xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_queue.c (revision b3ee1e4609512dfff642a96b34d7e5dfcdc92d05)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  */
24 
25 #include <linux/slab.h>
26 #include "kfd_priv.h"
27 #include "kfd_topology.h"
28 #include "kfd_svm.h"
29 
print_queue_properties(struct queue_properties * q)30 void print_queue_properties(struct queue_properties *q)
31 {
32 	if (!q)
33 		return;
34 
35 	pr_debug("Printing queue properties:\n");
36 	pr_debug("Queue Type: %u\n", q->type);
37 	pr_debug("Queue Size: %llu\n", q->queue_size);
38 	pr_debug("Queue percent: %u\n", q->queue_percent);
39 	pr_debug("Queue Address: 0x%llX\n", q->queue_address);
40 	pr_debug("Queue Id: %u\n", q->queue_id);
41 	pr_debug("Queue Process Vmid: %u\n", q->vmid);
42 	pr_debug("Queue Read Pointer: 0x%px\n", q->read_ptr);
43 	pr_debug("Queue Write Pointer: 0x%px\n", q->write_ptr);
44 	pr_debug("Queue Doorbell Pointer: 0x%p\n", q->doorbell_ptr);
45 	pr_debug("Queue Doorbell Offset: %u\n", q->doorbell_off);
46 }
47 
print_queue(struct queue * q)48 void print_queue(struct queue *q)
49 {
50 	if (!q)
51 		return;
52 	pr_debug("Printing queue:\n");
53 	pr_debug("Queue Type: %u\n", q->properties.type);
54 	pr_debug("Queue Size: %llu\n", q->properties.queue_size);
55 	pr_debug("Queue percent: %u\n", q->properties.queue_percent);
56 	pr_debug("Queue Address: 0x%llX\n", q->properties.queue_address);
57 	pr_debug("Queue Id: %u\n", q->properties.queue_id);
58 	pr_debug("Queue Process Vmid: %u\n", q->properties.vmid);
59 	pr_debug("Queue Read Pointer: 0x%px\n", q->properties.read_ptr);
60 	pr_debug("Queue Write Pointer: 0x%px\n", q->properties.write_ptr);
61 	pr_debug("Queue Doorbell Pointer: 0x%p\n", q->properties.doorbell_ptr);
62 	pr_debug("Queue Doorbell Offset: %u\n", q->properties.doorbell_off);
63 	pr_debug("Queue MQD Address: 0x%p\n", q->mqd);
64 	pr_debug("Queue MQD Gart: 0x%llX\n", q->gart_mqd_addr);
65 	pr_debug("Queue Process Address: 0x%p\n", q->process);
66 	pr_debug("Queue Device Address: 0x%p\n", q->device);
67 }
68 
init_queue(struct queue ** q,const struct queue_properties * properties)69 int init_queue(struct queue **q, const struct queue_properties *properties)
70 {
71 	struct queue *tmp_q;
72 
73 	tmp_q = kzalloc(sizeof(*tmp_q), GFP_KERNEL);
74 	if (!tmp_q)
75 		return -ENOMEM;
76 
77 	memcpy(&tmp_q->properties, properties, sizeof(*properties));
78 
79 	*q = tmp_q;
80 	return 0;
81 }
82 
uninit_queue(struct queue * q)83 void uninit_queue(struct queue *q)
84 {
85 	kfree(q);
86 }
87 
88 #if IS_ENABLED(CONFIG_HSA_AMD_SVM)
89 
kfd_queue_buffer_svm_get(struct kfd_process_device * pdd,u64 addr,u64 size)90 static int kfd_queue_buffer_svm_get(struct kfd_process_device *pdd, u64 addr, u64 size)
91 {
92 	struct kfd_process *p = pdd->process;
93 	struct list_head update_list;
94 	struct svm_range *prange;
95 	int ret = -EINVAL;
96 
97 	INIT_LIST_HEAD(&update_list);
98 	addr >>= PAGE_SHIFT;
99 	size >>= PAGE_SHIFT;
100 
101 	mutex_lock(&p->svms.lock);
102 
103 	/*
104 	 * range may split to multiple svm pranges aligned to granularity boundaery.
