memobj-r0drv-linux.c@ 92418

Last change on this file since 92418 was 92418, checked in by vboxsync, 3 years ago
IPRT/memobj-r0drv-linux: Implemented rtR0MemObjNativeAllocLarge using GFP_TRANSHUGE and made it fail with VERR_NOT_SUPPORTED if GFP_TRANSHUGE is not available (< 2.6.38). bugref:10093 bugref:5324
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev Revision`
File size: 70.4 KB

Line
1	/* $Id: memobj-r0drv-linux.c 92418 2021-11-15 08:38:55Z 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	/** This is non-zero if large page allocation. */
126	uint8_t cLargePageOrder;
127	#ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
128	/** Return from alloc_vm_area() that we now need to use for executable
129	* memory. */
130	struct vm_struct *pArea;
131	/** PTE array that goes along with pArea (must be freed). */
132	pte_t **papPtesForArea;
133	#endif
134	/** The pages in the apPages array. */
135	size_t cPages;
136	/** Array of struct page pointers. (variable size) */
137	struct page *apPages[1];
138	} RTR0MEMOBJLNX;
139	/** Pointer to the linux memory object. */
140	typedef RTR0MEMOBJLNX *PRTR0MEMOBJLNX;
141
142
143	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx);
144
145
146	/**
147	* Helper that converts from a RTR0PROCESS handle to a linux task.
148	*
149	* @returns The corresponding Linux task.
150	* @param R0Process IPRT ring-0 process handle.
151	*/
152	static struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
153	{
154	/** @todo fix rtR0ProcessToLinuxTask!! */
155	/** @todo many (all?) callers currently assume that we return 'current'! */
156	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
157	}
158
159
160	/**
161	* Compute order. Some functions allocate 2^order pages.
162	*
163	* @returns order.
164	* @param cPages Number of pages.
165	*/
166	static int rtR0MemObjLinuxOrder(size_t cPages)
167	{
168	int iOrder;
169	size_t cTmp;
170
171	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
172	;
173	if (cPages & ~((size_t)1 << iOrder))
174	++iOrder;
175
176	return iOrder;
177	}
178
179
180	/**
181	* Converts from RTMEM_PROT_* to Linux PAGE_*.
182	*
183	* @returns Linux page protection constant.
184	* @param fProt The IPRT protection mask.
185	* @param fKernel Whether it applies to kernel or user space.
186	*/
187	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
188	{
189	switch (fProt)
190	{
191	default:
192	AssertMsgFailed(("%#x %d\n", fProt, fKernel)); RT_FALL_THRU();
193	case RTMEM_PROT_NONE:
194	return PAGE_NONE;
195
196	case RTMEM_PROT_READ:
197	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
198
199	case RTMEM_PROT_WRITE:
200	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
201	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
202
203	case RTMEM_PROT_EXEC:
204	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
205	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
206	if (fKernel)
207	{
208	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
209	pgprot_val(fPg) &= ~_PAGE_RW;
210	return fPg;
211	}
212	return PAGE_READONLY_EXEC;
213	#else
214	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
215	#endif
216
217	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
218	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
219	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
220	}
221	}
222
223
224	/**
225	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
226	* an empty user space mapping.
227	*
228	* We acquire the mmap_sem/mmap_lock of the task!
229	*
230	* @returns Pointer to the mapping.
231	* (void *)-1 on failure.
232	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
233	* @param cb The size of the mapping.
234	* @param uAlignment The alignment of the mapping.
235	* @param pTask The Linux task to create this mapping in.
236	* @param fProt The RTMEM_PROT_* mask.
237	*/
238	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
239	{
240	unsigned fLnxProt;
241	unsigned long ulAddr;
242
243	Assert(pTask == current); /* do_mmap */
244	RT_NOREF_PV(pTask);
245
246	/*
247	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
248	*/
249	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
250	if (fProt == RTMEM_PROT_NONE)
251	fLnxProt = PROT_NONE;
252	else
253	{
254	fLnxProt = 0;
255	if (fProt & RTMEM_PROT_READ)
256	fLnxProt \|= PROT_READ;
257	if (fProt & RTMEM_PROT_WRITE)
258	fLnxProt \|= PROT_WRITE;
259	if (fProt & RTMEM_PROT_EXEC)
260	fLnxProt \|= PROT_EXEC;
261	}
262
263	if (R3PtrFixed != (RTR3PTR)-1)
264	{
265	#if RTLNX_VER_MIN(3,5,0)
266	ulAddr = vm_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
267	#else
268	LNX_MM_DOWN_WRITE(pTask->mm);
269	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
270	LNX_MM_UP_WRITE(pTask->mm);
271	#endif
272	}
273	else
274	{
275	#if RTLNX_VER_MIN(3,5,0)
276	ulAddr = vm_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
277	#else
278	LNX_MM_DOWN_WRITE(pTask->mm);
279	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
280	LNX_MM_UP_WRITE(pTask->mm);
281	#endif
282	if ( !(ulAddr & ~PAGE_MASK)
283	&& (ulAddr & (uAlignment - 1)))
284	{
285	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
286	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
287	* ourselves) and further by there begin two mmap strategies (top / bottom). */
288	/* For now, just ignore uAlignment requirements... */
289	}
290	}
291
292
293	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
294	return (void *)-1;
295	return (void *)ulAddr;
296	}
297
298
299	/**
300	* Worker that destroys a user space mapping.
301	* Undoes what rtR0MemObjLinuxDoMmap did.
302	*
303	* We acquire the mmap_sem/mmap_lock of the task!
304	*
305	* @param pv The ring-3 mapping.
306	* @param cb The size of the mapping.
307	* @param pTask The Linux task to destroy this mapping in.
308	*/
309	static void rtR0MemObjLinuxDoMunmap(void pv, size_t cb, struct task_struct pTask)
310	{
311	#if RTLNX_VER_MIN(3,5,0)
312	Assert(pTask == current); RT_NOREF_PV(pTask);
313	vm_munmap((unsigned long)pv, cb);
314	#elif defined(USE_RHEL4_MUNMAP)
315	LNX_MM_DOWN_WRITE(pTask->mm);
316	do_munmap(pTask->mm, (unsigned long)pv, cb, 0); /* should it be 1 or 0? */
317	LNX_MM_UP_WRITE(pTask->mm);
318	#else
319	LNX_MM_DOWN_WRITE(pTask->mm);
320	do_munmap(pTask->mm, (unsigned long)pv, cb);
321	LNX_MM_UP_WRITE(pTask->mm);
322	#endif
323	}
324
325
326	/**
327	* Internal worker that allocates physical pages and creates the memory object for them.
328	*
329	* @returns IPRT status code.
330	* @param ppMemLnx Where to store the memory object pointer.
331	* @param enmType The object type.
332	* @param cb The number of bytes to allocate.
333	* @param uAlignment The alignment of the physical memory.
334	* Only valid if fContiguous == true, ignored otherwise.
335	* @param fFlagsLnx The page allocation flags (GPFs).
336	* @param fContiguous Whether the allocation must be contiguous.
337	* @param fExecutable Whether the memory must be executable.
338	* @param rcNoMem What to return when we're out of pages.
339	* @param pszTag Allocation tag used for statistics and such.
340	*/
341	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
342	size_t uAlignment, gfp_t fFlagsLnx, bool fContiguous, bool fExecutable, int rcNoMem,
343	const char *pszTag)
344	{
345	size_t iPage;
346	size_t const cPages = cb >> PAGE_SHIFT;
347	struct page *paPages;
348
349	/*
350	* Allocate a memory object structure that's large enough to contain
351	* the page pointer array.
