memobj-r0drv-linux.c@ 4160

Last change on this file since 4160 was 4160, checked in by vboxsync, 17 years ago
Some bug fixes and allocation fallbacks.
Property svn:keywords set to `Id Rev`
File size: 35.9 KB

Line
1	/* $Revision: 4160 $ */
2	/** @file
3	* innotek Portable Runtime - Ring-0 Memory Objects, Linux.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 innotek GmbH
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 as published by the Free Software Foundation,
13	* in version 2 as it comes in the "COPYING" file of the VirtualBox OSE
14	* distribution. VirtualBox OSE is distributed in the hope that it will
15	* be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*******************************************************************************
20	* Header Files *
21	*******************************************************************************/
22	#include "the-linux-kernel.h"
23
24	#include <iprt/memobj.h>
25	#include <iprt/alloc.h>
26	#include <iprt/assert.h>
27	#include <iprt/log.h>
28	#include <iprt/string.h>
29	#include <iprt/process.h>
30	#include "internal/memobj.h"
31
32
33	/*******************************************************************************
34	* Structures and Typedefs *
35	*******************************************************************************/
36	/**
37	* The Darwin version of the memory object structure.
38	*/
39	typedef struct RTR0MEMOBJLNX
40	{
41	/** The core structure. */
42	RTR0MEMOBJINTERNAL Core;
43	/** Set if the allocation is contiguous.
44	* This means it has to be given back as one chunk. */
45	bool fContiguous;
46	/** Set if we've vmap'ed thed memory into ring-0. */
47	bool fMappedToRing0;
48	/** The pages in the apPages array. */
49	size_t cPages;
50	/** Array of struct page pointers. (variable size) */
51	struct page *apPages[1];
52	} RTR0MEMOBJLNX, *PRTR0MEMOBJLNX;
53
54
55	/**
56	* Helper that converts from a RTR0PROCESS handle to a linux task.
57	*
58	* @returns The corresponding Linux task.
59	* @param R0Process IPRT ring-0 process handle.
60	*/
61	struct task_struct *rtR0ProcessToLinuxTask(RTR0PROCESS R0Process)
62	{
63	/** @todo fix rtR0ProcessToLinuxTask!! */
64	return R0Process == RTR0ProcHandleSelf() ? current : NULL;
65	}
66
67
68	/**
69	* Compute order. Some functions allocate 2^order pages.
70	*
71	* @returns order.
72	* @param cPages Number of pages.
73	*/
74	static int rtR0MemObjLinuxOrder(size_t cPages)
75	{
76	int iOrder;
77	size_t cTmp;
78
79	for (iOrder = 0, cTmp = cPages; cTmp >>= 1; ++iOrder)
80	;
81	if (cPages & ~((size_t)1 << iOrder))
82	++iOrder;
83
84	return iOrder;
85	}
86
87
88	/**
89	* Converts from RTMEM_PROT_* to Linux PAGE_*.
90	*
91	* @returns Linux page protection constant.
92	* @param fProt The IPRT protection mask.
93	* @param fKernel Whether it applies to kernel or user space.
94	*/
95	static pgprot_t rtR0MemObjLinuxConvertProt(unsigned fProt, bool fKernel)
96	{
97	switch (fProt)
98	{
99	default:
100	AssertMsgFailed(("%#x %d\n", fProt, fKernel));
101	case RTMEM_PROT_NONE:
102	return PAGE_NONE;
103
104	case RTMEM_PROT_READ:
105	return fKernel ? PAGE_KERNEL_RO : PAGE_READONLY;
106
107	case RTMEM_PROT_WRITE:
108	case RTMEM_PROT_WRITE \| RTMEM_PROT_READ:
109	return fKernel ? PAGE_KERNEL : PAGE_SHARED;
110
111	case RTMEM_PROT_EXEC:
112	case RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
113	#if defined(RT_ARCH_X86) \|\| defined(RT_ARCH_AMD64)
114	if (fKernel)
115	{
116	pgprot_t fPg = MY_PAGE_KERNEL_EXEC;
117	pgprot_val(fPg) &= ~_PAGE_RW;
118	return fPg;
119	}
120	return PAGE_READONLY_EXEC;
121	#else
122	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_READONLY_EXEC;
123	#endif
124
125	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC:
126	case RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_READ:
127	return fKernel ? MY_PAGE_KERNEL_EXEC : PAGE_SHARED_EXEC;
128	}
129	}
130
131
132	/**
133	* Internal worker that allocates physical pages and creates the memory object for them.
134	*
135	* @returns IPRT status code.
136	* @param ppMemLnx Where to store the memory object pointer.
137	* @param enmType The object type.
138	* @param cb The number of bytes to allocate.
139	* @param fFlagsLnx The page allocation flags (GPFs).
140	* @param fContiguous Whether the allocation must be contiguous.
141	*/
142	static int rtR0MemObjLinuxAllocPages(PRTR0MEMOBJLNX *ppMemLnx, RTR0MEMOBJTYPE enmType, size_t cb, unsigned fFlagsLnx, bool fContiguous)
143	{
144	size_t iPage;
145	size_t cPages = cb >> PAGE_SHIFT;
146	struct page *paPages;
147
148	/*
149	* Allocate a memory object structure that's large enough to contain
150	* the page pointer array.
