drivers/infiniband/hw/ipath/ipath_cq.c

   1 /*
   2  * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/err.h>
  35 #include <linux/vmalloc.h>
  36
  37 #include "ipath_verbs.h"
  38
  39 /**
  40  * ipath_cq_enter - add a new entry to the completion queue
  41  * @cq: completion queue
  42  * @entry: work completion entry to add
  43  * @sig: true if @entry is a solicitated entry
  44  *
  45  * This may be called with qp->s_lock held.
  46  */
  47 void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
  48 {
  49         struct ipath_cq_wc *wc;
  50         unsigned long flags;
  51         u32 head;
  52         u32 next;
  53
  54         spin_lock_irqsave(&cq->lock, flags);
  55
  56         /*
  57          * Note that the head pointer might be writable by user processes.
  58          * Take care to verify it is a sane value.
  59          */
  60         wc = cq->queue;
  61         head = wc->head;
  62         if (head >= (unsigned) cq->ibcq.cqe) {
  63                 head = cq->ibcq.cqe;
  64                 next = 0;
  65         } else
  66                 next = head + 1;
  67         if (unlikely(next == wc->tail)) {
  68                 spin_unlock_irqrestore(&cq->lock, flags);
  69                 if (cq->ibcq.event_handler) {
  70                         struct ib_event ev;
  71
  72                         ev.device = cq->ibcq.device;
  73                         ev.element.cq = &cq->ibcq;
  74                         ev.event = IB_EVENT_CQ_ERR;
  75                         cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
  76                 }
  77                 return;
  78         }
  79         if (cq->ip) {
  80                 wc->uqueue[head].wr_id = entry->wr_id;
  81                 wc->uqueue[head].status = entry->status;
  82                 wc->uqueue[head].opcode = entry->opcode;
  83                 wc->uqueue[head].vendor_err = entry->vendor_err;
  84                 wc->uqueue[head].byte_len = entry->byte_len;
  85                 wc->uqueue[head].imm_data = (__u32 __force)entry->imm_data;
  86                 wc->uqueue[head].qp_num = entry->qp->qp_num;
  87                 wc->uqueue[head].src_qp = entry->src_qp;
  88                 wc->uqueue[head].wc_flags = entry->wc_flags;
  89                 wc->uqueue[head].pkey_index = entry->pkey_index;
  90                 wc->uqueue[head].slid = entry->slid;
  91                 wc->uqueue[head].sl = entry->sl;
  92                 wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits;
  93                 wc->uqueue[head].port_num = entry->port_num;
  94                 /* Make sure entry is written before the head index. */
  95                 smp_wmb();
  96         } else
  97                 wc->kqueue[head] = *entry;
  98         wc->head = next;
  99
 100         if (cq->notify == IB_CQ_NEXT_COMP ||
 101             (cq->notify == IB_CQ_SOLICITED && solicited)) {
 102                 cq->notify = IB_CQ_NONE;
 103                 cq->triggered++;
 104                 /*
 105                  * This will cause send_complete() to be called in
 106                  * another thread.
 107                  */
 108                 tasklet_hi_schedule(&cq->comptask);
 109         }
 110
 111         spin_unlock_irqrestore(&cq->lock, flags);
 112
 113         if (entry->status != IB_WC_SUCCESS)
 114                 to_idev(cq->ibcq.device)->n_wqe_errs++;
 115 }
 116
 117 /**
 118  * ipath_poll_cq - poll for work completion entries
 119  * @ibcq: the completion queue to poll
 120  * @num_entries: the maximum number of entries to return
 121  * @entry: pointer to array where work completions are placed
 122  *
 123  * Returns the number of completion entries polled.
 124  *
 125  * This may be called from interrupt context.  Also called by ib_poll_cq()
 126  * in the generic verbs code.
 127  */
 128 int ipath_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
 129 {
 130         struct ipath_cq *cq = to_icq(ibcq);
 131         struct ipath_cq_wc *wc;
 132         unsigned long flags;
 133         int npolled;
 134         u32 tail;
 135
 136         /* The kernel can only poll a kernel completion queue */
 137         if (cq->ip) {
 138                 npolled = -EINVAL;
 139                 goto bail;
 140         }
 141
 142         spin_lock_irqsave(&cq->lock, flags);
 143
 144         wc = cq->queue;
 145         tail = wc->tail;
 146         if (tail > (u32) cq->ibcq.cqe)
 147                 tail = (u32) cq->ibcq.cqe;
 148         for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
 149                 if (tail == wc->head)
 150                         break;
 151                 /* The kernel doesn't need a RMB since it has the lock. */
 152                 *entry = wc->kqueue[tail];
 153                 if (tail >= cq->ibcq.cqe)
 154                         tail = 0;
 155                 else
 156                         tail++;
 157         }
 158         wc->tail = tail;
 159
 160         spin_unlock_irqrestore(&cq->lock, flags);
 161
 162 bail:
 163         return npolled;
 164 }
 165
 166 static void send_complete(unsigned long data)
 167 {
 168         struct ipath_cq *cq = (struct ipath_cq *)data;
 169
 170         /*
 171          * The completion handler will most likely rearm the notification
 172          * and poll for all pending entries.  If a new completion entry
 173          * is added while we are in this routine, tasklet_hi_schedule()
 174          * won't call us again until we return so we check triggered to
 175          * see if we need to call the handler again.
 176          */
 177         for (;;) {
 178                 u8 triggered = cq->triggered;
 179
 180                 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 181
 182                 if (cq->triggered == triggered)
 183                         return;
 184         }
 185 }
 186
 187 /**
 188  * ipath_create_cq - create a completion queue
 189  * @ibdev: the device this completion queue is attached to
 190  * @entries: the minimum size of the completion queue
 191  * @context: unused by the InfiniPath driver
 192  * @udata: unused by the InfiniPath driver
 193  *
 194  * Returns a pointer to the completion queue or negative errno values
 195  * for failure.
 196  *
 197  * Called by ib_create_cq() in the generic verbs code.
 198  */
 199 struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries, int comp_vector,
 200                               struct ib_ucontext *context,
 201                               struct ib_udata *udata)
 202 {
 203         struct ipath_ibdev *dev = to_idev(ibdev);
 204         struct ipath_cq *cq;
 205         struct ipath_cq_wc *wc;
 206         struct ib_cq *ret;
 207         u32 sz;
 208
 209         if (entries < 1 || entries > ib_ipath_max_cqes) {
 210                 ret = ERR_PTR(-EINVAL);
 211                 goto done;
 212         }
 213
 214         /* Allocate the completion queue structure. */
 215         cq = kmalloc(sizeof(*cq), GFP_KERNEL);
 216         if (!cq) {
 217                 ret = ERR_PTR(-ENOMEM);
 218                 goto done;
 219         }
 220
 221         /*
 222          * Allocate the completion queue entries and head/tail pointers.
 223          * This is allocated separately so that it can be resized and
 224          * also mapped into user space.
 225          * We need to use vmalloc() in order to support mmap and large
 226          * numbers of entries.
 227          */
 228         sz = sizeof(*wc);
 229         if (udata && udata->outlen >= sizeof(__u64))
 230                 sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
 231         else
 232                 sz += sizeof(struct ib_wc) * (entries + 1);
 233         wc = vmalloc_user(sz);
 234         if (!wc) {
 235                 ret = ERR_PTR(-ENOMEM);
 236                 goto bail_cq;
 237         }
 238
 239         /*
 240          * Return the address of the WC as the offset to mmap.
 241          * See ipath_mmap() for details.