352	*/
353	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), enmType,
354	NULL, cb, pszTag);
355	if (!pMemLnx)
356	return VERR_NO_MEMORY;
357	pMemLnx->Core.fFlags \|= RTR0MEMOBJ_FLAGS_UNINITIALIZED_AT_ALLOC;
358	pMemLnx->cPages = cPages;
359
360	if (cPages > 255)
361	{
362	# ifdef __GFP_REPEAT
363	/* Try hard to allocate the memory, but the allocation attempt might fail. */
364	fFlagsLnx \|= __GFP_REPEAT;
365	# endif
366	# ifdef __GFP_NOMEMALLOC
367	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
368	fFlagsLnx \|= __GFP_NOMEMALLOC;
369	# endif
370	}
371
372	/*
373	* Allocate the pages.
374	* For small allocations we'll try contiguous first and then fall back on page by page.
375	*/
376	#if RTLNX_VER_MIN(2,4,22)
377	if ( fContiguous
378	\|\| cb <= PAGE_SIZE * 2)
379	{
380	# ifdef VBOX_USE_INSERT_PAGE
381	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
382	# else
383	paPages = alloc_pages(fFlagsLnx \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
384	# endif
385	if (paPages)
386	{
387	fContiguous = true;
388	for (iPage = 0; iPage < cPages; iPage++)
389	pMemLnx->apPages[iPage] = &paPages[iPage];
390	}
391	else if (fContiguous)
392	{
393	rtR0MemObjDelete(&pMemLnx->Core);
394	return rcNoMem;
395	}
396	}
397
398	if (!fContiguous)
399	{
400	/** @todo Try use alloc_pages_bulk_array when available, it should be faster
401	* than a alloc_page loop. Put it in #ifdefs similar to
402	* IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC. */
403	for (iPage = 0; iPage < cPages; iPage++)
404	{
405	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
406	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
407	{
408	while (iPage-- > 0)
409	__free_page(pMemLnx->apPages[iPage]);
410	rtR0MemObjDelete(&pMemLnx->Core);
411	return rcNoMem;
412	}
413	}
414	}
415
416	#else /* < 2.4.22 */
417	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
418	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
419	if (!paPages)
420	{
421	rtR0MemObjDelete(&pMemLnx->Core);
422	return rcNoMem;
423	}
424	for (iPage = 0; iPage < cPages; iPage++)
425	{
426	pMemLnx->apPages[iPage] = &paPages[iPage];
427	if (fExecutable)
428	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
429	if (PageHighMem(pMemLnx->apPages[iPage]))
430	BUG();
431	}
432
433	fContiguous = true;
434	#endif /* < 2.4.22 */
435	pMemLnx->fContiguous = fContiguous;
436	pMemLnx->fExecutable = fExecutable;
437
438	#if RTLNX_VER_MAX(4,5,0)
439	/*
440	* Reserve the pages.
441	*
442	* Linux >= 4.5 with CONFIG_DEBUG_VM panics when setting PG_reserved on compound
443	* pages. According to Michal Hocko this shouldn't be necessary anyway because
444	* as pages which are not on the LRU list are never evictable.
445	*/
446	for (iPage = 0; iPage < cPages; iPage++)
447	SetPageReserved(pMemLnx->apPages[iPage]);
448	#endif
449
450	/*
451	* Note that the physical address of memory allocated with alloc_pages(flags, order)
452	* is always 2^(PAGE_SHIFT+order)-aligned.
453	*/
454	if ( fContiguous
455	&& uAlignment > PAGE_SIZE)
456	{
457	/*
458	* Check for alignment constraints. The physical address of memory allocated with
459	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
460	*/
461	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
462	{
463	/*
464	* This should never happen!
465	*/
466	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
467	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
468	rtR0MemObjLinuxFreePages(pMemLnx);
469	return rcNoMem;
470	}
471	}
472
473	*ppMemLnx = pMemLnx;
474	return VINF_SUCCESS;
475	}
476
477
478	/**
479	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
480	*
481	* This method does NOT free the object.
482	*
483	* @param pMemLnx The object which physical pages should be freed.
484	*/
485	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
486	{
487	size_t iPage = pMemLnx->cPages;
488	if (iPage > 0)
489	{
490	/*
491	* Restore the page flags.
492	*/
493	while (iPage-- > 0)
494	{
495	#if RTLNX_VER_MAX(4,5,0)
496	/* See SetPageReserved() in rtR0MemObjLinuxAllocPages() */
497	ClearPageReserved(pMemLnx->apPages[iPage]);
498	#endif
499	#if RTLNX_VER_MAX(2,4,22)
500	if (pMemLnx->fExecutable)
501	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
502	#endif
503	}
504
505	/*
506	* Free the pages.
507	*/
508	#if RTLNX_VER_MIN(2,4,22)
509	if (!pMemLnx->fContiguous)
510	{
511	iPage = pMemLnx->cPages;
512	while (iPage-- > 0)
513	__free_page(pMemLnx->apPages[iPage]);
514	}
515	else
516	#endif
517	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
518
519	pMemLnx->cPages = 0;
520	}
521	}
522
523
524	#ifdef IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
525	/**
526	* User data passed to the apply_to_page_range() callback.
527	*/
528	typedef struct LNXAPPLYPGRANGE
529	{
530	/** Pointer to the memory object. */
531	PRTR0MEMOBJLNX pMemLnx;
532	/** The page protection flags to apply. */
533	pgprot_t fPg;
534	} LNXAPPLYPGRANGE;
535	/** Pointer to the user data. */
536	typedef LNXAPPLYPGRANGE *PLNXAPPLYPGRANGE;
537	/** Pointer to the const user data. */
538	typedef const LNXAPPLYPGRANGE *PCLNXAPPLYPGRANGE;
539
540	/**
541	* Callback called in apply_to_page_range().
542	*
543	* @returns Linux status code.
544	* @param pPte Pointer to the page table entry for the given address.
545	* @param uAddr The address to apply the new protection to.
546	* @param pvUser The opaque user data.
547	*/
548	static int rtR0MemObjLinuxApplyPageRange(pte_t pPte, unsigned long uAddr, void pvUser)
549	{
550	PCLNXAPPLYPGRANGE pArgs = (PCLNXAPPLYPGRANGE)pvUser;
551	PRTR0MEMOBJLNX pMemLnx = pArgs->pMemLnx;
552	size_t idxPg = (uAddr - (unsigned long)pMemLnx->Core.pv) >> PAGE_SHIFT;
553
554	set_pte(pPte, mk_pte(pMemLnx->apPages[idxPg], pArgs->fPg));
555	return 0;
556	}
557	#endif
558
559
560	/**
561	* Maps the allocation into ring-0.
562	*
563	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
564	*
565	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
566	* space, so we'll use that mapping if possible. If execute access is required, we'll
567	* play safe and do our own mapping.
568	*
569	* @returns IPRT status code.
570	* @param pMemLnx The linux memory object to map.
571	* @param fExecutable Whether execute access is required.
572	*/
573	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
574	{
575	int rc = VINF_SUCCESS;
576
577	/*
578	* Choose mapping strategy.
579	*/
580	bool fMustMap = fExecutable
581	\|\| !pMemLnx->fContiguous;
582	if (!fMustMap)
583	{
584	size_t iPage = pMemLnx->cPages;
585	while (iPage-- > 0)
586	if (PageHighMem(pMemLnx->apPages[iPage]))
587	{
588	fMustMap = true;
589	break;
590	}
591	}
592
593	Assert(!pMemLnx->Core.pv);
594	Assert(!pMemLnx->fMappedToRing0);
595
596	if (fMustMap)
597	{
598	/*
599	* Use vmap - 2.4.22 and later.
600	*/
601	#if RTLNX_VER_MIN(2,4,22)
602	pgprot_t fPg;
603	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
604	# ifdef _PAGE_NX
605	if (!fExecutable)
606	pgprot_val(fPg) \|= _PAGE_NX;
607	# endif
608
609	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
610	if (fExecutable)
611	{
612	# if RTLNX_VER_MIN(3,2,51)
613	pte_t papPtes = (pte_t )kmalloc_array(pMemLnx->cPages, sizeof(papPtes[0]), GFP_KERNEL);
614	# else
615	pte_t papPtes = (pte_t )kmalloc(pMemLnx->cPages * sizeof(papPtes[0]), GFP_KERNEL);
616	# endif
617	if (papPtes)
618	{
619	pMemLnx->pArea = alloc_vm_area(pMemLnx->Core.cb, papPtes); /* Note! pArea->nr_pages is not set. */
620	if (pMemLnx->pArea)
621	{
622	size_t i;
623	Assert(pMemLnx->pArea->size >= pMemLnx->Core.cb); /* Note! includes guard page. */
624	Assert(pMemLnx->pArea->addr);
625	# ifdef _PAGE_NX
626	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
627	# endif
628	pMemLnx->papPtesForArea = papPtes;
629	for (i = 0; i < pMemLnx->cPages; i++)
630	*papPtes[i] = mk_pte(pMemLnx->apPages[i], fPg);
631	pMemLnx->Core.pv = pMemLnx->pArea->addr;
632	pMemLnx->fMappedToRing0 = true;
633	}
634	else
635	{
636	kfree(papPtes);
637	rc = VERR_MAP_FAILED;
638	}
639	}
640	else
641	rc = VERR_MAP_FAILED;
642	}
643	else
644	# endif
645	{
646	# if defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
647	if (fExecutable)
648	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
649	# endif
650
651	# ifdef VM_MAP
652	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
653	# else
654	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
655	# endif
656	if (pMemLnx->Core.pv)
657	pMemLnx->fMappedToRing0 = true;
658	else
659	rc = VERR_MAP_FAILED;
660	}
661	#else /* < 2.4.22 */
662	rc = VERR_NOT_SUPPORTED;
663	#endif
664	}
665	else
666	{
667	/*
668	* Use the kernel RAM mapping.
669	*/
670	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
671	Assert(pMemLnx->Core.pv);
672	}
673
674	return rc;
675	}
676
677
678	/**
679	* Undoes what rtR0MemObjLinuxVMap() did.
680	*
681	* @param pMemLnx The linux memory object.
682	*/
683	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
684	{
685	#if RTLNX_VER_MIN(2,4,22)
686	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
687	if (pMemLnx->pArea)
688	{
689	# if 0
690	pte_t **papPtes = pMemLnx->papPtesForArea;
691	size_t i;
692	for (i = 0; i < pMemLnx->cPages; i++)
693	*papPtes[i] = 0;
694	# endif
695	free_vm_area(pMemLnx->pArea);
696	kfree(pMemLnx->papPtesForArea);
697	pMemLnx->pArea = NULL;
698	pMemLnx->papPtesForArea = NULL;
699	}
700	else
701	# endif
702	if (pMemLnx->fMappedToRing0)
703	{
704	Assert(pMemLnx->Core.pv);
705	vunmap(pMemLnx->Core.pv);
706	pMemLnx->fMappedToRing0 = false;
707	}
708	#else /* < 2.4.22 */
709	Assert(!pMemLnx->fMappedToRing0);
710	#endif
711	pMemLnx->Core.pv = NULL;
712	}
713
714
715	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
716	{
717	IPRT_LINUX_SAVE_EFL_AC();
718	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
719
720	/*
721	* Release any memory that we've allocated or locked.
722	*/
723	switch (pMemLnx->Core.enmType)
724	{
725	case RTR0MEMOBJTYPE_PAGE:
726	case RTR0MEMOBJTYPE_LOW:
727	case RTR0MEMOBJTYPE_CONT:
728	case RTR0MEMOBJTYPE_PHYS:
729	case RTR0MEMOBJTYPE_PHYS_NC:
730	rtR0MemObjLinuxVUnmap(pMemLnx);
731	rtR0MemObjLinuxFreePages(pMemLnx);
732	break;
733
734	case RTR0MEMOBJTYPE_LARGE_PAGE:
735	{
736	uint32_t const cLargePages = pMemLnx->Core.cb >> (pMemLnx->cLargePageOrder + PAGE_SHIFT);
737	uint32_t iLargePage;
738	for (iLargePage = 0; iLargePage < cLargePages; iLargePage++)
739	__free_pages(pMemLnx->apPages[iLargePage << pMemLnx->cLargePageOrder], pMemLnx->cLargePageOrder);
740	pMemLnx->cPages = 0;
741
742	#ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
743	Assert(!pMemLnx->pArea);
744	Assert(!pMemLnx->papPtesForArea);
745	#endif
746	break;
747	}
748
749	case RTR0MEMOBJTYPE_LOCK:
750	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
751	{
752	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
753	size_t iPage;
754	Assert(pTask);
755	if (pTask && pTask->mm)
756	LNX_MM_DOWN_READ(pTask->mm);
757
758	iPage = pMemLnx->cPages;
759	while (iPage-- > 0)
760	{
761	if (!PageReserved(pMemLnx->apPages[iPage]))
762	SetPageDirty(pMemLnx->apPages[iPage]);
763	#if RTLNX_VER_MIN(4,6,0)
764	put_page(pMemLnx->apPages[iPage]);
765	#else
766	page_cache_release(pMemLnx->apPages[iPage]);
767	#endif
768	}
769
770	if (pTask && pTask->mm)
771	LNX_MM_UP_READ(pTask->mm);
772	}
773	/* else: kernel memory - nothing to do here. */
774	break;
775
776	case RTR0MEMOBJTYPE_RES_VIRT:
777	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	{
787	vunmap(pMemLnx->Core.pv);
788
789	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
790	__free_page(pMemLnx->apPages[0]);
791	pMemLnx->apPages[0] = NULL;
792	pMemLnx->cPages = 0;
793	}
794	pMemLnx->Core.pv = NULL;
795	break;
796
797	case RTR0MEMOBJTYPE_MAPPING:
798	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
799	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
800	{
801	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
802	Assert(pTask);
803	if (pTask && pTask->mm)
804	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
805	}
806	else
807	vunmap(pMemLnx->Core.pv);
808	pMemLnx->Core.pv = NULL;
809	break;
810
811	default:
812	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
813	return VERR_INTERNAL_ERROR;
814	}
815	IPRT_LINUX_RESTORE_EFL_ONLY_AC();
816	return VINF_SUCCESS;
817	}
818
819
820	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
821	{
822	IPRT_LINUX_SAVE_EFL_AC();
823	PRTR0MEMOBJLNX pMemLnx;
824	int rc;
825
826	#if RTLNX_VER_MIN(2,4,22)
827	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
828	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, pszTag);
829	#else
830	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
831	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, pszTag);
832	#endif
833	if (RT_SUCCESS(rc))
834	{
835	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
836	if (RT_SUCCESS(rc))
837	{
838	*ppMem = &pMemLnx->Core;
839	IPRT_LINUX_RESTORE_EFL_AC();
840	return rc;
841	}
842
843	rtR0MemObjLinuxFreePages(pMemLnx);
844	rtR0MemObjDelete(&pMemLnx->Core);
845	}
846
847	IPRT_LINUX_RESTORE_EFL_AC();
848	return rc;
849	}
850
851
852	DECLHIDDEN(int) rtR0MemObjNativeAllocLarge(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, size_t cbLargePage, uint32_t fFlags,
853	const char *pszTag)
854	{
855	#ifdef GFP_TRANSHUGE
856	/*
857	* Allocate a memory object structure that's large enough to contain
858	* the page pointer array.
859	*/
860	# ifdef __GFP_MOVABLE
861	unsigned const fGfp = (GFP_TRANSHUGE \| __GFP_ZERO) & ~__GFP_MOVABLE;
862	# else
863	unsigned const fGfp = (GFP_TRANSHUGE \| __GFP_ZERO);
864	# endif
865	size_t const cPagesPerLarge = cbLargePage >> PAGE_SHIFT;
866	unsigned const cLargePageOrder = rtR0MemObjLinuxOrder(cPagesPerLarge);
867	size_t const cLargePages = cb >> (cLargePageOrder + PAGE_SHIFT);
868	size_t const cPages = cb >> PAGE_SHIFT;
869	PRTR0MEMOBJLNX pMemLnx;
870
871	Assert(RT_BIT_64(cLargePageOrder + PAGE_SHIFT) == cbLargePage);
872	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]),
873	RTR0MEMOBJTYPE_LARGE_PAGE, NULL, cb, pszTag);
874	if (pMemLnx)
875	{
876	size_t iLargePage;
877
878	pMemLnx->Core.fFlags \|= RTR0MEMOBJ_FLAGS_ZERO_AT_ALLOC;
879	pMemLnx->cLargePageOrder = cLargePageOrder;
880	pMemLnx->cPages = cPages;
881
882	/*
883	* Allocate the requested number of large pages.
884	*/
885	for (iLargePage = 0; iLargePage < cLargePages; iLargePage++)
886	{
887	struct page *paPages = alloc_pages(fGfp, cLargePageOrder);
888	if (paPages)
889	{
890	size_t const iPageBase = iLargePage << cLargePageOrder;
891	size_t iPage = cPagesPerLarge;
892	while (iPage-- > 0)
893	pMemLnx->apPages[iPageBase + iPage] = &paPages[iPage];
894	}
895	else
896	{
897	/*Log(("rtR0MemObjNativeAllocLarge: cb=%#zx cPages=%#zx cLargePages=%#zx cLargePageOrder=%u cPagesPerLarge=%#zx iLargePage=%#zx -> failed!\n",
898	cb, cPages, cLargePages, cLargePageOrder, cPagesPerLarge, iLargePage, paPages));*/
899	while (iLargePage-- > 0)
900	__free_pages(pMemLnx->apPages[iLargePage << (cLargePageOrder - PAGE_SHIFT)], cLargePageOrder);
901	rtR0MemObjDelete(&pMemLnx->Core);
902	return VERR_NO_MEMORY;
903	}
904	}
905	*ppMem = &pMemLnx->Core;
906	return VINF_SUCCESS;
907	}
908	return VERR_NO_MEMORY;
909
910	#else
911	/*
912	* We don't call rtR0MemObjFallbackAllocLarge here as it can be a really
913	* bad idea to trigger the swap daemon and whatnot. So, just fail.
914	*/
915	RT_NOREF(ppMem, cb, cbLargePage, fFlags, pszTag);
916	return VERR_NOT_SUPPORTED;
917	#endif
918	}
919
920
921	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
922	{
923	IPRT_LINUX_SAVE_EFL_AC();
924	PRTR0MEMOBJLNX pMemLnx;
925	int rc;
926
927	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
928	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
929	/* ZONE_DMA32: 0-4GB */
930	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
931	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
932	if (RT_FAILURE(rc))
933	#endif
934	#ifdef RT_ARCH_AMD64
935	/* ZONE_DMA: 0-16MB */
936	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
937	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
938	#else
939	# ifdef CONFIG_X86_PAE
940	# endif
941	/* ZONE_NORMAL: 0-896MB */
942	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
943	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
944	#endif
945	if (RT_SUCCESS(rc))
946	{
947	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
948	if (RT_SUCCESS(rc))
949	{
950	*ppMem = &pMemLnx->Core;
951	IPRT_LINUX_RESTORE_EFL_AC();
952	return rc;
953	}
954
955	rtR0MemObjLinuxFreePages(pMemLnx);
956	rtR0MemObjDelete(&pMemLnx->Core);
957	}
958
959	IPRT_LINUX_RESTORE_EFL_AC();
960	return rc;
961	}
962
963
964	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
965	{
966	IPRT_LINUX_SAVE_EFL_AC();
967	PRTR0MEMOBJLNX pMemLnx;
968	int rc;
969
970	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
971	/* ZONE_DMA32: 0-4GB */
972	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
973	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
974	if (RT_FAILURE(rc))
975	#endif
976	#ifdef RT_ARCH_AMD64
977	/* ZONE_DMA: 0-16MB */
978	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
979	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
980	#else
981	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
982	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
983	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
984	#endif
985	if (RT_SUCCESS(rc))
986	{
987	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
988	if (RT_SUCCESS(rc))
989	{
990	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
991	size_t iPage = pMemLnx->cPages;
992	while (iPage-- > 0)
993	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
994	#endif
995	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
996	*ppMem = &pMemLnx->Core;
997	IPRT_LINUX_RESTORE_EFL_AC();
998	return rc;
999	}
1000
1001	rtR0MemObjLinuxFreePages(pMemLnx);
1002	rtR0MemObjDelete(&pMemLnx->Core);
1003	}
1004
1005	IPRT_LINUX_RESTORE_EFL_AC();
1006	return rc;
1007	}
1008
1009
1010	/**
1011	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
1012	*
1013	* @returns IPRT status code.
1014	* @param ppMemLnx Where to
1015	* @param enmType The object type.
1016	* @param cb The size of the allocation.
1017	* @param uAlignment The alignment of the physical memory.
1018	* Only valid for fContiguous == true, ignored otherwise.
1019	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
1020	* @param pszTag Allocation tag used for statistics and such.
1021	* @param fGfp The Linux GFP flags to use for the allocation.
1022	*/
1023	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
1024	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag, gfp_t fGfp)
1025	{
1026	PRTR0MEMOBJLNX pMemLnx;
1027	int rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
1028	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
1029	false /fExecutable/, VERR_NO_PHYS_MEMORY, pszTag);
1030	if (RT_FAILURE(rc))
1031	return rc;
1032
1033	/*
1034	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
1035	*/
1036	if (PhysHighest != NIL_RTHCPHYS)
1037	{
1038	size_t iPage = pMemLnx->cPages;
1039	while (iPage-- > 0)
1040	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
1041	{
1042	rtR0MemObjLinuxFreePages(pMemLnx);
1043	rtR0MemObjDelete(&pMemLnx->Core);
1044	return VERR_NO_MEMORY;
1045	}
1046	}
1047
1048	/*
1049	* Complete the object.
1050	*/
1051	if (enmType == RTR0MEMOBJTYPE_PHYS)
1052	{
1053	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
1054	pMemLnx->Core.u.Phys.fAllocated = true;
1055	}
1056	*ppMem = &pMemLnx->Core;
1057	return rc;
1058	}
1059
1060
1061	/**
1062	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
1063	*
1064	* @returns IPRT status code.
1065	* @param ppMem Where to store the memory object pointer on success.
1066	* @param enmType The object type.
1067	* @param cb The size of the allocation.
1068	* @param uAlignment The alignment of the physical memory.
1069	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
1070	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
1071	* @param pszTag Allocation tag used for statistics and such.
1072	*/
1073	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
1074	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag)
1075	{
1076	int rc;
1077	IPRT_LINUX_SAVE_EFL_AC();
1078
1079	/*
1080	* There are two clear cases and that's the <=16MB and anything-goes ones.
1081	* When the physical address limit is somewhere in-between those two we'll
1082	* just have to try, starting with HIGHUSER and working our way thru the
1083	* different types, hoping we'll get lucky.
1084	*
1085	* We should probably move this physical address restriction logic up to
1086	* the page alloc function as it would be more efficient there. But since
1087	* we don't expect this to be a performance issue just yet it can wait.
1088	*/
1089	if (PhysHighest == NIL_RTHCPHYS)
1090	/* ZONE_HIGHMEM: the whole physical memory */
1091	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1092	else if (PhysHighest <= _1M * 16)
1093	/* ZONE_DMA: 0-16MB */
1094	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1095	else
1096	{
1097	rc = VERR_NO_MEMORY;
1098	if (RT_FAILURE(rc))
1099	/* ZONE_HIGHMEM: the whole physical memory */
1100	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1101	if (RT_FAILURE(rc))
1102	/* ZONE_NORMAL: 0-896MB */
1103	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_USER);
1104	#ifdef GFP_DMA32
1105	if (RT_FAILURE(rc))
1106	/* ZONE_DMA32: 0-4GB */
1107	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA32);
1108	#endif
1109	if (RT_FAILURE(rc))
1110	/* ZONE_DMA: 0-16MB */
1111	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1112	}
1113	IPRT_LINUX_RESTORE_EFL_AC();
1114	return rc;
1115	}
1116
1117
1118	/**
1119	* Translates a kernel virtual address to a linux page structure by walking the
1120	* page tables.
1121	*
1122	* @note We do assume that the page tables will not change as we are walking
1123	* them. This assumption is rather forced by the fact that I could not
1124	* immediately see any way of preventing this from happening. So, we
1125	* take some extra care when accessing them.
1126	*
1127	* Because of this, we don't want to use this function on memory where
1128	* attribute changes to nearby pages is likely to cause large pages to
1129	* be used or split up. So, don't use this for the linear mapping of
1130	* physical memory.
1131	*
1132	* @returns Pointer to the page structur or NULL if it could not be found.
1133	* @param pv The kernel virtual address.
1134	*/
1135	RTDECL(struct page ) rtR0MemObjLinuxVirtToPage(void pv)
1136	{
1137	unsigned long ulAddr = (unsigned long)pv;
1138	unsigned long pfn;
1139	struct page *pPage;
1140	pte_t *pEntry;
1141	union
1142	{
1143	pgd_t Global;
1144	#if RTLNX_VER_MIN(4,12,0)
1145	p4d_t Four;
1146	#endif
1147	#if RTLNX_VER_MIN(2,6,11)
1148	pud_t Upper;
1149	#endif
1150	pmd_t Middle;
1151	pte_t Entry;
1152	} u;
1153
1154	/* Should this happen in a situation this code will be called in? And if
1155	* so, can it change under our feet? See also
1156	* "Documentation/vm/active_mm.txt" in the kernel sources. */
1157	if (RT_UNLIKELY(!current->active_mm))
1158	return NULL;
1159	u.Global = *pgd_offset(current->active_mm, ulAddr);
1160	if (RT_UNLIKELY(pgd_none(u.Global)))
1161	return NULL;
1162	#if RTLNX_VER_MIN(2,6,11)
1163	# if RTLNX_VER_MIN(4,12,0)
1164	u.Four = *p4d_offset(&u.Global, ulAddr);
1165	if (RT_UNLIKELY(p4d_none(u.Four)))
1166	return NULL;
1167	if (p4d_large(u.Four))
1168	{
1169	pPage = p4d_page(u.Four);
1170	AssertReturn(pPage, NULL);
1171	pfn = page_to_pfn(pPage); /* doing the safe way... */
1172	AssertCompile(P4D_SHIFT - PAGE_SHIFT < 31);
1173	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (P4D_SHIFT - PAGE_SHIFT)) - 1);
1174	return pfn_to_page(pfn);
1175	}
1176	u.Upper = *pud_offset(&u.Four, ulAddr);
1177	# else /* < 4.12 */
1178	u.Upper = *pud_offset(&u.Global, ulAddr);
1179	# endif /* < 4.12 */
1180	if (RT_UNLIKELY(pud_none(u.Upper)))
1181	return NULL;
1182	# if RTLNX_VER_MIN(2,6,25)
1183	if (pud_large(u.Upper))
1184	{
1185	pPage = pud_page(u.Upper);
1186	AssertReturn(pPage, NULL);
1187	pfn = page_to_pfn(pPage); /* doing the safe way... */
1188	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
1189	return pfn_to_page(pfn);
1190	}
1191	# endif
1192	u.Middle = *pmd_offset(&u.Upper, ulAddr);
1193	#else /* < 2.6.11 */
1194	u.Middle = *pmd_offset(&u.Global, ulAddr);
1195	#endif /* < 2.6.11 */
1196	if (RT_UNLIKELY(pmd_none(u.Middle)))
1197	return NULL;
1198	#if RTLNX_VER_MIN(2,6,0)
1199	if (pmd_large(u.Middle))
1200	{
1201	pPage = pmd_page(u.Middle);
1202	AssertReturn(pPage, NULL);
1203	pfn = page_to_pfn(pPage); /* doing the safe way... */
1204	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
1205	return pfn_to_page(pfn);
1206	}
1207	#endif
1208
1209	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
1210	pEntry = pte_offset_map(&u.Middle, ulAddr);
1211	#else
1212	pEntry = pte_offset(&u.Middle, ulAddr);
1213	#endif
1214	if (RT_UNLIKELY(!pEntry))
1215	return NULL;
1216	u.Entry = *pEntry;
1217	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map)
1218	pte_unmap(pEntry);
1219	#endif
1220
1221	if (RT_UNLIKELY(!pte_present(u.Entry)))
1222	return NULL;
1223	return pte_page(u.Entry);
1224	}
1225	RT_EXPORT_SYMBOL(rtR0MemObjLinuxVirtToPage);
1226
1227
1228	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment,
1229	const char *pszTag)
1230	{
1231	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest, pszTag);
1232	}
1233
1234
1235	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, const char *pszTag)
1236	{
1237	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest, pszTag);
1238	}
1239
1240
1241	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy,
1242	const char *pszTag)
1243	{
1244	IPRT_LINUX_SAVE_EFL_AC();
1245
1246	/*
1247	* All we need to do here is to validate that we can use
1248	* ioremap on the specified address (32/64-bit dma_addr_t).
1249	*/
1250	PRTR0MEMOBJLNX pMemLnx;
1251	dma_addr_t PhysAddr = Phys;
1252	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
1253
1254	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb, pszTag);
1255	if (!pMemLnx)
1256	{
1257	IPRT_LINUX_RESTORE_EFL_AC();
1258	return VERR_NO_MEMORY;
1259	}
1260
1261	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
1262	pMemLnx->Core.u.Phys.fAllocated = false;
1263	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
1264	Assert(!pMemLnx->cPages);
1265	*ppMem = &pMemLnx->Core;
1266	IPRT_LINUX_RESTORE_EFL_AC();
1267	return VINF_SUCCESS;
1268	}
1269
1270	/* openSUSE Leap 42.3 detection :-/ */
1271	#if RTLNX_VER_RANGE(4,4,0, 4,6,0) && defined(FAULT_FLAG_REMOTE)
1272	# define GET_USER_PAGES_API KERNEL_VERSION(4, 10, 0) /* no typo! */
1273	#else
1274	# define GET_USER_PAGES_API LINUX_VERSION_CODE
1275	#endif
1276
1277	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess,
1278	RTR0PROCESS R0Process, const char *pszTag)
1279	{
1280	IPRT_LINUX_SAVE_EFL_AC();
1281	const int cPages = cb >> PAGE_SHIFT;
1282	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1283	struct vm_area_struct **papVMAs;
1284	PRTR0MEMOBJLNX pMemLnx;
1285	int rc = VERR_NO_MEMORY;
1286	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1287
1288	/*
1289	* Check for valid task and size overflows.
1290	*/
1291	if (!pTask)
1292	return VERR_NOT_SUPPORTED;
1293	if (((size_t)cPages << PAGE_SHIFT) != cb)
1294	return VERR_OUT_OF_RANGE;
1295
1296	/*
1297	* Allocate the memory object and a temporary buffer for the VMAs.
1298	*/
1299	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1300	(void *)R3Ptr, cb, pszTag);
1301	if (!pMemLnx)
1302	{
1303	IPRT_LINUX_RESTORE_EFL_AC();
1304	return VERR_NO_MEMORY;
1305	}
1306
1307	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1308	if (papVMAs)
1309	{
1310	LNX_MM_DOWN_READ(pTask->mm);
1311
1312	/*
1313	* Get user pages.
1314	*/
1315	/** @todo r=bird: Should we not force read access too? */
1316	#if GET_USER_PAGES_API >= KERNEL_VERSION(4, 6, 0)
1317	if (R0Process == RTR0ProcHandleSelf())
1318	rc = get_user_pages(R3Ptr, /* Where from. */
1319	cPages, /* How many pages. */
1320	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1321	fWrite ? FOLL_WRITE \| /* Write to memory. */
1322	FOLL_FORCE /* force write access. */
1323	: 0, /* Write to memory. */
1324	# else
1325	fWrite, /* Write to memory. */
1326	fWrite, /* force write access. */
1327	# endif
1328	&pMemLnx->apPages[0], /* Page array. */
1329	papVMAs); /* vmas */
1330	/*
1331	* Actually this should not happen at the moment as call this function
1332	* only for our own process.
1333	*/
1334	else
1335	rc = get_user_pages_remote(
1336	# if GET_USER_PAGES_API < KERNEL_VERSION(5, 9, 0)
1337	pTask, /* Task for fault accounting. */
1338	# endif
1339	pTask->mm, /* Whose pages. */
1340	R3Ptr, /* Where from. */
1341	cPages, /* How many pages. */
1342	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1343	fWrite ? FOLL_WRITE \| /* Write to memory. */
1344	FOLL_FORCE /* force write access. */
1345	: 0, /* Write to memory. */
1346	# else
1347	fWrite, /* Write to memory. */
1348	fWrite, /* force write access. */
1349	# endif
1350	&pMemLnx->apPages[0], /* Page array. */
1351	papVMAs /* vmas */
1352	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 10, 0)
1353	, NULL /* locked */
1354	# endif
1355	);
1356	#else /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1357	rc = get_user_pages(pTask, /* Task for fault accounting. */
1358	pTask->mm, /* Whose pages. */
1359	R3Ptr, /* Where from. */
1360	cPages, /* How many pages. */
1361	/* The get_user_pages API change was back-ported to 4.4.168. */
1362	# if RTLNX_VER_RANGE(4,4,168, 4,5,0)
1363	fWrite ? FOLL_WRITE \| /* Write to memory. */
1364	FOLL_FORCE /* force write access. */
1365	: 0, /* Write to memory. */
1366	# else
1367	fWrite, /* Write to memory. */
1368	fWrite, /* force write access. */
1369	# endif
1370	&pMemLnx->apPages[0], /* Page array. */
1371	papVMAs); /* vmas */
1372	#endif /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1373	if (rc == cPages)
1374	{
1375	/*
1376	* Flush dcache (required?), protect against fork and _really_ pin the page
1377	* table entries. get_user_pages() will protect against swapping out the
1378	* pages but it will NOT protect against removing page table entries. This
1379	* can be achieved with
1380	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1381	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1382	* Usual Linux distributions support only a limited size of locked pages
1383	* (e.g. 32KB).
1384	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1385	* or by
1386	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1387	* a range check.
1388	*/
1389	/** @todo The Linux fork() protection will require more work if this API
1390	* is to be used for anything but locking VM pages. */
1391	while (rc-- > 0)
1392	{
1393	flush_dcache_page(pMemLnx->apPages[rc]);
1394	papVMAs[rc]->vm_flags \|= VM_DONTCOPY \| VM_LOCKED;
1395	}
1396
1397	LNX_MM_UP_READ(pTask->mm);
1398
1399	RTMemFree(papVMAs);
1400
1401	pMemLnx->Core.u.Lock.R0Process = R0Process;
1402	pMemLnx->cPages = cPages;
1403	Assert(!pMemLnx->fMappedToRing0);
1404	*ppMem = &pMemLnx->Core;
1405
1406	IPRT_LINUX_RESTORE_EFL_AC();
1407	return VINF_SUCCESS;
1408	}
1409
1410	/*
1411	* Failed - we need to unlock any pages that we succeeded to lock.
1412	*/
1413	while (rc-- > 0)
1414	{
1415	if (!PageReserved(pMemLnx->apPages[rc]))
1416	SetPageDirty(pMemLnx->apPages[rc]);
1417	#if RTLNX_VER_MIN(4,6,0)
1418	put_page(pMemLnx->apPages[rc]);
1419	#else
1420	page_cache_release(pMemLnx->apPages[rc]);
1421	#endif
1422	}
1423
1424	LNX_MM_UP_READ(pTask->mm);
1425
1426	RTMemFree(papVMAs);
1427	rc = VERR_LOCK_FAILED;
1428	}
1429
1430	rtR0MemObjDelete(&pMemLnx->Core);
1431	IPRT_LINUX_RESTORE_EFL_AC();
1432	return rc;
1433	}
1434
1435
1436	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void pv, size_t cb, uint32_t fAccess, const char pszTag)
1437	{
1438	IPRT_LINUX_SAVE_EFL_AC();
1439	void pvLast = (uint8_t )pv + cb - 1;
1440	size_t const cPages = cb >> PAGE_SHIFT;
1441	PRTR0MEMOBJLNX pMemLnx;
1442	bool fLinearMapping;
1443	int rc;
1444	uint8_t *pbPage;
1445	size_t iPage;
1446	NOREF(fAccess);
1447
1448	if ( !RTR0MemKernelIsValidAddr(pv)
1449	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1450	return VERR_INVALID_PARAMETER;
1451
1452	/*
1453	* The lower part of the kernel memory has a linear mapping between
1454	* physical and virtual addresses. So we take a short cut here. This is
1455	* assumed to be the cleanest way to handle those addresses (and the code
1456	* is well tested, though the test for determining it is not very nice).
1457	* If we ever decide it isn't we can still remove it.
1458	*/
1459	#if 0
1460	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1461	#else
1462	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1463	&& (unsigned long)pvLast < (unsigned long)high_memory;
1464	#endif
1465
1466	/*
1467	* Allocate the memory object.
1468	*/
1469	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1470	pv, cb, pszTag);
1471	if (!pMemLnx)
1472	{
1473	IPRT_LINUX_RESTORE_EFL_AC();
1474	return VERR_NO_MEMORY;
1475	}
1476
1477	/*
1478	* Gather the pages.
1479	* We ASSUME all kernel pages are non-swappable and non-movable.
1480	*/
1481	rc = VINF_SUCCESS;
1482	pbPage = (uint8_t *)pvLast;
1483	iPage = cPages;
1484	if (!fLinearMapping)
1485	{
1486	while (iPage-- > 0)
1487	{
1488	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1489	if (RT_UNLIKELY(!pPage))
1490	{
1491	rc = VERR_LOCK_FAILED;
1492	break;
1493	}
1494	pMemLnx->apPages[iPage] = pPage;
1495	pbPage -= PAGE_SIZE;
1496	}
1497	}
1498	else
1499	{
1500	while (iPage-- > 0)
1501	{
1502	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1503	pbPage -= PAGE_SIZE;
1504	}
1505	}
1506	if (RT_SUCCESS(rc))
1507	{
1508	/*
1509	* Complete the memory object and return.
1510	*/
1511	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1512	pMemLnx->cPages = cPages;
1513	Assert(!pMemLnx->fMappedToRing0);
1514	*ppMem = &pMemLnx->Core;
1515
1516	IPRT_LINUX_RESTORE_EFL_AC();
1517	return VINF_SUCCESS;
1518	}
1519
1520	rtR0MemObjDelete(&pMemLnx->Core);
1521	IPRT_LINUX_RESTORE_EFL_AC();
1522	return rc;
1523	}
1524
1525
1526	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment,
1527	const char *pszTag)
1528	{
1529	#if RTLNX_VER_MIN(2,4,22)
1530	IPRT_LINUX_SAVE_EFL_AC();
1531	const size_t cPages = cb >> PAGE_SHIFT;
1532	struct page *pDummyPage;
1533	struct page **papPages;
1534
1535	/* check for unsupported stuff. */
1536	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1537	if (uAlignment > PAGE_SIZE)
1538	return VERR_NOT_SUPPORTED;
1539
1540	/*
1541	* Allocate a dummy page and create a page pointer array for vmap such that
1542	* the dummy page is mapped all over the reserved area.
1543	*/
1544	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1545	if (pDummyPage)
1546	{
1547	papPages = RTMemAlloc(sizeof(papPages) cPages);
1548	if (papPages)
1549	{
1550	void *pv;
1551	size_t iPage = cPages;
1552	while (iPage-- > 0)
1553	papPages[iPage] = pDummyPage;
1554	# ifdef VM_MAP
1555	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1556	# else
1557	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1558	# endif
1559	RTMemFree(papPages);
1560	if (pv)
1561	{
1562	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1563	if (pMemLnx)
1564	{
1565	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1566	pMemLnx->cPages = 1;
1567	pMemLnx->apPages[0] = pDummyPage;
1568	*ppMem = &pMemLnx->Core;
1569	IPRT_LINUX_RESTORE_EFL_AC();
1570	return VINF_SUCCESS;
1571	}
1572	vunmap(pv);
1573	}
1574	}
1575	__free_page(pDummyPage);
1576	}
1577	IPRT_LINUX_RESTORE_EFL_AC();
1578	return VERR_NO_MEMORY;
1579
1580	#else /* < 2.4.22 */
1581	/*
1582	* Could probably use ioremap here, but the caller is in a better position than us
1583	* to select some safe physical memory.
1584	*/
1585	return VERR_NOT_SUPPORTED;
1586	#endif
1587	}
1588
1589
1590	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment,
1591	RTR0PROCESS R0Process, const char *pszTag)
1592	{
1593	IPRT_LINUX_SAVE_EFL_AC();
1594	PRTR0MEMOBJLNX pMemLnx;
1595	void *pv;
1596	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1597	if (!pTask)
1598	return VERR_NOT_SUPPORTED;
1599
1600	/*
1601	* Check that the specified alignment is supported.
1602	*/
1603	if (uAlignment > PAGE_SIZE)
1604	return VERR_NOT_SUPPORTED;
1605
1606	/*
1607	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1608	*/
1609	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1610	if (pv == (void *)-1)
1611	{
1612	IPRT_LINUX_RESTORE_EFL_AC();
1613	return VERR_NO_MEMORY;
1614	}
1615
1616	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1617	if (!pMemLnx)
1618	{
1619	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1620	IPRT_LINUX_RESTORE_EFL_AC();
1621	return VERR_NO_MEMORY;
1622	}
1623
1624	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1625	*ppMem = &pMemLnx->Core;
1626	IPRT_LINUX_RESTORE_EFL_AC();
1627	return VINF_SUCCESS;
1628	}
1629
1630
1631	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment,
1632	unsigned fProt, size_t offSub, size_t cbSub, const char *pszTag)
1633	{
1634	int rc = VERR_NO_MEMORY;
1635	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1636	PRTR0MEMOBJLNX pMemLnx;
1637	IPRT_LINUX_SAVE_EFL_AC();
1638
1639	/* Fail if requested to do something we can't. */
1640	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1641	if (uAlignment > PAGE_SIZE)
1642	return VERR_NOT_SUPPORTED;
1643
1644	/*
1645	* Create the IPRT memory object.
1646	*/
1647	if (!cbSub)
1648	cbSub = pMemLnxToMap->Core.cb - offSub;
1649	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1650	if (pMemLnx)
1651	{
1652	if (pMemLnxToMap->cPages)
1653	{
1654	#if RTLNX_VER_MIN(2,4,22)
1655	/*
1656	* Use vmap - 2.4.22 and later.
1657	*/
1658	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1659	/** @todo We don't really care too much for EXEC here... 5.8 always adds NX. */
1660	Assert(((offSub + cbSub) >> PAGE_SHIFT) <= pMemLnxToMap->cPages);
1661	# ifdef VM_MAP
1662	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_MAP, fPg);
1663	# else
1664	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_ALLOC, fPg);
1665	# endif
1666	if (pMemLnx->Core.pv)
1667	{
1668	pMemLnx->fMappedToRing0 = true;
1669	rc = VINF_SUCCESS;
1670	}
1671	else
1672	rc = VERR_MAP_FAILED;
1673
1674	#else /* < 2.4.22 */
1675	/*
1676	* Only option here is to share mappings if possible and forget about fProt.
1677	*/
1678	if (rtR0MemObjIsRing3(pMemToMap))
1679	rc = VERR_NOT_SUPPORTED;
1680	else
1681	{
1682	rc = VINF_SUCCESS;
1683	if (!pMemLnxToMap->Core.pv)
1684	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1685	if (RT_SUCCESS(rc))
1686	{
1687	Assert(pMemLnxToMap->Core.pv);
1688	pMemLnx->Core.pv = (uint8_t *)pMemLnxToMap->Core.pv + offSub;
1689	}
1690	}
1691	#endif
1692	}
1693	else
1694	{
1695	/*
1696	* MMIO / physical memory.
1697	*/
1698	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1699	#if RTLNX_VER_MIN(2,6,25)
1700	/*
1701	* ioremap() defaults to no caching since the 2.6 kernels.
1702	* ioremap_nocache() has been removed finally in 5.6-rc1.
1703	*/
1704	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1705	? ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1706	: ioremap_cache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1707	#else /* KERNEL_VERSION < 2.6.25 */
1708	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1709	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1710	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1711	#endif /* KERNEL_VERSION < 2.6.25 */
1712	if (pMemLnx->Core.pv)
1713	{
1714	/** @todo fix protection. */
1715	rc = VINF_SUCCESS;
1716	}
1717	}
1718	if (RT_SUCCESS(rc))
1719	{
1720	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1721	*ppMem = &pMemLnx->Core;
1722	IPRT_LINUX_RESTORE_EFL_AC();
1723	return VINF_SUCCESS;
1724	}
1725	rtR0MemObjDelete(&pMemLnx->Core);
1726	}
1727
1728	IPRT_LINUX_RESTORE_EFL_AC();
1729	return rc;
1730	}
1731
1732
1733	#ifdef VBOX_USE_PAE_HACK
1734	/**
1735	* Replace the PFN of a PTE with the address of the actual page.
1736	*
1737	* The caller maps a reserved dummy page at the address with the desired access
1738	* and flags.
1739	*
1740	* This hack is required for older Linux kernels which don't provide
1741	* remap_pfn_range().
1742	*
1743	* @returns 0 on success, -ENOMEM on failure.
1744	* @param mm The memory context.
1745	* @param ulAddr The mapping address.
1746	* @param Phys The physical address of the page to map.
1747	*/
1748	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1749	{
1750	int rc = -ENOMEM;
1751	pgd_t *pgd;
1752
1753	spin_lock(&mm->page_table_lock);
1754
1755	pgd = pgd_offset(mm, ulAddr);
1756	if (!pgd_none(pgd) && !pgd_bad(pgd))
1757	{
1758	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1759	if (!pmd_none(*pmd))
1760	{
1761	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1762	if (ptep)
1763	{
1764	pte_t pte = *ptep;
1765	pte.pte_high &= 0xfff00000;
1766	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1767	pte.pte_low &= 0x00000fff;
1768	pte.pte_low \|= (Phys & 0xfffff000);
1769	set_pte(ptep, pte);
1770	pte_unmap(ptep);
1771	rc = 0;
1772	}
1773	}
1774	}
1775
1776	spin_unlock(&mm->page_table_lock);
1777	return rc;
1778	}
1779	#endif /* VBOX_USE_PAE_HACK */
1780
1781
1782	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment,
1783	unsigned fProt, RTR0PROCESS R0Process, size_t offSub, size_t cbSub, const char *pszTag)
1784	{
1785	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1786	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1787	int rc = VERR_NO_MEMORY;
1788	PRTR0MEMOBJLNX pMemLnx;
1789	#ifdef VBOX_USE_PAE_HACK
1790	struct page *pDummyPage;
1791	RTHCPHYS DummyPhys;
1792	#endif
1793	IPRT_LINUX_SAVE_EFL_AC();
1794
1795	/*
1796	* Check for restrictions.
1797	*/
1798	if (!pTask)
1799	return VERR_NOT_SUPPORTED;
1800	if (uAlignment > PAGE_SIZE)
1801	return VERR_NOT_SUPPORTED;
1802
1803	#ifdef VBOX_USE_PAE_HACK
1804	/*
1805	* Allocate a dummy page for use when mapping the memory.
1806	*/
1807	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1808	if (!pDummyPage)
1809	{
1810	IPRT_LINUX_RESTORE_EFL_AC();
1811	return VERR_NO_MEMORY;
1812	}
1813	SetPageReserved(pDummyPage);
1814	DummyPhys = page_to_phys(pDummyPage);
1815	#endif
1816
1817	/*
1818	* Create the IPRT memory object.
1819	*/
1820	Assert(!offSub \|\| cbSub);
1821	if (cbSub == 0)
1822	cbSub = pMemLnxToMap->Core.cb;
1823	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1824	if (pMemLnx)
1825	{
1826	/*
1827	* Allocate user space mapping.
1828	*/
1829	void *pv;
1830	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cbSub, uAlignment, pTask, fProt);
1831	if (pv != (void *)-1)
1832	{
1833	/*
1834	* Map page by page into the mmap area.
1835	* This is generic, paranoid and not very efficient.
1836	*/
1837	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1838	unsigned long ulAddrCur = (unsigned long)pv;
1839	const size_t cPages = (offSub + cbSub) >> PAGE_SHIFT;
1840	size_t iPage;
1841
1842	LNX_MM_DOWN_WRITE(pTask->mm);
1843
1844	rc = VINF_SUCCESS;
1845	if (pMemLnxToMap->cPages)
1846	{
1847	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1848	{
1849	#if RTLNX_VER_MAX(2,6,11)
1850	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1851	#endif
1852	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1853	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1854	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1855	#endif
1856	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1857	/* remap_page_range() limitation on x86 */
1858	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1859	#endif
1860
1861	#if defined(VBOX_USE_INSERT_PAGE) && RTLNX_VER_MIN(2,6,22)
1862	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1863	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1864	* See remap_pfn_range() in mm/memory.c */
1865	#if RTLNX_VER_MIN(3,7,0)
1866	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1867	#else
1868	vma->vm_flags \|= VM_RESERVED;
1869	#endif
1870	#elif RTLNX_VER_MIN(2,6,11)
1871	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1872	#elif defined(VBOX_USE_PAE_HACK)
1873	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1874	if (!rc)
1875	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1876	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1877	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1878	#else /* 2.4 */
1879	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1880	#endif
1881	if (rc)
1882	{
1883	rc = VERR_NO_MEMORY;
1884	break;
1885	}
1886	}
1887	}
1888	else
1889	{
1890	RTHCPHYS Phys;
1891	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1892	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1893	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1894	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1895	else
1896	{
1897	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1898	Phys = NIL_RTHCPHYS;
1899	}
1900	if (Phys != NIL_RTHCPHYS)
1901	{
1902	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1903	{
1904	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1905	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1906	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1907	#endif
1908	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1909	/* remap_page_range() limitation on x86 */
1910	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1911	#endif
1912
1913	#if RTLNX_VER_MIN(2,6,11)
1914	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1915	#elif defined(VBOX_USE_PAE_HACK)
1916	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1917	if (!rc)
1918	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1919	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1920	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1921	#else /* 2.4 */
1922	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1923	#endif
1924	if (rc)
1925	{
1926	rc = VERR_NO_MEMORY;
1927	break;
1928	}
1929	}
1930	}
1931	}
1932
1933	#ifdef CONFIG_NUMA_BALANCING
1934	# if RTLNX_VER_MAX(3,13,0) && RTLNX_RHEL_MAX(7,0)
1935	# define VBOX_NUMA_HACK_OLD
1936	# endif
1937	if (RT_SUCCESS(rc))
1938	{
1939	/** @todo Ugly hack! But right now we have no other means to
1940	* disable automatic NUMA page balancing. */
1941	# ifdef RT_OS_X86
1942	# ifdef VBOX_NUMA_HACK_OLD
1943	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1944	# endif
1945	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1946	# else
1947	# ifdef VBOX_NUMA_HACK_OLD
1948	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1949	# endif
1950	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1951	# endif
1952	}
1953	#endif /* CONFIG_NUMA_BALANCING */
1954
1955	LNX_MM_UP_WRITE(pTask->mm);
1956
1957	if (RT_SUCCESS(rc))
1958	{
1959	#ifdef VBOX_USE_PAE_HACK
1960	__free_page(pDummyPage);
1961	#endif
1962	pMemLnx->Core.pv = pv;
1963	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1964	*ppMem = &pMemLnx->Core;
1965	IPRT_LINUX_RESTORE_EFL_AC();
1966	return VINF_SUCCESS;
1967	}
1968
1969	/*
1970	* Bail out.
1971	*/
1972	rtR0MemObjLinuxDoMunmap(pv, cbSub, pTask);
1973	}
1974	rtR0MemObjDelete(&pMemLnx->Core);
1975	}
1976	#ifdef VBOX_USE_PAE_HACK
1977	__free_page(pDummyPage);
1978	#endif
1979
1980	IPRT_LINUX_RESTORE_EFL_AC();
1981	return rc;
1982	}
1983
1984
1985	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
1986	{
1987	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
1988	/*
1989	* Currently only supported when we've got addresses PTEs from the kernel.
1990	*/
1991	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1992	if (pMemLnx->pArea && pMemLnx->papPtesForArea)
1993	{
1994	pgprot_t const fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
1995	size_t const cPages = (offSub + cbSub) >> PAGE_SHIFT;
1996	pte_t **papPtes = pMemLnx->papPtesForArea;
1997	size_t i;
1998
1999	for (i = offSub >> PAGE_SHIFT; i < cPages; i++)
2000	{
2001	set_pte(papPtes[i], mk_pte(pMemLnx->apPages[i], fPg));
2002	}
2003	preempt_disable();
2004	__flush_tlb_all();
2005	preempt_enable();
2006	return VINF_SUCCESS;
2007	}
2008	# elif defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
2009	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
2010	if ( pMemLnx->fExecutable
2011	&& pMemLnx->fMappedToRing0)
2012	{
2013	LNXAPPLYPGRANGE Args;
2014	Args.pMemLnx = pMemLnx;
2015	Args.fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
2016	int rcLnx = apply_to_page_range(current->active_mm, (unsigned long)pMemLnx->Core.pv + offSub, cbSub,
2017	rtR0MemObjLinuxApplyPageRange, (void *)&Args);
2018	if (rcLnx)
2019	return VERR_NOT_SUPPORTED;
2020
2021	return VINF_SUCCESS;
2022	}
2023	# endif
2024
2025	NOREF(pMem);
2026	NOREF(offSub);
2027	NOREF(cbSub);
2028	NOREF(fProt);
2029	return VERR_NOT_SUPPORTED;
2030	}
2031
2032
2033	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
2034	{
2035	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
2036
2037	if (pMemLnx->cPages)
2038	return page_to_phys(pMemLnx->apPages[iPage]);
2039
2040	switch (pMemLnx->Core.enmType)
2041	{
2042	case RTR0MEMOBJTYPE_CONT:
2043	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
2044
2045	case RTR0MEMOBJTYPE_PHYS:
2046	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
2047
2048	/* the parent knows */
2049	case RTR0MEMOBJTYPE_MAPPING:
2050	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
2051
2052	/* cPages > 0 */
2053	case RTR0MEMOBJTYPE_LOW:
2054	case RTR0MEMOBJTYPE_LOCK:
2055	case RTR0MEMOBJTYPE_PHYS_NC:
2056	case RTR0MEMOBJTYPE_PAGE:
2057	case RTR0MEMOBJTYPE_LARGE_PAGE:
2058	default:
2059	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
2060	RT_FALL_THROUGH();
2061
2062	case RTR0MEMOBJTYPE_RES_VIRT:
2063	return NIL_RTHCPHYS;
2064	}
2065	}
2066

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 92418

Download in other formats: