Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (c) 2015 HGST, a Western Digital Company.
4 */
5#include <linux/module.h>
6#include <linux/err.h>
7#include <linux/slab.h>
8#include <rdma/ib_verbs.h>
9
10#include "core_priv.h"
11
12#include <trace/events/rdma_core.h>
13/* Max size for shared CQ, may require tuning */
14#define IB_MAX_SHARED_CQ_SZ 4096U
15
16/* # of WCs to poll for with a single call to ib_poll_cq */
17#define IB_POLL_BATCH 16
18#define IB_POLL_BATCH_DIRECT 8
19
20/* # of WCs to iterate over before yielding */
21#define IB_POLL_BUDGET_IRQ 256
22#define IB_POLL_BUDGET_WORKQUEUE 65536
23
24#define IB_POLL_FLAGS \
25 (IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)
26
27static const struct dim_cq_moder
28rdma_dim_prof[RDMA_DIM_PARAMS_NUM_PROFILES] = {
29 {1, 0, 1, 0},
30 {1, 0, 4, 0},
31 {2, 0, 4, 0},
32 {2, 0, 8, 0},
33 {4, 0, 8, 0},
34 {16, 0, 8, 0},
35 {16, 0, 16, 0},
36 {32, 0, 16, 0},
37 {32, 0, 32, 0},
38};
39
40static void ib_cq_rdma_dim_work(struct work_struct *w)
41{
42 struct dim *dim = container_of(w, struct dim, work);
43 struct ib_cq *cq = dim->priv;
44
45 u16 usec = rdma_dim_prof[dim->profile_ix].usec;
46 u16 comps = rdma_dim_prof[dim->profile_ix].comps;
47
48 dim->state = DIM_START_MEASURE;
49
50 trace_cq_modify(cq, comps, usec);
51 cq->device->ops.modify_cq(cq, comps, usec);
52}
53
54static void rdma_dim_init(struct ib_cq *cq)
55{
56 struct dim *dim;
57
58 if (!cq->device->ops.modify_cq || !cq->device->use_cq_dim ||
59 cq->poll_ctx == IB_POLL_DIRECT)
60 return;
61
62 dim = kzalloc(sizeof(struct dim), GFP_KERNEL);
63 if (!dim)
64 return;
65
66 dim->state = DIM_START_MEASURE;
67 dim->tune_state = DIM_GOING_RIGHT;
68 dim->profile_ix = RDMA_DIM_START_PROFILE;
69 dim->priv = cq;
70 cq->dim = dim;
71
72 INIT_WORK(&dim->work, ib_cq_rdma_dim_work);
73}
74
75static void rdma_dim_destroy(struct ib_cq *cq)
76{
77 if (!cq->dim)
78 return;
79
80 cancel_work_sync(&cq->dim->work);
81 kfree(cq->dim);
82}
83
84static int __poll_cq(struct ib_cq *cq, int num_entries, struct ib_wc *wc)
85{
86 int rc;
87
88 rc = ib_poll_cq(cq, num_entries, wc);
89 trace_cq_poll(cq, num_entries, rc);
90 return rc;
91}
92
93static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *wcs,
94 int batch)
95{
96 int i, n, completed = 0;
97
98 trace_cq_process(cq);
99
100 /*
101 * budget might be (-1) if the caller does not
102 * want to bound this call, thus we need unsigned
103 * minimum here.
104 */
105 while ((n = __poll_cq(cq, min_t(u32, batch,
106 budget - completed), wcs)) > 0) {
107 for (i = 0; i < n; i++) {
108 struct ib_wc *wc = &wcs[i];
109
110 if (wc->wr_cqe)
111 wc->wr_cqe->done(cq, wc);
112 else
113 WARN_ON_ONCE(wc->status == IB_WC_SUCCESS);
114 }
115
116 completed += n;
117
118 if (n != batch || (budget != -1 && completed >= budget))
119 break;
120 }
121
122 return completed;
123}
124
125/**
126 * ib_process_direct_cq - process a CQ in caller context
127 * @cq: CQ to process
128 * @budget: number of CQEs to poll for
129 *
130 * This function is used to process all outstanding CQ entries.
131 * It does not offload CQ processing to a different context and does
132 * not ask for completion interrupts from the HCA.
133 * Using direct processing on CQ with non IB_POLL_DIRECT type may trigger
134 * concurrent processing.
135 *
136 * Note: do not pass -1 as %budget unless it is guaranteed that the number
137 * of completions that will be processed is small.
