#include "drbd_req.h"
+static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size);
+
/* Update disk stats at start of I/O request */
static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
{
const int rw = bio_data_dir(bio);
int cpu;
cpu = part_stat_lock();
+ part_round_stats(cpu, &mdev->vdisk->part0);
part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
(void) cpu; /* The macro invocations above want the cpu argument, I do not like
INIT_LIST_HEAD(&req->tl_requests);
INIT_LIST_HEAD(&req->w.list);
+ /* one reference to be put by __drbd_make_request */
+ atomic_set(&req->completion_ref, 1);
+ /* one kref as long as completion_ref > 0 */
+ kref_init(&req->kref);
return req;
}
-static void drbd_req_free(struct drbd_request *req)
-{
- mempool_free(req, drbd_request_mempool);
-}
-
-/* rw is bio_data_dir(), only READ or WRITE */
-static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
+void drbd_req_destroy(struct kref *kref)
{
- const unsigned long s = req->rq_state;
+ struct drbd_request *req = container_of(kref, struct drbd_request, kref);
+ struct drbd_conf *mdev = req->w.mdev;
+ const unsigned s = req->rq_state;
+
+ if ((req->master_bio && !(s & RQ_POSTPONED)) ||
+ atomic_read(&req->completion_ref) ||
+ (s & RQ_LOCAL_PENDING) ||
+ ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
+ dev_err(DEV, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
+ s, atomic_read(&req->completion_ref));
+ return;
+ }
/* remove it from the transfer log.
* well, only if it had been there in the first
* and never sent), it should still be "empty" as
* initialized in drbd_req_new(), so we can list_del() it
* here unconditionally */
- list_del(&req->tl_requests);
+ list_del_init(&req->tl_requests);
/* if it was a write, we may have to set the corresponding
* bit(s) out-of-sync first. If it had a local part, we need to
* release the reference to the activity log. */
- if (rw == WRITE) {
+ if (s & RQ_WRITE) {
/* Set out-of-sync unless both OK flags are set
* (local only or remote failed).
* Other places where we set out-of-sync:
* READ with local io-error */
- if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
- drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
- if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
- drbd_set_in_sync(mdev, req->i.sector, req->i.size);
+ /* There is a special case:
+ * we may notice late that IO was suspended,
+ * and postpone, or schedule for retry, a write,
+ * before it even was submitted or sent.
+ * In that case we do not want to touch the bitmap at all.
+ */
+ if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
+ if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
+ drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
+
+ if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
+ drbd_set_in_sync(mdev, req->i.sector, req->i.size);
+ }
/* one might be tempted to move the drbd_al_complete_io
* to the local io completion callback drbd_request_endio.
* but after the extent has been dropped from the al,
* we would forget to resync the corresponding extent.
*/
- if (s & RQ_LOCAL_MASK) {
+ if (s & RQ_IN_ACT_LOG) {
if (get_ldev_if_state(mdev, D_FAILED)) {
- if (s & RQ_IN_ACT_LOG)
- drbd_al_complete_io(mdev, &req->i);
+ drbd_al_complete_io(mdev, &req->i);
put_ldev(mdev);
} else if (__ratelimit(&drbd_ratelimit_state)) {
dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
}
}
- drbd_req_free(req);
+ mempool_free(req, drbd_request_mempool);
}
-static void queue_barrier(struct drbd_conf *mdev)
-{
- struct drbd_tl_epoch *b;
-
- /* We are within the req_lock. Once we queued the barrier for sending,
- * we set the CREATE_BARRIER bit. It is cleared as soon as a new
- * barrier/epoch object is added. This is the only place this bit is
- * set. It indicates that the barrier for this epoch is already queued,
- * and no new epoch has been created yet. */
- if (test_bit(CREATE_BARRIER, &mdev->flags))
- return;
-
- b = mdev->tconn->newest_tle;
- b->w.cb = w_send_barrier;
- b->w.mdev = mdev;
- /* inc_ap_pending done here, so we won't
- * get imbalanced on connection loss.
- * dec_ap_pending will be done in got_BarrierAck
- * or (on connection loss) in tl_clear. */
- inc_ap_pending(mdev);
- drbd_queue_work(&mdev->tconn->data.work, &b->w);
- set_bit(CREATE_BARRIER, &mdev->flags);
+static void wake_all_senders(struct drbd_tconn *tconn) {
+ wake_up(&tconn->sender_work.q_wait);
}
-static void _about_to_complete_local_write(struct drbd_conf *mdev,
- struct drbd_request *req)
+/* must hold resource->req_lock */
+static void start_new_tl_epoch(struct drbd_tconn *tconn)
{
- const unsigned long s = req->rq_state;
+ /* no point closing an epoch, if it is empty, anyways. */
+ if (tconn->current_tle_writes == 0)
+ return;
- /* Before we can signal completion to the upper layers,
- * we may need to close the current epoch.
- * We can skip this, if this request has not even been sent, because we
- * did not have a fully established connection yet/anymore, during
- * bitmap exchange, or while we are C_AHEAD due to congestion policy.
- */
- if (mdev->state.conn >= C_CONNECTED &&
- (s & RQ_NET_SENT) != 0 &&
- req->epoch == mdev->tconn->newest_tle->br_number)
- queue_barrier(mdev);
+ tconn->current_tle_writes = 0;
+ atomic_inc(&tconn->current_tle_nr);
+ wake_all_senders(tconn);
}
void complete_master_bio(struct drbd_conf *mdev,
* if it has already been completed, or cannot be completed yet.
* If m->bio is set, the error status to be returned is placed in m->error.
*/
-void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
+static
+void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
{
- const unsigned long s = req->rq_state;
+ const unsigned s = req->rq_state;
struct drbd_conf *mdev = req->w.mdev;
- int rw = req->rq_state & RQ_WRITE ? WRITE : READ;
+ int rw;
+ int error, ok;
/* we must not complete the master bio, while it is
* still being processed by _drbd_send_zc_bio (drbd_send_dblock)
* the receiver,
* the bio_endio completion callbacks.
*/
- if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
+ if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
+ (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
+ (s & RQ_COMPLETION_SUSP)) {
+ dev_err(DEV, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
return;
- if (req->i.waiting) {
- /* Retry all conflicting peer requests. */
- wake_up(&mdev->misc_wait);
}
- if (s & RQ_NET_QUEUED)
- return;
- if (s & RQ_NET_PENDING)
+
+ if (!req->master_bio) {
+ dev_err(DEV, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
return;
+ }
- if (req->master_bio) {
- /* this is DATA_RECEIVED (remote read)
- * or protocol C P_WRITE_ACK
- * or protocol B P_RECV_ACK
- * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
- * or canceled or failed,
- * or killed from the transfer log due to connection loss.
- */
+ rw = bio_rw(req->master_bio);
+
+ /*
+ * figure out whether to report success or failure.
+ *
+ * report success when at least one of the operations succeeded.
+ * or, to put the other way,
+ * only report failure, when both operations failed.
+ *
+ * what to do about the failures is handled elsewhere.
+ * what we need to do here is just: complete the master_bio.
+ *
+ * local completion error, if any, has been stored as ERR_PTR
+ * in private_bio within drbd_request_endio.
+ */
+ ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
+ error = PTR_ERR(req->private_bio);
+
+ /* remove the request from the conflict detection
+ * respective block_id verification hash */
+ if (!drbd_interval_empty(&req->i)) {
+ struct rb_root *root;
- /*
- * figure out whether to report success or failure.
- *
- * report success when at least one of the operations succeeded.
- * or, to put the other way,
- * only report failure, when both operations failed.
- *
- * what to do about the failures is handled elsewhere.
- * what we need to do here is just: complete the master_bio.
- *
- * local completion error, if any, has been stored as ERR_PTR
- * in private_bio within drbd_request_endio.
- */
- int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
- int error = PTR_ERR(req->private_bio);
-
- /* remove the request from the conflict detection
- * respective block_id verification hash */
- if (!drbd_interval_empty(&req->i)) {
- struct rb_root *root;
-
- if (rw == WRITE)
- root = &mdev->write_requests;
- else
- root = &mdev->read_requests;
- drbd_remove_request_interval(root, req);
- } else if (!(s & RQ_POSTPONED))
- D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
-
- /* for writes we need to do some extra housekeeping */
if (rw == WRITE)
- _about_to_complete_local_write(mdev, req);
+ root = &mdev->write_requests;
+ else
+ root = &mdev->read_requests;
+ drbd_remove_request_interval(root, req);
+ } else if (!(s & RQ_POSTPONED))
+ D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
- /* Update disk stats */
- _drbd_end_io_acct(mdev, req);
+ /* Before we can signal completion to the upper layers,
+ * we may need to close the current transfer log epoch.
+ * We are within the request lock, so we can simply compare
+ * the request epoch number with the current transfer log
+ * epoch number. If they match, increase the current_tle_nr,
+ * and reset the transfer log epoch write_cnt.
+ */
+ if (rw == WRITE &&
+ req->epoch == atomic_read(&mdev->tconn->current_tle_nr))
+ start_new_tl_epoch(mdev->tconn);
- if (!(s & RQ_POSTPONED)) {
- m->error = ok ? 0 : (error ?: -EIO);
- m->bio = req->master_bio;
- }
+ /* Update disk stats */
+ _drbd_end_io_acct(mdev, req);
+
+ /* If READ failed,
+ * have it be pushed back to the retry work queue,
+ * so it will re-enter __drbd_make_request(),
+ * and be re-assigned to a suitable local or remote path,
+ * or failed if we do not have access to good data anymore.
+ *
+ * Unless it was failed early by __drbd_make_request(),
+ * because no path was available, in which case
+ * it was not even added to the transfer_log.
+ *
+ * READA may fail, and will not be retried.
+ *
+ * WRITE should have used all available paths already.
+ */
+ if (!ok && rw == READ && !list_empty(&req->tl_requests))
+ req->rq_state |= RQ_POSTPONED;
+
+ if (!(req->rq_state & RQ_POSTPONED)) {
+ m->error = ok ? 0 : (error ?: -EIO);
+ m->bio = req->master_bio;
req->master_bio = NULL;
}
+}
- if (s & RQ_LOCAL_PENDING)
- return;
+static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
+{
+ struct drbd_conf *mdev = req->w.mdev;
+ D_ASSERT(m || (req->rq_state & RQ_POSTPONED));
+
+ if (!atomic_sub_and_test(put, &req->completion_ref))
+ return 0;
- if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
- /* this is disconnected (local only) operation,
- * or protocol A, B, or C P_BARRIER_ACK,
- * or killed from the transfer log due to connection loss. */
- _req_is_done(mdev, req, rw);
+ drbd_req_complete(req, m);
+
+ if (req->rq_state & RQ_POSTPONED) {
+ /* don't destroy the req object just yet,
+ * but queue it for retry */
+ drbd_restart_request(req);
+ return 0;
}
- /* else: network part and not DONE yet. that is
- * protocol A, B, or C, barrier ack still pending... */
+
+ return 1;
}
-static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
+/* I'd like this to be the only place that manipulates
+ * req->completion_ref and req->kref. */
+static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
+ int clear, int set)
{
struct drbd_conf *mdev = req->w.mdev;
+ unsigned s = req->rq_state;
+ int c_put = 0;
+ int k_put = 0;
+
+ if (drbd_suspended(mdev) && !((s | clear) & RQ_COMPLETION_SUSP))
+ set |= RQ_COMPLETION_SUSP;
+
+ /* apply */
+
+ req->rq_state &= ~clear;
+ req->rq_state |= set;
+
+ /* no change? */
+ if (req->rq_state == s)
+ return;
+
+ /* intent: get references */
+
+ if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
+ atomic_inc(&req->completion_ref);
+
+ if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
+ inc_ap_pending(mdev);
+ atomic_inc(&req->completion_ref);
+ }
+
+ if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
+ atomic_inc(&req->completion_ref);
+
+ if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
+ kref_get(&req->kref); /* wait for the DONE */
+
+ if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
+ atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
+
+ if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
+ atomic_inc(&req->completion_ref);
+
+ /* progress: put references */
+
+ if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
+ ++c_put;
+
+ if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
+ D_ASSERT(req->rq_state & RQ_LOCAL_PENDING);
+ /* local completion may still come in later,
+ * we need to keep the req object around. */
+ kref_get(&req->kref);
+ ++c_put;
+ }
+
+ if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
+ if (req->rq_state & RQ_LOCAL_ABORTED)
+ ++k_put;
+ else
+ ++c_put;
+ }
+
+ if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
+ dec_ap_pending(mdev);
+ ++c_put;
+ }
- if (!drbd_suspended(mdev))
- _req_may_be_done(req, m);
+ if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
+ ++c_put;
+
+ if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
+ if (req->rq_state & RQ_NET_SENT)
+ atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
+ ++k_put;
+ }
+
+ /* potentially complete and destroy */
