1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37#include <linux/hardirq.h>
38#include <linux/export.h>
39
40#include <linux/mlx4/cmd.h>
41#include <linux/mlx4/cq.h>
42
43#include "mlx4.h"
44#include "icm.h"
45
46#define MLX4_CQ_STATUS_OK ( 0 << 28)
47#define MLX4_CQ_STATUS_OVERFLOW ( 9 << 28)
48#define MLX4_CQ_STATUS_WRITE_FAIL (10 << 28)
49#define MLX4_CQ_FLAG_CC ( 1 << 18)
50#define MLX4_CQ_FLAG_OI ( 1 << 17)
51#define MLX4_CQ_STATE_ARMED ( 9 << 8)
52#define MLX4_CQ_STATE_ARMED_SOL ( 6 << 8)
53#define MLX4_EQ_STATE_FIRED (10 << 8)
54
55#define TASKLET_MAX_TIME 2
56#define TASKLET_MAX_TIME_JIFFIES msecs_to_jiffies(TASKLET_MAX_TIME)
57
58void mlx4_cq_tasklet_cb(unsigned long data)
59{
60 unsigned long flags;
61 unsigned long end = jiffies + TASKLET_MAX_TIME_JIFFIES;
62 struct mlx4_eq_tasklet *ctx = (struct mlx4_eq_tasklet *)data;
63 struct mlx4_cq *mcq, *temp;
64
65 spin_lock_irqsave(&ctx->lock, flags);
66 list_splice_tail_init(&ctx->list, &ctx->process_list);
67 spin_unlock_irqrestore(&ctx->lock, flags);
68
69 list_for_each_entry_safe(mcq, temp, &ctx->process_list, tasklet_ctx.list) {
70 list_del_init(&mcq->tasklet_ctx.list);
71 mcq->tasklet_ctx.comp(mcq);
72 if (refcount_dec_and_test(&mcq->refcount))
73 complete(&mcq->free);
74 if (time_after(jiffies, end))
75 break;
76 }
77
78 if (!list_empty(&ctx->process_list))
79 tasklet_schedule(&ctx->task);
80}
81
82static void mlx4_add_cq_to_tasklet(struct mlx4_cq *cq)
83{
84 struct mlx4_eq_tasklet *tasklet_ctx = cq->tasklet_ctx.priv;
85 unsigned long flags;
86 bool kick;
87
88 spin_lock_irqsave(&tasklet_ctx->lock, flags);
89
90
91
92
93
94 if (list_empty_careful(&cq->tasklet_ctx.list)) {
95 refcount_inc(&cq->refcount);
96 kick = list_empty(&tasklet_ctx->list);
97 list_add_tail(&cq->tasklet_ctx.list, &tasklet_ctx->list);
98 if (kick)
99 tasklet_schedule(&tasklet_ctx->task);
100 }
101 spin_unlock_irqrestore(&tasklet_ctx->lock, flags);
102}
103
104void mlx4_cq_completion(struct mlx4_dev *dev, u32 cqn)
105{
106 struct mlx4_cq *cq;
107
108 rcu_read_lock();
109 cq = radix_tree_lookup(&mlx4_priv(dev)->cq_table.tree,
110 cqn & (dev->caps.num_cqs - 1));
111 rcu_read_unlock();
112
113 if (!cq) {
114 mlx4_dbg(dev, "Completion event for bogus CQ %08x\n", cqn);
115 return;
116 }
117
118
119
120
121 ++cq->arm_sn;
122
123 cq->comp(cq);
124}
125
126void mlx4_cq_event(struct mlx4_dev *dev, u32 cqn, int event_type)
127{
128 struct mlx4_cq_table *cq_table = &mlx4_priv(dev)->cq_table;
129 struct mlx4_cq *cq;
130
131 rcu_read_lock();
132 cq = radix_tree_lookup(&cq_table->tree, cqn & (dev->caps.num_cqs - 1));
133 rcu_read_unlock();
134
135 if (!cq) {
136 mlx4_dbg(dev, "Async event for bogus CQ %08x\n", cqn);
137 return;
138 }
139
140
141
142
143 cq->event(cq, event_type);
144}
145
146static int mlx4_SW2HW_CQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
147 int cq_num, u8 opmod)
148{
149 return mlx4_cmd(dev, mailbox->dma, cq_num, opmod,
150 MLX4_CMD_SW2HW_CQ, MLX4_CMD_TIME_CLASS_A,
151 MLX4_CMD_WRAPPED);
152}
153
154static int mlx4_MODIFY_CQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
155 int cq_num, u32 opmod)
156{
157 return mlx4_cmd(dev, mailbox->dma, cq_num, opmod, MLX4_CMD_MODIFY_CQ,
158 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
159}
160
161static int mlx4_HW2SW_CQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
162 int cq_num)
163{
164 return mlx4_cmd_box(dev, 0, mailbox ? mailbox->dma : 0,
165 cq_num, mailbox ? 0 : 1, MLX4_CMD_HW2SW_CQ,
166 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
167}
168
169int mlx4_cq_modify(struct mlx4_dev *dev, struct mlx4_cq *cq,
170 u16 count, u16 period)
171{
172 struct mlx4_cmd_mailbox *mailbox;
173 struct mlx4_cq_context *cq_context;
174 int err;
175
176 mailbox = mlx4_alloc_cmd_mailbox(dev);
177 if (IS_ERR(mailbox))
178 return PTR_ERR(mailbox);
179
180 cq_context = mailbox->buf;
181 cq_context->cq_max_count = cpu_to_be16(count);
182 cq_context->cq_period = cpu_to_be16(period);
183
184 err = mlx4_MODIFY_CQ(dev, mailbox, cq->cqn, 1);
185
186 mlx4_free_cmd_mailbox(dev, mailbox);
187 return err;
188}
189EXPORT_SYMBOL_GPL(mlx4_cq_modify);
190
191int mlx4_cq_resize(struct mlx4_dev *dev, struct mlx4_cq *cq,
192 int entries, struct mlx4_mtt *mtt)
193{
194 struct mlx4_cmd_mailbox *mailbox;
195 struct mlx4_cq_context *cq_context;
196 u64 mtt_addr;
197 int err;
198
199 mailbox = mlx4_alloc_cmd_mailbox(dev);
200 if (IS_ERR(mailbox))
201 return PTR_ERR(mailbox);
202
203 cq_context = mailbox->buf;
204 cq_context->logsize_usrpage = cpu_to_be32(ilog2(entries) << 24);
205 cq_context->log_page_size = mtt->page_shift - 12;
206 mtt_addr = mlx4_mtt_addr(dev, mtt);
207 cq_context->mtt_base_addr_h = mtt_addr >> 32;
208 cq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
209
210 err = mlx4_MODIFY_CQ(dev, mailbox, cq->cqn, 0);
211
212 mlx4_free_cmd_mailbox(dev, mailbox);
213 return err;
214}
215EXPORT_SYMBOL_GPL(mlx4_cq_resize);
216
217int __mlx4_cq_alloc_icm(struct mlx4_dev *dev, int *cqn)
218{
219 struct mlx4_priv *priv = mlx4_priv(dev);
220 struct mlx4_cq_table *cq_table = &priv->cq_table;
221 int err;
222
223 *cqn = mlx4_bitmap_alloc(&cq_table->bitmap);
224 if (*cqn == -1)
225 return -ENOMEM;
226
227 err = mlx4_table_get(dev, &cq_table->table, *cqn);
228 if (err)
229 goto err_out;
230
231 err = mlx4_table_get(dev, &cq_table->cmpt_table, *cqn);
232 if (err)
233 goto err_put;
234 return 0;
235
236err_put:
237 mlx4_table_put(dev, &cq_table->table, *cqn);
238
239err_out:
240 mlx4_bitmap_free(&cq_table->bitmap, *cqn, MLX4_NO_RR);
241 return err;
242}
243
244static int mlx4_cq_alloc_icm(struct mlx4_dev *dev, int *cqn, u8 usage)
245{
246 u32 in_modifier = RES_CQ | (((u32)usage & 3) << 30);
247 u64 out_param;
248 int err;
249
250 if (mlx4_is_mfunc(dev)) {
251 err = mlx4_cmd_imm(dev, 0, &out_param, in_modifier,
252 RES_OP_RESERVE_AND_MAP, MLX4_CMD_ALLOC_RES,
253 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
254 if (err)
255 return err;
256 else {
257 *cqn = get_param_l(&out_param);
258 return 0;
259 }
260 }
261 return __mlx4_cq_alloc_icm(dev, cqn);
262}
263
264void __mlx4_cq_free_icm(struct mlx4_dev *dev, int cqn)
265{
266 struct mlx4_priv *priv = mlx4_priv(dev);
267 struct mlx4_cq_table *cq_table = &priv->cq_table;
268
269 mlx4_table_put(dev, &cq_table->cmpt_table, cqn);
270 mlx4_table_put(dev, &cq_table->table, cqn);
271 mlx4_bitmap_free(&cq_table->bitmap, cqn, MLX4_NO_RR);
272}
273
274static void mlx4_cq_free_icm(struct mlx4_dev *dev, int cqn)
275{
276 u64 in_param = 0;
277 int err;
278
279 if (mlx4_is_mfunc(dev)) {
280 set_param_l(&in_param, cqn);
281 err = mlx4_cmd(dev, in_param, RES_CQ, RES_OP_RESERVE_AND_MAP,
282 MLX4_CMD_FREE_RES,
283 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
284 if (err)
285 mlx4_warn(dev, "Failed freeing cq:%d\n", cqn);
286 } else
287 __mlx4_cq_free_icm(dev, cqn);
288}
289
290static int mlx4_init_user_cqes(void *buf, int entries, int cqe_size)
291{
292 int entries_per_copy = PAGE_SIZE / cqe_size;
293 void *init_ents;
294 int err = 0;
295 int i;
296
297 init_ents = kmalloc(PAGE_SIZE, GFP_KERNEL);
298 if (!init_ents)
299 return -ENOMEM;
300
301
302
303
304
305 memset(init_ents, 0xcc, PAGE_SIZE);
306
307 if (entries_per_copy < entries) {
308 for (i = 0; i < entries / entries_per_copy; i++) {
309 err = copy_to_user((void __user *)buf, init_ents, PAGE_SIZE) ?
310 -EFAULT : 0;
311 if (err)
312 goto out;
313
314 buf += PAGE_SIZE;
315 }
316 } else {
317 err = copy_to_user((void __user *)buf, init_ents, entries * cqe_size) ?
318 -EFAULT : 0;
319 }
320
321out:
322 kfree(init_ents);
323
324 return err;
325}
326
327static void mlx4_init_kernel_cqes(struct mlx4_buf *buf,
328 int entries,
329 int cqe_size)
330{
331 int i;
332
333 if (buf->nbufs == 1)
334 memset(buf->direct.buf, 0xcc, entries * cqe_size);
335 else
336 for (i = 0; i < buf->npages; i++)
337 memset(buf->page_list[i].buf, 0xcc,
338 1UL << buf->page_shift);
339}
340
341int mlx4_cq_alloc(struct mlx4_dev *dev, int nent,
342 struct mlx4_mtt *mtt, struct mlx4_uar *uar, u64 db_rec,
343 struct mlx4_cq *cq, unsigned vector, int collapsed,
344 int timestamp_en, void *buf_addr, bool user_cq)
345{
346 bool sw_cq_init = dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SW_CQ_INIT;
347 struct mlx4_priv *priv = mlx4_priv(dev);
348 struct mlx4_cq_table *cq_table = &priv->cq_table;
349 struct mlx4_cmd_mailbox *mailbox;
350 struct mlx4_cq_context *cq_context;
351 u64 mtt_addr;
352 int err;
353
354 if (vector >= dev->caps.num_comp_vectors)
355 return -EINVAL;
356
357 cq->vector = vector;
358
359 err = mlx4_cq_alloc_icm(dev, &cq->cqn, cq->usage);
360 if (err)
361 return err;
362
363 spin_lock(&cq_table->lock);
364 err = radix_tree_insert(&cq_table->tree, cq->cqn, cq);
365 spin_unlock(&cq_table->lock);
366 if (err)
367 goto err_icm;
368
369 mailbox = mlx4_alloc_cmd_mailbox(dev);
370 if (IS_ERR(mailbox)) {
371 err = PTR_ERR(mailbox);
372 goto err_radix;
373 }
374
375 cq_context = mailbox->buf;
376 cq_context->flags = cpu_to_be32(!!collapsed << 18);
377 if (timestamp_en)
378 cq_context->flags |= cpu_to_be32(1 << 19);
379
380 cq_context->logsize_usrpage =
381 cpu_to_be32((ilog2(nent) << 24) |
382 mlx4_to_hw_uar_index(dev, uar->index));
383 cq_context->comp_eqn = priv->eq_table.eq[MLX4_CQ_TO_EQ_VECTOR(vector)].eqn;
384 cq_context->log_page_size = mtt->page_shift - MLX4_ICM_PAGE_SHIFT;
385
386 mtt_addr = mlx4_mtt_addr(dev, mtt);
387 cq_context->mtt_base_addr_h = mtt_addr >> 32;
388 cq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
389 cq_context->db_rec_addr = cpu_to_be64(db_rec);
390
391 if (sw_cq_init) {
392 if (user_cq) {
393 err = mlx4_init_user_cqes(buf_addr, nent,
394 dev->caps.cqe_size);
395 if (err)
396 sw_cq_init = false;
397 } else {
398 mlx4_init_kernel_cqes(buf_addr, nent,
399 dev->caps.cqe_size);
400 }
401 }
402
403 err = mlx4_SW2HW_CQ(dev, mailbox, cq->cqn, sw_cq_init);
404
405 mlx4_free_cmd_mailbox(dev, mailbox);
406 if (err)
407 goto err_radix;
408
409 cq->cons_index = 0;
410 cq->arm_sn = 1;
411 cq->uar = uar;
412 refcount_set(&cq->refcount, 1);
413 init_completion(&cq->free);
414 cq->comp = mlx4_add_cq_to_tasklet;
415 cq->tasklet_ctx.priv =
416 &priv->eq_table.eq[MLX4_CQ_TO_EQ_VECTOR(vector)].tasklet_ctx;
417 INIT_LIST_HEAD(&cq->tasklet_ctx.list);
418
419
420 cq->irq = priv->eq_table.eq[MLX4_CQ_TO_EQ_VECTOR(vector)].irq;
421 return 0;
422
423err_radix:
424 spin_lock(&cq_table->lock);
425 radix_tree_delete(&cq_table->tree, cq->cqn);
426 spin_unlock(&cq_table->lock);
427
428err_icm:
429 mlx4_cq_free_icm(dev, cq->cqn);
430
431 return err;
432}
433EXPORT_SYMBOL_GPL(mlx4_cq_alloc);
434
435void mlx4_cq_free(struct mlx4_dev *dev, struct mlx4_cq *cq)
436{
437 struct mlx4_priv *priv = mlx4_priv(dev);
438 struct mlx4_cq_table *cq_table = &priv->cq_table;
439 int err;
440
441 err = mlx4_HW2SW_CQ(dev, NULL, cq->cqn);
442 if (err)
443 mlx4_warn(dev, "HW2SW_CQ failed (%d) for CQN %06x\n", err, cq->cqn);
444
445 spin_lock(&cq_table->lock);
446 radix_tree_delete(&cq_table->tree, cq->cqn);
447 spin_unlock(&cq_table->lock);
448
449 synchronize_irq(priv->eq_table.eq[MLX4_CQ_TO_EQ_VECTOR(cq->vector)].irq);
450 if (priv->eq_table.eq[MLX4_CQ_TO_EQ_VECTOR(cq->vector)].irq !=
451 priv->eq_table.eq[MLX4_EQ_ASYNC].irq)
452 synchronize_irq(priv->eq_table.eq[MLX4_EQ_ASYNC].irq);
453
454 if (refcount_dec_and_test(&cq->refcount))
455 complete(&cq->free);
456 wait_for_completion(&cq->free);
457
458 mlx4_cq_free_icm(dev, cq->cqn);
459}
460EXPORT_SYMBOL_GPL(mlx4_cq_free);
461
462int mlx4_init_cq_table(struct mlx4_dev *dev)
463{
464 struct mlx4_cq_table *cq_table = &mlx4_priv(dev)->cq_table;
465 int err;
466
467 spin_lock_init(&cq_table->lock);
468 INIT_RADIX_TREE(&cq_table->tree, GFP_ATOMIC);
469 if (mlx4_is_slave(dev))
470 return 0;
471
472 err = mlx4_bitmap_init(&cq_table->bitmap, dev->caps.num_cqs,
473 dev->caps.num_cqs - 1, dev->caps.reserved_cqs, 0);
474 if (err)
475 return err;
476
477 return 0;
478}
479
480void mlx4_cleanup_cq_table(struct mlx4_dev *dev)
481{
482 if (mlx4_is_slave(dev))
483 return;
484
485 mlx4_bitmap_cleanup(&mlx4_priv(dev)->cq_table.bitmap);
486}
487