 242          */
 243         if (udata && udata->outlen >= sizeof(__u64)) {
 244                 int err;
 245
 246                 cq->ip = ipath_create_mmap_info(dev, sz, context, wc);
 247                 if (!cq->ip) {
 248                         ret = ERR_PTR(-ENOMEM);
 249                         goto bail_wc;
 250                 }
 251
 252                 err = ib_copy_to_udata(udata, &cq->ip->offset,
 253                                        sizeof(cq->ip->offset));
 254                 if (err) {
 255                         ret = ERR_PTR(err);
 256                         goto bail_ip;
 257                 }
 258         } else
 259                 cq->ip = NULL;
 260
 261         spin_lock(&dev->n_cqs_lock);
 262         if (dev->n_cqs_allocated == ib_ipath_max_cqs) {
 263                 spin_unlock(&dev->n_cqs_lock);
 264                 ret = ERR_PTR(-ENOMEM);
 265                 goto bail_ip;
 266         }
 267
 268         dev->n_cqs_allocated++;
 269         spin_unlock(&dev->n_cqs_lock);
 270
 271         if (cq->ip) {
 272                 spin_lock_irq(&dev->pending_lock);
 273                 list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps);
 274                 spin_unlock_irq(&dev->pending_lock);
 275         }
 276
 277         /*
 278          * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
 279          * The number of entries should be >= the number requested or return
 280          * an error.
 281          */
 282         cq->ibcq.cqe = entries;
 283         cq->notify = IB_CQ_NONE;
 284         cq->triggered = 0;
 285         spin_lock_init(&cq->lock);
 286         tasklet_init(&cq->comptask, send_complete, (unsigned long)cq);
 287         wc->head = 0;
 288         wc->tail = 0;
 289         cq->queue = wc;
 290
 291         ret = &cq->ibcq;
 292
 293         goto done;
 294
 295 bail_ip:
 296         kfree(cq->ip);
 297 bail_wc:
 298         vfree(wc);
 299 bail_cq:
 300         kfree(cq);
 301 done:
 302         return ret;
 303 }
 304
 305 /**
 306  * ipath_destroy_cq - destroy a completion queue
 307  * @ibcq: the completion queue to destroy.
 308  *
 309  * Returns 0 for success.
 310  *
 311  * Called by ib_destroy_cq() in the generic verbs code.
 312  */
 313 int ipath_destroy_cq(struct ib_cq *ibcq)
 314 {
 315         struct ipath_ibdev *dev = to_idev(ibcq->device);
 316         struct ipath_cq *cq = to_icq(ibcq);
 317
 318         tasklet_kill(&cq->comptask);
 319         spin_lock(&dev->n_cqs_lock);
 320         dev->n_cqs_allocated--;
 321         spin_unlock(&dev->n_cqs_lock);
 322         if (cq->ip)
 323                 kref_put(&cq->ip->ref, ipath_release_mmap_info);
 324         else
 325                 vfree(cq->queue);
 326         kfree(cq);
 327
 328         return 0;
 329 }
 330
 331 /**
 332  * ipath_req_notify_cq - change the notification type for a completion queue
 333  * @ibcq: the completion queue
 334  * @notify_flags: the type of notification to request
 335  *
 336  * Returns 0 for success.
 337  *
 338  * This may be called from interrupt context.  Also called by
 339  * ib_req_notify_cq() in the generic verbs code.
 340  */
 341 int ipath_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
 342 {
 343         struct ipath_cq *cq = to_icq(ibcq);
 344         unsigned long flags;
 345         int ret = 0;
 346
 347         spin_lock_irqsave(&cq->lock, flags);
 348         /*
 349          * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
 350          * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
 351          */
 352         if (cq->notify != IB_CQ_NEXT_COMP)
 353                 cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
 354
 355         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
 356             cq->queue->head != cq->queue->tail)
 357                 ret = 1;
 358
 359         spin_unlock_irqrestore(&cq->lock, flags);
 360
 361         return ret;
 362 }
 363
 364 /**
 365  * ipath_resize_cq - change the size of the CQ
 366  * @ibcq: the completion queue
 367  *
 368  * Returns 0 for success.
 369  */
 370 int ipath_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
 371 {
 372         struct ipath_cq *cq = to_icq(ibcq);
 373         struct ipath_cq_wc *old_wc;
 374         struct ipath_cq_wc *wc;
 375         u32 head, tail, n;
 376         int ret;
 377         u32 sz;
 378
 379         if (cqe < 1 || cqe > ib_ipath_max_cqes) {
 380                 ret = -EINVAL;
 381                 goto bail;
 382         }
 383
 384         /*
 385          * Need to use vmalloc() if we want to support large #s of entries.
 386          */
 387         sz = sizeof(*wc);
 388         if (udata && udata->outlen >= sizeof(__u64))
 389                 sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
 390         else
 391                 sz += sizeof(struct ib_wc) * (cqe + 1);
 392         wc = vmalloc_user(sz);
 393         if (!wc) {
 394                 ret = -ENOMEM;
 395                 goto bail;
 396         }
 397
 398         /* Check that we can write the offset to mmap. */
 399         if (udata && udata->outlen >= sizeof(__u64)) {
 400                 __u64 offset = 0;
 401
 402                 ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
 403                 if (ret)
 404                         goto bail_free;
 405         }
 406
 407         spin_lock_irq(&cq->lock);
 408         /*
 409          * Make sure head and tail are sane since they
 410          * might be user writable.
 411          */
 412         old_wc = cq->queue;
 413         head = old_wc->head;
 414         if (head > (u32) cq->ibcq.cqe)
 415                 head = (u32) cq->ibcq.cqe;
 416         tail = old_wc->tail;
 417         if (tail > (u32) cq->ibcq.cqe)
 418                 tail = (u32) cq->ibcq.cqe;
 419         if (head < tail)
 420                 n = cq->ibcq.cqe + 1 + head - tail;
 421         else
 422                 n = head - tail;
 423         if (unlikely((u32)cqe < n)) {
 424                 ret = -EINVAL;
 425                 goto bail_unlock;
 426         }
 427         for (n = 0; tail != head; n++) {
 428                 if (cq->ip)
 429                         wc->uqueue[n] = old_wc->uqueue[tail];
 430                 else
 431                         wc->kqueue[n] = old_wc->kqueue[tail];
 432                 if (tail == (u32) cq->ibcq.cqe)
 433                         tail = 0;
 434                 else
 435                         tail++;
 436         }
 437         cq->ibcq.cqe = cqe;
 438         wc->head = n;
 439         wc->tail = 0;
 440         cq->queue = wc;
 441         spin_unlock_irq(&cq->lock);
 442
 443         vfree(old_wc);
 444
 445         if (cq->ip) {
 446                 struct ipath_ibdev *dev = to_idev(ibcq->device);
 447                 struct ipath_mmap_info *ip = cq->ip;
 448
 449                 ipath_update_mmap_info(dev, ip, sz, wc);
 450
 451                 /*
 452                  * Return the offset to mmap.
 453                  * See ipath_mmap() for details.
 454                  */
 455                 if (udata && udata->outlen >= sizeof(__u64)) {
 456                         ret = ib_copy_to_udata(udata, &ip->offset,
 457                                                sizeof(ip->offset));
 458                         if (ret)
 459                                 goto bail;
 460                 }
 461
 462                 spin_lock_irq(&dev->pending_lock);
 463                 if (list_empty(&ip->pending_mmaps))
 464                         list_add(&ip->pending_mmaps, &dev->pending_mmaps);
 465                 spin_unlock_irq(&dev->pending_lock);
 466         }
 467
 468         ret = 0;
 469         goto bail;
 470
 471 bail_unlock:
 472         spin_unlock_irq(&cq->lock);
 473 bail_free:
 474         vfree(wc);
 475 bail:
 476         return ret;
 477 }