+
+ if (k_put || c_put) {
+ /* Completion does it's own kref_put. If we are going to
+ * kref_sub below, we need req to be still around then. */
+ int at_least = k_put + !!c_put;
+ int refcount = atomic_read(&req->kref.refcount);
+ if (refcount < at_least)
+ dev_err(DEV,
+ "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
+ s, req->rq_state, refcount, at_least);
+ }
+
+ /* If we made progress, retry conflicting peer requests, if any. */
+ if (req->i.waiting)
+ wake_up(&mdev->misc_wait);
+
+ if (c_put)
+ k_put += drbd_req_put_completion_ref(req, m, c_put);
+ if (k_put)
+ kref_sub(&req->kref, k_put, drbd_req_destroy);
+}
+
+static void drbd_report_io_error(struct drbd_conf *mdev, struct drbd_request *req)
+{
+ char b[BDEVNAME_SIZE];
+
+ if (!__ratelimit(&drbd_ratelimit_state))
+ return;
+
+ dev_warn(DEV, "local %s IO error sector %llu+%u on %s\n",
+ (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
+ (unsigned long long)req->i.sector,
+ req->i.size >> 9,
+ bdevname(mdev->ldev->backing_bdev, b));
}
/* obviously this could be coded as many single functions
/* reached via __drbd_make_request
* and from w_read_retry_remote */
D_ASSERT(!(req->rq_state & RQ_NET_MASK));
- req->rq_state |= RQ_NET_PENDING;
rcu_read_lock();
nc = rcu_dereference(mdev->tconn->net_conf);
p = nc->wire_protocol;
req->rq_state |=
p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
- inc_ap_pending(mdev);
+ mod_rq_state(req, m, 0, RQ_NET_PENDING);
break;
case TO_BE_SUBMITTED: /* locally */
/* reached via __drbd_make_request */
D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
- req->rq_state |= RQ_LOCAL_PENDING;
+ mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
break;
case COMPLETED_OK:
else
mdev->read_cnt += req->i.size >> 9;
- req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
- req->rq_state &= ~RQ_LOCAL_PENDING;
-
- _req_may_be_done_not_susp(req, m);
- put_ldev(mdev);
+ mod_rq_state(req, m, RQ_LOCAL_PENDING,
+ RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
break;
case ABORT_DISK_IO:
- req->rq_state |= RQ_LOCAL_ABORTED;
- if (req->rq_state & RQ_WRITE)
- _req_may_be_done_not_susp(req, m);
- else
- goto goto_queue_for_net_read;
+ mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
break;
case WRITE_COMPLETED_WITH_ERROR:
- req->rq_state |= RQ_LOCAL_COMPLETED;
- req->rq_state &= ~RQ_LOCAL_PENDING;
-
- __drbd_chk_io_error(mdev, false);
- _req_may_be_done_not_susp(req, m);
- put_ldev(mdev);
- break;
-
- case READ_AHEAD_COMPLETED_WITH_ERROR:
- /* it is legal to fail READA */
- req->rq_state |= RQ_LOCAL_COMPLETED;
- req->rq_state &= ~RQ_LOCAL_PENDING;
- _req_may_be_done_not_susp(req, m);
- put_ldev(mdev);
+ drbd_report_io_error(mdev, req);
+ __drbd_chk_io_error(mdev, DRBD_WRITE_ERROR);
+ mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
break;
case READ_COMPLETED_WITH_ERROR:
drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
-
- req->rq_state |= RQ_LOCAL_COMPLETED;
- req->rq_state &= ~RQ_LOCAL_PENDING;
-
- D_ASSERT(!(req->rq_state & RQ_NET_MASK));
-
- __drbd_chk_io_error(mdev, false);
- put_ldev(mdev);
-
- goto_queue_for_net_read:
-
- /* no point in retrying if there is no good remote data,
- * or we have no connection. */
- if (mdev->state.pdsk != D_UP_TO_DATE) {
- _req_may_be_done_not_susp(req, m);
- break;
- }
-
- /* _req_mod(req,TO_BE_SENT); oops, recursion... */
- req->rq_state |= RQ_NET_PENDING;
- inc_ap_pending(mdev);
- /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
+ drbd_report_io_error(mdev, req);
+ __drbd_chk_io_error(mdev, DRBD_READ_ERROR);
+ /* fall through. */
+ case READ_AHEAD_COMPLETED_WITH_ERROR:
+ /* it is legal to fail READA, no __drbd_chk_io_error in that case. */
+ mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
+ break;
case QUEUE_FOR_NET_READ:
/* READ or READA, and
/* from __drbd_make_request
* or from bio_endio during read io-error recovery */
- /* so we can verify the handle in the answer packet
- * corresponding hlist_del is in _req_may_be_done() */
+ /* So we can verify the handle in the answer packet.
+ * Corresponding drbd_remove_request_interval is in
+ * drbd_req_complete() */
D_ASSERT(drbd_interval_empty(&req->i));
drbd_insert_interval(&mdev->read_requests, &req->i);
set_bit(UNPLUG_REMOTE, &mdev->flags);
D_ASSERT(req->rq_state & RQ_NET_PENDING);
- req->rq_state |= RQ_NET_QUEUED;
- req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
- ? w_read_retry_remote
- : w_send_read_req;
- drbd_queue_work(&mdev->tconn->data.work, &req->w);
+ D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0);
+ mod_rq_state(req, m, 0, RQ_NET_QUEUED);
+ req->w.cb = w_send_read_req;
+ drbd_queue_work(&mdev->tconn->sender_work, &req->w);
break;
case QUEUE_FOR_NET_WRITE:
/* assert something? */
/* from __drbd_make_request only */
- /* corresponding hlist_del is in _req_may_be_done() */
+ /* Corresponding drbd_remove_request_interval is in
+ * drbd_req_complete() */
D_ASSERT(drbd_interval_empty(&req->i));
drbd_insert_interval(&mdev->write_requests, &req->i);
* hurting performance. */
set_bit(UNPLUG_REMOTE, &mdev->flags);
- /* see __drbd_make_request,
- * just after it grabs the req_lock */
- D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
-
- req->epoch = mdev->tconn->newest_tle->br_number;
-
- /* increment size of current epoch */
- mdev->tconn->newest_tle->n_writes++;
-
/* queue work item to send data */
D_ASSERT(req->rq_state & RQ_NET_PENDING);
- req->rq_state |= RQ_NET_QUEUED;
+ mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
req->w.cb = w_send_dblock;
- drbd_queue_work(&mdev->tconn->data.work, &req->w);
+ drbd_queue_work(&mdev->tconn->sender_work, &req->w);
/* close the epoch, in case it outgrew the limit */
rcu_read_lock();
nc = rcu_dereference(mdev->tconn->net_conf);
p = nc->max_epoch_size;
rcu_read_unlock();
- if (mdev->tconn->newest_tle->n_writes >= p)
- queue_barrier(mdev);
+ if (mdev->tconn->current_tle_writes >= p)
+ start_new_tl_epoch(mdev->tconn);
break;
case QUEUE_FOR_SEND_OOS:
- req->rq_state |= RQ_NET_QUEUED;
+ mod_rq_state(req, m, 0, RQ_NET_QUEUED);
req->w.cb = w_send_out_of_sync;
- drbd_queue_work(&mdev->tconn->data.work, &req->w);
+ drbd_queue_work(&mdev->tconn->sender_work, &req->w);
break;
- case OOS_HANDED_TO_NETWORK:
- /* actually the same */
+ case READ_RETRY_REMOTE_CANCELED:
case SEND_CANCELED:
- /* treat it the same */
case SEND_FAILED:
/* real cleanup will be done from tl_clear. just update flags
