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_cq_direct - 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 - 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 *
208 * This is the proper interface to allocate a CQ for in-kernel users. A
209 * CQ allocated with this interface will automatically be polled from the
210 * specified context. The ULP must use wr->wr_cqe instead of wr->wr_id
211 * to use this CQ abstraction.
212 */
213struct ib_cq *__ib_alloc_cq(struct ib_device *dev, void *private, int nr_cqe,
214 int comp_vector, enum ib_poll_context poll_ctx,
215 const char *caller)
216{
217 struct ib_cq_init_attr cq_attr = {
218 .cqe = nr_cqe,
219 .comp_vector = comp_vector,
220 };
221 struct ib_cq *cq;
222 int ret = -ENOMEM;
223
224 cq = rdma_zalloc_drv_obj(dev, ib_cq);
225 if (!cq)
226 return ERR_PTR(ret);
227
228 cq->device = dev;
229 cq->cq_context = private;
230 cq->poll_ctx = poll_ctx;
231 atomic_set(&cq->usecnt, 0);
232 cq->comp_vector = comp_vector;
233
234 cq->wc = kmalloc_array(IB_POLL_BATCH, sizeof(*cq->wc), GFP_KERNEL);
235 if (!cq->wc)
236 goto out_free_cq;
237
238 rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
239 rdma_restrack_set_name(&cq->res, caller);
240
241 ret = dev->ops.create_cq(cq, &cq_attr, NULL);
242 if (ret)
243 goto out_free_wc;
244
245 rdma_dim_init(cq);
246
247 switch (cq->poll_ctx) {
248 case IB_POLL_DIRECT:
249 cq->comp_handler = ib_cq_completion_direct;
250 break;
251 case IB_POLL_SOFTIRQ:
252 cq->comp_handler = ib_cq_completion_softirq;
253
254 irq_poll_init(&cq->iop, IB_POLL_BUDGET_IRQ, ib_poll_handler);
255 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
256 break;
257 case IB_POLL_WORKQUEUE:
258 case IB_POLL_UNBOUND_WORKQUEUE:
259 cq->comp_handler = ib_cq_completion_workqueue;
260 INIT_WORK(&cq->work, ib_cq_poll_work);
261 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
262 cq->comp_wq = (cq->poll_ctx == IB_POLL_WORKQUEUE) ?
263 ib_comp_wq : ib_comp_unbound_wq;
264 break;
265 default:
266 ret = -EINVAL;
267 goto out_destroy_cq;
268 }
269
270 rdma_restrack_add(&cq->res);
271 trace_cq_alloc(cq, nr_cqe, comp_vector, poll_ctx);
272 return cq;
273
274out_destroy_cq:
275 rdma_dim_destroy(cq);
276 cq->device->ops.destroy_cq(cq, NULL);
277out_free_wc:
278 rdma_restrack_put(&cq->res);
279 kfree(cq->wc);
280out_free_cq:
281 kfree(cq);
282 trace_cq_alloc_error(nr_cqe, comp_vector, poll_ctx, ret);
283 return ERR_PTR(ret);
284}
285EXPORT_SYMBOL(__ib_alloc_cq);
286
287/**
288 * __ib_alloc_cq_any - allocate a completion queue
289 * @dev: device to allocate the CQ for
290 * @private: driver private data, accessible from cq->cq_context
291 * @nr_cqe: number of CQEs to allocate
292 * @poll_ctx: context to poll the CQ from
293 * @caller: module owner name
294 *
295 * Attempt to spread ULP Completion Queues over each device's interrupt
296 * vectors. A simple best-effort mechanism is used.
297 */
298struct ib_cq *__ib_alloc_cq_any(struct ib_device *dev, void *private,
299 int nr_cqe, enum ib_poll_context poll_ctx,
300 const char *caller)
301{
302 static atomic_t counter;
303 int comp_vector = 0;
304
305 if (dev->num_comp_vectors > 1)
306 comp_vector =
307 atomic_inc_return(&counter) %
308 min_t(int, dev->num_comp_vectors, num_online_cpus());
309
310 return __ib_alloc_cq(dev, private, nr_cqe, comp_vector, poll_ctx,
311 caller);
312}
313EXPORT_SYMBOL(__ib_alloc_cq_any);
314
315/**
316 * ib_free_cq - free a completion queue
317 * @cq: completion queue to free.
318 */
319void ib_free_cq(struct ib_cq *cq)
320{
321 int ret;
322
323 if (WARN_ON_ONCE(atomic_read(&cq->usecnt)))
324 return;
325 if (WARN_ON_ONCE(cq->cqe_used))
326 return;
327
328 switch (cq->poll_ctx) {
329 case IB_POLL_DIRECT:
330 break;
331 case IB_POLL_SOFTIRQ:
332 irq_poll_disable(&cq->iop);
333 break;
334 case IB_POLL_WORKQUEUE:
335 case IB_POLL_UNBOUND_WORKQUEUE:
336 cancel_work_sync(&cq->work);
337 break;
338 default:
339 WARN_ON_ONCE(1);
340 }
341
342 rdma_dim_destroy(cq);
343 trace_cq_free(cq);
344 ret = cq->device->ops.destroy_cq(cq, NULL);
345 WARN_ONCE(ret, "Destroy of kernel CQ shouldn't fail");
346 rdma_restrack_del(&cq->res);
347 kfree(cq->wc);
348 kfree(cq);
349}
350EXPORT_SYMBOL(ib_free_cq);
351
352void ib_cq_pool_cleanup(struct ib_device *dev)
353{
354 struct ib_cq *cq, *n;
355 unsigned int i;
356
357 for (i = 0; i < ARRAY_SIZE(dev->cq_pools); i++) {
358 list_for_each_entry_safe(cq, n, &dev->cq_pools[i],
359 pool_entry) {
360 WARN_ON(cq->cqe_used);
361 list_del(&cq->pool_entry);
362 cq->shared = false;
363 ib_free_cq(cq);
364 }
365 }
366}
367
368static int ib_alloc_cqs(struct ib_device *dev, unsigned int nr_cqes,
369 enum ib_poll_context poll_ctx)
370{
371 LIST_HEAD(tmp_list);
372 unsigned int nr_cqs, i;
373 struct ib_cq *cq, *n;
374 int ret;
375
376 if (poll_ctx > IB_POLL_LAST_POOL_TYPE) {
377 WARN_ON_ONCE(poll_ctx > IB_POLL_LAST_POOL_TYPE);
378 return -EINVAL;
379 }
380
381 /*
382 * Allocate at least as many CQEs as requested, and otherwise
383 * a reasonable batch size so that we can share CQs between
384 * multiple users instead of allocating a larger number of CQs.
385 */
386 nr_cqes = min_t(unsigned int, dev->attrs.max_cqe,
387 max(nr_cqes, IB_MAX_SHARED_CQ_SZ));
388 nr_cqs = min_t(unsigned int, dev->num_comp_vectors, num_online_cpus());
389 for (i = 0; i < nr_cqs; i++) {
390 cq = ib_alloc_cq(dev, NULL, nr_cqes, i, poll_ctx);
391 if (IS_ERR(cq)) {
392 ret = PTR_ERR(cq);
393 goto out_free_cqs;
394 }
395 cq->shared = true;
396 list_add_tail(&cq->pool_entry, &tmp_list);
397 }
398
399 spin_lock_irq(&dev->cq_pools_lock);
400 list_splice(&tmp_list, &dev->cq_pools[poll_ctx]);
401 spin_unlock_irq(&dev->cq_pools_lock);
402
403 return 0;
404
405out_free_cqs:
406 list_for_each_entry_safe(cq, n, &tmp_list, pool_entry) {
407 cq->shared = false;
408 ib_free_cq(cq);
409 }
410 return ret;
411}
412
413/**
414 * ib_cq_pool_get() - Find the least used completion queue that matches
415 * a given cpu hint (or least used for wild card affinity) and fits
416 * nr_cqe.
417 * @dev: rdma device
418 * @nr_cqe: number of needed cqe entries
419 * @comp_vector_hint: completion vector hint (-1) for the driver to assign
420 * a comp vector based on internal counter
421 * @poll_ctx: cq polling context
422 *
423 * Finds a cq that satisfies @comp_vector_hint and @nr_cqe requirements and
424 * claim entries in it for us. In case there is no available cq, allocate
425 * a new cq with the requirements and add it to the device pool.
426 * IB_POLL_DIRECT cannot be used for shared cqs so it is not a valid value
427 * for @poll_ctx.
428 */
429struct ib_cq *ib_cq_pool_get(struct ib_device *dev, unsigned int nr_cqe,
430 int comp_vector_hint,
431 enum ib_poll_context poll_ctx)
432{
433 static unsigned int default_comp_vector;
434 unsigned int vector, num_comp_vectors;
435 struct ib_cq *cq, *found = NULL;
436 int ret;
437
438 if (poll_ctx > IB_POLL_LAST_POOL_TYPE) {
439 WARN_ON_ONCE(poll_ctx > IB_POLL_LAST_POOL_TYPE);
440 return ERR_PTR(-EINVAL);
441 }
442
443 num_comp_vectors =
444 min_t(unsigned int, dev->num_comp_vectors, num_online_cpus());
445 /* Project the affinty to the device completion vector range */
446 if (comp_vector_hint < 0) {
447 comp_vector_hint =
448 (READ_ONCE(default_comp_vector) + 1) % num_comp_vectors;
449 WRITE_ONCE(default_comp_vector, comp_vector_hint);
450 }
451 vector = comp_vector_hint % num_comp_vectors;
452
453 /*
454 * Find the least used CQ with correct affinity and
455 * enough free CQ entries
456 */
457 while (!found) {
458 spin_lock_irq(&dev->cq_pools_lock);
459 list_for_each_entry(cq, &dev->cq_pools[poll_ctx],
460 pool_entry) {
461 /*
462 * Check to see if we have found a CQ with the
463 * correct completion vector
464 */
465 if (vector != cq->comp_vector)
466 continue;
467 if (cq->cqe_used + nr_cqe > cq->cqe)
468 continue;
469 found = cq;
470 break;
471 }
472
473 if (found) {
474 found->cqe_used += nr_cqe;
475 spin_unlock_irq(&dev->cq_pools_lock);
476
477 return found;
478 }
479 spin_unlock_irq(&dev->cq_pools_lock);
480
481 /*
482 * Didn't find a match or ran out of CQs in the device
483 * pool, allocate a new array of CQs.
484 */
485 ret = ib_alloc_cqs(dev, nr_cqe, poll_ctx);
486 if (ret)
487 return ERR_PTR(ret);
488 }
489
490 return found;
491}
492EXPORT_SYMBOL(ib_cq_pool_get);
493
494/**
495 * ib_cq_pool_put - Return a CQ taken from a shared pool.
496 * @cq: The CQ to return.
497 * @nr_cqe: The max number of cqes that the user had requested.
498 */
499void ib_cq_pool_put(struct ib_cq *cq, unsigned int nr_cqe)
500{
501 if (WARN_ON_ONCE(nr_cqe > cq->cqe_used))
502 return;
503
504 spin_lock_irq(&cq->device->cq_pools_lock);
505 cq->cqe_used -= nr_cqe;
506 spin_unlock_irq(&cq->device->cq_pools_lock);
507}
508EXPORT_SYMBOL(ib_cq_pool_put);