Loading...
1/**************************************************************************
2 *
3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27/*
28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29 */
30
31#define pr_fmt(fmt) "[TTM] " fmt
32
33#include <drm/ttm/ttm_module.h>
34#include <drm/ttm/ttm_bo_driver.h>
35#include <drm/ttm/ttm_placement.h>
36#include <drm/drm_vma_manager.h>
37#include <linux/mm.h>
38#include <linux/pfn_t.h>
39#include <linux/rbtree.h>
40#include <linux/module.h>
41#include <linux/uaccess.h>
42#include <linux/mem_encrypt.h>
43
44#define TTM_BO_VM_NUM_PREFAULT 16
45
46static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
47 struct vm_fault *vmf)
48{
49 int ret = 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_sem
62 * held.
63 */
64 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
65 ret = VM_FAULT_RETRY;
66 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
67 goto out_unlock;
68
69 ttm_bo_reference(bo);
70 up_read(&vmf->vma->vm_mm->mmap_sem);
71 (void) dma_fence_wait(bo->moving, true);
72 ttm_bo_unreserve(bo);
73 ttm_bo_unref(&bo);
74 goto out_unlock;
75 }
76
77 /*
78 * Ordinary wait.
79 */
80 ret = dma_fence_wait(bo->moving, true);
81 if (unlikely(ret != 0)) {
82 ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
83 VM_FAULT_NOPAGE;
84 goto out_unlock;
85 }
86
87out_clear:
88 dma_fence_put(bo->moving);
89 bo->moving = NULL;
90
91out_unlock:
92 return ret;
93}
94
95static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
96 unsigned long page_offset)
97{
98 struct ttm_bo_device *bdev = bo->bdev;
99
100 if (bdev->driver->io_mem_pfn)
101 return bdev->driver->io_mem_pfn(bo, page_offset);
102
103 return ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT)
104 + page_offset;
105}
106
107static int ttm_bo_vm_fault(struct vm_fault *vmf)
108{
109 struct vm_area_struct *vma = vmf->vma;
110 struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
111 vma->vm_private_data;
112 struct ttm_bo_device *bdev = bo->bdev;
113 unsigned long page_offset;
114 unsigned long page_last;
115 unsigned long pfn;
116 struct ttm_tt *ttm = NULL;
117 struct page *page;
118 int ret;
119 int i;
120 unsigned long address = vmf->address;
121 struct ttm_mem_type_manager *man =
122 &bdev->man[bo->mem.mem_type];
123 struct vm_area_struct cvma;
124
125 /*
126 * Work around locking order reversal in fault / nopfn
127 * between mmap_sem and bo_reserve: Perform a trylock operation
128 * for reserve, and if it fails, retry the fault after waiting
129 * for the buffer to become unreserved.
130 */
131 ret = ttm_bo_reserve(bo, true, true, NULL);
132 if (unlikely(ret != 0)) {
133 if (ret != -EBUSY)
134 return VM_FAULT_NOPAGE;
135
136 if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
137 if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
138 ttm_bo_reference(bo);
139 up_read(&vmf->vma->vm_mm->mmap_sem);
140 (void) ttm_bo_wait_unreserved(bo);
141 ttm_bo_unref(&bo);
142 }
143
144 return VM_FAULT_RETRY;
145 }
146
147 /*
148 * If we'd want to change locking order to
149 * mmap_sem -> bo::reserve, we'd use a blocking reserve here
150 * instead of retrying the fault...
151 */
152 return VM_FAULT_NOPAGE;
153 }
154
155 /*
156 * Refuse to fault imported pages. This should be handled
157 * (if at all) by redirecting mmap to the exporter.
158 */
159 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
160 ret = VM_FAULT_SIGBUS;
161 goto out_unlock;
162 }
163
164 if (bdev->driver->fault_reserve_notify) {
165 ret = bdev->driver->fault_reserve_notify(bo);
166 switch (ret) {
167 case 0:
168 break;
169 case -EBUSY:
170 case -ERESTARTSYS:
171 ret = VM_FAULT_NOPAGE;
172 goto out_unlock;
173 default:
174 ret = VM_FAULT_SIGBUS;
175 goto out_unlock;
176 }
177 }
178
179 /*
180 * Wait for buffer data in transit, due to a pipelined
181 * move.
182 */
183 ret = ttm_bo_vm_fault_idle(bo, vmf);
184 if (unlikely(ret != 0)) {
185 if (ret == VM_FAULT_RETRY &&
186 !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
187 /* The BO has already been unreserved. */
188 return ret;
189 }
190
191 goto out_unlock;
192 }
193
194 ret = ttm_mem_io_lock(man, true);
195 if (unlikely(ret != 0)) {
196 ret = VM_FAULT_NOPAGE;
197 goto out_unlock;
198 }
199 ret = ttm_mem_io_reserve_vm(bo);
200 if (unlikely(ret != 0)) {
201 ret = VM_FAULT_SIGBUS;
202 goto out_io_unlock;
203 }
204
205 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
206 vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
207 page_last = vma_pages(vma) + vma->vm_pgoff -
208 drm_vma_node_start(&bo->vma_node);
209
210 if (unlikely(page_offset >= bo->num_pages)) {
211 ret = VM_FAULT_SIGBUS;
212 goto out_io_unlock;
213 }
214
215 /*
216 * Make a local vma copy to modify the page_prot member
217 * and vm_flags if necessary. The vma parameter is protected
218 * by mmap_sem in write mode.
219 */
220 cvma = *vma;
221 cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
222
223 if (bo->mem.bus.is_iomem) {
224 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
225 cvma.vm_page_prot);
226 } else {
227 struct ttm_operation_ctx ctx = {
228 .interruptible = false,
229 .no_wait_gpu = false,
230 .flags = TTM_OPT_FLAG_FORCE_ALLOC
231
232 };
233
234 ttm = bo->ttm;
235 cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
236 cvma.vm_page_prot);
237
238 /* Allocate all page at once, most common usage */
239 if (ttm_tt_populate(ttm, &ctx)) {
240 ret = VM_FAULT_OOM;
241 goto out_io_unlock;
242 }
243 }
244
245 /*
246 * Speculatively prefault a number of pages. Only error on
247 * first page.
248 */
249 for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
250 if (bo->mem.bus.is_iomem) {
251 /* Iomem should not be marked encrypted */
252 cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot);
253 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
254 } else {
255 page = ttm->pages[page_offset];
256 if (unlikely(!page && i == 0)) {
257 ret = VM_FAULT_OOM;
258 goto out_io_unlock;
259 } else if (unlikely(!page)) {
260 break;
261 }
262 page->index = drm_vma_node_start(&bo->vma_node) +
263 page_offset;
264 pfn = page_to_pfn(page);
265 }
266
267 if (vma->vm_flags & VM_MIXEDMAP)
268 ret = vm_insert_mixed(&cvma, address,
269 __pfn_to_pfn_t(pfn, PFN_DEV));
270 else
271 ret = vm_insert_pfn(&cvma, address, pfn);
272
273 /*
274 * Somebody beat us to this PTE or prefaulting to
275 * an already populated PTE, or prefaulting error.
276 */
277
278 if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
279 break;
280 else if (unlikely(ret != 0)) {
281 ret =
282 (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
283 goto out_io_unlock;
284 }
285
286 address += PAGE_SIZE;
287 if (unlikely(++page_offset >= page_last))
288 break;
289 }
290 ret = VM_FAULT_NOPAGE;
291out_io_unlock:
292 ttm_mem_io_unlock(man);
293out_unlock:
294 ttm_bo_unreserve(bo);
295 return ret;
296}
297
298static void ttm_bo_vm_open(struct vm_area_struct *vma)
299{
300 struct ttm_buffer_object *bo =
301 (struct ttm_buffer_object *)vma->vm_private_data;
302
303 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
304
305 (void)ttm_bo_reference(bo);
306}
307
308static void ttm_bo_vm_close(struct vm_area_struct *vma)
309{
310 struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
311
312 ttm_bo_unref(&bo);
313 vma->vm_private_data = NULL;
314}
315
316static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
317 unsigned long offset,
318 uint8_t *buf, int len, int write)
319{
320 unsigned long page = offset >> PAGE_SHIFT;
321 unsigned long bytes_left = len;
322 int ret;
323
324 /* Copy a page at a time, that way no extra virtual address
325 * mapping is needed
326 */
327 offset -= page << PAGE_SHIFT;
328 do {
329 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
330 struct ttm_bo_kmap_obj map;
331 void *ptr;
332 bool is_iomem;
333
334 ret = ttm_bo_kmap(bo, page, 1, &map);
335 if (ret)
336 return ret;
337
338 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
339 WARN_ON_ONCE(is_iomem);
340 if (write)
341 memcpy(ptr, buf, bytes);
342 else
343 memcpy(buf, ptr, bytes);
344 ttm_bo_kunmap(&map);
345
346 page++;
347 buf += bytes;
348 bytes_left -= bytes;
349 offset = 0;
350 } while (bytes_left);
351
352 return len;
353}
354
355static int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
356 void *buf, int len, int write)
357{
358 unsigned long offset = (addr) - vma->vm_start;
359 struct ttm_buffer_object *bo = vma->vm_private_data;
360 int ret;
361
362 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->num_pages)
363 return -EIO;
364
365 ret = ttm_bo_reserve(bo, true, false, NULL);
366 if (ret)
367 return ret;
368
369 switch (bo->mem.mem_type) {
370 case TTM_PL_SYSTEM:
371 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
372 ret = ttm_tt_swapin(bo->ttm);
373 if (unlikely(ret != 0))
374 return ret;
375 }
376 /* fall through */
377 case TTM_PL_TT:
378 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
379 break;
380 default:
381 if (bo->bdev->driver->access_memory)
382 ret = bo->bdev->driver->access_memory(
383 bo, offset, buf, len, write);
384 else
385 ret = -EIO;
386 }
387
388 ttm_bo_unreserve(bo);
389
390 return ret;
391}
392
393static const struct vm_operations_struct ttm_bo_vm_ops = {
394 .fault = ttm_bo_vm_fault,
395 .open = ttm_bo_vm_open,
396 .close = ttm_bo_vm_close,
397 .access = ttm_bo_vm_access
398};
399
400static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
401 unsigned long offset,
402 unsigned long pages)
403{
404 struct drm_vma_offset_node *node;
405 struct ttm_buffer_object *bo = NULL;
406
407 drm_vma_offset_lock_lookup(&bdev->vma_manager);
408
409 node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
410 if (likely(node)) {
411 bo = container_of(node, struct ttm_buffer_object, vma_node);
412 if (!kref_get_unless_zero(&bo->kref))
413 bo = NULL;
414 }
415
416 drm_vma_offset_unlock_lookup(&bdev->vma_manager);
417
418 if (!bo)
419 pr_err("Could not find buffer object to map\n");
420
421 return bo;
422}
423
424int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
425 struct ttm_bo_device *bdev)
426{
427 struct ttm_bo_driver *driver;
428 struct ttm_buffer_object *bo;
429 int ret;
430
431 bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
432 if (unlikely(!bo))
433 return -EINVAL;
434
435 driver = bo->bdev->driver;
436 if (unlikely(!driver->verify_access)) {
437 ret = -EPERM;
438 goto out_unref;
439 }
440 ret = driver->verify_access(bo, filp);
441 if (unlikely(ret != 0))
442 goto out_unref;
443
444 vma->vm_ops = &ttm_bo_vm_ops;
445
446 /*
447 * Note: We're transferring the bo reference to
448 * vma->vm_private_data here.
449 */
450
451 vma->vm_private_data = bo;
452
453 /*
454 * We'd like to use VM_PFNMAP on shared mappings, where
455 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
456 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
457 * bad for performance. Until that has been sorted out, use
458 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
459 */
460 vma->vm_flags |= VM_MIXEDMAP;
461 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
462 return 0;
463out_unref:
464 ttm_bo_unref(&bo);
465 return ret;
466}
467EXPORT_SYMBOL(ttm_bo_mmap);
468
469int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
470{
471 if (vma->vm_pgoff != 0)
472 return -EACCES;
473
474 vma->vm_ops = &ttm_bo_vm_ops;
475 vma->vm_private_data = ttm_bo_reference(bo);
476 vma->vm_flags |= VM_MIXEDMAP;
477 vma->vm_flags |= VM_IO | VM_DONTEXPAND;
478 return 0;
479}
480EXPORT_SYMBOL(ttm_fbdev_mmap);
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_driver.h>
35#include <drm/ttm/ttm_placement.h>
36#include <drm/drm_vma_manager.h>
37#include <drm/drm_drv.h>
38#include <drm/drm_managed.h>
39#include <linux/mm.h>
40#include <linux/pfn_t.h>
41#include <linux/rbtree.h>
42#include <linux/module.h>
43#include <linux/uaccess.h>
44#include <linux/mem_encrypt.h>
45
46static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
47 struct vm_fault *vmf)
48{
49 vm_fault_t ret = 0;
50 int err = 0;
51
52 if (likely(!bo->moving))
53 goto out_unlock;
54
55 /*
56 * Quick non-stalling check for idle.
57 */
58 if (dma_fence_is_signaled(bo->moving))
59 goto out_clear;
60
61 /*
62 * If possible, avoid waiting for GPU with mmap_lock
63 * held. We only do this if the fault allows retry and this
64 * is the first attempt.
65 */
66 if (fault_flag_allow_retry_first(vmf->flags)) {
67 ret = VM_FAULT_RETRY;
68 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
69 goto out_unlock;
70
71 ttm_bo_get(bo);
72 mmap_read_unlock(vmf->vma->vm_mm);
73 (void) dma_fence_wait(bo->moving, true);
74 dma_resv_unlock(bo->base.resv);
75 ttm_bo_put(bo);
76 goto out_unlock;
77 }
78
79 /*
80 * Ordinary wait.
81 */
82 err = dma_fence_wait(bo->moving, true);
83 if (unlikely(err != 0)) {
84 ret = (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
85 VM_FAULT_NOPAGE;
86 goto out_unlock;
87 }
88
89out_clear:
90 dma_fence_put(bo->moving);
91 bo->moving = NULL;
92
93out_unlock:
94 return ret;
95}
96
97static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
98 unsigned long page_offset)
99{
100 struct ttm_device *bdev = bo->bdev;
101
102 if (bdev->funcs->io_mem_pfn)
103 return bdev->funcs->io_mem_pfn(bo, page_offset);
104
105 return (bo->resource->bus.offset >> PAGE_SHIFT) + 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 /*
162 * Refuse to fault imported pages. This should be handled
163 * (if at all) by redirecting mmap to the exporter.
164 */
165 if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
166 dma_resv_unlock(bo->base.resv);
167 return VM_FAULT_SIGBUS;
168 }
169
170 return 0;
171}
172EXPORT_SYMBOL(ttm_bo_vm_reserve);
173
174#ifdef CONFIG_TRANSPARENT_HUGEPAGE
175/**
176 * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
177 * @vmf: Fault data
178 * @bo: The buffer object
179 * @page_offset: Page offset from bo start
180 * @fault_page_size: The size of the fault in pages.
181 * @pgprot: The page protections.
182 * Does additional checking whether it's possible to insert a PUD or PMD
183 * pfn and performs the insertion.
184 *
185 * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
186 * a huge fault was not possible, or on insertion error.
187 */
188static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
189 struct ttm_buffer_object *bo,
190 pgoff_t page_offset,
191 pgoff_t fault_page_size,
192 pgprot_t pgprot)
193{
194 pgoff_t i;
195 vm_fault_t ret;
196 unsigned long pfn;
197 pfn_t pfnt;
198 struct ttm_tt *ttm = bo->ttm;
199 bool write = vmf->flags & FAULT_FLAG_WRITE;
200
201 /* Fault should not cross bo boundary. */
202 page_offset &= ~(fault_page_size - 1);
203 if (page_offset + fault_page_size > bo->resource->num_pages)
204 goto out_fallback;
205
206 if (bo->resource->bus.is_iomem)
207 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
208 else
209 pfn = page_to_pfn(ttm->pages[page_offset]);
210
211 /* pfn must be fault_page_size aligned. */
212 if ((pfn & (fault_page_size - 1)) != 0)
213 goto out_fallback;
214
215 /* Check that memory is contiguous. */
216 if (!bo->resource->bus.is_iomem) {
217 for (i = 1; i < fault_page_size; ++i) {
218 if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
219 goto out_fallback;
220 }
221 } else if (bo->bdev->funcs->io_mem_pfn) {
222 for (i = 1; i < fault_page_size; ++i) {
223 if (ttm_bo_io_mem_pfn(bo, page_offset + i) != pfn + i)
224 goto out_fallback;
225 }
226 }
227
228 pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
229 if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
230 ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
231#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
232 else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
233 ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
234#endif
235 else
236 WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
237
238 if (ret != VM_FAULT_NOPAGE)
239 goto out_fallback;
240
241 return VM_FAULT_NOPAGE;
242out_fallback:
243 count_vm_event(THP_FAULT_FALLBACK);
244 return VM_FAULT_FALLBACK;
245}
246#else
247static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
248 struct ttm_buffer_object *bo,
249 pgoff_t page_offset,
250 pgoff_t fault_page_size,
251 pgprot_t pgprot)
252{
253 return VM_FAULT_FALLBACK;
254}
255#endif
256
257/**
258 * ttm_bo_vm_fault_reserved - TTM fault helper
259 * @vmf: The struct vm_fault given as argument to the fault callback
260 * @prot: The page protection to be used for this memory area.
261 * @num_prefault: Maximum number of prefault pages. The caller may want to
262 * specify this based on madvice settings and the size of the GPU object
263 * backed by the memory.
264 * @fault_page_size: The size of the fault in pages.
265 *
266 * This function inserts one or more page table entries pointing to the
267 * memory backing the buffer object, and then returns a return code
268 * instructing the caller to retry the page access.
269 *
270 * Return:
271 * VM_FAULT_NOPAGE on success or pending signal
272 * VM_FAULT_SIGBUS on unspecified error
273 * VM_FAULT_OOM on out-of-memory
274 * VM_FAULT_RETRY if retryable wait
275 */
276vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
277 pgprot_t prot,
278 pgoff_t num_prefault,
279 pgoff_t fault_page_size)
280{
281 struct vm_area_struct *vma = vmf->vma;
282 struct ttm_buffer_object *bo = vma->vm_private_data;
283 struct ttm_device *bdev = bo->bdev;
284 unsigned long page_offset;
285 unsigned long page_last;
286 unsigned long pfn;
287 struct ttm_tt *ttm = NULL;
288 struct page *page;
289 int err;
290 pgoff_t i;
291 vm_fault_t ret = VM_FAULT_NOPAGE;
292 unsigned long address = vmf->address;
293
294 /*
295 * Wait for buffer data in transit, due to a pipelined
296 * move.
297 */
298 ret = ttm_bo_vm_fault_idle(bo, vmf);
299 if (unlikely(ret != 0))
300 return ret;
301
302 err = ttm_mem_io_reserve(bdev, bo->resource);
303 if (unlikely(err != 0))
304 return VM_FAULT_SIGBUS;
305
306 page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
307 vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
308 page_last = vma_pages(vma) + vma->vm_pgoff -
309 drm_vma_node_start(&bo->base.vma_node);
310
311 if (unlikely(page_offset >= bo->resource->num_pages))
312 return VM_FAULT_SIGBUS;
313
314 prot = ttm_io_prot(bo, bo->resource, prot);
315 if (!bo->resource->bus.is_iomem) {
316 struct ttm_operation_ctx ctx = {
317 .interruptible = false,
318 .no_wait_gpu = false,
319 .force_alloc = true
320 };
321
322 ttm = bo->ttm;
323 if (ttm_tt_populate(bdev, bo->ttm, &ctx))
324 return VM_FAULT_OOM;
325 } else {
326 /* Iomem should not be marked encrypted */
327 prot = pgprot_decrypted(prot);
328 }
329
330 /* We don't prefault on huge faults. Yet. */
331 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1)
332 return ttm_bo_vm_insert_huge(vmf, bo, page_offset,
333 fault_page_size, prot);
334
335 /*
336 * Speculatively prefault a number of pages. Only error on
337 * first page.
338 */
339 for (i = 0; i < num_prefault; ++i) {
340 if (bo->resource->bus.is_iomem) {
341 pfn = ttm_bo_io_mem_pfn(bo, page_offset);
342 } else {
343 page = ttm->pages[page_offset];
344 if (unlikely(!page && i == 0)) {
345 return VM_FAULT_OOM;
346 } else if (unlikely(!page)) {
347 break;
348 }
349 page->index = drm_vma_node_start(&bo->base.vma_node) +
350 page_offset;
351 pfn = page_to_pfn(page);
352 }
353
354 /*
355 * Note that the value of @prot at this point may differ from
356 * the value of @vma->vm_page_prot in the caching- and
357 * encryption bits. This is because the exact location of the
358 * data may not be known at mmap() time and may also change
359 * at arbitrary times while the data is mmap'ed.
360 * See vmf_insert_mixed_prot() for a discussion.
361 */
362 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
363
364 /* Never error on prefaulted PTEs */
365 if (unlikely((ret & VM_FAULT_ERROR))) {
366 if (i == 0)
367 return VM_FAULT_NOPAGE;
368 else
369 break;
370 }
371
372 address += PAGE_SIZE;
373 if (unlikely(++page_offset >= page_last))
374 break;
375 }
376 return ret;
377}
378EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
379
380static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
381{
382 struct page *dummy_page = (struct page *)res;
383
384 __free_page(dummy_page);
385}
386
387vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
388{
389 struct vm_area_struct *vma = vmf->vma;
390 struct ttm_buffer_object *bo = vma->vm_private_data;
391 struct drm_device *ddev = bo->base.dev;
392 vm_fault_t ret = VM_FAULT_NOPAGE;
393 unsigned long address;
394 unsigned long pfn;
395 struct page *page;
396
397 /* Allocate new dummy page to map all the VA range in this VMA to it*/
398 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
399 if (!page)
400 return VM_FAULT_OOM;
401
402 /* Set the page to be freed using drmm release action */
403 if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
404 return VM_FAULT_OOM;
405
406 pfn = page_to_pfn(page);
407
408 /* Prefault the entire VMA range right away to avoid further faults */
409 for (address = vma->vm_start; address < vma->vm_end;
410 address += PAGE_SIZE)
411 ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
412
413 return ret;
414}
415EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
416
417vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
418{
419 struct vm_area_struct *vma = vmf->vma;
420 pgprot_t prot;
421 struct ttm_buffer_object *bo = vma->vm_private_data;
422 struct drm_device *ddev = bo->base.dev;
423 vm_fault_t ret;
424 int idx;
425
426 ret = ttm_bo_vm_reserve(bo, vmf);
427 if (ret)
428 return ret;
429
430 prot = vma->vm_page_prot;
431 if (drm_dev_enter(ddev, &idx)) {
432 ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
433 drm_dev_exit(idx);
434 } else {
435 ret = ttm_bo_vm_dummy_page(vmf, prot);
436 }
437 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
438 return ret;
439
440 dma_resv_unlock(bo->base.resv);
441
442 return ret;
443}
444EXPORT_SYMBOL(ttm_bo_vm_fault);
445
446void ttm_bo_vm_open(struct vm_area_struct *vma)
447{
448 struct ttm_buffer_object *bo = vma->vm_private_data;
449
450 WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
451
452 ttm_bo_get(bo);
453}
454EXPORT_SYMBOL(ttm_bo_vm_open);
455
456void ttm_bo_vm_close(struct vm_area_struct *vma)
457{
458 struct ttm_buffer_object *bo = vma->vm_private_data;
459
460 ttm_bo_put(bo);
461 vma->vm_private_data = NULL;
462}
463EXPORT_SYMBOL(ttm_bo_vm_close);
464
465static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
466 unsigned long offset,
467 uint8_t *buf, int len, int write)
468{
469 unsigned long page = offset >> PAGE_SHIFT;
470 unsigned long bytes_left = len;
471 int ret;
472
473 /* Copy a page at a time, that way no extra virtual address
474 * mapping is needed
475 */
476 offset -= page << PAGE_SHIFT;
477 do {
478 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
479 struct ttm_bo_kmap_obj map;
480 void *ptr;
481 bool is_iomem;
482
483 ret = ttm_bo_kmap(bo, page, 1, &map);
484 if (ret)
485 return ret;
486
487 ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
488 WARN_ON_ONCE(is_iomem);
489 if (write)
490 memcpy(ptr, buf, bytes);
491 else
492 memcpy(buf, ptr, bytes);
493 ttm_bo_kunmap(&map);
494
495 page++;
496 buf += bytes;
497 bytes_left -= bytes;
498 offset = 0;
499 } while (bytes_left);
500
501 return len;
502}
503
504int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
505 void *buf, int len, int write)
506{
507 struct ttm_buffer_object *bo = vma->vm_private_data;
508 unsigned long offset = (addr) - vma->vm_start +
509 ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
510 << PAGE_SHIFT);
511 int ret;
512
513 if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->resource->num_pages)
514 return -EIO;
515
516 ret = ttm_bo_reserve(bo, true, false, NULL);
517 if (ret)
518 return ret;
519
520 switch (bo->resource->mem_type) {
521 case TTM_PL_SYSTEM:
522 if (unlikely(bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
523 ret = ttm_tt_swapin(bo->ttm);
524 if (unlikely(ret != 0))
525 return ret;
526 }
527 fallthrough;
528 case TTM_PL_TT:
529 ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
530 break;
531 default:
532 if (bo->bdev->funcs->access_memory)
533 ret = bo->bdev->funcs->access_memory(
534 bo, offset, buf, len, write);
535 else
536 ret = -EIO;
537 }
538
539 ttm_bo_unreserve(bo);
540
541 return ret;
542}
543EXPORT_SYMBOL(ttm_bo_vm_access);
544
545static const struct vm_operations_struct ttm_bo_vm_ops = {
546 .fault = ttm_bo_vm_fault,
547 .open = ttm_bo_vm_open,
548 .close = ttm_bo_vm_close,
549 .access = ttm_bo_vm_access,
550};
551
552int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
553{
554 /* Enforce no COW since would have really strange behavior with it. */
555 if (is_cow_mapping(vma->vm_flags))
556 return -EINVAL;
557
558 ttm_bo_get(bo);
559
560 /*
561 * Drivers may want to override the vm_ops field. Otherwise we
562 * use TTM's default callbacks.
563 */
564 if (!vma->vm_ops)
565 vma->vm_ops = &ttm_bo_vm_ops;
566
567 /*
568 * Note: We're transferring the bo reference to
569 * vma->vm_private_data here.
570 */
571
572 vma->vm_private_data = bo;
573
574 vma->vm_flags |= VM_PFNMAP;
575 vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
576 return 0;
577}
578EXPORT_SYMBOL(ttm_bo_mmap_obj);