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 static 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("kfd: 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->get_mqd_manager(dev->dqm,
60 						KFD_MQD_TYPE_CIK_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 		goto err_get_kernel_doorbell;
74 
75 	retval = kfd2kgd->allocate_mem(dev->kgd,
76 					queue_size,
77 					PAGE_SIZE,
78 					KFD_MEMPOOL_SYSTEM_WRITECOMBINE,
79 					(struct kgd_mem **) &kq->pq);
80 
81 	if (retval != 0)
82 		goto err_pq_allocate_vidmem;
83 
84 	kq->pq_kernel_addr = kq->pq->cpu_ptr;
85 	kq->pq_gpu_addr = kq->pq->gpu_addr;
86 
87 	retval = kfd2kgd->allocate_mem(dev->kgd,
88 					sizeof(*kq->rptr_kernel),
89 					32,
90 					KFD_MEMPOOL_SYSTEM_WRITECOMBINE,
91 					(struct kgd_mem **) &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 = kfd2kgd->allocate_mem(dev->kgd,
100 					sizeof(*kq->wptr_kernel),
101 					32,
102 					KFD_MEMPOOL_SYSTEM_WRITECOMBINE,
103 					(struct kgd_mem **) &kq->wptr_mem);
104 
105 	if (retval != 0)
106 		goto err_wptr_allocate_vidmem;
107 
108 	kq->wptr_kernel = kq->wptr_mem->cpu_ptr;
109 	kq->wptr_gpu_addr = kq->wptr_mem->gpu_addr;
110 
111 	memset(kq->pq_kernel_addr, 0, queue_size);
112 	memset(kq->rptr_kernel, 0, sizeof(*kq->rptr_kernel));
113 	memset(kq->wptr_kernel, 0, sizeof(*kq->wptr_kernel));
114 
115 	prop.queue_size = queue_size;
116 	prop.is_interop = false;
117 	prop.priority = 1;
118 	prop.queue_percent = 100;
119 	prop.type = type;
120 	prop.vmid = 0;
121 	prop.queue_address = kq->pq_gpu_addr;
122 	prop.read_ptr = (uint32_t *) kq->rptr_gpu_addr;
123 	prop.write_ptr = (uint32_t *) kq->wptr_gpu_addr;
124 
125 	if (init_queue(&kq->queue, prop) != 0)
126 		goto err_init_queue;
127 
128 	kq->queue->device = dev;
129 	kq->queue->process = kfd_get_process(current);
130 
131 	retval = kq->mqd->init_mqd(kq->mqd, &kq->queue->mqd,
132 					&kq->queue->mqd_mem_obj,
133 					&kq->queue->gart_mqd_addr,
134 					&kq->queue->properties);
135 	if (retval != 0)
136 		goto err_init_mqd;
137 
138 	/* assign HIQ to HQD */
139 	if (type == KFD_QUEUE_TYPE_HIQ) {
140 		pr_debug("assigning hiq to hqd\n");
141 		kq->queue->pipe = KFD_CIK_HIQ_PIPE;
142 		kq->queue->queue = KFD_CIK_HIQ_QUEUE;
143 		kq->mqd->load_mqd(kq->mqd, kq->queue->mqd, kq->queue->pipe,
144 					kq->queue->queue, NULL);
145 	} else {
146 		/* allocate fence for DIQ */
147 
148 		retval = kfd2kgd->allocate_mem(dev->kgd,
149 					sizeof(uint32_t),
150 					32,
151 					KFD_MEMPOOL_SYSTEM_WRITECOMBINE,
152 					(struct kgd_mem **) &kq->fence_mem_obj);
153 
154 		if (retval != 0)
155 			goto err_alloc_fence;
156 
157 		kq->fence_kernel_address = kq->fence_mem_obj->cpu_ptr;
158 		kq->fence_gpu_addr = kq->fence_mem_obj->gpu_addr;
159 	}
160 
161 	print_queue(kq->queue);
162 
163 	return true;
164 err_alloc_fence:
165 err_init_mqd:
166 	uninit_queue(kq->queue);
167 err_init_queue:
168 	kfd2kgd->free_mem(dev->kgd, (struct kgd_mem *) kq->wptr_mem);
169 err_wptr_allocate_vidmem:
170 	kfd2kgd->free_mem(dev->kgd, (struct kgd_mem *) kq->rptr_mem);
171 err_rptr_allocate_vidmem:
172 	kfd2kgd->free_mem(dev->kgd, (struct kgd_mem *) kq->pq);
173 err_pq_allocate_vidmem:
174 	pr_err("kfd: error init pq\n");
175 	kfd_release_kernel_doorbell(dev, prop.doorbell_ptr);
176 err_get_kernel_doorbell:
177 	pr_err("kfd: error init doorbell");
178 	return false;
179 
180 }
181 
182 static void uninitialize(struct kernel_queue *kq)
183 {
184 	BUG_ON(!kq);
185 
186 	if (kq->queue->properties.type == KFD_QUEUE_TYPE_HIQ)
187 		kq->mqd->destroy_mqd(kq->mqd,
188 					NULL,
189 					false,
190 					QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
191 					kq->queue->pipe,
192 					kq->queue->queue);
193 
194 	kfd2kgd->free_mem(kq->dev->kgd, (struct kgd_mem *) kq->rptr_mem);
195 	kfd2kgd->free_mem(kq->dev->kgd, (struct kgd_mem *) kq->wptr_mem);
196 	kfd2kgd->free_mem(kq->dev->kgd, (struct kgd_mem *) kq->pq);
197 	kfd_release_kernel_doorbell(kq->dev,
198 					kq->queue->properties.doorbell_ptr);
199 	uninit_queue(kq->queue);
200 }
201 
202 static int acquire_packet_buffer(struct kernel_queue *kq,
203 		size_t packet_size_in_dwords, unsigned int **buffer_ptr)
204 {
205 	size_t available_size;
206 	size_t queue_size_dwords;
207 	uint32_t wptr, rptr;
208 	unsigned int *queue_address;
209 
210 	BUG_ON(!kq || !buffer_ptr);
211 
212 	rptr = *kq->rptr_kernel;
213 	wptr = *kq->wptr_kernel;
214 	queue_address = (unsigned int *)kq->pq_kernel_addr;
215 	queue_size_dwords = kq->queue->properties.queue_size / sizeof(uint32_t);
216 
217 	pr_debug("kfd: In func %s\nrptr: %d\nwptr: %d\nqueue_address 0x%p\n",
218 			__func__, rptr, wptr, queue_address);
219 
220 	available_size = (rptr - 1 - wptr + queue_size_dwords) %
221 							queue_size_dwords;
222 
223 	if (packet_size_in_dwords >= queue_size_dwords ||
224 			packet_size_in_dwords >= available_size) {
225 		/*
226 		 * make sure calling functions know
227 		 * acquire_packet_buffer() failed
228 		 */
229 		*buffer_ptr = NULL;
230 		return -ENOMEM;
231 	}
232 
233 	if (wptr + packet_size_in_dwords >= queue_size_dwords) {
234 		while (wptr > 0) {
235 			queue_address[wptr] = kq->nop_packet;
236 			wptr = (wptr + 1) % queue_size_dwords;
237 		}
238 	}
239 
240 	*buffer_ptr = &queue_address[wptr];
241 	kq->pending_wptr = wptr + packet_size_in_dwords;
242 
243 	return 0;
244 }
245 
246 static void submit_packet(struct kernel_queue *kq)
247 {
248 #ifdef DEBUG
249 	int i;
250 #endif
251 
252 	BUG_ON(!kq);
253 
254 #ifdef DEBUG
255 	for (i = *kq->wptr_kernel; i < kq->pending_wptr; i++) {
256 		pr_debug("0x%2X ", kq->pq_kernel_addr[i]);
257 		if (i % 15 == 0)
258 			pr_debug("\n");
259 	}
260 	pr_debug("\n");
261 #endif
262 
263 	*kq->wptr_kernel = kq->pending_wptr;
264 	write_kernel_doorbell(kq->queue->properties.doorbell_ptr,
265 				kq->pending_wptr);
266 }
267 
268 static int sync_with_hw(struct kernel_queue *kq, unsigned long timeout_ms)
269 {
270 	unsigned long org_timeout_ms;
271 
272 	BUG_ON(!kq);
273 
274 	org_timeout_ms = timeout_ms;
275 	timeout_ms += jiffies * 1000 / HZ;
276 	while (*kq->wptr_kernel != *kq->rptr_kernel) {
277 		if (time_after(jiffies * 1000 / HZ, timeout_ms)) {
278 			pr_err("kfd: kernel_queue %s timeout expired %lu\n",
279 				__func__, org_timeout_ms);
280 			pr_err("kfd: wptr: %d rptr: %d\n",
281 				*kq->wptr_kernel, *kq->rptr_kernel);
282 			return -ETIME;
283 		}
284 		schedule();
285 	}
286 
287 	return 0;
288 }
289 
290 static void rollback_packet(struct kernel_queue *kq)
291 {
292 	BUG_ON(!kq);
293 	kq->pending_wptr = *kq->queue->properties.write_ptr;
294 }
295 
296 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
297 					enum kfd_queue_type type)
298 {
299 	struct kernel_queue *kq;
300 
301 	BUG_ON(!dev);
302 
303 	kq = kzalloc(sizeof(struct kernel_queue), GFP_KERNEL);
304 	if (!kq)
305 		return NULL;
306 
307 	kq->initialize = initialize;
308 	kq->uninitialize = uninitialize;
309 	kq->acquire_packet_buffer = acquire_packet_buffer;
310 	kq->submit_packet = submit_packet;
311 	kq->sync_with_hw = sync_with_hw;
312 	kq->rollback_packet = rollback_packet;
313 
314 	if (kq->initialize(kq, dev, type, KFD_KERNEL_QUEUE_SIZE) == false) {
315 		pr_err("kfd: failed to init kernel queue\n");
316 		kfree(kq);
317 		return NULL;
318 	}
319 	return kq;
320 }
321 
322 void kernel_queue_uninit(struct kernel_queue *kq)
323 {
324 	BUG_ON(!kq);
325 
326 	kq->uninitialize(kq);
327 	kfree(kq);
328 }
329 
330 static __attribute__((unused)) void test_kq(struct kfd_dev *dev)
331 {
332 	struct kernel_queue *kq;
333 	uint32_t *buffer, i;
334 	int retval;
335 
336 	BUG_ON(!dev);
337 
338 	pr_debug("kfd: starting kernel queue test\n");
339 
340 	kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_HIQ);
341 	BUG_ON(!kq);
342 
343 	retval = kq->acquire_packet_buffer(kq, 5, &buffer);
344 	BUG_ON(retval != 0);
345 	for (i = 0; i < 5; i++)
346 		buffer[i] = kq->nop_packet;
347 	kq->submit_packet(kq);
348 	kq->sync_with_hw(kq, 1000);
349 
350 	pr_debug("kfd: ending kernel queue test\n");
351 }
352 
353 
354