/* 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;
}
/* If called from mod_rq_state (expected normal case) or *drbd_send_and_submit(thelesslikelynormalpath),thisholdsthe *req_lock,andreq->tl_requestswilltypicalybeon->transfer_log, *thoughitmaybestillempty(neveraddedtothetransferlog). * *Ifcalledfromdo_retry(),wedoNOTholdthereq_lock,butweare *stillallowedtounconditionallylist_del(&req->tl_requests),
* because it will be on a local on-stack list only. */
list_del_init(&req->tl_requests);
/* finally remove the request from the conflict detection
* respective block_id verification interval tree. */ if (!drbd_interval_empty(&req->i)) { struct rb_root *root;
/* if it was a write, we may have to set the corresponding *bit(s)out-of-syncfirst.Ifithadalocalpart,weneedto
* release the reference to the activity log. */ if (s & RQ_WRITE) { /* Set out-of-sync unless both OK flags are set *(localonlyorremotefailed). *Otherplaceswherewesetout-of-sync:
* READ with local io-error */
/* There is a special case: *wemaynoticelatethatIOwassuspended, *andpostpone,orscheduleforretry,awrite, *beforeitevenwassubmittedorsent. *Inthatcasewedonotwanttotouchthebitmapatall.
*/ struct drbd_peer_device *peer_device = first_peer_device(device); 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(peer_device, req->i.sector, req->i.size);
if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
drbd_set_in_sync(peer_device, req->i.sector, req->i.size);
}
/* one might be tempted to move the drbd_al_complete_io *tothelocaliocompletioncallbackdrbd_request_endio. *but,ifthiswasamirrorwrite,wemayonly *drbd_al_complete_ioafterthisisRQ_NET_DONE, *otherwisetheextentcouldbedroppedfromtheal *beforeithasactuallybeenwrittenonthepeer. *ifwecrashbeforeourpeerknowsabouttherequest, *butaftertheextenthasbeendroppedfromtheal, *wewouldforgettoresyncthecorrespondingextent.
*/ if (s & RQ_IN_ACT_LOG) { if (get_ldev_if_state(device, D_FAILED)) {
drbd_al_complete_io(device, &req->i);
put_ldev(device);
} elseif (drbd_ratelimit()) {
drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), " "but my Disk seems to have failed :(\n",
(unsignedlonglong) req->i.sector, req->i.size);
}
}
}
/* must hold resource->req_lock */ void start_new_tl_epoch(struct drbd_connection *connection)
{ /* no point closing an epoch, if it is empty, anyways. */ if (connection->current_tle_writes == 0) return;
/* Helper for __req_mod(). *Setm->biotothemasterbio,ifitisfittobecompleted, *orleaveitalone(itisinitializedtoNULLin__req_mod), *ifithasalreadybeencompleted,orcannotbecompletedyet. *Ifm->bioisset,theerrorstatustobereturnedisplacedinm->error.
*/ static void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
{ constunsigned s = req->rq_state; struct drbd_device *device = req->device; int error, ok;
/* we must not complete the master bio, while it is *stillbeingprocessedby_drbd_send_zc_bio(drbd_send_dblock) *notyetacknowledgedbythepeer *notyetcompletedbythelocaliosubsystem *theseflagsmaygetclearedinanyorderby *theworker, *thereceiver, *thebio_endiocompletioncallbacks.
*/ if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
(s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
(s & RQ_COMPLETION_SUSP)) {
drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s); return;
}
/* *figureoutwhethertoreportsuccessorfailure. * *reportsuccesswhenatleastoneoftheoperationssucceeded. *or,toputtheotherway, *onlyreportfailure,whenbothoperationsfailed. * *whattodoaboutthefailuresishandledelsewhere. *whatweneedtodohereisjust:completethemaster_bio. * *localcompletionerror,ifany,hasbeenstoredasERR_PTR *inprivate_biowithindrbd_request_endio.
*/
ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
error = PTR_ERR(req->private_bio);
/* Before we can signal completion to the upper layers, *wemayneedtoclosethecurrenttransferlogepoch. *Wearewithintherequestlock,sowecansimplycompare *therequestepochnumberwiththecurrenttransferlog *epochnumber.Iftheymatch,increasethecurrent_tle_nr, *andresetthetransferlogepochwrite_cnt.
