]>
Commit | Line | Data |
---|---|---|
d38ceaf9 AD |
1 | /* |
2 | * Copyright 2008 Advanced Micro Devices, Inc. | |
3 | * Copyright 2008 Red Hat Inc. | |
4 | * Copyright 2009 Jerome Glisse. | |
5 | * | |
6 | * Permission is hereby granted, free of charge, to any person obtaining a | |
7 | * copy of this software and associated documentation files (the "Software"), | |
8 | * to deal in the Software without restriction, including without limitation | |
9 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
10 | * and/or sell copies of the Software, and to permit persons to whom the | |
11 | * Software is furnished to do so, subject to the following conditions: | |
12 | * | |
13 | * The above copyright notice and this permission notice shall be included in | |
14 | * all copies or substantial portions of the Software. | |
15 | * | |
16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
19 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR | |
20 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | |
21 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | |
22 | * OTHER DEALINGS IN THE SOFTWARE. | |
23 | * | |
24 | * Authors: Dave Airlie | |
25 | * Alex Deucher | |
26 | * Jerome Glisse | |
27 | */ | |
28 | #include <drm/drmP.h> | |
29 | #include <drm/amdgpu_drm.h> | |
30 | #include "amdgpu.h" | |
31 | #include "amdgpu_trace.h" | |
32 | ||
33 | /* | |
34 | * GPUVM | |
35 | * GPUVM is similar to the legacy gart on older asics, however | |
36 | * rather than there being a single global gart table | |
37 | * for the entire GPU, there are multiple VM page tables active | |
38 | * at any given time. The VM page tables can contain a mix | |
39 | * vram pages and system memory pages and system memory pages | |
40 | * can be mapped as snooped (cached system pages) or unsnooped | |
41 | * (uncached system pages). | |
42 | * Each VM has an ID associated with it and there is a page table | |
43 | * associated with each VMID. When execting a command buffer, | |
44 | * the kernel tells the the ring what VMID to use for that command | |
45 | * buffer. VMIDs are allocated dynamically as commands are submitted. | |
46 | * The userspace drivers maintain their own address space and the kernel | |
47 | * sets up their pages tables accordingly when they submit their | |
48 | * command buffers and a VMID is assigned. | |
49 | * Cayman/Trinity support up to 8 active VMs at any given time; | |
50 | * SI supports 16. | |
51 | */ | |
52 | ||
53 | /** | |
54 | * amdgpu_vm_num_pde - return the number of page directory entries | |
55 | * | |
56 | * @adev: amdgpu_device pointer | |
57 | * | |
8843dbbb | 58 | * Calculate the number of page directory entries. |
d38ceaf9 AD |
59 | */ |
60 | static unsigned amdgpu_vm_num_pdes(struct amdgpu_device *adev) | |
61 | { | |
62 | return adev->vm_manager.max_pfn >> amdgpu_vm_block_size; | |
63 | } | |
64 | ||
65 | /** | |
66 | * amdgpu_vm_directory_size - returns the size of the page directory in bytes | |
67 | * | |
68 | * @adev: amdgpu_device pointer | |
69 | * | |
8843dbbb | 70 | * Calculate the size of the page directory in bytes. |
d38ceaf9 AD |
71 | */ |
72 | static unsigned amdgpu_vm_directory_size(struct amdgpu_device *adev) | |
73 | { | |
74 | return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_pdes(adev) * 8); | |
75 | } | |
76 | ||
77 | /** | |
56467ebf | 78 | * amdgpu_vm_get_pd_bo - add the VM PD to a validation list |
d38ceaf9 AD |
79 | * |
80 | * @vm: vm providing the BOs | |
3c0eea6c | 81 | * @validated: head of validation list |
56467ebf | 82 | * @entry: entry to add |
d38ceaf9 AD |
83 | * |
84 | * Add the page directory to the list of BOs to | |
56467ebf | 85 | * validate for command submission. |
d38ceaf9 | 86 | */ |
56467ebf CK |
87 | void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, |
88 | struct list_head *validated, | |
89 | struct amdgpu_bo_list_entry *entry) | |
d38ceaf9 | 90 | { |
56467ebf | 91 | entry->robj = vm->page_directory; |
56467ebf CK |
92 | entry->priority = 0; |
93 | entry->tv.bo = &vm->page_directory->tbo; | |
94 | entry->tv.shared = true; | |
95 | list_add(&entry->tv.head, validated); | |
96 | } | |
d38ceaf9 | 97 | |
56467ebf | 98 | /** |
ee1782c3 | 99 | * amdgpu_vm_get_bos - add the vm BOs to a duplicates list |
56467ebf CK |
100 | * |
101 | * @vm: vm providing the BOs | |
3c0eea6c | 102 | * @duplicates: head of duplicates list |
d38ceaf9 | 103 | * |
ee1782c3 CK |
104 | * Add the page directory to the BO duplicates list |
105 | * for command submission. | |
d38ceaf9 | 106 | */ |
ee1782c3 | 107 | void amdgpu_vm_get_pt_bos(struct amdgpu_vm *vm, struct list_head *duplicates) |
d38ceaf9 | 108 | { |
ee1782c3 | 109 | unsigned i; |
d38ceaf9 AD |
110 | |
111 | /* add the vm page table to the list */ | |
ee1782c3 CK |
112 | for (i = 0; i <= vm->max_pde_used; ++i) { |
113 | struct amdgpu_bo_list_entry *entry = &vm->page_tables[i].entry; | |
114 | ||
115 | if (!entry->robj) | |
d38ceaf9 AD |
116 | continue; |
117 | ||
ee1782c3 | 118 | list_add(&entry->tv.head, duplicates); |
d38ceaf9 | 119 | } |
eceb8a15 CK |
120 | |
121 | } | |
122 | ||
123 | /** | |
124 | * amdgpu_vm_move_pt_bos_in_lru - move the PT BOs to the LRU tail | |
125 | * | |
126 | * @adev: amdgpu device instance | |
127 | * @vm: vm providing the BOs | |
128 | * | |
129 | * Move the PT BOs to the tail of the LRU. | |
130 | */ | |
131 | void amdgpu_vm_move_pt_bos_in_lru(struct amdgpu_device *adev, | |
132 | struct amdgpu_vm *vm) | |
133 | { | |
134 | struct ttm_bo_global *glob = adev->mman.bdev.glob; | |
135 | unsigned i; | |
136 | ||
137 | spin_lock(&glob->lru_lock); | |
138 | for (i = 0; i <= vm->max_pde_used; ++i) { | |
139 | struct amdgpu_bo_list_entry *entry = &vm->page_tables[i].entry; | |
140 | ||
141 | if (!entry->robj) | |
142 | continue; | |
143 | ||
144 | ttm_bo_move_to_lru_tail(&entry->robj->tbo); | |
145 | } | |
146 | spin_unlock(&glob->lru_lock); | |
d38ceaf9 AD |
147 | } |
148 | ||
149 | /** | |
150 | * amdgpu_vm_grab_id - allocate the next free VMID | |
151 | * | |
d38ceaf9 | 152 | * @vm: vm to allocate id for |
7f8a5290 CK |
153 | * @ring: ring we want to submit job to |
154 | * @sync: sync object where we add dependencies | |
94dd0a4a | 155 | * @fence: fence protecting ID from reuse |
d38ceaf9 | 156 | * |
7f8a5290 | 157 | * Allocate an id for the vm, adding fences to the sync obj as necessary. |
d38ceaf9 | 158 | */ |
7f8a5290 | 159 | int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring, |
94dd0a4a | 160 | struct amdgpu_sync *sync, struct fence *fence) |
d38ceaf9 | 161 | { |
d38ceaf9 AD |
162 | struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx]; |
163 | struct amdgpu_device *adev = ring->adev; | |
a9a78b32 CK |
164 | struct amdgpu_vm_manager_id *id; |
165 | int r; | |
d38ceaf9 | 166 | |
94dd0a4a CK |
167 | mutex_lock(&adev->vm_manager.lock); |
168 | ||
d38ceaf9 | 169 | /* check if the id is still valid */ |
1c16c0a7 | 170 | if (vm_id->id) { |
1c16c0a7 CK |
171 | long owner; |
172 | ||
a9a78b32 CK |
173 | id = &adev->vm_manager.ids[vm_id->id]; |
174 | owner = atomic_long_read(&id->owner); | |
1c16c0a7 | 175 | if (owner == (long)vm) { |
a9a78b32 | 176 | list_move_tail(&id->list, &adev->vm_manager.ids_lru); |
165e4e07 | 177 | trace_amdgpu_vm_grab_id(vm, vm_id->id, ring->idx); |
d38ceaf9 | 178 | |
a9a78b32 CK |
179 | fence_put(id->active); |
180 | id->active = fence_get(fence); | |
d38ceaf9 | 181 | |
94dd0a4a | 182 | mutex_unlock(&adev->vm_manager.lock); |
7f8a5290 | 183 | return 0; |
d38ceaf9 | 184 | } |
d38ceaf9 AD |
185 | } |
186 | ||
a9a78b32 CK |
187 | /* we definately need to flush */ |
188 | vm_id->pd_gpu_addr = ~0ll; | |
7f8a5290 | 189 | |
a9a78b32 CK |
190 | id = list_first_entry(&adev->vm_manager.ids_lru, |
191 | struct amdgpu_vm_manager_id, | |
192 | list); | |
193 | list_move_tail(&id->list, &adev->vm_manager.ids_lru); | |
194 | atomic_long_set(&id->owner, (long)vm); | |
94dd0a4a | 195 | |
a9a78b32 CK |
196 | vm_id->id = id - adev->vm_manager.ids; |
197 | trace_amdgpu_vm_grab_id(vm, vm_id->id, ring->idx); | |
94dd0a4a | 198 | |
a9a78b32 | 199 | r = amdgpu_sync_fence(ring->adev, sync, id->active); |
94dd0a4a | 200 | |
a9a78b32 CK |
201 | if (!r) { |
202 | fence_put(id->active); | |
203 | id->active = fence_get(fence); | |
d38ceaf9 AD |
204 | } |
205 | ||
94dd0a4a | 206 | mutex_unlock(&adev->vm_manager.lock); |
a9a78b32 | 207 | return r; |
d38ceaf9 AD |
208 | } |
209 | ||
210 | /** | |
211 | * amdgpu_vm_flush - hardware flush the vm | |
212 | * | |
213 | * @ring: ring to use for flush | |
214 | * @vm: vm we want to flush | |
215 | * @updates: last vm update that we waited for | |
216 | * | |
8843dbbb | 217 | * Flush the vm. |
d38ceaf9 AD |
218 | */ |
219 | void amdgpu_vm_flush(struct amdgpu_ring *ring, | |
220 | struct amdgpu_vm *vm, | |
3c62338c | 221 | struct fence *updates) |
d38ceaf9 AD |
222 | { |
223 | uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory); | |
224 | struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx]; | |
3c62338c | 225 | struct fence *flushed_updates = vm_id->flushed_updates; |
b56c2285 | 226 | bool is_later; |
3c62338c | 227 | |
b56c2285 CK |
228 | if (!flushed_updates) |
229 | is_later = true; | |
230 | else if (!updates) | |
231 | is_later = false; | |
232 | else | |
233 | is_later = fence_is_later(updates, flushed_updates); | |
d38ceaf9 | 234 | |
b56c2285 | 235 | if (pd_addr != vm_id->pd_gpu_addr || is_later) { |
d38ceaf9 | 236 | trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id); |
b56c2285 | 237 | if (is_later) { |
3c62338c CZ |
238 | vm_id->flushed_updates = fence_get(updates); |
239 | fence_put(flushed_updates); | |
240 | } | |
d38ceaf9 AD |
241 | vm_id->pd_gpu_addr = pd_addr; |
242 | amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr); | |
243 | } | |
244 | } | |
245 | ||
d38ceaf9 AD |
246 | /** |
247 | * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo | |
248 | * | |
249 | * @vm: requested vm | |
250 | * @bo: requested buffer object | |
251 | * | |
8843dbbb | 252 | * Find @bo inside the requested vm. |
d38ceaf9 AD |
253 | * Search inside the @bos vm list for the requested vm |
254 | * Returns the found bo_va or NULL if none is found | |
255 | * | |
256 | * Object has to be reserved! | |
257 | */ | |
258 | struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm, | |
259 | struct amdgpu_bo *bo) | |
260 | { | |
261 | struct amdgpu_bo_va *bo_va; | |
262 | ||
263 | list_for_each_entry(bo_va, &bo->va, bo_list) { | |
264 | if (bo_va->vm == vm) { | |
265 | return bo_va; | |
266 | } | |
267 | } | |
268 | return NULL; | |
269 | } | |
270 | ||
271 | /** | |
272 | * amdgpu_vm_update_pages - helper to call the right asic function | |
273 | * | |
274 | * @adev: amdgpu_device pointer | |
9ab21462 CK |
275 | * @gtt: GART instance to use for mapping |
276 | * @gtt_flags: GTT hw access flags | |
d38ceaf9 AD |
277 | * @ib: indirect buffer to fill with commands |
278 | * @pe: addr of the page entry | |
279 | * @addr: dst addr to write into pe | |
280 | * @count: number of page entries to update | |
281 | * @incr: increase next addr by incr bytes | |
282 | * @flags: hw access flags | |
d38ceaf9 AD |
283 | * |
284 | * Traces the parameters and calls the right asic functions | |
285 | * to setup the page table using the DMA. | |
286 | */ | |
287 | static void amdgpu_vm_update_pages(struct amdgpu_device *adev, | |
9ab21462 CK |
288 | struct amdgpu_gart *gtt, |
289 | uint32_t gtt_flags, | |
d38ceaf9 AD |
290 | struct amdgpu_ib *ib, |
291 | uint64_t pe, uint64_t addr, | |
292 | unsigned count, uint32_t incr, | |
9ab21462 | 293 | uint32_t flags) |
d38ceaf9 AD |
294 | { |
295 | trace_amdgpu_vm_set_page(pe, addr, count, incr, flags); | |
296 | ||
9ab21462 CK |
297 | if ((gtt == &adev->gart) && (flags == gtt_flags)) { |
298 | uint64_t src = gtt->table_addr + (addr >> 12) * 8; | |
d38ceaf9 AD |
299 | amdgpu_vm_copy_pte(adev, ib, pe, src, count); |
300 | ||
9ab21462 CK |
301 | } else if (gtt) { |
302 | dma_addr_t *pages_addr = gtt->pages_addr; | |
b07c9d2a CK |
303 | amdgpu_vm_write_pte(adev, ib, pages_addr, pe, addr, |
304 | count, incr, flags); | |
305 | ||
306 | } else if (count < 3) { | |
307 | amdgpu_vm_write_pte(adev, ib, NULL, pe, addr, | |
308 | count, incr, flags); | |
d38ceaf9 AD |
309 | |
310 | } else { | |
311 | amdgpu_vm_set_pte_pde(adev, ib, pe, addr, | |
312 | count, incr, flags); | |
313 | } | |
314 | } | |
315 | ||
4c7eb91c | 316 | int amdgpu_vm_free_job(struct amdgpu_job *job) |
d5fc5e82 CZ |
317 | { |
318 | int i; | |
4c7eb91c JZ |
319 | for (i = 0; i < job->num_ibs; i++) |
320 | amdgpu_ib_free(job->adev, &job->ibs[i]); | |
321 | kfree(job->ibs); | |
d5fc5e82 CZ |
322 | return 0; |
323 | } | |
324 | ||
d38ceaf9 AD |
325 | /** |
326 | * amdgpu_vm_clear_bo - initially clear the page dir/table | |
327 | * | |
328 | * @adev: amdgpu_device pointer | |
329 | * @bo: bo to clear | |
ef9f0a83 CZ |
330 | * |
331 | * need to reserve bo first before calling it. | |
d38ceaf9 AD |
332 | */ |
333 | static int amdgpu_vm_clear_bo(struct amdgpu_device *adev, | |
334 | struct amdgpu_bo *bo) | |
335 | { | |
336 | struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring; | |
4af9f07c | 337 | struct fence *fence = NULL; |
d5fc5e82 | 338 | struct amdgpu_ib *ib; |
d38ceaf9 AD |
339 | unsigned entries; |
340 | uint64_t addr; | |
341 | int r; | |
342 | ||
ca952613 | 343 | r = reservation_object_reserve_shared(bo->tbo.resv); |
344 | if (r) | |
345 | return r; | |
346 | ||
d38ceaf9 AD |
347 | r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false); |
348 | if (r) | |
ef9f0a83 | 349 | goto error; |
d38ceaf9 AD |
350 | |
351 | addr = amdgpu_bo_gpu_offset(bo); | |
352 | entries = amdgpu_bo_size(bo) / 8; | |
353 | ||
d5fc5e82 CZ |
354 | ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL); |
355 | if (!ib) | |
ef9f0a83 | 356 | goto error; |
d38ceaf9 | 357 | |
b6ea2f37 | 358 | r = amdgpu_ib_get(ring, NULL, 64, ib); |
d38ceaf9 AD |
359 | if (r) |
360 | goto error_free; | |
361 | ||
d5fc5e82 CZ |
362 | ib->length_dw = 0; |
363 | ||
9ab21462 | 364 | amdgpu_vm_update_pages(adev, NULL, 0, ib, addr, 0, entries, 0, 0); |
d5fc5e82 | 365 | amdgpu_vm_pad_ib(adev, ib); |
b6ea2f37 | 366 | |
d5fc5e82 | 367 | WARN_ON(ib->length_dw > 64); |
4af9f07c CZ |
368 | r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1, |
369 | &amdgpu_vm_free_job, | |
370 | AMDGPU_FENCE_OWNER_VM, | |
371 | &fence); | |
372 | if (!r) | |
373 | amdgpu_bo_fence(bo, fence, true); | |
281b4223 | 374 | fence_put(fence); |
cadf97b1 | 375 | return 0; |
ef9f0a83 | 376 | |
d38ceaf9 | 377 | error_free: |
d5fc5e82 CZ |
378 | amdgpu_ib_free(adev, ib); |
379 | kfree(ib); | |
d38ceaf9 | 380 | |
ef9f0a83 | 381 | error: |
d38ceaf9 AD |
382 | return r; |
383 | } | |
384 | ||
385 | /** | |
b07c9d2a | 386 | * amdgpu_vm_map_gart - Resolve gart mapping of addr |
d38ceaf9 | 387 | * |
b07c9d2a | 388 | * @pages_addr: optional DMA address to use for lookup |
d38ceaf9 AD |
389 | * @addr: the unmapped addr |
390 | * | |
391 | * Look up the physical address of the page that the pte resolves | |
b07c9d2a | 392 | * to and return the pointer for the page table entry. |
d38ceaf9 | 393 | */ |
b07c9d2a | 394 | uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr) |
d38ceaf9 AD |
395 | { |
396 | uint64_t result; | |
397 | ||
b07c9d2a CK |
398 | if (pages_addr) { |
399 | /* page table offset */ | |
400 | result = pages_addr[addr >> PAGE_SHIFT]; | |
401 | ||
402 | /* in case cpu page size != gpu page size*/ | |
403 | result |= addr & (~PAGE_MASK); | |
404 | ||
405 | } else { | |
406 | /* No mapping required */ | |
407 | result = addr; | |
408 | } | |
d38ceaf9 | 409 | |
b07c9d2a | 410 | result &= 0xFFFFFFFFFFFFF000ULL; |
d38ceaf9 AD |
411 | |
412 | return result; | |
413 | } | |
414 | ||
415 | /** | |
416 | * amdgpu_vm_update_pdes - make sure that page directory is valid | |
417 | * | |
418 | * @adev: amdgpu_device pointer | |
419 | * @vm: requested vm | |
420 | * @start: start of GPU address range | |
421 | * @end: end of GPU address range | |
422 | * | |
423 | * Allocates new page tables if necessary | |
8843dbbb | 424 | * and updates the page directory. |
d38ceaf9 | 425 | * Returns 0 for success, error for failure. |
d38ceaf9 AD |
426 | */ |
427 | int amdgpu_vm_update_page_directory(struct amdgpu_device *adev, | |
428 | struct amdgpu_vm *vm) | |
429 | { | |
430 | struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring; | |
431 | struct amdgpu_bo *pd = vm->page_directory; | |
432 | uint64_t pd_addr = amdgpu_bo_gpu_offset(pd); | |
433 | uint32_t incr = AMDGPU_VM_PTE_COUNT * 8; | |
434 | uint64_t last_pde = ~0, last_pt = ~0; | |
435 | unsigned count = 0, pt_idx, ndw; | |
d5fc5e82 | 436 | struct amdgpu_ib *ib; |
4af9f07c | 437 | struct fence *fence = NULL; |
d5fc5e82 | 438 | |
d38ceaf9 AD |
439 | int r; |
440 | ||
441 | /* padding, etc. */ | |
442 | ndw = 64; | |
443 | ||
444 | /* assume the worst case */ | |
445 | ndw += vm->max_pde_used * 6; | |
446 | ||
d5fc5e82 CZ |
447 | ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL); |
448 | if (!ib) | |
449 | return -ENOMEM; | |
450 | ||
451 | r = amdgpu_ib_get(ring, NULL, ndw * 4, ib); | |
7a574557 SM |
452 | if (r) { |
453 | kfree(ib); | |
d38ceaf9 | 454 | return r; |
7a574557 | 455 | } |
d5fc5e82 | 456 | ib->length_dw = 0; |
d38ceaf9 AD |
457 | |
458 | /* walk over the address space and update the page directory */ | |
459 | for (pt_idx = 0; pt_idx <= vm->max_pde_used; ++pt_idx) { | |
ee1782c3 | 460 | struct amdgpu_bo *bo = vm->page_tables[pt_idx].entry.robj; |
d38ceaf9 AD |
461 | uint64_t pde, pt; |
462 | ||
463 | if (bo == NULL) | |
464 | continue; | |
465 | ||
466 | pt = amdgpu_bo_gpu_offset(bo); | |
467 | if (vm->page_tables[pt_idx].addr == pt) | |
468 | continue; | |
469 | vm->page_tables[pt_idx].addr = pt; | |
470 | ||
471 | pde = pd_addr + pt_idx * 8; | |
472 | if (((last_pde + 8 * count) != pde) || | |
473 | ((last_pt + incr * count) != pt)) { | |
474 | ||
475 | if (count) { | |
9ab21462 CK |
476 | amdgpu_vm_update_pages(adev, NULL, 0, ib, |
477 | last_pde, last_pt, | |
478 | count, incr, | |
479 | AMDGPU_PTE_VALID); | |
d38ceaf9 AD |
480 | } |
481 | ||
482 | count = 1; | |
483 | last_pde = pde; | |
484 | last_pt = pt; | |
485 | } else { | |
486 | ++count; | |
487 | } | |
488 | } | |
489 | ||
490 | if (count) | |
9ab21462 CK |
491 | amdgpu_vm_update_pages(adev, NULL, 0, ib, last_pde, last_pt, |
492 | count, incr, AMDGPU_PTE_VALID); | |
d38ceaf9 | 493 | |
d5fc5e82 CZ |
494 | if (ib->length_dw != 0) { |
495 | amdgpu_vm_pad_ib(adev, ib); | |
496 | amdgpu_sync_resv(adev, &ib->sync, pd->tbo.resv, AMDGPU_FENCE_OWNER_VM); | |
497 | WARN_ON(ib->length_dw > ndw); | |
4af9f07c CZ |
498 | r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1, |
499 | &amdgpu_vm_free_job, | |
500 | AMDGPU_FENCE_OWNER_VM, | |
501 | &fence); | |
502 | if (r) | |
503 | goto error_free; | |
05906dec | 504 | |
4af9f07c | 505 | amdgpu_bo_fence(pd, fence, true); |
05906dec BN |
506 | fence_put(vm->page_directory_fence); |
507 | vm->page_directory_fence = fence_get(fence); | |
281b4223 | 508 | fence_put(fence); |
d38ceaf9 | 509 | } |
d5fc5e82 | 510 | |
cadf97b1 | 511 | if (ib->length_dw == 0) { |
d5fc5e82 CZ |
512 | amdgpu_ib_free(adev, ib); |
513 | kfree(ib); | |
514 | } | |
d38ceaf9 AD |
515 | |
516 | return 0; | |
d5fc5e82 CZ |
517 | |
518 | error_free: | |
d5fc5e82 CZ |
519 | amdgpu_ib_free(adev, ib); |
520 | kfree(ib); | |
4af9f07c | 521 | return r; |
d38ceaf9 AD |
522 | } |
523 | ||
524 | /** | |
525 | * amdgpu_vm_frag_ptes - add fragment information to PTEs | |
526 | * | |
527 | * @adev: amdgpu_device pointer | |
9ab21462 CK |
528 | * @gtt: GART instance to use for mapping |
529 | * @gtt_flags: GTT hw mapping flags | |
d38ceaf9 AD |
530 | * @ib: IB for the update |
531 | * @pe_start: first PTE to handle | |
532 | * @pe_end: last PTE to handle | |
533 | * @addr: addr those PTEs should point to | |
534 | * @flags: hw mapping flags | |
d38ceaf9 AD |
535 | */ |
536 | static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev, | |
9ab21462 CK |
537 | struct amdgpu_gart *gtt, |
538 | uint32_t gtt_flags, | |
d38ceaf9 AD |
539 | struct amdgpu_ib *ib, |
540 | uint64_t pe_start, uint64_t pe_end, | |
9ab21462 | 541 | uint64_t addr, uint32_t flags) |
d38ceaf9 AD |
542 | { |
543 | /** | |
544 | * The MC L1 TLB supports variable sized pages, based on a fragment | |
545 | * field in the PTE. When this field is set to a non-zero value, page | |
546 | * granularity is increased from 4KB to (1 << (12 + frag)). The PTE | |
547 | * flags are considered valid for all PTEs within the fragment range | |
548 | * and corresponding mappings are assumed to be physically contiguous. | |
549 | * | |
550 | * The L1 TLB can store a single PTE for the whole fragment, | |
551 | * significantly increasing the space available for translation | |
552 | * caching. This leads to large improvements in throughput when the | |
553 | * TLB is under pressure. | |
554 | * | |
555 | * The L2 TLB distributes small and large fragments into two | |
556 | * asymmetric partitions. The large fragment cache is significantly | |
557 | * larger. Thus, we try to use large fragments wherever possible. | |
558 | * Userspace can support this by aligning virtual base address and | |
559 | * allocation size to the fragment size. | |
560 | */ | |
561 | ||
562 | /* SI and newer are optimized for 64KB */ | |
563 | uint64_t frag_flags = AMDGPU_PTE_FRAG_64KB; | |
564 | uint64_t frag_align = 0x80; | |
565 | ||
566 | uint64_t frag_start = ALIGN(pe_start, frag_align); | |
567 | uint64_t frag_end = pe_end & ~(frag_align - 1); | |
568 | ||
569 | unsigned count; | |
570 | ||
31f6c1fe CK |
571 | /* Abort early if there isn't anything to do */ |
572 | if (pe_start == pe_end) | |
573 | return; | |
574 | ||
d38ceaf9 | 575 | /* system pages are non continuously */ |
9ab21462 | 576 | if (gtt || !(flags & AMDGPU_PTE_VALID) || (frag_start >= frag_end)) { |
d38ceaf9 AD |
577 | |
578 | count = (pe_end - pe_start) / 8; | |
9ab21462 CK |
579 | amdgpu_vm_update_pages(adev, gtt, gtt_flags, ib, pe_start, |
580 | addr, count, AMDGPU_GPU_PAGE_SIZE, | |
581 | flags); | |
d38ceaf9 AD |
582 | return; |
583 | } | |
584 | ||
585 | /* handle the 4K area at the beginning */ | |
586 | if (pe_start != frag_start) { | |
587 | count = (frag_start - pe_start) / 8; | |
9ab21462 CK |
588 | amdgpu_vm_update_pages(adev, NULL, 0, ib, pe_start, addr, |
589 | count, AMDGPU_GPU_PAGE_SIZE, flags); | |
d38ceaf9 AD |
590 | addr += AMDGPU_GPU_PAGE_SIZE * count; |
591 | } | |
592 | ||
593 | /* handle the area in the middle */ | |
594 | count = (frag_end - frag_start) / 8; | |
9ab21462 CK |
595 | amdgpu_vm_update_pages(adev, NULL, 0, ib, frag_start, addr, count, |
596 | AMDGPU_GPU_PAGE_SIZE, flags | frag_flags); | |
d38ceaf9 AD |
597 | |
598 | /* handle the 4K area at the end */ | |
599 | if (frag_end != pe_end) { | |
600 | addr += AMDGPU_GPU_PAGE_SIZE * count; | |
601 | count = (pe_end - frag_end) / 8; | |
9ab21462 CK |
602 | amdgpu_vm_update_pages(adev, NULL, 0, ib, frag_end, addr, |
603 | count, AMDGPU_GPU_PAGE_SIZE, flags); | |
d38ceaf9 AD |
604 | } |
605 | } | |
606 | ||
607 | /** | |
608 | * amdgpu_vm_update_ptes - make sure that page tables are valid | |
609 | * | |
610 | * @adev: amdgpu_device pointer | |
9ab21462 CK |
611 | * @gtt: GART instance to use for mapping |
612 | * @gtt_flags: GTT hw mapping flags | |
d38ceaf9 AD |
613 | * @vm: requested vm |
614 | * @start: start of GPU address range | |
615 | * @end: end of GPU address range | |
616 | * @dst: destination address to map to | |
617 | * @flags: mapping flags | |
618 | * | |
8843dbbb | 619 | * Update the page tables in the range @start - @end. |
d38ceaf9 | 620 | */ |
a1e08d3b CK |
621 | static void amdgpu_vm_update_ptes(struct amdgpu_device *adev, |
622 | struct amdgpu_gart *gtt, | |
623 | uint32_t gtt_flags, | |
624 | struct amdgpu_vm *vm, | |
625 | struct amdgpu_ib *ib, | |
626 | uint64_t start, uint64_t end, | |
627 | uint64_t dst, uint32_t flags) | |
d38ceaf9 | 628 | { |
31f6c1fe CK |
629 | const uint64_t mask = AMDGPU_VM_PTE_COUNT - 1; |
630 | ||
631 | uint64_t last_pe_start = ~0, last_pe_end = ~0, last_dst = ~0; | |
d38ceaf9 AD |
632 | uint64_t addr; |
633 | ||
634 | /* walk over the address space and update the page tables */ | |
635 | for (addr = start; addr < end; ) { | |
636 | uint64_t pt_idx = addr >> amdgpu_vm_block_size; | |
ee1782c3 | 637 | struct amdgpu_bo *pt = vm->page_tables[pt_idx].entry.robj; |
d38ceaf9 | 638 | unsigned nptes; |
31f6c1fe | 639 | uint64_t pe_start; |
d38ceaf9 AD |
640 | |
641 | if ((addr & ~mask) == (end & ~mask)) | |
642 | nptes = end - addr; | |
643 | else | |
644 | nptes = AMDGPU_VM_PTE_COUNT - (addr & mask); | |
645 | ||
31f6c1fe CK |
646 | pe_start = amdgpu_bo_gpu_offset(pt); |
647 | pe_start += (addr & mask) * 8; | |
d38ceaf9 | 648 | |
31f6c1fe | 649 | if (last_pe_end != pe_start) { |
d38ceaf9 | 650 | |
31f6c1fe CK |
651 | amdgpu_vm_frag_ptes(adev, gtt, gtt_flags, ib, |
652 | last_pe_start, last_pe_end, | |
653 | last_dst, flags); | |
d38ceaf9 | 654 | |
31f6c1fe CK |
655 | last_pe_start = pe_start; |
656 | last_pe_end = pe_start + 8 * nptes; | |
d38ceaf9 AD |
657 | last_dst = dst; |
658 | } else { | |
31f6c1fe | 659 | last_pe_end += 8 * nptes; |
d38ceaf9 AD |
660 | } |
661 | ||
662 | addr += nptes; | |
663 | dst += nptes * AMDGPU_GPU_PAGE_SIZE; | |
664 | } | |
665 | ||
31f6c1fe CK |
666 | amdgpu_vm_frag_ptes(adev, gtt, gtt_flags, ib, |
667 | last_pe_start, last_pe_end, | |
668 | last_dst, flags); | |
d38ceaf9 AD |
669 | } |
670 | ||
d38ceaf9 AD |
671 | /** |
672 | * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table | |
673 | * | |
674 | * @adev: amdgpu_device pointer | |
9ab21462 | 675 | * @gtt: GART instance to use for mapping |
a14faa65 | 676 | * @gtt_flags: flags as they are used for GTT |
d38ceaf9 | 677 | * @vm: requested vm |
a14faa65 CK |
678 | * @start: start of mapped range |
679 | * @last: last mapped entry | |
680 | * @flags: flags for the entries | |
d38ceaf9 | 681 | * @addr: addr to set the area to |
d38ceaf9 AD |
682 | * @fence: optional resulting fence |
683 | * | |
a14faa65 | 684 | * Fill in the page table entries between @start and @last. |
d38ceaf9 | 685 | * Returns 0 for success, -EINVAL for failure. |
d38ceaf9 AD |
686 | */ |
687 | static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, | |
9ab21462 CK |
688 | struct amdgpu_gart *gtt, |
689 | uint32_t gtt_flags, | |
d38ceaf9 | 690 | struct amdgpu_vm *vm, |
a14faa65 CK |
691 | uint64_t start, uint64_t last, |
692 | uint32_t flags, uint64_t addr, | |
693 | struct fence **fence) | |
d38ceaf9 AD |
694 | { |
695 | struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring; | |
a1e08d3b | 696 | void *owner = AMDGPU_FENCE_OWNER_VM; |
d38ceaf9 | 697 | unsigned nptes, ncmds, ndw; |
d5fc5e82 | 698 | struct amdgpu_ib *ib; |
4af9f07c | 699 | struct fence *f = NULL; |
d38ceaf9 AD |
700 | int r; |
701 | ||
a1e08d3b CK |
702 | /* sync to everything on unmapping */ |
703 | if (!(flags & AMDGPU_PTE_VALID)) | |
704 | owner = AMDGPU_FENCE_OWNER_UNDEFINED; | |
705 | ||
a14faa65 | 706 | nptes = last - start + 1; |
d38ceaf9 AD |
707 | |
708 | /* | |
709 | * reserve space for one command every (1 << BLOCK_SIZE) | |
710 | * entries or 2k dwords (whatever is smaller) | |
711 | */ | |
712 | ncmds = (nptes >> min(amdgpu_vm_block_size, 11)) + 1; | |
713 | ||
714 | /* padding, etc. */ | |
715 | ndw = 64; | |
716 | ||
9ab21462 | 717 | if ((gtt == &adev->gart) && (flags == gtt_flags)) { |
d38ceaf9 AD |
718 | /* only copy commands needed */ |
719 | ndw += ncmds * 7; | |
720 | ||
9ab21462 | 721 | } else if (gtt) { |
d38ceaf9 AD |
722 | /* header for write data commands */ |
723 | ndw += ncmds * 4; | |
724 | ||
725 | /* body of write data command */ | |
726 | ndw += nptes * 2; | |
727 | ||
728 | } else { | |
729 | /* set page commands needed */ | |
730 | ndw += ncmds * 10; | |
731 | ||
732 | /* two extra commands for begin/end of fragment */ | |
733 | ndw += 2 * 10; | |
734 | } | |
735 | ||
d5fc5e82 CZ |
736 | ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL); |
737 | if (!ib) | |
738 | return -ENOMEM; | |
739 | ||
740 | r = amdgpu_ib_get(ring, NULL, ndw * 4, ib); | |
741 | if (r) { | |
742 | kfree(ib); | |
d38ceaf9 | 743 | return r; |
d5fc5e82 CZ |
744 | } |
745 | ||
a1e08d3b CK |
746 | r = amdgpu_sync_resv(adev, &ib->sync, vm->page_directory->tbo.resv, |
747 | owner); | |
748 | if (r) | |
749 | goto error_free; | |
d38ceaf9 | 750 | |
a1e08d3b CK |
751 | r = reservation_object_reserve_shared(vm->page_directory->tbo.resv); |
752 | if (r) | |
753 | goto error_free; | |
754 | ||
755 | amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib, start, last + 1, | |
756 | addr, flags); | |
d38ceaf9 | 757 | |
d5fc5e82 CZ |
758 | amdgpu_vm_pad_ib(adev, ib); |
759 | WARN_ON(ib->length_dw > ndw); | |
4af9f07c CZ |
760 | r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1, |
761 | &amdgpu_vm_free_job, | |
762 | AMDGPU_FENCE_OWNER_VM, | |
763 | &f); | |
764 | if (r) | |
765 | goto error_free; | |
d38ceaf9 | 766 | |
bf60efd3 | 767 | amdgpu_bo_fence(vm->page_directory, f, true); |
4af9f07c CZ |
768 | if (fence) { |
769 | fence_put(*fence); | |
770 | *fence = fence_get(f); | |
771 | } | |
281b4223 | 772 | fence_put(f); |
d38ceaf9 | 773 | return 0; |
d5fc5e82 CZ |
774 | |
775 | error_free: | |
d5fc5e82 CZ |
776 | amdgpu_ib_free(adev, ib); |
777 | kfree(ib); | |
4af9f07c | 778 | return r; |
d38ceaf9 AD |
779 | } |
780 | ||
a14faa65 CK |
781 | /** |
782 | * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks | |
783 | * | |
784 | * @adev: amdgpu_device pointer | |
785 | * @gtt: GART instance to use for mapping | |
786 | * @vm: requested vm | |
787 | * @mapping: mapped range and flags to use for the update | |
788 | * @addr: addr to set the area to | |
789 | * @gtt_flags: flags as they are used for GTT | |
790 | * @fence: optional resulting fence | |
791 | * | |
792 | * Split the mapping into smaller chunks so that each update fits | |
793 | * into a SDMA IB. | |
794 | * Returns 0 for success, -EINVAL for failure. | |
795 | */ | |
796 | static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev, | |
797 | struct amdgpu_gart *gtt, | |
798 | uint32_t gtt_flags, | |
799 | struct amdgpu_vm *vm, | |
800 | struct amdgpu_bo_va_mapping *mapping, | |
801 | uint64_t addr, struct fence **fence) | |
802 | { | |
803 | const uint64_t max_size = 64ULL * 1024ULL * 1024ULL / AMDGPU_GPU_PAGE_SIZE; | |
804 | ||
805 | uint64_t start = mapping->it.start; | |
806 | uint32_t flags = gtt_flags; | |
807 | int r; | |
808 | ||
809 | /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here | |
810 | * but in case of something, we filter the flags in first place | |
811 | */ | |
812 | if (!(mapping->flags & AMDGPU_PTE_READABLE)) | |
813 | flags &= ~AMDGPU_PTE_READABLE; | |
814 | if (!(mapping->flags & AMDGPU_PTE_WRITEABLE)) | |
815 | flags &= ~AMDGPU_PTE_WRITEABLE; | |
816 | ||
817 | trace_amdgpu_vm_bo_update(mapping); | |
818 | ||
819 | addr += mapping->offset; | |
820 | ||
821 | if (!gtt || ((gtt == &adev->gart) && (flags == gtt_flags))) | |
822 | return amdgpu_vm_bo_update_mapping(adev, gtt, gtt_flags, vm, | |
823 | start, mapping->it.last, | |
824 | flags, addr, fence); | |
825 | ||
826 | while (start != mapping->it.last + 1) { | |
827 | uint64_t last; | |
828 | ||
829 | last = min((uint64_t)mapping->it.last, start + max_size); | |
830 | r = amdgpu_vm_bo_update_mapping(adev, gtt, gtt_flags, vm, | |
831 | start, last, flags, addr, | |
832 | fence); | |
833 | if (r) | |
834 | return r; | |
835 | ||
836 | start = last + 1; | |
837 | addr += max_size; | |
838 | } | |
839 | ||
840 | return 0; | |
841 | } | |
842 | ||
d38ceaf9 AD |
843 | /** |
844 | * amdgpu_vm_bo_update - update all BO mappings in the vm page table | |
845 | * | |
846 | * @adev: amdgpu_device pointer | |
847 | * @bo_va: requested BO and VM object | |
848 | * @mem: ttm mem | |
849 | * | |
850 | * Fill in the page table entries for @bo_va. | |
851 | * Returns 0 for success, -EINVAL for failure. | |
852 | * | |
853 | * Object have to be reserved and mutex must be locked! | |
854 | */ | |
855 | int amdgpu_vm_bo_update(struct amdgpu_device *adev, | |
856 | struct amdgpu_bo_va *bo_va, | |
857 | struct ttm_mem_reg *mem) | |
858 | { | |
859 | struct amdgpu_vm *vm = bo_va->vm; | |
860 | struct amdgpu_bo_va_mapping *mapping; | |
9ab21462 | 861 | struct amdgpu_gart *gtt = NULL; |
d38ceaf9 AD |
862 | uint32_t flags; |
863 | uint64_t addr; | |
864 | int r; | |
865 | ||
866 | if (mem) { | |
b7d698d7 | 867 | addr = (u64)mem->start << PAGE_SHIFT; |
9ab21462 CK |
868 | switch (mem->mem_type) { |
869 | case TTM_PL_TT: | |
870 | gtt = &bo_va->bo->adev->gart; | |
871 | break; | |
872 | ||
873 | case TTM_PL_VRAM: | |
d38ceaf9 | 874 | addr += adev->vm_manager.vram_base_offset; |
9ab21462 CK |
875 | break; |
876 | ||
877 | default: | |
878 | break; | |
879 | } | |
d38ceaf9 AD |
880 | } else { |
881 | addr = 0; | |
882 | } | |
883 | ||
d38ceaf9 AD |
884 | flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem); |
885 | ||
7fc11959 CK |
886 | spin_lock(&vm->status_lock); |
887 | if (!list_empty(&bo_va->vm_status)) | |
888 | list_splice_init(&bo_va->valids, &bo_va->invalids); | |
889 | spin_unlock(&vm->status_lock); | |
890 | ||
891 | list_for_each_entry(mapping, &bo_va->invalids, list) { | |
a14faa65 CK |
892 | r = amdgpu_vm_bo_split_mapping(adev, gtt, flags, vm, mapping, addr, |
893 | &bo_va->last_pt_update); | |
d38ceaf9 AD |
894 | if (r) |
895 | return r; | |
896 | } | |
897 | ||
d6c10f6b CK |
898 | if (trace_amdgpu_vm_bo_mapping_enabled()) { |
899 | list_for_each_entry(mapping, &bo_va->valids, list) | |
900 | trace_amdgpu_vm_bo_mapping(mapping); | |
901 | ||
902 | list_for_each_entry(mapping, &bo_va->invalids, list) | |
903 | trace_amdgpu_vm_bo_mapping(mapping); | |
904 | } | |
905 | ||
d38ceaf9 | 906 | spin_lock(&vm->status_lock); |
6d1d0ef7 | 907 | list_splice_init(&bo_va->invalids, &bo_va->valids); |
d38ceaf9 | 908 | list_del_init(&bo_va->vm_status); |
7fc11959 CK |
909 | if (!mem) |
910 | list_add(&bo_va->vm_status, &vm->cleared); | |
d38ceaf9 AD |
911 | spin_unlock(&vm->status_lock); |
912 | ||
913 | return 0; | |
914 | } | |
915 | ||
916 | /** | |
917 | * amdgpu_vm_clear_freed - clear freed BOs in the PT | |
918 | * | |
919 | * @adev: amdgpu_device pointer | |
920 | * @vm: requested vm | |
921 | * | |
922 | * Make sure all freed BOs are cleared in the PT. | |
923 | * Returns 0 for success. | |
924 | * | |
925 | * PTs have to be reserved and mutex must be locked! | |
926 | */ | |
927 | int amdgpu_vm_clear_freed(struct amdgpu_device *adev, | |
928 | struct amdgpu_vm *vm) | |
929 | { | |
930 | struct amdgpu_bo_va_mapping *mapping; | |
931 | int r; | |
932 | ||
81d75a30 | 933 | spin_lock(&vm->freed_lock); |
d38ceaf9 AD |
934 | while (!list_empty(&vm->freed)) { |
935 | mapping = list_first_entry(&vm->freed, | |
936 | struct amdgpu_bo_va_mapping, list); | |
937 | list_del(&mapping->list); | |
81d75a30 | 938 | spin_unlock(&vm->freed_lock); |
a14faa65 CK |
939 | r = amdgpu_vm_bo_split_mapping(adev, NULL, 0, vm, mapping, |
940 | 0, NULL); | |
d38ceaf9 AD |
941 | kfree(mapping); |
942 | if (r) | |
943 | return r; | |
944 | ||
81d75a30 | 945 | spin_lock(&vm->freed_lock); |
d38ceaf9 | 946 | } |
81d75a30 | 947 | spin_unlock(&vm->freed_lock); |
948 | ||
d38ceaf9 AD |
949 | return 0; |
950 | ||
951 | } | |
952 | ||
953 | /** | |
954 | * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT | |
955 | * | |
956 | * @adev: amdgpu_device pointer | |
957 | * @vm: requested vm | |
958 | * | |
959 | * Make sure all invalidated BOs are cleared in the PT. | |
960 | * Returns 0 for success. | |
961 | * | |
962 | * PTs have to be reserved and mutex must be locked! | |
963 | */ | |
964 | int amdgpu_vm_clear_invalids(struct amdgpu_device *adev, | |
cfe2c978 | 965 | struct amdgpu_vm *vm, struct amdgpu_sync *sync) |
d38ceaf9 | 966 | { |
cfe2c978 | 967 | struct amdgpu_bo_va *bo_va = NULL; |
91e1a520 | 968 | int r = 0; |
d38ceaf9 AD |
969 | |
970 | spin_lock(&vm->status_lock); | |
971 | while (!list_empty(&vm->invalidated)) { | |
972 | bo_va = list_first_entry(&vm->invalidated, | |
973 | struct amdgpu_bo_va, vm_status); | |
974 | spin_unlock(&vm->status_lock); | |
69b576a1 | 975 | mutex_lock(&bo_va->mutex); |
d38ceaf9 | 976 | r = amdgpu_vm_bo_update(adev, bo_va, NULL); |
69b576a1 | 977 | mutex_unlock(&bo_va->mutex); |
d38ceaf9 AD |
978 | if (r) |
979 | return r; | |
980 | ||
981 | spin_lock(&vm->status_lock); | |
982 | } | |
983 | spin_unlock(&vm->status_lock); | |
984 | ||
cfe2c978 | 985 | if (bo_va) |
bb1e38a4 | 986 | r = amdgpu_sync_fence(adev, sync, bo_va->last_pt_update); |
91e1a520 CK |
987 | |
988 | return r; | |
d38ceaf9 AD |
989 | } |
990 | ||
991 | /** | |
992 | * amdgpu_vm_bo_add - add a bo to a specific vm | |
993 | * | |
994 | * @adev: amdgpu_device pointer | |
995 | * @vm: requested vm | |
996 | * @bo: amdgpu buffer object | |
997 | * | |
8843dbbb | 998 | * Add @bo into the requested vm. |
d38ceaf9 AD |
999 | * Add @bo to the list of bos associated with the vm |
1000 | * Returns newly added bo_va or NULL for failure | |
1001 | * | |
1002 | * Object has to be reserved! | |
1003 | */ | |
1004 | struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, | |
1005 | struct amdgpu_vm *vm, | |
1006 | struct amdgpu_bo *bo) | |
1007 | { | |
1008 | struct amdgpu_bo_va *bo_va; | |
1009 | ||
1010 | bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL); | |
1011 | if (bo_va == NULL) { | |
1012 | return NULL; | |
1013 | } | |
1014 | bo_va->vm = vm; | |
1015 | bo_va->bo = bo; | |
d38ceaf9 AD |
1016 | bo_va->ref_count = 1; |
1017 | INIT_LIST_HEAD(&bo_va->bo_list); | |
7fc11959 CK |
1018 | INIT_LIST_HEAD(&bo_va->valids); |
1019 | INIT_LIST_HEAD(&bo_va->invalids); | |
d38ceaf9 | 1020 | INIT_LIST_HEAD(&bo_va->vm_status); |
69b576a1 | 1021 | mutex_init(&bo_va->mutex); |
d38ceaf9 | 1022 | list_add_tail(&bo_va->bo_list, &bo->va); |
d38ceaf9 AD |
1023 | |
1024 | return bo_va; | |
1025 | } | |
1026 | ||
1027 | /** | |
1028 | * amdgpu_vm_bo_map - map bo inside a vm | |
1029 | * | |
1030 | * @adev: amdgpu_device pointer | |
1031 | * @bo_va: bo_va to store the address | |
1032 | * @saddr: where to map the BO | |
1033 | * @offset: requested offset in the BO | |
1034 | * @flags: attributes of pages (read/write/valid/etc.) | |
1035 | * | |
1036 | * Add a mapping of the BO at the specefied addr into the VM. | |
1037 | * Returns 0 for success, error for failure. | |
1038 | * | |
49b02b18 | 1039 | * Object has to be reserved and unreserved outside! |
d38ceaf9 AD |
1040 | */ |
1041 | int amdgpu_vm_bo_map(struct amdgpu_device *adev, | |
1042 | struct amdgpu_bo_va *bo_va, | |
1043 | uint64_t saddr, uint64_t offset, | |
1044 | uint64_t size, uint32_t flags) | |
1045 | { | |
1046 | struct amdgpu_bo_va_mapping *mapping; | |
1047 | struct amdgpu_vm *vm = bo_va->vm; | |
1048 | struct interval_tree_node *it; | |
1049 | unsigned last_pfn, pt_idx; | |
1050 | uint64_t eaddr; | |
1051 | int r; | |
1052 | ||
0be52de9 CK |
1053 | /* validate the parameters */ |
1054 | if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK || | |
49b02b18 | 1055 | size == 0 || size & AMDGPU_GPU_PAGE_MASK) |
0be52de9 | 1056 | return -EINVAL; |
0be52de9 | 1057 | |
d38ceaf9 | 1058 | /* make sure object fit at this offset */ |
005ae95e | 1059 | eaddr = saddr + size - 1; |
49b02b18 | 1060 | if ((saddr >= eaddr) || (offset + size > amdgpu_bo_size(bo_va->bo))) |
d38ceaf9 | 1061 | return -EINVAL; |
d38ceaf9 AD |
1062 | |
1063 | last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE; | |
005ae95e FK |
1064 | if (last_pfn >= adev->vm_manager.max_pfn) { |
1065 | dev_err(adev->dev, "va above limit (0x%08X >= 0x%08X)\n", | |
d38ceaf9 | 1066 | last_pfn, adev->vm_manager.max_pfn); |
d38ceaf9 AD |
1067 | return -EINVAL; |
1068 | } | |
1069 | ||
d38ceaf9 AD |
1070 | saddr /= AMDGPU_GPU_PAGE_SIZE; |
1071 | eaddr /= AMDGPU_GPU_PAGE_SIZE; | |
1072 | ||
c25867df | 1073 | spin_lock(&vm->it_lock); |
005ae95e | 1074 | it = interval_tree_iter_first(&vm->va, saddr, eaddr); |
c25867df | 1075 | spin_unlock(&vm->it_lock); |
d38ceaf9 AD |
1076 | if (it) { |
1077 | struct amdgpu_bo_va_mapping *tmp; | |
1078 | tmp = container_of(it, struct amdgpu_bo_va_mapping, it); | |
1079 | /* bo and tmp overlap, invalid addr */ | |
1080 | dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with " | |
1081 | "0x%010lx-0x%010lx\n", bo_va->bo, saddr, eaddr, | |
1082 | tmp->it.start, tmp->it.last + 1); | |
d38ceaf9 | 1083 | r = -EINVAL; |
f48b2659 | 1084 | goto error; |
d38ceaf9 AD |
1085 | } |
1086 | ||
1087 | mapping = kmalloc(sizeof(*mapping), GFP_KERNEL); | |
1088 | if (!mapping) { | |
d38ceaf9 | 1089 | r = -ENOMEM; |
f48b2659 | 1090 | goto error; |
d38ceaf9 AD |
1091 | } |
1092 | ||
1093 | INIT_LIST_HEAD(&mapping->list); | |
1094 | mapping->it.start = saddr; | |
005ae95e | 1095 | mapping->it.last = eaddr; |
d38ceaf9 AD |
1096 | mapping->offset = offset; |
1097 | mapping->flags = flags; | |
1098 | ||
69b576a1 | 1099 | mutex_lock(&bo_va->mutex); |
7fc11959 | 1100 | list_add(&mapping->list, &bo_va->invalids); |
69b576a1 | 1101 | mutex_unlock(&bo_va->mutex); |
c25867df | 1102 | spin_lock(&vm->it_lock); |
d38ceaf9 | 1103 | interval_tree_insert(&mapping->it, &vm->va); |
c25867df | 1104 | spin_unlock(&vm->it_lock); |
93e3e438 | 1105 | trace_amdgpu_vm_bo_map(bo_va, mapping); |
d38ceaf9 AD |
1106 | |
1107 | /* Make sure the page tables are allocated */ | |
1108 | saddr >>= amdgpu_vm_block_size; | |
1109 | eaddr >>= amdgpu_vm_block_size; | |
1110 | ||
1111 | BUG_ON(eaddr >= amdgpu_vm_num_pdes(adev)); | |
1112 | ||
1113 | if (eaddr > vm->max_pde_used) | |
1114 | vm->max_pde_used = eaddr; | |
1115 | ||
d38ceaf9 AD |
1116 | /* walk over the address space and allocate the page tables */ |
1117 | for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) { | |
bf60efd3 | 1118 | struct reservation_object *resv = vm->page_directory->tbo.resv; |
ee1782c3 | 1119 | struct amdgpu_bo_list_entry *entry; |
d38ceaf9 AD |
1120 | struct amdgpu_bo *pt; |
1121 | ||
ee1782c3 CK |
1122 | entry = &vm->page_tables[pt_idx].entry; |
1123 | if (entry->robj) | |
d38ceaf9 AD |
1124 | continue; |
1125 | ||
d38ceaf9 AD |
1126 | r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8, |
1127 | AMDGPU_GPU_PAGE_SIZE, true, | |
857d913d AD |
1128 | AMDGPU_GEM_DOMAIN_VRAM, |
1129 | AMDGPU_GEM_CREATE_NO_CPU_ACCESS, | |
bf60efd3 | 1130 | NULL, resv, &pt); |
49b02b18 | 1131 | if (r) |
d38ceaf9 | 1132 | goto error_free; |
49b02b18 | 1133 | |
82b9c55b CK |
1134 | /* Keep a reference to the page table to avoid freeing |
1135 | * them up in the wrong order. | |
1136 | */ | |
1137 | pt->parent = amdgpu_bo_ref(vm->page_directory); | |
1138 | ||
d38ceaf9 AD |
1139 | r = amdgpu_vm_clear_bo(adev, pt); |
1140 | if (r) { | |
1141 | amdgpu_bo_unref(&pt); | |
1142 | goto error_free; | |
1143 | } | |
1144 | ||
ee1782c3 | 1145 | entry->robj = pt; |
ee1782c3 CK |
1146 | entry->priority = 0; |
1147 | entry->tv.bo = &entry->robj->tbo; | |
1148 | entry->tv.shared = true; | |
d38ceaf9 | 1149 | vm->page_tables[pt_idx].addr = 0; |
d38ceaf9 AD |
1150 | } |
1151 | ||
d38ceaf9 AD |
1152 | return 0; |
1153 | ||
1154 | error_free: | |
d38ceaf9 | 1155 | list_del(&mapping->list); |
c25867df | 1156 | spin_lock(&vm->it_lock); |
d38ceaf9 | 1157 | interval_tree_remove(&mapping->it, &vm->va); |
c25867df | 1158 | spin_unlock(&vm->it_lock); |
93e3e438 | 1159 | trace_amdgpu_vm_bo_unmap(bo_va, mapping); |
d38ceaf9 AD |
1160 | kfree(mapping); |
1161 | ||
f48b2659 | 1162 | error: |
d38ceaf9 AD |
1163 | return r; |
1164 | } | |
1165 | ||
1166 | /** | |
1167 | * amdgpu_vm_bo_unmap - remove bo mapping from vm | |
1168 | * | |
1169 | * @adev: amdgpu_device pointer | |
1170 | * @bo_va: bo_va to remove the address from | |
1171 | * @saddr: where to the BO is mapped | |
1172 | * | |
1173 | * Remove a mapping of the BO at the specefied addr from the VM. | |
1174 | * Returns 0 for success, error for failure. | |
1175 | * | |
49b02b18 | 1176 | * Object has to be reserved and unreserved outside! |
d38ceaf9 AD |
1177 | */ |
1178 | int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, | |
1179 | struct amdgpu_bo_va *bo_va, | |
1180 | uint64_t saddr) | |
1181 | { | |
1182 | struct amdgpu_bo_va_mapping *mapping; | |
1183 | struct amdgpu_vm *vm = bo_va->vm; | |
7fc11959 | 1184 | bool valid = true; |
d38ceaf9 | 1185 | |
6c7fc503 | 1186 | saddr /= AMDGPU_GPU_PAGE_SIZE; |
69b576a1 | 1187 | mutex_lock(&bo_va->mutex); |
7fc11959 | 1188 | list_for_each_entry(mapping, &bo_va->valids, list) { |
d38ceaf9 AD |
1189 | if (mapping->it.start == saddr) |
1190 | break; | |
1191 | } | |
1192 | ||
7fc11959 CK |
1193 | if (&mapping->list == &bo_va->valids) { |
1194 | valid = false; | |
1195 | ||
1196 | list_for_each_entry(mapping, &bo_va->invalids, list) { | |
1197 | if (mapping->it.start == saddr) | |
1198 | break; | |
1199 | } | |
1200 | ||
69b576a1 CZ |
1201 | if (&mapping->list == &bo_va->invalids) { |
1202 | mutex_unlock(&bo_va->mutex); | |
7fc11959 | 1203 | return -ENOENT; |
69b576a1 | 1204 | } |
d38ceaf9 | 1205 | } |
69b576a1 | 1206 | mutex_unlock(&bo_va->mutex); |
d38ceaf9 | 1207 | list_del(&mapping->list); |
c25867df | 1208 | spin_lock(&vm->it_lock); |
d38ceaf9 | 1209 | interval_tree_remove(&mapping->it, &vm->va); |
c25867df | 1210 | spin_unlock(&vm->it_lock); |
93e3e438 | 1211 | trace_amdgpu_vm_bo_unmap(bo_va, mapping); |
d38ceaf9 | 1212 | |
81d75a30 | 1213 | if (valid) { |
1214 | spin_lock(&vm->freed_lock); | |
d38ceaf9 | 1215 | list_add(&mapping->list, &vm->freed); |
81d75a30 | 1216 | spin_unlock(&vm->freed_lock); |
1217 | } else { | |
d38ceaf9 | 1218 | kfree(mapping); |
81d75a30 | 1219 | } |
d38ceaf9 AD |
1220 | |
1221 | return 0; | |
1222 | } | |
1223 | ||
1224 | /** | |
1225 | * amdgpu_vm_bo_rmv - remove a bo to a specific vm | |
1226 | * | |
1227 | * @adev: amdgpu_device pointer | |
1228 | * @bo_va: requested bo_va | |
1229 | * | |
8843dbbb | 1230 | * Remove @bo_va->bo from the requested vm. |
d38ceaf9 AD |
1231 | * |
1232 | * Object have to be reserved! | |
1233 | */ | |
1234 | void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, | |
1235 | struct amdgpu_bo_va *bo_va) | |
1236 | { | |
1237 | struct amdgpu_bo_va_mapping *mapping, *next; | |
1238 | struct amdgpu_vm *vm = bo_va->vm; | |
1239 | ||
1240 | list_del(&bo_va->bo_list); | |
1241 | ||
d38ceaf9 AD |
1242 | spin_lock(&vm->status_lock); |
1243 | list_del(&bo_va->vm_status); | |
1244 | spin_unlock(&vm->status_lock); | |
1245 | ||
7fc11959 | 1246 | list_for_each_entry_safe(mapping, next, &bo_va->valids, list) { |
d38ceaf9 | 1247 | list_del(&mapping->list); |
c25867df | 1248 | spin_lock(&vm->it_lock); |
d38ceaf9 | 1249 | interval_tree_remove(&mapping->it, &vm->va); |
c25867df | 1250 | spin_unlock(&vm->it_lock); |
93e3e438 | 1251 | trace_amdgpu_vm_bo_unmap(bo_va, mapping); |
81d75a30 | 1252 | spin_lock(&vm->freed_lock); |
7fc11959 | 1253 | list_add(&mapping->list, &vm->freed); |
81d75a30 | 1254 | spin_unlock(&vm->freed_lock); |
7fc11959 CK |
1255 | } |
1256 | list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) { | |
1257 | list_del(&mapping->list); | |
c25867df | 1258 | spin_lock(&vm->it_lock); |
7fc11959 | 1259 | interval_tree_remove(&mapping->it, &vm->va); |
c25867df | 1260 | spin_unlock(&vm->it_lock); |
7fc11959 | 1261 | kfree(mapping); |
d38ceaf9 | 1262 | } |
bb1e38a4 | 1263 | fence_put(bo_va->last_pt_update); |
69b576a1 | 1264 | mutex_destroy(&bo_va->mutex); |
d38ceaf9 | 1265 | kfree(bo_va); |
d38ceaf9 AD |
1266 | } |
1267 | ||
1268 | /** | |
1269 | * amdgpu_vm_bo_invalidate - mark the bo as invalid | |
1270 | * | |
1271 | * @adev: amdgpu_device pointer | |
1272 | * @vm: requested vm | |
1273 | * @bo: amdgpu buffer object | |
1274 | * | |
8843dbbb | 1275 | * Mark @bo as invalid. |
d38ceaf9 AD |
1276 | */ |
1277 | void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, | |
1278 | struct amdgpu_bo *bo) | |
1279 | { | |
1280 | struct amdgpu_bo_va *bo_va; | |
1281 | ||
1282 | list_for_each_entry(bo_va, &bo->va, bo_list) { | |
7fc11959 CK |
1283 | spin_lock(&bo_va->vm->status_lock); |
1284 | if (list_empty(&bo_va->vm_status)) | |
d38ceaf9 | 1285 | list_add(&bo_va->vm_status, &bo_va->vm->invalidated); |
7fc11959 | 1286 | spin_unlock(&bo_va->vm->status_lock); |
d38ceaf9 AD |
1287 | } |
1288 | } | |
1289 | ||
1290 | /** | |
1291 | * amdgpu_vm_init - initialize a vm instance | |
1292 | * | |
1293 | * @adev: amdgpu_device pointer | |
1294 | * @vm: requested vm | |
1295 | * | |
8843dbbb | 1296 | * Init @vm fields. |
d38ceaf9 AD |
1297 | */ |
1298 | int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm) | |
1299 | { | |
1300 | const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE, | |
1301 | AMDGPU_VM_PTE_COUNT * 8); | |
9571e1d8 | 1302 | unsigned pd_size, pd_entries; |
d38ceaf9 AD |
1303 | int i, r; |
1304 | ||
1305 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
1306 | vm->ids[i].id = 0; | |
1307 | vm->ids[i].flushed_updates = NULL; | |
d38ceaf9 | 1308 | } |
d38ceaf9 AD |
1309 | vm->va = RB_ROOT; |
1310 | spin_lock_init(&vm->status_lock); | |
1311 | INIT_LIST_HEAD(&vm->invalidated); | |
7fc11959 | 1312 | INIT_LIST_HEAD(&vm->cleared); |
d38ceaf9 | 1313 | INIT_LIST_HEAD(&vm->freed); |
c25867df | 1314 | spin_lock_init(&vm->it_lock); |
81d75a30 | 1315 | spin_lock_init(&vm->freed_lock); |
d38ceaf9 AD |
1316 | pd_size = amdgpu_vm_directory_size(adev); |
1317 | pd_entries = amdgpu_vm_num_pdes(adev); | |
1318 | ||
1319 | /* allocate page table array */ | |
9571e1d8 | 1320 | vm->page_tables = drm_calloc_large(pd_entries, sizeof(struct amdgpu_vm_pt)); |
d38ceaf9 AD |
1321 | if (vm->page_tables == NULL) { |
1322 | DRM_ERROR("Cannot allocate memory for page table array\n"); | |
1323 | return -ENOMEM; | |
1324 | } | |
1325 | ||
05906dec BN |
1326 | vm->page_directory_fence = NULL; |
1327 | ||
d38ceaf9 | 1328 | r = amdgpu_bo_create(adev, pd_size, align, true, |
857d913d AD |
1329 | AMDGPU_GEM_DOMAIN_VRAM, |
1330 | AMDGPU_GEM_CREATE_NO_CPU_ACCESS, | |
72d7668b | 1331 | NULL, NULL, &vm->page_directory); |
d38ceaf9 AD |
1332 | if (r) |
1333 | return r; | |
ef9f0a83 CZ |
1334 | r = amdgpu_bo_reserve(vm->page_directory, false); |
1335 | if (r) { | |
1336 | amdgpu_bo_unref(&vm->page_directory); | |
1337 | vm->page_directory = NULL; | |
1338 | return r; | |
1339 | } | |
d38ceaf9 | 1340 | r = amdgpu_vm_clear_bo(adev, vm->page_directory); |
ef9f0a83 | 1341 | amdgpu_bo_unreserve(vm->page_directory); |
d38ceaf9 AD |
1342 | if (r) { |
1343 | amdgpu_bo_unref(&vm->page_directory); | |
1344 | vm->page_directory = NULL; | |
1345 | return r; | |
1346 | } | |
1347 | ||
1348 | return 0; | |
1349 | } | |
1350 | ||
1351 | /** | |
1352 | * amdgpu_vm_fini - tear down a vm instance | |
1353 | * | |
1354 | * @adev: amdgpu_device pointer | |
1355 | * @vm: requested vm | |
1356 | * | |
8843dbbb | 1357 | * Tear down @vm. |
d38ceaf9 AD |
1358 | * Unbind the VM and remove all bos from the vm bo list |
1359 | */ | |
1360 | void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) | |
1361 | { | |
1362 | struct amdgpu_bo_va_mapping *mapping, *tmp; | |
1363 | int i; | |
1364 | ||
1365 | if (!RB_EMPTY_ROOT(&vm->va)) { | |
1366 | dev_err(adev->dev, "still active bo inside vm\n"); | |
1367 | } | |
1368 | rbtree_postorder_for_each_entry_safe(mapping, tmp, &vm->va, it.rb) { | |
1369 | list_del(&mapping->list); | |
1370 | interval_tree_remove(&mapping->it, &vm->va); | |
1371 | kfree(mapping); | |
1372 | } | |
1373 | list_for_each_entry_safe(mapping, tmp, &vm->freed, list) { | |
1374 | list_del(&mapping->list); | |
1375 | kfree(mapping); | |
1376 | } | |
1377 | ||
1378 | for (i = 0; i < amdgpu_vm_num_pdes(adev); i++) | |
ee1782c3 | 1379 | amdgpu_bo_unref(&vm->page_tables[i].entry.robj); |
9571e1d8 | 1380 | drm_free_large(vm->page_tables); |
d38ceaf9 AD |
1381 | |
1382 | amdgpu_bo_unref(&vm->page_directory); | |
05906dec | 1383 | fence_put(vm->page_directory_fence); |
d38ceaf9 | 1384 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
1c16c0a7 CK |
1385 | unsigned id = vm->ids[i].id; |
1386 | ||
1387 | atomic_long_cmpxchg(&adev->vm_manager.ids[id].owner, | |
1388 | (long)vm, 0); | |
3c62338c | 1389 | fence_put(vm->ids[i].flushed_updates); |
d38ceaf9 AD |
1390 | } |
1391 | ||
d38ceaf9 | 1392 | } |
ea89f8c9 | 1393 | |
a9a78b32 CK |
1394 | /** |
1395 | * amdgpu_vm_manager_init - init the VM manager | |
1396 | * | |
1397 | * @adev: amdgpu_device pointer | |
1398 | * | |
1399 | * Initialize the VM manager structures | |
1400 | */ | |
1401 | void amdgpu_vm_manager_init(struct amdgpu_device *adev) | |
1402 | { | |
1403 | unsigned i; | |
1404 | ||
1405 | INIT_LIST_HEAD(&adev->vm_manager.ids_lru); | |
1406 | ||
1407 | /* skip over VMID 0, since it is the system VM */ | |
1408 | for (i = 1; i < adev->vm_manager.num_ids; ++i) | |
1409 | list_add_tail(&adev->vm_manager.ids[i].list, | |
1410 | &adev->vm_manager.ids_lru); | |
1411 | } | |
1412 | ||
ea89f8c9 CK |
1413 | /** |
1414 | * amdgpu_vm_manager_fini - cleanup VM manager | |
1415 | * | |
1416 | * @adev: amdgpu_device pointer | |
1417 | * | |
1418 | * Cleanup the VM manager and free resources. | |
1419 | */ | |
1420 | void amdgpu_vm_manager_fini(struct amdgpu_device *adev) | |
1421 | { | |
1422 | unsigned i; | |
1423 | ||
1424 | for (i = 0; i < AMDGPU_NUM_VM; ++i) | |
1c16c0a7 | 1425 | fence_put(adev->vm_manager.ids[i].active); |
ea89f8c9 | 1426 | } |