fs/orangefs/waitqueue.c

   1 /*
   2  * (C) 2001 Clemson University and The University of Chicago
   3  * (C) 2011 Omnibond Systems
   4  *
   5  * Changes by Acxiom Corporation to implement generic service_operation()
   6  * function, Copyright Acxiom Corporation, 2005.
   7  *
   8  * See COPYING in top-level directory.
   9  */
  10
  11 /*
  12  *  In-kernel waitqueue operations.
  13  */
  14
  15 #include "protocol.h"
  16 #include "orangefs-kernel.h"
  17 #include "orangefs-bufmap.h"
  18
  19 static int wait_for_matching_downcall(struct orangefs_kernel_op_s *);
  20 static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *);
  21
  22 /*
  23  * What we do in this function is to walk the list of operations that are
  24  * present in the request queue and mark them as purged.
  25  * NOTE: This is called from the device close after client-core has
  26  * guaranteed that no new operations could appear on the list since the
  27  * client-core is anyway going to exit.
  28  */
  29 void purge_waiting_ops(void)
  30 {
  31         struct orangefs_kernel_op_s *op;
  32
  33         spin_lock(&orangefs_request_list_lock);
  34         list_for_each_entry(op, &orangefs_request_list, list) {
  35                 gossip_debug(GOSSIP_WAIT_DEBUG,
  36                              "pvfs2-client-core: purging op tag %llu %s\n",
  37                              llu(op->tag),
  38                              get_opname_string(op));
  39                 set_op_state_purged(op);
  40         }
  41         spin_unlock(&orangefs_request_list_lock);
  42 }
  43
  44 /*
  45  * submits a ORANGEFS operation and waits for it to complete
  46  *
  47  * Note op->downcall.status will contain the status of the operation (in
  48  * errno format), whether provided by pvfs2-client or a result of failure to
  49  * service the operation.  If the caller wishes to distinguish, then
  50  * op->state can be checked to see if it was serviced or not.
  51  *
  52  * Returns contents of op->downcall.status for convenience
  53  */
  54 int service_operation(struct orangefs_kernel_op_s *op,
  55                       const char *op_name,
  56                       int flags)
  57 {
  58         /* flags to modify behavior */
  59         sigset_t orig_sigset;
  60         int ret = 0;
  61
  62         DEFINE_WAIT(wait_entry);
  63
  64         op->upcall.tgid = current->tgid;
  65         op->upcall.pid = current->pid;
  66
  67 retry_servicing:
  68         op->downcall.status = 0;
  69         gossip_debug(GOSSIP_WAIT_DEBUG,
  70                      "orangefs: service_operation: %s %p\n",
  71                      op_name,
  72                      op);
  73         gossip_debug(GOSSIP_WAIT_DEBUG,
  74                      "orangefs: operation posted by process: %s, pid: %i\n",
  75                      current->comm,
  76                      current->pid);
  77
  78         /* mask out signals if this operation is not to be interrupted */
  79         if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
  80                 orangefs_block_signals(&orig_sigset);
  81
  82         if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) {
  83                 ret = mutex_lock_interruptible(&request_mutex);
  84                 /*
  85                  * check to see if we were interrupted while waiting for
  86                  * semaphore
  87                  */
  88                 if (ret < 0) {
  89                         if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
  90                                 orangefs_set_signals(&orig_sigset);
  91                         op->downcall.status = ret;
  92                         gossip_debug(GOSSIP_WAIT_DEBUG,
  93                                      "orangefs: service_operation interrupted.\n");
  94                         return ret;
  95                 }
  96         }
  97
  98         /* queue up the operation */
  99         spin_lock(&orangefs_request_list_lock);
 100         spin_lock(&op->lock);
 101         set_op_state_waiting(op);
 102         if (flags & ORANGEFS_OP_PRIORITY)
 103                 list_add(&op->list, &orangefs_request_list);
 104         else
 105                 list_add_tail(&op->list, &orangefs_request_list);
 106         spin_unlock(&op->lock);
 107         wake_up_interruptible(&orangefs_request_list_waitq);
 108         if (!__is_daemon_in_service()) {
 109                 /*
 110                  * By incrementing the per-operation attempt counter, we
 111                  * directly go into the timeout logic while waiting for
 112                  * the matching downcall to be read
 113                  */
 114                 gossip_debug(GOSSIP_WAIT_DEBUG,
 115                              "%s:client core is NOT in service.\n",
 116                              __func__);
 117                 op->attempts++;
 118         }
 119         spin_unlock(&orangefs_request_list_lock);
 120
 121         if (!(flags & ORANGEFS_OP_NO_SEMAPHORE))
 122                 mutex_unlock(&request_mutex);
 123
 124         /*
 125          * If we are asked to service an asynchronous operation from
 126          * VFS perspective, we are done.
 127          */
 128         if (flags & ORANGEFS_OP_ASYNC)
 129                 return 0;
 130
 131         ret = wait_for_matching_downcall(op);
 132
 133         if (ret < 0) {
 134                 /* failed to get matching downcall */
 135                 if (ret == -ETIMEDOUT) {
 136                         gossip_err("orangefs: %s -- wait timed out; aborting attempt.\n",
 137                                    op_name);
 138                 }
 139                 orangefs_clean_up_interrupted_operation(op);
 140                 op->downcall.status = ret;
 141         } else {
 142                 spin_unlock(&op->lock);
 143                 /* got matching downcall; make sure status is in errno format */
 144                 op->downcall.status =
 145                     orangefs_normalize_to_errno(op->downcall.status);
 146                 ret = op->downcall.status;
 147         }
 148
 149         if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
 150                 orangefs_set_signals(&orig_sigset);
 151
 152         BUG_ON(ret != op->downcall.status);
 153         /* retry if operation has not been serviced and if requested */
 154         if (!op_state_serviced(op) && op->downcall.status == -EAGAIN) {
 155                 gossip_debug(GOSSIP_WAIT_DEBUG,
 156                              "orangefs: tag %llu (%s)"
 157                              " -- operation to be retried (%d attempt)\n",
 158                              llu(op->tag),
 159                              op_name,
 160                              op->attempts + 1);
 161
 162                 if (!op->uses_shared_memory)
 163                         /*
 164                          * this operation doesn't use the shared memory
 165                          * system
 166                          */
 167                         goto retry_servicing;
 168
 169                 /* op uses shared memory */
 170                 if (orangefs_get_bufmap_init() == 0) {
 171                         WARN_ON(1);
 172                         /*
 173                          * This operation uses the shared memory system AND
 174                          * the system is not yet ready. This situation occurs
 175                          * when the client-core is restarted AND there were
 176                          * operations waiting to be processed or were already
 177                          * in process.
 178                          */
 179                         gossip_debug(GOSSIP_WAIT_DEBUG,
 180                                      "uses_shared_memory is true.\n");
 181                         gossip_debug(GOSSIP_WAIT_DEBUG,
 182                                      "Client core in-service status(%d).\n",
 183                                      is_daemon_in_service());
 184                         gossip_debug(GOSSIP_WAIT_DEBUG, "bufmap_init:%d.\n",
 185                                      orangefs_get_bufmap_init());
 186                         gossip_debug(GOSSIP_WAIT_DEBUG,
 187                                      "operation's status is 0x%0x.\n",
 188                                      op->op_state);
 189
 190                         /*
 191                          * let process sleep for a few seconds so shared
 192                          * memory system can be initialized.
 193                          */
 194                         prepare_to_wait(&orangefs_bufmap_init_waitq,
 195                                         &wait_entry,
 196                                         TASK_INTERRUPTIBLE);
 197
 198                         /*
 199                          * Wait for orangefs_bufmap_initialize() to wake me up
 200                          * within the allotted time.
 201                          */
 202                         ret = schedule_timeout(
 203                                 ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ);
 204
 205                         gossip_debug(GOSSIP_WAIT_DEBUG,
 206                                      "Value returned from schedule_timeout:"
 207                                      "%d.\n",
 208                                      ret);
 209                         gossip_debug(GOSSIP_WAIT_DEBUG,
 210                                      "Is shared memory available? (%d).\n",
 211                                      orangefs_get_bufmap_init());
 212
 213                         finish_wait(&orangefs_bufmap_init_waitq, &wait_entry);
 214
 215                         if (orangefs_get_bufmap_init() == 0) {
 216                                 gossip_err("%s:The shared memory system has not started in %d seconds after the client core restarted.  Aborting user's request(%s).\n",
 217                                            __func__,
 218                                            ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS,
 219                                            get_opname_string(op));
 220                                 return -EIO;
 221                         }
 222
 223                         /*
 224                          * Return to the calling function and re-populate a
 225                          * shared memory buffer.
 226                          */
 227                         return -EAGAIN;
 228                 }
 229         }
 230
 231         gossip_debug(GOSSIP_WAIT_DEBUG,
 232                      "orangefs: service_operation %s returning: %d for %p.\n",
 233                      op_name,
 234                      ret,
 235                      op);
 236         return ret;
 237 }
 238
 239 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op)
 240 {
 241         u64 tag = op->tag;
 242         if (!op_state_in_progress(op))
 243                 return false;
 244
 245         op->slot_to_free = op->upcall.req.io.buf_index;
 246         memset(&op->upcall, 0, sizeof(op->upcall));
 247         memset(&op->downcall, 0, sizeof(op->downcall));
 248         op->upcall.type = ORANGEFS_VFS_OP_CANCEL;
 249         op->upcall.req.cancel.op_tag = tag;
 250         op->downcall.type = ORANGEFS_VFS_OP_INVALID;
 251         op->downcall.status = -1;
 252         orangefs_new_tag(op);
 253
 254         spin_lock(&orangefs_request_list_lock);
 255         /* orangefs_request_list_lock is enough of a barrier here */
 256         if (!__is_daemon_in_service()) {
 257                 spin_unlock(&orangefs_request_list_lock);
 258                 return false;
 259         }
 260         spin_lock(&op->lock);
 261         set_op_state_waiting(op);
 262         list_add(&op->list, &orangefs_request_list);
 263         spin_unlock(&op->lock);
 264         spin_unlock(&orangefs_request_list_lock);
 265
 266         gossip_debug(GOSSIP_UTILS_DEBUG,
 267                      "Attempting ORANGEFS operation cancellation of tag %llu\n",
 268                      llu(tag));
 269         return true;
 270 }
 271
 272 static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
 273 {
 274         /*
 275          * handle interrupted cases depending on what state we were in when
 276          * the interruption is detected.  there is a coarse grained lock
 277          * across the operation.
 278          *
 279          * Called with op->lock held.
 280          */
 281         op->op_state |= OP_VFS_STATE_GIVEN_UP;
 282
 283         if (op_state_waiting(op)) {
 284                 /*
 285                  * upcall hasn't been read; remove op from upcall request
 286                  * list.
 287                  */
 288                 spin_unlock(&op->lock);
 289                 spin_lock(&orangefs_request_list_lock);
 290                 list_del(&op->list);
 291                 spin_unlock(&orangefs_request_list_lock);
 292                 gossip_debug(GOSSIP_WAIT_DEBUG,
 293                              "Interrupted: Removed op %p from request_list\n",
 294                              op);
 295         } else if (op_state_in_progress(op)) {
 296                 /* op must be removed from the in progress htable */
 297                 spin_unlock(&op->lock);
 298                 spin_lock(&htable_ops_in_progress_lock);
 299                 list_del(&op->list);
 300                 spin_unlock(&htable_ops_in_progress_lock);
 301                 gossip_debug(GOSSIP_WAIT_DEBUG,
 302                              "Interrupted: Removed op %p"
 303                              " from htable_ops_in_progress\n",
 304                              op);
 305         } else if (!op_state_serviced(op)) {
 306                 spin_unlock(&op->lock);
 307                 gossip_err("interrupted operation is in a weird state 0x%x\n",
 308                            op->op_state);
 309         } else {
 310                 /*
 311                  * It is not intended for execution to flow here,
 312                  * but having this unlock here makes sparse happy.
 313                  */
 314                 gossip_err("%s: can't get here.\n", __func__);
 315                 spin_unlock(&op->lock);
 316         }
 317         reinit_completion(&op->waitq);
 318 }
 319
 320 /*
 321  * sleeps on waitqueue waiting for matching downcall.
 322  * if client-core finishes servicing, then we are good to go.
 323  * else if client-core exits, we get woken up here, and retry with a timeout
 324  *
 325  * Post when this call returns to the caller, the specified op will no
 326  * longer be on any list or htable.
 327  *
 328  * Returns 0 on success and -errno on failure
 329  * Errors are:
 330  * EAGAIN in case we want the caller to requeue and try again..
 331  * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
 332  * operation since client-core seems to be exiting too often
 333  * or if we were interrupted.
 334  *
 335  * Returns with op->lock taken.
 336  */
 337 static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op)
 338 {
 339         long timeout, n;
 340
 341         timeout = op->attempts ? op_timeout_secs * HZ : MAX_SCHEDULE_TIMEOUT;
 342         n = wait_for_completion_interruptible_timeout(&op->waitq, timeout);
 343         spin_lock(&op->lock);
 344
 345         if (op_state_serviced(op))
 346                 return 0;
 347
 348         if (unlikely(n < 0)) {
 349                 gossip_debug(GOSSIP_WAIT_DEBUG,
 350                              "*** %s:"
 351                              " operation interrupted by a signal (tag "
 352                              "%llu, op %p)\n",
 353                              __func__,
 354                              llu(op->tag),
 355                              op);
 356                 return -EINTR;
 357         }
 358         op->attempts++;
 359         if (op_state_purged(op)) {
 360                 gossip_debug(GOSSIP_WAIT_DEBUG,
 361                              "*** %s:"
 362                              " operation purged (tag "
 363                              "%llu, %p, att %d)\n",
 364                              __func__,
 365                              llu(op->tag),
 366                              op,
 367                              op->attempts);
 368                 return (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
 369                          -EAGAIN :
 370                          -EIO;
 371         }
 372         /* must have timed out, then... */
 373         gossip_debug(GOSSIP_WAIT_DEBUG,
 374                      "*** %s:"
 375                      " operation timed out (tag"
 376                      " %llu, %p, att %d)\n",
 377                      __func__,
 378                      llu(op->tag),
 379                      op,
 380                      op->attempts);
 381         return -ETIMEDOUT;
 382 }