* so it is no longer marked as on the worker queue */
- req->rq_state &= ~RQ_NET_QUEUED;
- /* if we did it right, tl_clear should be scheduled only after
- * this, so this should not be necessary! */
- _req_may_be_done_not_susp(req, m);
+ mod_rq_state(req, m, RQ_NET_QUEUED, 0);
break;
case HANDED_OVER_TO_NETWORK:
/* assert something? */
- if (bio_data_dir(req->master_bio) == WRITE)
- atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
-
if (bio_data_dir(req->master_bio) == WRITE &&
!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
/* this is what is dangerous about protocol A:
* pretend it was successfully written on the peer. */
- if (req->rq_state & RQ_NET_PENDING) {
- dec_ap_pending(mdev);
- req->rq_state &= ~RQ_NET_PENDING;
- req->rq_state |= RQ_NET_OK;
- } /* else: neg-ack was faster... */
+ if (req->rq_state & RQ_NET_PENDING)
+ mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
+ /* else: neg-ack was faster... */
/* it is still not yet RQ_NET_DONE until the
* corresponding epoch barrier got acked as well,
* so we know what to dirty on connection loss */
}
- req->rq_state &= ~RQ_NET_QUEUED;
- req->rq_state |= RQ_NET_SENT;
- /* because _drbd_send_zc_bio could sleep, and may want to
- * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
- * "COMPLETED_OK" events came in, once we return from
- * _drbd_send_zc_bio (drbd_send_dblock), we have to check
- * whether it is done already, and end it. */
- _req_may_be_done_not_susp(req, m);
+ mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
+ break;
+
+ case OOS_HANDED_TO_NETWORK:
+ /* Was not set PENDING, no longer QUEUED, so is now DONE
+ * as far as this connection is concerned. */
+ mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
break;
- case READ_RETRY_REMOTE_CANCELED:
- req->rq_state &= ~RQ_NET_QUEUED;
- /* fall through, in case we raced with drbd_disconnect */
case CONNECTION_LOST_WHILE_PENDING:
/* transfer log cleanup after connection loss */
- /* assert something? */
- if (req->rq_state & RQ_NET_PENDING)
- dec_ap_pending(mdev);
- req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
- req->rq_state |= RQ_NET_DONE;
- if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
- atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
+ mod_rq_state(req, m,
+ RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
+ RQ_NET_DONE);
+ break;
- /* if it is still queued, we may not complete it here.
- * it will be canceled soon. */
- if (!(req->rq_state & RQ_NET_QUEUED))
- _req_may_be_done(req, m); /* Allowed while state.susp */
+ case CONFLICT_RESOLVED:
+ /* for superseded conflicting writes of multiple primaries,
+ * there is no need to keep anything in the tl, potential
+ * node crashes are covered by the activity log.
+ *
+ * If this request had been marked as RQ_POSTPONED before,
+ * it will actually not be completed, but "restarted",
+ * resubmitted from the retry worker context. */
+ D_ASSERT(req->rq_state & RQ_NET_PENDING);
+ D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
+ mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
break;
case WRITE_ACKED_BY_PEER_AND_SIS:
req->rq_state |= RQ_NET_SIS;
- case DISCARD_WRITE:
- /* for discarded conflicting writes of multiple primaries,
- * there is no need to keep anything in the tl, potential
- * node crashes are covered by the activity log. */
- req->rq_state |= RQ_NET_DONE;
- /* fall through */
case WRITE_ACKED_BY_PEER:
D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
/* protocol C; successfully written on peer.
- * Nothing to do here.
+ * Nothing more to do here.
* We want to keep the tl in place for all protocols, to cater
- * for volatile write-back caches on lower level devices.
- *
- * A barrier request is expected to have forced all prior
- * requests onto stable storage, so completion of a barrier
- * request could set NET_DONE right here, and not wait for the
- * P_BARRIER_ACK, but that is an unnecessary optimization. */
+ * for volatile write-back caches on lower level devices. */
goto ack_common;
- /* this makes it effectively the same as for: */
case RECV_ACKED_BY_PEER:
D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
/* protocol B; pretends to be successfully written on peer.
* see also notes above in HANDED_OVER_TO_NETWORK about
* protocol != C */
ack_common:
- req->rq_state |= RQ_NET_OK;
D_ASSERT(req->rq_state & RQ_NET_PENDING);
- dec_ap_pending(mdev);
- atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
- req->rq_state &= ~RQ_NET_PENDING;
- _req_may_be_done_not_susp(req, m);
+ mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
break;
case POSTPONE_WRITE:
*/
D_ASSERT(req->rq_state & RQ_NET_PENDING);
req->rq_state |= RQ_POSTPONED;
- _req_may_be_done_not_susp(req, m);
+ if (req->i.waiting)
+ wake_up(&mdev->misc_wait);
+ /* Do not clear RQ_NET_PENDING. This request will make further
+ * progress via restart_conflicting_writes() or
+ * fail_postponed_requests(). Hopefully. */
break;
case NEG_ACKED:
- /* assert something? */
- if (req->rq_state & RQ_NET_PENDING) {
- dec_ap_pending(mdev);
- atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
- }
- req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
-
- req->rq_state |= RQ_NET_DONE;
- _req_may_be_done_not_susp(req, m);
- /* else: done by HANDED_OVER_TO_NETWORK */
+ mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
break;
case FAIL_FROZEN_DISK_IO:
if (!(req->rq_state & RQ_LOCAL_COMPLETED))
break;
-
- _req_may_be_done(req, m); /* Allowed while state.susp */
+ mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
break;
case RESTART_FROZEN_DISK_IO:
if (!(req->rq_state & RQ_LOCAL_COMPLETED))
break;
- req->rq_state &= ~RQ_LOCAL_COMPLETED;
+ mod_rq_state(req, m,
+ RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
+ RQ_LOCAL_PENDING);
rv = MR_READ;
if (bio_data_dir(req->master_bio) == WRITE)
rv = MR_WRITE;
- get_ldev(mdev);
+ get_ldev(mdev); /* always succeeds in this call path */
req->w.cb = w_restart_disk_io;
- drbd_queue_work(&mdev->tconn->data.work, &req->w);
+ drbd_queue_work(&mdev->tconn->sender_work, &req->w);
break;
case RESEND:
+ /* Simply complete (local only) READs. */
+ if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
+ mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
+ break;
+ }
+
/* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
- before the connection loss (B&C only); only P_BARRIER_ACK was missing.
- Trowing them out of the TL here by pretending we got a BARRIER_ACK
- We ensure that the peer was not rebooted */
+ before the connection loss (B&C only); only P_BARRIER_ACK
+ (or the local completion?) was missing when we suspended.
+ Throwing them out of the TL here by pretending we got a BARRIER_ACK.
+ During connection handshake, we ensure that the peer was not rebooted. */
if (!(req->rq_state & RQ_NET_OK)) {
+ /* FIXME could this possibly be a req->w.cb == w_send_out_of_sync?
+ * in that case we must not set RQ_NET_PENDING. */
+
+ mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
if (req->w.cb) {
- drbd_queue_work(&mdev->tconn->data.work, &req->w);
+ drbd_queue_work(&mdev->tconn->sender_work, &req->w);
rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
- }
+ } /* else: FIXME can this happen? */
break;
}
/* else, fall through to BARRIER_ACKED */
case BARRIER_ACKED:
+ /* barrier ack for READ requests does not make sense */
if (!(req->rq_state & RQ_WRITE))
break;
* this is bad, because if the connection is lost now,
* we won't be able to clean them up... */
dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
- list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests);
- }
- if ((req->rq_state & RQ_NET_MASK) != 0) {
- req->rq_state |= RQ_NET_DONE;
- if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK)))
- atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
}
- _req_may_be_done(req, m); /* Allowed while state.susp */
+ /* Allowed to complete requests, even while suspended.
+ * As this is called for all requests within a matching epoch,
+ * we need to filter, and only set RQ_NET_DONE for those that
+ * have actually been on the wire. */
+ mod_rq_state(req, m, RQ_COMPLETION_SUSP,
+ (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
break;
case DATA_RECEIVED:
D_ASSERT(req->rq_state & RQ_NET_PENDING);
- dec_ap_pending(mdev);
- req->rq_state &= ~RQ_NET_PENDING;
- req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
- _req_may_be_done_not_susp(req, m);
+ mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
break;
};
return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
}
+static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector,
+ enum drbd_read_balancing rbm)
+{
+ struct backing_dev_info *bdi;
+ int stripe_shift;
+
+ switch (rbm) {
+ case RB_CONGESTED_REMOTE:
+ bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
+ return bdi_read_congested(bdi);
+ case RB_LEAST_PENDING:
+ return atomic_read(&mdev->local_cnt) >
+ atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
+ case RB_32K_STRIPING: /* stripe_shift = 15 */
+ case RB_64K_STRIPING:
+ case RB_128K_STRIPING:
+ case RB_256K_STRIPING:
+ case RB_512K_STRIPING:
+ case RB_1M_STRIPING: /* stripe_shift = 20 */
+ stripe_shift = (rbm - RB_32K_STRIPING + 15);
+ return (sector >> (stripe_shift - 9)) & 1;
+ case RB_ROUND_ROBIN:
+ return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
+ case RB_PREFER_REMOTE:
+ return true;
+ case RB_PREFER_LOCAL:
+ default:
+ return false;
+ }
+}
+
/*
* complete_conflicting_writes - wait for any conflicting write requests
*
* The write_requests tree contains all active write requests which we
* currently know about. Wait for any requests to complete which conflict with
* the new one.
+ *
+ * Only way out: remove the conflicting intervals from the tree.
*/
-static int complete_conflicting_writes(struct drbd_conf *mdev,
- sector_t sector, int size)
+static void complete_conflicting_writes(struct drbd_request *req)
{
- for(;;) {
- struct drbd_interval *i;
- int err;
+ DEFINE_WAIT(wait);
+ struct drbd_conf *mdev = req->w.mdev;
+ struct drbd_interval *i;
+ sector_t sector = req->i.sector;
+ int size = req->i.size;
+ i = drbd_find_overlap(&mdev->write_requests, sector, size);
+ if (!i)
+ return;
+
+ for (;;) {
+ prepare_to_wait(&mdev->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
i = drbd_find_overlap(&mdev->write_requests, sector, size);
if (!i)
- return 0;
- err = drbd_wait_misc(mdev, i);
- if (err)
- return err;
+ break;
+ /* Indicate to wake up device->misc_wait on progress. */
+ i->waiting = true;
+ spin_unlock_irq(&mdev->tconn->req_lock);
+ schedule();
+ spin_lock_irq(&mdev->tconn->req_lock);
}
+ finish_wait(&mdev->misc_wait, &wait);
}
-int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
+/* called within req_lock and rcu_read_lock() */
+static void maybe_pull_ahead(struct drbd_conf *mdev)
{
+ struct drbd_tconn *tconn = mdev->tconn;
+ struct net_conf *nc;
+ bool congested = false;
+ enum drbd_on_congestion on_congestion;
+
+ nc = rcu_dereference(tconn->net_conf);
+ on_congestion = nc ? nc->on_congestion : OC_BLOCK;
+ if (on_congestion == OC_BLOCK ||
+ tconn->agreed_pro_version < 96)
+ return;
+
+ /* If I don't even have good local storage, we can not reasonably try
+ * to pull ahead of the peer. We also need the local reference to make
+ * sure mdev->act_log is there.
+ */
+ if (!get_ldev_if_state(mdev, D_UP_TO_DATE))
+ return;
+
+ if (nc->cong_fill &&
+ atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
+ dev_info(DEV, "Congestion-fill threshold reached\n");
+ congested = true;
+ }
+
+ if (mdev->act_log->used >= nc->cong_extents) {
+ dev_info(DEV, "Congestion-extents threshold reached\n");
+ congested = true;
+ }
+
+ if (congested) {
+ /* start a new epoch for non-mirrored writes */
+ start_new_tl_epoch(mdev->tconn);
+
+ if (on_congestion == OC_PULL_AHEAD)
+ _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
+ else /*nc->on_congestion == OC_DISCONNECT */
+ _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
+ }
+ put_ldev(mdev);
+}
+
+/* If this returns false, and req->private_bio is still set,
+ * this should be submitted locally.
+ *
+ * If it returns false, but req->private_bio is not set,
+ * we do not have access to good data :(
+ *
+ * Otherwise, this destroys req->private_bio, if any,
+ * and returns true.
+ */
+static bool do_remote_read(struct drbd_request *req)
+{
+ struct drbd_conf *mdev = req->w.mdev;
+ enum drbd_read_balancing rbm;
+
+ if (req->private_bio) {
+ if (!drbd_may_do_local_read(mdev,
+ req->i.sector, req->i.size)) {
+ bio_put(req->private_bio);
+ req->private_bio = NULL;
+ put_ldev(mdev);
+ }
+ }
+
+ if (mdev->state.pdsk != D_UP_TO_DATE)
+ return false;
+
+ if (req->private_bio == NULL)
+ return true;
+
+ /* TODO: improve read balancing decisions, take into account drbd
+ * protocol, pending requests etc. */
+
+ rcu_read_lock();
+ rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
+ rcu_read_unlock();
+
+ if (rbm == RB_PREFER_LOCAL && req->private_bio)
+ return false; /* submit locally */
+
+ if (remote_due_to_read_balancing(mdev, req->i.sector, rbm)) {
+ if (req->private_bio) {
+ bio_put(req->private_bio);
+ req->private_bio = NULL;
+ put_ldev(mdev);
+ }
+ return true;
+ }
+
+ return false;
+}
+
+/* returns number of connections (== 1, for drbd 8.4)
+ * expected to actually write this data,
+ * which does NOT include those that we are L_AHEAD for. */
+static int drbd_process_write_request(struct drbd_request *req)
+{
+ struct drbd_conf *mdev = req->w.mdev;
+ int remote, send_oos;
+
+ rcu_read_lock();
+ remote = drbd_should_do_remote(mdev->state);
+ if (remote) {
+ maybe_pull_ahead(mdev);
+ remote = drbd_should_do_remote(mdev->state);
+ }
+ send_oos = drbd_should_send_out_of_sync(mdev->state);
+ rcu_read_unlock();
+
+ /* Need to replicate writes. Unless it is an empty flush,
+ * which is better mapped to a DRBD P_BARRIER packet,
+ * also for drbd wire protocol compatibility reasons.
+ * If this was a flush, just start a new epoch.
+ * Unless the current epoch was empty anyways, or we are not currently
+ * replicating, in which case there is no point. */
+ if (unlikely(req->i.size == 0)) {
+ /* The only size==0 bios we expect are empty flushes. */
+ D_ASSERT(req->master_bio->bi_rw & REQ_FLUSH);
+ if (remote)
+ start_new_tl_epoch(mdev->tconn);
+ return 0;
+ }
+
+ if (!remote && !send_oos)
+ return 0;
+
+ D_ASSERT(!(remote && send_oos));
+
+ if (remote) {
+ _req_mod(req, TO_BE_SENT);
+ _req_mod(req, QUEUE_FOR_NET_WRITE);
+ } else if (drbd_set_out_of_sync(mdev, req->i.sector, req->i.size))
+ _req_mod(req, QUEUE_FOR_SEND_OOS);
+
+ return remote;
+}
+
+static void
+drbd_submit_req_private_bio(struct drbd_request *req)
+{
+ struct drbd_conf *mdev = req->w.mdev;
+ struct bio *bio = req->private_bio;
const int rw = bio_rw(bio);
- const int size = bio->bi_size;
- const sector_t sector = bio->bi_sector;
- struct drbd_tl_epoch *b = NULL;
+
+ bio->bi_bdev = mdev->ldev->backing_bdev;
+
+ /* State may have changed since we grabbed our reference on the
+ * ->ldev member. Double check, and short-circuit to endio.
+ * In case the last activity log transaction failed to get on
+ * stable storage, and this is a WRITE, we may not even submit
+ * this bio. */
+ if (get_ldev(mdev)) {
+ if (drbd_insert_fault(mdev,
+ rw == WRITE ? DRBD_FAULT_DT_WR
+ : rw == READ ? DRBD_FAULT_DT_RD
+ : DRBD_FAULT_DT_RA))
+ bio_endio(bio, -EIO);
+ else
+ generic_make_request(bio);
+ put_ldev(mdev);
+ } else
+ bio_endio(bio, -EIO);
+}
+
+void __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
+{
+ const int rw = bio_rw(bio);
+ struct bio_and_error m = { NULL, };
struct drbd_request *req;
- struct net_conf *nc;
- int local, remote, send_oos = 0;
- int err;
- int ret = 0;
+ bool no_remote = false;
/* allocate outside of all locks; */
req = drbd_req_new(mdev, bio);
* if user cannot handle io errors, that's not our business. */
dev_err(DEV, "could not kmalloc() req\n");
bio_endio(bio, -ENOMEM);
- return 0;
+ return;
}
req->start_time = start_time;
- local = get_ldev(mdev);
- if (!local) {
- bio_put(req->private_bio); /* or we get a bio leak */
+ if (!get_ldev(mdev)) {
+ bio_put(req->private_bio);
req->private_bio = NULL;
}
- if (rw == WRITE) {
- remote = 1;
- } else {
- /* READ || READA */
- if (local) {
- if (!drbd_may_do_local_read(mdev, sector, size)) {
- /* we could kick the syncer to
- * sync this extent asap, wait for
- * it, then continue locally.
- * Or just issue the request remotely.
- */
- local = 0;
- bio_put(req->private_bio);
- req->private_bio = NULL;
- put_ldev(mdev);
- }
- }
- remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
- }
-
- /* If we have a disk, but a READA request is mapped to remote,
- * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
- * Just fail that READA request right here.
- *
- * THINK: maybe fail all READA when not local?
- * or make this configurable...
- * if network is slow, READA won't do any good.
- */
- if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
- err = -EWOULDBLOCK;
- goto fail_and_free_req;
- }
/* For WRITES going to the local disk, grab a reference on the target
* extent. This waits for any resync activity in the corresponding
* resync extent to finish, and, if necessary, pulls in the target
* extent into the activity log, which involves further disk io because
- * of transactional on-disk meta data updates. */
- if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
+ * of transactional on-disk meta data updates.
