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 *, long, bool);
  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         long timeout = MAX_SCHEDULE_TIMEOUT;
  59         /* flags to modify behavior */
  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         if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) {
  79                 if (flags & ORANGEFS_OP_INTERRUPTIBLE)
  80                         ret = mutex_lock_interruptible(&request_mutex);
  81                 else
  82                         ret = mutex_lock_killable(&request_mutex);
  83                 /*
  84                  * check to see if we were interrupted while waiting for
  85                  * semaphore
  86                  */
  87                 if (ret < 0) {
  88                         op->downcall.status = ret;
  89                         gossip_debug(GOSSIP_WAIT_DEBUG,
  90                                      "orangefs: service_operation interrupted.\n");
  91                         return ret;
  92                 }
  93         }
  94
  95         /* queue up the operation */
  96         spin_lock(&orangefs_request_list_lock);
  97         spin_lock(&op->lock);
  98         set_op_state_waiting(op);
  99         if (flags & ORANGEFS_OP_PRIORITY)
 100                 list_add(&op->list, &orangefs_request_list);
 101         else
 102                 list_add_tail(&op->list, &orangefs_request_list);
 103         spin_unlock(&op->lock);
 104         wake_up_interruptible(&orangefs_request_list_waitq);
 105         if (!__is_daemon_in_service()) {
 106                 gossip_debug(GOSSIP_WAIT_DEBUG,
 107                              "%s:client core is NOT in service.\n",
 108                              __func__);
 109                 timeout = op_timeout_secs * HZ;
 110         }
 111         spin_unlock(&orangefs_request_list_lock);
 112
 113         if (!(flags & ORANGEFS_OP_NO_SEMAPHORE))
 114                 mutex_unlock(&request_mutex);
 115
 116         /*
 117          * If we are asked to service an asynchronous operation from
 118          * VFS perspective, we are done.
 119          */
 120         if (flags & ORANGEFS_OP_ASYNC)
 121                 return 0;
 122
 123         ret = wait_for_matching_downcall(op, timeout,
 124                                          flags & ORANGEFS_OP_INTERRUPTIBLE);
 125         if (!ret) {
 126                 spin_unlock(&op->lock);
 127                 /* got matching downcall; make sure status is in errno format */
 128                 op->downcall.status =
 129                     orangefs_normalize_to_errno(op->downcall.status);
 130                 ret = op->downcall.status;
 131                 goto out;
 132         }
 133
 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         /* retry if operation has not been serviced and if requested */
 142         if (ret == -EAGAIN) {
 143                 op->attempts++;
 144                 timeout = op_timeout_secs * HZ;
 145                 gossip_debug(GOSSIP_WAIT_DEBUG,
 146                              "orangefs: tag %llu (%s)"
 147                              " -- operation to be retried (%d attempt)\n",
 148                              llu(op->tag),
 149                              op_name,
 150                              op->attempts);
 151
 152                 if (!op->uses_shared_memory)
 153                         /*
 154                          * this operation doesn't use the shared memory
 155                          * system
 156                          */
 157                         goto retry_servicing;
 158
 159                 /* op uses shared memory */
 160                 if (orangefs_get_bufmap_init() == 0) {
 161                         WARN_ON(1);
 162                         /*
 163                          * This operation uses the shared memory system AND
 164                          * the system is not yet ready. This situation occurs
 165                          * when the client-core is restarted AND there were
 166                          * operations waiting to be processed or were already
 167                          * in process.
 168                          */
 169                         gossip_debug(GOSSIP_WAIT_DEBUG,
 170                                      "uses_shared_memory is true.\n");
 171                         gossip_debug(GOSSIP_WAIT_DEBUG,
 172                                      "Client core in-service status(%d).\n",
 173                                      is_daemon_in_service());
 174                         gossip_debug(GOSSIP_WAIT_DEBUG, "bufmap_init:%d.\n",
 175                                      orangefs_get_bufmap_init());
 176                         gossip_debug(GOSSIP_WAIT_DEBUG,
 177                                      "operation's status is 0x%0x.\n",
 178                                      op->op_state);
 179
 180                         /*
 181                          * let process sleep for a few seconds so shared
 182                          * memory system can be initialized.
 183                          */
 184                         prepare_to_wait(&orangefs_bufmap_init_waitq,
 185                                         &wait_entry,
 186                                         TASK_INTERRUPTIBLE);
 187
 188                         /*
 189                          * Wait for orangefs_bufmap_initialize() to wake me up
 190                          * within the allotted time.
 191                          */
 192                         ret = schedule_timeout(
 193                                 ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS * HZ);
 194
 195                         gossip_debug(GOSSIP_WAIT_DEBUG,
 196                                      "Value returned from schedule_timeout:"
 197                                      "%d.\n",
 198                                      ret);
 199                         gossip_debug(GOSSIP_WAIT_DEBUG,
 200                                      "Is shared memory available? (%d).\n",
 201                                      orangefs_get_bufmap_init());
 202
 203                         finish_wait(&orangefs_bufmap_init_waitq, &wait_entry);
 204
 205                         if (orangefs_get_bufmap_init() == 0) {
 206                                 gossip_err("%s:The shared memory system has not started in %d seconds after the client core restarted.  Aborting user's request(%s).\n",
 207                                            __func__,
 208                                            ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS,
 209                                            get_opname_string(op));
 210                                 return -EIO;
 211                         }
 212
 213                         /*
 214                          * Return to the calling function and re-populate a
 215                          * shared memory buffer.
 216                          */
 217                         return -EAGAIN;
 218                 }
 219         }
 220
 221 out:
 222         gossip_debug(GOSSIP_WAIT_DEBUG,
 223                      "orangefs: service_operation %s returning: %d for %p.\n",
 224                      op_name,
 225                      ret,
 226                      op);
 227         return ret;
 228 }
 229
 230 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op)
 231 {
 232         u64 tag = op->tag;
 233         if (!op_state_in_progress(op))
 234                 return false;
 235
 236         op->slot_to_free = op->upcall.req.io.buf_index;
 237         memset(&op->upcall, 0, sizeof(op->upcall));
 238         memset(&op->downcall, 0, sizeof(op->downcall));
 239         op->upcall.type = ORANGEFS_VFS_OP_CANCEL;
 240         op->upcall.req.cancel.op_tag = tag;
 241         op->downcall.type = ORANGEFS_VFS_OP_INVALID;
 242         op->downcall.status = -1;
 243         orangefs_new_tag(op);
 244
 245         spin_lock(&orangefs_request_list_lock);
 246         /* orangefs_request_list_lock is enough of a barrier here */
 247         if (!__is_daemon_in_service()) {
 248                 spin_unlock(&orangefs_request_list_lock);
 249                 return false;
 250         }
 251         spin_lock(&op->lock);
 252         set_op_state_waiting(op);
 253         list_add(&op->list, &orangefs_request_list);
 254         spin_unlock(&op->lock);
 255         spin_unlock(&orangefs_request_list_lock);
 256
 257         gossip_debug(GOSSIP_UTILS_DEBUG,
 258                      "Attempting ORANGEFS operation cancellation of tag %llu\n",
 259                      llu(tag));
 260         return true;
 261 }
 262
 263 static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
 264 {
 265         /*
 266          * handle interrupted cases depending on what state we were in when
 267          * the interruption is detected.  there is a coarse grained lock
 268          * across the operation.
 269          *
 270          * Called with op->lock held.
 271          */
 272         op->op_state |= OP_VFS_STATE_GIVEN_UP;
 273
 274         if (op_state_waiting(op)) {
 275                 /*
 276                  * upcall hasn't been read; remove op from upcall request
 277                  * list.
 278                  */
 279                 spin_unlock(&op->lock);
 280                 spin_lock(&orangefs_request_list_lock);
 281                 list_del(&op->list);
 282                 spin_unlock(&orangefs_request_list_lock);
 283                 gossip_debug(GOSSIP_WAIT_DEBUG,
 284                              "Interrupted: Removed op %p from request_list\n",
 285                              op);
 286         } else if (op_state_in_progress(op)) {
 287                 /* op must be removed from the in progress htable */
 288                 spin_unlock(&op->lock);
 289                 spin_lock(&htable_ops_in_progress_lock);
 290                 list_del(&op->list);
 291                 spin_unlock(&htable_ops_in_progress_lock);
 292                 gossip_debug(GOSSIP_WAIT_DEBUG,
 293                              "Interrupted: Removed op %p"
 294                              " from htable_ops_in_progress\n",
 295                              op);
 296         } else if (!op_state_serviced(op)) {
 297                 spin_unlock(&op->lock);
 298                 gossip_err("interrupted operation is in a weird state 0x%x\n",
 299                            op->op_state);
 300         } else {
 301                 /*
 302                  * It is not intended for execution to flow here,
 303                  * but having this unlock here makes sparse happy.
 304                  */
 305                 gossip_err("%s: can't get here.\n", __func__);
 306                 spin_unlock(&op->lock);
 307         }
 308         reinit_completion(&op->waitq);
 309 }
 310
 311 /*
 312  * sleeps on waitqueue waiting for matching downcall.
 313  * if client-core finishes servicing, then we are good to go.
 314  * else if client-core exits, we get woken up here, and retry with a timeout
 315  *
 316  * Post when this call returns to the caller, the specified op will no
 317  * longer be on any list or htable.
 318  *
 319  * Returns 0 on success and -errno on failure
 320  * Errors are:
 321  * EAGAIN in case we want the caller to requeue and try again..
 322  * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
 323  * operation since client-core seems to be exiting too often
 324  * or if we were interrupted.
 325  *
 326  * Returns with op->lock taken.
 327  */
 328 static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op,
 329                                       long timeout,
 330                                       bool interruptible)
 331 {
 332         long n;
 333
 334         if (interruptible)
 335                 n = wait_for_completion_interruptible_timeout(&op->waitq, timeout);
 336         else
 337                 n = wait_for_completion_killable_timeout(&op->waitq, timeout);
 338
 339         spin_lock(&op->lock);
 340
 341         if (op_state_serviced(op))
 342                 return 0;
 343
 344         if (unlikely(n < 0)) {
 345                 gossip_debug(GOSSIP_WAIT_DEBUG,
 346                              "*** %s:"
 347                              " operation interrupted by a signal (tag "
 348                              "%llu, op %p)\n",
 349                              __func__,
 350                              llu(op->tag),
 351                              op);
 352                 return -EINTR;
 353         }
 354         if (op_state_purged(op)) {
 355                 gossip_debug(GOSSIP_WAIT_DEBUG,
 356                              "*** %s:"
 357                              " operation purged (tag "
 358                              "%llu, %p, att %d)\n",
 359                              __func__,
 360                              llu(op->tag),
 361                              op,
 362                              op->attempts);
 363                 return (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
 364                          -EAGAIN :
 365                          -EIO;
 366         }
 367         /* must have timed out, then... */
 368         gossip_debug(GOSSIP_WAIT_DEBUG,
 369                      "*** %s:"
 370                      " operation timed out (tag"
 371                      " %llu, %p, att %d)\n",
 372                      __func__,
 373                      llu(op->tag),
 374                      op,
 375                      op->attempts);
 376         return -ETIMEDOUT;
 377 }