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 	ktime_t now = ktime_get();
145 
146 	spin_lock(&user_cq->wait_list_lock);
147 	list_for_each_entry(pend, &user_cq->wait_list_head, wait_list_node) {
148 		pend->fence.timestamp = now;
149 		complete_all(&pend->fence.completion);
150 	}
151 	spin_unlock(&user_cq->wait_list_lock);
152 }
153 
154 /**
155  * hl_irq_handler_user_cq - irq handler for user completion queues
156  *
157  * @irq: irq number
158  * @arg: pointer to user interrupt structure
159  *
160  */
hl_irq_handler_user_cq(int irq,void * arg)161 irqreturn_t hl_irq_handler_user_cq(int irq, void *arg)
162 {
163 	struct hl_user_interrupt *user_cq = arg;
164 	struct hl_device *hdev = user_cq->hdev;
165 
166 	dev_dbg(hdev->dev,
167 		"got user completion interrupt id %u",
168 		user_cq->interrupt_id);
169 
170 	/* Handle user cq interrupts registered on all interrupts */
171 	handle_user_cq(hdev, &hdev->common_user_interrupt);
172 
173 	/* Handle user cq interrupts registered on this specific interrupt */
174 	handle_user_cq(hdev, user_cq);
175 
176 	return IRQ_HANDLED;
177 }
178 
179 /**
180  * hl_irq_handler_default - default irq handler
181  *
182  * @irq: irq number
183  * @arg: pointer to user interrupt structure
184  *
185  */
hl_irq_handler_default(int irq,void * arg)186 irqreturn_t hl_irq_handler_default(int irq, void *arg)
187 {
188 	struct hl_user_interrupt *user_interrupt = arg;
189 	struct hl_device *hdev = user_interrupt->hdev;
190 	u32 interrupt_id = user_interrupt->interrupt_id;
191 
192 	dev_err(hdev->dev,
193 		"got invalid user interrupt %u",
194 		interrupt_id);
195 
196 	return IRQ_HANDLED;
197 }
198 
199 /**
200  * hl_irq_handler_eq - irq handler for event queue
201  *
202  * @irq: irq number
203  * @arg: pointer to event queue structure
204  *
205  */
hl_irq_handler_eq(int irq,void * arg)206 irqreturn_t hl_irq_handler_eq(int irq, void *arg)
207 {
208 	struct hl_eq *eq = arg;
209 	struct hl_device *hdev = eq->hdev;
210 	struct hl_eq_entry *eq_entry;
211 	struct hl_eq_entry *eq_base;
212 	struct hl_eqe_work *handle_eqe_work;
213 	bool entry_ready;
214 	u32 cur_eqe;
215 	u16 cur_eqe_index;
216 
217 	eq_base = eq->kernel_address;
218 
219 	while (1) {
220 		cur_eqe = le32_to_cpu(eq_base[eq->ci].hdr.ctl);
221 		entry_ready = !!FIELD_GET(EQ_CTL_READY_MASK, cur_eqe);
222 
223 		if (!entry_ready)
224 			break;
225 
226 		cur_eqe_index = FIELD_GET(EQ_CTL_INDEX_MASK, cur_eqe);
227 		if ((hdev->event_queue.check_eqe_index) &&
228 				(((eq->prev_eqe_index + 1) & EQ_CTL_INDEX_MASK)
229 							!= cur_eqe_index)) {
230 			dev_dbg(hdev->dev,
231 				"EQE 0x%x in queue is ready but index does not match %d!=%d",
232 				eq_base[eq->ci].hdr.ctl,
233 				((eq->prev_eqe_index + 1) & EQ_CTL_INDEX_MASK),
234 				cur_eqe_index);
235 			break;
236 		}
237 
238 		eq->prev_eqe_index++;
239 
240 		eq_entry = &eq_base[eq->ci];
241 
242 		/*
243 		 * Make sure we read EQ entry contents after we've
244 		 * checked the ownership bit.
245 		 */
246 		dma_rmb();
247 
248 		if (hdev->disabled) {
249 			dev_warn(hdev->dev,
250 				"Device disabled but received IRQ %d for EQ\n",
251 					irq);
252 			goto skip_irq;
253 		}
254 
255 		handle_eqe_work = kmalloc(sizeof(*handle_eqe_work), GFP_ATOMIC);
256 		if (handle_eqe_work) {
257 			INIT_WORK(&handle_eqe_work->eq_work, irq_handle_eqe);
258 			handle_eqe_work->hdev = hdev;
259 
260 			memcpy(&handle_eqe_work->eq_entry, eq_entry,
261 					sizeof(*eq_entry));
262 
263 			queue_work(hdev->eq_wq, &handle_eqe_work->eq_work);
264 		}
265 skip_irq:
266 		/* Clear EQ entry ready bit */
267 		eq_entry->hdr.ctl =
268 			cpu_to_le32(le32_to_cpu(eq_entry->hdr.ctl) &
269 							~EQ_CTL_READY_MASK);
270 
271 		eq->ci = hl_eq_inc_ptr(eq->ci);
272 
273 		hdev->asic_funcs->update_eq_ci(hdev, eq->ci);
274 	}
275 
276 	return IRQ_HANDLED;
277 }
278 
279 /**
280  * hl_cq_init - main initialization function for an cq object
281  *
282  * @hdev: pointer to device structure
283  * @q: pointer to cq structure
284  * @hw_queue_id: The H/W queue ID this completion queue belongs to
285  *
286  * Allocate dma-able memory for the completion queue and initialize fields
287  * Returns 0 on success
288  */
hl_cq_init(struct hl_device * hdev,struct hl_cq * q,u32 hw_queue_id)289 int hl_cq_init(struct hl_device *hdev, struct hl_cq *q, u32 hw_queue_id)
290 {
291 	void *p;
292 
293 	p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev, HL_CQ_SIZE_IN_BYTES,
294 				&q->bus_address, GFP_KERNEL | __GFP_ZERO);
295 	if (!p)
296 		return -ENOMEM;
297 
298 	q->hdev = hdev;
299 	q->kernel_address = p;
300 	q->hw_queue_id = hw_queue_id;
301 	q->ci = 0;
302 	q->pi = 0;
303 
304 	atomic_set(&q->free_slots_cnt, HL_CQ_LENGTH);
305 
306 	return 0;
307 }
308 
309 /**
310  * hl_cq_fini - destroy completion queue
311  *
312  * @hdev: pointer to device structure
313  * @q: pointer to cq structure
314  *
315  * Free the completion queue memory
316  */
hl_cq_fini(struct hl_device * hdev,struct hl_cq * q)317 void hl_cq_fini(struct hl_device *hdev, struct hl_cq *q)
318 {
319 	hdev->asic_funcs->asic_dma_free_coherent(hdev, HL_CQ_SIZE_IN_BYTES,
320 						 q->kernel_address,
321 						 q->bus_address);
322 }
323 
hl_cq_reset(struct hl_device * hdev,struct hl_cq * q)324 void hl_cq_reset(struct hl_device *hdev, struct hl_cq *q)
325 {
326 	q->ci = 0;
327 	q->pi = 0;
328 
329 	atomic_set(&q->free_slots_cnt, HL_CQ_LENGTH);
330 
331 	/*
332 	 * It's not enough to just reset the PI/CI because the H/W may have
333 	 * written valid completion entries before it was halted and therefore
334 	 * we need to clean the actual queues so we won't process old entries
335 	 * when the device is operational again
336 	 */
337 
338 	memset(q->kernel_address, 0, HL_CQ_SIZE_IN_BYTES);
339 }
340 
341 /**
342  * hl_eq_init - main initialization function for an event queue object
343  *
344  * @hdev: pointer to device structure
345  * @q: pointer to eq structure
346  *
347  * Allocate dma-able memory for the event queue and initialize fields
348  * Returns 0 on success
349  */
hl_eq_init(struct hl_device * hdev,struct hl_eq * q)350 int hl_eq_init(struct hl_device *hdev, struct hl_eq *q)
351 {
352 	void *p;
353 
354 	p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
355 							HL_EQ_SIZE_IN_BYTES,
356 							&q->bus_address);
357 	if (!p)
358 		return -ENOMEM;
359 
360 	q->hdev = hdev;
361 	q->kernel_address = p;
362 	q->ci = 0;
363 	q->prev_eqe_index = 0;
364 
365 	return 0;
366 }
367 
368 /**
369  * hl_eq_fini - destroy event queue
370  *
371  * @hdev: pointer to device structure
372  * @q: pointer to eq structure
373  *
374  * Free the event queue memory
375  */
hl_eq_fini(struct hl_device * hdev,struct hl_eq * q)376 void hl_eq_fini(struct hl_device *hdev, struct hl_eq *q)
377 {
378 	flush_workqueue(hdev->eq_wq);
379 
380 	hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
381 					HL_EQ_SIZE_IN_BYTES,
382 					q->kernel_address);
383 }
384 
hl_eq_reset(struct hl_device * hdev,struct hl_eq * q)385 void hl_eq_reset(struct hl_device *hdev, struct hl_eq *q)
386 {
387 	q->ci = 0;
388 	q->prev_eqe_index = 0;
389 
390 	/*
391 	 * It's not enough to just reset the PI/CI because the H/W may have
392 	 * written valid completion entries before it was halted and therefore
393 	 * we need to clean the actual queues so we won't process old entries
394 	 * when the device is operational again
395 	 */
396 
397 	memset(q->kernel_address, 0, HL_EQ_SIZE_IN_BYTES);
398 }
399