ttm_bo_vm.c

  1/* SPDX-License-Identifier: GPL-2.0 OR MIT */
  2/**************************************************************************
  3 *
  4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
  5 * All Rights Reserved.
  6 *
  7 * Permission is hereby granted, free of charge, to any person obtaining a
  8 * copy of this software and associated documentation files (the
  9 * "Software"), to deal in the Software without restriction, including
 10 * without limitation the rights to use, copy, modify, merge, publish,
 11 * distribute, sub license, and/or sell copies of the Software, and to
 12 * permit persons to whom the Software is furnished to do so, subject to
 13 * the following conditions:
 14 *
 15 * The above copyright notice and this permission notice (including the
 16 * next paragraph) shall be included in all copies or substantial portions
 17 * of the Software.
 18 *
 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 26 *
 27 **************************************************************************/
 28/*
 29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
 30 */
 31
 32#define pr_fmt(fmt) "[TTM] " fmt
 33
 34#include <linux/export.h>
 35
 36#include <drm/ttm/ttm_bo.h>
 37#include <drm/ttm/ttm_placement.h>
 38#include <drm/ttm/ttm_tt.h>
 39
 40#include <drm/drm_drv.h>
 41#include <drm/drm_managed.h>
 42
 43static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
 44				struct vm_fault *vmf)
 45{
 46	long err = 0;
 47
 48	/*
 49	 * Quick non-stalling check for idle.
 50	 */
 51	if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_KERNEL))
 52		return 0;
 53
 54	/*
 55	 * If possible, avoid waiting for GPU with mmap_lock
 56	 * held.  We only do this if the fault allows retry and this
 57	 * is the first attempt.
 58	 */
 59	if (fault_flag_allow_retry_first(vmf->flags)) {
 60		if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
 61			return VM_FAULT_RETRY;
 62
 63		drm_gem_object_get(&bo->base);
 64		mmap_read_unlock(vmf->vma->vm_mm);
 65		(void)dma_resv_wait_timeout(bo->base.resv,
 66					    DMA_RESV_USAGE_KERNEL, true,
 67					    MAX_SCHEDULE_TIMEOUT);
 68		dma_resv_unlock(bo->base.resv);
 69		drm_gem_object_put(&bo->base);
 70		return VM_FAULT_RETRY;
 71	}
 72
 73	/*
 74	 * Ordinary wait.
 75	 */
 76	err = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_KERNEL, true,
 77				    MAX_SCHEDULE_TIMEOUT);
 78	if (unlikely(err < 0)) {
 79		return (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
 80			VM_FAULT_NOPAGE;
 81	}
 82
 83	return 0;
 84}
 85
 86static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
 87				       unsigned long page_offset)
 88{
 89	struct ttm_device *bdev = bo->bdev;
 90
 91	if (bdev->funcs->io_mem_pfn)
 92		return bdev->funcs->io_mem_pfn(bo, page_offset);
 93
 94	return (bo->resource->bus.offset >> PAGE_SHIFT) + page_offset;
 95}
 96
 97/**
 98 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
 99 * @bo: The buffer object
100 * @vmf: The fault structure handed to the callback
101 *
102 * vm callbacks like fault() and *_mkwrite() allow for the mmap_lock to be dropped
103 * during long waits, and after the wait the callback will be restarted. This
104 * is to allow other threads using the same virtual memory space concurrent
105 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
106 * object reservations sometimes wait for GPU and should therefore be
107 * considered long waits. This function reserves the buffer object interruptibly
108 * taking this into account. Starvation is avoided by the vm system not
109 * allowing too many repeated restarts.
110 * This function is intended to be used in customized fault() and _mkwrite()
111 * handlers.
112 *
113 * Return:
114 *    0 on success and the bo was reserved.
115 *    VM_FAULT_RETRY if blocking wait.
116 *    VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
117 */
118vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
119			     struct vm_fault *vmf)
120{
121	/*
122	 * Work around locking order reversal in fault / nopfn
123	 * between mmap_lock and bo_reserve: Perform a trylock operation
124	 * for reserve, and if it fails, retry the fault after waiting
125	 * for the buffer to become unreserved.
126	 */
127	if (unlikely(!dma_resv_trylock(bo->base.resv))) {
128		/*
129		 * If the fault allows retry and this is the first
130		 * fault attempt, we try to release the mmap_lock
131		 * before waiting
132		 */
133		if (fault_flag_allow_retry_first(vmf->flags)) {
134			if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
135				drm_gem_object_get(&bo->base);
136				mmap_read_unlock(vmf->vma->vm_mm);
137				if (!dma_resv_lock_interruptible(bo->base.resv,
138								 NULL))
139					dma_resv_unlock(bo->base.resv);
140				drm_gem_object_put(&bo->base);
141			}
142
143			return VM_FAULT_RETRY;
144		}
145
146		if (dma_resv_lock_interruptible(bo->base.resv, NULL))
147			return VM_FAULT_NOPAGE;
148	}
149
150	/*
151	 * Refuse to fault imported pages. This should be handled
152	 * (if at all) by redirecting mmap to the exporter.
153	 */
154	if (bo->ttm && (bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL)) {
155		if (!(bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL_MAPPABLE)) {
156			dma_resv_unlock(bo->base.resv);
157			return VM_FAULT_SIGBUS;
158		}
159	}
160
161	return 0;
162}
163EXPORT_SYMBOL(ttm_bo_vm_reserve);
164
165/**
166 * ttm_bo_vm_fault_reserved - TTM fault helper
167 * @vmf: The struct vm_fault given as argument to the fault callback
168 * @prot: The page protection to be used for this memory area.
169 * @num_prefault: Maximum number of prefault pages. The caller may want to
170 * specify this based on madvice settings and the size of the GPU object
171 * backed by the memory.
172 *
173 * This function inserts one or more page table entries pointing to the
174 * memory backing the buffer object, and then returns a return code
175 * instructing the caller to retry the page access.
176 *
177 * Return:
178 *   VM_FAULT_NOPAGE on success or pending signal
179 *   VM_FAULT_SIGBUS on unspecified error
180 *   VM_FAULT_OOM on out-of-memory
181 *   VM_FAULT_RETRY if retryable wait
182 */
183vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
184				    pgprot_t prot,
185				    pgoff_t num_prefault)
186{
187	struct vm_area_struct *vma = vmf->vma;
188	struct ttm_buffer_object *bo = vma->vm_private_data;
189	struct ttm_device *bdev = bo->bdev;
190	unsigned long page_offset;
191	unsigned long page_last;
192	unsigned long pfn;
193	struct ttm_tt *ttm = NULL;
194	struct page *page;
195	int err;
196	pgoff_t i;
197	vm_fault_t ret = VM_FAULT_NOPAGE;
198	unsigned long address = vmf->address;
199
200	/*
201	 * Wait for buffer data in transit, due to a pipelined
202	 * move.
203	 */
204	ret = ttm_bo_vm_fault_idle(bo, vmf);
205	if (unlikely(ret != 0))
206		return ret;
207
208	err = ttm_mem_io_reserve(bdev, bo->resource);
209	if (unlikely(err != 0))
210		return VM_FAULT_SIGBUS;
211
212	page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
213		vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
214	page_last = vma_pages(vma) + vma->vm_pgoff -
215		drm_vma_node_start(&bo->base.vma_node);
216
217	if (unlikely(page_offset >= PFN_UP(bo->base.size)))
218		return VM_FAULT_SIGBUS;
219
220	prot = ttm_io_prot(bo, bo->resource, prot);
221	if (!bo->resource->bus.is_iomem) {
222		struct ttm_operation_ctx ctx = {
223			.interruptible = true,
224			.no_wait_gpu = false,
225		};
226
227		ttm = bo->ttm;
228		err = ttm_bo_populate(bo, &ctx);
229		if (err) {
230			if (err == -EINTR || err == -ERESTARTSYS ||
231			    err == -EAGAIN)
232				return VM_FAULT_NOPAGE;
233
234			pr_debug("TTM fault hit %pe.\n", ERR_PTR(err));
235			return VM_FAULT_SIGBUS;
236		}
237	} else {
238		/* Iomem should not be marked encrypted */
239		prot = pgprot_decrypted(prot);
240	}
241
242	/*
243	 * Speculatively prefault a number of pages. Only error on
244	 * first page.
245	 */
246	for (i = 0; i < num_prefault; ++i) {
247		if (bo->resource->bus.is_iomem) {
248			pfn = ttm_bo_io_mem_pfn(bo, page_offset);
249		} else {
250			page = ttm->pages[page_offset];
251			if (unlikely(!page && i == 0)) {
252				return VM_FAULT_OOM;
253			} else if (unlikely(!page)) {
254				break;
255			}
256			pfn = page_to_pfn(page);
257		}
258
259		/*
260		 * Note that the value of @prot at this point may differ from
261		 * the value of @vma->vm_page_prot in the caching- and
262		 * encryption bits. This is because the exact location of the
263		 * data may not be known at mmap() time and may also change
264		 * at arbitrary times while the data is mmap'ed.
265		 * See vmf_insert_pfn_prot() for a discussion.
266		 */
267		ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
268
269		/* Never error on prefaulted PTEs */
270		if (unlikely((ret & VM_FAULT_ERROR))) {
271			if (i == 0)
272				return VM_FAULT_NOPAGE;
273			else
274				break;
275		}
276
277		address += PAGE_SIZE;
278		if (unlikely(++page_offset >= page_last))
279			break;
280	}
281	return ret;
282}
283EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
284
285static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
286{
287	struct page *dummy_page = (struct page *)res;
288
289	__free_page(dummy_page);
290}
291
292vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
293{
294	struct vm_area_struct *vma = vmf->vma;
295	struct ttm_buffer_object *bo = vma->vm_private_data;
296	struct drm_device *ddev = bo->base.dev;
297	vm_fault_t ret = VM_FAULT_NOPAGE;
298	unsigned long address;
299	unsigned long pfn;
300	struct page *page;
301
302	/* Allocate new dummy page to map all the VA range in this VMA to it*/
303	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
304	if (!page)
305		return VM_FAULT_OOM;
306
307	/* Set the page to be freed using drmm release action */
308	if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
309		return VM_FAULT_OOM;
310
311	pfn = page_to_pfn(page);
312
313	/* Prefault the entire VMA range right away to avoid further faults */
314	for (address = vma->vm_start; address < vma->vm_end;
315	     address += PAGE_SIZE)
316		ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
317
318	return ret;
319}
320EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
321
322vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
323{
324	struct vm_area_struct *vma = vmf->vma;
325	pgprot_t prot;
326	struct ttm_buffer_object *bo = vma->vm_private_data;
327	struct drm_device *ddev = bo->base.dev;
328	vm_fault_t ret;
329	int idx;
330
331	ret = ttm_bo_vm_reserve(bo, vmf);
332	if (ret)
333		return ret;
334
335	prot = vma->vm_page_prot;
336	if (drm_dev_enter(ddev, &idx)) {
337		ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
338		drm_dev_exit(idx);
339	} else {
340		ret = ttm_bo_vm_dummy_page(vmf, prot);
341	}
342	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
343		return ret;
344
345	dma_resv_unlock(bo->base.resv);
346
347	return ret;
348}
349EXPORT_SYMBOL(ttm_bo_vm_fault);
350
351void ttm_bo_vm_open(struct vm_area_struct *vma)
352{
353	struct ttm_buffer_object *bo = vma->vm_private_data;
354
355	WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
356
357	drm_gem_object_get(&bo->base);
358}
359EXPORT_SYMBOL(ttm_bo_vm_open);
360
361void ttm_bo_vm_close(struct vm_area_struct *vma)
362{
363	struct ttm_buffer_object *bo = vma->vm_private_data;
364
365	drm_gem_object_put(&bo->base);
366	vma->vm_private_data = NULL;
367}
368EXPORT_SYMBOL(ttm_bo_vm_close);
369
370static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
371				 unsigned long offset,
372				 uint8_t *buf, int len, int write)
373{
374	unsigned long page = offset >> PAGE_SHIFT;
375	unsigned long bytes_left = len;
376	int ret;
377
378	/* Copy a page at a time, that way no extra virtual address
379	 * mapping is needed
380	 */
381	offset -= page << PAGE_SHIFT;
382	do {
383		unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
384		struct ttm_bo_kmap_obj map;
385		void *ptr;
386		bool is_iomem;
387
388		ret = ttm_bo_kmap(bo, page, 1, &map);
389		if (ret)
390			return ret;
391
392		ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
393		WARN_ON_ONCE(is_iomem);
394		if (write)
395			memcpy(ptr, buf, bytes);
396		else
397			memcpy(buf, ptr, bytes);
398		ttm_bo_kunmap(&map);
399
400		page++;
401		buf += bytes;
402		bytes_left -= bytes;
403		offset = 0;
404	} while (bytes_left);
405
406	return len;
407}
408
409/**
410 * ttm_bo_access - Helper to access a buffer object
411 *
412 * @bo: ttm buffer object
413 * @offset: access offset into buffer object
414 * @buf: pointer to caller memory to read into or write from
415 * @len: length of access
416 * @write: write access
417 *
418 * Utility function to access a buffer object. Useful when buffer object cannot
419 * be easily mapped (non-contiguous, non-visible, etc...). Should not directly
420 * be exported to user space via a peak / poke interface.
421 *
422 * Returns:
423 * @len if successful, negative error code on failure.
424 */
425int ttm_bo_access(struct ttm_buffer_object *bo, unsigned long offset,
426		  void *buf, int len, int write)
427{
428	int ret;
429
430	if (len < 1 || (offset + len) > bo->base.size)
431		return -EIO;
432
433	ret = ttm_bo_reserve(bo, true, false, NULL);
434	if (ret)
435		return ret;
436
437	if (!bo->resource) {
438		ret = -ENODATA;
439		goto unlock;
440	}
441
442	switch (bo->resource->mem_type) {
443	case TTM_PL_SYSTEM:
444		fallthrough;
445	case TTM_PL_TT:
446		ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
447		break;
448	default:
449		if (bo->bdev->funcs->access_memory)
450			ret = bo->bdev->funcs->access_memory
451				(bo, offset, buf, len, write);
452		else
453			ret = -EIO;
454	}
455
456unlock:
457	ttm_bo_unreserve(bo);
458
459	return ret;
460}
461EXPORT_SYMBOL(ttm_bo_access);
462
463int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
464		     void *buf, int len, int write)
465{
466	struct ttm_buffer_object *bo = vma->vm_private_data;
467	unsigned long offset = (addr) - vma->vm_start +
468		((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
469		 << PAGE_SHIFT);
470
471	return ttm_bo_access(bo, offset, buf, len, write);
472}
473EXPORT_SYMBOL(ttm_bo_vm_access);
474
475static const struct vm_operations_struct ttm_bo_vm_ops = {
476	.fault = ttm_bo_vm_fault,
477	.open = ttm_bo_vm_open,
478	.close = ttm_bo_vm_close,
479	.access = ttm_bo_vm_access,
480};
481
482/**
483 * ttm_bo_mmap_obj - mmap memory backed by a ttm buffer object.
484 *
485 * @vma:       vma as input from the fbdev mmap method.
486 * @bo:        The bo backing the address space.
487 *
488 * Maps a buffer object.
489 */
490int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
491{
492	/* Enforce no COW since would have really strange behavior with it. */
493	if (is_cow_mapping(vma->vm_flags))
494		return -EINVAL;
495
496	drm_gem_object_get(&bo->base);
497
498	/*
499	 * Drivers may want to override the vm_ops field. Otherwise we
500	 * use TTM's default callbacks.
501	 */
502	if (!vma->vm_ops)
503		vma->vm_ops = &ttm_bo_vm_ops;
504
505	/*
506	 * Note: We're transferring the bo reference to
507	 * vma->vm_private_data here.
508	 */
509
510	vma->vm_private_data = bo;
511
512	vm_flags_set(vma, VM_PFNMAP | VM_IO | VM_DONTEXPAND | VM_DONTDUMP);
513	return 0;
514}
515EXPORT_SYMBOL(ttm_bo_mmap_obj);