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                 gossip_debug(GOSSIP_DEV_DEBUG,
  41                              "%s: op:%s: op_state:%d: process:%s:\n",
  42                              __func__,
  43                              get_opname_string(op),
  44                              op->op_state,
  45                              current->comm);
  46         }
  47         spin_unlock(&orangefs_request_list_lock);
  48 }
  49
  50 /*
  51  * submits a ORANGEFS operation and waits for it to complete
  52  *
  53  * Note op->downcall.status will contain the status of the operation (in
  54  * errno format), whether provided by pvfs2-client or a result of failure to
  55  * service the operation.  If the caller wishes to distinguish, then
  56  * op->state can be checked to see if it was serviced or not.
  57  *
  58  * Returns contents of op->downcall.status for convenience
  59  */
  60 int service_operation(struct orangefs_kernel_op_s *op,
  61                       const char *op_name,
  62                       int flags)
  63 {
  64         long timeout = MAX_SCHEDULE_TIMEOUT;
  65         int ret = 0;
  66
  67         DEFINE_WAIT(wait_entry);
  68
  69         op->upcall.tgid = current->tgid;
  70         op->upcall.pid = current->pid;
  71
  72 retry_servicing:
  73         op->downcall.status = 0;
  74         gossip_debug(GOSSIP_WAIT_DEBUG,
  75                      "%s: %s op:%p: process:%s: pid:%d:\n",
  76                      __func__,
  77                      op_name,
  78                      op,
  79                      current->comm,
  80                      current->pid);
  81
  82         /*
  83          * If ORANGEFS_OP_NO_MUTEX was set in flags, we need to avoid
  84          * acquiring the request_mutex because we're servicing a
  85          * high priority remount operation and the request_mutex is
  86          * already taken.
  87          */
  88         if (!(flags & ORANGEFS_OP_NO_MUTEX)) {
  89                 if (flags & ORANGEFS_OP_INTERRUPTIBLE)
  90                         ret = mutex_lock_interruptible(&orangefs_request_mutex);
  91                 else
  92                         ret = mutex_lock_killable(&orangefs_request_mutex);
  93                 /*
  94                  * check to see if we were interrupted while waiting for
  95                  * mutex
  96                  */
  97                 if (ret < 0) {
  98                         op->downcall.status = ret;
  99                         gossip_debug(GOSSIP_WAIT_DEBUG,
 100                                      "%s: service_operation interrupted.\n",
 101                                      __func__);
 102                         return ret;
 103                 }
 104         }
 105
 106         /* queue up the operation */
 107         spin_lock(&orangefs_request_list_lock);
 108         spin_lock(&op->lock);
 109         set_op_state_waiting(op);
 110         gossip_debug(GOSSIP_DEV_DEBUG,
 111                      "%s: op:%s: op_state:%d: process:%s:\n",
 112                      __func__,
 113                      get_opname_string(op),
 114                      op->op_state,
 115                      current->comm);
 116         /* add high priority remount op to the front of the line. */
 117         if (flags & ORANGEFS_OP_PRIORITY)
 118                 list_add(&op->list, &orangefs_request_list);
 119         else
 120                 list_add_tail(&op->list, &orangefs_request_list);
 121         spin_unlock(&op->lock);
 122         wake_up_interruptible(&orangefs_request_list_waitq);
 123         if (!__is_daemon_in_service()) {
 124                 gossip_debug(GOSSIP_WAIT_DEBUG,
 125                              "%s:client core is NOT in service.\n",
 126                              __func__);
 127                 /*
 128                  * Don't wait for the userspace component to return if
 129                  * the filesystem is being umounted anyway.