105 	 */
106 	while (size) {
107 		uint32_t gpuid, gpuidx;
108 		int r;
109 
110 		prange = svm_range_from_addr(&p->svms, addr, NULL);
111 		if (!prange)
112 			break;
113 
114 		if (!prange->mapped_to_gpu)
115 			break;
116 
117 		r = kfd_process_gpuid_from_node(p, pdd->dev, &gpuid, &gpuidx);
118 		if (r < 0)
119 			break;
120 		if (!test_bit(gpuidx, prange->bitmap_access) &&
121 		    !test_bit(gpuidx, prange->bitmap_aip))
122 			break;
123 
124 		if (!(prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED))
125 			break;
126 
127 		list_add(&prange->update_list, &update_list);
128 
129 		if (prange->last - prange->start + 1 >= size) {
130 			size = 0;
131 			break;
132 		}
133 
134 		size -= prange->last - prange->start + 1;
135 		addr += prange->last - prange->start + 1;
136 	}
137 	if (size) {
138 		pr_debug("[0x%llx 0x%llx] not registered\n", addr, addr + size - 1);
139 		goto out_unlock;
140 	}
141 
142 	list_for_each_entry(prange, &update_list, update_list)
143 		atomic_inc(&prange->queue_refcount);
144 	ret = 0;
145 
146 out_unlock:
147 	mutex_unlock(&p->svms.lock);
148 	return ret;
149 }
150 
kfd_queue_buffer_svm_put(struct kfd_process_device * pdd,u64 addr,u64 size)151 static void kfd_queue_buffer_svm_put(struct kfd_process_device *pdd, u64 addr, u64 size)
152 {
153 	struct kfd_process *p = pdd->process;
154 	struct svm_range *prange, *pchild;
155 	struct interval_tree_node *node;
156 	unsigned long last;
157 
158 	addr >>= PAGE_SHIFT;
159 	last = addr + (size >> PAGE_SHIFT) - 1;
160 
161 	mutex_lock(&p->svms.lock);
162 
163 	node = interval_tree_iter_first(&p->svms.objects, addr, last);
164 	while (node) {
165 		struct interval_tree_node *next_node;
166 		unsigned long next_start;
167 
168 		prange = container_of(node, struct svm_range, it_node);
169 		next_node = interval_tree_iter_next(node, addr, last);
170 		next_start = min(node->last, last) + 1;
171 
172 		if (atomic_add_unless(&prange->queue_refcount, -1, 0)) {
173 			list_for_each_entry(pchild, &prange->child_list, child_list)
174 				atomic_add_unless(&pchild->queue_refcount, -1, 0);
175 		}
176 
177 		node = next_node;
178 		addr = next_start;
179 	}
180 
181 	mutex_unlock(&p->svms.lock);
182 }
183 #else
184 
kfd_queue_buffer_svm_get(struct kfd_process_device * pdd,u64 addr,u64 size)185 static int kfd_queue_buffer_svm_get(struct kfd_process_device *pdd, u64 addr, u64 size)
186 {
187 	return -EINVAL;
188 }
189 
kfd_queue_buffer_svm_put(struct kfd_process_device * pdd,u64 addr,u64 size)190 static void kfd_queue_buffer_svm_put(struct kfd_process_device *pdd, u64 addr, u64 size)
191 {
192 }
193 
194 #endif
195 
kfd_queue_buffer_get(struct amdgpu_vm * vm,void __user * addr,struct amdgpu_bo ** pbo,u64 expected_size)196 int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo,
197 			 u64 expected_size)
198 {
199 	struct amdgpu_bo_va_mapping *mapping;
200 	u64 user_addr;
201 	u64 size;
202 
203 	user_addr = (u64)addr >> AMDGPU_GPU_PAGE_SHIFT;
204 	size = expected_size >> AMDGPU_GPU_PAGE_SHIFT;
205 
206 	mapping = amdgpu_vm_bo_lookup_mapping(vm, user_addr);
207 	if (!mapping)
208 		goto out_err;
209 
210 	if (user_addr != mapping->start ||
211 	    (size != 0 && user_addr + size - 1 != mapping->last)) {
212 		pr_debug("expected size 0x%llx not equal to mapping addr 0x%llx size 0x%llx\n",
213 			expected_size, mapping->start << AMDGPU_GPU_PAGE_SHIFT,
214 			(mapping->last - mapping->start + 1) << AMDGPU_GPU_PAGE_SHIFT);
215 		goto out_err;
216 	}
217 
218 	*pbo = amdgpu_bo_ref(mapping->bo_va->base.bo);
219 	mapping->bo_va->queue_refcount++;
220 	return 0;
221 
222 out_err:
223 	*pbo = NULL;
224 	return -EINVAL;
225 }
226 
227 /* FIXME: remove this function, just call amdgpu_bo_unref directly */
kfd_queue_buffer_put(struct amdgpu_bo ** bo)228 void kfd_queue_buffer_put(struct amdgpu_bo **bo)
229 {
230 	amdgpu_bo_unref(bo);
231 }
232 
kfd_queue_acquire_buffers(struct kfd_process_device * pdd,struct queue_properties * properties)233 int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties)
234 {
235 	struct kfd_topology_device *topo_dev;
236 	u64 expected_queue_size;
237 	struct amdgpu_vm *vm;
238 	u32 total_cwsr_size;
239 	int err;
240 
241 	topo_dev = kfd_topology_device_by_id(pdd->dev->id);
242 	if (!topo_dev)
243 		return -EINVAL;
244 
245 	/* AQL queues on GFX7 and GFX8 appear twice their actual size */
246 	if (properties->type == KFD_QUEUE_TYPE_COMPUTE &&
247 	    properties->format == KFD_QUEUE_FORMAT_AQL &&
248 	    topo_dev->node_props.gfx_target_version >= 70000 &&
249 	    topo_dev->node_props.gfx_target_version < 90000)
250 		expected_queue_size = properties->queue_size / 2;
251 	else
252 		expected_queue_size = properties->queue_size;
253 
254 	vm = drm_priv_to_vm(pdd->drm_priv);
255 	err = amdgpu_bo_reserve(vm->root.bo, false);
256 	if (err)
257 		return err;
258 
259 	err = kfd_queue_buffer_get(vm, properties->write_ptr, &properties->wptr_bo, PAGE_SIZE);
260 	if (err)
261 		goto out_err_unreserve;
262 
263 	err = kfd_queue_buffer_get(vm, properties->read_ptr, &properties->rptr_bo, PAGE_SIZE);
264 	if (err)
265 		goto out_err_unreserve;
266 
267 	err = kfd_queue_buffer_get(vm, (void *)properties->queue_address,
268 				   &properties->ring_bo, expected_queue_size);
269 	if (err)
270 		goto out_err_unreserve;
271 
272 	/* only compute queue requires EOP buffer and CWSR area */
273 	if (properties->type != KFD_QUEUE_TYPE_COMPUTE)
274 		goto out_unreserve;
275 
276 	/* EOP buffer is not required for all ASICs */
277 	if (properties->eop_ring_buffer_address) {
278 		if (properties->eop_ring_buffer_size != topo_dev->node_props.eop_buffer_size) {
279 			pr_debug("queue eop bo size 0x%x not equal to node eop buf size 0x%x\n",
280 				properties->eop_ring_buffer_size,
281 				topo_dev->node_props.eop_buffer_size);
282 			err = -EINVAL;
283 			goto out_err_unreserve;
284 		}
285 		err = kfd_queue_buffer_get(vm, (void *)properties->eop_ring_buffer_address,
286 					   &properties->eop_buf_bo,
287 					   properties->eop_ring_buffer_size);
288 		if (err)
289 			goto out_err_unreserve;
290 	}
291 
292 	if (properties->ctl_stack_size != topo_dev->node_props.ctl_stack_size) {
293 		pr_debug("queue ctl stack size 0x%x not equal to node ctl stack size 0x%x\n",
294 			properties->ctl_stack_size,
295 			topo_dev->node_props.ctl_stack_size);
296 		err = -EINVAL;
297 		goto out_err_unreserve;
298 	}
299 
300 	if (properties->ctx_save_restore_area_size != topo_dev->node_props.cwsr_size) {
301 		pr_debug("queue cwsr size 0x%x not equal to node cwsr size 0x%x\n",
302 			properties->ctx_save_restore_area_size,
303 			topo_dev->node_props.cwsr_size);
304 		err = -EINVAL;
305 		goto out_err_unreserve;
306 	}
307 
308 	total_cwsr_size = (topo_dev->node_props.cwsr_size + topo_dev->node_props.debug_memory_size)
309 			  * NUM_XCC(pdd->dev->xcc_mask);
310 	total_cwsr_size = ALIGN(total_cwsr_size, PAGE_SIZE);
311 
312 	err = kfd_queue_buffer_get(vm, (void *)properties->ctx_save_restore_area_address,
313 				   &properties->cwsr_bo, total_cwsr_size);
314 	if (!err)
315 		goto out_unreserve;
316 
317 	amdgpu_bo_unreserve(vm->root.bo);
318 
319 	err = kfd_queue_buffer_svm_get(pdd, properties->ctx_save_restore_area_address,
320 				       total_cwsr_size);
321 	if (err)
322 		goto out_err_release;
323 
324 	return 0;
325 
326 out_unreserve:
327 	amdgpu_bo_unreserve(vm->root.bo);
328 	return 0;
329 
330 out_err_unreserve:
331 	amdgpu_bo_unreserve(vm->root.bo);
332 out_err_release:
333 	/* FIXME: make a _locked version of this that can be called before
334 	 * dropping the VM reservation.
335 	 */
336 	kfd_queue_unref_bo_vas(pdd, properties);
337 	kfd_queue_release_buffers(pdd, properties);
338 	return err;
339 }
340 
kfd_queue_release_buffers(struct kfd_process_device * pdd,struct queue_properties * properties)341 int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties)
342 {
343 	struct kfd_topology_device *topo_dev;
344 	u32 total_cwsr_size;
345 
346 	kfd_queue_buffer_put(&properties->wptr_bo);
347 	kfd_queue_buffer_put(&properties->rptr_bo);
348 	kfd_queue_buffer_put(&properties->ring_bo);
349 	kfd_queue_buffer_put(&properties->eop_buf_bo);
350 	kfd_queue_buffer_put(&properties->cwsr_bo);
351 
352 	topo_dev = kfd_topology_device_by_id(pdd->dev->id);
353 	if (!topo_dev)
354 		return -EINVAL;
355 	total_cwsr_size = (topo_dev->node_props.cwsr_size + topo_dev->node_props.debug_memory_size)
356 			  * NUM_XCC(pdd->dev->xcc_mask);
357 	total_cwsr_size = ALIGN(total_cwsr_size, PAGE_SIZE);
358 
359 	kfd_queue_buffer_svm_put(pdd, properties->ctx_save_restore_area_address, total_cwsr_size);
360 	return 0;
361 }
362 
kfd_queue_unref_bo_va(struct amdgpu_vm * vm,struct amdgpu_bo ** bo)363 void kfd_queue_unref_bo_va(struct amdgpu_vm *vm, struct amdgpu_bo **bo)
364 {
365 	if (*bo) {
366 		struct amdgpu_bo_va *bo_va;
367 
368 		bo_va = amdgpu_vm_bo_find(vm, *bo);
369 		if (bo_va && bo_va->queue_refcount)
370 			bo_va->queue_refcount--;
371 	}
372 }
373 
kfd_queue_unref_bo_vas(struct kfd_process_device * pdd,struct queue_properties * properties)374 int kfd_queue_unref_bo_vas(struct kfd_process_device *pdd,
375 			   struct queue_properties *properties)
376 {
377 	struct amdgpu_vm *vm;
378 	int err;
379 
380 	vm = drm_priv_to_vm(pdd->drm_priv);
381 	err = amdgpu_bo_reserve(vm->root.bo, false);
382 	if (err)
383 		return err;
384 
385 	kfd_queue_unref_bo_va(vm, &properties->wptr_bo);
386 	kfd_queue_unref_bo_va(vm, &properties->rptr_bo);
387 	kfd_queue_unref_bo_va(vm, &properties->ring_bo);
388 	kfd_queue_unref_bo_va(vm, &properties->eop_buf_bo);
389 	kfd_queue_unref_bo_va(vm, &properties->cwsr_bo);
390 
391 	amdgpu_bo_unreserve(vm->root.bo);
392 	return 0;
393 }
394 
395 #define SGPR_SIZE_PER_CU	0x4000
396 #define LDS_SIZE_PER_CU		0x10000
397 #define HWREG_SIZE_PER_CU	0x1000
398 #define DEBUGGER_BYTES_ALIGN	64
399 #define DEBUGGER_BYTES_PER_WAVE	32
400 
kfd_get_vgpr_size_per_cu(u32 gfxv)401 static u32 kfd_get_vgpr_size_per_cu(u32 gfxv)
402 {
403 	u32 vgpr_size = 0x40000;
404 
405 	if ((gfxv / 100 * 100) == 90400 ||	/* GFX_VERSION_AQUA_VANJARAM */
406 	    gfxv == 90010 ||			/* GFX_VERSION_ALDEBARAN */
407 	    gfxv == 90008 ||			/* GFX_VERSION_ARCTURUS */
408 	    gfxv == 90500)
409 		vgpr_size = 0x80000;
410 	else if (gfxv == 110000 ||		/* GFX_VERSION_PLUM_BONITO */
411 		 gfxv == 110001 ||		/* GFX_VERSION_WHEAT_NAS */
412 		 gfxv == 120000 ||		/* GFX_VERSION_GFX1200 */
413 		 gfxv == 120001)		/* GFX_VERSION_GFX1201 */
414 		vgpr_size = 0x60000;
415 
416 	return vgpr_size;
417 }
418 
419 #define WG_CONTEXT_DATA_SIZE_PER_CU(gfxv, props)	\
420 	(kfd_get_vgpr_size_per_cu(gfxv) + SGPR_SIZE_PER_CU +\
421 	 (((gfxv) == 90500) ? (props->lds_size_in_kb << 10) : LDS_SIZE_PER_CU) +\
422 	 HWREG_SIZE_PER_CU)
423 
424 #define CNTL_STACK_BYTES_PER_WAVE(gfxv)	\
425 	((gfxv) >= 100100 ? 12 : 8)	/* GFX_VERSION_NAVI10*/
426 
427 #define SIZEOF_HSA_USER_CONTEXT_SAVE_AREA_HEADER 40
428 
kfd_queue_ctx_save_restore_size(struct kfd_topology_device * dev)429 void kfd_queue_ctx_save_restore_size(struct kfd_topology_device *dev)
430 {
431 	struct kfd_node_properties *props = &dev->node_props;
432 	u32 gfxv = props->gfx_target_version;
433 	u32 ctl_stack_size;
434 	u32 wg_data_size;
435 	u32 wave_num;
436 	u32 cu_num;
437 
438 	if (gfxv < 80001)	/* GFX_VERSION_CARRIZO */
439 		return;
440 
441 	cu_num = props->simd_count / props->simd_per_cu / NUM_XCC(dev->gpu->xcc_mask);
442 	wave_num = (gfxv < 100100) ?	/* GFX_VERSION_NAVI10 */
443 		    min(cu_num * 40, props->array_count / props->simd_arrays_per_engine * 512)
444 		    : cu_num * 32;
445 
446 	wg_data_size = ALIGN(cu_num * WG_CONTEXT_DATA_SIZE_PER_CU(gfxv, props), PAGE_SIZE);
447 	ctl_stack_size = wave_num * CNTL_STACK_BYTES_PER_WAVE(gfxv) + 8;
448 	ctl_stack_size = ALIGN(SIZEOF_HSA_USER_CONTEXT_SAVE_AREA_HEADER + ctl_stack_size,
449 			       PAGE_SIZE);
450 
451 	if ((gfxv / 10000 * 10000) == 100000) {
452 		/* HW design limits control stack size to 0x7000.
453 		 * This is insufficient for theoretical PM4 cases
454 		 * but sufficient for AQL, limited by SPI events.
455 		 */
456 		ctl_stack_size = min(ctl_stack_size, 0x7000);
457 	}
458 
459 	props->ctl_stack_size = ctl_stack_size;
460 	props->debug_memory_size = ALIGN(wave_num * DEBUGGER_BYTES_PER_WAVE, DEBUGGER_BYTES_ALIGN);
461 	props->cwsr_size = ctl_stack_size + wg_data_size;
462 
463 	if (gfxv == 80002)	/* GFX_VERSION_TONGA */
464 		props->eop_buffer_size = 0x8000;
465 	else if ((gfxv / 100 * 100) == 90400)	/* GFX_VERSION_AQUA_VANJARAM */
466 		props->eop_buffer_size = 4096;
467 	else if (gfxv >= 80000)
468 		props->eop_buffer_size = 4096;
469 }
470