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/slab.h>
  36 #include <linux/vmalloc.h>
  37
  38 #include "ipath_verbs.h"
  39
  40 /**
  41  * ipath_cq_enter - add a new entry to the completion queue
  42  * @cq: completion queue
  43  * @entry: work completion entry to add
  44  * @sig: true if @entry is a solicitated entry
  45  *
  46  * This may be called with qp->s_lock held.
  47  */
  48 void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
  49 {
  50         struct ipath_cq_wc *wc;
  51         unsigned long flags;
  52         u32 head;
  53         u32 next;
  54
  55         spin_lock_irqsave(&cq->lock, flags);
  56
  57         /*
  58          * Note that the head pointer might be writable by user processes.
  59          * Take care to verify it is a sane value.
  60          */
  61         wc = cq->queue;
  62         head = wc->head;
  63         if (head >= (unsigned) cq->ibcq.cqe) {
  64                 head = cq->ibcq.cqe;
  65                 next = 0;
  66         } else
  67                 next = head + 1;
  68         if (unlikely(next == wc->tail)) {
  69                 spin_unlock_irqrestore(&cq->lock, flags);
  70                 if (cq->ibcq.event_handler) {
  71                         struct ib_event ev;
  72
  73                         ev.device = cq->ibcq.device;
  74                         ev.element.cq = &cq->ibcq;
  75                         ev.event = IB_EVENT_CQ_ERR;
  76                         cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
  77                 }
  78                 return;
  79         }
  80         if (cq->ip) {
  81                 wc->uqueue[head].wr_id = entry->wr_id;
  82                 wc->uqueue[head].status = entry->status;
  83                 wc->uqueue[head].opcode = entry->opcode;
  84                 wc->uqueue[head].vendor_err = entry->vendor_err;
  85                 wc->uqueue[head].byte_len = entry->byte_len;
  86                 wc->uqueue[head].ex.imm_data = (__u32 __force) entry->ex.imm_data;
  87                 wc->uqueue[head].qp_num = entry->qp->qp_num;
  88                 wc->uqueue[head].src_qp = entry->src_qp;
  89                 wc->uqueue[head].wc_flags = entry->wc_flags;
  90                 wc->uqueue[head].pkey_index = entry->pkey_index;
  91                 wc->uqueue[head].slid = entry->slid;
  92                 wc->uqueue[head].sl = entry->sl;
  93                 wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits;
  94                 wc->uqueue[head].port_num = entry->port_num;
  95                 /* Make sure entry is written before the head index. */
  96                 smp_wmb();
  97         } else
  98                 wc->kqueue[head] = *entry;
  99         wc->head = next;
 100
 101         if (cq->notify == IB_CQ_NEXT_COMP ||
 102             (cq->notify == IB_CQ_SOLICITED && solicited)) {
 103                 cq->notify = IB_CQ_NONE;
 104                 cq->triggered++;
 105                 /*
 106                  * This will cause send_complete() to be called in
 107                  * another thread.
 108                  */
 109                 tasklet_hi_schedule(&cq->comptask);
 110         }
 111
 112         spin_unlock_irqrestore(&cq->lock, flags);
 113
 114         if (entry->status != IB_WC_SUCCESS)
 115                 to_idev(cq->ibcq.device)->n_wqe_errs++;
 116 }
 117
 118 /**
 119  * ipath_poll_cq - poll for work completion entries
 120  * @ibcq: the completion queue to poll
 121  * @num_entries: the maximum number of entries to return
 122  * @entry: pointer to array where work completions are placed
 123  *
 124  * Returns the number of completion entries polled.
 125  *
 126  * This may be called from interrupt context.  Also called by ib_poll_cq()
 127  * in the generic verbs code.
 128  */
 129 int ipath_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
 130 {
 131         struct ipath_cq *cq = to_icq(ibcq);
 132         struct ipath_cq_wc *wc;
 133         unsigned long flags;
 134         int npolled;
 135         u32 tail;
 136
 137         /* The kernel can only poll a kernel completion queue */
 138         if (cq->ip) {
 139                 npolled = -EINVAL;
 140                 goto bail;
 141         }
 142
 143         spin_lock_irqsave(&cq->lock, flags);
 144
 145         wc = cq->queue;
 146         tail = wc->tail;
 147         if (tail > (u32) cq->ibcq.cqe)
 148                 tail = (u32) cq->ibcq.cqe;
 149         for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
 150                 if (tail == wc->head)
 151                         break;
 152                 /* The kernel doesn't need a RMB since it has the lock. */
 153                 *entry = wc->kqueue[tail];
 154                 if (tail >= cq->ibcq.cqe)
 155                         tail = 0;
 156                 else
 157                         tail++;
 158         }
 159         wc->tail = tail;
 160
 161         spin_unlock_irqrestore(&cq->lock, flags);
 162
 163 bail:
 164         return npolled;
 165 }
 166
 167 static void send_complete(unsigned long data)
 168 {
 169         struct ipath_cq *cq = (struct ipath_cq *)data;
 170
 171         /*
 172          * The completion handler will most likely rearm the notification
 173          * and poll for all pending entries.  If a new completion entry
 174          * is added while we are in this routine, tasklet_hi_schedule()
 175          * won't call us again until we return so we check triggered to
 176          * see if we need to call the handler again.
 177          */
 178         for (;;) {
 179                 u8 triggered = cq->triggered;
 180
 181                 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 182
 183                 if (cq->triggered == triggered)
 184                         return;
 185         }
 186 }
 187
 188 /**
 189  * ipath_create_cq - create a completion queue
 190  * @ibdev: the device this completion queue is attached to
 191  * @attr: creation attributes
 192  * @context: unused by the InfiniPath driver
 193  * @udata: unused by the InfiniPath driver
 194  *
 195  * Returns a pointer to the completion queue or negative errno values
 196  * for failure.
 197  *
 198  * Called by ib_create_cq() in the generic verbs code.
 199  */
 200 struct ib_cq *ipath_create_cq(struct ib_device *ibdev,
 201                               const struct ib_cq_init_attr *attr,
 202                               struct ib_ucontext *context,
 203                               struct ib_udata *udata)
 204 {
 205         int entries = attr->cqe;
 206         struct ipath_ibdev *dev = to_idev(ibdev);
 207         struct ipath_cq *cq;
 208         struct ipath_cq_wc *wc;
 209         struct ib_cq *ret;
 210         u32 sz;
 211
 212         if (attr->flags)
 213                 return ERR_PTR(-EINVAL);
 214
 215         if (entries < 1 || entries > ib_ipath_max_cqes) {
 216                 ret = ERR_PTR(-EINVAL);
 217                 goto done;
 218         }
 219
 220         /* Allocate the completion queue structure. */
 221         cq = kmalloc(sizeof(*cq), GFP_KERNEL);
 222         if (!cq) {
 223                 ret = ERR_PTR(-ENOMEM);
 224                 goto done;
 225         }
 226
 227         /*
 228          * Allocate the completion queue entries and head/tail pointers.
 229          * This is allocated separately so that it can be resized and
 230          * also mapped into user space.
 231          * We need to use vmalloc() in order to support mmap and large
 232          * numbers of entries.
 233          */
 234         sz = sizeof(*wc);
 235         if (udata && udata->outlen >= sizeof(__u64))
 236                 sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
 237         else
 238                 sz += sizeof(struct ib_wc) * (entries + 1);
 239         wc = vmalloc_user(sz);
 240         if (!wc) {
 241                 ret = ERR_PTR(-ENOMEM);
 242                 goto bail_cq;
 243         }
 244
 245         /*
 246          * Return the address of the WC as the offset to mmap.
 247          * See ipath_mmap() for details.
 248          */
 249         if (udata && udata->outlen >= sizeof(__u64)) {
 250                 int err;
 251
 252                 cq->ip = ipath_create_mmap_info(dev, sz, context, wc);
 253                 if (!cq->ip) {
 254                         ret = ERR_PTR(-ENOMEM);
 255                         goto bail_wc;
 256                 }
 257
 258                 err = ib_copy_to_udata(udata, &cq->ip->offset,
 259                                        sizeof(cq->ip->offset));
 260                 if (err) {
 261                         ret = ERR_PTR(err);
 262                         goto bail_ip;
 263                 }
 264         } else
 265                 cq->ip = NULL;
 266
 267         spin_lock(&dev->n_cqs_lock);
 268         if (dev->n_cqs_allocated == ib_ipath_max_cqs) {
 269                 spin_unlock(&dev->n_cqs_lock);
 270                 ret = ERR_PTR(-ENOMEM);
 271                 goto bail_ip;
 272         }
 273
 274         dev->n_cqs_allocated++;
 275         spin_unlock(&dev->n_cqs_lock);
 276
 277         if (cq->ip) {
 278                 spin_lock_irq(&dev->pending_lock);
 279                 list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps);
 280                 spin_unlock_irq(&dev->pending_lock);
 281         }
 282
 283         /*
 284          * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
 285          * The number of entries should be >= the number requested or return
 286          * an error.
 287          */
 288         cq->ibcq.cqe = entries;
 289         cq->notify = IB_CQ_NONE;
 290         cq->triggered = 0;
 291         spin_lock_init(&cq->lock);
 292         tasklet_init(&cq->comptask, send_complete, (unsigned long)cq);
 293         wc->head = 0;
 294         wc->tail = 0;
 295         cq->queue = wc;
 296
 297         ret = &cq->ibcq;
 298
 299         goto done;
 300
 301 bail_ip:
 302         kfree(cq->ip);
 303 bail_wc:
 304         vfree(wc);
 305 bail_cq:
 306         kfree(cq);
 307 done:
 308         return ret;
 309 }
 310
 311 /**
 312  * ipath_destroy_cq - destroy a completion queue
 313  * @ibcq: the completion queue to destroy.
 314  *
 315  * Returns 0 for success.
 316  *
 317  * Called by ib_destroy_cq() in the generic verbs code.
 318  */
 319 int ipath_destroy_cq(struct ib_cq *ibcq)
 320 {
 321         struct ipath_ibdev *dev = to_idev(ibcq->device);
 322         struct ipath_cq *cq = to_icq(ibcq);
 323
 324         tasklet_kill(&cq->comptask);
 325         spin_lock(&dev->n_cqs_lock);
 326         dev->n_cqs_allocated--;
 327         spin_unlock(&dev->n_cqs_lock);
 328         if (cq->ip)
 329                 kref_put(&cq->ip->ref, ipath_release_mmap_info);
 330         else
 331                 vfree(cq->queue);
 332         kfree(cq);
 333
 334         return 0;
 335 }
 336
 337 /**
 338  * ipath_req_notify_cq - change the notification type for a completion queue
 339  * @ibcq: the completion queue
 340  * @notify_flags: the type of notification to request
 341  *
 342  * Returns 0 for success.
 343  *
 344  * This may be called from interrupt context.  Also called by
 345  * ib_req_notify_cq() in the generic verbs code.
 346  */
 347 int ipath_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
 348 {
 349         struct ipath_cq *cq = to_icq(ibcq);
 350         unsigned long flags;
 351         int ret = 0;
 352
 353         spin_lock_irqsave(&cq->lock, flags);
 354         /*
 355          * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
 356          * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
 357          */
 358         if (cq->notify != IB_CQ_NEXT_COMP)
 359                 cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
 360
 361         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
 362             cq->queue->head != cq->queue->tail)
 363                 ret = 1;
 364
 365         spin_unlock_irqrestore(&cq->lock, flags);
 366
 367         return ret;
 368 }
 369
 370 /**
 371  * ipath_resize_cq - change the size of the CQ
 372  * @ibcq: the completion queue
 373  *
 374  * Returns 0 for success.