*/ if (op_is_write(bio_op(req->master_bio)) &&
req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
start_new_tl_epoch(first_peer_device(device)->connection);
/* Update disk stats */
bio_end_io_acct(req->master_bio, req->start_jif);
if (!(req->rq_state & RQ_POSTPONED)) {
m->error = ok ? 0 : (error ?: -EIO);
m->bio = req->master_bio;
req->master_bio = NULL; /* We leave it in the tree, to be able to verify later *write-acksinprotocol!=Cduringresync. *Butwemarkitas"complete",soitwon'tbecountedas
* conflict in a multi-primary setup. */
req->i.completed = true;
}
if (req->i.waiting)
wake_up(&device->misc_wait);
/* Either we are about to complete to upper layers, *orwewillrestartthisrequest. *Ineithercase,therequestobjectwillbedestroyedsoon,
* so better remove it from all lists. */
list_del_init(&req->req_pending_master_completion);
}
/* still holds resource->req_lock */ staticvoid drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
{ struct drbd_device *device = req->device;
D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
if (!put) return;
if (!atomic_sub_and_test(put, &req->completion_ref)) return;
drbd_req_complete(req, m);
/* local completion may still come in later,
* we need to keep the req object around. */ if (req->rq_state & RQ_LOCAL_ABORTED) return;
if (req->rq_state & RQ_POSTPONED) { /* don't destroy the req object just yet,
* but queue it for retry */
drbd_restart_request(req); return;
}
/* I'd like this to be the only place that manipulates
* req->completion_ref and req->kref. */ staticvoid mod_rq_state(struct drbd_request *req, struct bio_and_error *m, int clear, int set)
{ struct drbd_device *device = req->device; struct drbd_peer_device *peer_device = first_peer_device(device); unsigned s = req->rq_state; int c_put = 0;
if (drbd_suspended(device) && !((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 */
kref_get(&req->kref);
if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
atomic_inc(&req->completion_ref);
if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) { if (s & RQ_NET_SENT)
atomic_sub(req->i.size >> 9, &device->ap_in_flight); if (s & RQ_EXP_BARR_ACK)
kref_put(&req->kref, drbd_req_destroy);
req->net_done_jif = jiffies;
/* in ahead/behind mode, or just in case, *beforewefinallydestroythisrequest,
* the caching pointers must not reference it anymore */
advance_conn_req_next(peer_device, req);
advance_conn_req_ack_pending(peer_device, req);
advance_conn_req_not_net_done(peer_device, req);
}
/* potentially complete and destroy */
/* If we made progress, retry conflicting peer requests, if any. */ if (req->i.waiting)
wake_up(&device->misc_wait);
drbd_req_put_completion_ref(req, m, c_put);
kref_put(&req->kref, drbd_req_destroy);
}
/* does not happen... *initializationdoneindrbd_req_new caseCREATED: break;
*/
case TO_BE_SENT: /* via network */ /* reached via __drbd_make_request
* and from w_read_retry_remote */
D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
rcu_read_lock();
nc = rcu_dereference(connection->net_conf);
p = nc->wire_protocol;
rcu_read_unlock();
req->rq_state |=
p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
mod_rq_state(req, m, 0, RQ_NET_PENDING); break;
case TO_BE_SUBMITTED: /* locally */ /* reached via __drbd_make_request */
D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
mod_rq_state(req, m, 0, RQ_LOCAL_PENDING); break;
case COMPLETED_OK: if (req->rq_state & RQ_WRITE)
device->writ_cnt += req->i.size >> 9; else
device->read_cnt += req->i.size >> 9;
mod_rq_state(req, m, RQ_LOCAL_PENDING,
RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); break;
case ABORT_DISK_IO:
mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED); break;
case WRITE_COMPLETED_WITH_ERROR:
drbd_report_io_error(device, req);
__drbd_chk_io_error(device, 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(peer_device, req->i.sector, req->i.size);
drbd_report_io_error(device, req);
__drbd_chk_io_error(device, DRBD_READ_ERROR);
fallthrough; case READ_AHEAD_COMPLETED_WITH_ERROR: /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); break;
case DISCARD_COMPLETED_NOTSUPP: case DISCARD_COMPLETED_WITH_ERROR: /* I'd rather not detach from local disk just because it
