	{
1607	Assert(pMemLnxToMap->Core.pv);
1608	pMemLnx->Core.pv = (uint8_t *)pMemLnxToMap->Core.pv + offSub;
1609	}
1610	}
1611	#endif
1612	}
1613	else
1614	{
1615	/*
1616	* MMIO / physical memory.
1617	*/
1618	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1619	#if RTLNX_VER_MIN(2,6,25)
1620	/*
1621	* ioremap() defaults to no caching since the 2.6 kernels.
1622	* ioremap_nocache() has been removed finally in 5.6-rc1.
1623	*/
1624	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1625	? ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1626	: ioremap_cache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1627	#else /* KERNEL_VERSION < 2.6.25 */
1628	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1629	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1630	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1631	#endif /* KERNEL_VERSION < 2.6.25 */
1632	if (pMemLnx->Core.pv)
1633	{
1634	/** @todo fix protection. */
1635	rc = VINF_SUCCESS;
1636	}
1637	}
1638	if (RT_SUCCESS(rc))
1639	{
1640	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1641	*ppMem = &pMemLnx->Core;
1642	IPRT_LINUX_RESTORE_EFL_AC();
1643	return VINF_SUCCESS;
1644	}
1645	rtR0MemObjDelete(&pMemLnx->Core);
1646	}
1647
1648	IPRT_LINUX_RESTORE_EFL_AC();
1649	return rc;
1650	}
1651
1652
1653	#ifdef VBOX_USE_PAE_HACK
1654	/**
1655	* Replace the PFN of a PTE with the address of the actual page.
1656	*
1657	* The caller maps a reserved dummy page at the address with the desired access
1658	* and flags.
1659	*
1660	* This hack is required for older Linux kernels which don't provide
1661	* remap_pfn_range().
1662	*
1663	* @returns 0 on success, -ENOMEM on failure.
1664	* @param mm The memory context.
1665	* @param ulAddr The mapping address.
1666	* @param Phys The physical address of the page to map.
1667	*/
1668	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1669	{
1670	int rc = -ENOMEM;
1671	pgd_t *pgd;
1672
1673	spin_lock(&mm->page_table_lock);
1674
1675	pgd = pgd_offset(mm, ulAddr);
1676	if (!pgd_none(pgd) && !pgd_bad(pgd))
1677	{
1678	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1679	if (!pmd_none(*pmd))
1680	{
1681	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1682	if (ptep)
1683	{
1684	pte_t pte = *ptep;
1685	pte.pte_high &= 0xfff00000;
1686	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1687	pte.pte_low &= 0x00000fff;
1688	pte.pte_low \|= (Phys & 0xfffff000);
1689	set_pte(ptep, pte);
1690	pte_unmap(ptep);
1691	rc = 0;
1692	}
1693	}
1694	}
1695
1696	spin_unlock(&mm->page_table_lock);
1697	return rc;
1698	}
1699	#endif /* VBOX_USE_PAE_HACK */
1700
1701
1702	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment,
1703	unsigned fProt, RTR0PROCESS R0Process, size_t offSub, size_t cbSub, const char *pszTag)
1704	{
1705	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1706	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1707	int rc = VERR_NO_MEMORY;
1708	PRTR0MEMOBJLNX pMemLnx;
1709	#ifdef VBOX_USE_PAE_HACK
1710	struct page *pDummyPage;
1711	RTHCPHYS DummyPhys;
1712	#endif
1713	IPRT_LINUX_SAVE_EFL_AC();
1714
1715	/*
1716	* Check for restrictions.
1717	*/
1718	if (!pTask)
1719	return VERR_NOT_SUPPORTED;
1720	if (uAlignment > PAGE_SIZE)
1721	return VERR_NOT_SUPPORTED;
1722
1723	#ifdef VBOX_USE_PAE_HACK
1724	/*
1725	* Allocate a dummy page for use when mapping the memory.
1726	*/
1727	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1728	if (!pDummyPage)
1729	{
1730	IPRT_LINUX_RESTORE_EFL_AC();
1731	return VERR_NO_MEMORY;
1732	}
1733	SetPageReserved(pDummyPage);
1734	DummyPhys = page_to_phys(pDummyPage);
1735	#endif
1736
1737	/*
1738	* Create the IPRT memory object.
1739	*/
1740	Assert(!offSub \|\| cbSub);
1741	if (cbSub == 0)
1742	cbSub = pMemLnxToMap->Core.cb;
1743	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1744	if (pMemLnx)
1745	{
1746	/*
1747	* Allocate user space mapping.
1748	*/
1749	void *pv;
1750	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cbSub, uAlignment, pTask, fProt);
1751	if (pv != (void *)-1)
1752	{
1753	/*
1754	* Map page by page into the mmap area.
1755	* This is generic, paranoid and not very efficient.
1756	*/
1757	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1758	unsigned long ulAddrCur = (unsigned long)pv;
1759	const size_t cPages = (offSub + cbSub) >> PAGE_SHIFT;
1760	size_t iPage;
1761
1762	LNX_MM_DOWN_WRITE(pTask->mm);
1763
1764	rc = VINF_SUCCESS;
1765	if (pMemLnxToMap->cPages)
1766	{
1767	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1768	{
1769	#if RTLNX_VER_MAX(2,6,11)
1770	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1771	#endif
1772	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1773	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1774	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1775	#endif
1776	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1777	/* remap_page_range() limitation on x86 */
1778	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1779	#endif
1780
1781	#if defined(VBOX_USE_INSERT_PAGE) && RTLNX_VER_MIN(2,6,22)
1782	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1783	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1784	* See remap_pfn_range() in mm/memory.c */
1785	#if RTLNX_VER_MIN(3,7,0)
1786	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1787	#else
1788	vma->vm_flags \|= VM_RESERVED;
1789	#endif
1790	#elif RTLNX_VER_MIN(2,6,11)
1791	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1792	#elif defined(VBOX_USE_PAE_HACK)
1793	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1794	if (!rc)
1795	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1796	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1797	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1798	#else /* 2.4 */
1799	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1800	#endif
1801	if (rc)
1802	{
1803	rc = VERR_NO_MEMORY;
1804	break;
1805	}
1806	}
1807	}
1808	else
1809	{
1810	RTHCPHYS Phys;
1811	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1812	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1813	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1814	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1815	else
1816	{
1817	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1818	Phys = NIL_RTHCPHYS;
1819	}
1820	if (Phys != NIL_RTHCPHYS)
1821	{
1822	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1823	{
1824	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1825	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1826	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1827	#endif
1828	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1829	/* remap_page_range() limitation on x86 */
1830	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1831	#endif
1832
1833	#if RTLNX_VER_MIN(2,6,11)
1834	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1835	#elif defined(VBOX_USE_PAE_HACK)
1836	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1837	if (!rc)
1838	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1839	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1840	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1841	#else /* 2.4 */
1842	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1843	#endif
1844	if (rc)
1845	{
1846	rc = VERR_NO_MEMORY;
1847	break;
1848	}
1849	}
1850	}
1851	}
1852
1853	#ifdef CONFIG_NUMA_BALANCING
1854	# if RTLNX_VER_MAX(3,13,0) && RTLNX_RHEL_MAX(7,0)
1855	# define VBOX_NUMA_HACK_OLD
1856	# endif
1857	if (RT_SUCCESS(rc))
1858	{
1859	/** @todo Ugly hack! But right now we have no other means to
1860	* disable automatic NUMA page balancing. */
1861	# ifdef RT_OS_X86
1862	# ifdef VBOX_NUMA_HACK_OLD
1863	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1864	# endif
1865	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1866	# else
1867	# ifdef VBOX_NUMA_HACK_OLD
1868	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1869	# endif
1870	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1871	# endif
1872	}
1873	#endif /* CONFIG_NUMA_BALANCING */
1874
1875	LNX_MM_UP_WRITE(pTask->mm);
1876
1877	if (RT_SUCCESS(rc))
1878	{
1879	#ifdef VBOX_USE_PAE_HACK
1880	__free_page(pDummyPage);
1881	#endif
1882	pMemLnx->Core.pv = pv;
1883	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1884	*ppMem = &pMemLnx->Core;
1885	IPRT_LINUX_RESTORE_EFL_AC();
1886	return VINF_SUCCESS;
1887	}
1888
1889	/*
1890	* Bail out.
1891	*/
1892	rtR0MemObjLinuxDoMunmap(pv, cbSub, pTask);
1893	}
1894	rtR0MemObjDelete(&pMemLnx->Core);
1895	}
1896	#ifdef VBOX_USE_PAE_HACK
1897	__free_page(pDummyPage);
1898	#endif
1899
1900	IPRT_LINUX_RESTORE_EFL_AC();
1901	return rc;
1902	}
1903
1904
1905	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1906	{
1907	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
1908	/*
1909	* Currently only supported when we've got addresses PTEs from the kernel.
1910	*/
1911	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1912	if (pMemLnx->pArea && pMemLnx->papPtesForArea)
1913	{
1914	pgprot_t const fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1915	size_t const cPages = (offSub + cbSub) >> PAGE_SHIFT;
1916	pte_t **papPtes = pMemLnx->papPtesForArea;
1917	size_t i;
1918
1919	for (i = offSub >> PAGE_SHIFT; i < cPages; i++)
1920	{
1921	set_pte(papPtes[i], mk_pte(pMemLnx->apPages[i], fPg));
1922	}
1923	preempt_disable();
1924	__flush_tlb_all();
1925	preempt_enable();
1926	return VINF_SUCCESS;
1927	}
1928	# elif defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
1929	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1930	if ( pMemLnx->fExecutable
1931	&& pMemLnx->fMappedToRing0)
1932	{
1933	LNXAPPLYPGRANGE Args;
1934	Args.pMemLnx = pMemLnx;
1935	Args.fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1936	int rcLnx = apply_to_page_range(current->active_mm, (unsigned long)pMemLnx->Core.pv + offSub, cbSub,
1937	rtR0MemObjLinuxApplyPageRange, (void *)&Args);
1938	if (rcLnx)
1939	return VERR_NOT_SUPPORTED;
1940
1941	return VINF_SUCCESS;
1942	}
1943	# endif
1944
1945	NOREF(pMem);
1946	NOREF(offSub);
1947	NOREF(cbSub);
1948	NOREF(fProt);
1949	return VERR_NOT_SUPPORTED;
1950	}
1951
1952
1953	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1954	{
1955	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1956
1957	if (pMemLnx->cPages)
1958	return page_to_phys(pMemLnx->apPages[iPage]);
1959
1960	switch (pMemLnx->Core.enmType)
1961	{
1962	case RTR0MEMOBJTYPE_CONT:
1963	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1964
1965	case RTR0MEMOBJTYPE_PHYS:
1966	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1967
1968	/* the parent knows */
1969	case RTR0MEMOBJTYPE_MAPPING:
1970	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1971
1972	/* cPages > 0 */
1973	case RTR0MEMOBJTYPE_LOW:
1974	case RTR0MEMOBJTYPE_LOCK:
1975	case RTR0MEMOBJTYPE_PHYS_NC:
1976	case RTR0MEMOBJTYPE_PAGE:
1977	default:
1978	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1979	RT_FALL_THROUGH();
1980
1981	case RTR0MEMOBJTYPE_RES_VIRT:
1982	return NIL_RTHCPHYS;
1983	}
1984	}
1985

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 91482

Download in other formats: