memobj-r0drv-linux.c@ 99286

Last change on this file since 99286 was 98867, checked in by vboxsync, 22 months ago
Linux: vboxdrv: Introduce initial support for kernel 6.3.x series, bugref:10381.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Rev Revision`
File size: 70.9 KB

Line
1	/* $Id: memobj-r0drv-linux.c 98867 2023-03-07 17:18:31Z vboxsync $ */
2	/** @file
3	* IPRT - Ring-0 Memory Objects, Linux.
4	*/
5
6	/*
7	* Copyright (C) 2006-2023 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* The contents of this file may alternatively be used under the terms
26	* of the Common Development and Distribution License Version 1.0
27	* (CDDL), a copy of it is provided in the "COPYING.CDDL" file included
28	* in the VirtualBox distribution, in which case the provisions of the
29	* CDDL are applicable instead of those of the GPL.
30	*
31	* You may elect to license modified versions of this file under the
32	* terms and conditions of either the GPL or the CDDL or both.
33	*
34	* SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0
35	*/
36
37
38	/*********************************************************************************************************************************
39	* Header Files *
40	*********************************************************************************************************************************/
41	#include "the-linux-kernel.h"
42
43	#include <iprt/memobj.h>
44	#include <iprt/assert.h>
45	#include <iprt/err.h>
46	#include <iprt/log.h>
47	#include <iprt/mem.h>
48	#include <iprt/process.h>
49	#include <iprt/string.h>
50	#include "internal/memobj.h"
51	#include "internal/iprt.h"
52
53
54	/*********************************************************************************************************************************
55	* Defined Constants And Macros *
56	*********************************************************************************************************************************/
57	/* early 2.6 kernels */
58	#ifndef PAGE_SHARED_EXEC
59	# define PAGE_SHARED_EXEC PAGE_SHARED
60	#endif
61	#ifndef PAGE_READONLY_EXEC
62	# define PAGE_READONLY_EXEC PAGE_READONLY
63	#endif
64
65	/** @def IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
66	* Whether we use alloc_vm_area (3.2+) for executable memory.
67	* This is a must for 5.8+, but we enable it all the way back to 3.2.x for
68	* better W^R compliance (fExecutable flag). */
69	#if RTLNX_VER_RANGE(3,2,0, 5,10,0) \|\| defined(DOXYGEN_RUNNING)
70	# define IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
71	#endif
72	/** @def IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
73	* alloc_vm_area was removed with 5.10 so we have to resort to a different way
74	* to allocate executable memory.
75	* It would be possible to remove IPRT_USE_ALLOC_VM_AREA_FOR_EXEC and use
76	* this path execlusively for 3.2+ but no time to test it really works on every
77	* supported kernel, so better play safe for now.
78	*/
79	#if RTLNX_VER_MIN(5,10,0) \|\| defined(DOXYGEN_RUNNING)
80	# define IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
81	#endif
82
83	/*
84	* 2.6.29+ kernels don't work with remap_pfn_range() anymore because
85	* track_pfn_vma_new() is apparently not defined for non-RAM pages.
86	* It should be safe to use vm_insert_page() older kernels as well.
87	*/
88	#if RTLNX_VER_MIN(2,6,23)
89	# define VBOX_USE_INSERT_PAGE
90	#endif
91	#if defined(CONFIG_X86_PAE) \
92	&& ( defined(HAVE_26_STYLE_REMAP_PAGE_RANGE) \
93	\|\| RTLNX_VER_RANGE(2,6,0, 2,6,11) )
94	# define VBOX_USE_PAE_HACK
95	#endif
96
97	/* gfp_t was introduced in 2.6.14, define it for earlier. */
98	#if RTLNX_VER_MAX(2,6,14)
99	# define gfp_t unsigned
100	#endif
101
102	/*
103	* Wrappers around mmap_lock/mmap_sem difference.
104	*/
105	#if RTLNX_VER_MIN(5,8,0)
106	# define LNX_MM_DOWN_READ(a_pMm) down_read(&(a_pMm)->mmap_lock)
107	# define LNX_MM_UP_READ(a_pMm) up_read(&(a_pMm)->mmap_lock)
108	# define LNX_MM_DOWN_WRITE(a_pMm) down_write(&(a_pMm)->mmap_lock)
109	# define LNX_MM_UP_WRITE(a_pMm) up_write(&(a_pMm)->mmap_lock)
110	#else
111	# define LNX_MM_DOWN_READ(a_pMm) down_read(&(a_pMm)->mmap_sem)
112	# define LNX_MM_UP_READ(a_pMm) up_read(&(a_pMm)->mmap_sem)
113	# define LNX_MM_DOWN_WRITE(a_pMm) down_write(&(a_pMm)->mmap_sem)
114	# define LNX_MM_UP_WRITE(a_pMm) up_write(&(a_pMm)->mmap_sem)
115	#endif
116
117
118	/*********************************************************************************************************************************
119	* Structures and Typedefs *
120	*********************************************************************************************************************************/
121	/**
122	* The Linux version of the memory object structure.
123	*/
124	typedef struct RTR0MEMOBJLNX
125	{
126	/** The core structure. */
127	RTR0MEMOBJINTERNAL Core;
128	/** Set if the allocation is contiguous.
129	* This means it has to be given back as one chunk. */
130	bool fContiguous;
131	/** Set if executable allocation. */
132	bool fExecutable;
133	/** Set if we've vmap'ed the memory into ring-0. */
134	bool fMappedToRing0;
135	/** This is non-zero if large page allocation. */
136	uint8_t cLargePageOrder;
137	#ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
138	/** Return from alloc_vm_area() that we now need to use for executable
139	* memory. */
140	struct vm_struct *pArea;
141	/** PTE array that goes along with pArea (must be freed). */
142	pte_t **papPtesForArea;
143	#endif
144	/** The pages in the apPages array. */
145	size_t cPages;
146	/** Array of struct page pointers. (variable size) */
147	struct page *apPages[1];
148	} RTR0MEMOBJLNX;
149	/** Pointer to the linux memory object. */
150	typedef RTR0MEMOBJLNX *PRTR0MEMOBJLNX;
151
152
153	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx);
154
155
156	/**
157	* Helper that converts from a RTR0PROCESS handle to a linux task.
158	*
159	* @returns The corresponding Linux task.
160	* @param R0Process IPRT ring-0 process handle.
161	*/
162	static struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
163	{
164	/** @todo fix rtR0ProcessToLinuxTask!! */
165	/** @todo many (all?) callers currently assume that we return 'current'! */
166	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
167	}
168
169
170	/**
171	* Compute order. Some functions allocate 2^order pages.
172	*
173	* @returns order.
174	* @param cPages Number of pages.
175	*/
176	static int rtR0MemObjLinuxOrder(size_t cPages)
177	{
178	int iOrder;
179	size_t cTmp;
180
181	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
182	;
183	if (cPages & ~((size_t)1 << iOrder))
184	++iOrder;
185
186	return iOrder;
187	}
188
189
190	/**
191	* Converts from RTMEM_PROT_* to Linux PAGE_*.
192	*
193	* @returns Linux page protection constant.
194	* @param fProt The IPRT protection mask.
195	* @param fKernel Whether it applies to kernel or user space.
196	*/
197	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
198	{
199	switch (fProt)
200	{
201	default:
202	AssertMsgFailed(("%#x %d\n", fProt, fKernel)); RT_FALL_THRU();
203	case RTMEM_PROT_NONE:
204	return PAGE_NONE;
205
206	case RTMEM_PROT_READ:
207	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
208
209	case RTMEM_PROT_WRITE:
210	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
211	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
212
213	case RTMEM_PROT_EXEC:
214	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
215	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
216	if (fKernel)
217	{
218	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
219	pgprot_val(fPg) &= ~_PAGE_RW;
220	return fPg;
221	}
222	return PAGE_READONLY_EXEC;
223	#else
224	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
225	#endif
226
227	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
228	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
229	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
230	}
231	}
232
233
234	/**
235	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
236	* an empty user space mapping.
237	*
238	* We acquire the mmap_sem/mmap_lock of the task!
239	*
240	* @returns Pointer to the mapping.
241	* (void *)-1 on failure.
242	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
243	* @param cb The size of the mapping.
244	* @param uAlignment The alignment of the mapping.
245	* @param pTask The Linux task to create this mapping in.
246	* @param fProt The RTMEM_PROT_* mask.
247	*/
248	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
249	{
250	unsigned fLnxProt;
251	unsigned long ulAddr;
252
253	Assert(pTask == current); /* do_mmap */
254	RT_NOREF_PV(pTask);
255
256	/*
257	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
258	*/
259	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
260	if (fProt == RTMEM_PROT_NONE)
261	fLnxProt = PROT_NONE;
262	else
263	{
264	fLnxProt = 0;
265	if (fProt & RTMEM_PROT_READ)
266	fLnxProt \|= PROT_READ;
267	if (fProt & RTMEM_PROT_WRITE)
268	fLnxProt \|= PROT_WRITE;
269	if (fProt & RTMEM_PROT_EXEC)
270	fLnxProt \|= PROT_EXEC;
271	}
272
273	if (R3PtrFixed != (RTR3PTR)-1)
274	{
275	#if RTLNX_VER_MIN(3,5,0)
276	ulAddr = vm_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
277	#else
278	LNX_MM_DOWN_WRITE(pTask->mm);
279	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
280	LNX_MM_UP_WRITE(pTask->mm);
281	#endif
282	}
283	else
284	{
285	#if RTLNX_VER_MIN(3,5,0)
286	ulAddr = vm_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
287	#else
288	LNX_MM_DOWN_WRITE(pTask->mm);
289	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
290	LNX_MM_UP_WRITE(pTask->mm);
291	#endif
292	if ( !(ulAddr & ~PAGE_MASK)
293	&& (ulAddr & (uAlignment - 1)))
294	{
295	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
296	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
297	* ourselves) and further by there begin two mmap strategies (top / bottom). */
298	/* For now, just ignore uAlignment requirements... */
299	}
300	}
301
302
303	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
304	return (void *)-1;
305	return (void *)ulAddr;
306	}
307
308
309	/**
310	* Worker that destroys a user space mapping.
311	* Undoes what rtR0MemObjLinuxDoMmap did.
312	*
313	* We acquire the mmap_sem/mmap_lock of the task!
314	*
315	* @param pv The ring-3 mapping.
316	* @param cb The size of the mapping.
317	* @param pTask The Linux task to destroy this mapping in.
318	*/
319	static void rtR0MemObjLinuxDoMunmap(void pv, size_t cb, struct task_struct pTask)
320	{
321	#if RTLNX_VER_MIN(3,5,0)
322	Assert(pTask == current); RT_NOREF_PV(pTask);
323	vm_munmap((unsigned long)pv, cb);
324	#elif defined(USE_RHEL4_MUNMAP)
325	LNX_MM_DOWN_WRITE(pTask->mm);
326	do_munmap(pTask->mm, (unsigned long)pv, cb, 0); /* should it be 1 or 0? */
327	LNX_MM_UP_WRITE(pTask->mm);
328	#else
329	LNX_MM_DOWN_WRITE(pTask->mm);
330	do_munmap(pTask->mm, (unsigned long)pv, cb);
331	LNX_MM_UP_WRITE(pTask->mm);
332	#endif
333	}
334
335
336	/**
337	* Internal worker that allocates physical pages and creates the memory object for them.
338	*
339	* @returns IPRT status code.
340	* @param ppMemLnx Where to store the memory object pointer.
341	* @param enmType The object type.
342	* @param cb The number of bytes to allocate.
343	* @param uAlignment The alignment of the physical memory.
344	* Only valid if fContiguous == true, ignored otherwise.
345	* @param fFlagsLnx The page allocation flags (GPFs).
346	* @param fContiguous Whether the allocation must be contiguous.
347	* @param fExecutable Whether the memory must be executable.
348	* @param rcNoMem What to return when we're out of pages.
349	* @param pszTag Allocation tag used for statistics and such.
350	*/
351	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb,
352	size_t uAlignment, gfp_t fFlagsLnx, bool fContiguous, bool fExecutable, int rcNoMem,
353	const char *pszTag)
354	{
355	size_t iPage;
356	size_t const cPages = cb >> PAGE_SHIFT;
357	struct page *paPages;
358
359	/*
360	* Allocate a memory object structure that's large enough to contain
361	* the page pointer array.
362	*/
363	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), enmType,
364	NULL, cb, pszTag);
365	if (!pMemLnx)
366	return VERR_NO_MEMORY;
367	pMemLnx->Core.fFlags \|= RTR0MEMOBJ_FLAGS_UNINITIALIZED_AT_ALLOC;
368	pMemLnx->cPages = cPages;
369
370	if (cPages > 255)
371	{
372	# ifdef __GFP_REPEAT
373	/* Try hard to allocate the memory, but the allocation attempt might fail. */
374	fFlagsLnx \|= __GFP_REPEAT;
375	# endif
376	# ifdef __GFP_NOMEMALLOC
377	/* Introduced with Linux 2.6.12: Don't use emergency reserves */
378	fFlagsLnx \|= __GFP_NOMEMALLOC;
379	# endif
380	}
381
382	/*
383	* Allocate the pages.
384	* For small allocations we'll try contiguous first and then fall back on page by page.
385	*/
386	#if RTLNX_VER_MIN(2,4,22)
387	if ( fContiguous
388	\|\| cb <= PAGE_SIZE * 2)
389	{
390	# ifdef VBOX_USE_INSERT_PAGE
391	paPages = alloc_pages(fFlagsLnx \| __GFP_COMP \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
392	# else
393	paPages = alloc_pages(fFlagsLnx \| __GFP_NOWARN, rtR0MemObjLinuxOrder(cPages));
394	# endif
395	if (paPages)
396	{
397	fContiguous = true;
398	for (iPage = 0; iPage < cPages; iPage++)
399	pMemLnx->apPages[iPage] = &paPages[iPage];
400	}
401	else if (fContiguous)
402	{
403	rtR0MemObjDelete(&pMemLnx->Core);
404	return rcNoMem;
405	}
406	}
407
408	if (!fContiguous)
409	{
410	/** @todo Try use alloc_pages_bulk_array when available, it should be faster
411	* than a alloc_page loop. Put it in #ifdefs similar to
412	* IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC. */
413	for (iPage = 0; iPage < cPages; iPage++)
414	{
415	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx \| __GFP_NOWARN);
416	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
417	{
418	while (iPage-- > 0)
419	__free_page(pMemLnx->apPages[iPage]);
420	rtR0MemObjDelete(&pMemLnx->Core);
421	return rcNoMem;
422	}
423	}
424	}
425
426	#else /* < 2.4.22 */
427	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
428	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cPages));
429	if (!paPages)
430	{
431	rtR0MemObjDelete(&pMemLnx->Core);
432	return rcNoMem;
433	}
434	for (iPage = 0; iPage < cPages; iPage++)
435	{
436	pMemLnx->apPages[iPage] = &paPages[iPage];
437	if (fExecutable)
438	MY_SET_PAGES_EXEC(pMemLnx->apPages[iPage], 1);
439	if (PageHighMem(pMemLnx->apPages[iPage]))
440	BUG();
441	}
442
443	fContiguous = true;
444	#endif /* < 2.4.22 */
445	pMemLnx->fContiguous = fContiguous;
446	pMemLnx->fExecutable = fExecutable;
447
448	#if RTLNX_VER_MAX(4,5,0)
449	/*
450	* Reserve the pages.
451	*
452	* Linux >= 4.5 with CONFIG_DEBUG_VM panics when setting PG_reserved on compound
453	* pages. According to Michal Hocko this shouldn't be necessary anyway because
454	* as pages which are not on the LRU list are never evictable.
455	*/
456	for (iPage = 0; iPage < cPages; iPage++)
457	SetPageReserved(pMemLnx->apPages[iPage]);
458	#endif
459
460	/*
461	* Note that the physical address of memory allocated with alloc_pages(flags, order)
462	* is always 2^(PAGE_SHIFT+order)-aligned.
463	*/
464	if ( fContiguous
465	&& uAlignment > PAGE_SIZE)
466	{
467	/*
468	* Check for alignment constraints. The physical address of memory allocated with
469	* alloc_pages(flags, order) is always 2^(PAGE_SHIFT+order)-aligned.
470	*/
471	if (RT_UNLIKELY(page_to_phys(pMemLnx->apPages[0]) & (uAlignment - 1)))
472	{
473	/*
474	* This should never happen!
475	*/
476	printk("rtR0MemObjLinuxAllocPages(cb=0x%lx, uAlignment=0x%lx): alloc_pages(..., %d) returned physical memory at 0x%lx!\n",
477	(unsigned long)cb, (unsigned long)uAlignment, rtR0MemObjLinuxOrder(cPages), (unsigned long)page_to_phys(pMemLnx->apPages[0]));
478	rtR0MemObjLinuxFreePages(pMemLnx);
479	return rcNoMem;
480	}
481	}
482
483	*ppMemLnx = pMemLnx;
484	return VINF_SUCCESS;
485	}
486
487
488	/**
489	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
490	*
491	* This method does NOT free the object.
492	*
493	* @param pMemLnx The object which physical pages should be freed.
494	*/
495	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
496	{
497	size_t iPage = pMemLnx->cPages;
498	if (iPage > 0)
499	{
500	/*
501	* Restore the page flags.
502	*/
503	while (iPage-- > 0)
504	{
505	#if RTLNX_VER_MAX(4,5,0)
506	/* See SetPageReserved() in rtR0MemObjLinuxAllocPages() */
507	ClearPageReserved(pMemLnx->apPages[iPage]);
508	#endif
509	#if RTLNX_VER_MAX(2,4,22)
510	if (pMemLnx->fExecutable)
511	MY_SET_PAGES_NOEXEC(pMemLnx->apPages[iPage], 1);
512	#endif
513	}
514
515	/*
516	* Free the pages.
517	*/
518	#if RTLNX_VER_MIN(2,4,22)
519	if (!pMemLnx->fContiguous)
520	{
521	iPage = pMemLnx->cPages;
522	while (iPage-- > 0)
523	__free_page(pMemLnx->apPages[iPage]);
524	}
525	else
526	#endif
527	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
528
529	pMemLnx->cPages = 0;
530	}
531	}
532
533
534	#ifdef IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC
535	/**
536	* User data passed to the apply_to_page_range() callback.
537	*/
538	typedef struct LNXAPPLYPGRANGE
539	{
540	/** Pointer to the memory object. */
541	PRTR0MEMOBJLNX pMemLnx;
542	/** The page protection flags to apply. */
543	pgprot_t fPg;
544	} LNXAPPLYPGRANGE;
545	/** Pointer to the user data. */
546	typedef LNXAPPLYPGRANGE *PLNXAPPLYPGRANGE;
547	/** Pointer to the const user data. */
548	typedef const LNXAPPLYPGRANGE *PCLNXAPPLYPGRANGE;
549
550	/**
551	* Callback called in apply_to_page_range().
552	*
553	* @returns Linux status code.
554	* @param pPte Pointer to the page table entry for the given address.
555	* @param uAddr The address to apply the new protection to.
556	* @param pvUser The opaque user data.
557	*/
558	static int rtR0MemObjLinuxApplyPageRange(pte_t pPte, unsigned long uAddr, void pvUser)
559	{
560	PCLNXAPPLYPGRANGE pArgs = (PCLNXAPPLYPGRANGE)pvUser;
561	PRTR0MEMOBJLNX pMemLnx = pArgs->pMemLnx;
562	size_t idxPg = (uAddr - (unsigned long)pMemLnx->Core.pv) >> PAGE_SHIFT;
563
564	set_pte(pPte, mk_pte(pMemLnx->apPages[idxPg], pArgs->fPg));
565	return 0;
566	}
567	#endif
568
569
570	/**
571	* Maps the allocation into ring-0.
572	*
573	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
574	*
575	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
576	* space, so we'll use that mapping if possible. If execute access is required, we'll
577	* play safe and do our own mapping.
578	*
579	* @returns IPRT status code.
580	* @param pMemLnx The linux memory object to map.
581	* @param fExecutable Whether execute access is required.
582	*/
583	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
584	{
585	int rc = VINF_SUCCESS;
586
587	/*
588	* Choose mapping strategy.
589	*/
590	bool fMustMap = fExecutable
591	\|\| !pMemLnx->fContiguous;
592	if (!fMustMap)
593	{
594	size_t iPage = pMemLnx->cPages;
595	while (iPage-- > 0)
596	if (PageHighMem(pMemLnx->apPages[iPage]))
597	{
598	fMustMap = true;
599	break;
600	}
601	}
602
603	Assert(!pMemLnx->Core.pv);
604	Assert(!pMemLnx->fMappedToRing0);
605
606	if (fMustMap)
607	{
608	/*
609	* Use vmap - 2.4.22 and later.
610	*/
611	#if RTLNX_VER_MIN(2,4,22)
612	pgprot_t fPg;
613	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
614	# ifdef _PAGE_NX
615	if (!fExecutable)
616	pgprot_val(fPg) \|= _PAGE_NX;
617	# endif
618
619	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
620	if (fExecutable)
621	{
622	# if RTLNX_VER_MIN(3,2,51)
623	pte_t papPtes = (pte_t )kmalloc_array(pMemLnx->cPages, sizeof(papPtes[0]), GFP_KERNEL);
624	# else
625	pte_t papPtes = (pte_t )kmalloc(pMemLnx->cPages * sizeof(papPtes[0]), GFP_KERNEL);
626	# endif
627	if (papPtes)
628	{
629	pMemLnx->pArea = alloc_vm_area(pMemLnx->Core.cb, papPtes); /* Note! pArea->nr_pages is not set. */
630	if (pMemLnx->pArea)
631	{
632	size_t i;
633	Assert(pMemLnx->pArea->size >= pMemLnx->Core.cb); /* Note! includes guard page. */
634	Assert(pMemLnx->pArea->addr);
635	# ifdef _PAGE_NX
636	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
637	# endif
638	pMemLnx->papPtesForArea = papPtes;
639	for (i = 0; i < pMemLnx->cPages; i++)
640	*papPtes[i] = mk_pte(pMemLnx->apPages[i], fPg);
641	pMemLnx->Core.pv = pMemLnx->pArea->addr;
642	pMemLnx->fMappedToRing0 = true;
643	}
644	else
645	{
646	kfree(papPtes);
647	rc = VERR_MAP_FAILED;
648	}
649	}
650	else
651	rc = VERR_MAP_FAILED;
652	}
653	else
654	# endif
655	{
656	# if defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
657	if (fExecutable)
658	pgprot_val(fPg) \|= _PAGE_NX; /* Uses RTR0MemObjProtect to clear NX when memory ready, W^X fashion. */
659	# endif
660
661	# ifdef VM_MAP
662	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
663	# else
664	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
665	# endif
666	if (pMemLnx->Core.pv)
667	pMemLnx->fMappedToRing0 = true;
668	else
669	rc = VERR_MAP_FAILED;
670	}
671	#else /* < 2.4.22 */
672	rc = VERR_NOT_SUPPORTED;
673	#endif
674	}
675	else
676	{
677	/*
678	* Use the kernel RAM mapping.
679	*/
680	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
681	Assert(pMemLnx->Core.pv);
682	}
683
684	return rc;
685	}
686
687
688	/**
689	* Undoes what rtR0MemObjLinuxVMap() did.
690	*
691	* @param pMemLnx The linux memory object.
692	*/
693	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
694	{
695	#if RTLNX_VER_MIN(2,4,22)
696	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
697	if (pMemLnx->pArea)
698	{
699	# if 0
700	pte_t **papPtes = pMemLnx->papPtesForArea;
701	size_t i;
702	for (i = 0; i < pMemLnx->cPages; i++)
703	*papPtes[i] = 0;
704	# endif
705	free_vm_area(pMemLnx->pArea);
706	kfree(pMemLnx->papPtesForArea);
707	pMemLnx->pArea = NULL;
708	pMemLnx->papPtesForArea = NULL;
709	}
710	else
711	# endif
712	if (pMemLnx->fMappedToRing0)
713	{
714	Assert(pMemLnx->Core.pv);
715	vunmap(pMemLnx->Core.pv);
716	pMemLnx->fMappedToRing0 = false;
717	}
718	#else /* < 2.4.22 */
719	Assert(!pMemLnx->fMappedToRing0);
720	#endif
721	pMemLnx->Core.pv = NULL;
722	}
723
724
725	DECLHIDDEN(int) rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
726	{
727	IPRT_LINUX_SAVE_EFL_AC();
728	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
729
730	/*
731	* Release any memory that we've allocated or locked.
732	*/
733	switch (pMemLnx->Core.enmType)
734	{
735	case RTR0MEMOBJTYPE_PAGE:
736	case RTR0MEMOBJTYPE_LOW:
737	case RTR0MEMOBJTYPE_CONT:
738	case RTR0MEMOBJTYPE_PHYS:
739	case RTR0MEMOBJTYPE_PHYS_NC:
740	rtR0MemObjLinuxVUnmap(pMemLnx);
741	rtR0MemObjLinuxFreePages(pMemLnx);
742	break;
743
744	case RTR0MEMOBJTYPE_LARGE_PAGE:
745	{
746	uint32_t const cLargePages = pMemLnx->Core.cb >> (pMemLnx->cLargePageOrder + PAGE_SHIFT);
747	uint32_t iLargePage;
748	for (iLargePage = 0; iLargePage < cLargePages; iLargePage++)
749	__free_pages(pMemLnx->apPages[iLargePage << pMemLnx->cLargePageOrder], pMemLnx->cLargePageOrder);
750	pMemLnx->cPages = 0;
751
752	#ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
753	Assert(!pMemLnx->pArea);
754	Assert(!pMemLnx->papPtesForArea);
755	#endif
756	break;
757	}
758
759	case RTR0MEMOBJTYPE_LOCK:
760	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
761	{
762	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
763	size_t iPage;
764	Assert(pTask);
765	if (pTask && pTask->mm)
766	LNX_MM_DOWN_READ(pTask->mm);
767
768	iPage = pMemLnx->cPages;
769	while (iPage-- > 0)
770	{
771	if (!PageReserved(pMemLnx->apPages[iPage]))
772	SetPageDirty(pMemLnx->apPages[iPage]);
773	#if RTLNX_VER_MIN(4,6,0)
774	put_page(pMemLnx->apPages[iPage]);
775	#else
776	page_cache_release(pMemLnx->apPages[iPage]);
777	#endif
778	}
779
780	if (pTask && pTask->mm)
781	LNX_MM_UP_READ(pTask->mm);
782	}
783	/* else: kernel memory - nothing to do here. */
784	break;
785
786	case RTR0MEMOBJTYPE_RES_VIRT:
787	Assert(pMemLnx->Core.pv);
788	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
789	{
790	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
791	Assert(pTask);
792	if (pTask && pTask->mm)
793	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
794	}
795	else
796	{
797	vunmap(pMemLnx->Core.pv);
798
799	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
800	__free_page(pMemLnx->apPages[0]);
801	pMemLnx->apPages[0] = NULL;
802	pMemLnx->cPages = 0;
803	}
804	pMemLnx->Core.pv = NULL;
805	break;
806
807	case RTR0MEMOBJTYPE_MAPPING:
808	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
809	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
810	{
811	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
812	Assert(pTask);
813	if (pTask && pTask->mm)
814	rtR0MemObjLinuxDoMunmap(pMemLnx->Core.pv, pMemLnx->Core.cb, pTask);
815	}
816	else
817	vunmap(pMemLnx->Core.pv);
818	pMemLnx->Core.pv = NULL;
819	break;
820
821	default:
822	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
823	return VERR_INTERNAL_ERROR;
824	}
825	IPRT_LINUX_RESTORE_EFL_ONLY_AC();
826	return VINF_SUCCESS;
827	}
828
829
830	DECLHIDDEN(int) rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
831	{
832	IPRT_LINUX_SAVE_EFL_AC();
833	PRTR0MEMOBJLNX pMemLnx;
834	int rc;
835
836	#if RTLNX_VER_MIN(2,4,22)
837	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_HIGHUSER,
838	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, pszTag);
839	#else
840	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, PAGE_SIZE, GFP_USER,
841	false /* non-contiguous */, fExecutable, VERR_NO_MEMORY, pszTag);
842	#endif
843	if (RT_SUCCESS(rc))
844	{
845	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
846	if (RT_SUCCESS(rc))
847	{
848	*ppMem = &pMemLnx->Core;
849	IPRT_LINUX_RESTORE_EFL_AC();
850	return rc;
851	}
852
853	rtR0MemObjLinuxFreePages(pMemLnx);
854	rtR0MemObjDelete(&pMemLnx->Core);
855	}
856
857	IPRT_LINUX_RESTORE_EFL_AC();
858	return rc;
859	}
860
861
862	DECLHIDDEN(int) rtR0MemObjNativeAllocLarge(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, size_t cbLargePage, uint32_t fFlags,
863	const char *pszTag)
864	{
865	#ifdef GFP_TRANSHUGE
866	/*
867	* Allocate a memory object structure that's large enough to contain
868	* the page pointer array.
869	*/
870	# ifdef __GFP_MOVABLE
871	unsigned const fGfp = (GFP_TRANSHUGE \| __GFP_ZERO) & ~__GFP_MOVABLE;
872	# else
873	unsigned const fGfp = (GFP_TRANSHUGE \| __GFP_ZERO);
874	# endif
875	size_t const cPagesPerLarge = cbLargePage >> PAGE_SHIFT;
876	unsigned const cLargePageOrder = rtR0MemObjLinuxOrder(cPagesPerLarge);
877	size_t const cLargePages = cb >> (cLargePageOrder + PAGE_SHIFT);
878	size_t const cPages = cb >> PAGE_SHIFT;
879	PRTR0MEMOBJLNX pMemLnx;
880
881	Assert(RT_BIT_64(cLargePageOrder + PAGE_SHIFT) == cbLargePage);
882	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]),
883	RTR0MEMOBJTYPE_LARGE_PAGE, NULL, cb, pszTag);
884	if (pMemLnx)
885	{
886	size_t iLargePage;
887
888	pMemLnx->Core.fFlags \|= RTR0MEMOBJ_FLAGS_ZERO_AT_ALLOC;
889	pMemLnx->cLargePageOrder = cLargePageOrder;
890	pMemLnx->cPages = cPages;
891
892	/*
893	* Allocate the requested number of large pages.
894	*/
895	for (iLargePage = 0; iLargePage < cLargePages; iLargePage++)
896	{
897	struct page *paPages = alloc_pages(fGfp, cLargePageOrder);
898	if (paPages)
899	{
900	size_t const iPageBase = iLargePage << cLargePageOrder;
901	size_t iPage = cPagesPerLarge;
902	while (iPage-- > 0)
903	pMemLnx->apPages[iPageBase + iPage] = &paPages[iPage];
904	}
905	else
906	{
907	/*Log(("rtR0MemObjNativeAllocLarge: cb=%#zx cPages=%#zx cLargePages=%#zx cLargePageOrder=%u cPagesPerLarge=%#zx iLargePage=%#zx -> failed!\n",
908	cb, cPages, cLargePages, cLargePageOrder, cPagesPerLarge, iLargePage, paPages));*/
909	while (iLargePage-- > 0)
910	__free_pages(pMemLnx->apPages[iLargePage << (cLargePageOrder - PAGE_SHIFT)], cLargePageOrder);
911	rtR0MemObjDelete(&pMemLnx->Core);
912	return VERR_NO_MEMORY;
913	}
914	}
915	*ppMem = &pMemLnx->Core;
916	return VINF_SUCCESS;
917	}
918	return VERR_NO_MEMORY;
919
920	#else
921	/*
922	* We don't call rtR0MemObjFallbackAllocLarge here as it can be a really
923	* bad idea to trigger the swap daemon and whatnot. So, just fail.
924	*/
925	RT_NOREF(ppMem, cb, cbLargePage, fFlags, pszTag);
926	return VERR_NOT_SUPPORTED;
927	#endif
928	}
929
930
931	DECLHIDDEN(int) rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
932	{
933	IPRT_LINUX_SAVE_EFL_AC();
934	PRTR0MEMOBJLNX pMemLnx;
935	int rc;
936
937	/* Try to avoid GFP_DMA. GFM_DMA32 was introduced with Linux 2.6.15. */
938	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
939	/* ZONE_DMA32: 0-4GB */
940	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA32,
941	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
942	if (RT_FAILURE(rc))
943	#endif
944	#ifdef RT_ARCH_AMD64
945	/* ZONE_DMA: 0-16MB */
946	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_DMA,
947	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
948	#else
949	# ifdef CONFIG_X86_PAE
950	# endif
951	/* ZONE_NORMAL: 0-896MB */
952	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, PAGE_SIZE, GFP_USER,
953	false /* non-contiguous */, fExecutable, VERR_NO_LOW_MEMORY, pszTag);
954	#endif
955	if (RT_SUCCESS(rc))
956	{
957	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
958	if (RT_SUCCESS(rc))
959	{
960	*ppMem = &pMemLnx->Core;
961	IPRT_LINUX_RESTORE_EFL_AC();
962	return rc;
963	}
964
965	rtR0MemObjLinuxFreePages(pMemLnx);
966	rtR0MemObjDelete(&pMemLnx->Core);
967	}
968
969	IPRT_LINUX_RESTORE_EFL_AC();
970	return rc;
971	}
972
973
974	DECLHIDDEN(int) rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable, const char *pszTag)
975	{
976	IPRT_LINUX_SAVE_EFL_AC();
977	PRTR0MEMOBJLNX pMemLnx;
978	int rc;
979
980	#if (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_X86_PAE)) && defined(GFP_DMA32)
981	/* ZONE_DMA32: 0-4GB */
982	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA32,
983	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
984	if (RT_FAILURE(rc))
985	#endif
986	#ifdef RT_ARCH_AMD64
987	/* ZONE_DMA: 0-16MB */
988	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_DMA,
989	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
990	#else
991	/* ZONE_NORMAL (32-bit hosts): 0-896MB */
992	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, PAGE_SIZE, GFP_USER,
993	true /* contiguous */, fExecutable, VERR_NO_CONT_MEMORY, pszTag);
994	#endif
995	if (RT_SUCCESS(rc))
996	{
997	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
998	if (RT_SUCCESS(rc))
999	{
1000	#if defined(RT_STRICT) && (defined(RT_ARCH_AMD64) \|\| defined(CONFIG_HIGHMEM64G))
1001	size_t iPage = pMemLnx->cPages;
1002	while (iPage-- > 0)
1003	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
1004	#endif
1005	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
1006	*ppMem = &pMemLnx->Core;
1007	IPRT_LINUX_RESTORE_EFL_AC();
1008	return rc;
1009	}
1010
1011	rtR0MemObjLinuxFreePages(pMemLnx);
1012	rtR0MemObjDelete(&pMemLnx->Core);
1013	}
1014
1015	IPRT_LINUX_RESTORE_EFL_AC();
1016	return rc;
1017	}
1018
1019
1020	/**
1021	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
1022	*
1023	* @returns IPRT status code.
1024	* @param ppMemLnx Where to
1025	* @param enmType The object type.
1026	* @param cb The size of the allocation.
1027	* @param uAlignment The alignment of the physical memory.
1028	* Only valid for fContiguous == true, ignored otherwise.
1029	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
1030	* @param pszTag Allocation tag used for statistics and such.
1031	* @param fGfp The Linux GFP flags to use for the allocation.
1032	*/
1033	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
1034	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag, gfp_t fGfp)
1035	{
1036	PRTR0MEMOBJLNX pMemLnx;
1037	int rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, uAlignment, fGfp,
1038	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */,
1039	false /fExecutable/, VERR_NO_PHYS_MEMORY, pszTag);
1040	if (RT_FAILURE(rc))
1041	return rc;
1042
1043	/*
1044	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
1045	*/
1046	if (PhysHighest != NIL_RTHCPHYS)
1047	{
1048	size_t iPage = pMemLnx->cPages;
1049	while (iPage-- > 0)
1050	if (page_to_phys(pMemLnx->apPages[iPage]) > PhysHighest)
1051	{
1052	rtR0MemObjLinuxFreePages(pMemLnx);
1053	rtR0MemObjDelete(&pMemLnx->Core);
1054	return VERR_NO_MEMORY;
1055	}
1056	}
1057
1058	/*
1059	* Complete the object.
1060	*/
1061	if (enmType == RTR0MEMOBJTYPE_PHYS)
1062	{
1063	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
1064	pMemLnx->Core.u.Phys.fAllocated = true;
1065	}
1066	*ppMem = &pMemLnx->Core;
1067	return rc;
1068	}
1069
1070
1071	/**
1072	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
1073	*
1074	* @returns IPRT status code.
1075	* @param ppMem Where to store the memory object pointer on success.
1076	* @param enmType The object type.
1077	* @param cb The size of the allocation.
1078	* @param uAlignment The alignment of the physical memory.
1079	* Only valid for enmType == RTR0MEMOBJTYPE_PHYS, ignored otherwise.
1080	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
1081	* @param pszTag Allocation tag used for statistics and such.
1082	*/
1083	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType,
1084	size_t cb, size_t uAlignment, RTHCPHYS PhysHighest, const char *pszTag)
1085	{
1086	int rc;
1087	IPRT_LINUX_SAVE_EFL_AC();
1088
1089	/*
1090	* There are two clear cases and that's the <=16MB and anything-goes ones.
1091	* When the physical address limit is somewhere in-between those two we'll
1092	* just have to try, starting with HIGHUSER and working our way thru the
1093	* different types, hoping we'll get lucky.
1094	*
1095	* We should probably move this physical address restriction logic up to
1096	* the page alloc function as it would be more efficient there. But since
1097	* we don't expect this to be a performance issue just yet it can wait.
1098	*/
1099	if (PhysHighest == NIL_RTHCPHYS)
1100	/* ZONE_HIGHMEM: the whole physical memory */
1101	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1102	else if (PhysHighest <= _1M * 16)
1103	/* ZONE_DMA: 0-16MB */
1104	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1105	else
1106	{
1107	rc = VERR_NO_MEMORY;
1108	if (RT_FAILURE(rc))
1109	/* ZONE_HIGHMEM: the whole physical memory */
1110	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_HIGHUSER);
1111	if (RT_FAILURE(rc))
1112	/* ZONE_NORMAL: 0-896MB */
1113	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_USER);
1114	#ifdef GFP_DMA32
1115	if (RT_FAILURE(rc))
1116	/* ZONE_DMA32: 0-4GB */
1117	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA32);
1118	#endif
1119	if (RT_FAILURE(rc))
1120	/* ZONE_DMA: 0-16MB */
1121	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, uAlignment, PhysHighest, pszTag, GFP_DMA);
1122	}
1123	IPRT_LINUX_RESTORE_EFL_AC();
1124	return rc;
1125	}
1126
1127
1128	/**
1129	* Translates a kernel virtual address to a linux page structure by walking the
1130	* page tables.
1131	*
1132	* @note We do assume that the page tables will not change as we are walking
1133	* them. This assumption is rather forced by the fact that I could not
1134	* immediately see any way of preventing this from happening. So, we
1135	* take some extra care when accessing them.
1136	*
1137	* Because of this, we don't want to use this function on memory where
1138	* attribute changes to nearby pages is likely to cause large pages to
1139	* be used or split up. So, don't use this for the linear mapping of
1140	* physical memory.
1141	*
1142	* @returns Pointer to the page structur or NULL if it could not be found.
1143	* @param pv The kernel virtual address.
1144	*/
1145	RTDECL(struct page ) rtR0MemObjLinuxVirtToPage(void pv)
1146	{
1147	unsigned long ulAddr = (unsigned long)pv;
1148	unsigned long pfn;
1149	struct page *pPage;
1150	pte_t *pEntry;
1151	union
1152	{
1153	pgd_t Global;
1154	#if RTLNX_VER_MIN(4,12,0)
1155	p4d_t Four;
1156	#endif
1157	#if RTLNX_VER_MIN(2,6,11)
1158	pud_t Upper;
1159	#endif
1160	pmd_t Middle;
1161	pte_t Entry;
1162	} u;
1163
1164	/* Should this happen in a situation this code will be called in? And if
1165	* so, can it change under our feet? See also
1166	* "Documentation/vm/active_mm.txt" in the kernel sources. */
1167	if (RT_UNLIKELY(!current->active_mm))
1168	return NULL;
1169	u.Global = *pgd_offset(current->active_mm, ulAddr);
1170	if (RT_UNLIKELY(pgd_none(u.Global)))
1171	return NULL;
1172	#if RTLNX_VER_MIN(2,6,11)
1173	# if RTLNX_VER_MIN(4,12,0)
1174	u.Four = *p4d_offset(&u.Global, ulAddr);
1175	if (RT_UNLIKELY(p4d_none(u.Four)))
1176	return NULL;
1177	if (p4d_large(u.Four))
1178	{
1179	pPage = p4d_page(u.Four);
1180	AssertReturn(pPage, NULL);
1181	pfn = page_to_pfn(pPage); /* doing the safe way... */
1182	AssertCompile(P4D_SHIFT - PAGE_SHIFT < 31);
1183	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (P4D_SHIFT - PAGE_SHIFT)) - 1);
1184	return pfn_to_page(pfn);
1185	}
1186	u.Upper = *pud_offset(&u.Four, ulAddr);
1187	# else /* < 4.12 */
1188	u.Upper = *pud_offset(&u.Global, ulAddr);
1189	# endif /* < 4.12 */
1190	if (RT_UNLIKELY(pud_none(u.Upper)))
1191	return NULL;
1192	# if RTLNX_VER_MIN(2,6,25)
1193	if (pud_large(u.Upper))
1194	{
1195	pPage = pud_page(u.Upper);
1196	AssertReturn(pPage, NULL);
1197	pfn = page_to_pfn(pPage); /* doing the safe way... */
1198	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PUD_SHIFT - PAGE_SHIFT)) - 1);
1199	return pfn_to_page(pfn);
1200	}
1201	# endif
1202	u.Middle = *pmd_offset(&u.Upper, ulAddr);
1203	#else /* < 2.6.11 */
1204	u.Middle = *pmd_offset(&u.Global, ulAddr);
1205	#endif /* < 2.6.11 */
1206	if (RT_UNLIKELY(pmd_none(u.Middle)))
1207	return NULL;
1208	#if RTLNX_VER_MIN(2,6,0)
1209	if (pmd_large(u.Middle))
1210	{
1211	pPage = pmd_page(u.Middle);
1212	AssertReturn(pPage, NULL);
1213	pfn = page_to_pfn(pPage); /* doing the safe way... */
1214	pfn += (ulAddr >> PAGE_SHIFT) & ((UINT32_C(1) << (PMD_SHIFT - PAGE_SHIFT)) - 1);
1215	return pfn_to_page(pfn);
1216	}
1217	#endif
1218
1219	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map) /* As usual, RHEL 3 had pte_offset_map earlier. */
1220	pEntry = pte_offset_map(&u.Middle, ulAddr);
1221	#else
1222	pEntry = pte_offset(&u.Middle, ulAddr);
1223	#endif
1224	if (RT_UNLIKELY(!pEntry))
1225	return NULL;
1226	u.Entry = *pEntry;
1227	#if RTLNX_VER_MIN(2,5,5) \|\| defined(pte_offset_map)
1228	pte_unmap(pEntry);
1229	#endif
1230
1231	if (RT_UNLIKELY(!pte_present(u.Entry)))
1232	return NULL;
1233	return pte_page(u.Entry);
1234	}
1235	RT_EXPORT_SYMBOL(rtR0MemObjLinuxVirtToPage);
1236
1237
1238	DECLHIDDEN(int) rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, size_t uAlignment,
1239	const char *pszTag)
1240	{
1241	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, uAlignment, PhysHighest, pszTag);
1242	}
1243
1244
1245	DECLHIDDEN(int) rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest, const char *pszTag)
1246	{
1247	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PAGE_SIZE, PhysHighest, pszTag);
1248	}
1249
1250
1251	DECLHIDDEN(int) rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb, uint32_t uCachePolicy,
1252	const char *pszTag)
1253	{
1254	IPRT_LINUX_SAVE_EFL_AC();
1255
1256	/*
1257	* All we need to do here is to validate that we can use
1258	* ioremap on the specified address (32/64-bit dma_addr_t).
1259	*/
1260	PRTR0MEMOBJLNX pMemLnx;
1261	dma_addr_t PhysAddr = Phys;
1262	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
1263
1264	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb, pszTag);
1265	if (!pMemLnx)
1266	{
1267	IPRT_LINUX_RESTORE_EFL_AC();
1268	return VERR_NO_MEMORY;
1269	}
1270
1271	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
1272	pMemLnx->Core.u.Phys.fAllocated = false;
1273	pMemLnx->Core.u.Phys.uCachePolicy = uCachePolicy;
1274	Assert(!pMemLnx->cPages);
1275	*ppMem = &pMemLnx->Core;
1276	IPRT_LINUX_RESTORE_EFL_AC();
1277	return VINF_SUCCESS;
1278	}
1279
1280	/* openSUSE Leap 42.3 detection :-/ */
1281	#if RTLNX_VER_RANGE(4,4,0, 4,6,0) && defined(FAULT_FLAG_REMOTE)
1282	# define GET_USER_PAGES_API KERNEL_VERSION(4, 10, 0) /* no typo! */
1283	#else
1284	# define GET_USER_PAGES_API LINUX_VERSION_CODE
1285	#endif
1286
1287	DECLHIDDEN(int) rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, uint32_t fAccess,
1288	RTR0PROCESS R0Process, const char *pszTag)
1289	{
1290	IPRT_LINUX_SAVE_EFL_AC();
1291	const int cPages = cb >> PAGE_SHIFT;
1292	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1293	struct vm_area_struct **papVMAs;
1294	PRTR0MEMOBJLNX pMemLnx;
1295	int rc = VERR_NO_MEMORY;
1296	int const fWrite = fAccess & RTMEM_PROT_WRITE ? 1 : 0;
1297
1298	/*
1299	* Check for valid task and size overflows.
1300	*/
1301	if (!pTask)
1302	return VERR_NOT_SUPPORTED;
1303	if (((size_t)cPages << PAGE_SHIFT) != cb)
1304	return VERR_OUT_OF_RANGE;
1305
1306	/*
1307	* Allocate the memory object and a temporary buffer for the VMAs.
1308	*/
1309	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1310	(void *)R3Ptr, cb, pszTag);
1311	if (!pMemLnx)
1312	{
1313	IPRT_LINUX_RESTORE_EFL_AC();
1314	return VERR_NO_MEMORY;
1315	}
1316
1317	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
1318	if (papVMAs)
1319	{
1320	LNX_MM_DOWN_READ(pTask->mm);
1321
1322	/*
1323	* Get user pages.
1324	*/
1325	/** @todo r=bird: Should we not force read access too? */
1326	#if GET_USER_PAGES_API >= KERNEL_VERSION(4, 6, 0)
1327	if (R0Process == RTR0ProcHandleSelf())
1328	rc = get_user_pages(R3Ptr, /* Where from. */
1329	cPages, /* How many pages. */
1330	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1331	fWrite ? FOLL_WRITE \| /* Write to memory. */
1332	FOLL_FORCE /* force write access. */
1333	: 0, /* Write to memory. */
1334	# else
1335	fWrite, /* Write to memory. */
1336	fWrite, /* force write access. */
1337	# endif
1338	&pMemLnx->apPages[0], /* Page array. */
1339	papVMAs); /* vmas */
1340	/*
1341	* Actually this should not happen at the moment as call this function
1342	* only for our own process.
1343	*/
1344	else
1345	rc = get_user_pages_remote(
1346	# if GET_USER_PAGES_API < KERNEL_VERSION(5, 9, 0)
1347	pTask, /* Task for fault accounting. */
1348	# endif
1349	pTask->mm, /* Whose pages. */
1350	R3Ptr, /* Where from. */
1351	cPages, /* How many pages. */
1352	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 9, 0)
1353	fWrite ? FOLL_WRITE \| /* Write to memory. */
1354	FOLL_FORCE /* force write access. */
1355	: 0, /* Write to memory. */
1356	# else
1357	fWrite, /* Write to memory. */
1358	fWrite, /* force write access. */
1359	# endif
1360	&pMemLnx->apPages[0], /* Page array. */
1361	papVMAs /* vmas */
1362	# if GET_USER_PAGES_API >= KERNEL_VERSION(4, 10, 0)
1363	, NULL /* locked */
1364	# endif
1365	);
1366	#else /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1367	rc = get_user_pages(pTask, /* Task for fault accounting. */
1368	pTask->mm, /* Whose pages. */
1369	R3Ptr, /* Where from. */
1370	cPages, /* How many pages. */
1371	/* The get_user_pages API change was back-ported to 4.4.168. */
1372	# if RTLNX_VER_RANGE(4,4,168, 4,5,0)
1373	fWrite ? FOLL_WRITE \| /* Write to memory. */
1374	FOLL_FORCE /* force write access. */
1375	: 0, /* Write to memory. */
1376	# else
1377	fWrite, /* Write to memory. */
1378	fWrite, /* force write access. */
1379	# endif
1380	&pMemLnx->apPages[0], /* Page array. */
1381	papVMAs); /* vmas */
1382	#endif /* GET_USER_PAGES_API < KERNEL_VERSION(4, 6, 0) */
1383	if (rc == cPages)
1384	{
1385	/*
1386	* Flush dcache (required?), protect against fork and _really_ pin the page
1387	* table entries. get_user_pages() will protect against swapping out the
1388	* pages but it will NOT protect against removing page table entries. This
1389	* can be achieved with
1390	* - using mlock / mmap(..., MAP_LOCKED, ...) from userland. This requires
1391	* an appropriate limit set up with setrlimit(..., RLIMIT_MEMLOCK, ...).
1392	* Usual Linux distributions support only a limited size of locked pages
1393	* (e.g. 32KB).
1394	* - setting the PageReserved bit (as we do in rtR0MemObjLinuxAllocPages()
1395	* or by
1396	* - setting the VM_LOCKED flag. This is the same as doing mlock() without
1397	* a range check.
1398	*/
1399	/** @todo The Linux fork() protection will require more work if this API
1400	* is to be used for anything but locking VM pages. */
1401	while (rc-- > 0)
1402	{
1403	flush_dcache_page(pMemLnx->apPages[rc]);
1404	#if RTLNX_VER_MIN(6,3,0)
1405	vm_flags_set(papVMAs[rc], VM_DONTCOPY \| VM_LOCKED);
1406	#else
1407	papVMAs[rc]->vm_flags \|= VM_DONTCOPY \| VM_LOCKED;
1408	#endif
1409	}
1410
1411	LNX_MM_UP_READ(pTask->mm);
1412
1413	RTMemFree(papVMAs);
1414
1415	pMemLnx->Core.u.Lock.R0Process = R0Process;
1416	pMemLnx->cPages = cPages;
1417	Assert(!pMemLnx->fMappedToRing0);
1418	*ppMem = &pMemLnx->Core;
1419
1420	IPRT_LINUX_RESTORE_EFL_AC();
1421	return VINF_SUCCESS;
1422	}
1423
1424	/*
1425	* Failed - we need to unlock any pages that we succeeded to lock.
1426	*/
1427	while (rc-- > 0)
1428	{
1429	if (!PageReserved(pMemLnx->apPages[rc]))
1430	SetPageDirty(pMemLnx->apPages[rc]);
1431	#if RTLNX_VER_MIN(4,6,0)
1432	put_page(pMemLnx->apPages[rc]);
1433	#else
1434	page_cache_release(pMemLnx->apPages[rc]);
1435	#endif
1436	}
1437
1438	LNX_MM_UP_READ(pTask->mm);
1439
1440	RTMemFree(papVMAs);
1441	rc = VERR_LOCK_FAILED;
1442	}
1443
1444	rtR0MemObjDelete(&pMemLnx->Core);
1445	IPRT_LINUX_RESTORE_EFL_AC();
1446	return rc;
1447	}
1448
1449
1450	DECLHIDDEN(int) rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void pv, size_t cb, uint32_t fAccess, const char pszTag)
1451	{
1452	IPRT_LINUX_SAVE_EFL_AC();
1453	void pvLast = (uint8_t )pv + cb - 1;
1454	size_t const cPages = cb >> PAGE_SHIFT;
1455	PRTR0MEMOBJLNX pMemLnx;
1456	bool fLinearMapping;
1457	int rc;
1458	uint8_t *pbPage;
1459	size_t iPage;
1460	NOREF(fAccess);
1461
1462	if ( !RTR0MemKernelIsValidAddr(pv)
1463	\|\| !RTR0MemKernelIsValidAddr(pv + cb))
1464	return VERR_INVALID_PARAMETER;
1465
1466	/*
1467	* The lower part of the kernel memory has a linear mapping between
1468	* physical and virtual addresses. So we take a short cut here. This is
1469	* assumed to be the cleanest way to handle those addresses (and the code
1470	* is well tested, though the test for determining it is not very nice).
1471	* If we ever decide it isn't we can still remove it.
1472	*/
1473	#if 0
1474	fLinearMapping = (unsigned long)pvLast < VMALLOC_START;
1475	#else
1476	fLinearMapping = (unsigned long)pv >= (unsigned long)__va(0)
1477	&& (unsigned long)pvLast < (unsigned long)high_memory;
1478	#endif
1479
1480	/*
1481	* Allocate the memory object.
1482	*/
1483	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_UOFFSETOF_DYN(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK,
1484	pv, cb, pszTag);
1485	if (!pMemLnx)
1486	{
1487	IPRT_LINUX_RESTORE_EFL_AC();
1488	return VERR_NO_MEMORY;
1489	}
1490
1491	/*
1492	* Gather the pages.
1493	* We ASSUME all kernel pages are non-swappable and non-movable.
1494	*/
1495	rc = VINF_SUCCESS;
1496	pbPage = (uint8_t *)pvLast;
1497	iPage = cPages;
1498	if (!fLinearMapping)
1499	{
1500	while (iPage-- > 0)
1501	{
1502	struct page *pPage = rtR0MemObjLinuxVirtToPage(pbPage);
1503	if (RT_UNLIKELY(!pPage))
1504	{
1505	rc = VERR_LOCK_FAILED;
1506	break;
1507	}
1508	pMemLnx->apPages[iPage] = pPage;
1509	pbPage -= PAGE_SIZE;
1510	}
1511	}
1512	else
1513	{
1514	while (iPage-- > 0)
1515	{
1516	pMemLnx->apPages[iPage] = virt_to_page(pbPage);
1517	pbPage -= PAGE_SIZE;
1518	}
1519	}
1520	if (RT_SUCCESS(rc))
1521	{
1522	/*
1523	* Complete the memory object and return.
1524	*/
1525	pMemLnx->Core.u.Lock.R0Process = NIL_RTR0PROCESS;
1526	pMemLnx->cPages = cPages;
1527	Assert(!pMemLnx->fMappedToRing0);
1528	*ppMem = &pMemLnx->Core;
1529
1530	IPRT_LINUX_RESTORE_EFL_AC();
1531	return VINF_SUCCESS;
1532	}
1533
1534	rtR0MemObjDelete(&pMemLnx->Core);
1535	IPRT_LINUX_RESTORE_EFL_AC();
1536	return rc;
1537	}
1538
1539
1540	DECLHIDDEN(int) rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment,
1541	const char *pszTag)
1542	{
1543	#if RTLNX_VER_MIN(2,4,22)
1544	IPRT_LINUX_SAVE_EFL_AC();
1545	const size_t cPages = cb >> PAGE_SHIFT;
1546	struct page *pDummyPage;
1547	struct page **papPages;
1548
1549	/* check for unsupported stuff. */
1550	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1551	if (uAlignment > PAGE_SIZE)
1552	return VERR_NOT_SUPPORTED;
1553
1554	/*
1555	* Allocate a dummy page and create a page pointer array for vmap such that
1556	* the dummy page is mapped all over the reserved area.
1557	*/
1558	pDummyPage = alloc_page(GFP_HIGHUSER \| __GFP_NOWARN);
1559	if (pDummyPage)
1560	{
1561	papPages = RTMemAlloc(sizeof(papPages) cPages);
1562	if (papPages)
1563	{
1564	void *pv;
1565	size_t iPage = cPages;
1566	while (iPage-- > 0)
1567	papPages[iPage] = pDummyPage;
1568	# ifdef VM_MAP
1569	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
1570	# else
1571	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
1572	# endif
1573	RTMemFree(papPages);
1574	if (pv)
1575	{
1576	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1577	if (pMemLnx)
1578	{
1579	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
1580	pMemLnx->cPages = 1;
1581	pMemLnx->apPages[0] = pDummyPage;
1582	*ppMem = &pMemLnx->Core;
1583	IPRT_LINUX_RESTORE_EFL_AC();
1584	return VINF_SUCCESS;
1585	}
1586	vunmap(pv);
1587	}
1588	}
1589	__free_page(pDummyPage);
1590	}
1591	IPRT_LINUX_RESTORE_EFL_AC();
1592	return VERR_NO_MEMORY;
1593
1594	#else /* < 2.4.22 */
1595	/*
1596	* Could probably use ioremap here, but the caller is in a better position than us
1597	* to select some safe physical memory.
1598	*/
1599	return VERR_NOT_SUPPORTED;
1600	#endif
1601	}
1602
1603
1604	DECLHIDDEN(int) rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment,
1605	RTR0PROCESS R0Process, const char *pszTag)
1606	{
1607	IPRT_LINUX_SAVE_EFL_AC();
1608	PRTR0MEMOBJLNX pMemLnx;
1609	void *pv;
1610	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1611	if (!pTask)
1612	return VERR_NOT_SUPPORTED;
1613
1614	/*
1615	* Check that the specified alignment is supported.
1616	*/
1617	if (uAlignment > PAGE_SIZE)
1618	return VERR_NOT_SUPPORTED;
1619
1620	/*
1621	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
1622	*/
1623	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
1624	if (pv == (void *)-1)
1625	{
1626	IPRT_LINUX_RESTORE_EFL_AC();
1627	return VERR_NO_MEMORY;
1628	}
1629
1630	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb, pszTag);
1631	if (!pMemLnx)
1632	{
1633	rtR0MemObjLinuxDoMunmap(pv, cb, pTask);
1634	IPRT_LINUX_RESTORE_EFL_AC();
1635	return VERR_NO_MEMORY;
1636	}
1637
1638	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
1639	*ppMem = &pMemLnx->Core;
1640	IPRT_LINUX_RESTORE_EFL_AC();
1641	return VINF_SUCCESS;
1642	}
1643
1644
1645	DECLHIDDEN(int) rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment,
1646	unsigned fProt, size_t offSub, size_t cbSub, const char *pszTag)
1647	{
1648	int rc = VERR_NO_MEMORY;
1649	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1650	PRTR0MEMOBJLNX pMemLnx;
1651	IPRT_LINUX_SAVE_EFL_AC();
1652
1653	/* Fail if requested to do something we can't. */
1654	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
1655	if (uAlignment > PAGE_SIZE)
1656	return VERR_NOT_SUPPORTED;
1657
1658	/*
1659	* Create the IPRT memory object.
1660	*/
1661	if (!cbSub)
1662	cbSub = pMemLnxToMap->Core.cb - offSub;
1663	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1664	if (pMemLnx)
1665	{
1666	if (pMemLnxToMap->cPages)
1667	{
1668	#if RTLNX_VER_MIN(2,4,22)
1669	/*
1670	* Use vmap - 2.4.22 and later.
1671	*/
1672	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
1673	/** @todo We don't really care too much for EXEC here... 5.8 always adds NX. */
1674	Assert(((offSub + cbSub) >> PAGE_SHIFT) <= pMemLnxToMap->cPages);
1675	# ifdef VM_MAP
1676	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_MAP, fPg);
1677	# else
1678	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[offSub >> PAGE_SHIFT], cbSub >> PAGE_SHIFT, VM_ALLOC, fPg);
1679	# endif
1680	if (pMemLnx->Core.pv)
1681	{
1682	pMemLnx->fMappedToRing0 = true;
1683	rc = VINF_SUCCESS;
1684	}
1685	else
1686	rc = VERR_MAP_FAILED;
1687
1688	#else /* < 2.4.22 */
1689	/*
1690	* Only option here is to share mappings if possible and forget about fProt.
1691	*/
1692	if (rtR0MemObjIsRing3(pMemToMap))
1693	rc = VERR_NOT_SUPPORTED;
1694	else
1695	{
1696	rc = VINF_SUCCESS;
1697	if (!pMemLnxToMap->Core.pv)
1698	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
1699	if (RT_SUCCESS(rc))
1700	{
1701	Assert(pMemLnxToMap->Core.pv);
1702	pMemLnx->Core.pv = (uint8_t *)pMemLnxToMap->Core.pv + offSub;
1703	}
1704	}
1705	#endif
1706	}
1707	else
1708	{
1709	/*
1710	* MMIO / physical memory.
1711	*/
1712	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
1713	#if RTLNX_VER_MIN(2,6,25)
1714	/*
1715	* ioremap() defaults to no caching since the 2.6 kernels.
1716	* ioremap_nocache() has been removed finally in 5.6-rc1.
1717	*/
1718	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1719	? ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1720	: ioremap_cache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1721	#else /* KERNEL_VERSION < 2.6.25 */
1722	pMemLnx->Core.pv = pMemLnxToMap->Core.u.Phys.uCachePolicy == RTMEM_CACHE_POLICY_MMIO
1723	? ioremap_nocache(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub)
1724	: ioremap(pMemLnxToMap->Core.u.Phys.PhysBase + offSub, cbSub);
1725	#endif /* KERNEL_VERSION < 2.6.25 */
1726	if (pMemLnx->Core.pv)
1727	{
1728	/** @todo fix protection. */
1729	rc = VINF_SUCCESS;
1730	}
1731	}
1732	if (RT_SUCCESS(rc))
1733	{
1734	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
1735	*ppMem = &pMemLnx->Core;
1736	IPRT_LINUX_RESTORE_EFL_AC();
1737	return VINF_SUCCESS;
1738	}
1739	rtR0MemObjDelete(&pMemLnx->Core);
1740	}
1741
1742	IPRT_LINUX_RESTORE_EFL_AC();
1743	return rc;
1744	}
1745
1746
1747	#ifdef VBOX_USE_PAE_HACK
1748	/**
1749	* Replace the PFN of a PTE with the address of the actual page.
1750	*
1751	* The caller maps a reserved dummy page at the address with the desired access
1752	* and flags.
1753	*
1754	* This hack is required for older Linux kernels which don't provide
1755	* remap_pfn_range().
1756	*
1757	* @returns 0 on success, -ENOMEM on failure.
1758	* @param mm The memory context.
1759	* @param ulAddr The mapping address.
1760	* @param Phys The physical address of the page to map.
1761	*/
1762	static int rtR0MemObjLinuxFixPte(struct mm_struct *mm, unsigned long ulAddr, RTHCPHYS Phys)
1763	{
1764	int rc = -ENOMEM;
1765	pgd_t *pgd;
1766
1767	spin_lock(&mm->page_table_lock);
1768
1769	pgd = pgd_offset(mm, ulAddr);
1770	if (!pgd_none(pgd) && !pgd_bad(pgd))
1771	{
1772	pmd_t *pmd = pmd_offset(pgd, ulAddr);
1773	if (!pmd_none(*pmd))
1774	{
1775	pte_t *ptep = pte_offset_map(pmd, ulAddr);
1776	if (ptep)
1777	{
1778	pte_t pte = *ptep;
1779	pte.pte_high &= 0xfff00000;
1780	pte.pte_high \|= ((Phys >> 32) & 0x000fffff);
1781	pte.pte_low &= 0x00000fff;
1782	pte.pte_low \|= (Phys & 0xfffff000);
1783	set_pte(ptep, pte);
1784	pte_unmap(ptep);
1785	rc = 0;
1786	}
1787	}
1788	}
1789
1790	spin_unlock(&mm->page_table_lock);
1791	return rc;
1792	}
1793	#endif /* VBOX_USE_PAE_HACK */
1794
1795
1796	DECLHIDDEN(int) rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment,
1797	unsigned fProt, RTR0PROCESS R0Process, size_t offSub, size_t cbSub, const char *pszTag)
1798	{
1799	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1800	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1801	int rc = VERR_NO_MEMORY;
1802	PRTR0MEMOBJLNX pMemLnx;
1803	#ifdef VBOX_USE_PAE_HACK
1804	struct page *pDummyPage;
1805	RTHCPHYS DummyPhys;
1806	#endif
1807	IPRT_LINUX_SAVE_EFL_AC();
1808
1809	/*
1810	* Check for restrictions.
1811	*/
1812	if (!pTask)
1813	return VERR_NOT_SUPPORTED;
1814	if (uAlignment > PAGE_SIZE)
1815	return VERR_NOT_SUPPORTED;
1816
1817	#ifdef VBOX_USE_PAE_HACK
1818	/*
1819	* Allocate a dummy page for use when mapping the memory.
1820	*/
1821	pDummyPage = alloc_page(GFP_USER \| __GFP_NOWARN);
1822	if (!pDummyPage)
1823	{
1824	IPRT_LINUX_RESTORE_EFL_AC();
1825	return VERR_NO_MEMORY;
1826	}
1827	SetPageReserved(pDummyPage);
1828	DummyPhys = page_to_phys(pDummyPage);
1829	#endif
1830
1831	/*
1832	* Create the IPRT memory object.
1833	*/
1834	Assert(!offSub \|\| cbSub);
1835	if (cbSub == 0)
1836	cbSub = pMemLnxToMap->Core.cb;
1837	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, cbSub, pszTag);
1838	if (pMemLnx)
1839	{
1840	/*
1841	* Allocate user space mapping.
1842	*/
1843	void *pv;
1844	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cbSub, uAlignment, pTask, fProt);
1845	if (pv != (void *)-1)
1846	{
1847	/*
1848	* Map page by page into the mmap area.
1849	* This is generic, paranoid and not very efficient.
1850	*/
1851	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1852	unsigned long ulAddrCur = (unsigned long)pv;
1853	const size_t cPages = (offSub + cbSub) >> PAGE_SHIFT;
1854	size_t iPage;
1855
1856	LNX_MM_DOWN_WRITE(pTask->mm);
1857
1858	rc = VINF_SUCCESS;
1859	if (pMemLnxToMap->cPages)
1860	{
1861	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1862	{
1863	#if RTLNX_VER_MAX(2,6,11)
1864	RTHCPHYS Phys = page_to_phys(pMemLnxToMap->apPages[iPage]);
1865	#endif
1866	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1867	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1868	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1869	#endif
1870	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1871	/* remap_page_range() limitation on x86 */
1872	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1873	#endif
1874
1875	#if defined(VBOX_USE_INSERT_PAGE) && RTLNX_VER_MIN(2,6,22)
1876	rc = vm_insert_page(vma, ulAddrCur, pMemLnxToMap->apPages[iPage]);
1877	/* Thes flags help making 100% sure some bad stuff wont happen (swap, core, ++).
1878	* See remap_pfn_range() in mm/memory.c */
1879
1880	#if RTLNX_VER_MIN(6,3,0)
1881	vm_flags_set(vma, VM_DONTEXPAND \| VM_DONTDUMP);
1882	#elif RTLNX_VER_MIN(3,7,0)
1883	vma->vm_flags \|= VM_DONTEXPAND \| VM_DONTDUMP;
1884	#else
1885	vma->vm_flags \|= VM_RESERVED;
1886	#endif
1887	#elif RTLNX_VER_MIN(2,6,11)
1888	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1889	#elif defined(VBOX_USE_PAE_HACK)
1890	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1891	if (!rc)
1892	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1893	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1894	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1895	#else /* 2.4 */
1896	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1897	#endif
1898	if (rc)
1899	{
1900	rc = VERR_NO_MEMORY;
1901	break;
1902	}
1903	}
1904	}
1905	else
1906	{
1907	RTHCPHYS Phys;
1908	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1909	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1910	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1911	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1912	else
1913	{
1914	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1915	Phys = NIL_RTHCPHYS;
1916	}
1917	if (Phys != NIL_RTHCPHYS)
1918	{
1919	for (iPage = offSub >> PAGE_SHIFT; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1920	{
1921	#if RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1922	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1923	AssertBreakStmt(vma, rc = VERR_INTERNAL_ERROR);
1924	#endif
1925	#if RTLNX_VER_MAX(2,6,0) && defined(RT_ARCH_X86)
1926	/* remap_page_range() limitation on x86 */
1927	AssertBreakStmt(Phys < _4G, rc = VERR_NO_MEMORY);
1928	#endif
1929
1930	#if RTLNX_VER_MIN(2,6,11)
1931	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1932	#elif defined(VBOX_USE_PAE_HACK)
1933	rc = remap_page_range(vma, ulAddrCur, DummyPhys, PAGE_SIZE, fPg);
1934	if (!rc)
1935	rc = rtR0MemObjLinuxFixPte(pTask->mm, ulAddrCur, Phys);
1936	#elif RTLNX_VER_MIN(2,6,0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1937	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1938	#else /* 2.4 */
1939	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1940	#endif
1941	if (rc)
1942	{
1943	rc = VERR_NO_MEMORY;
1944	break;
1945	}
1946	}
1947	}
1948	}
1949
1950	#ifdef CONFIG_NUMA_BALANCING
1951	# if RTLNX_VER_MAX(3,13,0) && RTLNX_RHEL_MAX(7,0)
1952	# define VBOX_NUMA_HACK_OLD
1953	# endif
1954	if (RT_SUCCESS(rc))
1955	{
1956	/** @todo Ugly hack! But right now we have no other means to
1957	* disable automatic NUMA page balancing. */
1958	# ifdef RT_OS_X86
1959	# ifdef VBOX_NUMA_HACK_OLD
1960	pTask->mm->numa_next_reset = jiffies + 0x7fffffffUL;
1961	# endif
1962	pTask->mm->numa_next_scan = jiffies + 0x7fffffffUL;
1963	# else
1964	# ifdef VBOX_NUMA_HACK_OLD
1965	pTask->mm->numa_next_reset = jiffies + 0x7fffffffffffffffUL;
1966	# endif
1967	pTask->mm->numa_next_scan = jiffies + 0x7fffffffffffffffUL;
1968	# endif
1969	}
1970	#endif /* CONFIG_NUMA_BALANCING */
1971
1972	LNX_MM_UP_WRITE(pTask->mm);
1973
1974	if (RT_SUCCESS(rc))
1975	{
1976	#ifdef VBOX_USE_PAE_HACK
1977	__free_page(pDummyPage);
1978	#endif
1979	pMemLnx->Core.pv = pv;
1980	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1981	*ppMem = &pMemLnx->Core;
1982	IPRT_LINUX_RESTORE_EFL_AC();
1983	return VINF_SUCCESS;
1984	}
1985
1986	/*
1987	* Bail out.
1988	*/
1989	rtR0MemObjLinuxDoMunmap(pv, cbSub, pTask);
1990	}
1991	rtR0MemObjDelete(&pMemLnx->Core);
1992	}
1993	#ifdef VBOX_USE_PAE_HACK
1994	__free_page(pDummyPage);
1995	#endif
1996
1997	IPRT_LINUX_RESTORE_EFL_AC();
1998	return rc;
1999	}
2000
2001
2002	DECLHIDDEN(int) rtR0MemObjNativeProtect(PRTR0MEMOBJINTERNAL pMem, size_t offSub, size_t cbSub, uint32_t fProt)
2003	{
2004	# ifdef IPRT_USE_ALLOC_VM_AREA_FOR_EXEC
2005	/*
2006	* Currently only supported when we've got addresses PTEs from the kernel.
2007	*/
2008	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
2009	if (pMemLnx->pArea && pMemLnx->papPtesForArea)
2010	{
2011	pgprot_t const fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
2012	size_t const cPages = (offSub + cbSub) >> PAGE_SHIFT;
2013	pte_t **papPtes = pMemLnx->papPtesForArea;
2014	size_t i;
2015
2016	for (i = offSub >> PAGE_SHIFT; i < cPages; i++)
2017	{
2018	set_pte(papPtes[i], mk_pte(pMemLnx->apPages[i], fPg));
2019	}
2020	preempt_disable();
2021	__flush_tlb_all();
2022	preempt_enable();
2023	return VINF_SUCCESS;
2024	}
2025	# elif defined(IPRT_USE_APPLY_TO_PAGE_RANGE_FOR_EXEC)
2026	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
2027	if ( pMemLnx->fExecutable
2028	&& pMemLnx->fMappedToRing0)
2029	{
2030	LNXAPPLYPGRANGE Args;
2031	Args.pMemLnx = pMemLnx;
2032	Args.fPg = rtR0MemObjLinuxConvertProt(fProt, true /fKernel/);
2033	int rcLnx = apply_to_page_range(current->active_mm, (unsigned long)pMemLnx->Core.pv + offSub, cbSub,
2034	rtR0MemObjLinuxApplyPageRange, (void *)&Args);
2035	if (rcLnx)
2036	return VERR_NOT_SUPPORTED;
2037
2038	return VINF_SUCCESS;
2039	}
2040	# endif
2041
2042	NOREF(pMem);
2043	NOREF(offSub);
2044	NOREF(cbSub);
2045	NOREF(fProt);
2046	return VERR_NOT_SUPPORTED;
2047	}
2048
2049
2050	DECLHIDDEN(RTHCPHYS) rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
2051	{
2052	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
2053
2054	if (pMemLnx->cPages)
2055	return page_to_phys(pMemLnx->apPages[iPage]);
2056
2057	switch (pMemLnx->Core.enmType)
2058	{
2059	case RTR0MEMOBJTYPE_CONT:
2060	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
2061
2062	case RTR0MEMOBJTYPE_PHYS:
2063	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
2064
2065	/* the parent knows */
2066	case RTR0MEMOBJTYPE_MAPPING:
2067	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
2068
2069	/* cPages > 0 */
2070	case RTR0MEMOBJTYPE_LOW:
2071	case RTR0MEMOBJTYPE_LOCK:
2072	case RTR0MEMOBJTYPE_PHYS_NC:
2073	case RTR0MEMOBJTYPE_PAGE:
2074	case RTR0MEMOBJTYPE_LARGE_PAGE:
2075	default:
2076	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
2077	RT_FALL_THROUGH();
2078
2079	case RTR0MEMOBJTYPE_RES_VIRT:
2080	return NIL_RTHCPHYS;
2081	}
2082	}
2083

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 99286

Download in other formats: