• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /*
4  * Copyright 2016-2019 HabanaLabs, Ltd.
5  * All Rights Reserved.
6  */
7 
8 #include "habanalabs.h"
9 
10 #include <linux/slab.h>
11 
12 /**
13  * struct hl_eqe_work - This structure is used to schedule work of EQ
14  *                      entry and cpucp_reset event
15  *
16  * @eq_work:          workqueue object to run when EQ entry is received
17  * @hdev:             pointer to device structure
18  * @eq_entry:         copy of the EQ entry
19  */
20 struct hl_eqe_work {
21 	struct work_struct	eq_work;
22 	struct hl_device	*hdev;
23 	struct hl_eq_entry	eq_entry;
24 };
25 
26 /**
27  * hl_cq_inc_ptr - increment ci or pi of cq
28  *
29  * @ptr: the current ci or pi value of the completion queue
30  *
31  * Increment ptr by 1. If it reaches the number of completion queue
32  * entries, set it to 0
33  */
hl_cq_inc_ptr(u32 ptr)34 inline u32 hl_cq_inc_ptr(u32 ptr)
35 {
36 	ptr++;
37 	if (unlikely(ptr == HL_CQ_LENGTH))
38 		ptr = 0;
39 	return ptr;
40 }
41 
42 /**
43  * hl_eq_inc_ptr - increment ci of eq
44  *
45  * @ptr: the current ci value of the event queue
46  *
47  * Increment ptr by 1. If it reaches the number of event queue
48  * entries, set it to 0
49  */
hl_eq_inc_ptr(u32 ptr)50 static inline u32 hl_eq_inc_ptr(u32 ptr)
51 {
52 	ptr++;
53 	if (unlikely(ptr == HL_EQ_LENGTH))
54 		ptr = 0;
55 	return ptr;
56 }
57 
irq_handle_eqe(struct work_struct * work)58 static void irq_handle_eqe(struct work_struct *work)
59 {
60 	struct hl_eqe_work *eqe_work = container_of(work, struct hl_eqe_work,
61 							eq_work);
62 	struct hl_device *hdev = eqe_work->hdev;
63 
64 	hdev->asic_funcs->handle_eqe(hdev, &eqe_work->eq_entry);
65 
66 	kfree(eqe_work);
67 }
68 
69 /**
70  * hl_irq_handler_cq - irq handler for completion queue
71  *
72  * @irq: irq number
73  * @arg: pointer to completion queue structure
74  *
75  */
hl_irq_handler_cq(int irq,void * arg)76 irqreturn_t hl_irq_handler_cq(int irq, void *arg)
77 {
78 	struct hl_cq *cq = arg;
79 	struct hl_device *hdev = cq->hdev;
80 	struct hl_hw_queue *queue;
81 	struct hl_cs_job *job;
82 	bool shadow_index_valid;
83 	u16 shadow_index;
84 	struct hl_cq_entry *cq_entry, *cq_base;
85 
86 	if (hdev->disabled) {
87 		dev_dbg(hdev->dev,
88 			"Device disabled but received IRQ %d for CQ %d\n",
89 			irq, cq->hw_queue_id);
90 		return IRQ_HANDLED;
91 	}
92 
93 	cq_base = cq->kernel_address;
94 
95 	while (1) {
96 		bool entry_ready = ((le32_to_cpu(cq_base[cq->ci].data) &
97 					CQ_ENTRY_READY_MASK)
98 						>> CQ_ENTRY_READY_SHIFT);
99 
100 		if (!entry_ready)
101 			break;
102 
103 		cq_entry = (struct hl_cq_entry *) &cq_base[cq->ci];
104 
105 		/* Make sure we read CQ entry contents after we've
106 		 * checked the ownership bit.
107 		 */
108 		dma_rmb();
109 
110 		shadow_index_valid = ((le32_to_cpu(cq_entry->data) &
111 					CQ_ENTRY_SHADOW_INDEX_VALID_MASK)
112 					>> CQ_ENTRY_SHADOW_INDEX_VALID_SHIFT);
113 
114 		shadow_index = (u16) ((le32_to_cpu(cq_entry->data) &
115 					CQ_ENTRY_SHADOW_INDEX_MASK)
116 					>> CQ_ENTRY_SHADOW_INDEX_SHIFT);
117 
118 		queue = &hdev->kernel_queues[cq->hw_queue_id];
119 
120 		if ((shadow_index_valid) && (!hdev->disabled)) {
121 			job = queue->shadow_queue[hl_pi_2_offset(shadow_index)];
122 			queue_work(hdev->cq_wq[cq->cq_idx], &job->finish_work);
123 		}
124 
125 		atomic_inc(&queue->ci);
126 
127 		/* Clear CQ entry ready bit */
128 		cq_entry->data = cpu_to_le32(le32_to_cpu(cq_entry->data) &
129 						~CQ_ENTRY_READY_MASK);
130 
131 		cq->ci = hl_cq_inc_ptr(cq->ci);
132 
133 		/* Increment free slots */
134 		atomic_inc(&cq->free_slots_cnt);
135 	}
136 
137 	return IRQ_HANDLED;
138 }
139 
handle_user_cq(struct hl_device * hdev,struct hl_user_interrupt * user_cq)140 static void handle_user_cq(struct hl_device *hdev,
141 			struct hl_user_interrupt *user_cq)
142 {
143 	struct hl_user_pending_interrupt *pend;
144 
145 	spin_lock(&user_cq->wait_list_lock);
146 	list_for_each_entry(pend, &user_cq->wait_list_head, wait_list_node)
147 		complete_all(&pend->fence.completion);
148 	spin_unlock(&user_cq->wait_list_lock);
149 }
150 
151 /**
152  * hl_irq_handler_user_cq - irq handler for user completion queues
153  *
154  * @irq: irq number
155  * @arg: pointer to user interrupt structure
156  *
157  */
hl_irq_handler_user_cq(int irq,void * arg)158 irqreturn_t hl_irq_handler_user_cq(int irq, void *arg)
159 {
160 	struct hl_user_interrupt *user_cq = arg;
161 	struct hl_device *hdev = user_cq->hdev;
162 
163 	dev_dbg(hdev->dev,
164 		"got user completion interrupt id %u",
165 		user_cq->interrupt_id);
166 
167 	/* Handle user cq interrupts registered on all interrupts */
168 	handle_user_cq(hdev, &hdev->common_user_interrupt);
169 
170 	/* Handle user cq interrupts registered on this specific interrupt */
171 	handle_user_cq(hdev, user_cq);
172 
173 	return IRQ_HANDLED;
174 }
175 
176 /**
177  * hl_irq_handler_default - default irq handler
178  *
179  * @irq: irq number
180  * @arg: pointer to user interrupt structure
181  *
182  */
hl_irq_handler_default(int irq,void * arg)183 irqreturn_t hl_irq_handler_default(int irq, void *arg)
184 {
185 	struct hl_user_interrupt *user_interrupt = arg;
186 	struct hl_device *hdev = user_interrupt->hdev;
187 	u32 interrupt_id = user_interrupt->interrupt_id;
188 
189 	dev_err(hdev->dev,
190 		"got invalid user interrupt %u",
191 		interrupt_id);
192 
193 	return IRQ_HANDLED;
194 }
195 
196 /**
197  * hl_irq_handler_eq - irq handler for event queue
198  *
199  * @irq: irq number
200  * @arg: pointer to event queue structure
201  *
202  */
hl_irq_handler_eq(int irq,void * arg)203 irqreturn_t hl_irq_handler_eq(int irq, void *arg)
204 {
205 	struct hl_eq *eq = arg;
206 	struct hl_device *hdev = eq->hdev;
207 	struct hl_eq_entry *eq_entry;
208 	struct hl_eq_entry *eq_base;
209 	struct hl_eqe_work *handle_eqe_work;
210 	bool entry_ready;
211 	u32 cur_eqe;
212 	u16 cur_eqe_index;
213 
214 	eq_base = eq->kernel_address;
215 
216 	while (1) {
217 		cur_eqe = le32_to_cpu(eq_base[eq->ci].hdr.ctl);
218 		entry_ready = !!FIELD_GET(EQ_CTL_READY_MASK, cur_eqe);
219 
220 		if (!entry_ready)
221 			break;
222 
223 		cur_eqe_index = FIELD_GET(EQ_CTL_INDEX_MASK, cur_eqe);
224 		if ((hdev->event_queue.check_eqe_index) &&
225 				(((eq->prev_eqe_index + 1) & EQ_CTL_INDEX_MASK)
226 							!= cur_eqe_index)) {
227 			dev_dbg(hdev->dev,
228 				"EQE 0x%x in queue is ready but index does not match %d!=%d",
229 				eq_base[eq->ci].hdr.ctl,
230 				((eq->prev_eqe_index + 1) & EQ_CTL_INDEX_MASK),
231 				cur_eqe_index);
232 			break;
233 		}
234 
235 		eq->prev_eqe_index++;
236 
237 		eq_entry = &eq_base[eq->ci];
238 
239 		/*
240 		 * Make sure we read EQ entry contents after we've
241 		 * checked the ownership bit.
242 		 */
243 		dma_rmb();
244 
245 		if (hdev->disabled) {
246 			dev_warn(hdev->dev,
247 				"Device disabled but received IRQ %d for EQ\n",
248 					irq);
249 			goto skip_irq;
250 		}
251 
252 		handle_eqe_work = kmalloc(sizeof(*handle_eqe_work), GFP_ATOMIC);
253 		if (handle_eqe_work) {
254 			INIT_WORK(&handle_eqe_work->eq_work, irq_handle_eqe);
255 			handle_eqe_work->hdev = hdev;
256 
257 			memcpy(&handle_eqe_work->eq_entry, eq_entry,
258 					sizeof(*eq_entry));
259 
260 			queue_work(hdev->eq_wq, &handle_eqe_work->eq_work);
261 		}
262 skip_irq:
263 		/* Clear EQ entry ready bit */
264 		eq_entry->hdr.ctl =
265 			cpu_to_le32(le32_to_cpu(eq_entry->hdr.ctl) &
266 							~EQ_CTL_READY_MASK);
267 
268 		eq->ci = hl_eq_inc_ptr(eq->ci);
269 
270 		hdev->asic_funcs->update_eq_ci(hdev, eq->ci);
271 	}
272 
273 	return IRQ_HANDLED;
274 }
275 
276 /**
277  * hl_cq_init - main initialization function for an cq object
278  *
279  * @hdev: pointer to device structure
280  * @q: pointer to cq structure
281  * @hw_queue_id: The H/W queue ID this completion queue belongs to
282  *
283  * Allocate dma-able memory for the completion queue and initialize fields
284  * Returns 0 on success
285  */
hl_cq_init(struct hl_device * hdev,struct hl_cq * q,u32 hw_queue_id)286 int hl_cq_init(struct hl_device *hdev, struct hl_cq *q, u32 hw_queue_id)
287 {
288 	void *p;
289 
290 	p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev, HL_CQ_SIZE_IN_BYTES,
291 				&q->bus_address, GFP_KERNEL | __GFP_ZERO);
292 	if (!p)
293 		return -ENOMEM;
294 
295 	q->hdev = hdev;
296 	q->kernel_address = p;
297 	q->hw_queue_id = hw_queue_id;
298 	q->ci = 0;
299 	q->pi = 0;
300 
301 	atomic_set(&q->free_slots_cnt, HL_CQ_LENGTH);
302 
303 	return 0;
304 }
305 
306 /**
307  * hl_cq_fini - destroy completion queue
308  *
309  * @hdev: pointer to device structure
310  * @q: pointer to cq structure
311  *
312  * Free the completion queue memory
313  */
hl_cq_fini(struct hl_device * hdev,struct hl_cq * q)314 void hl_cq_fini(struct hl_device *hdev, struct hl_cq *q)
315 {
316 	hdev->asic_funcs->asic_dma_free_coherent(hdev, HL_CQ_SIZE_IN_BYTES,
317 						 q->kernel_address,
318 						 q->bus_address);
319 }
320 
hl_cq_reset(struct hl_device * hdev,struct hl_cq * q)321 void hl_cq_reset(struct hl_device *hdev, struct hl_cq *q)
322 {
323 	q->ci = 0;
324 	q->pi = 0;
325 
326 	atomic_set(&q->free_slots_cnt, HL_CQ_LENGTH);
327 
328 	/*
329 	 * It's not enough to just reset the PI/CI because the H/W may have
330 	 * written valid completion entries before it was halted and therefore
331 	 * we need to clean the actual queues so we won't process old entries
332 	 * when the device is operational again
333 	 */
334 
335 	memset(q->kernel_address, 0, HL_CQ_SIZE_IN_BYTES);
336 }
337 
338 /**
339  * hl_eq_init - main initialization function for an event queue object
340  *
341  * @hdev: pointer to device structure
342  * @q: pointer to eq structure
343  *
344  * Allocate dma-able memory for the event queue and initialize fields
345  * Returns 0 on success
346  */
hl_eq_init(struct hl_device * hdev,struct hl_eq * q)347 int hl_eq_init(struct hl_device *hdev, struct hl_eq *q)
348 {
349 	void *p;
350 
351 	p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
352 							HL_EQ_SIZE_IN_BYTES,
353 							&q->bus_address);
354 	if (!p)
355 		return -ENOMEM;
356 
357 	q->hdev = hdev;
358 	q->kernel_address = p;
359 	q->ci = 0;
360 	q->prev_eqe_index = 0;
361 
362 	return 0;
363 }
364 
365 /**
366  * hl_eq_fini - destroy event queue
367  *
368  * @hdev: pointer to device structure
369  * @q: pointer to eq structure
370  *
371  * Free the event queue memory
372  */
hl_eq_fini(struct hl_device * hdev,struct hl_eq * q)373 void hl_eq_fini(struct hl_device *hdev, struct hl_eq *q)
374 {
375 	flush_workqueue(hdev->eq_wq);
376 
377 	hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
378 					HL_EQ_SIZE_IN_BYTES,
379 					q->kernel_address);
380 }
381 
hl_eq_reset(struct hl_device * hdev,struct hl_eq * q)382 void hl_eq_reset(struct hl_device *hdev, struct hl_eq *q)
383 {
384 	q->ci = 0;
385 	q->prev_eqe_index = 0;
386 
387 	/*
388 	 * It's not enough to just reset the PI/CI because the H/W may have
389 	 * written valid completion entries before it was halted and therefore
390 	 * we need to clean the actual queues so we won't process old entries
391 	 * when the device is operational again
392 	 */
393 
394 	memset(q->kernel_address, 0, HL_EQ_SIZE_IN_BYTES);
395 }
396