1// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 *
23 */
24
25#include <linux/types.h>
26#include <linux/mutex.h>
27#include <linux/slab.h>
28#include <linux/printk.h>
29#include <linux/sched.h>
30#include "kfd_kernel_queue.h"
31#include "kfd_priv.h"
32#include "kfd_device_queue_manager.h"
33#include "kfd_pm4_headers.h"
34#include "kfd_pm4_opcodes.h"
35#include "amdgpu_reset.h"
36
37#define PM4_COUNT_ZERO (((1 << 15) - 1) << 16)
38
39/* Initialize a kernel queue, including allocations of GART memory
40 * needed for the queue.
41 */
42static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev,
43 enum kfd_queue_type type, unsigned int queue_size)
44{
45 struct queue_properties prop;
46 int retval;
47 union PM4_MES_TYPE_3_HEADER nop;
48
49 pr_debug("Initializing queue type %d size %d\n", type, queue_size);
50
51 memset(&prop, 0, sizeof(prop));
52 memset(&nop, 0, sizeof(nop));
53
54 nop.opcode = IT_NOP;
55 nop.type = PM4_TYPE_3;
56 nop.u32all |= PM4_COUNT_ZERO;
57
58 kq->dev = dev;
59 kq->nop_packet = nop.u32all;
60 switch (type) {
61 case KFD_QUEUE_TYPE_DIQ:
62 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ];
63 break;
64 case KFD_QUEUE_TYPE_HIQ:
65 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
66 break;
67 default:
68 WARN(1, "Invalid queue type %d\n", type);
69 dev_err(dev->adev->dev, "Invalid queue type %d\n", type);
70 return false;
71 }
72
73 if (!kq->mqd_mgr)
74 return false;
75
76 prop.doorbell_ptr = kfd_get_kernel_doorbell(kfd: dev->kfd, doorbell_off: &prop.doorbell_off);
77
78 if (!prop.doorbell_ptr) {
79 dev_err(dev->adev->dev, "Failed to initialize doorbell");
80 goto err_get_kernel_doorbell;
81 }
82
83 retval = kfd_gtt_sa_allocate(node: dev, size: queue_size, mem_obj: &kq->pq);
84 if (retval != 0) {
85 dev_err(dev->adev->dev, "Failed to init pq queues size %d\n",
86 queue_size);
87 goto err_pq_allocate_vidmem;
88 }
89
90 kq->pq_kernel_addr = kq->pq->cpu_ptr;
91 kq->pq_gpu_addr = kq->pq->gpu_addr;
92
93 /* For CIK family asics, kq->eop_mem is not needed */
94 if (dev->adev->asic_type > CHIP_MULLINS) {
95 retval = kfd_gtt_sa_allocate(node: dev, PAGE_SIZE, mem_obj: &kq->eop_mem);
96 if (retval != 0)
97 goto err_eop_allocate_vidmem;
98
99 kq->eop_gpu_addr = kq->eop_mem->gpu_addr;
100 kq->eop_kernel_addr = kq->eop_mem->cpu_ptr;
101
102 memset(kq->eop_kernel_addr, 0, PAGE_SIZE);
103 }
104
105 retval = kfd_gtt_sa_allocate(node: dev, size: sizeof(*kq->rptr_kernel),
106 mem_obj: &kq->rptr_mem);
107
108 if (retval != 0)
109 goto err_rptr_allocate_vidmem;
110
111 kq->rptr_kernel = kq->rptr_mem->cpu_ptr;
112 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr;
113
114 retval = kfd_gtt_sa_allocate(node: dev, size: dev->kfd->device_info.doorbell_size,
115 mem_obj: &kq->wptr_mem);
116
117 if (retval != 0)
118 goto err_wptr_allocate_vidmem;
119
120 kq->wptr_kernel = kq->wptr_mem->cpu_ptr;
121 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr;
122
123 memset(kq->pq_kernel_addr, 0, queue_size);
124 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel));
125 memset(kq->wptr_kernel, 0, dev->kfd->device_info.doorbell_size);
126
127 prop.queue_size = queue_size;
128 prop.is_interop = false;
129 prop.is_gws = false;
130 prop.priority = 1;
131 prop.queue_percent = 100;
132 prop.type = type;
133 prop.vmid = 0;
134 prop.queue_address = kq->pq_gpu_addr;
135 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr;
136 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr;
137 prop.eop_ring_buffer_address = kq->eop_gpu_addr;
138 prop.eop_ring_buffer_size = PAGE_SIZE;
139
140 if (init_queue(q: &kq->queue, properties: &prop) != 0)
141 goto err_init_queue;
142
143 kq->queue->device = dev;
144 kq->queue->process = kfd_get_process(current);
145
146 kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev,
147 &kq->queue->properties);
148 if (!kq->queue->mqd_mem_obj)
149 goto err_allocate_mqd;
150 kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd,
151 kq->queue->mqd_mem_obj,
152 &kq->queue->gart_mqd_addr,
153 &kq->queue->properties);
154 /* assign HIQ to HQD */
155 if (type == KFD_QUEUE_TYPE_HIQ) {
156 pr_debug("Assigning hiq to hqd\n");
157 kq->queue->pipe = KFD_CIK_HIQ_PIPE;
158 kq->queue->queue = KFD_CIK_HIQ_QUEUE;
159 kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd,
160 kq->queue->pipe, kq->queue->queue,
161 &kq->queue->properties, NULL);
162 } else {
163 /* allocate fence for DIQ */
164
165 retval = kfd_gtt_sa_allocate(node: dev, size: sizeof(uint32_t),
166 mem_obj: &kq->fence_mem_obj);
167
168 if (retval != 0)
169 goto err_alloc_fence;
170
171 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
172 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
173 }
174
175 print_queue(q: kq->queue);
176
177 return true;
178err_alloc_fence:
179 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj);
180err_allocate_mqd:
181 uninit_queue(q: kq->queue);
182err_init_queue:
183 kfd_gtt_sa_free(node: dev, mem_obj: kq->wptr_mem);
184err_wptr_allocate_vidmem:
185 kfd_gtt_sa_free(node: dev, mem_obj: kq->rptr_mem);
186err_rptr_allocate_vidmem:
187 kfd_gtt_sa_free(node: dev, mem_obj: kq->eop_mem);
188err_eop_allocate_vidmem:
189 kfd_gtt_sa_free(node: dev, mem_obj: kq->pq);
190err_pq_allocate_vidmem:
191 kfd_release_kernel_doorbell(kfd: dev->kfd, db_addr: prop.doorbell_ptr);
192err_get_kernel_doorbell:
193 return false;
194
195}
196
197/* Uninitialize a kernel queue and free all its memory usages. */
198static void kq_uninitialize(struct kernel_queue *kq)
199{
200 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && down_read_trylock(sem: &kq->dev->adev->reset_domain->sem)) {
201 kq->mqd_mgr->destroy_mqd(kq->mqd_mgr,
202 kq->queue->mqd,
203 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
204 KFD_UNMAP_LATENCY_MS,
205 kq->queue->pipe,
206 kq->queue->queue);
207 up_read(sem: &kq->dev->adev->reset_domain->sem);
208 }
209 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ)
210 kfd_gtt_sa_free(node: kq->dev, mem_obj: kq->fence_mem_obj);
211
212 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd,
213 kq->queue->mqd_mem_obj);
214
215 kfd_gtt_sa_free(node: kq->dev, mem_obj: kq->rptr_mem);
216 kfd_gtt_sa_free(node: kq->dev, mem_obj: kq->wptr_mem);
217
218 /* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free()
219 * is able to handle NULL properly.
220 */
221 kfd_gtt_sa_free(node: kq->dev, mem_obj: kq->eop_mem);
222
223 kfd_gtt_sa_free(node: kq->dev, mem_obj: kq->pq);
224 kfd_release_kernel_doorbell(kfd: kq->dev->kfd,
225 db_addr: kq->queue->properties.doorbell_ptr);
226 uninit_queue(q: kq->queue);
227}
228
229int kq_acquire_packet_buffer(struct kernel_queue *kq,
230 size_t packet_size_in_dwords, unsigned int **buffer_ptr)
231{
232 size_t available_size;
233 size_t queue_size_dwords;
234 uint32_t wptr, rptr;
235 uint64_t wptr64;
236 unsigned int *queue_address;
237
238 /* When rptr == wptr, the buffer is empty.
239 * When rptr == wptr + 1, the buffer is full.
240 * It is always rptr that advances to the position of wptr, rather than
241 * the opposite. So we can only use up to queue_size_dwords - 1 dwords.
242 */
243 rptr = *kq->rptr_kernel;
244 wptr = kq->pending_wptr;
245 wptr64 = kq->pending_wptr64;
246 queue_address = (unsigned int *)kq->pq_kernel_addr;
247 queue_size_dwords = kq->queue->properties.queue_size / 4;
248
249 pr_debug("rptr: %d\n", rptr);
250 pr_debug("wptr: %d\n", wptr);
251 pr_debug("queue_address 0x%p\n", queue_address);
252
253 available_size = (rptr + queue_size_dwords - 1 - wptr) %
254 queue_size_dwords;
255
256 if (packet_size_in_dwords > available_size) {
257 /*
258 * make sure calling functions know
259 * acquire_packet_buffer() failed
260 */
261 goto err_no_space;
262 }
263
264 if (wptr + packet_size_in_dwords >= queue_size_dwords) {
265 /* make sure after rolling back to position 0, there is
266 * still enough space.
267 */
268 if (packet_size_in_dwords >= rptr)
269 goto err_no_space;
270
271 /* fill nops, roll back and start at position 0 */
272 while (wptr > 0) {
273 queue_address[wptr] = kq->nop_packet;
274 wptr = (wptr + 1) % queue_size_dwords;
275 wptr64++;
276 }
277 }
278
279 *buffer_ptr = &queue_address[wptr];
280 kq->pending_wptr = wptr + packet_size_in_dwords;
281 kq->pending_wptr64 = wptr64 + packet_size_in_dwords;
282
283 return 0;
284
285err_no_space:
286 *buffer_ptr = NULL;
287 return -ENOMEM;
288}
289
290int kq_submit_packet(struct kernel_queue *kq)
291{
292#ifdef DEBUG
293 int i;
294
295 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) {
296 pr_debug("0x%2X ", kq->pq_kernel_addr[i]);
297 if (i % 15 == 0)
298 pr_debug("\n");
299 }
300 pr_debug("\n");
301#endif
302 /* Fatal err detected, packet submission won't go through */
303 if (amdgpu_amdkfd_is_fed(adev: kq->dev->adev))
304 return -EIO;
305
306 /* Make sure ring buffer is updated before wptr updated */
307 mb();
308
309 if (kq->dev->kfd->device_info.doorbell_size == 8) {
310 *kq->wptr64_kernel = kq->pending_wptr64;
311 mb(); /* Make sure wptr updated before ring doorbell */
312 write_kernel_doorbell64(db: kq->queue->properties.doorbell_ptr,
313 value: kq->pending_wptr64);
314 } else {
315 *kq->wptr_kernel = kq->pending_wptr;
316 mb(); /* Make sure wptr updated before ring doorbell */
317 write_kernel_doorbell(db: kq->queue->properties.doorbell_ptr,
318 value: kq->pending_wptr);
319 }
320
321 return 0;
322}
323
324void kq_rollback_packet(struct kernel_queue *kq)
325{
326 if (kq->dev->kfd->device_info.doorbell_size == 8) {
327 kq->pending_wptr64 = *kq->wptr64_kernel;
328 kq->pending_wptr = *kq->wptr_kernel %
329 (kq->queue->properties.queue_size / 4);
330 } else {
331 kq->pending_wptr = *kq->wptr_kernel;
332 }
333}
334
335struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
336 enum kfd_queue_type type)
337{
338 struct kernel_queue *kq;
339
340 kq = kzalloc(sizeof(*kq), GFP_KERNEL);
341 if (!kq)
342 return NULL;
343
344 if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE))
345 return kq;
346
347 dev_err(dev->adev->dev, "Failed to init kernel queue\n");
348
349 kfree(objp: kq);
350 return NULL;
351}
352
353void kernel_queue_uninit(struct kernel_queue *kq)
354{
355 kq_uninitialize(kq);
356 kfree(objp: kq);
357}
358
359/* FIXME: Can this test be removed? */
360static __attribute__((unused)) void test_kq(struct kfd_node *dev)
361{
362 struct kernel_queue *kq;
363 uint32_t *buffer, i;
364 int retval;
365
366 dev_err(dev->adev->dev, "Starting kernel queue test\n");
367
368 kq = kernel_queue_init(dev, type: KFD_QUEUE_TYPE_HIQ);
369 if (unlikely(!kq)) {
370 dev_err(dev->adev->dev, " Failed to initialize HIQ\n");
371 dev_err(dev->adev->dev, "Kernel queue test failed\n");
372 return;
373 }
374
375 retval = kq_acquire_packet_buffer(kq, packet_size_in_dwords: 5, buffer_ptr: &buffer);
376 if (unlikely(retval != 0)) {
377 dev_err(dev->adev->dev, " Failed to acquire packet buffer\n");
378 dev_err(dev->adev->dev, "Kernel queue test failed\n");
379 return;
380 }
381 for (i = 0; i < 5; i++)
382 buffer[i] = kq->nop_packet;
383 kq_submit_packet(kq);
384
385 dev_err(dev->adev->dev, "Ending kernel queue test\n");
386}
387
388
389

source code of linux/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c