+ * Empty flushes don't need to go into the activity log, they can only
+ * flush data for pending writes which are already in there. */
+ if (rw == WRITE && req->private_bio && req->i.size
+ && !test_bit(AL_SUSPENDED, &mdev->flags)) {
req->rq_state |= RQ_IN_ACT_LOG;
drbd_al_begin_io(mdev, &req->i);
}
- remote = remote && drbd_should_do_remote(mdev->state);
- send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
- D_ASSERT(!(remote && send_oos));
-
- if (!(local || remote) && !drbd_suspended(mdev)) {
- if (__ratelimit(&drbd_ratelimit_state))
- dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
- err = -EIO;
- goto fail_free_complete;
- }
-
- /* For WRITE request, we have to make sure that we have an
- * unused_spare_tle, in case we need to start a new epoch.
- * I try to be smart and avoid to pre-allocate always "just in case",
- * but there is a race between testing the bit and pointer outside the
- * spinlock, and grabbing the spinlock.
- * if we lost that race, we retry. */
- if (rw == WRITE && (remote || send_oos) &&
- mdev->tconn->unused_spare_tle == NULL &&
- test_bit(CREATE_BARRIER, &mdev->flags)) {
-allocate_barrier:
- b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
- if (!b) {
- dev_err(DEV, "Failed to alloc barrier.\n");
- err = -ENOMEM;
- goto fail_free_complete;
- }
- }
-
- /* GOOD, everything prepared, grab the spin_lock */
spin_lock_irq(&mdev->tconn->req_lock);
-
if (rw == WRITE) {
- err = complete_conflicting_writes(mdev, sector, size);
- if (err) {
- if (err != -ERESTARTSYS)
- _conn_request_state(mdev->tconn,
- NS(conn, C_TIMEOUT),
- CS_HARD);
- spin_unlock_irq(&mdev->tconn->req_lock);
- err = -EIO;
- goto fail_free_complete;
- }
+ /* This may temporarily give up the req_lock,
+ * but will re-aquire it before it returns here.
+ * Needs to be before the check on drbd_suspended() */
+ complete_conflicting_writes(req);
}
- if (drbd_suspended(mdev)) {
- /* If we got suspended, use the retry mechanism of
- generic_make_request() to restart processing of this
- bio. In the next call to drbd_make_request
- we sleep in inc_ap_bio() */
- ret = 1;
- spin_unlock_irq(&mdev->tconn->req_lock);
- goto fail_free_complete;
- }
+ /* no more giving up req_lock from now on! */
- if (remote || send_oos) {
- remote = drbd_should_do_remote(mdev->state);
- send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
- D_ASSERT(!(remote && send_oos));
-
- if (!(remote || send_oos))
- dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
- if (!(local || remote)) {
- dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
- spin_unlock_irq(&mdev->tconn->req_lock);
- err = -EIO;
- goto fail_free_complete;
+ if (drbd_suspended(mdev)) {
+ /* push back and retry: */
+ req->rq_state |= RQ_POSTPONED;
+ if (req->private_bio) {
+ bio_put(req->private_bio);
+ req->private_bio = NULL;
+ put_ldev(mdev);
}
+ goto out;
}
- if (b && mdev->tconn->unused_spare_tle == NULL) {
- mdev->tconn->unused_spare_tle = b;
- b = NULL;
- }
- if (rw == WRITE && (remote || send_oos) &&
- mdev->tconn->unused_spare_tle == NULL &&
- test_bit(CREATE_BARRIER, &mdev->flags)) {
- /* someone closed the current epoch
- * while we were grabbing the spinlock */
- spin_unlock_irq(&mdev->tconn->req_lock);
- goto allocate_barrier;
- }
-
-
/* Update disk stats */
_drbd_start_io_acct(mdev, req, bio);
- /* _maybe_start_new_epoch(mdev);
- * If we need to generate a write barrier packet, we have to add the
- * new epoch (barrier) object, and queue the barrier packet for sending,
- * and queue the req's data after it _within the same lock_, otherwise
- * we have race conditions were the reorder domains could be mixed up.
- *
- * Even read requests may start a new epoch and queue the corresponding
- * barrier packet. To get the write ordering right, we only have to
- * make sure that, if this is a write request and it triggered a
- * barrier packet, this request is queued within the same spinlock. */
- if ((remote || send_oos) && mdev->tconn->unused_spare_tle &&
- test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
- _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle);
- mdev->tconn->unused_spare_tle = NULL;
- } else {
- D_ASSERT(!(remote && rw == WRITE &&
- test_bit(CREATE_BARRIER, &mdev->flags)));
+ /* We fail READ/READA early, if we can not serve it.
+ * We must do this before req is registered on any lists.
+ * Otherwise, drbd_req_complete() will queue failed READ for retry. */
+ if (rw != WRITE) {
+ if (!do_remote_read(req) && !req->private_bio)
+ goto nodata;
}
- /* NOTE
- * Actually, 'local' may be wrong here already, since we may have failed
- * to write to the meta data, and may become wrong anytime because of
- * local io-error for some other request, which would lead to us
- * "detaching" the local disk.
- *
- * 'remote' may become wrong any time because the network could fail.
- *
- * This is a harmless race condition, though, since it is handled
- * correctly at the appropriate places; so it just defers the failure
- * of the respective operation.
- */
+ /* which transfer log epoch does this belong to? */
+ req->epoch = atomic_read(&mdev->tconn->current_tle_nr);
- /* mark them early for readability.
- * this just sets some state flags. */
- if (remote)
- _req_mod(req, TO_BE_SENT);
- if (local)
- _req_mod(req, TO_BE_SUBMITTED);
-
- list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests);
-
- /* NOTE remote first: to get the concurrent write detection right,
- * we must register the request before start of local IO. */
- if (remote) {
- /* either WRITE and C_CONNECTED,
- * or READ, and no local disk,
- * or READ, but not in sync.
- */
- _req_mod(req, (rw == WRITE)
- ? QUEUE_FOR_NET_WRITE
- : QUEUE_FOR_NET_READ);
- }
- if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
- _req_mod(req, QUEUE_FOR_SEND_OOS);
-
- rcu_read_lock();
- nc = rcu_dereference(mdev->tconn->net_conf);
- if (remote &&
- nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) {
- int congested = 0;
-
- if (nc->cong_fill &&
- atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
- dev_info(DEV, "Congestion-fill threshold reached\n");
- congested = 1;
- }
-
- if (mdev->act_log->used >= nc->cong_extents) {
- dev_info(DEV, "Congestion-extents threshold reached\n");
- congested = 1;
- }
-
- if (congested) {
- queue_barrier(mdev); /* last barrier, after mirrored writes */
+ /* no point in adding empty flushes to the transfer log,
+ * they are mapped to drbd barriers already. */
+ if (likely(req->i.size!=0)) {
+ if (rw == WRITE)
+ mdev->tconn->current_tle_writes++;
- if (nc->on_congestion == OC_PULL_AHEAD)
- _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
- else /*nc->on_congestion == OC_DISCONNECT */
- _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
- }
+ list_add_tail(&req->tl_requests, &mdev->tconn->transfer_log);
}
- rcu_read_unlock();
- spin_unlock_irq(&mdev->tconn->req_lock);
- kfree(b); /* if someone else has beaten us to it... */
-
- if (local) {
- req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
-
- /* State may have changed since we grabbed our reference on the
- * mdev->ldev member. Double check, and short-circuit to endio.
- * In case the last activity log transaction failed to get on
- * stable storage, and this is a WRITE, we may not even submit
- * this bio. */
- if (get_ldev(mdev)) {
- if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
- : rw == READ ? DRBD_FAULT_DT_RD
- : DRBD_FAULT_DT_RA))
- bio_endio(req->private_bio, -EIO);
- else
- generic_make_request(req->private_bio);
- put_ldev(mdev);
+ if (rw == WRITE) {
+ if (!drbd_process_write_request(req))
+ no_remote = true;
+ } else {
+ /* We either have a private_bio, or we can read from remote.
+ * Otherwise we had done the goto nodata above. */
+ if (req->private_bio == NULL) {
+ _req_mod(req, TO_BE_SENT);
+ _req_mod(req, QUEUE_FOR_NET_READ);
} else
- bio_endio(req->private_bio, -EIO);
+ no_remote = true;
}
- return 0;
-
-fail_free_complete:
- if (req->rq_state & RQ_IN_ACT_LOG)
- drbd_al_complete_io(mdev, &req->i);
-fail_and_free_req:
- if (local) {
- bio_put(req->private_bio);
- req->private_bio = NULL;
- put_ldev(mdev);
+ if (req->private_bio) {
+ /* needs to be marked within the same spinlock */
+ _req_mod(req, TO_BE_SUBMITTED);
+ /* but we need to give up the spinlock to submit */
+ spin_unlock_irq(&mdev->tconn->req_lock);
+ drbd_submit_req_private_bio(req);
+ spin_lock_irq(&mdev->tconn->req_lock);
+ } else if (no_remote) {
+nodata:
+ if (__ratelimit(&drbd_ratelimit_state))
+ dev_err(DEV, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
+ (unsigned long long)req->i.sector, req->i.size >> 9);
+ /* A write may have been queued for send_oos, however.
+ * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
}
- if (!ret)
- bio_endio(bio, err);
- drbd_req_free(req);
- dec_ap_bio(mdev);
- kfree(b);
+out:
+ if (drbd_req_put_completion_ref(req, &m, 1))
+ kref_put(&req->kref, drbd_req_destroy);
+ spin_unlock_irq(&mdev->tconn->req_lock);
- return ret;
+ if (m.bio)
+ complete_master_bio(mdev, &m);
+ return;
}
-int drbd_make_request(struct request_queue *q, struct bio *bio)
+void drbd_make_request(struct request_queue *q, struct bio *bio)
{
struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
unsigned long start_time;
/*
* what we "blindly" assume:
*/
- D_ASSERT(bio->bi_size > 0);
D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
inc_ap_bio(mdev);
- return __drbd_make_request(mdev, bio, start_time);
+ __drbd_make_request(mdev, bio, start_time);
}
/* This is called by bio_add_page().
return limit;
}
+struct drbd_request *find_oldest_request(struct drbd_tconn *tconn)
+{
+ /* Walk the transfer log,
+ * and find the oldest not yet completed request */
+ struct drbd_request *r;
+ list_for_each_entry(r, &tconn->transfer_log, tl_requests) {
+ if (atomic_read(&r->completion_ref))
+ return r;
+ }
+ return NULL;
+}
+
void request_timer_fn(unsigned long data)
{
struct drbd_conf *mdev = (struct drbd_conf *) data;
struct drbd_tconn *tconn = mdev->tconn;
struct drbd_request *req; /* oldest request */
- struct list_head *le;
struct net_conf *nc;
unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
+ unsigned long now;
rcu_read_lock();
nc = rcu_dereference(tconn->net_conf);
- ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0;
+ if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS)
+ ent = nc->timeout * HZ/10 * nc->ko_count;
- if (get_ldev(mdev)) {
+ if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */
dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
put_ldev(mdev);
}
et = min_not_zero(dt, ent);
- if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED))
+ if (!et)
return; /* Recurring timer stopped */
+ now = jiffies;
+
spin_lock_irq(&tconn->req_lock);
- le = &tconn->oldest_tle->requests;
- if (list_empty(le)) {
+ req = find_oldest_request(tconn);
+ if (!req) {
spin_unlock_irq(&tconn->req_lock);
- mod_timer(&mdev->request_timer, jiffies + et);
+ mod_timer(&mdev->request_timer, now + et);
return;
}
- le = le->prev;
- req = list_entry(le, struct drbd_request, tl_requests);
- if (ent && req->rq_state & RQ_NET_PENDING) {
- if (time_is_before_eq_jiffies(req->start_time + ent)) {
- dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
- _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
- }
+ /* The request is considered timed out, if
+ * - we have some effective timeout from the configuration,
+ * with above state restrictions applied,
+ * - the oldest request is waiting for a response from the network
+ * resp. the local disk,
+ * - the oldest request is in fact older than the effective timeout,
+ * - the connection was established (resp. disk was attached)
+ * for longer than the timeout already.
+ * Note that for 32bit jiffies and very stable connections/disks,
+ * we may have a wrap around, which is catched by
+ * !time_in_range(now, last_..._jif, last_..._jif + timeout).
+ *
+ * Side effect: once per 32bit wrap-around interval, which means every
+ * ~198 days with 250 HZ, we have a window where the timeout would need
+ * to expire twice (worst case) to become effective. Good enough.
+ */
+ if (ent && req->rq_state & RQ_NET_PENDING &&
+ time_after(now, req->start_time + ent) &&
+ !time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) {
+ dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
+ _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
}
- if (dt && req->rq_state & RQ_LOCAL_PENDING) {
- if (time_is_before_eq_jiffies(req->start_time + dt)) {
- dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
- __drbd_chk_io_error(mdev, 1);
- }
+ if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev &&
+ time_after(now, req->start_time + dt) &&
+ !time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) {
+ dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
+ __drbd_chk_io_error(mdev, DRBD_FORCE_DETACH);
}
- nt = (time_is_before_eq_jiffies(req->start_time + et) ? jiffies : req->start_time) + et;
+ nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
spin_unlock_irq(&tconn->req_lock);
mod_timer(&mdev->request_timer, nt);
}