amd/amdkfd/kfd_kernel_queue.c

0001 // SPDX-License-Identifier: GPL-2.0 OR MIT
0002 /*
0003  * Copyright 2014-2022 Advanced Micro Devices, Inc.
0004  *
0005  * Permission is hereby granted, free of charge, to any person obtaining a
0006  * copy of this software and associated documentation files (the "Software"),
0007  * to deal in the Software without restriction, including without limitation
0008  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
0009  * and/or sell copies of the Software, and to permit persons to whom the
0010  * Software is furnished to do so, subject to the following conditions:
0011  *
0012  * The above copyright notice and this permission notice shall be included in
0013  * all copies or substantial portions of the Software.
0014  *
0015  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
0016  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
0017  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
0018  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
0019  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
0020  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
0021  * OTHER DEALINGS IN THE SOFTWARE.
0022  *
0023  */
0024
0025 #include <linux/types.h>
0026 #include <linux/mutex.h>
0027 #include <linux/slab.h>
0028 #include <linux/printk.h>
0029 #include <linux/sched.h>
0030 #include "kfd_kernel_queue.h"
0031 #include "kfd_priv.h"
0032 #include "kfd_device_queue_manager.h"
0033 #include "kfd_pm4_headers.h"
0034 #include "kfd_pm4_opcodes.h"
0035
0036 #define PM4_COUNT_ZERO (((1 << 15) - 1) << 16)
0037
0038 /* Initialize a kernel queue, including allocations of GART memory
0039  * needed for the queue.
0040  */
0041 static bool kq_initialize(struct kernel_queue *kq, struct kfd_dev *dev,
0042         enum kfd_queue_type type, unsigned int queue_size)
0043 {
0044     struct queue_properties prop;
0045     int retval;
0046     union PM4_MES_TYPE_3_HEADER nop;
0047
0048     if (WARN_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ))
0049         return false;
0050
0051     pr_debug("Initializing queue type %d size %d\n", KFD_QUEUE_TYPE_HIQ,
0052             queue_size);
0053
0054     memset(&prop, 0, sizeof(prop));
0055     memset(&nop, 0, sizeof(nop));
0056
0057     nop.opcode = IT_NOP;
0058     nop.type = PM4_TYPE_3;
0059     nop.u32all |= PM4_COUNT_ZERO;
0060
0061     kq->dev = dev;
0062     kq->nop_packet = nop.u32all;
0063     switch (type) {
0064     case KFD_QUEUE_TYPE_DIQ:
0065         kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ];
0066         break;
0067     case KFD_QUEUE_TYPE_HIQ:
0068         kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
0069         break;
0070     default:
0071         pr_err("Invalid queue type %d\n", type);
0072         return false;
0073     }
0074
0075     if (!kq->mqd_mgr)
0076         return false;
0077
0078     prop.doorbell_ptr = kfd_get_kernel_doorbell(dev, &prop.doorbell_off);
0079
0080     if (!prop.doorbell_ptr) {
0081         pr_err("Failed to initialize doorbell");
0082         goto err_get_kernel_doorbell;
0083     }
0084
0085     retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq);
0086     if (retval != 0) {
0087         pr_err("Failed to init pq queues size %d\n", queue_size);
0088         goto err_pq_allocate_vidmem;
0089     }
0090
0091     kq->pq_kernel_addr = kq->pq->cpu_ptr;
0092     kq->pq_gpu_addr = kq->pq->gpu_addr;
0093
0094     /* For CIK family asics, kq->eop_mem is not needed */
0095     if (dev->adev->asic_type > CHIP_MULLINS) {
0096         retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &kq->eop_mem);
0097         if (retval != 0)
0098             goto err_eop_allocate_vidmem;
0099
0100         kq->eop_gpu_addr = kq->eop_mem->gpu_addr;
0101         kq->eop_kernel_addr = kq->eop_mem->cpu_ptr;
0102
0103         memset(kq->eop_kernel_addr, 0, PAGE_SIZE);
0104     }
0105
0106     retval = kfd_gtt_sa_allocate(dev, sizeof(*kq->rptr_kernel),
0107                     &kq->rptr_mem);
0108
0109     if (retval != 0)
0110         goto err_rptr_allocate_vidmem;
0111
0112     kq->rptr_kernel = kq->rptr_mem->cpu_ptr;
0113     kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr;
0114
0115     retval = kfd_gtt_sa_allocate(dev, dev->device_info.doorbell_size,
0116                     &kq->wptr_mem);
0117
0118     if (retval != 0)
0119         goto err_wptr_allocate_vidmem;
0120
0121     kq->wptr_kernel = kq->wptr_mem->cpu_ptr;
0122     kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr;
0123
0124     memset(kq->pq_kernel_addr, 0, queue_size);
0125     memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel));
0126     memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel));
0127
0128     prop.queue_size = queue_size;
0129     prop.is_interop = false;
0130     prop.is_gws = false;
0131     prop.priority = 1;
0132     prop.queue_percent = 100;
0133     prop.type = type;
0134     prop.vmid = 0;
0135     prop.queue_address = kq->pq_gpu_addr;
0136     prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr;
0137     prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr;
0138     prop.eop_ring_buffer_address = kq->eop_gpu_addr;
0139     prop.eop_ring_buffer_size = PAGE_SIZE;
0140
0141     if (init_queue(&kq->queue, &prop) != 0)
0142         goto err_init_queue;
0143
0144     kq->queue->device = dev;
0145     kq->queue->process = kfd_get_process(current);
0146
0147     kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev,
0148                     &kq->queue->properties);
0149     if (!kq->queue->mqd_mem_obj)
0150         goto err_allocate_mqd;
0151     kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd,
0152                     kq->queue->mqd_mem_obj,
0153                     &kq->queue->gart_mqd_addr,
0154                     &kq->queue->properties);
0155     /* assign HIQ to HQD */
0156     if (type == KFD_QUEUE_TYPE_HIQ) {
0157         pr_debug("Assigning hiq to hqd\n");
0158         kq->queue->pipe = KFD_CIK_HIQ_PIPE;
0159         kq->queue->queue = KFD_CIK_HIQ_QUEUE;
0160         kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd,
0161                 kq->queue->pipe, kq->queue->queue,
0162                 &kq->queue->properties, NULL);
0163     } else {
0164         /* allocate fence for DIQ */
0165
0166         retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t),
0167                         &kq->fence_mem_obj);
0168
0169         if (retval != 0)
0170             goto err_alloc_fence;
0171
0172         kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
0173         kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
0174     }
0175
0176     print_queue(kq->queue);
0177
0178     return true;
0179 err_alloc_fence:
0180     kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj);
0181 err_allocate_mqd:
0182     uninit_queue(kq->queue);
0183 err_init_queue:
0184     kfd_gtt_sa_free(dev, kq->wptr_mem);
0185 err_wptr_allocate_vidmem:
0186     kfd_gtt_sa_free(dev, kq->rptr_mem);
0187 err_rptr_allocate_vidmem:
0188     kfd_gtt_sa_free(dev, kq->eop_mem);
0189 err_eop_allocate_vidmem:
0190     kfd_gtt_sa_free(dev, kq->pq);
0191 err_pq_allocate_vidmem:
0192     kfd_release_kernel_doorbell(dev, prop.doorbell_ptr);
0193 err_get_kernel_doorbell:
0194     return false;
0195
0196 }
0197
0198 /* Uninitialize a kernel queue and free all its memory usages. */
0199 static void kq_uninitialize(struct kernel_queue *kq, bool hanging)
0200 {
0201     if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && !hanging)
0202         kq->mqd_mgr->destroy_mqd(kq->mqd_mgr,
0203                     kq->queue->mqd,
0204                     KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
0205                     KFD_UNMAP_LATENCY_MS,
0206                     kq->queue->pipe,
0207                     kq->queue->queue);
0208     else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ)
0209         kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj);
0210
0211     kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd,
0212                 kq->queue->mqd_mem_obj);
0213
0214     kfd_gtt_sa_free(kq->dev, kq->rptr_mem);
0215     kfd_gtt_sa_free(kq->dev, kq->wptr_mem);
0216
0217     /* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free()
0218      * is able to handle NULL properly.
0219      */
0220     kfd_gtt_sa_free(kq->dev, kq->eop_mem);
0221
0222     kfd_gtt_sa_free(kq->dev, kq->pq);
0223     kfd_release_kernel_doorbell(kq->dev,
0224                     kq->queue->properties.doorbell_ptr);
0225     uninit_queue(kq->queue);
0226 }
0227
0228 int kq_acquire_packet_buffer(struct kernel_queue *kq,
0229         size_t packet_size_in_dwords, unsigned int **buffer_ptr)
0230 {
0231     size_t available_size;
0232     size_t queue_size_dwords;
0233     uint32_t wptr, rptr;
0234     uint64_t wptr64;
0235     unsigned int *queue_address;
0236
0237     /* When rptr == wptr, the buffer is empty.
0238      * When rptr == wptr + 1, the buffer is full.
0239      * It is always rptr that advances to the position of wptr, rather than
0240      * the opposite. So we can only use up to queue_size_dwords - 1 dwords.
0241      */
0242     rptr = *kq->rptr_kernel;
0243     wptr = kq->pending_wptr;
0244     wptr64 = kq->pending_wptr64;
0245     queue_address = (unsigned int *)kq->pq_kernel_addr;
0246     queue_size_dwords = kq->queue->properties.queue_size / 4;
0247
0248     pr_debug("rptr: %d\n", rptr);
0249     pr_debug("wptr: %d\n", wptr);
0250     pr_debug("queue_address 0x%p\n", queue_address);
0251
0252     available_size = (rptr + queue_size_dwords - 1 - wptr) %
0253                             queue_size_dwords;
0254
0255     if (packet_size_in_dwords > available_size) {
0256         /*
0257          * make sure calling functions know
0258          * acquire_packet_buffer() failed
0259          */
0260         goto err_no_space;
0261     }
0262
0263     if (wptr + packet_size_in_dwords >= queue_size_dwords) {
0264         /* make sure after rolling back to position 0, there is
0265          * still enough space.
0266          */
0267         if (packet_size_in_dwords >= rptr)
0268             goto err_no_space;
0269
0270         /* fill nops, roll back and start at position 0 */
0271         while (wptr > 0) {
0272             queue_address[wptr] = kq->nop_packet;
0273             wptr = (wptr + 1) % queue_size_dwords;
0274             wptr64++;
0275         }
0276     }
0277
0278     *buffer_ptr = &queue_address[wptr];
0279     kq->pending_wptr = wptr + packet_size_in_dwords;
0280     kq->pending_wptr64 = wptr64 + packet_size_in_dwords;
0281
0282     return 0;
0283
0284 err_no_space:
0285     *buffer_ptr = NULL;
0286     return -ENOMEM;
0287 }
0288
0289 void kq_submit_packet(struct kernel_queue *kq)
0290 {
0291 #ifdef DEBUG
0292     int i;
0293
0294     for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) {
0295         pr_debug("0x%2X ", kq->pq_kernel_addr[i]);
0296         if (i % 15 == 0)
0297             pr_debug("\n");
0298     }
0299     pr_debug("\n");
0300 #endif
0301     if (kq->dev->device_info.doorbell_size == 8) {
0302         *kq->wptr64_kernel = kq->pending_wptr64;
0303         write_kernel_doorbell64(kq->queue->properties.doorbell_ptr,
0304                     kq->pending_wptr64);
0305     } else {
0306         *kq->wptr_kernel = kq->pending_wptr;
0307         write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
0308                     kq->pending_wptr);
0309     }
0310 }
0311
0312 void kq_rollback_packet(struct kernel_queue *kq)
0313 {
0314     if (kq->dev->device_info.doorbell_size == 8) {
0315         kq->pending_wptr64 = *kq->wptr64_kernel;
0316         kq->pending_wptr = *kq->wptr_kernel %
0317             (kq->queue->properties.queue_size / 4);
0318     } else {
0319         kq->pending_wptr = *kq->wptr_kernel;
0320     }
0321 }
0322
0323 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
0324                     enum kfd_queue_type type)
0325 {
0326     struct kernel_queue *kq;
0327
0328     kq = kzalloc(sizeof(*kq), GFP_KERNEL);
0329     if (!kq)
0330         return NULL;
0331
0332     if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE))
0333         return kq;
0334
0335     pr_err("Failed to init kernel queue\n");
0336
0337     kfree(kq);
0338     return NULL;
0339 }
0340
0341 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging)
0342 {
0343     kq_uninitialize(kq, hanging);
0344     kfree(kq);
0345 }
0346
0347 /* FIXME: Can this test be removed? */
0348 static __attribute__((unused)) void test_kq(struct kfd_dev *dev)
0349 {
0350     struct kernel_queue *kq;
0351     uint32_t *buffer, i;
0352     int retval;
0353
0354     pr_err("Starting kernel queue test\n");
0355
0356     kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ);
0357     if (unlikely(!kq)) {
0358         pr_err("  Failed to initialize HIQ\n");
0359         pr_err("Kernel queue test failed\n");
0360         return;
0361     }
0362
0363     retval = kq_acquire_packet_buffer(kq, 5, &buffer);
0364     if (unlikely(retval != 0)) {
0365         pr_err("  Failed to acquire packet buffer\n");
0366         pr_err("Kernel queue test failed\n");
0367         return;
0368     }
0369     for (i = 0; i < 5; i++)
0370         buffer[i] = kq->nop_packet;
0371     kq_submit_packet(kq);
0372
0373     pr_err("Ending kernel queue test\n");
0374 }
0375
0376