2 * Copyright 2019 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
24 #ifndef __AMDGPU_MES_H__
25 #define __AMDGPU_MES_H__
27 #include "amdgpu_irq.h"
28 #include "kgd_kfd_interface.h"
29 #include "amdgpu_gfx.h"
30 #include <linux/sched/mm.h>
32 #define AMDGPU_MES_MAX_COMPUTE_PIPES 8
33 #define AMDGPU_MES_MAX_GFX_PIPES 2
34 #define AMDGPU_MES_MAX_SDMA_PIPES 2
36 #define AMDGPU_MES_API_VERSION_SHIFT 12
37 #define AMDGPU_MES_FEAT_VERSION_SHIFT 24
39 #define AMDGPU_MES_VERSION_MASK 0x00000fff
40 #define AMDGPU_MES_API_VERSION_MASK 0x00fff000
41 #define AMDGPU_MES_FEAT_VERSION_MASK 0xff000000
43 enum amdgpu_mes_priority_level {
44 AMDGPU_MES_PRIORITY_LEVEL_LOW = 0,
45 AMDGPU_MES_PRIORITY_LEVEL_NORMAL = 1,
46 AMDGPU_MES_PRIORITY_LEVEL_MEDIUM = 2,
47 AMDGPU_MES_PRIORITY_LEVEL_HIGH = 3,
48 AMDGPU_MES_PRIORITY_LEVEL_REALTIME = 4,
49 AMDGPU_MES_PRIORITY_NUM_LEVELS
52 #define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */
53 #define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */
55 struct amdgpu_mes_funcs;
57 enum admgpu_mes_pipe {
58 AMDGPU_MES_SCHED_PIPE = 0,
60 AMDGPU_MAX_MES_PIPES = 2,
64 struct amdgpu_device *adev;
66 struct mutex mutex_hidden;
69 struct idr gang_id_idr;
70 struct idr queue_id_idr;
71 struct ida doorbell_ida;
73 spinlock_t queue_id_lock;
75 uint32_t sched_version;
78 uint32_t total_max_queue;
79 uint32_t doorbell_id_offset;
80 uint32_t max_doorbell_slices;
82 uint64_t default_process_quantum;
83 uint64_t default_gang_quantum;
85 struct amdgpu_ring ring;
88 const struct firmware *fw[AMDGPU_MAX_MES_PIPES];
91 struct amdgpu_bo *ucode_fw_obj[AMDGPU_MAX_MES_PIPES];
92 uint64_t ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
93 uint32_t *ucode_fw_ptr[AMDGPU_MAX_MES_PIPES];
94 uint32_t ucode_fw_version[AMDGPU_MAX_MES_PIPES];
95 uint64_t uc_start_addr[AMDGPU_MAX_MES_PIPES];
98 struct amdgpu_bo *data_fw_obj[AMDGPU_MAX_MES_PIPES];
99 uint64_t data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
100 uint32_t *data_fw_ptr[AMDGPU_MAX_MES_PIPES];
101 uint32_t data_fw_version[AMDGPU_MAX_MES_PIPES];
102 uint64_t data_start_addr[AMDGPU_MAX_MES_PIPES];
105 struct amdgpu_bo *eop_gpu_obj[AMDGPU_MAX_MES_PIPES];
106 uint64_t eop_gpu_addr[AMDGPU_MAX_MES_PIPES];
108 void *mqd_backup[AMDGPU_MAX_MES_PIPES];
109 struct amdgpu_irq_src irq[AMDGPU_MAX_MES_PIPES];
111 uint32_t vmid_mask_gfxhub;
112 uint32_t vmid_mask_mmhub;
113 uint32_t compute_hqd_mask[AMDGPU_MES_MAX_COMPUTE_PIPES];
114 uint32_t gfx_hqd_mask[AMDGPU_MES_MAX_GFX_PIPES];
115 uint32_t sdma_hqd_mask[AMDGPU_MES_MAX_SDMA_PIPES];
116 uint32_t aggregated_doorbells[AMDGPU_MES_PRIORITY_NUM_LEVELS];
117 uint32_t sch_ctx_offs;
118 uint64_t sch_ctx_gpu_addr;
119 uint64_t *sch_ctx_ptr;
120 uint32_t query_status_fence_offs;
121 uint64_t query_status_fence_gpu_addr;
122 uint64_t *query_status_fence_ptr;
123 uint32_t read_val_offs;
124 uint64_t read_val_gpu_addr;
125 uint32_t *read_val_ptr;
127 uint32_t saved_flags;
129 /* initialize kiq pipe */
130 int (*kiq_hw_init)(struct amdgpu_device *adev);
131 int (*kiq_hw_fini)(struct amdgpu_device *adev);
133 /* ip specific functions */
134 const struct amdgpu_mes_funcs *funcs;
137 struct amdgpu_mes_process {
139 struct amdgpu_vm *vm;
140 uint64_t pd_gpu_addr;
141 struct amdgpu_bo *proc_ctx_bo;
142 uint64_t proc_ctx_gpu_addr;
143 void *proc_ctx_cpu_ptr;
144 uint64_t process_quantum;
145 struct list_head gang_list;
146 uint32_t doorbell_index;
147 unsigned long *doorbell_bitmap;
148 struct mutex doorbell_lock;
151 struct amdgpu_mes_gang {
154 int inprocess_gang_priority;
155 int global_priority_level;
156 struct list_head list;
157 struct amdgpu_mes_process *process;
158 struct amdgpu_bo *gang_ctx_bo;
159 uint64_t gang_ctx_gpu_addr;
160 void *gang_ctx_cpu_ptr;
161 uint64_t gang_quantum;
162 struct list_head queue_list;
165 struct amdgpu_mes_queue {
166 struct list_head list;
167 struct amdgpu_mes_gang *gang;
169 uint64_t doorbell_off;
170 struct amdgpu_bo *mqd_obj;
172 uint64_t mqd_gpu_addr;
173 uint64_t wptr_gpu_addr;
176 struct amdgpu_ring *ring;
179 struct amdgpu_mes_queue_properties {
181 uint64_t hqd_base_gpu_addr;
182 uint64_t rptr_gpu_addr;
183 uint64_t wptr_gpu_addr;
184 uint64_t wptr_mc_addr;
186 uint64_t eop_gpu_addr;
187 uint32_t hqd_pipe_priority;
188 uint32_t hqd_queue_priority;
190 struct amdgpu_ring *ring;
192 uint64_t doorbell_off;
195 struct amdgpu_mes_gang_properties {
197 uint32_t gang_quantum;
198 uint32_t inprocess_gang_priority;
199 uint32_t priority_level;
200 int global_priority_level;
203 struct mes_add_queue_input {
205 uint64_t page_table_base_addr;
206 uint64_t process_va_start;
207 uint64_t process_va_end;
208 uint64_t process_quantum;
209 uint64_t process_context_addr;
210 uint64_t gang_quantum;
211 uint64_t gang_context_addr;
212 uint32_t inprocess_gang_priority;
213 uint32_t gang_global_priority_level;
214 uint32_t doorbell_offset;
217 uint64_t wptr_mc_addr;
224 uint32_t is_kfd_process;
225 uint32_t is_aql_queue;
229 struct mes_remove_queue_input {
230 uint32_t doorbell_offset;
231 uint64_t gang_context_addr;
234 struct mes_unmap_legacy_queue_input {
235 enum amdgpu_unmap_queues_action action;
237 uint32_t doorbell_offset;
240 uint64_t trail_fence_addr;
241 uint64_t trail_fence_data;
244 struct mes_suspend_gang_input {
245 bool suspend_all_gangs;
246 uint64_t gang_context_addr;
247 uint64_t suspend_fence_addr;
248 uint32_t suspend_fence_value;
251 struct mes_resume_gang_input {
252 bool resume_all_gangs;
253 uint64_t gang_context_addr;
256 enum mes_misc_opcode {
257 MES_MISC_OP_WRITE_REG,
258 MES_MISC_OP_READ_REG,
259 MES_MISC_OP_WRM_REG_WAIT,
260 MES_MISC_OP_WRM_REG_WR_WAIT,
263 struct mes_misc_op_input {
264 enum mes_misc_opcode op;
269 uint64_t buffer_addr;
286 struct amdgpu_mes_funcs {
287 int (*add_hw_queue)(struct amdgpu_mes *mes,
288 struct mes_add_queue_input *input);
290 int (*remove_hw_queue)(struct amdgpu_mes *mes,
291 struct mes_remove_queue_input *input);
293 int (*unmap_legacy_queue)(struct amdgpu_mes *mes,
294 struct mes_unmap_legacy_queue_input *input);
296 int (*suspend_gang)(struct amdgpu_mes *mes,
297 struct mes_suspend_gang_input *input);
299 int (*resume_gang)(struct amdgpu_mes *mes,
300 struct mes_resume_gang_input *input);
302 int (*misc_op)(struct amdgpu_mes *mes,
303 struct mes_misc_op_input *input);
306 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
307 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev))
309 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs);
311 int amdgpu_mes_init(struct amdgpu_device *adev);
312 void amdgpu_mes_fini(struct amdgpu_device *adev);
314 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
315 struct amdgpu_vm *vm);
316 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid);
318 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
319 struct amdgpu_mes_gang_properties *gprops,
321 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id);
323 int amdgpu_mes_suspend(struct amdgpu_device *adev);
324 int amdgpu_mes_resume(struct amdgpu_device *adev);
326 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
327 struct amdgpu_mes_queue_properties *qprops,
329 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id);
331 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
332 struct amdgpu_ring *ring,
333 enum amdgpu_unmap_queues_action action,
334 u64 gpu_addr, u64 seq);
336 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
337 int amdgpu_mes_wreg(struct amdgpu_device *adev,
338 uint32_t reg, uint32_t val);
339 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
340 uint32_t val, uint32_t mask);
341 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
342 uint32_t reg0, uint32_t reg1,
343 uint32_t ref, uint32_t mask);
345 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
346 int queue_type, int idx,
347 struct amdgpu_mes_ctx_data *ctx_data,
348 struct amdgpu_ring **out);
349 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
350 struct amdgpu_ring *ring);
352 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
353 enum amdgpu_mes_priority_level prio);
355 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
356 struct amdgpu_mes_ctx_data *ctx_data);
357 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data);
358 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
359 struct amdgpu_vm *vm,
360 struct amdgpu_mes_ctx_data *ctx_data);
361 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
362 struct amdgpu_mes_ctx_data *ctx_data);
364 int amdgpu_mes_self_test(struct amdgpu_device *adev);
366 int amdgpu_mes_alloc_process_doorbells(struct amdgpu_device *adev,
367 unsigned int *doorbell_index);
368 void amdgpu_mes_free_process_doorbells(struct amdgpu_device *adev,
369 unsigned int doorbell_index);
370 unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar(
371 struct amdgpu_device *adev,
372 uint32_t doorbell_index,
373 unsigned int doorbell_id);
374 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev);
377 * MES lock can be taken in MMU notifiers.
379 * A bit more detail about why to set no-FS reclaim with MES lock:
381 * The purpose of the MMU notifier is to stop GPU access to memory so
382 * that the Linux VM subsystem can move pages around safely. This is
383 * done by preempting user mode queues for the affected process. When
384 * MES is used, MES lock needs to be taken to preempt the queues.
386 * The MMU notifier callback entry point in the driver is
387 * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from
389 * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm ->
390 * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues
392 * The last part of the chain is a function pointer where we take the
395 * The problem with taking locks in the MMU notifier is, that MMU
396 * notifiers can be called in reclaim-FS context. That's where the
397 * kernel frees up pages to make room for new page allocations under
398 * memory pressure. While we are running in reclaim-FS context, we must
399 * not trigger another memory reclaim operation because that would
400 * recursively reenter the reclaim code and cause a deadlock. The
401 * memalloc_nofs_save/restore calls guarantee that.
403 * In addition we also need to avoid lock dependencies on other locks taken
404 * under the MES lock, for example reservation locks. Here is a possible
405 * scenario of a deadlock:
406 * Thread A: takes and holds reservation lock | triggers reclaim-FS |
407 * MMU notifier | blocks trying to take MES lock
408 * Thread B: takes and holds MES lock | blocks trying to take reservation lock
410 * In this scenario Thread B gets involved in a deadlock even without
411 * triggering a reclaim-FS operation itself.
412 * To fix this and break the lock dependency chain you'd need to either:
413 * 1. protect reservation locks with memalloc_nofs_save/restore, or
414 * 2. avoid taking reservation locks under the MES lock.
416 * Reservation locks are taken all over the kernel in different subsystems, we
417 * have no control over them and their lock dependencies.So the only workable
418 * solution is to avoid taking other locks under the MES lock.
419 * As a result, make sure no reclaim-FS happens while holding this lock anywhere
420 * to prevent deadlocks when an MMU notifier runs in reclaim-FS context.
422 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes)
424 mutex_lock(&mes->mutex_hidden);
425 mes->saved_flags = memalloc_noreclaim_save();
428 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes)
430 memalloc_noreclaim_restore(mes->saved_flags);
431 mutex_unlock(&mes->mutex_hidden);
433 #endif /* __AMDGPU_MES_H__ */