Loading...
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 <drm/ttm/ttm_bo.h>
35#include <drm/ttm/ttm_placement.h>
36#include <drm/ttm/ttm_tt.h>
37
38#include <drm/drm_drv.h>
39#include <drm/drm_managed.h>
40
41static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
42 struct vm_fault *vmf)
43{
44 long err = 0;
45
46 /*
47 * Quick non-stalling check for idle.
48 */
49 if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_KERNEL))
50 return 0;
51
52 /*
53 * If possible, avoid waiting for GPU with mmap_lock
54 * held. We only do this if the fault allows retry and this
55 * is the first attempt.
56 */
57 if (fault_flag_allow_retry_first(vmf->flags)) {
58 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
59 return VM_FAULT_RETRY;
60
61 ttm_bo_get(bo);
62 mmap_read_unlock(vmf->vma->vm_mm);
63 (void)dma_resv_wait_timeout(bo->base.resv,
64 DMA_RESV_USAGE_KERNEL, true,
65 MAX_SCHEDULE_TIMEOUT);
66 dma_resv_unlock(bo->base.resv);
67 ttm_bo_put(bo);
68 return VM_FAULT_RETRY;
69 }
70
71 /*
72 * Ordinary wait.
73 */
74 err = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_KERNEL, true,
75 MAX_SCHEDULE_TIMEOUT);
76 if (unlikely(err < 0)) {
77 return (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
78 VM_FAULT_NOPAGE;
79 }
80
81 return 0;
82}
83
84static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
85 unsigned long page_offset)
86{
87 struct ttm_device *bdev = bo->bdev;
88
89 if (bdev->funcs->io_mem_pfn)
90 return bdev->funcs->io_mem_pfn(bo, page_offset);
91
92 return (bo->resource->bus.offset >> PAGE_SHIFT) + page_offset;
93}
94
95/**
96 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
97 * @bo: The buffer object
98 * @vmf: The fault structure handed to the callback
99 *
100 * vm callbacks like fault() and *_mkwrite() allow for the mmap_lock to be dropped
101 * during long waits, and after the wait the callback will be restarted. This
102 * is to allow other threads using the same virtual memory space concurrent
103 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
104 * object reservations sometimes wait for GPU and should therefore be
105 * considered long waits. This function reserves the buffer object interruptibly
106 * taking this into account. Starvation is avoided by the vm system not
107 * allowing too many repeated restarts.
108 * This function is intended to be used in customized fault() and _mkwrite()
109 * handlers.
110 *
111 * Return:
112 * 0 on success and the bo was reserved.
113 * VM_FAULT_RETRY if blocking wait.
114 * VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
115 */
116vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
117 struct vm_fault *vmf)
118{
119 /*
120 * Work around locking order reversal in fault / nopfn
121 * between mmap_lock and bo_reserve: Perform a trylock operation
122 * for reserve, and if it fails, retry the fault after waiting
123 * for the buffer to become unreserved.
124 */
125 if (unlikely(!dma_resv_trylock(bo->base.resv))) {
126 /*
127 * If the fault allows retry and this is the first
128 * fault attempt, we try to release the mmap_lock
129 * before waiting
130 */
131 if (fault_flag_allow_retry_first(vmf->flags)) {
132 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
133 ttm_bo_get(bo);
134 mmap_read_unlock(vmf->vma->vm_mm);
135 if (!dma_resv_lock_interruptible(bo->base.resv,
136 NULL))
137 dma_resv_unlock(bo->base.resv);
138 ttm_bo_put(bo);
139 }
140
141 return VM_FAULT_RETRY;
142 }
143
144 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
145 return VM_FAULT_NOPAGE;
146 }
147
148 /*
149 * Refuse to fault imported pages. This should be handled
150 * (if at all) by redirecting mmap to the exporter.
151 */
152 if (bo->ttm && (bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL)) {
153 if (!(bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL_MAPPABLE)) {
154 dma_resv_unlock(bo->base.resv);
155 return VM_FAULT_SIGBUS;
156 }
157 }
158
159 return 0;
160}
161EXPORT_SYMBOL(ttm_bo_vm_reserve);
162
163/**
164 * ttm_bo_vm_fault_reserved - TTM fault helper
165 * @vmf: The struct vm_fault given as argument to the fault callback
166 * @prot: The page protection to be used for this memory area.
167 * @num_prefault: Maximum number of prefault pages. The caller may want to
168 * specify this based on madvice settings and the size of the GPU object
169 * backed by the memory.
170 *
171 * This function inserts one or more page table entries pointing to the
172 * memory backing the buffer object, and then returns a return code
173 * instructing the caller to retry the page access.
174 *
175 * Return:
176 * VM_FAULT_NOPAGE on success or pending signal
177 * VM_FAULT_SIGBUS on unspecified error
178 * VM_FAULT_OOM on out-of-memory
179 * VM_FAULT_RETRY if retryable wait
180 */
181vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
182 pgprot_t prot,
183 pgoff_t num_prefault)
184{
185 struct vm_area_struct *vma = vmf->vma;
186 struct ttm_buffer_object *bo = vma->vm_private_data;
187 struct ttm_device *bdev = bo->bdev;
188 unsigned long page_offset;
189 unsigned long page_last;
190 unsigned long pfn;
191 struct ttm_tt *ttm = NULL;
192 struct page *page;
193 int err;
194 pgoff_t i;
195 vm_fault_t ret = VM_FAULT_NOPAGE;
196 unsigned long address = vmf->address;
197
198 /*
199 * Wait for buffer data in transit, due to a pipelined
200 * move.
201 */
202 ret = ttm_bo_vm_fault_idle(bo, vmf);
203 if (unlikely(ret != 0))
204 return ret;
205
206 err = ttm_mem_io_reserve(bdev, bo->resource);
207 if (unlikely(err != 0))
208 return VM_FAULT_SIGBUS;
209
210 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
211 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
212 page_last = vma_pages(vma) + vma->vm_pgoff -
213 drm_vma_node_start(&bo->base.vma_node);
214
215 if (unlikely(page_offset >= PFN_UP(bo->base.size)))
216 return VM_FAULT_SIGBUS;
217
218 prot = ttm_io_prot(bo, bo->resource, prot);
219 if (!bo->resource->bus.is_iomem) {
220 struct ttm_operation_ctx ctx = {
221 .interruptible = true,
222 .no_wait_gpu = false,
223 .force_alloc = true
224 };
225
226 ttm = bo->ttm;
227 err = ttm_bo_populate(bo, &ctx);
228 if (err) {
229 if (err == -EINTR || err == -ERESTARTSYS ||
230 err == -EAGAIN)
231 return VM_FAULT_NOPAGE;
232
233 pr_debug("TTM fault hit %pe.\n", ERR_PTR(err));
234 return VM_FAULT_SIGBUS;
235 }
236 } else {
237 /* Iomem should not be marked encrypted */
238 prot = pgprot_decrypted(prot);
239 }
240
241 /*
242 * Speculatively prefault a number of pages. Only error on
243 * first page.
244 */
245 for (i = 0; i < num_prefault; ++i) {
246 if (bo->resource->bus.is_iomem) {
247 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
248 } else {
249 page = ttm->pages[page_offset];
250 if (unlikely(!page && i == 0)) {
251 return VM_FAULT_OOM;
252 } else if (unlikely(!page)) {
253 break;
254 }
255 pfn = page_to_pfn(page);
256 }
257
258 /*
259 * Note that the value of @prot at this point may differ from
260 * the value of @vma->vm_page_prot in the caching- and
261 * encryption bits. This is because the exact location of the
262 * data may not be known at mmap() time and may also change
263 * at arbitrary times while the data is mmap'ed.
264 * See vmf_insert_pfn_prot() for a discussion.
265 */
266 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
267
268 /* Never error on prefaulted PTEs */
269 if (unlikely((ret & VM_FAULT_ERROR))) {
270 if (i == 0)
271 return VM_FAULT_NOPAGE;
272 else
273 break;
274 }
275
276 address += PAGE_SIZE;
277 if (unlikely(++page_offset >= page_last))
278 break;
279 }
280 return ret;
281}
282EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
283
284static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
285{
286 struct page *dummy_page = (struct page *)res;
287
288 __free_page(dummy_page);
289}
290
291vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
292{
293 struct vm_area_struct *vma = vmf->vma;
294 struct ttm_buffer_object *bo = vma->vm_private_data;
295 struct drm_device *ddev = bo->base.dev;
296 vm_fault_t ret = VM_FAULT_NOPAGE;
297 unsigned long address;
298 unsigned long pfn;
299 struct page *page;
300
301 /* Allocate new dummy page to map all the VA range in this VMA to it*/
302 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
303 if (!page)
304 return VM_FAULT_OOM;
305
306 /* Set the page to be freed using drmm release action */
307 if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
308 return VM_FAULT_OOM;
309
310 pfn = page_to_pfn(page);
311
312 /* Prefault the entire VMA range right away to avoid further faults */
313 for (address = vma->vm_start; address < vma->vm_end;
314 address += PAGE_SIZE)
315 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
316
317 return ret;
318}
319EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
320
321vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
322{
323 struct vm_area_struct *vma = vmf->vma;
324 pgprot_t prot;
325 struct ttm_buffer_object *bo = vma->vm_private_data;
326 struct drm_device *ddev = bo->base.dev;
327 vm_fault_t ret;
328 int idx;
329
330 ret = ttm_bo_vm_reserve(bo, vmf);
331 if (ret)
332 return ret;
333
334 prot = vma->vm_page_prot;
335 if (drm_dev_enter(ddev, &idx)) {
336 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
337 drm_dev_exit(idx);
338 } else {
339 ret = ttm_bo_vm_dummy_page(vmf, prot);
340 }
341 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
342 return ret;
343
344 dma_resv_unlock(bo->base.resv);
345
346 return ret;
347}
348EXPORT_SYMBOL(ttm_bo_vm_fault);
349
350void ttm_bo_vm_open(struct vm_area_struct *vma)
351{
352 struct ttm_buffer_object *bo = vma->vm_private_data;
353
354 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
355
356 ttm_bo_get(bo);
357}
358EXPORT_SYMBOL(ttm_bo_vm_open);
359
360void ttm_bo_vm_close(struct vm_area_struct *vma)
361{
362 struct ttm_buffer_object *bo = vma->vm_private_data;
363
364 ttm_bo_put(bo);
365 vma->vm_private_data = NULL;
366}
367EXPORT_SYMBOL(ttm_bo_vm_close);
368
369static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
370 unsigned long offset,
371 uint8_t *buf, int len, int write)
372{
373 unsigned long page = offset >> PAGE_SHIFT;
374 unsigned long bytes_left = len;
375 int ret;
376
377 /* Copy a page at a time, that way no extra virtual address
378 * mapping is needed
379 */
380 offset -= page << PAGE_SHIFT;
381 do {
382 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
383 struct ttm_bo_kmap_obj map;
384 void *ptr;
385 bool is_iomem;
386
387 ret = ttm_bo_kmap(bo, page, 1, &map);
388 if (ret)
389 return ret;
390
391 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
392 WARN_ON_ONCE(is_iomem);
393 if (write)
394 memcpy(ptr, buf, bytes);
395 else
396 memcpy(buf, ptr, bytes);
397 ttm_bo_kunmap(&map);
398
399 page++;
400 buf += bytes;
401 bytes_left -= bytes;
402 offset = 0;
403 } while (bytes_left);
404
405 return len;
406}
407
408int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
409 void *buf, int len, int write)
410{
411 struct ttm_buffer_object *bo = vma->vm_private_data;
412 unsigned long offset = (addr) - vma->vm_start +
413 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
414 << PAGE_SHIFT);
415 int ret;
416
417 if (len < 1 || (offset + len) > bo->base.size)
418 return -EIO;
419
420 ret = ttm_bo_reserve(bo, true, false, NULL);
421 if (ret)
422 return ret;
423
424 switch (bo->resource->mem_type) {
425 case TTM_PL_SYSTEM:
426 fallthrough;
427 case TTM_PL_TT:
428 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
429 break;
430 default:
431 if (bo->bdev->funcs->access_memory)
432 ret = bo->bdev->funcs->access_memory(
433 bo, offset, buf, len, write);
434 else
435 ret = -EIO;
436 }
437
438 ttm_bo_unreserve(bo);
439
440 return ret;
441}
442EXPORT_SYMBOL(ttm_bo_vm_access);
443
444static const struct vm_operations_struct ttm_bo_vm_ops = {
445 .fault = ttm_bo_vm_fault,
446 .open = ttm_bo_vm_open,
447 .close = ttm_bo_vm_close,
448 .access = ttm_bo_vm_access,
449};
450
451/**
452 * ttm_bo_mmap_obj - mmap memory backed by a ttm buffer object.
453 *
454 * @vma: vma as input from the fbdev mmap method.
455 * @bo: The bo backing the address space.
456 *
457 * Maps a buffer object.
458 */
459int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
460{
461 /* Enforce no COW since would have really strange behavior with it. */
462 if (is_cow_mapping(vma->vm_flags))
463 return -EINVAL;
464
465 ttm_bo_get(bo);
466
467 /*
468 * Drivers may want to override the vm_ops field. Otherwise we
469 * use TTM's default callbacks.
470 */
471 if (!vma->vm_ops)
472 vma->vm_ops = &ttm_bo_vm_ops;
473
474 /*
475 * Note: We're transferring the bo reference to
476 * vma->vm_private_data here.
477 */
478
479 vma->vm_private_data = bo;
480
481 vm_flags_set(vma, VM_PFNMAP | VM_IO | VM_DONTEXPAND | VM_DONTDUMP);
482 return 0;
483}
484EXPORT_SYMBOL(ttm_bo_mmap_obj);
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 <drm/ttm/ttm_module.h>
35#include <drm/ttm/ttm_bo_driver.h>
36#include <drm/ttm/ttm_placement.h>
37#include <drm/drm_vma_manager.h>
38#include <linux/mm.h>
39#include <linux/pfn_t.h>
40#include <linux/rbtree.h>
41#include <linux/module.h>
42#include <linux/uaccess.h>
43#include <linux/mem_encrypt.h>
44
45static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
46 struct vm_fault *vmf)
47{
48 vm_fault_t ret = 0;
49 int err = 0;
50
51 if (likely(!bo->moving))
52 goto out_unlock;
53
54 /*
55 * Quick non-stalling check for idle.
56 */
57 if (dma_fence_is_signaled(bo->moving))
58 goto out_clear;
59
60 /*
61 * If possible, avoid waiting for GPU with mmap_lock
62 * held. We only do this if the fault allows retry and this
63 * is the first attempt.
64 */
65 if (fault_flag_allow_retry_first(vmf->flags)) {
66 ret = VM_FAULT_RETRY;
67 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
68 goto out_unlock;
69
70 ttm_bo_get(bo);
71 mmap_read_unlock(vmf->vma->vm_mm);
72 (void) dma_fence_wait(bo->moving, true);
73 dma_resv_unlock(bo->base.resv);
74 ttm_bo_put(bo);
75 goto out_unlock;
76 }
77
78 /*
79 * Ordinary wait.
80 */
81 err = dma_fence_wait(bo->moving, true);
82 if (unlikely(err != 0)) {
83 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
84 VM_FAULT_NOPAGE;
85 goto out_unlock;
86 }
87
88out_clear:
89 dma_fence_put(bo->moving);
90 bo->moving = NULL;
91
92out_unlock:
93 return ret;
94}
95
96static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
97 unsigned long page_offset)
98{
99 struct ttm_bo_device *bdev = bo->bdev;
100
101 if (bdev->driver->io_mem_pfn)
102 return bdev->driver->io_mem_pfn(bo, page_offset);
103
104 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
105 + page_offset;
106}
107
108/**
109 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
110 * @bo: The buffer object
111 * @vmf: The fault structure handed to the callback
112 *
113 * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
114 * during long waits, and after the wait the callback will be restarted. This
115 * is to allow other threads using the same virtual memory space concurrent
116 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
117 * object reservations sometimes wait for GPU and should therefore be
118 * considered long waits. This function reserves the buffer object interruptibly
119 * taking this into account. Starvation is avoided by the vm system not
120 * allowing too many repeated restarts.
121 * This function is intended to be used in customized fault() and _mkwrite()
122 * handlers.
123 *
124 * Return:
125 * 0 on success and the bo was reserved.
126 * VM_FAULT_RETRY if blocking wait.
127 * VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
128 */
129vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
130 struct vm_fault *vmf)
131{
132 /*
133 * Work around locking order reversal in fault / nopfn
134 * between mmap_lock and bo_reserve: Perform a trylock operation
135 * for reserve, and if it fails, retry the fault after waiting
136 * for the buffer to become unreserved.
137 */
138 if (unlikely(!dma_resv_trylock(bo->base.resv))) {
139 /*
140 * If the fault allows retry and this is the first
141 * fault attempt, we try to release the mmap_lock
142 * before waiting
143 */
144 if (fault_flag_allow_retry_first(vmf->flags)) {
145 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
146 ttm_bo_get(bo);
147 mmap_read_unlock(vmf->vma->vm_mm);
148 if (!dma_resv_lock_interruptible(bo->base.resv,
149 NULL))
150 dma_resv_unlock(bo->base.resv);
151 ttm_bo_put(bo);
152 }
153
154 return VM_FAULT_RETRY;
155 }
156
157 if (dma_resv_lock_interruptible(bo->base.resv, NULL))
158 return VM_FAULT_NOPAGE;
159 }
160
161 return 0;
162}
163EXPORT_SYMBOL(ttm_bo_vm_reserve);
164
165#ifdef CONFIG_TRANSPARENT_HUGEPAGE
166/**
167 * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
168 * @vmf: Fault data
169 * @bo: The buffer object
170 * @page_offset: Page offset from bo start
171 * @fault_page_size: The size of the fault in pages.
172 * @pgprot: The page protections.
173 * Does additional checking whether it's possible to insert a PUD or PMD
174 * pfn and performs the insertion.
175 *
176 * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
177 * a huge fault was not possible, or on insertion error.
178 */
179static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
180 struct ttm_buffer_object *bo,
181 pgoff_t page_offset,
182 pgoff_t fault_page_size,
183 pgprot_t pgprot)
184{
185 pgoff_t i;
186 vm_fault_t ret;
187 unsigned long pfn;
188 pfn_t pfnt;
189 struct ttm_tt *ttm = bo->ttm;
190 bool write = vmf->flags & FAULT_FLAG_WRITE;
191
192 /* Fault should not cross bo boundary. */
193 page_offset &= ~(fault_page_size - 1);
194 if (page_offset + fault_page_size > bo->num_pages)
195 goto out_fallback;
196
197 if (bo->mem.bus.is_iomem)
198 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
199 else
200 pfn = page_to_pfn(ttm->pages[page_offset]);
201
202 /* pfn must be fault_page_size aligned. */
203 if ((pfn & (fault_page_size - 1)) != 0)
204 goto out_fallback;
205
206 /* Check that memory is contiguous. */
207 if (!bo->mem.bus.is_iomem) {
208 for (i = 1; i < fault_page_size; ++i) {
209 if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
210 goto out_fallback;
211 }
212 } else if (bo->bdev->driver->io_mem_pfn) {
213 for (i = 1; i < fault_page_size; ++i) {
214 if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
215 goto out_fallback;
216 }
217 }
218
219 pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
220 if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
221 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
222#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
223 else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
224 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
225#endif
226 else
227 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
228
229 if (ret != VM_FAULT_NOPAGE)
230 goto out_fallback;
231
232 return VM_FAULT_NOPAGE;
233out_fallback:
234 count_vm_event(THP_FAULT_FALLBACK);
235 return VM_FAULT_FALLBACK;
236}
237#else
238static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
239 struct ttm_buffer_object *bo,
240 pgoff_t page_offset,
241 pgoff_t fault_page_size,
242 pgprot_t pgprot)
243{
244 return VM_FAULT_FALLBACK;
245}
246#endif
247
248/**
249 * ttm_bo_vm_fault_reserved - TTM fault helper
250 * @vmf: The struct vm_fault given as argument to the fault callback
251 * @prot: The page protection to be used for this memory area.
252 * @num_prefault: Maximum number of prefault pages. The caller may want to
253 * specify this based on madvice settings and the size of the GPU object
254 * backed by the memory.
255 * @fault_page_size: The size of the fault in pages.
256 *
257 * This function inserts one or more page table entries pointing to the
258 * memory backing the buffer object, and then returns a return code
259 * instructing the caller to retry the page access.
260 *
261 * Return:
262 * VM_FAULT_NOPAGE on success or pending signal
263 * VM_FAULT_SIGBUS on unspecified error
264 * VM_FAULT_OOM on out-of-memory
265 * VM_FAULT_RETRY if retryable wait
266 */
267vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
268 pgprot_t prot,
269 pgoff_t num_prefault,
270 pgoff_t fault_page_size)
271{
272 struct vm_area_struct *vma = vmf->vma;
273 struct ttm_buffer_object *bo = vma->vm_private_data;
274 struct ttm_bo_device *bdev = bo->bdev;
275 unsigned long page_offset;
276 unsigned long page_last;
277 unsigned long pfn;
278 struct ttm_tt *ttm = NULL;
279 struct page *page;
280 int err;
281 pgoff_t i;
282 vm_fault_t ret = VM_FAULT_NOPAGE;
283 unsigned long address = vmf->address;
284 struct ttm_mem_type_manager *man =
285 &bdev->man[bo->mem.mem_type];
286
287 /*
288 * Refuse to fault imported pages. This should be handled
289 * (if at all) by redirecting mmap to the exporter.
290 */
291 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
292 return VM_FAULT_SIGBUS;
293
294 if (bdev->driver->fault_reserve_notify) {
295 struct dma_fence *moving = dma_fence_get(bo->moving);
296
297 err = bdev->driver->fault_reserve_notify(bo);
298 switch (err) {
299 case 0:
300 break;
301 case -EBUSY:
302 case -ERESTARTSYS:
303 dma_fence_put(moving);
304 return VM_FAULT_NOPAGE;
305 default:
306 dma_fence_put(moving);
307 return VM_FAULT_SIGBUS;
308 }
309
310 if (bo->moving != moving) {
311 spin_lock(&ttm_bo_glob.lru_lock);
312 ttm_bo_move_to_lru_tail(bo, NULL);
313 spin_unlock(&ttm_bo_glob.lru_lock);
314 }
315 dma_fence_put(moving);
316 }
317
318 /*
319 * Wait for buffer data in transit, due to a pipelined
320 * move.
321 */
322 ret = ttm_bo_vm_fault_idle(bo, vmf);
323 if (unlikely(ret != 0))
324 return ret;
325
326 err = ttm_mem_io_lock(man, true);
327 if (unlikely(err != 0))
328 return VM_FAULT_NOPAGE;
329 err = ttm_mem_io_reserve_vm(bo);
330 if (unlikely(err != 0)) {
331 ret = VM_FAULT_SIGBUS;
332 goto out_io_unlock;
333 }
334
335 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
336 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
337 page_last = vma_pages(vma) + vma->vm_pgoff -
338 drm_vma_node_start(&bo->base.vma_node);
339
340 if (unlikely(page_offset >= bo->num_pages)) {
341 ret = VM_FAULT_SIGBUS;
342 goto out_io_unlock;
343 }
344
345 prot = ttm_io_prot(bo->mem.placement, prot);
346 if (!bo->mem.bus.is_iomem) {
347 struct ttm_operation_ctx ctx = {
348 .interruptible = false,
349 .no_wait_gpu = false,
350 .flags = TTM_OPT_FLAG_FORCE_ALLOC
351
352 };
353
354 ttm = bo->ttm;
355 if (ttm_tt_populate(bo->ttm, &ctx)) {
356 ret = VM_FAULT_OOM;
357 goto out_io_unlock;
358 }
359 } else {
360 /* Iomem should not be marked encrypted */
361 prot = pgprot_decrypted(prot);
362 }
363
364 /* We don't prefault on huge faults. Yet. */
365 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1) {
366 ret = ttm_bo_vm_insert_huge(vmf, bo, page_offset,
367 fault_page_size, prot);
368 goto out_io_unlock;
369 }
370
371 /*
372 * Speculatively prefault a number of pages. Only error on
373 * first page.
374 */
375 for (i = 0; i < num_prefault; ++i) {
376 if (bo->mem.bus.is_iomem) {
377 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
378 } else {
379 page = ttm->pages[page_offset];
380 if (unlikely(!page && i == 0)) {
381 ret = VM_FAULT_OOM;
382 goto out_io_unlock;
383 } else if (unlikely(!page)) {
384 break;
385 }
386 page->index = drm_vma_node_start(&bo->base.vma_node) +
387 page_offset;
388 pfn = page_to_pfn(page);
389 }
390
391 /*
392 * Note that the value of @prot at this point may differ from
393 * the value of @vma->vm_page_prot in the caching- and
394 * encryption bits. This is because the exact location of the
395 * data may not be known at mmap() time and may also change
396 * at arbitrary times while the data is mmap'ed.
397 * See vmf_insert_mixed_prot() for a discussion.
398 */
399 if (vma->vm_flags & VM_MIXEDMAP)
400 ret = vmf_insert_mixed_prot(vma, address,
401 __pfn_to_pfn_t(pfn, PFN_DEV),
402 prot);
403 else
404 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
405
406 /* Never error on prefaulted PTEs */
407 if (unlikely((ret & VM_FAULT_ERROR))) {
408 if (i == 0)
409 goto out_io_unlock;
410 else
411 break;
412 }
413
414 address += PAGE_SIZE;
415 if (unlikely(++page_offset >= page_last))
416 break;
417 }
418 ret = VM_FAULT_NOPAGE;
419out_io_unlock:
420 ttm_mem_io_unlock(man);
421 return ret;
422}
423EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
424
425vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
426{
427 struct vm_area_struct *vma = vmf->vma;
428 pgprot_t prot;
429 struct ttm_buffer_object *bo = vma->vm_private_data;
430 vm_fault_t ret;
431
432 ret = ttm_bo_vm_reserve(bo, vmf);
433 if (ret)
434 return ret;
435
436 prot = vma->vm_page_prot;
437 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
438 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
439 return ret;
440
441 dma_resv_unlock(bo->base.resv);
442
443 return ret;
444}
445EXPORT_SYMBOL(ttm_bo_vm_fault);
446
447void ttm_bo_vm_open(struct vm_area_struct *vma)
448{
449 struct ttm_buffer_object *bo = vma->vm_private_data;
450
451 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
452
453 ttm_bo_get(bo);
454}
455EXPORT_SYMBOL(ttm_bo_vm_open);
456
457void ttm_bo_vm_close(struct vm_area_struct *vma)
458{
459 struct ttm_buffer_object *bo = vma->vm_private_data;
460
461 ttm_bo_put(bo);
462 vma->vm_private_data = NULL;
463}
464EXPORT_SYMBOL(ttm_bo_vm_close);
465
466static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
467 unsigned long offset,
468 uint8_t *buf, int len, int write)
469{
470 unsigned long page = offset >> PAGE_SHIFT;
471 unsigned long bytes_left = len;
472 int ret;
473
474 /* Copy a page at a time, that way no extra virtual address
475 * mapping is needed
476 */
477 offset -= page << PAGE_SHIFT;
478 do {
479 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
480 struct ttm_bo_kmap_obj map;
481 void *ptr;
482 bool is_iomem;
483
484 ret = ttm_bo_kmap(bo, page, 1, &map);
485 if (ret)
486 return ret;
487
488 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
489 WARN_ON_ONCE(is_iomem);
490 if (write)
491 memcpy(ptr, buf, bytes);
492 else
493 memcpy(buf, ptr, bytes);
494 ttm_bo_kunmap(&map);
495
496 page++;
497 buf += bytes;
498 bytes_left -= bytes;
499 offset = 0;
500 } while (bytes_left);
501
502 return len;
503}
504
505int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
506 void *buf, int len, int write)
507{
508 struct ttm_buffer_object *bo = vma->vm_private_data;
509 unsigned long offset = (addr) - vma->vm_start +
510 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
511 << PAGE_SHIFT);
512 int ret;
513
514 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
515 return -EIO;
516
517 ret = ttm_bo_reserve(bo, true, false, NULL);
518 if (ret)
519 return ret;
520
521 switch (bo->mem.mem_type) {
522 case TTM_PL_SYSTEM:
523 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
524 ret = ttm_tt_swapin(bo->ttm);
525 if (unlikely(ret != 0))
526 return ret;
527 }
528 fallthrough;
529 case TTM_PL_TT:
530 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
531 break;
532 default:
533 if (bo->bdev->driver->access_memory)
534 ret = bo->bdev->driver->access_memory(
535 bo, offset, buf, len, write);
536 else
537 ret = -EIO;
538 }
539
540 ttm_bo_unreserve(bo);
541
542 return ret;
543}
544EXPORT_SYMBOL(ttm_bo_vm_access);
545
546static const struct vm_operations_struct ttm_bo_vm_ops = {
547 .fault = ttm_bo_vm_fault,
548 .open = ttm_bo_vm_open,
549 .close = ttm_bo_vm_close,
550 .access = ttm_bo_vm_access,
551};
552
553static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
554 unsigned long offset,
555 unsigned long pages)
556{
557 struct drm_vma_offset_node *node;
558 struct ttm_buffer_object *bo = NULL;
559
560 drm_vma_offset_lock_lookup(bdev->vma_manager);
561
562 node = drm_vma_offset_lookup_locked(bdev->vma_manager, offset, pages);
563 if (likely(node)) {
564 bo = container_of(node, struct ttm_buffer_object,
565 base.vma_node);
566 bo = ttm_bo_get_unless_zero(bo);
567 }
568
569 drm_vma_offset_unlock_lookup(bdev->vma_manager);
570
571 if (!bo)
572 pr_err("Could not find buffer object to map\n");
573
574 return bo;
575}
576
577static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_struct *vma)
578{
579 vma->vm_ops = &ttm_bo_vm_ops;
580
581 /*
582 * Note: We're transferring the bo reference to
583 * vma->vm_private_data here.
584 */
585
586 vma->vm_private_data = bo;
587
588 /*
589 * We'd like to use VM_PFNMAP on shared mappings, where
590 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
591 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
592 * bad for performance. Until that has been sorted out, use
593 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
594 */
595 vma->vm_flags |= VM_MIXEDMAP;
596 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
597}
598
599int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
600 struct ttm_bo_device *bdev)
601{
602 struct ttm_bo_driver *driver;
603 struct ttm_buffer_object *bo;
604 int ret;
605
606 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
607 return -EINVAL;
608
609 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
610 if (unlikely(!bo))
611 return -EINVAL;
612
613 driver = bo->bdev->driver;
614 if (unlikely(!driver->verify_access)) {
615 ret = -EPERM;
616 goto out_unref;
617 }
618 ret = driver->verify_access(bo, filp);
619 if (unlikely(ret != 0))
620 goto out_unref;
621
622 ttm_bo_mmap_vma_setup(bo, vma);
623 return 0;
624out_unref:
625 ttm_bo_put(bo);
626 return ret;
627}
628EXPORT_SYMBOL(ttm_bo_mmap);
629
630int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
631{
632 ttm_bo_get(bo);
633 ttm_bo_mmap_vma_setup(bo, vma);
634 return 0;
635}
636EXPORT_SYMBOL(ttm_bo_mmap_obj);