2 * Copyright 2014 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 #include <linux/slab.h>
25 #include <linux/mutex.h>
26 #include "kfd_device_queue_manager.h"
27 #include "kfd_kernel_queue.h"
30 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
31 unsigned int buffer_size_bytes)
33 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
35 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
36 "Runlist IB overflow");
40 static void pm_calc_rlib_size(struct packet_manager *pm,
41 unsigned int *rlib_size,
42 bool *over_subscription)
44 unsigned int process_count, queue_count, compute_queue_count;
45 unsigned int map_queue_size;
46 unsigned int max_proc_per_quantum = 1;
47 struct kfd_dev *dev = pm->dqm->dev;
49 process_count = pm->dqm->processes_count;
50 queue_count = pm->dqm->queue_count;
51 compute_queue_count = queue_count - pm->dqm->sdma_queue_count;
53 /* check if there is over subscription
54 * Note: the arbitration between the number of VMIDs and
55 * hws_max_conc_proc has been done in
56 * kgd2kfd_device_init().
58 *over_subscription = false;
60 if (dev->max_proc_per_quantum > 1)
61 max_proc_per_quantum = dev->max_proc_per_quantum;
63 if ((process_count > max_proc_per_quantum) ||
64 compute_queue_count > get_queues_num(pm->dqm)) {
65 *over_subscription = true;
66 pr_debug("Over subscribed runlist\n");
69 map_queue_size = pm->pmf->map_queues_size;
70 /* calculate run list ib allocation size */
71 *rlib_size = process_count * pm->pmf->map_process_size +
72 queue_count * map_queue_size;
75 * Increase the allocation size in case we need a chained run list
76 * when over subscription
78 if (*over_subscription)
79 *rlib_size += pm->pmf->runlist_size;
81 pr_debug("runlist ib size %d\n", *rlib_size);
84 static int pm_allocate_runlist_ib(struct packet_manager *pm,
85 unsigned int **rl_buffer,
86 uint64_t *rl_gpu_buffer,
87 unsigned int *rl_buffer_size,
88 bool *is_over_subscription)
92 if (WARN_ON(pm->allocated))
95 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
97 mutex_lock(&pm->lock);
99 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
103 pr_err("Failed to allocate runlist IB\n");
107 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
108 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
110 memset(*rl_buffer, 0, *rl_buffer_size);
111 pm->allocated = true;
114 mutex_unlock(&pm->lock);
118 static int pm_create_runlist_ib(struct packet_manager *pm,
119 struct list_head *queues,
120 uint64_t *rl_gpu_addr,
121 size_t *rl_size_bytes)
123 unsigned int alloc_size_bytes;
124 unsigned int *rl_buffer, rl_wptr, i;
125 int retval, proccesses_mapped;
126 struct device_process_node *cur;
127 struct qcm_process_device *qpd;
129 struct kernel_queue *kq;
130 bool is_over_subscription;
132 rl_wptr = retval = proccesses_mapped = 0;
134 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
135 &alloc_size_bytes, &is_over_subscription);
139 *rl_size_bytes = alloc_size_bytes;
140 pm->ib_size_bytes = alloc_size_bytes;
142 pr_debug("Building runlist ib process count: %d queues count %d\n",
143 pm->dqm->processes_count, pm->dqm->queue_count);
145 /* build the run list ib packet */
146 list_for_each_entry(cur, queues, list) {
148 /* build map process packet */
149 if (proccesses_mapped >= pm->dqm->processes_count) {
150 pr_debug("Not enough space left in runlist IB\n");
155 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
160 inc_wptr(&rl_wptr, pm->pmf->map_process_size,
163 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
164 if (!kq->queue->properties.is_active)
167 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
168 kq->queue->queue, qpd->is_debug);
170 retval = pm->pmf->map_queues(pm,
178 pm->pmf->map_queues_size,
182 list_for_each_entry(q, &qpd->queues_list, list) {
183 if (!q->properties.is_active)
186 pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
187 q->queue, qpd->is_debug);
189 retval = pm->pmf->map_queues(pm,
198 pm->pmf->map_queues_size,
203 pr_debug("Finished map process and queues to runlist\n");
205 if (is_over_subscription)
206 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
208 alloc_size_bytes / sizeof(uint32_t),
211 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
212 pr_debug("0x%2X ", rl_buffer[i]);
218 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
220 switch (dqm->dev->device_info->asic_family) {
223 /* PM4 packet structures on CIK are the same as on VI */
229 pm->pmf = &kfd_vi_pm_funcs;
233 pm->pmf = &kfd_v9_pm_funcs;
236 WARN(1, "Unexpected ASIC family %u",
237 dqm->dev->device_info->asic_family);
242 mutex_init(&pm->lock);
243 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
244 if (!pm->priv_queue) {
245 mutex_destroy(&pm->lock);
248 pm->allocated = false;
253 void pm_uninit(struct packet_manager *pm)
255 mutex_destroy(&pm->lock);
256 kernel_queue_uninit(pm->priv_queue);
259 int pm_send_set_resources(struct packet_manager *pm,
260 struct scheduling_resources *res)
262 uint32_t *buffer, size;
265 size = pm->pmf->set_resources_size;
266 mutex_lock(&pm->lock);
267 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
268 size / sizeof(uint32_t),
269 (unsigned int **)&buffer);
271 pr_err("Failed to allocate buffer on kernel queue\n");
276 retval = pm->pmf->set_resources(pm, buffer, res);
278 pm->priv_queue->ops.submit_packet(pm->priv_queue);
280 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
283 mutex_unlock(&pm->lock);
288 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
290 uint64_t rl_gpu_ib_addr;
292 size_t rl_ib_size, packet_size_dwords;
295 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
298 goto fail_create_runlist_ib;
300 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
302 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
303 mutex_lock(&pm->lock);
305 retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
306 packet_size_dwords, &rl_buffer);
308 goto fail_acquire_packet_buffer;
310 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
311 rl_ib_size / sizeof(uint32_t), false);
313 goto fail_create_runlist;
315 pm->priv_queue->ops.submit_packet(pm->priv_queue);
317 mutex_unlock(&pm->lock);
322 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
323 fail_acquire_packet_buffer:
324 mutex_unlock(&pm->lock);
325 fail_create_runlist_ib:
330 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
331 uint32_t fence_value)
333 uint32_t *buffer, size;
336 if (WARN_ON(!fence_address))
339 size = pm->pmf->query_status_size;
340 mutex_lock(&pm->lock);
341 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
342 size / sizeof(uint32_t), (unsigned int **)&buffer);
344 pr_err("Failed to allocate buffer on kernel queue\n");
349 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
351 pm->priv_queue->ops.submit_packet(pm->priv_queue);
353 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
356 mutex_unlock(&pm->lock);
360 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
361 enum kfd_unmap_queues_filter filter,
362 uint32_t filter_param, bool reset,
363 unsigned int sdma_engine)
365 uint32_t *buffer, size;
368 size = pm->pmf->unmap_queues_size;
369 mutex_lock(&pm->lock);
370 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
371 size / sizeof(uint32_t), (unsigned int **)&buffer);
373 pr_err("Failed to allocate buffer on kernel queue\n");
378 retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param,
381 pm->priv_queue->ops.submit_packet(pm->priv_queue);
383 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
386 mutex_unlock(&pm->lock);
390 void pm_release_ib(struct packet_manager *pm)
392 mutex_lock(&pm->lock);
394 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
395 pm->allocated = false;
397 mutex_unlock(&pm->lock);
400 #if defined(CONFIG_DEBUG_FS)
402 int pm_debugfs_runlist(struct seq_file *m, void *data)
404 struct packet_manager *pm = data;
406 mutex_lock(&pm->lock);
408 if (!pm->allocated) {
409 seq_puts(m, " No active runlist\n");
413 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
414 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
417 mutex_unlock(&pm->lock);
421 int pm_debugfs_hang_hws(struct packet_manager *pm)
423 uint32_t *buffer, size;
426 size = pm->pmf->query_status_size;
427 mutex_lock(&pm->lock);
428 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
429 size / sizeof(uint32_t), (unsigned int **)&buffer);
431 pr_err("Failed to allocate buffer on kernel queue\n");
435 memset(buffer, 0x55, size);
436 pm->priv_queue->ops.submit_packet(pm->priv_queue);
438 pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
439 buffer[0], buffer[1], buffer[2], buffer[3],
440 buffer[4], buffer[5], buffer[6]);
442 mutex_unlock(&pm->lock);