 130                  */
 131                 if (op->upcall.type == ORANGEFS_VFS_OP_FS_UMOUNT)
 132                         timeout = 0;
 133                 else
 134                         timeout = op_timeout_secs * HZ;
 135         }
 136         spin_unlock(&orangefs_request_list_lock);
 137
 138         if (!(flags & ORANGEFS_OP_NO_MUTEX))
 139                 mutex_unlock(&orangefs_request_mutex);
 140
 141         ret = wait_for_matching_downcall(op, timeout,
 142                                          flags & ORANGEFS_OP_INTERRUPTIBLE);
 143
 144         gossip_debug(GOSSIP_WAIT_DEBUG,
 145                      "%s: wait_for_matching_downcall returned %d for %p\n",
 146                      __func__,
 147                      ret,
 148                      op);
 149
 150         /* got matching downcall; make sure status is in errno format */
 151         if (!ret) {
 152                 spin_unlock(&op->lock);
 153                 op->downcall.status =
 154                     orangefs_normalize_to_errno(op->downcall.status);
 155                 ret = op->downcall.status;
 156                 goto out;
 157         }
 158
 159         /* failed to get matching downcall */
 160         if (ret == -ETIMEDOUT) {
 161                 gossip_err("%s: %s -- wait timed out; aborting attempt.\n",
 162                            __func__,
 163                            op_name);
 164         }
 165
 166         /*
 167          * remove a waiting op from the request list or
 168          * remove an in-progress op from the in-progress list.
 169          */
 170         orangefs_clean_up_interrupted_operation(op);
 171
 172         op->downcall.status = ret;
 173         /* retry if operation has not been serviced and if requested */
 174         if (ret == -EAGAIN) {
 175                 op->attempts++;
 176                 timeout = op_timeout_secs * HZ;
 177                 gossip_debug(GOSSIP_WAIT_DEBUG,
 178                              "orangefs: tag %llu (%s)"
 179                              " -- operation to be retried (%d attempt)\n",
 180                              llu(op->tag),
 181                              op_name,
 182                              op->attempts);
 183
 184                 /*
 185                  * io ops (ops that use the shared memory buffer) have
 186                  * to be returned to their caller for a retry. Other ops
 187                  * can just be recycled here.
 188                  */
 189                 if (!op->uses_shared_memory)
 190                         goto retry_servicing;
 191         }
 192
 193 out:
 194         gossip_debug(GOSSIP_WAIT_DEBUG,
 195                      "%s: %s returning: %d for %p.\n",
 196                      __func__,
 197                      op_name,
 198                      ret,
 199                      op);
 200         return ret;
 201 }
 202
 203 /* This can get called on an I/O op if it had a bad service_operation. */
 204 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op)
 205 {
 206         u64 tag = op->tag;
 207         if (!op_state_in_progress(op))
 208                 return false;
 209
 210         op->slot_to_free = op->upcall.req.io.buf_index;
 211         memset(&op->upcall, 0, sizeof(op->upcall));
 212         memset(&op->downcall, 0, sizeof(op->downcall));
 213         op->upcall.type = ORANGEFS_VFS_OP_CANCEL;
 214         op->upcall.req.cancel.op_tag = tag;
 215         op->downcall.type = ORANGEFS_VFS_OP_INVALID;
 216         op->downcall.status = -1;
 217         orangefs_new_tag(op);
 218
 219         spin_lock(&orangefs_request_list_lock);
 220         /* orangefs_request_list_lock is enough of a barrier here */
 221         if (!__is_daemon_in_service()) {
 222                 spin_unlock(&orangefs_request_list_lock);
 223                 return false;
 224         }
 225         spin_lock(&op->lock);
 226         set_op_state_waiting(op);
 227         gossip_debug(GOSSIP_DEV_DEBUG,
 228                      "%s: op:%s: op_state:%d: process:%s:\n",
 229                      __func__,
 230                      get_opname_string(op),
 231                      op->op_state,
 232                      current->comm);
 233         list_add(&op->list, &orangefs_request_list);
 234         spin_unlock(&op->lock);
 235         spin_unlock(&orangefs_request_list_lock);
 236
 237         gossip_debug(GOSSIP_WAIT_DEBUG,
 238                      "Attempting ORANGEFS operation cancellation of tag %llu\n",
 239                      llu(tag));
 240         return true;
 241 }
 242
 243 /*
 244  * Change an op to the "given up" state and remove it from its list.
 245  */
 246 static void
 247         orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
 248 {
 249         /*
 250          * handle interrupted cases depending on what state we were in when
 251          * the interruption is detected.
 252          *
 253          * Called with op->lock held.
