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/* # of WCs to poll for with a single call to ib_poll_cq */
 11#define IB_POLL_BATCH			16
 12#define IB_POLL_BATCH_DIRECT		8
 13
 14/* # of WCs to iterate over before yielding */
 15#define IB_POLL_BUDGET_IRQ		256
 16#define IB_POLL_BUDGET_WORKQUEUE	65536
 17
 18#define IB_POLL_FLAGS \
 19	(IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)
 20
 21static const struct dim_cq_moder
 22rdma_dim_prof[RDMA_DIM_PARAMS_NUM_PROFILES] = {
 23	{1,   0, 1,  0},
 24	{1,   0, 4,  0},
 25	{2,   0, 4,  0},
 26	{2,   0, 8,  0},
 27	{4,   0, 8,  0},
 28	{16,  0, 8,  0},
 29	{16,  0, 16, 0},
 30	{32,  0, 16, 0},
 31	{32,  0, 32, 0},
 32};
 33
 34static void ib_cq_rdma_dim_work(struct work_struct *w)
 35{
 36	struct dim *dim = container_of(w, struct dim, work);
 37	struct ib_cq *cq = dim->priv;
 38
 39	u16 usec = rdma_dim_prof[dim->profile_ix].usec;
 40	u16 comps = rdma_dim_prof[dim->profile_ix].comps;
 41
 42	dim->state = DIM_START_MEASURE;
 43
 44	cq->device->ops.modify_cq(cq, comps, usec);
 45}
 46
 47static void rdma_dim_init(struct ib_cq *cq)
 48{
 49	struct dim *dim;
 50
 51	if (!cq->device->ops.modify_cq || !cq->device->use_cq_dim ||
 52	    cq->poll_ctx == IB_POLL_DIRECT)
 53		return;
 54
 55	dim = kzalloc(sizeof(struct dim), GFP_KERNEL);
 56	if (!dim)
 57		return;
 58
 59	dim->state = DIM_START_MEASURE;
 60	dim->tune_state = DIM_GOING_RIGHT;
 61	dim->profile_ix = RDMA_DIM_START_PROFILE;
 62	dim->priv = cq;
 63	cq->dim = dim;
 64
 65	INIT_WORK(&dim->work, ib_cq_rdma_dim_work);
 66}
 67
 68static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *wcs,
 69			   int batch)
 70{
 71	int i, n, completed = 0;
 72
 73	/*
 74	 * budget might be (-1) if the caller does not
 75	 * want to bound this call, thus we need unsigned
 76	 * minimum here.
 77	 */
 78	while ((n = ib_poll_cq(cq, min_t(u32, batch,
 79					 budget - completed), wcs)) > 0) {
 80		for (i = 0; i < n; i++) {
 81			struct ib_wc *wc = &wcs[i];
 82
 83			if (wc->wr_cqe)
 84				wc->wr_cqe->done(cq, wc);
 85			else
 86				WARN_ON_ONCE(wc->status == IB_WC_SUCCESS);
 87		}
 88
 89		completed += n;
 90
 91		if (n != batch || (budget != -1 && completed >= budget))
 92			break;
 93	}
 94
 95	return completed;
 96}
 97
 98/**
 99 * ib_process_direct_cq - process a CQ in caller context
100 * @cq:		CQ to process
101 * @budget:	number of CQEs to poll for
102 *
103 * This function is used to process all outstanding CQ entries.
104 * It does not offload CQ processing to a different context and does
105 * not ask for completion interrupts from the HCA.
106 * Using direct processing on CQ with non IB_POLL_DIRECT type may trigger
107 * concurrent processing.
108 *
109 * Note: do not pass -1 as %budget unless it is guaranteed that the number
110 * of completions that will be processed is small.
111 */
112int ib_process_cq_direct(struct ib_cq *cq, int budget)
113{
114	struct ib_wc wcs[IB_POLL_BATCH_DIRECT];
115
116	return __ib_process_cq(cq, budget, wcs, IB_POLL_BATCH_DIRECT);
117}
118EXPORT_SYMBOL(ib_process_cq_direct);
119
120static void ib_cq_completion_direct(struct ib_cq *cq, void *private)
121{
122	WARN_ONCE(1, "got unsolicited completion for CQ 0x%p\n", cq);
123}
124
125static int ib_poll_handler(struct irq_poll *iop, int budget)
126{
127	struct ib_cq *cq = container_of(iop, struct ib_cq, iop);
128	struct dim *dim = cq->dim;
129	int completed;
130
131	completed = __ib_process_cq(cq, budget, cq->wc, IB_POLL_BATCH);
132	if (completed < budget) {
133		irq_poll_complete(&cq->iop);
134		if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
135			irq_poll_sched(&cq->iop);
136	}
137
138	if (dim)
139		rdma_dim(dim, completed);
140
141	return completed;
142}
143
144static void ib_cq_completion_softirq(struct ib_cq *cq, void *private)
145{
146	irq_poll_sched(&cq->iop);
147}
148
149static void ib_cq_poll_work(struct work_struct *work)
150{
151	struct ib_cq *cq = container_of(work, struct ib_cq, work);
152	int completed;
153
154	completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE, cq->wc,
155				    IB_POLL_BATCH);
156	if (completed >= IB_POLL_BUDGET_WORKQUEUE ||
157	    ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
158		queue_work(cq->comp_wq, &cq->work);
159	else if (cq->dim)
160		rdma_dim(cq->dim, completed);
161}
162
163static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private)
164{
165	queue_work(cq->comp_wq, &cq->work);
166}
167
168/**
169 * __ib_alloc_cq_user - allocate a completion queue
170 * @dev:		device to allocate the CQ for
171 * @private:		driver private data, accessible from cq->cq_context
172 * @nr_cqe:		number of CQEs to allocate
173 * @comp_vector:	HCA completion vectors for this CQ
174 * @poll_ctx:		context to poll the CQ from.
175 * @caller:		module owner name.
176 * @udata:		Valid user data or NULL for kernel object
177 *
178 * This is the proper interface to allocate a CQ for in-kernel users. A
179 * CQ allocated with this interface will automatically be polled from the
180 * specified context. The ULP must use wr->wr_cqe instead of wr->wr_id
181 * to use this CQ abstraction.
182 */
183struct ib_cq *__ib_alloc_cq_user(struct ib_device *dev, void *private,
184				 int nr_cqe, int comp_vector,
185				 enum ib_poll_context poll_ctx,
186				 const char *caller, struct ib_udata *udata)
187{
188	struct ib_cq_init_attr cq_attr = {
189		.cqe		= nr_cqe,
190		.comp_vector	= comp_vector,
191	};
192	struct ib_cq *cq;
193	int ret = -ENOMEM;
194
195	cq = rdma_zalloc_drv_obj(dev, ib_cq);
196	if (!cq)
197		return ERR_PTR(ret);
198
199	cq->device = dev;
200	cq->cq_context = private;
201	cq->poll_ctx = poll_ctx;
202	atomic_set(&cq->usecnt, 0);
203
204	cq->wc = kmalloc_array(IB_POLL_BATCH, sizeof(*cq->wc), GFP_KERNEL);
205	if (!cq->wc)
206		goto out_free_cq;
207
208	cq->res.type = RDMA_RESTRACK_CQ;
209	rdma_restrack_set_task(&cq->res, caller);
210
211	ret = dev->ops.create_cq(cq, &cq_attr, NULL);
212	if (ret)
213		goto out_free_wc;
214
215	rdma_restrack_kadd(&cq->res);
216
217	rdma_dim_init(cq);
218
219	switch (cq->poll_ctx) {
220	case IB_POLL_DIRECT:
221		cq->comp_handler = ib_cq_completion_direct;
222		break;
223	case IB_POLL_SOFTIRQ:
224		cq->comp_handler = ib_cq_completion_softirq;
225
226		irq_poll_init(&cq->iop, IB_POLL_BUDGET_IRQ, ib_poll_handler);
227		ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
228		break;
229	case IB_POLL_WORKQUEUE:
230	case IB_POLL_UNBOUND_WORKQUEUE:
231		cq->comp_handler = ib_cq_completion_workqueue;
232		INIT_WORK(&cq->work, ib_cq_poll_work);
233		ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
234		cq->comp_wq = (cq->poll_ctx == IB_POLL_WORKQUEUE) ?
235				ib_comp_wq : ib_comp_unbound_wq;
236		break;
237	default:
238		ret = -EINVAL;
239		goto out_destroy_cq;
240	}
241
242	return cq;
243
244out_destroy_cq:
245	rdma_restrack_del(&cq->res);
246	cq->device->ops.destroy_cq(cq, udata);
247out_free_wc:
248	kfree(cq->wc);
249out_free_cq:
250	kfree(cq);
251	return ERR_PTR(ret);
252}
253EXPORT_SYMBOL(__ib_alloc_cq_user);
254
255/**
256 * __ib_alloc_cq_any - allocate a completion queue
257 * @dev:		device to allocate the CQ for
258 * @private:		driver private data, accessible from cq->cq_context
259 * @nr_cqe:		number of CQEs to allocate
260 * @poll_ctx:		context to poll the CQ from
261 * @caller:		module owner name
262 *
263 * Attempt to spread ULP Completion Queues over each device's interrupt
264 * vectors. A simple best-effort mechanism is used.
265 */
266struct ib_cq *__ib_alloc_cq_any(struct ib_device *dev, void *private,
267				int nr_cqe, enum ib_poll_context poll_ctx,
268				const char *caller)
269{
270	static atomic_t counter;
271	int comp_vector = 0;
272
273	if (dev->num_comp_vectors > 1)
274		comp_vector =
275			atomic_inc_return(&counter) %
276			min_t(int, dev->num_comp_vectors, num_online_cpus());
277
278	return __ib_alloc_cq_user(dev, private, nr_cqe, comp_vector, poll_ctx,
279				  caller, NULL);
280}
281EXPORT_SYMBOL(__ib_alloc_cq_any);
282
283/**
284 * ib_free_cq_user - free a completion queue
285 * @cq:		completion queue to free.
286 * @udata:	User data or NULL for kernel object
287 */
288void ib_free_cq_user(struct ib_cq *cq, struct ib_udata *udata)
289{
290	if (WARN_ON_ONCE(atomic_read(&cq->usecnt)))
291		return;
292
293	switch (cq->poll_ctx) {
294	case IB_POLL_DIRECT:
295		break;
296	case IB_POLL_SOFTIRQ:
297		irq_poll_disable(&cq->iop);
298		break;
299	case IB_POLL_WORKQUEUE:
300	case IB_POLL_UNBOUND_WORKQUEUE:
301		cancel_work_sync(&cq->work);
302		break;
303	default:
304		WARN_ON_ONCE(1);
305	}
306
307	rdma_restrack_del(&cq->res);
308	cq->device->ops.destroy_cq(cq, udata);
309	if (cq->dim)
310		cancel_work_sync(&cq->dim->work);
311	kfree(cq->dim);
312	kfree(cq->wc);
313	kfree(cq);
314}
315EXPORT_SYMBOL(ib_free_cq_user);