* failed a REQ_OP_DISCARD. */
mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); break;
case QUEUE_FOR_NET_READ: /* READ, and *nolocaldisk, *ortargetareamarkedasinvalid,
* or just got an io-error. */ /* from __drbd_make_request
* or from bio_endio during read io-error recovery */
/* So we can verify the handle in the answer packet. *Correspondingdrbd_remove_request_intervalisin
* drbd_req_complete() */
D_ASSERT(device, drbd_interval_empty(&req->i));
drbd_insert_interval(&device->read_requests, &req->i);
case QUEUE_FOR_NET_WRITE: /* assert something? */ /* from __drbd_make_request only */
/* Corresponding drbd_remove_request_interval is in
* drbd_req_complete() */
D_ASSERT(device, drbd_interval_empty(&req->i));
drbd_insert_interval(&device->write_requests, &req->i);
/* NOTE *Incasethereqendeduponthetransferlogbeforebeing *queuedontheworker,itcouldleadtothisrequestbeing *missedduringcleanupafterconnectionloss. *Sowehavetodobothoperationshere, *withinthesamelockthatprotectsthetransferlog. * *_req_add_to_epoch(req);thishastobeafterthe *_maybe_start_new_epoch(req);whichhappenedin *__drbd_make_request,becausewenowmaysetthebit *againourselvestoclosethecurrentepoch. *
* Add req to the (now) current epoch (barrier). */
/* otherwise we may lose an unplug, which may cause some remote *io-schedulertimeouttoexpire,increasingmaximumlatency,
* hurting performance. */
set_bit(UNPLUG_REMOTE, &device->flags);
/* queue work item to send data */
D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
req->w.cb = w_send_dblock;
drbd_queue_work(&connection->sender_work,
&req->w);
/* close the epoch, in case it outgrew the limit */
rcu_read_lock();
nc = rcu_dereference(connection->net_conf);
p = nc->max_epoch_size;
rcu_read_unlock(); if (connection->current_tle_writes >= p)
start_new_tl_epoch(connection);
break;
case QUEUE_FOR_SEND_OOS:
mod_rq_state(req, m, 0, RQ_NET_QUEUED);
req->w.cb = w_send_out_of_sync;
drbd_queue_work(&connection->sender_work,
&req->w); break;
case READ_RETRY_REMOTE_CANCELED: case SEND_CANCELED: 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 */
mod_rq_state(req, m, RQ_NET_QUEUED, 0); break;
case HANDED_OVER_TO_NETWORK: /* assert something? */ if (is_pending_write_protocol_A(req)) /* this is what is dangerous about protocol A:
* pretend it was successfully written on the peer. */
mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
RQ_NET_SENT|RQ_NET_OK); else
mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT); /* It is still not yet RQ_NET_DONE until the *correspondingepochbarriergotackedaswell,
* so we know what to dirty on connection loss. */ 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 CONNECTION_LOST_WHILE_PENDING: /* transfer log cleanup after connection loss */
mod_rq_state(req, m,
RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
RQ_NET_DONE); break;
case CONFLICT_RESOLVED: /* for superseded conflicting writes of multiple primaries, *thereisnoneedtokeepanythinginthetl,potential *nodecrashesarecoveredbytheactivitylog. * *IfthisrequesthadbeenmarkedasRQ_POSTPONEDbefore, *itwillactuallynotbecompleted,but"restarted",
* resubmitted from the retry worker context. */
D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
D_ASSERT(device, 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;
fallthrough; case WRITE_ACKED_BY_PEER: /* Normal operation protocol C: successfully written on peer. *Duringresync,eveninprotocol!=C, *werequestedanexplicitwriteackanyways. *Whichmeanswecannotevenassertanythinghere. *Nothingmoretodohere. *Wewanttokeepthetlinplaceforallprotocols,tocater
* for volatile write-back caches on lower level devices. */ goto ack_common; case RECV_ACKED_BY_PEER:
D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK); /* protocol B; pretends to be successfully written on peer. *seealsonotesaboveinHANDED_OVER_TO_NETWORKabout
* protocol != C */
ack_common:
mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); break;
case POSTPONE_WRITE:
D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); /* If this node has already detected the write conflict, the *workerwillbewaitingonmisc_wait.Wakeituponcethis *requesthascompletedlocally.
*/
D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
req->rq_state |= RQ_POSTPONED; if (req->i.waiting)
wake_up(&device->misc_wait); /* Do not clear RQ_NET_PENDING. This request will make further *progressviarestart_conflicting_writes()or
* fail_postponed_requests(). Hopefully. */ break;
case NEG_ACKED:
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;
mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); break;
case RESTART_FROZEN_DISK_IO: if (!(req->rq_state & RQ_LOCAL_COMPLETED)) break;
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(device); /* always succeeds in this call path */
req->w.cb = w_restart_disk_io;
drbd_queue_work(&connection->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 beforetheconnectionloss(B&Conly);onlyP_BARRIER_ACK (orthelocalcompletion?)wasmissingwhenwesuspended. ThrowingthemoutoftheTLherebypretendingwegotaBARRIER_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->dw.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) { /* w.cb expected to be w_send_dblock, or w_send_read_req */
drbd_queue_work(&connection->sender_work,
&req->w);
rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
} /* else: FIXME can this happen? */ break;
}
fallthrough; /* to BARRIER_ACKED */
case BARRIER_ACKED: /* barrier ack for READ requests does not make sense */ if (!(req->rq_state & RQ_WRITE)) break;
if (req->rq_state & RQ_NET_PENDING) { /* barrier came in before all requests were acked. *thisisbad,becauseiftheconnectionislostnow,
* we won't be able to clean them up... */
drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
} /* Allowed to complete requests, even while suspended. *Asthisiscalledforallrequestswithinamatchingepoch, *weneedtofilter,andonlysetRQ_NET_DONEforthosethat
* 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(device, req->rq_state & RQ_NET_PENDING);
mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE); break;
case QUEUE_AS_DRBD_BARRIER:
start_new_tl_epoch(connection);
mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE); break;
}
return rv;
}
/* we may do a local read if: *-weareconsistent(ofcourse), *-orwearegenerallyinconsistent, *BUTwearestill/alreadyINSYNCforthisarea. *sincesizemaybebiggerthanBM_BLOCK_SIZE, *wemayneedtocheckseveralbits.
*/ staticbool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
{ unsignedlong sbnr, ebnr;
sector_t esector, nr_sectors;
for (;;) {
drbd_for_each_overlap(i, &device->write_requests, sector, size) { /* Ignore, if already completed to upper layers. */ if (i->completed) continue; /* Handle the first found overlap. After the schedule
* we have to restart the tree walk. */ break;
} if (!i) /* if any */ break;
/* Indicate to wake up device->misc_wait on progress. */
prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
i->waiting = true;
spin_unlock_irq(&device->resource->req_lock);
schedule();
spin_lock_irq(&device->resource->req_lock);
}
finish_wait(&device->misc_wait, &wait);
}
if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD) return; /* nothing to do ... */
/* If I don't even have good local storage, we can not reasonably try *topullaheadofthepeer.Wealsoneedthelocalreferencetomake *suredevice->act_logisthere.
