drivers/infiniband/core/cq.c

   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 <trace/events/rdma_core.h>
  11
  12 /* # of WCs to poll for with a single call to ib_poll_cq */
  13 #define IB_POLL_BATCH                   16
  14 #define IB_POLL_BATCH_DIRECT            8
  15
  16 /* # of WCs to iterate over before yielding */
  17 #define IB_POLL_BUDGET_IRQ              256
  18 #define IB_POLL_BUDGET_WORKQUEUE        65536
  19
  20 #define IB_POLL_FLAGS \
  21         (IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)
  22
  23 static const struct dim_cq_moder
  24 rdma_dim_prof[RDMA_DIM_PARAMS_NUM_PROFILES] = {
  25         {1,   0, 1,  0},
  26         {1,   0, 4,  0},
  27         {2,   0, 4,  0},
  28         {2,   0, 8,  0},
  29         {4,   0, 8,  0},
  30         {16,  0, 8,  0},
  31         {16,  0, 16, 0},
  32         {32,  0, 16, 0},
  33         {32,  0, 32, 0},
  34 };
  35
  36 static void ib_cq_rdma_dim_work(struct work_struct *w)
  37 {
  38         struct dim *dim = container_of(w, struct dim, work);
  39         struct ib_cq *cq = dim->priv;
  40
  41         u16 usec = rdma_dim_prof[dim->profile_ix].usec;
  42         u16 comps = rdma_dim_prof[dim->profile_ix].comps;
  43
  44         dim->state = DIM_START_MEASURE;
  45
  46         trace_cq_modify(cq, comps, usec);
  47         cq->device->ops.modify_cq(cq, comps, usec);
  48 }
  49
  50 static void rdma_dim_init(struct ib_cq *cq)
  51 {
  52         struct dim *dim;
  53
  54         if (!cq->device->ops.modify_cq || !cq->device->use_cq_dim ||
  55             cq->poll_ctx == IB_POLL_DIRECT)
  56                 return;
  57
  58         dim = kzalloc(sizeof(struct dim), GFP_KERNEL);
  59         if (!dim)
  60                 return;
  61
  62         dim->state = DIM_START_MEASURE;
  63         dim->tune_state = DIM_GOING_RIGHT;
  64         dim->profile_ix = RDMA_DIM_START_PROFILE;
  65         dim->priv = cq;
  66         cq->dim = dim;
  67
  68         INIT_WORK(&dim->work, ib_cq_rdma_dim_work);
  69 }
  70
  71 static int __poll_cq(struct ib_cq *cq, int num_entries, struct ib_wc *wc)
  72 {
  73         int rc;
  74
  75         rc = ib_poll_cq(cq, num_entries, wc);
  76         trace_cq_poll(cq, num_entries, rc);
  77         return rc;
  78 }
  79
  80 static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *wcs,
  81                            int batch)
  82 {
  83         int i, n, completed = 0;
  84
  85         trace_cq_process(cq);
  86
  87         /*
  88          * budget might be (-1) if the caller does not
  89          * want to bound this call, thus we need unsigned
  90          * minimum here.
  91          */
  92         while ((n = __poll_cq(cq, min_t(u32, batch,
  93                                         budget - completed), wcs)) > 0) {
  94                 for (i = 0; i < n; i++) {
  95                         struct ib_wc *wc = &wcs[i];
  96
  97                         if (wc->wr_cqe)
  98                                 wc->wr_cqe->done(cq, wc);
  99                         else
 100                                 WARN_ON_ONCE(wc->status == IB_WC_SUCCESS);
 101                 }
 102
 103                 completed += n;
 104
 105                 if (n != batch || (budget != -1 && completed >= budget))
 106                         break;
 107         }
 108
 109         return completed;
 110 }
 111
 112 /**
 113  * ib_process_direct_cq - process a CQ in caller context
 114  * @cq:         CQ to process
 115  * @budget:     number of CQEs to poll for
 116  *
 117  * This function is used to process all outstanding CQ entries.
 118  * It does not offload CQ processing to a different context and does
 119  * not ask for completion interrupts from the HCA.
 120  * Using direct processing on CQ with non IB_POLL_DIRECT type may trigger
 121  * concurrent processing.
 122  *
 123  * Note: do not pass -1 as %budget unless it is guaranteed that the number
 124  * of completions that will be processed is small.
 125  */
 126 int ib_process_cq_direct(struct ib_cq *cq, int budget)
 127 {
 128         struct ib_wc wcs[IB_POLL_BATCH_DIRECT];
 129
 130         return __ib_process_cq(cq, budget, wcs, IB_POLL_BATCH_DIRECT);
 131 }
 132 EXPORT_SYMBOL(ib_process_cq_direct);
 133
 134 static void ib_cq_completion_direct(struct ib_cq *cq, void *private)
 135 {
 136         WARN_ONCE(1, "got unsolicited completion for CQ 0x%p\n", cq);
 137 }
 138
 139 static int ib_poll_handler(struct irq_poll *iop, int budget)
 140 {
 141         struct ib_cq *cq = container_of(iop, struct ib_cq, iop);
 142         struct dim *dim = cq->dim;
 143         int completed;
 144
 145         completed = __ib_process_cq(cq, budget, cq->wc, IB_POLL_BATCH);
 146         if (completed < budget) {
 147                 irq_poll_complete(&cq->iop);
 148                 if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) {
 149                         trace_cq_reschedule(cq);
 150                         irq_poll_sched(&cq->iop);
 151                 }
 152         }
 153
 154         if (dim)
 155                 rdma_dim(dim, completed);
 156
 157         return completed;
 158 }
 159
 160 static void ib_cq_completion_softirq(struct ib_cq *cq, void *private)
 161 {
 162         trace_cq_schedule(cq);
 163         irq_poll_sched(&cq->iop);
 164 }
 165
 166 static void ib_cq_poll_work(struct work_struct *work)
 167 {
 168         struct ib_cq *cq = container_of(work, struct ib_cq, work);
 169         int completed;
 170
 171         completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE, cq->wc,
 172                                     IB_POLL_BATCH);
 173         if (completed >= IB_POLL_BUDGET_WORKQUEUE ||
 174             ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
 175                 queue_work(cq->comp_wq, &cq->work);
 176         else if (cq->dim)
 177                 rdma_dim(cq->dim, completed);
 178 }
 179
 180 static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private)
 181 {
 182         trace_cq_schedule(cq);
 183         queue_work(cq->comp_wq, &cq->work);
 184 }
 185
 186 /**
 187  * __ib_alloc_cq_user - allocate a completion queue
 188  * @dev:                device to allocate the CQ for
 189  * @private:            driver private data, accessible from cq->cq_context
 190  * @nr_cqe:             number of CQEs to allocate
 191  * @comp_vector:        HCA completion vectors for this CQ
 192  * @poll_ctx:           context to poll the CQ from.
 193  * @caller:             module owner name.
 194  * @udata:              Valid user data or NULL for kernel object
 195  *
 196  * This is the proper interface to allocate a CQ for in-kernel users. A
 197  * CQ allocated with this interface will automatically be polled from the
 198  * specified context. The ULP must use wr->wr_cqe instead of wr->wr_id
 199  * to use this CQ abstraction.
 200  */
 201 struct ib_cq *__ib_alloc_cq_user(struct ib_device *dev, void *private,
 202                                  int nr_cqe, int comp_vector,
 203                                  enum ib_poll_context poll_ctx,
 204                                  const char *caller, struct ib_udata *udata)
 205 {
 206         struct ib_cq_init_attr cq_attr = {
 207                 .cqe            = nr_cqe,
 208                 .comp_vector    = comp_vector,
 209         };
 210         struct ib_cq *cq;
 211         int ret = -ENOMEM;
 212
 213         cq = rdma_zalloc_drv_obj(dev, ib_cq);
 214         if (!cq)
 215                 return ERR_PTR(ret);
 216
 217         cq->device = dev;
 218         cq->cq_context = private;
 219         cq->poll_ctx = poll_ctx;
 220         atomic_set(&cq->usecnt, 0);
 221
 222         cq->wc = kmalloc_array(IB_POLL_BATCH, sizeof(*cq->wc), GFP_KERNEL);
 223         if (!cq->wc)
 224                 goto out_free_cq;
 225
 226         cq->res.type = RDMA_RESTRACK_CQ;
 227         rdma_restrack_set_task(&cq->res, caller);
 228
 229         ret = dev->ops.create_cq(cq, &cq_attr, NULL);
 230         if (ret)
 231                 goto out_free_wc;
 232
 233         rdma_restrack_kadd(&cq->res);
 234
 235         rdma_dim_init(cq);
 236
 237         switch (cq->poll_ctx) {
 238         case IB_POLL_DIRECT:
 239                 cq->comp_handler = ib_cq_completion_direct;
 240                 break;
 241         case IB_POLL_SOFTIRQ:
 242                 cq->comp_handler = ib_cq_completion_softirq;
 243
 244                 irq_poll_init(&cq->iop, IB_POLL_BUDGET_IRQ, ib_poll_handler);
 245                 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 246                 break;
 247         case IB_POLL_WORKQUEUE:
 248         case IB_POLL_UNBOUND_WORKQUEUE:
 249                 cq->comp_handler = ib_cq_completion_workqueue;
 250                 INIT_WORK(&cq->work, ib_cq_poll_work);
 251                 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 252                 cq->comp_wq = (cq->poll_ctx == IB_POLL_WORKQUEUE) ?
 253                                 ib_comp_wq : ib_comp_unbound_wq;
 254                 break;
 255         default:
 256                 ret = -EINVAL;
 257                 goto out_destroy_cq;
 258         }
 259
 260         trace_cq_alloc(cq, nr_cqe, comp_vector, poll_ctx);
 261         return cq;
 262
 263 out_destroy_cq:
 264         rdma_restrack_del(&cq->res);
 265         cq->device->ops.destroy_cq(cq, udata);
 266 out_free_wc:
 267         kfree(cq->wc);
 268 out_free_cq:
 269         kfree(cq);
 270         trace_cq_alloc_error(nr_cqe, comp_vector, poll_ctx, ret);
 271         return ERR_PTR(ret);
 272 }
 273 EXPORT_SYMBOL(__ib_alloc_cq_user);
 274
 275 /**
 276  * __ib_alloc_cq_any - allocate a completion queue
 277  * @dev:                device to allocate the CQ for
 278  * @private:            driver private data, accessible from cq->cq_context
 279  * @nr_cqe:             number of CQEs to allocate
 280  * @poll_ctx:           context to poll the CQ from
 281  * @caller:             module owner name
 282  *
 283  * Attempt to spread ULP Completion Queues over each device's interrupt
 284  * vectors. A simple best-effort mechanism is used.
 285  */
 286 struct ib_cq *__ib_alloc_cq_any(struct ib_device *dev, void *private,
 287                                 int nr_cqe, enum ib_poll_context poll_ctx,
 288                                 const char *caller)
 289 {
 290         static atomic_t counter;
 291         int comp_vector = 0;
 292
 293         if (dev->num_comp_vectors > 1)
 294                 comp_vector =
 295                         atomic_inc_return(&counter) %
 296                         min_t(int, dev->num_comp_vectors, num_online_cpus());
 297
 298         return __ib_alloc_cq_user(dev, private, nr_cqe, comp_vector, poll_ctx,
 299                                   caller, NULL);
 300 }
 301 EXPORT_SYMBOL(__ib_alloc_cq_any);
 302
 303 /**
 304  * ib_free_cq_user - free a completion queue
 305  * @cq:         completion queue to free.
 306  * @udata:      User data or NULL for kernel object
 307  */
 308 void ib_free_cq_user(struct ib_cq *cq, struct ib_udata *udata)
 309 {
 310         if (WARN_ON_ONCE(atomic_read(&cq->usecnt)))
 311                 return;
 312
 313         switch (cq->poll_ctx) {
 314         case IB_POLL_DIRECT:
 315                 break;
 316         case IB_POLL_SOFTIRQ:
 317                 irq_poll_disable(&cq->iop);
 318                 break;
 319         case IB_POLL_WORKQUEUE:
 320         case IB_POLL_UNBOUND_WORKQUEUE:
 321                 cancel_work_sync(&cq->work);
 322                 break;
 323         default:
 324                 WARN_ON_ONCE(1);
 325         }
 326
 327         trace_cq_free(cq);
 328         rdma_restrack_del(&cq->res);
 329         cq->device->ops.destroy_cq(cq, udata);
 330         if (cq->dim)
 331                 cancel_work_sync(&cq->dim->work);
 332         kfree(cq->dim);
 333         kfree(cq->wc);
 334         kfree(cq);
 335 }
 336 EXPORT_SYMBOL(ib_free_cq_user);