 254          */
 255
 256         /*
 257          * List manipulation code elsewhere will ignore ops that
 258          * have been given up upon.
 259          */
 260         op->op_state |= OP_VFS_STATE_GIVEN_UP;
 261
 262         if (list_empty(&op->list)) {
 263                 /* caught copying to/from daemon */
 264                 BUG_ON(op_state_serviced(op));
 265                 spin_unlock(&op->lock);
 266                 wait_for_completion(&op->waitq);
 267         } else if (op_state_waiting(op)) {
 268                 /*
 269                  * upcall hasn't been read; remove op from upcall request
 270                  * list.
 271                  */
 272                 spin_unlock(&op->lock);
 273                 spin_lock(&orangefs_request_list_lock);
 274                 list_del_init(&op->list);
 275                 spin_unlock(&orangefs_request_list_lock);
 276                 gossip_debug(GOSSIP_WAIT_DEBUG,
 277                              "Interrupted: Removed op %p from request_list\n",
 278                              op);
 279         } else if (op_state_in_progress(op)) {
 280                 /* op must be removed from the in progress htable */
 281                 spin_unlock(&op->lock);
 282                 spin_lock(&orangefs_htable_ops_in_progress_lock);
 283                 list_del_init(&op->list);
 284                 spin_unlock(&orangefs_htable_ops_in_progress_lock);
 285                 gossip_debug(GOSSIP_WAIT_DEBUG,
 286                              "Interrupted: Removed op %p"
 287                              " from htable_ops_in_progress\n",
 288                              op);
 289         } else {
 290                 spin_unlock(&op->lock);
 291                 gossip_err("interrupted operation is in a weird state 0x%x\n",
 292                            op->op_state);
 293         }
 294         reinit_completion(&op->waitq);
 295 }
 296
 297 /*
 298  * Sleeps on waitqueue waiting for matching downcall.
 299  * If client-core finishes servicing, then we are good to go.
 300  * else if client-core exits, we get woken up here, and retry with a timeout
 301  *
 302  * When this call returns to the caller, the specified op will no
 303  * longer be in either the in_progress hash table or on the request list.
 304  *
 305  * Returns 0 on success and -errno on failure
 306  * Errors are:
 307  * EAGAIN in case we want the caller to requeue and try again..
 308  * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
 309  * operation since client-core seems to be exiting too often
 310  * or if we were interrupted.
 311  *
 312  * Returns with op->lock taken.
 313  */
 314 static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op,
 315                                       long timeout,
 316                                       bool interruptible)
 317 {
 318         long n;
 319
 320         /*
 321          * There's a "schedule_timeout" inside of these wait
 322          * primitives, during which the op is out of the hands of the
 323          * user process that needs something done and is being
 324          * manipulated by the client-core process.
 325          */
 326         if (interruptible)
 327                 n = wait_for_completion_interruptible_timeout(&op->waitq,
 328                                                               timeout);
 329         else
 330                 n = wait_for_completion_killable_timeout(&op->waitq, timeout);
 331
 332         spin_lock(&op->lock);
 333
 334         if (op_state_serviced(op))
 335                 return 0;
 336
 337         if (unlikely(n < 0)) {
 338                 gossip_debug(GOSSIP_WAIT_DEBUG,
 339                              "%s: operation interrupted, tag %llu, %p\n",
 340                              __func__,
 341                              llu(op->tag),
 342                              op);
 343                 return -EINTR;
 344         }
 345         if (op_state_purged(op)) {
 346                 gossip_debug(GOSSIP_WAIT_DEBUG,
 347                              "%s: operation purged, tag %llu, %p, %d\n",
 348                              __func__,
 349                              llu(op->tag),
 350                              op,
 351                              op->attempts);
 352                 return (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
 353                          -EAGAIN :
 354                          -EIO;
 355         }
 356         /* must have timed out, then... */
 357         gossip_debug(GOSSIP_WAIT_DEBUG,
 358                      "%s: operation timed out, tag %llu, %p, %d)\n",
 359                      __func__,
 360                      llu(op->tag),
 361                      op,
 362                      op->attempts);
 363         return -ETIMEDOUT;
 364 }