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 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
140 /**
141 * hl_irq_handler_eq - irq handler for event queue
142 *
143 * @irq: irq number
144 * @arg: pointer to event queue structure
145 *
146 */
hl_irq_handler_eq(int irq,void * arg)147 irqreturn_t hl_irq_handler_eq(int irq, void *arg)
148 {
149 struct hl_eq *eq = arg;
150 struct hl_device *hdev = eq->hdev;
151 struct hl_eq_entry *eq_entry;
152 struct hl_eq_entry *eq_base;
153 struct hl_eqe_work *handle_eqe_work;
154
155 eq_base = eq->kernel_address;
156
157 while (1) {
158 bool entry_ready =
159 ((le32_to_cpu(eq_base[eq->ci].hdr.ctl) &
160 EQ_CTL_READY_MASK) >> EQ_CTL_READY_SHIFT);
161
162 if (!entry_ready)
163 break;
164
165 eq_entry = &eq_base[eq->ci];
166
167 /*
168 * Make sure we read EQ entry contents after we've
169 * checked the ownership bit.
170 */
171 dma_rmb();
172
173 if (hdev->disabled) {
174 dev_warn(hdev->dev,
175 "Device disabled but received IRQ %d for EQ\n",
176 irq);
177 goto skip_irq;
178 }
179
180 handle_eqe_work = kmalloc(sizeof(*handle_eqe_work), GFP_ATOMIC);
181 if (handle_eqe_work) {
182 INIT_WORK(&handle_eqe_work->eq_work, irq_handle_eqe);
183 handle_eqe_work->hdev = hdev;
184
185 memcpy(&handle_eqe_work->eq_entry, eq_entry,
186 sizeof(*eq_entry));
187
188 queue_work(hdev->eq_wq, &handle_eqe_work->eq_work);
189 }
190 skip_irq:
191 /* Clear EQ entry ready bit */
192 eq_entry->hdr.ctl =
193 cpu_to_le32(le32_to_cpu(eq_entry->hdr.ctl) &
194 ~EQ_CTL_READY_MASK);
195
196 eq->ci = hl_eq_inc_ptr(eq->ci);
197
198 hdev->asic_funcs->update_eq_ci(hdev, eq->ci);
199 }
200
201 return IRQ_HANDLED;
202 }
203
204 /**
205 * hl_cq_init - main initialization function for an cq object
206 *
207 * @hdev: pointer to device structure
208 * @q: pointer to cq structure
209 * @hw_queue_id: The H/W queue ID this completion queue belongs to
210 *
211 * Allocate dma-able memory for the completion queue and initialize fields
212 * Returns 0 on success
213 */
hl_cq_init(struct hl_device * hdev,struct hl_cq * q,u32 hw_queue_id)214 int hl_cq_init(struct hl_device *hdev, struct hl_cq *q, u32 hw_queue_id)
215 {
216 void *p;
217
218 p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev, HL_CQ_SIZE_IN_BYTES,
219 &q->bus_address, GFP_KERNEL | __GFP_ZERO);
220 if (!p)
221 return -ENOMEM;
222
223 q->hdev = hdev;
224 q->kernel_address = p;
225 q->hw_queue_id = hw_queue_id;
226 q->ci = 0;
227 q->pi = 0;
228
229 atomic_set(&q->free_slots_cnt, HL_CQ_LENGTH);
230
231 return 0;
232 }
233
234 /**
235 * hl_cq_fini - destroy completion queue
236 *
237 * @hdev: pointer to device structure
238 * @q: pointer to cq structure
239 *
240 * Free the completion queue memory
241 */
hl_cq_fini(struct hl_device * hdev,struct hl_cq * q)242 void hl_cq_fini(struct hl_device *hdev, struct hl_cq *q)
243 {
244 hdev->asic_funcs->asic_dma_free_coherent(hdev, HL_CQ_SIZE_IN_BYTES,
245 q->kernel_address,
246 q->bus_address);
247 }
248
hl_cq_reset(struct hl_device * hdev,struct hl_cq * q)249 void hl_cq_reset(struct hl_device *hdev, struct hl_cq *q)
250 {
251 q->ci = 0;
252 q->pi = 0;
253
254 atomic_set(&q->free_slots_cnt, HL_CQ_LENGTH);
255
256 /*
257 * It's not enough to just reset the PI/CI because the H/W may have
258 * written valid completion entries before it was halted and therefore
259 * we need to clean the actual queues so we won't process old entries
260 * when the device is operational again
261 */
262
263 memset(q->kernel_address, 0, HL_CQ_SIZE_IN_BYTES);
264 }
265
266 /**
267 * hl_eq_init - main initialization function for an event queue object
268 *
269 * @hdev: pointer to device structure
270 * @q: pointer to eq structure
271 *
272 * Allocate dma-able memory for the event queue and initialize fields
273 * Returns 0 on success
274 */
hl_eq_init(struct hl_device * hdev,struct hl_eq * q)275 int hl_eq_init(struct hl_device *hdev, struct hl_eq *q)
276 {
277 void *p;
278
279 p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
280 HL_EQ_SIZE_IN_BYTES,
281 &q->bus_address);
282 if (!p)
283 return -ENOMEM;
284
285 q->hdev = hdev;
286 q->kernel_address = p;
287 q->ci = 0;
288
289 return 0;
290 }
291
292 /**
293 * hl_eq_fini - destroy event queue
294 *
295 * @hdev: pointer to device structure
296 * @q: pointer to eq structure
297 *
298 * Free the event queue memory
299 */
hl_eq_fini(struct hl_device * hdev,struct hl_eq * q)300 void hl_eq_fini(struct hl_device *hdev, struct hl_eq *q)
301 {
302 flush_workqueue(hdev->eq_wq);
303
304 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
305 HL_EQ_SIZE_IN_BYTES,
306 q->kernel_address);
307 }
308
hl_eq_reset(struct hl_device * hdev,struct hl_eq * q)309 void hl_eq_reset(struct hl_device *hdev, struct hl_eq *q)
310 {
311 q->ci = 0;
312
313 /*
314 * It's not enough to just reset the PI/CI because the H/W may have
315 * written valid completion entries before it was halted and therefore
316 * we need to clean the actual queues so we won't process old entries
317 * when the device is operational again
318 */
319
320 memset(q->kernel_address, 0, HL_EQ_SIZE_IN_BYTES);
321 }
322