151	*/
152	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), enmType, NULL, cb);
153	if (!pMemLnx)
154	return VERR_NO_MEMORY;
155	pMemLnx->cPages = cPages;
156
157	/*
158	* Allocate the pages.
159	* For small allocations we'll try contiguous first and then fall back on page by page.
160	*/
161	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
162	if ( fContiguous
163	\|\| cb <= PAGE_SIZE * 2)
164	{
165	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cb >> PAGE_SHIFT));
166	if (paPages)
167	{
168	fContiguous = true;
169	for (iPage = 0; iPage < cPages; iPage++)
170	pMemLnx->apPages[iPage] = &paPages[iPage];
171	}
172	else if (fContiguous)
173	{
174	rtR0MemObjDelete(&pMemLnx->Core);
175	return VERR_NO_MEMORY;
176	}
177	}
178
179	if (!fContiguous)
180	{
181	for (iPage = 0; iPage < cPages; iPage++)
182	{
183	pMemLnx->apPages[iPage] = alloc_page(fFlagsLnx);
184	if (RT_UNLIKELY(!pMemLnx->apPages[iPage]))
185	{
186	while (iPage-- > 0)
187	__free_page(pMemLnx->apPages[iPage]);
188	rtR0MemObjDelete(&pMemLnx->Core);
189	return VERR_NO_MEMORY;
190	}
191	}
192	}
193
194	#else /* < 2.4.22 */
195	/** @todo figure out why we didn't allocate page-by-page on 2.4.21 and older... */
196	paPages = alloc_pages(fFlagsLnx, rtR0MemObjLinuxOrder(cb >> PAGE_SHIFT));
197	if (!paPages)
198	{
199	rtR0MemObjDelete(&pMemLnx->Core);
200	return VERR_NO_MEMORY;
201	}
202	for (iPage = 0; iPage < cPages; iPage++)
203	{
204	pMemLnx->apPages[iPage] = &paPages[iPage];
205	if (pgprot_val(MY_PAGE_KERNEL_EXEC) != pgprot_val(PAGE_KERNEL))
206	MY_CHANGE_PAGE_ATTR(pMemLnx->apPages[iPage], 1, MY_PAGE_KERNEL_EXEC);
207	if (PageHighMem(pMemLnx->apPages[iPage]))
208	BUG();
209	}
210
211	fContiguous = true;
212	#endif /* < 2.4.22 */
213	pMemLnx->fContiguous = fContiguous;
214
215	/*
216	* Reserve the pages.
217	*/
218	for (iPage = 0; iPage < cPages; iPage++)
219	SetPageReserved(pMemLnx->apPages[iPage]);
220
221	*ppMemLnx = pMemLnx;
222	return VINF_SUCCESS;
223	}
224
225
226	/**
227	* Frees the physical pages allocated by the rtR0MemObjLinuxAllocPages() call.
228	*
229	* This method does NOT free the object.
230	*
231	* @param pMemLnx The object which physical pages should be freed.
232	*/
233	static void rtR0MemObjLinuxFreePages(PRTR0MEMOBJLNX pMemLnx)
234	{
235	size_t iPage = pMemLnx->cPages;
236	if (iPage > 0)
237	{
238	/*
239	* Restore the page flags.
240	*/
241	while (iPage-- > 0)
242	{
243	ClearPageReserved(pMemLnx->apPages[iPage]);
244	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
245	#else
246	if (pgprot_val(MY_PAGE_KERNEL_EXEC) != pgprot_val(PAGE_KERNEL))
247	MY_CHANGE_PAGE_ATTR(pMemLnx->apPages[iPage], 1, PAGE_KERNEL);
248	#endif
249	}
250
251	/*
252	* Free the pages.
253	*/
254	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
255	if (!pMemLnx->fContiguous)
256	{
257	iPage = pMemLnx->cPages;
258	while (iPage-- > 0)
259	__free_page(pMemLnx->apPages[iPage]);
260	}
261	else
262	#endif
263	__free_pages(pMemLnx->apPages[0], rtR0MemObjLinuxOrder(pMemLnx->cPages));
264
265	pMemLnx->cPages = 0;
266	}
267	}
268
269
270	/**
271	* Maps the allocation into ring-0.
272	*
273	* This will update the RTR0MEMOBJLNX::Core.pv and RTR0MEMOBJ::fMappedToRing0 members.
274	*
275	* Contiguous mappings that isn't in 'high' memory will already be mapped into kernel
276	* space, so we'll use that mapping if possible. If execute access is required, we'll
277	* play safe and do our own mapping.
278	*
279	* @returns IPRT status code.
280	* @param pMemLnx The linux memory object to map.
281	* @param fExecutable Whether execute access is required.
282	*/
283	static int rtR0MemObjLinuxVMap(PRTR0MEMOBJLNX pMemLnx, bool fExecutable)
284	{
285	int rc = VINF_SUCCESS;
286
287	/*
288	* Choose mapping strategy.
289	*/
290	bool fMustMap = fExecutable
291	\|\| !pMemLnx->fContiguous;
292	if (!fMustMap)
293	{
294	size_t iPage = pMemLnx->cPages;
295	while (iPage-- > 0)
296	if (PageHighMem(pMemLnx->apPages[iPage]))
297	{
298	fMustMap = true;
299	break;
300	}
301	}
302
303	Assert(!pMemLnx->Core.pv);
304	Assert(!pMemLnx->fMappedToRing0);
305
306	if (fMustMap)
307	{
308	/*
309	* Use vmap - 2.4.22 and later.
310	*/
311	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
312	pgprot_t fPg;
313	pgprot_val(fPg) = _PAGE_PRESENT \| _PAGE_RW;
314	# ifdef _PAGE_NX
315	if (!fExecutable)
316	pgprot_val(fPg) \|= _PAGE_NX;
317	# endif
318
319	# ifdef VM_MAP
320	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_MAP, fPg);
321	# else
322	pMemLnx->Core.pv = vmap(&pMemLnx->apPages[0], pMemLnx->cPages, VM_ALLOC, fPg);
323	# endif
324	if (pMemLnx->Core.pv)
325	pMemLnx->fMappedToRing0 = true;
326	else
327	rc = VERR_MAP_FAILED;
328	#else /* < 2.4.22 */
329	rc = VERR_NOT_SUPPORTED;
330	#endif
331	}
332	else
333	{
334	/*
335	* Use the kernel RAM mapping.
336	*/
337	pMemLnx->Core.pv = phys_to_virt(page_to_phys(pMemLnx->apPages[0]));
338	Assert(pMemLnx->Core.pv);
339	}
340
341	return rc;
342	}
343
344
345	/**
346	* Undos what rtR0MemObjLinuxVMap() did.
347	*
348	* @param pMemLnx The linux memory object.
349	*/
350	static void rtR0MemObjLinuxVUnmap(PRTR0MEMOBJLNX pMemLnx)
351	{
352	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
353	if (pMemLnx->fMappedToRing0)
354	{
355	Assert(pMemLnx->Core.pv);
356	vunmap(pMemLnx->Core.pv);
357	pMemLnx->fMappedToRing0 = false;
358	}
359	#else /* < 2.4.22 */
360	Assert(!pMemLnx->fMappedToRing0);
361	#endif
362	pMemLnx->Core.pv = NULL;
363	}
364
365
366	int rtR0MemObjNativeFree(RTR0MEMOBJ pMem)
367	{
368	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
369
370	/*
371	* Release any memory that we've allocated or locked.
372	*/
373	switch (pMemLnx->Core.enmType)
374	{
375	case RTR0MEMOBJTYPE_LOW:
376	case RTR0MEMOBJTYPE_PAGE:
377	case RTR0MEMOBJTYPE_CONT:
378	case RTR0MEMOBJTYPE_PHYS:
379	rtR0MemObjLinuxVUnmap(pMemLnx);
380	rtR0MemObjLinuxFreePages(pMemLnx);
381	break;
382
383	case RTR0MEMOBJTYPE_LOCK:
384	if (pMemLnx->Core.u.Lock.R0Process != NIL_RTR0PROCESS)
385	{
386	size_t iPage;
387	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
388	Assert(pTask);
389	if (pTask && pTask->mm)
390	down_read(&pTask->mm->mmap_sem);
391
392	iPage = pMemLnx->cPages;
393	while (iPage-- > 0)
394	{
395	if (!PageReserved(pMemLnx->apPages[iPage]))
396	SetPageDirty(pMemLnx->apPages[iPage]);
397	page_cache_release(pMemLnx->apPages[iPage]);
398	}
399
400	if (pTask && pTask->mm)
401	up_read(&pTask->mm->mmap_sem);
402	}
403	else
404	AssertFailed(); /* not implemented for R0 */
405	break;
406
407	case RTR0MEMOBJTYPE_RES_VIRT:
408	Assert(pMemLnx->Core.pv);
409	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
410	{
411	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
412	Assert(pTask);
413	if (pTask && pTask->mm)
414	{
415	down_write(&pTask->mm->mmap_sem);
416	MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
417	up_write(&pTask->mm->mmap_sem);
418	}
419	}
420	else
421	{
422	vunmap(pMemLnx->Core.pv);
423
424	Assert(pMemLnx->cPages == 1 && pMemLnx->apPages[0] != NULL);
425	__free_page(pMemLnx->apPages[0]);
426	pMemLnx->apPages[0] = NULL;
427	pMemLnx->cPages = 0;
428	}
429	pMemLnx->Core.pv = NULL;
430	break;
431
432	case RTR0MEMOBJTYPE_MAPPING:
433	Assert(pMemLnx->cPages == 0); Assert(pMemLnx->Core.pv);
434	if (pMemLnx->Core.u.ResVirt.R0Process != NIL_RTR0PROCESS)
435	{
436	struct task_struct *pTask = rtR0ProcessToLinuxTask(pMemLnx->Core.u.Lock.R0Process);
437	Assert(pTask);
438	if (pTask && pTask->mm)
439	{
440	down_write(&pTask->mm->mmap_sem);
441	MY_DO_MUNMAP(pTask->mm, (unsigned long)pMemLnx->Core.pv, pMemLnx->Core.cb);
442	up_write(&pTask->mm->mmap_sem);
443	}
444	}
445	else
446	vunmap(pMemLnx->Core.pv);
447	pMemLnx->Core.pv = NULL;
448	break;
449
450	default:
451	AssertMsgFailed(("enmType=%d\n", pMemLnx->Core.enmType));
452	return VERR_INTERNAL_ERROR;
453	}
454	return VINF_SUCCESS;
455	}
456
457
458	int rtR0MemObjNativeAllocPage(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
459	{
460	PRTR0MEMOBJLNX pMemLnx;
461	int rc;
462
463	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
464	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, GFP_HIGHUSER, false /* non-contiguous */);
465	#else
466	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_PAGE, cb, GFP_USER, false /* non-contiguous */);
467	#endif
468	if (RT_SUCCESS(rc))
469	{
470	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
471	if (RT_SUCCESS(rc))
472	{
473	*ppMem = &pMemLnx->Core;
474	return rc;
475	}
476
477	rtR0MemObjLinuxFreePages(pMemLnx);
478	rtR0MemObjDelete(&pMemLnx->Core);
479	}
480
481	return rc;
482	}
483
484
485	int rtR0MemObjNativeAllocLow(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
486	{
487	PRTR0MEMOBJLNX pMemLnx;
488	int rc;
489
490	#ifdef RT_ARCH_AMD64
491	# ifdef GFP_DMA32
492	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, GFP_DMA32, false /* non-contiguous */);
493	if (RT_FAILURE(rc))
494	# endif
495	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, GFP_DMA, false /* non-contiguous */);
496	#else
497	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_LOW, cb, GFP_USER, false /* non-contiguous */);
498	#endif
499	if (RT_SUCCESS(rc))
500	{
501	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
502	if (RT_SUCCESS(rc))
503	{
504	*ppMem = &pMemLnx->Core;
505	return rc;
506	}
507
508	rtR0MemObjLinuxFreePages(pMemLnx);
509	rtR0MemObjDelete(&pMemLnx->Core);
510	}
511
512	return rc;
513	}
514
515
516	int rtR0MemObjNativeAllocCont(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, bool fExecutable)
517	{
518	PRTR0MEMOBJLNX pMemLnx;
519	int rc;
520
521	#ifdef RT_ARCH_AMD64
522	# ifdef GFP_DMA32
523	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, GFP_DMA32, true /* contiguous */);
524	if (RT_FAILURE(rc))
525	# endif
526	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, GFP_DMA, true /* contiguous */);
527	#else
528	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, RTR0MEMOBJTYPE_CONT, cb, GFP_USER, true /* contiguous */);
529	#endif
530	if (RT_SUCCESS(rc))
531	{
532	rc = rtR0MemObjLinuxVMap(pMemLnx, fExecutable);
533	if (RT_SUCCESS(rc))
534	{
535	#ifdef RT_STRICT
536	size_t iPage = pMemLnx->cPages;
537	while (iPage-- > 0)
538	Assert(page_to_phys(pMemLnx->apPages[iPage]) < _4G);
539	#endif
540	pMemLnx->Core.u.Cont.Phys = page_to_phys(pMemLnx->apPages[0]);
541	*ppMem = &pMemLnx->Core;
542	return rc;
543	}
544
545	rtR0MemObjLinuxFreePages(pMemLnx);
546	rtR0MemObjDelete(&pMemLnx->Core);
547	}
548
549	return rc;
550	}
551
552
553	/**
554	* Worker for rtR0MemObjLinuxAllocPhysSub that tries one allocation strategy.
555	*
556	* @returns IPRT status.
557	* @param ppMemLnx Where to
558	* @param enmType The object type.
559	* @param cb The size of the allocation.
560	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
561	* @param fGfp The Linux GFP flags to use for the allocation.
562	*/
563	static int rtR0MemObjLinuxAllocPhysSub2(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType, size_t cb, RTHCPHYS PhysHighest, unsigned fGfp)
564	{
565	PRTR0MEMOBJLNX pMemLnx;
566	int rc;
567
568	rc = rtR0MemObjLinuxAllocPages(&pMemLnx, enmType, cb, fGfp,
569	enmType == RTR0MEMOBJTYPE_PHYS /* contiguous / non-contiguous */);
570	if (RT_FAILURE(rc))
571	return rc;
572
573	/*
574	* Check the addresses if necessary. (Can be optimized a bit for PHYS.)
575	*/
576	if (PhysHighest != NIL_RTHCPHYS)
577	{
578	size_t iPage = pMemLnx->cPages;
579	while (iPage-- > 0)
580	if (page_to_phys(pMemLnx->apPages[iPage]) >= PhysHighest)
581	{
582	rtR0MemObjLinuxFreePages(pMemLnx);
583	rtR0MemObjDelete(&pMemLnx->Core);
584	return VERR_NO_MEMORY;
585	}
586	}
587
588	/*
589	* Complete the object.
590	*/
591	if (enmType == RTR0MEMOBJTYPE_PHYS)
592	{
593	pMemLnx->Core.u.Phys.PhysBase = page_to_phys(pMemLnx->apPages[0]);
594	pMemLnx->Core.u.Phys.fAllocated = true;
595	}
596	*ppMem = &pMemLnx->Core;
597	return rc;
598	}
599
600
601	/**
602	* Worker for rtR0MemObjNativeAllocPhys and rtR0MemObjNativeAllocPhysNC.
603	*
604	* @returns IPRT status.
605	* @param ppMem Where to store the memory object pointer on success.
606	* @param enmType The object type.
607	* @param cb The size of the allocation.
608	* @param PhysHighest See rtR0MemObjNativeAllocPhys.
609	*/
610	static int rtR0MemObjLinuxAllocPhysSub(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJTYPE enmType, size_t cb, RTHCPHYS PhysHighest)
611	{
612	int rc;
613
614	/*
615	* There are two clear cases and that's the <=16MB and anything-goes ones.
616	* When the physical address limit is somewhere inbetween those two we'll
617	* just have to try, starting with HIGHUSER and working our way thru the
618	* different types, hoping we'll get lucky.
619	*
620	* We should probably move this physical address restriction logic up to
621	* the page alloc function as it would be more efficient there. But since
622	* we don't expect this to be a performance issue just yet it can wait.
623	*/
624	if (PhysHighest == NIL_RTHCPHYS)
625	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, PhysHighest, GFP_HIGHUSER);
626	else if (PhysHighest <= _1M * 16)
627	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, PhysHighest, GFP_DMA);
628	else
629	{
630	rc = VERR_NO_MEMORY;
631	if (RT_FAILURE(rc))
632	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, PhysHighest, GFP_HIGHUSER);
633	if (RT_FAILURE(rc))
634	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, PhysHighest, GFP_USER);
635	#ifdef GFP_DMA32
636	if (RT_FAILURE(rc))
637	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, PhysHighest, GFP_DMA32);
638	#endif
639	if (RT_FAILURE(rc))
640	rc = rtR0MemObjLinuxAllocPhysSub2(ppMem, enmType, cb, PhysHighest, GFP_DMA);
641	}
642	return rc;
643	}
644
645
646	int rtR0MemObjNativeAllocPhys(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
647	{
648	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS, cb, PhysHighest);
649	}
650
651
652	int rtR0MemObjNativeAllocPhysNC(PPRTR0MEMOBJINTERNAL ppMem, size_t cb, RTHCPHYS PhysHighest)
653	{
654	return rtR0MemObjLinuxAllocPhysSub(ppMem, RTR0MEMOBJTYPE_PHYS_NC, cb, PhysHighest);
655	}
656
657
658	int rtR0MemObjNativeEnterPhys(PPRTR0MEMOBJINTERNAL ppMem, RTHCPHYS Phys, size_t cb)
659	{
660	/*
661	* All we need to do here is to validate that we can use
662	* ioremap on the specified address (32/64-bit dma_addr_t).
663	*/
664	PRTR0MEMOBJLNX pMemLnx;
665	dma_addr_t PhysAddr = Phys;
666	AssertMsgReturn(PhysAddr == Phys, ("%#llx\n", (unsigned long long)Phys), VERR_ADDRESS_TOO_BIG);
667
668	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_PHYS, NULL, cb);
669	if (!pMemLnx)
670	return VERR_NO_MEMORY;
671
672	pMemLnx->Core.u.Phys.PhysBase = PhysAddr;
673	pMemLnx->Core.u.Phys.fAllocated = false;
674	Assert(!pMemLnx->cPages);
675	*ppMem = &pMemLnx->Core;
676	return VINF_SUCCESS;
677	}
678
679
680	int rtR0MemObjNativeLockUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3Ptr, size_t cb, RTR0PROCESS R0Process)
681	{
682	const int cPages = cb >> PAGE_SHIFT;
683	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
684	struct vm_area_struct **papVMAs;
685	PRTR0MEMOBJLNX pMemLnx;
686	int rc;
687
688	/*
689	* Check for valid task and size overflows.
690	*/
691	if (!pTask)
692	return VERR_NOT_SUPPORTED;
693	if (((size_t)cPages << PAGE_SHIFT) != cb)
694	return VERR_OUT_OF_RANGE;
695
696	/*
697	* Allocate the memory object and a temporary buffer for the VMAs.
698	*/
699	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(RT_OFFSETOF(RTR0MEMOBJLNX, apPages[cPages]), RTR0MEMOBJTYPE_LOCK, (void *)R3Ptr, cb);
700	if (!pMemLnx)
701	return VERR_NO_MEMORY;
702
703	papVMAs = (struct vm_area_struct *)RTMemAlloc(sizeof(papVMAs) * cPages);
704	if (papVMAs)
705	{
706	down_read(&pTask->mm->mmap_sem);
707
708	/*
709	* Get user pages.
710	*/
711	rc = get_user_pages(pTask, /* Task for fault acounting. */
712	pTask->mm, /* Whose pages. */
713	R3Ptr, /* Where from. */
714	cPages, /* How many pages. */
715	1, /* Write to memory. */
716	0, /* force. */
717	&pMemLnx->apPages[0], /* Page array. */
718	papVMAs); /* vmas */
719	if (rc == cPages)
720	{
721	/*
722	* Flush dcache (required?) and protect against fork.
723	*/
724	/** @todo The Linux fork() protection will require more work if this API
725	* is to be used for anything but locking VM pages. */
726	while (rc-- > 0)
727	{
728	flush_dcache_page(pMemLnx->apPages[rc]);
729	papVMAs[rc]->vm_flags \|= VM_DONTCOPY;
730	}
731
732	up_read(&pTask->mm->mmap_sem);
733
734	RTMemFree(papVMAs);
735
736	pMemLnx->Core.u.Lock.R0Process = R0Process;
737	pMemLnx->cPages = cPages;
738	Assert(!pMemLnx->fMappedToRing0);
739	*ppMem = &pMemLnx->Core;
740
741	return VINF_SUCCESS;
742	}
743
744	/*
745	* Failed - we need to unlock any pages that we succeeded to lock.
746	*/
747	while (rc-- > 0)
748	{
749	if (!PageReserved(pMemLnx->apPages[rc]))
750	SetPageDirty(pMemLnx->apPages[rc]);
751	page_cache_release(pMemLnx->apPages[rc]);
752	}
753
754	up_read(&pTask->mm->mmap_sem);
755
756	RTMemFree(papVMAs);
757	rc = VERR_LOCK_FAILED;
758	}
759
760	rtR0MemObjDelete(&pMemLnx->Core);
761	return rc;
762	}
763
764
765	int rtR0MemObjNativeLockKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pv, size_t cb)
766	{
767	/* What is there to lock? Should/Can we fake this? */
768	return VERR_NOT_SUPPORTED;
769	}
770
771
772	int rtR0MemObjNativeReserveKernel(PPRTR0MEMOBJINTERNAL ppMem, void *pvFixed, size_t cb, size_t uAlignment)
773	{
774	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
775	const size_t cPages = cb >> PAGE_SHIFT;
776	struct page *pDummyPage;
777	struct page **papPages;
778
779	/* check for unsupported stuff. */
780	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
781	AssertMsgReturn(uAlignment <= PAGE_SIZE, ("%#x\n", uAlignment), VERR_NOT_SUPPORTED);
782
783	/*
784	* Allocate a dummy page and create a page pointer array for vmap such that
785	* the dummy page is mapped all over the reserved area.
786	*/
787	pDummyPage = alloc_page(GFP_HIGHUSER);
788	if (!pDummyPage)
789	return VERR_NO_MEMORY;
790	papPages = RTMemAlloc(sizeof(papPages) cPages);
791	if (papPages)
792	{
793	void *pv;
794	size_t iPage = cPages;
795	while (iPage-- > 0)
796	papPages[iPage] = pDummyPage;
797	# ifdef VM_MAP
798	pv = vmap(papPages, cPages, VM_MAP, PAGE_KERNEL_RO);
799	# else
800	pv = vmap(papPages, cPages, VM_ALLOC, PAGE_KERNEL_RO);
801	# endif
802	RTMemFree(papPages);
803	if (pv)
804	{
805	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
806	if (pMemLnx)
807	{
808	pMemLnx->Core.u.ResVirt.R0Process = NIL_RTR0PROCESS;
809	pMemLnx->cPages = 1;
810	pMemLnx->apPages[0] = pDummyPage;
811	*ppMem = &pMemLnx->Core;
812	return VINF_SUCCESS;
813	}
814	vunmap(pv);
815	}
816	}
817	__free_page(pDummyPage);
818	return VERR_NO_MEMORY;
819
820	#else /* < 2.4.22 */
821	/*
822	* Could probably use ioremap here, but the caller is in a better position than us
823	* to select some safe physical memory.
824	*/
825	return VERR_NOT_SUPPORTED;
826	#endif
827	}
828
829
830	/**
831	* Worker for rtR0MemObjNativeReserveUser and rtR0MemObjNativerMapUser that creates
832	* an empty user space mapping.
833	*
834	* The caller takes care of acquiring the mmap_sem of the task.
835	*
836	* @returns Pointer to the mapping.
837	* (void *)-1 on failure.
838	* @param R3PtrFixed (RTR3PTR)-1 if anywhere, otherwise a specific location.
839	* @param cb The size of the mapping.
840	* @param uAlignment The alignment of the mapping.
841	* @param pTask The Linux task to create this mapping in.
842	* @param fProt The RTMEM_PROT_* mask.
843	*/
844	static void rtR0MemObjLinuxDoMmap(RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, struct task_struct pTask, unsigned fProt)
845	{
846	unsigned fLnxProt;
847	unsigned long ulAddr;
848
849	/*
850	* Convert from IPRT protection to mman.h PROT_ and call do_mmap.
851	*/
852	fProt &= (RTMEM_PROT_NONE \| RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC);
853	if (fProt == RTMEM_PROT_NONE)
854	fLnxProt = PROT_NONE;
855	else
856	{
857	fLnxProt = 0;
858	if (fProt & RTMEM_PROT_READ)
859	fLnxProt \|= PROT_READ;
860	if (fProt & RTMEM_PROT_WRITE)
861	fLnxProt \|= PROT_WRITE;
862	if (fProt & RTMEM_PROT_EXEC)
863	fLnxProt \|= PROT_EXEC;
864	}
865
866	if (R3PtrFixed != (RTR3PTR)-1)
867	ulAddr = do_mmap(NULL, R3PtrFixed, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS \| MAP_FIXED, 0);
868	else
869	{
870	ulAddr = do_mmap(NULL, 0, cb, fLnxProt, MAP_SHARED \| MAP_ANONYMOUS, 0);
871	if ( !(ulAddr & ~PAGE_MASK)
872	&& (ulAddr & (uAlignment - 1)))
873	{
874	/** @todo implement uAlignment properly... We'll probably need to make some dummy mappings to fill
875	* up alignment gaps. This is of course complicated by fragmentation (which we might have cause
876	* ourselves) and further by there begin two mmap strategies (top / bottom). */
877	/* For now, just ignore uAlignment requirements... */
878	}
879	}
880	if (ulAddr & ~PAGE_MASK) /* ~PAGE_MASK == PAGE_OFFSET_MASK */
881	return (void *)-1;
882	return (void *)ulAddr;
883	}
884
885
886	int rtR0MemObjNativeReserveUser(PPRTR0MEMOBJINTERNAL ppMem, RTR3PTR R3PtrFixed, size_t cb, size_t uAlignment, RTR0PROCESS R0Process)
887	{
888	PRTR0MEMOBJLNX pMemLnx;
889	void *pv;
890	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
891	if (!pTask)
892	return VERR_NOT_SUPPORTED;
893
894	/*
895	* Let rtR0MemObjLinuxDoMmap do the difficult bits.
896	*/
897	down_write(&pTask->mm->mmap_sem);
898	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, cb, uAlignment, pTask, RTMEM_PROT_NONE);
899	up_write(&pTask->mm->mmap_sem);
900	if (pv == (void *)-1)
901	return VERR_NO_MEMORY;
902
903	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_RES_VIRT, pv, cb);
904	if (!pMemLnx)
905	{
906	down_write(&pTask->mm->mmap_sem);
907	MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, cb);
908	up_write(&pTask->mm->mmap_sem);
909	return VERR_NO_MEMORY;
910	}
911
912	pMemLnx->Core.u.ResVirt.R0Process = R0Process;
913	*ppMem = &pMemLnx->Core;
914	return VINF_SUCCESS;
915	}
916
917
918	int rtR0MemObjNativeMapKernel(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, void *pvFixed, size_t uAlignment, unsigned fProt)
919	{
920	int rc = VERR_NO_MEMORY;
921	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
922	PRTR0MEMOBJLNX pMemLnx;
923
924	/* Fail if requested to do something we can't. */
925	AssertMsgReturn(pvFixed == (void *)-1, ("%p\n", pvFixed), VERR_NOT_SUPPORTED);
926	AssertMsgReturn(uAlignment <= PAGE_SIZE, ("%#x\n", uAlignment), VERR_NOT_SUPPORTED);
927
928	/*
929	* Create the IPRT memory object.
930	*/
931	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
932	if (pMemLnx)
933	{
934	if (pMemLnxToMap->cPages)
935	{
936	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 4, 22)
937	/*
938	* Use vmap - 2.4.22 and later.
939	*/
940	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, true /* kernel */);
941	# ifdef VM_MAP
942	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_MAP, fPg);
943	# else
944	pMemLnx->Core.pv = vmap(&pMemLnxToMap->apPages[0], pMemLnxToMap->cPages, VM_ALLOC, fPg);
945	# endif
946	if (pMemLnx->Core.pv)
947	{
948	pMemLnx->fMappedToRing0 = true;
949	rc = VINF_SUCCESS;
950	}
951	else
952	rc = VERR_MAP_FAILED;
953
954	#else /* < 2.4.22 */
955	/*
956	* Only option here is to share mappings if possible and forget about fProt.
957	*/
958	if (rtR0MemObjIsRing3(pMemToMap))
959	rc = VERR_NOT_SUPPORTED;
960	else
961	{
962	rc = VINF_SUCCESS;
963	if (!pMemLnxToMap->Core.pv)
964	rc = rtR0MemObjLinuxVMap(pMemLnxToMap, !!(fProt & RTMEM_PROT_EXEC));
965	if (RT_SUCCESS(rc))
966	{
967	Assert(pMemLnxToMap->Core.pv);
968	pMemLnx->Core.pv = pMemLnxToMap->Core.pv;
969	}
970	}
971	#endif
972	}
973	else
974	{
975	/*
976	* MMIO / physical memory.
977	*/
978	Assert(pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS && !pMemLnxToMap->Core.u.Phys.fAllocated);
979	pMemLnx->Core.pv = ioremap(pMemLnxToMap->Core.u.Phys.PhysBase, pMemLnxToMap->Core.cb);
980	if (pMemLnx->Core.pv)
981	{
982	/** @todo fix protection. */
983	rc = VINF_SUCCESS;
984	}
985	}
986	if (RT_SUCCESS(rc))
987	{
988	pMemLnx->Core.u.Mapping.R0Process = NIL_RTR0PROCESS;
989	*ppMem = &pMemLnx->Core;
990	return VINF_SUCCESS;
991	}
992	rtR0MemObjDelete(&pMemLnx->Core);
993	}
994
995	return rc;
996	}
997
998
999	int rtR0MemObjNativeMapUser(PPRTR0MEMOBJINTERNAL ppMem, RTR0MEMOBJ pMemToMap, RTR3PTR R3PtrFixed, size_t uAlignment, unsigned fProt, RTR0PROCESS R0Process)
1000	{
1001	struct task_struct *pTask = rtR0ProcessToLinuxTask(R0Process);
1002	PRTR0MEMOBJLNX pMemLnxToMap = (PRTR0MEMOBJLNX)pMemToMap;
1003	int rc = VERR_NO_MEMORY;
1004	PRTR0MEMOBJLNX pMemLnx;
1005
1006	/*
1007	* Check for restrictions.
1008	*/
1009	if (!pTask)
1010	return VERR_NOT_SUPPORTED;
1011
1012	/*
1013	* Create the IPRT memory object.
1014	*/
1015	pMemLnx = (PRTR0MEMOBJLNX)rtR0MemObjNew(sizeof(*pMemLnx), RTR0MEMOBJTYPE_MAPPING, NULL, pMemLnxToMap->Core.cb);
1016	if (pMemLnx)
1017	{
1018	/*
1019	* Allocate user space mapping.
1020	*/
1021	void *pv;
1022	down_write(&pTask->mm->mmap_sem);
1023	pv = rtR0MemObjLinuxDoMmap(R3PtrFixed, pMemLnxToMap->Core.cb, uAlignment, pTask, fProt);
1024	if (pv != (void *)-1)
1025	{
1026	/*
1027	* Map page by page into the mmap area.
1028	* This is generic, paranoid and not very efficient.
1029	*/
1030	pgprot_t fPg = rtR0MemObjLinuxConvertProt(fProt, false /* user */);
1031	unsigned long ulAddrCur = (unsigned long)pv;
1032	const size_t cPages = pMemLnxToMap->Core.cb >> PAGE_SHIFT;
1033	size_t iPage;
1034	rc = 0;
1035	if (pMemLnxToMap->cPages)
1036	{
1037	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE)
1038	{
1039	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1040	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1041	AssertBreak(vma, rc = VERR_INTERNAL_ERROR);
1042	#endif
1043
1044	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1045	rc = remap_pfn_range(vma, ulAddrCur, page_to_pfn(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1046	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1047	rc = remap_page_range(vma, ulAddrCur, page_to_phys(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1048	#else /* 2.4 */
1049	rc = remap_page_range(ulAddrCur, page_to_phys(pMemLnxToMap->apPages[iPage]), PAGE_SIZE, fPg);
1050	#endif
1051	if (rc)
1052	break;
1053	}
1054	}
1055	else
1056	{
1057	RTHCPHYS Phys;
1058	if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_PHYS)
1059	Phys = pMemLnxToMap->Core.u.Phys.PhysBase;
1060	else if (pMemLnxToMap->Core.enmType == RTR0MEMOBJTYPE_CONT)
1061	Phys = pMemLnxToMap->Core.u.Cont.Phys;
1062	else
1063	{
1064	AssertMsgFailed(("%d\n", pMemLnxToMap->Core.enmType));
1065	Phys = NIL_RTHCPHYS;
1066	}
1067	if (Phys != NIL_RTHCPHYS)
1068	{
1069	for (iPage = 0; iPage < cPages; iPage++, ulAddrCur += PAGE_SIZE, Phys += PAGE_SIZE)
1070	{
1071	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1072	struct vm_area_struct vma = find_vma(pTask->mm, ulAddrCur); / this is probably the same for all the pages... */
1073	AssertBreak(vma, rc = VERR_INTERNAL_ERROR);
1074	#endif
1075
1076	#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 11)
1077	rc = remap_pfn_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1078	#elif LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 0) \|\| defined(HAVE_26_STYLE_REMAP_PAGE_RANGE)
1079	rc = remap_page_range(vma, ulAddrCur, Phys, PAGE_SIZE, fPg);
1080	#else /* 2.4 */
1081	rc = remap_page_range(ulAddrCur, Phys, PAGE_SIZE, fPg);
1082	#endif
1083	if (rc)
1084	break;
1085	}
1086	}
1087	}
1088	if (!rc)
1089	{
1090	up_write(&pTask->mm->mmap_sem);
1091
1092	pMemLnx->Core.pv = pv;
1093	pMemLnx->Core.u.Mapping.R0Process = R0Process;
1094	*ppMem = &pMemLnx->Core;
1095	return VINF_SUCCESS;
1096	}
1097
1098	/*
1099	* Bail out.
1100	*/
1101	MY_DO_MUNMAP(pTask->mm, (unsigned long)pv, pMemLnxToMap->Core.cb);
1102	if (rc != VERR_INTERNAL_ERROR)
1103	rc = VERR_NO_MEMORY;
1104	}
1105
1106	up_write(&pTask->mm->mmap_sem);
1107
1108	rtR0MemObjDelete(&pMemLnx->Core);
1109	}
1110
1111	return rc;
1112	}
1113
1114
1115	RTHCPHYS rtR0MemObjNativeGetPagePhysAddr(PRTR0MEMOBJINTERNAL pMem, size_t iPage)
1116	{
1117	PRTR0MEMOBJLNX pMemLnx = (PRTR0MEMOBJLNX)pMem;
1118
1119	if (pMemLnx->cPages)
1120	return page_to_phys(pMemLnx->apPages[iPage]);
1121
1122	switch (pMemLnx->Core.enmType)
1123	{
1124	case RTR0MEMOBJTYPE_CONT:
1125	return pMemLnx->Core.u.Cont.Phys + (iPage << PAGE_SHIFT);
1126
1127	case RTR0MEMOBJTYPE_PHYS:
1128	return pMemLnx->Core.u.Phys.PhysBase + (iPage << PAGE_SHIFT);
1129
1130	/* the parent knows */
1131	case RTR0MEMOBJTYPE_MAPPING:
1132	return rtR0MemObjNativeGetPagePhysAddr(pMemLnx->Core.uRel.Child.pParent, iPage);
1133
1134	/* cPages > 0 */
1135	case RTR0MEMOBJTYPE_LOW:
1136	case RTR0MEMOBJTYPE_LOCK:
1137	case RTR0MEMOBJTYPE_PHYS_NC:
1138	case RTR0MEMOBJTYPE_PAGE:
1139	default:
1140	AssertMsgFailed(("%d\n", pMemLnx->Core.enmType));
1141	/* fall thru */
1142
1143	case RTR0MEMOBJTYPE_RES_VIRT:
1144	return NIL_RTHCPHYS;
1145	}
1146	}
1147

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

source: vbox/trunk/src/VBox/Runtime/r0drv/linux/memobj-r0drv-linux.c@ 4160

Download in other formats: