Loading...
1/*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
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:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
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.
21 *
22 */
23
24#include <linux/types.h>
25#include <linux/mutex.h>
26#include <linux/slab.h>
27#include <linux/printk.h>
28#include <linux/sched.h>
29#include "kfd_kernel_queue.h"
30#include "kfd_priv.h"
31#include "kfd_device_queue_manager.h"
32#include "kfd_pm4_headers.h"
33#include "kfd_pm4_opcodes.h"
34
35#define PM4_COUNT_ZERO (((1 << 15) - 1) << 16)
36
37static bool initialize(struct kernel_queue *kq, struct kfd_dev *dev,
38 enum kfd_queue_type type, unsigned int queue_size)
39{
40 struct queue_properties prop;
41 int retval;
42 union PM4_MES_TYPE_3_HEADER nop;
43
44 BUG_ON(!kq || !dev);
45 BUG_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ);
46
47 pr_debug("amdkfd: In func %s initializing queue type %d size %d\n",
48 __func__, KFD_QUEUE_TYPE_HIQ, queue_size);
49
50 nop.opcode = IT_NOP;
51 nop.type = PM4_TYPE_3;
52 nop.u32all |= PM4_COUNT_ZERO;
53
54 kq->dev = dev;
55 kq->nop_packet = nop.u32all;
56 switch (type) {
57 case KFD_QUEUE_TYPE_DIQ:
58 case KFD_QUEUE_TYPE_HIQ:
59 kq->mqd = dev->dqm->ops.get_mqd_manager(dev->dqm,
60 KFD_MQD_TYPE_HIQ);
61 break;
62 default:
63 BUG();
64 break;
65 }
66
67 if (kq->mqd == NULL)
68 return false;
69
70 prop.doorbell_ptr = kfd_get_kernel_doorbell(dev, &prop.doorbell_off);
71
72 if (prop.doorbell_ptr == NULL) {
73 pr_err("amdkfd: error init doorbell");
74 goto err_get_kernel_doorbell;
75 }
76
77 retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq);
78 if (retval != 0) {
79 pr_err("amdkfd: error init pq queues size (%d)\n", queue_size);
80 goto err_pq_allocate_vidmem;
81 }
82
83 kq->pq_kernel_addr = kq->pq->cpu_ptr;
84 kq->pq_gpu_addr = kq->pq->gpu_addr;
85
86 retval = kq->ops_asic_specific.initialize(kq, dev, type, queue_size);
87 if (retval == false)
88 goto err_eop_allocate_vidmem;
89
90 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel),
91 &kq->rptr_mem);
92
93 if (retval != 0)
94 goto err_rptr_allocate_vidmem;
95
96 kq->rptr_kernel = kq->rptr_mem->cpu_ptr;
97 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr;
98
99 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->wptr_kernel),
100 &kq->wptr_mem);
101
102 if (retval != 0)
103 goto err_wptr_allocate_vidmem;
104
105 kq->wptr_kernel = kq->wptr_mem->cpu_ptr;
106 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr;
107
108 memset(kq->pq_kernel_addr, 0, queue_size);
109 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel));
110 memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel));
111
112 prop.queue_size = queue_size;
113 prop.is_interop = false;
114 prop.priority = 1;
115 prop.queue_percent = 100;
116 prop.type = type;
117 prop.vmid = 0;
118 prop.queue_address = kq->pq_gpu_addr;
119 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr;
120 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr;
121 prop.eop_ring_buffer_address = kq->eop_gpu_addr;
122 prop.eop_ring_buffer_size = PAGE_SIZE;
123
124 if (init_queue(&kq->queue, prop) != 0)
125 goto err_init_queue;
126
127 kq->queue->device = dev;
128 kq->queue->process = kfd_get_process(current);
129
130 retval = kq->mqd->init_mqd(kq->mqd, &kq->queue->mqd,
131 &kq->queue->mqd_mem_obj,
132 &kq->queue->gart_mqd_addr,
133 &kq->queue->properties);
134 if (retval != 0)
135 goto err_init_mqd;
136
137 /* assign HIQ to HQD */
138 if (type == KFD_QUEUE_TYPE_HIQ) {
139 pr_debug("assigning hiq to hqd\n");
140 kq->queue->pipe = KFD_CIK_HIQ_PIPE;
141 kq->queue->queue = KFD_CIK_HIQ_QUEUE;
142 kq->mqd->load_mqd(kq->mqd, kq->queue->mqd, kq->queue->pipe,
143 kq->queue->queue, NULL);
144 } else {
145 /* allocate fence for DIQ */
146
147 retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t),
148 &kq->fence_mem_obj);
149
150 if (retval != 0)
151 goto err_alloc_fence;
152
153 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
154 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
155 }
156
157 print_queue(kq->queue);
158
159 return true;
160err_alloc_fence:
161err_init_mqd:
162 uninit_queue(kq->queue);
163err_init_queue:
164 kfd_gtt_sa_free(dev, kq->wptr_mem);
165err_wptr_allocate_vidmem:
166 kfd_gtt_sa_free(dev, kq->rptr_mem);
167err_rptr_allocate_vidmem:
168 kfd_gtt_sa_free(dev, kq->eop_mem);
169err_eop_allocate_vidmem:
170 kfd_gtt_sa_free(dev, kq->pq);
171err_pq_allocate_vidmem:
172 kfd_release_kernel_doorbell(dev, prop.doorbell_ptr);
173err_get_kernel_doorbell:
174 return false;
175
176}
177
178static void uninitialize(struct kernel_queue *kq)
179{
180 BUG_ON(!kq);
181
182 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ)
183 kq->mqd->destroy_mqd(kq->mqd,
184 NULL,
185 false,
186 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
187 kq->queue->pipe,
188 kq->queue->queue);
189 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ)
190 kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj);
191
192 kq->mqd->uninit_mqd(kq->mqd, kq->queue->mqd, kq->queue->mqd_mem_obj);
193
194 kfd_gtt_sa_free(kq->dev, kq->rptr_mem);
195 kfd_gtt_sa_free(kq->dev, kq->wptr_mem);
196 kq->ops_asic_specific.uninitialize(kq);
197 kfd_gtt_sa_free(kq->dev, kq->pq);
198 kfd_release_kernel_doorbell(kq->dev,
199 kq->queue->properties.doorbell_ptr);
200 uninit_queue(kq->queue);
201}
202
203static int acquire_packet_buffer(struct kernel_queue *kq,
204 size_t packet_size_in_dwords, unsigned int **buffer_ptr)
205{
206 size_t available_size;
207 size_t queue_size_dwords;
208 uint32_t wptr, rptr;
209 unsigned int *queue_address;
210
211 BUG_ON(!kq || !buffer_ptr);
212
213 rptr = *kq->rptr_kernel;
214 wptr = *kq->wptr_kernel;
215 queue_address = (unsigned int *)kq->pq_kernel_addr;
216 queue_size_dwords = kq->queue->properties.queue_size / sizeof(uint32_t);
217
218 pr_debug("rptr: %d\n", rptr);
219 pr_debug("wptr: %d\n", wptr);
220 pr_debug("queue_address 0x%p\n", queue_address);
221
222 available_size = (rptr - 1 - wptr + queue_size_dwords) %
223 queue_size_dwords;
224
225 if (packet_size_in_dwords >= queue_size_dwords ||
226 packet_size_in_dwords >= available_size) {
227 /*
228 * make sure calling functions know
229 * acquire_packet_buffer() failed
230 */
231 *buffer_ptr = NULL;
232 return -ENOMEM;
233 }
234
235 if (wptr + packet_size_in_dwords >= queue_size_dwords) {
236 while (wptr > 0) {
237 queue_address[wptr] = kq->nop_packet;
238 wptr = (wptr + 1) % queue_size_dwords;
239 }
240 }
241
242 *buffer_ptr = &queue_address[wptr];
243 kq->pending_wptr = wptr + packet_size_in_dwords;
244
245 return 0;
246}
247
248static void submit_packet(struct kernel_queue *kq)
249{
250#ifdef DEBUG
251 int i;
252#endif
253
254 BUG_ON(!kq);
255
256#ifdef DEBUG
257 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) {
258 pr_debug("0x%2X ", kq->pq_kernel_addr[i]);
259 if (i % 15 == 0)
260 pr_debug("\n");
261 }
262 pr_debug("\n");
263#endif
264
265 *kq->wptr_kernel = kq->pending_wptr;
266 write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
267 kq->pending_wptr);
268}
269
270static void rollback_packet(struct kernel_queue *kq)
271{
272 BUG_ON(!kq);
273 kq->pending_wptr = *kq->queue->properties.write_ptr;
274}
275
276struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
277 enum kfd_queue_type type)
278{
279 struct kernel_queue *kq;
280
281 BUG_ON(!dev);
282
283 kq = kzalloc(sizeof(struct kernel_queue), GFP_KERNEL);
284 if (!kq)
285 return NULL;
286
287 kq->ops.initialize = initialize;
288 kq->ops.uninitialize = uninitialize;
289 kq->ops.acquire_packet_buffer = acquire_packet_buffer;
290 kq->ops.submit_packet = submit_packet;
291 kq->ops.rollback_packet = rollback_packet;
292
293 switch (dev->device_info->asic_family) {
294 case CHIP_CARRIZO:
295 kernel_queue_init_vi(&kq->ops_asic_specific);
296 break;
297
298 case CHIP_KAVERI:
299 kernel_queue_init_cik(&kq->ops_asic_specific);
300 break;
301 }
302
303 if (kq->ops.initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE) == false) {
304 pr_err("amdkfd: failed to init kernel queue\n");
305 kfree(kq);
306 return NULL;
307 }
308 return kq;
309}
310
311void kernel_queue_uninit(struct kernel_queue *kq)
312{
313 BUG_ON(!kq);
314
315 kq->ops.uninitialize(kq);
316 kfree(kq);
317}
318
319static __attribute__((unused)) void test_kq(struct kfd_dev *dev)
320{
321 struct kernel_queue *kq;
322 uint32_t *buffer, i;
323 int retval;
324
325 BUG_ON(!dev);
326
327 pr_err("amdkfd: starting kernel queue test\n");
328
329 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ);
330 BUG_ON(!kq);
331
332 retval = kq->ops.acquire_packet_buffer(kq, 5, &buffer);
333 BUG_ON(retval != 0);
334 for (i = 0; i < 5; i++)
335 buffer[i] = kq->nop_packet;
336 kq->ops.submit_packet(kq);
337
338 pr_err("amdkfd: ending kernel queue test\n");
339}
340
341
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 if (WARN_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ))
50 return false;
51
52 pr_debug("Initializing queue type %d size %d\n", KFD_QUEUE_TYPE_HIQ,
53 queue_size);
54
55 memset(&prop, 0, sizeof(prop));
56 memset(&nop, 0, sizeof(nop));
57
58 nop.opcode = IT_NOP;
59 nop.type = PM4_TYPE_3;
60 nop.u32all |= PM4_COUNT_ZERO;
61
62 kq->dev = dev;
63 kq->nop_packet = nop.u32all;
64 switch (type) {
65 case KFD_QUEUE_TYPE_DIQ:
66 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ];
67 break;
68 case KFD_QUEUE_TYPE_HIQ:
69 kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
70 break;
71 default:
72 dev_err(dev->adev->dev, "Invalid queue type %d\n", type);
73 return false;
74 }
75
76 if (!kq->mqd_mgr)
77 return false;
78
79 prop.doorbell_ptr = kfd_get_kernel_doorbell(dev->kfd, &prop.doorbell_off);
80
81 if (!prop.doorbell_ptr) {
82 dev_err(dev->adev->dev, "Failed to initialize doorbell");
83 goto err_get_kernel_doorbell;
84 }
85
86 retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq);
87 if (retval != 0) {
88 dev_err(dev->adev->dev, "Failed to init pq queues size %d\n",
89 queue_size);
90 goto err_pq_allocate_vidmem;
91 }
92
93 kq->pq_kernel_addr = kq->pq->cpu_ptr;
94 kq->pq_gpu_addr = kq->pq->gpu_addr;
95
96 /* For CIK family asics, kq->eop_mem is not needed */
97 if (dev->adev->asic_type > CHIP_MULLINS) {
98 retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem);
99 if (retval != 0)
100 goto err_eop_allocate_vidmem;
101
102 kq->eop_gpu_addr = kq->eop_mem->gpu_addr;
103 kq->eop_kernel_addr = kq->eop_mem->cpu_ptr;
104
105 memset(kq->eop_kernel_addr, 0, PAGE_SIZE);
106 }
107
108 retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel),
109 &kq->rptr_mem);
110
111 if (retval != 0)
112 goto err_rptr_allocate_vidmem;
113
114 kq->rptr_kernel = kq->rptr_mem->cpu_ptr;
115 kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr;
116
117 retval = kfd_gtt_sa_allocate(dev, dev->kfd->device_info.doorbell_size,
118 &kq->wptr_mem);
119
120 if (retval != 0)
121 goto err_wptr_allocate_vidmem;
122
123 kq->wptr_kernel = kq->wptr_mem->cpu_ptr;
124 kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr;
125
126 memset(kq->pq_kernel_addr, 0, queue_size);
127 memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel));
128 memset(kq->wptr_kernel, 0, dev->kfd->device_info.doorbell_size);
129
130 prop.queue_size = queue_size;
131 prop.is_interop = false;
132 prop.is_gws = false;
133 prop.priority = 1;
134 prop.queue_percent = 100;
135 prop.type = type;
136 prop.vmid = 0;
137 prop.queue_address = kq->pq_gpu_addr;
138 prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr;
139 prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr;
140 prop.eop_ring_buffer_address = kq->eop_gpu_addr;
141 prop.eop_ring_buffer_size = PAGE_SIZE;
142
143 if (init_queue(&kq->queue, &prop) != 0)
144 goto err_init_queue;
145
146 kq->queue->device = dev;
147 kq->queue->process = kfd_get_process(current);
148
149 kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev,
150 &kq->queue->properties);
151 if (!kq->queue->mqd_mem_obj)
152 goto err_allocate_mqd;
153 kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd,
154 kq->queue->mqd_mem_obj,
155 &kq->queue->gart_mqd_addr,
156 &kq->queue->properties);
157 /* assign HIQ to HQD */
158 if (type == KFD_QUEUE_TYPE_HIQ) {
159 pr_debug("Assigning hiq to hqd\n");
160 kq->queue->pipe = KFD_CIK_HIQ_PIPE;
161 kq->queue->queue = KFD_CIK_HIQ_QUEUE;
162 kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd,
163 kq->queue->pipe, kq->queue->queue,
164 &kq->queue->properties, NULL);
165 } else {
166 /* allocate fence for DIQ */
167
168 retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t),
169 &kq->fence_mem_obj);
170
171 if (retval != 0)
172 goto err_alloc_fence;
173
174 kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
175 kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
176 }
177
178 print_queue(kq->queue);
179
180 return true;
181err_alloc_fence:
182 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj);
183err_allocate_mqd:
184 uninit_queue(kq->queue);
185err_init_queue:
186 kfd_gtt_sa_free(dev, kq->wptr_mem);
187err_wptr_allocate_vidmem:
188 kfd_gtt_sa_free(dev, kq->rptr_mem);
189err_rptr_allocate_vidmem:
190 kfd_gtt_sa_free(dev, kq->eop_mem);
191err_eop_allocate_vidmem:
192 kfd_gtt_sa_free(dev, kq->pq);
193err_pq_allocate_vidmem:
194 kfd_release_kernel_doorbell(dev->kfd, prop.doorbell_ptr);
195err_get_kernel_doorbell:
196 return false;
197
198}
199
200/* Uninitialize a kernel queue and free all its memory usages. */
201static void kq_uninitialize(struct kernel_queue *kq)
202{
203 if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && down_read_trylock(&kq->dev->adev->reset_domain->sem)) {
204 kq->mqd_mgr->destroy_mqd(kq->mqd_mgr,
205 kq->queue->mqd,
206 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
207 KFD_UNMAP_LATENCY_MS,
208 kq->queue->pipe,
209 kq->queue->queue);
210 up_read(&kq->dev->adev->reset_domain->sem);
211 }
212 else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ)
213 kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj);
214
215 kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd,
216 kq->queue->mqd_mem_obj);
217
218 kfd_gtt_sa_free(kq->dev, kq->rptr_mem);
219 kfd_gtt_sa_free(kq->dev, kq->wptr_mem);
220
221 /* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free()
222 * is able to handle NULL properly.
223 */
224 kfd_gtt_sa_free(kq->dev, kq->eop_mem);
225
226 kfd_gtt_sa_free(kq->dev, kq->pq);
227 kfd_release_kernel_doorbell(kq->dev->kfd,
228 kq->queue->properties.doorbell_ptr);
229 uninit_queue(kq->queue);
230}
231
232int kq_acquire_packet_buffer(struct kernel_queue *kq,
233 size_t packet_size_in_dwords, unsigned int **buffer_ptr)
234{
235 size_t available_size;
236 size_t queue_size_dwords;
237 uint32_t wptr, rptr;
238 uint64_t wptr64;
239 unsigned int *queue_address;
240
241 /* When rptr == wptr, the buffer is empty.
242 * When rptr == wptr + 1, the buffer is full.
243 * It is always rptr that advances to the position of wptr, rather than
244 * the opposite. So we can only use up to queue_size_dwords - 1 dwords.
245 */
246 rptr = *kq->rptr_kernel;
247 wptr = kq->pending_wptr;
248 wptr64 = kq->pending_wptr64;
249 queue_address = (unsigned int *)kq->pq_kernel_addr;
250 queue_size_dwords = kq->queue->properties.queue_size / 4;
251
252 pr_debug("rptr: %d\n", rptr);
253 pr_debug("wptr: %d\n", wptr);
254 pr_debug("queue_address 0x%p\n", queue_address);
255
256 available_size = (rptr + queue_size_dwords - 1 - wptr) %
257 queue_size_dwords;
258
259 if (packet_size_in_dwords > available_size) {
260 /*
261 * make sure calling functions know
262 * acquire_packet_buffer() failed
263 */
264 goto err_no_space;
265 }
266
267 if (wptr + packet_size_in_dwords >= queue_size_dwords) {
268 /* make sure after rolling back to position 0, there is
269 * still enough space.
270 */
271 if (packet_size_in_dwords >= rptr)
272 goto err_no_space;
273
274 /* fill nops, roll back and start at position 0 */
275 while (wptr > 0) {
276 queue_address[wptr] = kq->nop_packet;
277 wptr = (wptr + 1) % queue_size_dwords;
278 wptr64++;
279 }
280 }
281
282 *buffer_ptr = &queue_address[wptr];
283 kq->pending_wptr = wptr + packet_size_in_dwords;
284 kq->pending_wptr64 = wptr64 + packet_size_in_dwords;
285
286 return 0;
287
288err_no_space:
289 *buffer_ptr = NULL;
290 return -ENOMEM;
291}
292
293int kq_submit_packet(struct kernel_queue *kq)
294{
295#ifdef DEBUG
296 int i;
297
298 for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) {
299 pr_debug("0x%2X ", kq->pq_kernel_addr[i]);
300 if (i % 15 == 0)
301 pr_debug("\n");
302 }
303 pr_debug("\n");
304#endif
305 /* Fatal err detected, packet submission won't go through */
306 if (amdgpu_amdkfd_is_fed(kq->dev->adev))
307 return -EIO;
308
309 /* Make sure ring buffer is updated before wptr updated */
310 mb();
311
312 if (kq->dev->kfd->device_info.doorbell_size == 8) {
313 *kq->wptr64_kernel = kq->pending_wptr64;
314 mb(); /* Make sure wptr updated before ring doorbell */
315 write_kernel_doorbell64(kq->queue->properties.doorbell_ptr,
316 kq->pending_wptr64);
317 } else {
318 *kq->wptr_kernel = kq->pending_wptr;
319 mb(); /* Make sure wptr updated before ring doorbell */
320 write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
321 kq->pending_wptr);
322 }
323
324 return 0;
325}
326
327void kq_rollback_packet(struct kernel_queue *kq)
328{
329 if (kq->dev->kfd->device_info.doorbell_size == 8) {
330 kq->pending_wptr64 = *kq->wptr64_kernel;
331 kq->pending_wptr = *kq->wptr_kernel %
332 (kq->queue->properties.queue_size / 4);
333 } else {
334 kq->pending_wptr = *kq->wptr_kernel;
335 }
336}
337
338struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
339 enum kfd_queue_type type)
340{
341 struct kernel_queue *kq;
342
343 kq = kzalloc(sizeof(*kq), GFP_KERNEL);
344 if (!kq)
345 return NULL;
346
347 if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE))
348 return kq;
349
350 dev_err(dev->adev->dev, "Failed to init kernel queue\n");
351
352 kfree(kq);
353 return NULL;
354}
355
356void kernel_queue_uninit(struct kernel_queue *kq)
357{
358 kq_uninitialize(kq);
359 kfree(kq);
360}
361
362/* FIXME: Can this test be removed? */
363static __attribute__((unused)) void test_kq(struct kfd_node *dev)
364{
365 struct kernel_queue *kq;
366 uint32_t *buffer, i;
367 int retval;
368
369 dev_err(dev->adev->dev, "Starting kernel queue test\n");
370
371 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ);
372 if (unlikely(!kq)) {
373 dev_err(dev->adev->dev, " Failed to initialize HIQ\n");
374 dev_err(dev->adev->dev, "Kernel queue test failed\n");
375 return;
376 }
377
378 retval = kq_acquire_packet_buffer(kq, 5, &buffer);
379 if (unlikely(retval != 0)) {
380 dev_err(dev->adev->dev, " Failed to acquire packet buffer\n");
381 dev_err(dev->adev->dev, "Kernel queue test failed\n");
382 return;
383 }
384 for (i = 0; i < 5; i++)
385 buffer[i] = kq->nop_packet;
386 kq_submit_packet(kq);
387
388 dev_err(dev->adev->dev, "Ending kernel queue test\n");
389}
390
391