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
 37/* Initialize a kernel queue, including allocations of GART memory
 38 * needed for the queue.
 39 */
 40static bool kq_initialize(struct kernel_queue *kq, struct kfd_dev *dev,
 41		enum kfd_queue_type type, unsigned int queue_size)
 42{
 43	struct queue_properties prop;
 44	int retval;
 45	union PM4_MES_TYPE_3_HEADER nop;
 46
 47	if (WARN_ON(type != KFD_QUEUE_TYPE_DIQ && type != KFD_QUEUE_TYPE_HIQ))
 48		return false;
 49
 50	pr_debug("Initializing queue type %d size %d\n", KFD_QUEUE_TYPE_HIQ,
 51			queue_size);
 52
 53	memset(&prop, 0, sizeof(prop));
 54	memset(&nop, 0, sizeof(nop));
 55
 56	nop.opcode = IT_NOP;
 57	nop.type = PM4_TYPE_3;
 58	nop.u32all |= PM4_COUNT_ZERO;
 59
 60	kq->dev = dev;
 61	kq->nop_packet = nop.u32all;
 62	switch (type) {
 63	case KFD_QUEUE_TYPE_DIQ:
 64		kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_DIQ];
 65		break;
 66	case KFD_QUEUE_TYPE_HIQ:
 67		kq->mqd_mgr = dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
 68		break;
 69	default:
 70		pr_err("Invalid queue type %d\n", type);
 71		return false;
 72	}
 73
 74	if (!kq->mqd_mgr)
 75		return false;
 76
 77	prop.doorbell_ptr = kfd_get_kernel_doorbell(dev, &prop.doorbell_off);
 78
 79	if (!prop.doorbell_ptr) {
 80		pr_err("Failed to initialize doorbell");
 81		goto err_get_kernel_doorbell;
 82	}
 83
 84	retval = kfd_gtt_sa_allocate(dev, queue_size, &kq->pq);
 85	if (retval != 0) {
 86		pr_err("Failed to init pq queues size %d\n", 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->device_info->asic_family > CHIP_MULLINS) {
 95		retval = kfd_gtt_sa_allocate(dev, PAGE_SIZE, &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(dev, sizeof(*kq->rptr_kernel),
106					&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(dev, dev->device_info->doorbell_size,
115					&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, sizeof(*kq->wptr_kernel));
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	prop.cu_mask = NULL;
140
141	if (init_queue(&kq->queue, &prop) != 0)
142		goto err_init_queue;
143
144	kq->queue->device = dev;
145	kq->queue->process = kfd_get_process(current);
146
147	kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev,
148					&kq->queue->properties);
149	if (!kq->queue->mqd_mem_obj)
150		goto err_allocate_mqd;
151	kq->mqd_mgr->init_mqd(kq->mqd_mgr, &kq->queue->mqd,
152					kq->queue->mqd_mem_obj,
153					&kq->queue->gart_mqd_addr,
154					&kq->queue->properties);
155	/* assign HIQ to HQD */
156	if (type == KFD_QUEUE_TYPE_HIQ) {
157		pr_debug("Assigning hiq to hqd\n");
158		kq->queue->pipe = KFD_CIK_HIQ_PIPE;
159		kq->queue->queue = KFD_CIK_HIQ_QUEUE;
160		kq->mqd_mgr->load_mqd(kq->mqd_mgr, kq->queue->mqd,
161				kq->queue->pipe, kq->queue->queue,
162				&kq->queue->properties, NULL);
163	} else {
164		/* allocate fence for DIQ */
165
166		retval = kfd_gtt_sa_allocate(dev, sizeof(uint32_t),
167						&kq->fence_mem_obj);
168
169		if (retval != 0)
170			goto err_alloc_fence;
171
172		kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
173		kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
174	}
175
176	print_queue(kq->queue);
177
178	return true;
179err_alloc_fence:
180	kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd, kq->queue->mqd_mem_obj);
181err_allocate_mqd:
182	uninit_queue(kq->queue);
183err_init_queue:
184	kfd_gtt_sa_free(dev, kq->wptr_mem);
185err_wptr_allocate_vidmem:
186	kfd_gtt_sa_free(dev, kq->rptr_mem);
187err_rptr_allocate_vidmem:
188	kfd_gtt_sa_free(dev, kq->eop_mem);
189err_eop_allocate_vidmem:
190	kfd_gtt_sa_free(dev, kq->pq);
191err_pq_allocate_vidmem:
192	kfd_release_kernel_doorbell(dev, prop.doorbell_ptr);
193err_get_kernel_doorbell:
194	return false;
195
196}
197
198/* Uninitialize a kernel queue and free all its memory usages. */
199static void kq_uninitialize(struct kernel_queue *kq, bool hanging)
200{
201	if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ && !hanging)
202		kq->mqd_mgr->destroy_mqd(kq->mqd_mgr,
203					kq->queue->mqd,
204					KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
205					KFD_UNMAP_LATENCY_MS,
206					kq->queue->pipe,
207					kq->queue->queue);
208	else if (kq->queue->properties.type == KFD_QUEUE_TYPE_DIQ)
209		kfd_gtt_sa_free(kq->dev, kq->fence_mem_obj);
210
211	kq->mqd_mgr->free_mqd(kq->mqd_mgr, kq->queue->mqd,
212				kq->queue->mqd_mem_obj);
213
214	kfd_gtt_sa_free(kq->dev, kq->rptr_mem);
215	kfd_gtt_sa_free(kq->dev, kq->wptr_mem);
216
217	/* For CIK family asics, kq->eop_mem is Null, kfd_gtt_sa_free()
218	 * is able to handle NULL properly.
219	 */
220	kfd_gtt_sa_free(kq->dev, kq->eop_mem);
221
222	kfd_gtt_sa_free(kq->dev, kq->pq);
223	kfd_release_kernel_doorbell(kq->dev,
224					kq->queue->properties.doorbell_ptr);
225	uninit_queue(kq->queue);
226}
227
228int kq_acquire_packet_buffer(struct kernel_queue *kq,
229		size_t packet_size_in_dwords, unsigned int **buffer_ptr)
230{
231	size_t available_size;
232	size_t queue_size_dwords;
233	uint32_t wptr, rptr;
234	uint64_t wptr64;
235	unsigned int *queue_address;
236
237	/* When rptr == wptr, the buffer is empty.
238	 * When rptr == wptr + 1, the buffer is full.
239	 * It is always rptr that advances to the position of wptr, rather than
240	 * the opposite. So we can only use up to queue_size_dwords - 1 dwords.
241	 */
242	rptr = *kq->rptr_kernel;
243	wptr = kq->pending_wptr;
244	wptr64 = kq->pending_wptr64;
245	queue_address = (unsigned int *)kq->pq_kernel_addr;
246	queue_size_dwords = kq->queue->properties.queue_size / 4;
247
248	pr_debug("rptr: %d\n", rptr);
249	pr_debug("wptr: %d\n", wptr);
250	pr_debug("queue_address 0x%p\n", queue_address);
251
252	available_size = (rptr + queue_size_dwords - 1 - wptr) %
253							queue_size_dwords;
254
255	if (packet_size_in_dwords > available_size) {
256		/*
257		 * make sure calling functions know
258		 * acquire_packet_buffer() failed
259		 */
260		goto err_no_space;
261	}
262
263	if (wptr + packet_size_in_dwords >= queue_size_dwords) {
264		/* make sure after rolling back to position 0, there is
265		 * still enough space.
266		 */
267		if (packet_size_in_dwords >= rptr)
268			goto err_no_space;
269
270		/* fill nops, roll back and start at position 0 */
271		while (wptr > 0) {
272			queue_address[wptr] = kq->nop_packet;
273			wptr = (wptr + 1) % queue_size_dwords;
274			wptr64++;
275		}
276	}
277
278	*buffer_ptr = &queue_address[wptr];
279	kq->pending_wptr = wptr + packet_size_in_dwords;
280	kq->pending_wptr64 = wptr64 + packet_size_in_dwords;
281
282	return 0;
283
284err_no_space:
285	*buffer_ptr = NULL;
286	return -ENOMEM;
287}
288
289void kq_submit_packet(struct kernel_queue *kq)
290{
291#ifdef DEBUG
292	int i;
293
294	for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) {
295		pr_debug("0x%2X ", kq->pq_kernel_addr[i]);
296		if (i % 15 == 0)
297			pr_debug("\n");
298	}
299	pr_debug("\n");
300#endif
301	if (kq->dev->device_info->doorbell_size == 8) {
302		*kq->wptr64_kernel = kq->pending_wptr64;
303		write_kernel_doorbell64(kq->queue->properties.doorbell_ptr,
304					kq->pending_wptr64);
305	} else {
306		*kq->wptr_kernel = kq->pending_wptr;
307		write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
308					kq->pending_wptr);
309	}
310}
311
312void kq_rollback_packet(struct kernel_queue *kq)
313{
314	if (kq->dev->device_info->doorbell_size == 8) {
315		kq->pending_wptr64 = *kq->wptr64_kernel;
316		kq->pending_wptr = *kq->wptr_kernel %
317			(kq->queue->properties.queue_size / 4);
318	} else {
319		kq->pending_wptr = *kq->wptr_kernel;
320	}
321}
322
323struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
324					enum kfd_queue_type type)
325{
326	struct kernel_queue *kq;
327
328	kq = kzalloc(sizeof(*kq), GFP_KERNEL);
329	if (!kq)
330		return NULL;
331
332	if (kq_initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE))
333		return kq;
334
335	pr_err("Failed to init kernel queue\n");
336
337	kfree(kq);
338	return NULL;
339}
340
341void kernel_queue_uninit(struct kernel_queue *kq, bool hanging)
342{
343	kq_uninitialize(kq, hanging);
344	kfree(kq);
345}
346
347/* FIXME: Can this test be removed? */
348static __attribute__((unused)) void test_kq(struct kfd_dev *dev)
349{
350	struct kernel_queue *kq;
351	uint32_t *buffer, i;
352	int retval;
353
354	pr_err("Starting kernel queue test\n");
355
356	kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ);
357	if (unlikely(!kq)) {
358		pr_err("  Failed to initialize HIQ\n");
359		pr_err("Kernel queue test failed\n");
360		return;
361	}
362
363	retval = kq_acquire_packet_buffer(kq, 5, &buffer);
364	if (unlikely(retval != 0)) {
365		pr_err("  Failed to acquire packet buffer\n");
366		pr_err("Kernel queue test failed\n");
367		return;
368	}
369	for (i = 0; i < 5; i++)
370		buffer[i] = kq->nop_packet;
371	kq_submit_packet(kq);
372
373	pr_err("Ending kernel queue test\n");
374}
375
376