138 */
139int ib_process_cq_direct(struct ib_cq *cq, int budget)
140{
141 struct ib_wc wcs[IB_POLL_BATCH_DIRECT];
142
143 return __ib_process_cq(cq, budget, wcs, IB_POLL_BATCH_DIRECT);
144}
145EXPORT_SYMBOL(ib_process_cq_direct);
146
147static void ib_cq_completion_direct(struct ib_cq *cq, void *private)
148{
149 WARN_ONCE(1, "got unsolicited completion for CQ 0x%p\n", cq);
150}
151
152static int ib_poll_handler(struct irq_poll *iop, int budget)
153{
154 struct ib_cq *cq = container_of(iop, struct ib_cq, iop);
155 struct dim *dim = cq->dim;
156 int completed;
157
158 completed = __ib_process_cq(cq, budget, cq->wc, IB_POLL_BATCH);
159 if (completed < budget) {
160 irq_poll_complete(&cq->iop);
161 if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) {
162 trace_cq_reschedule(cq);
163 irq_poll_sched(&cq->iop);
164 }
165 }
166
167 if (dim)
168 rdma_dim(dim, completed);
169
170 return completed;
171}
172
173static void ib_cq_completion_softirq(struct ib_cq *cq, void *private)
174{
175 trace_cq_schedule(cq);
176 irq_poll_sched(&cq->iop);
177}
178
179static void ib_cq_poll_work(struct work_struct *work)
180{
181 struct ib_cq *cq = container_of(work, struct ib_cq, work);
182 int completed;
183
184 completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE, cq->wc,
185 IB_POLL_BATCH);
186 if (completed >= IB_POLL_BUDGET_WORKQUEUE ||
187 ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
188 queue_work(cq->comp_wq, &cq->work);
189 else if (cq->dim)
190 rdma_dim(cq->dim, completed);
191}
192
193static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private)
194{
195 trace_cq_schedule(cq);
196 queue_work(cq->comp_wq, &cq->work);
197}
198
199/**
200 * __ib_alloc_cq_user - allocate a completion queue
201 * @dev: device to allocate the CQ for
202 * @private: driver private data, accessible from cq->cq_context
203 * @nr_cqe: number of CQEs to allocate
204 * @comp_vector: HCA completion vectors for this CQ
205 * @poll_ctx: context to poll the CQ from.
206 * @caller: module owner name.
207 * @udata: Valid user data or NULL for kernel object
208 *
209 * This is the proper interface to allocate a CQ for in-kernel users. A
210 * CQ allocated with this interface will automatically be polled from the
211 * specified context. The ULP must use wr->wr_cqe instead of wr->wr_id
212 * to use this CQ abstraction.
213 */
214struct ib_cq *__ib_alloc_cq_user(struct ib_device *dev, void *private,
215 int nr_cqe, int comp_vector,
216 enum ib_poll_context poll_ctx,
217 const char *caller, struct ib_udata *udata)
218{
219 struct ib_cq_init_attr cq_attr = {
220 .cqe = nr_cqe,
221 .comp_vector = comp_vector,
222 };
223 struct ib_cq *cq;
224 int ret = -ENOMEM;
225
226 cq = rdma_zalloc_drv_obj(dev, ib_cq);
227 if (!cq)
228 return ERR_PTR(ret);
229
230 cq->device = dev;
231 cq->cq_context = private;
232 cq->poll_ctx = poll_ctx;
233 atomic_set(&cq->usecnt, 0);
234 cq->comp_vector = comp_vector;
235
236 cq->wc = kmalloc_array(IB_POLL_BATCH, sizeof(*cq->wc), GFP_KERNEL);
237 if (!cq->wc)
238 goto out_free_cq;
239
240 cq->res.type = RDMA_RESTRACK_CQ;
241 rdma_restrack_set_task(&cq->res, caller);
242
243 ret = dev->ops.create_cq(cq, &cq_attr, NULL);
244 if (ret)
245 goto out_free_wc;
246
247 rdma_restrack_kadd(&cq->res);
248
249 rdma_dim_init(cq);
250
251 switch (cq->poll_ctx) {
252 case IB_POLL_DIRECT:
253 cq->comp_handler = ib_cq_completion_direct;
254 break;
255 case IB_POLL_SOFTIRQ:
256 cq->comp_handler = ib_cq_completion_softirq;
257
258 irq_poll_init(&cq->iop, IB_POLL_BUDGET_IRQ, ib_poll_handler);
259 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
260 break;
261 case IB_POLL_WORKQUEUE:
262 case IB_POLL_UNBOUND_WORKQUEUE:
263 cq->comp_handler = ib_cq_completion_workqueue;
264 INIT_WORK(&cq->work, ib_cq_poll_work);
265 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
266 cq->comp_wq = (cq->poll_ctx == IB_POLL_WORKQUEUE) ?
267 ib_comp_wq : ib_comp_unbound_wq;
268 break;
269 default:
270 ret = -EINVAL;
271 goto out_destroy_cq;
272 }
273
274 trace_cq_alloc(cq, nr_cqe, comp_vector, poll_ctx);
275 return cq;
276
277out_destroy_cq:
278 rdma_dim_destroy(cq);
279 rdma_restrack_del(&cq->res);
280 cq->device->ops.destroy_cq(cq, udata);
281out_free_wc:
282 kfree(cq->wc);
283out_free_cq:
284 kfree(cq);
285 trace_cq_alloc_error(nr_cqe, comp_vector, poll_ctx, ret);
286 return ERR_PTR(ret);
287}
288EXPORT_SYMBOL(__ib_alloc_cq_user);
289
290/**
291 * __ib_alloc_cq_any - allocate a completion queue
292 * @dev: device to allocate the CQ for
293 * @private: driver private data, accessible from cq->cq_context
294 * @nr_cqe: number of CQEs to allocate
295 * @poll_ctx: context to poll the CQ from
296 * @caller: module owner name
297 *
298 * Attempt to spread ULP Completion Queues over each device's interrupt
299 * vectors. A simple best-effort mechanism is used.
300 */
301struct ib_cq *__ib_alloc_cq_any(struct ib_device *dev, void *private,
302 int nr_cqe, enum ib_poll_context poll_ctx,
303 const char *caller)
304{
305 static atomic_t counter;
306 int comp_vector = 0;
307
308 if (dev->num_comp_vectors > 1)
309 comp_vector =
310 atomic_inc_return(&counter) %
311 min_t(int, dev->num_comp_vectors, num_online_cpus());
312
313 return __ib_alloc_cq_user(dev, private, nr_cqe, comp_vector, poll_ctx,
314 caller, NULL);
315}
316EXPORT_SYMBOL(__ib_alloc_cq_any);
317
318/**
319 * ib_free_cq_user - free a completion queue
320 * @cq: completion queue to free.
321 * @udata: User data or NULL for kernel object
322 */
323void ib_free_cq_user(struct ib_cq *cq, struct ib_udata *udata)
324{
325 if (WARN_ON_ONCE(atomic_read(&cq->usecnt)))
326 return;
327 if (WARN_ON_ONCE(cq->cqe_used))
328 return;
329
330 switch (cq->poll_ctx) {
331 case IB_POLL_DIRECT:
332 break;
333 case IB_POLL_SOFTIRQ:
334 irq_poll_disable(&cq->iop);
335 break;
336 case IB_POLL_WORKQUEUE:
337 case IB_POLL_UNBOUND_WORKQUEUE:
338 cancel_work_sync(&cq->work);
339 break;
340 default:
341 WARN_ON_ONCE(1);
342 }
343
344 rdma_dim_destroy(cq);
345 trace_cq_free(cq);
346 rdma_restrack_del(&cq->res);
347 cq->device->ops.destroy_cq(cq, udata);
348 kfree(cq->wc);
349 kfree(cq);
350}
351EXPORT_SYMBOL(ib_free_cq_user);
352
353void ib_cq_pool_init(struct ib_device *dev)
354{
355 unsigned int i;
356
357 spin_lock_init(&dev->cq_pools_lock);
358 for (i = 0; i < ARRAY_SIZE(dev->cq_pools); i++)
359 INIT_LIST_HEAD(&dev->cq_pools[i]);
360}
361
362void ib_cq_pool_destroy(struct ib_device *dev)
363{
364 struct ib_cq *cq, *n;
365 unsigned int i;
366
367 for (i = 0; i < ARRAY_SIZE(dev->cq_pools); i++) {
368 list_for_each_entry_safe(cq, n, &dev->cq_pools[i],
369 pool_entry) {
370 WARN_ON(cq->cqe_used);
371 cq->shared = false;
372 ib_free_cq(cq);
373 }
374 }
375}
376
377static int ib_alloc_cqs(struct ib_device *dev, unsigned int nr_cqes,
378 enum ib_poll_context poll_ctx)
379{
380 LIST_HEAD(tmp_list);
381 unsigned int nr_cqs, i;
382 struct ib_cq *cq, *n;
383 int ret;
384
385 if (poll_ctx > IB_POLL_LAST_POOL_TYPE) {
386 WARN_ON_ONCE(poll_ctx > IB_POLL_LAST_POOL_TYPE);
387 return -EINVAL;
388 }
389
390 /*
391 * Allocate at least as many CQEs as requested, and otherwise
392 * a reasonable batch size so that we can share CQs between
393 * multiple users instead of allocating a larger number of CQs.
394 */
395 nr_cqes = min_t(unsigned int, dev->attrs.max_cqe,
396 max(nr_cqes, IB_MAX_SHARED_CQ_SZ));
397 nr_cqs = min_t(unsigned int, dev->num_comp_vectors, num_online_cpus());
398 for (i = 0; i < nr_cqs; i++) {
399 cq = ib_alloc_cq(dev, NULL, nr_cqes, i, poll_ctx);
400 if (IS_ERR(cq)) {
401 ret = PTR_ERR(cq);
402 goto out_free_cqs;
403 }
404 cq->shared = true;
405 list_add_tail(&cq->pool_entry, &tmp_list);
406 }
407
408 spin_lock_irq(&dev->cq_pools_lock);
409 list_splice(&tmp_list, &dev->cq_pools[poll_ctx]);
410 spin_unlock_irq(&dev->cq_pools_lock);
411
412 return 0;
413
414out_free_cqs:
415 list_for_each_entry_safe(cq, n, &tmp_list, pool_entry) {
416 cq->shared = false;
417 ib_free_cq(cq);
418 }
419 return ret;
420}
421
422/**
423 * ib_cq_pool_get() - Find the least used completion queue that matches
424 * a given cpu hint (or least used for wild card affinity) and fits
425 * nr_cqe.
426 * @dev: rdma device
427 * @nr_cqe: number of needed cqe entries
428 * @comp_vector_hint: completion vector hint (-1) for the driver to assign
429 * a comp vector based on internal counter
430 * @poll_ctx: cq polling context
431 *
432 * Finds a cq that satisfies @comp_vector_hint and @nr_cqe requirements and
433 * claim entries in it for us. In case there is no available cq, allocate
434 * a new cq with the requirements and add it to the device pool.
435 * IB_POLL_DIRECT cannot be used for shared cqs so it is not a valid value
436 * for @poll_ctx.
437 */
438struct ib_cq *ib_cq_pool_get(struct ib_device *dev, unsigned int nr_cqe,
439 int comp_vector_hint,
440 enum ib_poll_context poll_ctx)
441{
442 static unsigned int default_comp_vector;
443 unsigned int vector, num_comp_vectors;
444 struct ib_cq *cq, *found = NULL;
445 int ret;
446
447 if (poll_ctx > IB_POLL_LAST_POOL_TYPE) {
448 WARN_ON_ONCE(poll_ctx > IB_POLL_LAST_POOL_TYPE);
449 return ERR_PTR(-EINVAL);
450 }
451
452 num_comp_vectors =
453 min_t(unsigned int, dev->num_comp_vectors, num_online_cpus());
454 /* Project the affinty to the device completion vector range */
455 if (comp_vector_hint < 0) {
456 comp_vector_hint =
457 (READ_ONCE(default_comp_vector) + 1) % num_comp_vectors;
458 WRITE_ONCE(default_comp_vector, comp_vector_hint);
459 }
460 vector = comp_vector_hint % num_comp_vectors;
461
462 /*
463 * Find the least used CQ with correct affinity and
464 * enough free CQ entries
465 */
466 while (!found) {
467 spin_lock_irq(&dev->cq_pools_lock);
468 list_for_each_entry(cq, &dev->cq_pools[poll_ctx],
469 pool_entry) {
470 /*
471 * Check to see if we have found a CQ with the
472 * correct completion vector
473 */
474 if (vector != cq->comp_vector)
475 continue;
476 if (cq->cqe_used + nr_cqe > cq->cqe)
477 continue;
478 found = cq;
479 break;
480 }
481
482 if (found) {
483 found->cqe_used += nr_cqe;
484 spin_unlock_irq(&dev->cq_pools_lock);
485
486 return found;
487 }
488 spin_unlock_irq(&dev->cq_pools_lock);
489
490 /*
491 * Didn't find a match or ran out of CQs in the device
492 * pool, allocate a new array of CQs.
493 */
494 ret = ib_alloc_cqs(dev, nr_cqe, poll_ctx);
495 if (ret)
496 return ERR_PTR(ret);
497 }
498
499 return found;
500}
501EXPORT_SYMBOL(ib_cq_pool_get);
502
503/**
504 * ib_cq_pool_put - Return a CQ taken from a shared pool.
505 * @cq: The CQ to return.
506 * @nr_cqe: The max number of cqes that the user had requested.
507 */
508void ib_cq_pool_put(struct ib_cq *cq, unsigned int nr_cqe)
509{
510 if (WARN_ON_ONCE(nr_cqe > cq->cqe_used))
511 return;
512
513 spin_lock_irq(&cq->device->cq_pools_lock);
514 cq->cqe_used -= nr_cqe;
515 spin_unlock_irq(&cq->device->cq_pools_lock);
516}
517EXPORT_SYMBOL(ib_cq_pool_put);