1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
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:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
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.
27 **************************************************************************/
29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
32 #define pr_fmt(fmt) "[TTM] " fmt
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>
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>
45 static vm_fault_t
ttm_bo_vm_fault_idle(struct ttm_buffer_object
*bo
,
51 if (likely(!bo
->moving
))
55 * Quick non-stalling check for idle.
57 if (dma_fence_is_signaled(bo
->moving
))
61 * If possible, avoid waiting for GPU with mmap_sem
64 if (vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
) {
66 if (vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)
70 up_read(&vmf
->vma
->vm_mm
->mmap_sem
);
71 (void) dma_fence_wait(bo
->moving
, true);
72 dma_resv_unlock(bo
->base
.resv
);
80 err
= dma_fence_wait(bo
->moving
, true);
81 if (unlikely(err
!= 0)) {
82 ret
= (err
!= -ERESTARTSYS
) ? VM_FAULT_SIGBUS
:
88 dma_fence_put(bo
->moving
);
95 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object
*bo
,
96 unsigned long page_offset
)
98 struct ttm_bo_device
*bdev
= bo
->bdev
;
100 if (bdev
->driver
->io_mem_pfn
)
101 return bdev
->driver
->io_mem_pfn(bo
, page_offset
);
103 return ((bo
->mem
.bus
.base
+ bo
->mem
.bus
.offset
) >> PAGE_SHIFT
)
108 * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
109 * @bo: The buffer object
110 * @vmf: The fault structure handed to the callback
112 * vm callbacks like fault() and *_mkwrite() allow for the mm_sem to be dropped
113 * during long waits, and after the wait the callback will be restarted. This
114 * is to allow other threads using the same virtual memory space concurrent
115 * access to map(), unmap() completely unrelated buffer objects. TTM buffer
116 * object reservations sometimes wait for GPU and should therefore be
117 * considered long waits. This function reserves the buffer object interruptibly
118 * taking this into account. Starvation is avoided by the vm system not
119 * allowing too many repeated restarts.
120 * This function is intended to be used in customized fault() and _mkwrite()
124 * 0 on success and the bo was reserved.
125 * VM_FAULT_RETRY if blocking wait.
126 * VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
128 vm_fault_t
ttm_bo_vm_reserve(struct ttm_buffer_object
*bo
,
129 struct vm_fault
*vmf
)
132 * Work around locking order reversal in fault / nopfn
133 * between mmap_sem and bo_reserve: Perform a trylock operation
134 * for reserve, and if it fails, retry the fault after waiting
135 * for the buffer to become unreserved.
137 if (unlikely(!dma_resv_trylock(bo
->base
.resv
))) {
138 if (vmf
->flags
& FAULT_FLAG_ALLOW_RETRY
) {
139 if (!(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
)) {
141 up_read(&vmf
->vma
->vm_mm
->mmap_sem
);
142 if (!dma_resv_lock_interruptible(bo
->base
.resv
,
144 dma_resv_unlock(bo
->base
.resv
);
148 return VM_FAULT_RETRY
;
151 if (dma_resv_lock_interruptible(bo
->base
.resv
, NULL
))
152 return VM_FAULT_NOPAGE
;
157 EXPORT_SYMBOL(ttm_bo_vm_reserve
);
160 * ttm_bo_vm_fault_reserved - TTM fault helper
161 * @vmf: The struct vm_fault given as argument to the fault callback
162 * @prot: The page protection to be used for this memory area.
163 * @num_prefault: Maximum number of prefault pages. The caller may want to
164 * specify this based on madvice settings and the size of the GPU object
165 * backed by the memory.
167 * This function inserts one or more page table entries pointing to the
168 * memory backing the buffer object, and then returns a return code
169 * instructing the caller to retry the page access.
172 * VM_FAULT_NOPAGE on success or pending signal
173 * VM_FAULT_SIGBUS on unspecified error
174 * VM_FAULT_OOM on out-of-memory
175 * VM_FAULT_RETRY if retryable wait
177 vm_fault_t
ttm_bo_vm_fault_reserved(struct vm_fault
*vmf
,
179 pgoff_t num_prefault
)
181 struct vm_area_struct
*vma
= vmf
->vma
;
182 struct vm_area_struct cvma
= *vma
;
183 struct ttm_buffer_object
*bo
= vma
->vm_private_data
;
184 struct ttm_bo_device
*bdev
= bo
->bdev
;
185 unsigned long page_offset
;
186 unsigned long page_last
;
188 struct ttm_tt
*ttm
= NULL
;
192 vm_fault_t ret
= VM_FAULT_NOPAGE
;
193 unsigned long address
= vmf
->address
;
194 struct ttm_mem_type_manager
*man
=
195 &bdev
->man
[bo
->mem
.mem_type
];
198 * Refuse to fault imported pages. This should be handled
199 * (if at all) by redirecting mmap to the exporter.
201 if (bo
->ttm
&& (bo
->ttm
->page_flags
& TTM_PAGE_FLAG_SG
))
202 return VM_FAULT_SIGBUS
;
204 if (bdev
->driver
->fault_reserve_notify
) {
205 struct dma_fence
*moving
= dma_fence_get(bo
->moving
);
207 err
= bdev
->driver
->fault_reserve_notify(bo
);
213 return VM_FAULT_NOPAGE
;
215 return VM_FAULT_SIGBUS
;
218 if (bo
->moving
!= moving
) {
219 spin_lock(&ttm_bo_glob
.lru_lock
);
220 ttm_bo_move_to_lru_tail(bo
, NULL
);
221 spin_unlock(&ttm_bo_glob
.lru_lock
);
223 dma_fence_put(moving
);
227 * Wait for buffer data in transit, due to a pipelined
230 ret
= ttm_bo_vm_fault_idle(bo
, vmf
);
231 if (unlikely(ret
!= 0))
234 err
= ttm_mem_io_lock(man
, true);
235 if (unlikely(err
!= 0))
236 return VM_FAULT_NOPAGE
;
237 err
= ttm_mem_io_reserve_vm(bo
);
238 if (unlikely(err
!= 0)) {
239 ret
= VM_FAULT_SIGBUS
;
243 page_offset
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
244 vma
->vm_pgoff
- drm_vma_node_start(&bo
->base
.vma_node
);
245 page_last
= vma_pages(vma
) + vma
->vm_pgoff
-
246 drm_vma_node_start(&bo
->base
.vma_node
);
248 if (unlikely(page_offset
>= bo
->num_pages
)) {
249 ret
= VM_FAULT_SIGBUS
;
253 cvma
.vm_page_prot
= ttm_io_prot(bo
->mem
.placement
, prot
);
254 if (!bo
->mem
.bus
.is_iomem
) {
255 struct ttm_operation_ctx ctx
= {
256 .interruptible
= false,
257 .no_wait_gpu
= false,
258 .flags
= TTM_OPT_FLAG_FORCE_ALLOC
263 if (ttm_tt_populate(bo
->ttm
, &ctx
)) {
268 /* Iomem should not be marked encrypted */
269 cvma
.vm_page_prot
= pgprot_decrypted(cvma
.vm_page_prot
);
273 * Speculatively prefault a number of pages. Only error on
276 for (i
= 0; i
< num_prefault
; ++i
) {
277 if (bo
->mem
.bus
.is_iomem
) {
278 pfn
= ttm_bo_io_mem_pfn(bo
, page_offset
);
280 page
= ttm
->pages
[page_offset
];
281 if (unlikely(!page
&& i
== 0)) {
284 } else if (unlikely(!page
)) {
287 page
->index
= drm_vma_node_start(&bo
->base
.vma_node
) +
289 pfn
= page_to_pfn(page
);
292 if (vma
->vm_flags
& VM_MIXEDMAP
)
293 ret
= vmf_insert_mixed(&cvma
, address
,
294 __pfn_to_pfn_t(pfn
, PFN_DEV
));
296 ret
= vmf_insert_pfn(&cvma
, address
, pfn
);
298 /* Never error on prefaulted PTEs */
299 if (unlikely((ret
& VM_FAULT_ERROR
))) {
306 address
+= PAGE_SIZE
;
307 if (unlikely(++page_offset
>= page_last
))
310 ret
= VM_FAULT_NOPAGE
;
312 ttm_mem_io_unlock(man
);
315 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved
);
317 vm_fault_t
ttm_bo_vm_fault(struct vm_fault
*vmf
)
319 struct vm_area_struct
*vma
= vmf
->vma
;
321 struct ttm_buffer_object
*bo
= vma
->vm_private_data
;
324 ret
= ttm_bo_vm_reserve(bo
, vmf
);
328 prot
= vm_get_page_prot(vma
->vm_flags
);
329 ret
= ttm_bo_vm_fault_reserved(vmf
, prot
, TTM_BO_VM_NUM_PREFAULT
);
330 if (ret
== VM_FAULT_RETRY
&& !(vmf
->flags
& FAULT_FLAG_RETRY_NOWAIT
))
333 dma_resv_unlock(bo
->base
.resv
);
337 EXPORT_SYMBOL(ttm_bo_vm_fault
);
339 void ttm_bo_vm_open(struct vm_area_struct
*vma
)
341 struct ttm_buffer_object
*bo
= vma
->vm_private_data
;
343 WARN_ON(bo
->bdev
->dev_mapping
!= vma
->vm_file
->f_mapping
);
347 EXPORT_SYMBOL(ttm_bo_vm_open
);
349 void ttm_bo_vm_close(struct vm_area_struct
*vma
)
351 struct ttm_buffer_object
*bo
= vma
->vm_private_data
;
354 vma
->vm_private_data
= NULL
;
356 EXPORT_SYMBOL(ttm_bo_vm_close
);
358 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object
*bo
,
359 unsigned long offset
,
360 uint8_t *buf
, int len
, int write
)
362 unsigned long page
= offset
>> PAGE_SHIFT
;
363 unsigned long bytes_left
= len
;
366 /* Copy a page at a time, that way no extra virtual address
369 offset
-= page
<< PAGE_SHIFT
;
371 unsigned long bytes
= min(bytes_left
, PAGE_SIZE
- offset
);
372 struct ttm_bo_kmap_obj map
;
376 ret
= ttm_bo_kmap(bo
, page
, 1, &map
);
380 ptr
= (uint8_t *)ttm_kmap_obj_virtual(&map
, &is_iomem
) + offset
;
381 WARN_ON_ONCE(is_iomem
);
383 memcpy(ptr
, buf
, bytes
);
385 memcpy(buf
, ptr
, bytes
);
392 } while (bytes_left
);
397 int ttm_bo_vm_access(struct vm_area_struct
*vma
, unsigned long addr
,
398 void *buf
, int len
, int write
)
400 unsigned long offset
= (addr
) - vma
->vm_start
;
401 struct ttm_buffer_object
*bo
= vma
->vm_private_data
;
404 if (len
< 1 || (offset
+ len
) >> PAGE_SHIFT
> bo
->num_pages
)
407 ret
= ttm_bo_reserve(bo
, true, false, NULL
);
411 switch (bo
->mem
.mem_type
) {
413 if (unlikely(bo
->ttm
->page_flags
& TTM_PAGE_FLAG_SWAPPED
)) {
414 ret
= ttm_tt_swapin(bo
->ttm
);
415 if (unlikely(ret
!= 0))
420 ret
= ttm_bo_vm_access_kmap(bo
, offset
, buf
, len
, write
);
423 if (bo
->bdev
->driver
->access_memory
)
424 ret
= bo
->bdev
->driver
->access_memory(
425 bo
, offset
, buf
, len
, write
);
430 ttm_bo_unreserve(bo
);
434 EXPORT_SYMBOL(ttm_bo_vm_access
);
436 static const struct vm_operations_struct ttm_bo_vm_ops
= {
437 .fault
= ttm_bo_vm_fault
,
438 .open
= ttm_bo_vm_open
,
439 .close
= ttm_bo_vm_close
,
440 .access
= ttm_bo_vm_access
443 static struct ttm_buffer_object
*ttm_bo_vm_lookup(struct ttm_bo_device
*bdev
,
444 unsigned long offset
,
447 struct drm_vma_offset_node
*node
;
448 struct ttm_buffer_object
*bo
= NULL
;
450 drm_vma_offset_lock_lookup(bdev
->vma_manager
);
452 node
= drm_vma_offset_lookup_locked(bdev
->vma_manager
, offset
, pages
);
454 bo
= container_of(node
, struct ttm_buffer_object
,
456 bo
= ttm_bo_get_unless_zero(bo
);
459 drm_vma_offset_unlock_lookup(bdev
->vma_manager
);
462 pr_err("Could not find buffer object to map\n");
467 static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object
*bo
, struct vm_area_struct
*vma
)
469 vma
->vm_ops
= &ttm_bo_vm_ops
;
472 * Note: We're transferring the bo reference to
473 * vma->vm_private_data here.
476 vma
->vm_private_data
= bo
;
479 * We'd like to use VM_PFNMAP on shared mappings, where
480 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
481 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
482 * bad for performance. Until that has been sorted out, use
483 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
485 vma
->vm_flags
|= VM_MIXEDMAP
;
486 vma
->vm_flags
|= VM_IO
| VM_DONTEXPAND
| VM_DONTDUMP
;
489 int ttm_bo_mmap(struct file
*filp
, struct vm_area_struct
*vma
,
490 struct ttm_bo_device
*bdev
)
492 struct ttm_bo_driver
*driver
;
493 struct ttm_buffer_object
*bo
;
496 if (unlikely(vma
->vm_pgoff
< DRM_FILE_PAGE_OFFSET_START
))
499 bo
= ttm_bo_vm_lookup(bdev
, vma
->vm_pgoff
, vma_pages(vma
));
503 driver
= bo
->bdev
->driver
;
504 if (unlikely(!driver
->verify_access
)) {
508 ret
= driver
->verify_access(bo
, filp
);
509 if (unlikely(ret
!= 0))
512 ttm_bo_mmap_vma_setup(bo
, vma
);
518 EXPORT_SYMBOL(ttm_bo_mmap
);
520 int ttm_bo_mmap_obj(struct vm_area_struct
*vma
, struct ttm_buffer_object
*bo
)
523 ttm_bo_mmap_vma_setup(bo
, vma
);
526 EXPORT_SYMBOL(ttm_bo_mmap_obj
);