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 -
52 pm->dqm->xgmi_sdma_queue_count;
54 /* check if there is over subscription
55 * Note: the arbitration between the number of VMIDs and
56 * hws_max_conc_proc has been done in
57 * kgd2kfd_device_init().
59 *over_subscription = false;
61 if (dev->max_proc_per_quantum > 1)
62 max_proc_per_quantum = dev->max_proc_per_quantum;
64 if ((process_count > max_proc_per_quantum) ||
65 compute_queue_count > get_queues_num(pm->dqm)) {
66 *over_subscription = true;
67 pr_debug("Over subscribed runlist\n");
70 map_queue_size = pm->pmf->map_queues_size;
71 /* calculate run list ib allocation size */
72 *rlib_size = process_count * pm->pmf->map_process_size +
73 queue_count * map_queue_size;
76 * Increase the allocation size in case we need a chained run list
77 * when over subscription
79 if (*over_subscription)
80 *rlib_size += pm->pmf->runlist_size;
82 pr_debug("runlist ib size %d\n", *rlib_size);
85 static int pm_allocate_runlist_ib(struct packet_manager *pm,
86 unsigned int **rl_buffer,
87 uint64_t *rl_gpu_buffer,
88 unsigned int *rl_buffer_size,
89 bool *is_over_subscription)
93 if (WARN_ON(pm->allocated))
96 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
98 mutex_lock(&pm->lock);
100 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
104 pr_err("Failed to allocate runlist IB\n");
108 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
109 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
111 memset(*rl_buffer, 0, *rl_buffer_size);
112 pm->allocated = true;
115 mutex_unlock(&pm->lock);
119 static int pm_create_runlist_ib(struct packet_manager *pm,
120 struct list_head *queues,
121 uint64_t *rl_gpu_addr,
122 size_t *rl_size_bytes)
124 unsigned int alloc_size_bytes;
125 unsigned int *rl_buffer, rl_wptr, i;
126 int retval, proccesses_mapped;
127 struct device_process_node *cur;
128 struct qcm_process_device *qpd;
130 struct kernel_queue *kq;
131 bool is_over_subscription;
133 rl_wptr = retval = proccesses_mapped = 0;
135 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
136 &alloc_size_bytes, &is_over_subscription);
140 *rl_size_bytes = alloc_size_bytes;
141 pm->ib_size_bytes = alloc_size_bytes;
143 pr_debug("Building runlist ib process count: %d queues count %d\n",
144 pm->dqm->processes_count, pm->dqm->queue_count);
146 /* build the run list ib packet */
147 list_for_each_entry(cur, queues, list) {
149 /* build map process packet */
150 if (proccesses_mapped >= pm->dqm->processes_count) {
151 pr_debug("Not enough space left in runlist IB\n");
156 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
161 inc_wptr(&rl_wptr, pm->pmf->map_process_size,
164 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
165 if (!kq->queue->properties.is_active)
168 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
169 kq->queue->queue, qpd->is_debug);
171 retval = pm->pmf->map_queues(pm,
179 pm->pmf->map_queues_size,
183 list_for_each_entry(q, &qpd->queues_list, list) {
184 if (!q->properties.is_active)
187 pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
188 q->queue, qpd->is_debug);
190 retval = pm->pmf->map_queues(pm,
199 pm->pmf->map_queues_size,
204 pr_debug("Finished map process and queues to runlist\n");
206 if (is_over_subscription) {
207 if (!pm->is_over_subscription)
208 pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
209 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
211 alloc_size_bytes / sizeof(uint32_t),
214 pm->is_over_subscription = is_over_subscription;
216 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
217 pr_debug("0x%2X ", rl_buffer[i]);
223 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
225 switch (dqm->dev->device_info->asic_family) {
228 /* PM4 packet structures on CIK are the same as on VI */
236 pm->pmf = &kfd_vi_pm_funcs;
244 pm->pmf = &kfd_v9_pm_funcs;
249 pm->pmf = &kfd_v10_pm_funcs;
252 WARN(1, "Unexpected ASIC family %u",
253 dqm->dev->device_info->asic_family);
258 mutex_init(&pm->lock);
259 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
260 if (!pm->priv_queue) {
261 mutex_destroy(&pm->lock);
264 pm->allocated = false;
269 void pm_uninit(struct packet_manager *pm)
271 mutex_destroy(&pm->lock);
272 kernel_queue_uninit(pm->priv_queue);
275 int pm_send_set_resources(struct packet_manager *pm,
276 struct scheduling_resources *res)
278 uint32_t *buffer, size;
281 size = pm->pmf->set_resources_size;
282 mutex_lock(&pm->lock);
283 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
284 size / sizeof(uint32_t),
285 (unsigned int **)&buffer);
287 pr_err("Failed to allocate buffer on kernel queue\n");
292 retval = pm->pmf->set_resources(pm, buffer, res);
294 pm->priv_queue->ops.submit_packet(pm->priv_queue);
296 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
299 mutex_unlock(&pm->lock);
304 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
306 uint64_t rl_gpu_ib_addr;
308 size_t rl_ib_size, packet_size_dwords;
311 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
314 goto fail_create_runlist_ib;
316 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
318 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
319 mutex_lock(&pm->lock);
321 retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
322 packet_size_dwords, &rl_buffer);
324 goto fail_acquire_packet_buffer;
326 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
327 rl_ib_size / sizeof(uint32_t), false);
329 goto fail_create_runlist;
331 pm->priv_queue->ops.submit_packet(pm->priv_queue);
333 mutex_unlock(&pm->lock);
338 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
339 fail_acquire_packet_buffer:
340 mutex_unlock(&pm->lock);
341 fail_create_runlist_ib:
346 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
347 uint32_t fence_value)
349 uint32_t *buffer, size;
352 if (WARN_ON(!fence_address))
355 size = pm->pmf->query_status_size;
356 mutex_lock(&pm->lock);
357 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
358 size / sizeof(uint32_t), (unsigned int **)&buffer);
360 pr_err("Failed to allocate buffer on kernel queue\n");
365 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
367 pm->priv_queue->ops.submit_packet(pm->priv_queue);
369 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
372 mutex_unlock(&pm->lock);
376 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
377 enum kfd_unmap_queues_filter filter,
378 uint32_t filter_param, bool reset,
379 unsigned int sdma_engine)
381 uint32_t *buffer, size;
384 size = pm->pmf->unmap_queues_size;
385 mutex_lock(&pm->lock);
386 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
387 size / sizeof(uint32_t), (unsigned int **)&buffer);
389 pr_err("Failed to allocate buffer on kernel queue\n");
394 retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param,
397 pm->priv_queue->ops.submit_packet(pm->priv_queue);
399 pm->priv_queue->ops.rollback_packet(pm->priv_queue);
402 mutex_unlock(&pm->lock);
406 void pm_release_ib(struct packet_manager *pm)
408 mutex_lock(&pm->lock);
410 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
411 pm->allocated = false;
413 mutex_unlock(&pm->lock);
416 #if defined(CONFIG_DEBUG_FS)
418 int pm_debugfs_runlist(struct seq_file *m, void *data)
420 struct packet_manager *pm = data;
422 mutex_lock(&pm->lock);
424 if (!pm->allocated) {
425 seq_puts(m, " No active runlist\n");
429 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
430 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
433 mutex_unlock(&pm->lock);
437 int pm_debugfs_hang_hws(struct packet_manager *pm)
439 uint32_t *buffer, size;
442 size = pm->pmf->query_status_size;
443 mutex_lock(&pm->lock);
444 pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
445 size / sizeof(uint32_t), (unsigned int **)&buffer);
447 pr_err("Failed to allocate buffer on kernel queue\n");
451 memset(buffer, 0x55, size);
452 pm->priv_queue->ops.submit_packet(pm->priv_queue);
454 pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
455 buffer[0], buffer[1], buffer[2], buffer[3],
456 buffer[4], buffer[5], buffer[6]);
458 mutex_unlock(&pm->lock);