*/ if (!get_ldev_if_state(device, D_UP_TO_DATE)) return;
if (remote_due_to_read_balancing(device, req->i.sector, rbm)) { if (req->private_bio) {
bio_put(req->private_bio);
req->private_bio = NULL;
put_ldev(device);
} returntrue;
}
returnfalse;
}
bool drbd_should_do_remote(union drbd_dev_state s)
{ return s.pdsk == D_UP_TO_DATE ||
(s.pdsk >= D_INCONSISTENT &&
s.conn >= C_WF_BITMAP_T &&
s.conn < C_AHEAD); /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T. Thatisequivalentsincebefore96IOwasfrozenintheC_WF_BITMAP*
states. */
}
staticbool drbd_should_send_out_of_sync(union drbd_dev_state s)
{ return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S; /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
since we enter state C_AHEAD only if proto >= 96 */
}
/* returns number of connections (== 1, for drbd 8.4) *expectedtoactuallywritethisdata,
* which does NOT include those that we are L_AHEAD for. */ staticint drbd_process_write_request(struct drbd_request *req)
{ struct drbd_device *device = req->device; struct drbd_peer_device *peer_device = first_peer_device(device); int remote, send_oos;
/* Need to replicate writes. Unless it is an empty flush, *whichisbettermappedtoaDRBDP_BARRIERpacket, *alsofordrbdwireprotocolcompatibilityreasons. *Ifthiswasaflush,juststartanewepoch. *Unlessthecurrentepochwasemptyanyways,orwearenotcurrently
* 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(device, req->master_bio->bi_opf & REQ_PREFLUSH); if (remote)
_req_mod(req, QUEUE_AS_DRBD_BARRIER, peer_device); return remote;
}
if (bio_op(bio) != REQ_OP_READ)
type = DRBD_FAULT_DT_WR; elseif (bio->bi_opf & REQ_RAHEAD)
type = DRBD_FAULT_DT_RA; else
type = DRBD_FAULT_DT_RD;
/* State may have changed since we grabbed our reference on the *->ldevmember.Doublecheck,andshort-circuittoendio. *Incasethelastactivitylogtransactionfailedtogeton *stablestorage,andthisisaWRITE,wemaynotevensubmit
* this bio. */ if (get_ldev(device)) { if (drbd_insert_fault(device, type))
bio_io_error(bio); elseif (bio_op(bio) == REQ_OP_WRITE_ZEROES)
drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM)); elseif (bio_op(bio) == REQ_OP_DISCARD)
drbd_process_discard_or_zeroes_req(req, EE_TRIM); else
submit_bio_noacct(bio);
put_ldev(device);
} else
bio_io_error(bio);
}
staticvoid drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
{
spin_lock_irq(&device->resource->req_lock);
list_add_tail(&req->tl_requests, &device->submit.writes);
list_add_tail(&req->req_pending_master_completion,
&device->pending_master_completion[1/* WRITE */]);
spin_unlock_irq(&device->resource->req_lock);
queue_work(device->submit.wq, &device->submit.worker); /* do_submit() may sleep internally on al_wait, too */
wake_up(&device->al_wait);
}
/* returns the new drbd_request pointer, if the caller is expected to *drbd_send_and_submit()it(tosavelatency),orNULLifwequeuedthe *requestonthesubmitterthread. *ReturnsERR_PTR(-ENOMEM)ifwecannotallocateadrbd_request.
*/ staticstruct drbd_request *
drbd_request_prepare(struct drbd_device *device, struct bio *bio)
{ constint rw = bio_data_dir(bio); struct drbd_request *req;
/* allocate outside of all locks; */
req = drbd_req_new(device, bio); if (!req) {
dec_ap_bio(device); /* only pass the error to the upper layers.
* if user cannot handle io errors, that's not our business. */
drbd_err(device, "could not kmalloc() req\n");
bio->bi_status = BLK_STS_RESOURCE;
bio_endio(bio); return ERR_PTR(-ENOMEM);
}
/* Update disk stats */
req->start_jif = bio_start_io_acct(req->master_bio);
spin_lock_irq(&resource->req_lock); /* In case the sender did not process it yet, raise the flag to
* have it followed with P_UNPLUG_REMOTE just after. */
req->rq_state |= RQ_UNPLUG; /* but also queue a generic unplug */
drbd_queue_unplug(req->device);
kref_put(&req->kref, drbd_req_destroy);
spin_unlock_irq(&resource->req_lock);
}
staticstruct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
{ /* A lot of text to say
* return (struct drbd_plug_cb*)blk_check_plugged(); */ struct drbd_plug_cb *plug; struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
staticvoid drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
{ struct drbd_request *tmp = plug->most_recent_req; /* Will be sent to some peer.
* Remember to tag it with UNPLUG_REMOTE on unplug */
kref_get(&req->kref);
plug->most_recent_req = req; if (tmp)
kref_put(&tmp->kref, drbd_req_destroy);
}
spin_lock_irq(&resource->req_lock); if (rw == WRITE) { /* This may temporarily give up the req_lock, *butwillre-aquireitbeforeitreturnshere.
* Needs to be before the check on drbd_suspended() */
complete_conflicting_writes(req); /* no more giving up req_lock from now on! */
/* check for congestion, and potentially stop sending
* full data updates, but start sending "dirty bits" only. */
maybe_pull_ahead(device);
}
if (drbd_suspended(device)) { /* push back and retry: */
req->rq_state |= RQ_POSTPONED; if (req->private_bio) {
bio_put(req->private_bio);
req->private_bio = NULL;
put_ldev(device);
} goto out;
}
/* We fail READ early, if we can not serve it. *Wemustdothisbeforereqisregisteredonanylists.
* Otherwise, drbd_req_complete() will queue failed READ for retry. */ if (rw != WRITE) { if (!do_remote_read(req) && !req->private_bio) goto nodata;
}
/* which transfer log epoch does this belong to? */
req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
/* 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)
first_peer_device(device)->connection->current_tle_writes++;
if (rw == WRITE) { if (req->private_bio && !may_do_writes(device)) {
bio_put(req->private_bio);
req->private_bio = NULL;
put_ldev(device); goto nodata;
} 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, peer_device);
_req_mod(req, QUEUE_FOR_NET_READ, peer_device);
} else
no_remote = true;
}
if (no_remote == false) { struct drbd_plug_cb *plug = drbd_check_plugged(resource); if (plug)
drbd_update_plug(plug, req);
}
/* If it took the fast path in drbd_request_prepare, add it here.
* The slow path has added it already. */ if (list_empty(&req->req_pending_master_completion))
list_add_tail(&req->req_pending_master_completion,
&device->pending_master_completion[rw == WRITE]); if (req->private_bio) { /* needs to be marked within the same spinlock */
req->pre_submit_jif = jiffies;
list_add_tail(&req->req_pending_local,
&device->pending_completion[rw == WRITE]);
_req_mod(req, TO_BE_SUBMITTED, NULL); /* but we need to give up the spinlock to submit */
submit_private_bio = true;
} elseif (no_remote) {
nodata: if (drbd_ratelimit())
drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
(unsignedlonglong)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() */
}
/* Even though above is a kref_put(), this is safe. *Aslongaswestillneedtosubmitourprivatebio, *weholdacompletionref,andtherequestcannotdisappear. *Ifhoweverthisrequestdidnotevenhaveaprivatebiotosubmit *(e.g.remoteread),reqmayalreadybeinvalidnow.
* That's why we cannot check on req->private_bio. */ if (submit_private_bio)
drbd_submit_req_private_bio(req); if (m.bio)
complete_master_bio(device, &m);
}
void __drbd_make_request(struct drbd_device *device, struct bio *bio)
{ struct drbd_request *req = drbd_request_prepare(device, bio); if (IS_ERR_OR_NULL(req)) return;
drbd_send_and_submit(device, req);
}
if (rw == WRITE /* rw != WRITE should not even end up here! */
&& req->private_bio && req->i.size
&& !test_bit(AL_SUSPENDED, &device->flags)) { if (!drbd_al_begin_io_fastpath(device, &req->i)) continue;
void do_submit(struct work_struct *ws)
{ struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
LIST_HEAD(incoming); /* from drbd_make_request() */
LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */
LIST_HEAD(busy); /* blocked by resync requests */
/* grab new incoming requests */
spin_lock_irq(&device->resource->req_lock);
list_splice_tail_init(&device->submit.writes, &incoming);
spin_unlock_irq(&device->resource->req_lock);
for (;;) {
DEFINE_WAIT(wait);
/* move used-to-be-busy back to front of incoming */
list_splice_init(&busy, &incoming);
submit_fast_path(device, &incoming); if (list_empty(&incoming)) break;
for (;;) {
prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
list_splice_init(&busy, &incoming);
prepare_al_transaction_nonblock(device, &incoming, &pending, &busy); if (!list_empty(&pending)) break;
schedule();
/* If all currently "hot" activity log extents are kept busy by *incomingrequests,westillmustnottotallystarvenew *requeststo"cold"extents. *Somethinglefton&incomingmeanstherehadnotbeen *enoughupdateslotsavailable,andtheactivitylog *hasbeenmarkedas"starving". * *Tryagainnow,withoutlookingfornewrequests, *effectivelyblockingallnewrequestsuntilwemade *atleast_some_progresswithwhatwecurrentlyhave.
*/ if (!list_empty(&incoming)) continue;
/* Nothing moved to pending, but nothing left *onincoming:allmovedtobusy!
* Grab new and iterate. */
spin_lock_irq(&device->resource->req_lock);
list_splice_tail_init(&device->submit.writes, &incoming);
spin_unlock_irq(&device->resource->req_lock);
}
finish_wait(&device->al_wait, &wait);
/* If the transaction was full, before all incoming requests *hadbeenprocessed,skipaheadtocommit,anditerate *withoutsplicinginmoreincomingrequestsfromupperlayers. * *Else,ifallincominghavebeenprocessed, *theyhavebecomeeither"pending"(tobesubmittedafter *nexttransactioncommit)or"busy"(blockedbyresync). * *Maybemorewasqueued,whilewepreparedthetransaction? *Trytostuffthoseintothistransactionaswell. *Bestrictlynon-blockinghere, *wealreadyhavesomethingtocommit. * *Commitifwedon'tmakeanymoreprogres.
*/
while (list_empty(&incoming)) {
LIST_HEAD(more_pending);
LIST_HEAD(more_incoming); bool made_progress;
/* It is ok to look outside the lock,
* it's only an optimization anyways */ if (list_empty(&device->submit.writes)) break;
if (!time_after(now, net_req->pre_send_jif + ent)) returnfalse;
if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) returnfalse;
if (net_req->rq_state & RQ_NET_PENDING) {
drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); returntrue;
}
/* We received an ACK already (or are using protocol A), *butarewaitingfortheepochclosingbarrierack. *Checkifwesentthebarrieralready.Weshouldnotblamethepeer
* for being unresponsive, if we did not even ask it yet. */ if (net_req->epoch == connection->send.current_epoch_nr) {
drbd_warn(device, "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); returnfalse;
}
/* Worst case: we may have been blocked for whatever reason, then *suddenlyareabletosendalotofrequests(andepochseparating *barriers)inquicksuccession. *Thetimestampofthenet_reqmaybemuchtoooldandnotcorrespond *tothesendingtimeoftherelevantunack'edbarrierpacket,so *wouldtriggeraspurioustimeout.Thelatestbarrierpacketmay *haveatoorecenttimestamptotriggerthetimeout,potentiallymiss *atimeout.Rightnowwedon'thaveaplacetoconvenientlystore *thesetimestamps. *Butinthisparticularsituation,theapplicationrequestsarestill *completedtoupperlayers,DRBDshouldstill"feel"responsive. *Noneedyettokillthisconnection,itmaystillrecover. *Ifnot,eventuallywewillhavequeuedenoughintothenetworkfor *ustoblock.Fromthatpointofview,thetimestampofthelastsent *barrierpacketisrelevantenough.
*/ if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
connection->send.last_sent_barrier_jif, now,
jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout); returntrue;
} returnfalse;
}
/* A request is considered timed out, if *-wehavesomeeffectivetimeoutfromtheconfiguration, *withsomestaterestrictionsapplied, *-theoldestrequestiswaitingforaresponsefromthenetwork *resp.thelocaldisk, *-theoldestrequestisinfactolderthantheeffectivetimeout, *-theconnectionwasestablished(resp.diskwasattached) *forlongerthanthetimeoutalready. *Notethatfor32bitjiffiesandverystableconnections/disks, *wemayhaveawraparound,whichiscatchedby *!time_in_range(now,last_..._jif,last_..._jif+timeout). * *Sideeffect:onceper32bitwrap-aroundinterval,whichmeansevery *~198dayswith250HZ,wehaveawindowwherethetimeoutwouldneed *toexpiretwice(worstcase)tobecomeeffective.Goodenough.
*/
/* maybe the oldest request waiting for the peer is in fact still *blockingintcpsendmsg.That'sok,though,that'shandledviathe *socketsendtimeout,requestingaping,andbumpingko-countin *we_should_drop_the_connection().
*/
/* check the oldest request we did successfully sent,
* but which is still waiting for an ACK. */
req_peer = connection->req_ack_pending;
/* if we don't have such request (e.g. protocoll A) *checktheoldestrequestswhichisstillwaitingonitsepoch
* closing barrier ack. */ if (!req_peer)
req_peer = connection->req_not_net_done;
/* evaluate the oldest peer request only in one timer! */ if (req_peer && req_peer->device != device)
req_peer = NULL;
/* do we have something to evaluate? */ if (req_peer == NULL && req_write == NULL && req_read == NULL) goto out;
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.