 375  */
 376 int ipath_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
 377 {
 378         struct ipath_cq *cq = to_icq(ibcq);
 379         struct ipath_cq_wc *old_wc;
 380         struct ipath_cq_wc *wc;
 381         u32 head, tail, n;
 382         int ret;
 383         u32 sz;
 384
 385         if (cqe < 1 || cqe > ib_ipath_max_cqes) {
 386                 ret = -EINVAL;
 387                 goto bail;
 388         }
 389
 390         /*
 391          * Need to use vmalloc() if we want to support large #s of entries.
 392          */
 393         sz = sizeof(*wc);
 394         if (udata && udata->outlen >= sizeof(__u64))
 395                 sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
 396         else
 397                 sz += sizeof(struct ib_wc) * (cqe + 1);
 398         wc = vmalloc_user(sz);
 399         if (!wc) {
 400                 ret = -ENOMEM;
 401                 goto bail;
 402         }
 403
 404         /* Check that we can write the offset to mmap. */
 405         if (udata && udata->outlen >= sizeof(__u64)) {
 406                 __u64 offset = 0;
 407
 408                 ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
 409                 if (ret)
 410                         goto bail_free;
 411         }
 412
 413         spin_lock_irq(&cq->lock);
 414         /*
 415          * Make sure head and tail are sane since they
 416          * might be user writable.
 417          */
 418         old_wc = cq->queue;
 419         head = old_wc->head;
 420         if (head > (u32) cq->ibcq.cqe)
 421                 head = (u32) cq->ibcq.cqe;
 422         tail = old_wc->tail;
 423         if (tail > (u32) cq->ibcq.cqe)
 424                 tail = (u32) cq->ibcq.cqe;
 425         if (head < tail)
 426                 n = cq->ibcq.cqe + 1 + head - tail;
 427         else
 428                 n = head - tail;
 429         if (unlikely((u32)cqe < n)) {
 430                 ret = -EINVAL;
 431                 goto bail_unlock;
 432         }
 433         for (n = 0; tail != head; n++) {
 434                 if (cq->ip)
 435                         wc->uqueue[n] = old_wc->uqueue[tail];
 436                 else
 437                         wc->kqueue[n] = old_wc->kqueue[tail];
 438                 if (tail == (u32) cq->ibcq.cqe)
 439                         tail = 0;
 440                 else
 441                         tail++;
 442         }
 443         cq->ibcq.cqe = cqe;
 444         wc->head = n;
 445         wc->tail = 0;
 446         cq->queue = wc;
 447         spin_unlock_irq(&cq->lock);
 448
 449         vfree(old_wc);
 450
 451         if (cq->ip) {
 452                 struct ipath_ibdev *dev = to_idev(ibcq->device);
 453                 struct ipath_mmap_info *ip = cq->ip;
 454
 455                 ipath_update_mmap_info(dev, ip, sz, wc);
 456
 457                 /*
 458                  * Return the offset to mmap.
 459                  * See ipath_mmap() for details.
 460                  */
 461                 if (udata && udata->outlen >= sizeof(__u64)) {
 462                         ret = ib_copy_to_udata(udata, &ip->offset,
 463                                                sizeof(ip->offset));
 464                         if (ret)
 465                                 goto bail;
 466                 }
 467
 468                 spin_lock_irq(&dev->pending_lock);
 469                 if (list_empty(&ip->pending_mmaps))
 470                         list_add(&ip->pending_mmaps, &dev->pending_mmaps);
 471                 spin_unlock_irq(&dev->pending_lock);
 472         }
 473
 474         ret = 0;
 475         goto bail;
 476
 477 bail_unlock:
 478         spin_unlock_irq(&cq->lock);
 479 bail_free:
 480         vfree(wc);
 481 bail:
 482         return ret;
 483 }