1 #include "ceph_debug.h"
4 #include <linux/highmem.h>
6 #include <linux/pagemap.h>
7 #include <linux/slab.h>
8 #include <linux/uaccess.h>
11 #include "osd_client.h"
12 #include "messenger.h"
16 const static struct ceph_connection_operations osd_con_ops;
18 static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
21 * Implement client access to distributed object storage cluster.
23 * All data objects are stored within a cluster/cloud of OSDs, or
24 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
25 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
26 * remote daemons serving up and coordinating consistent and safe
29 * Cluster membership and the mapping of data objects onto storage devices
30 * are described by the osd map.
32 * We keep track of pending OSD requests (read, write), resubmit
33 * requests to different OSDs when the cluster topology/data layout
34 * change, or retry the affected requests when the communications
35 * channel with an OSD is reset.
39 * calculate the mapping of a file extent onto an object, and fill out the
40 * request accordingly. shorten extent as necessary if it crosses an
43 * fill osd op in request message.
45 static void calc_layout(struct ceph_osd_client *osdc,
46 struct ceph_vino vino, struct ceph_file_layout *layout,
48 struct ceph_osd_request *req)
50 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
51 struct ceph_osd_op *op = (void *)(reqhead + 1);
53 u64 objoff, objlen; /* extent in object */
56 reqhead->snapid = cpu_to_le64(vino.snap);
59 ceph_calc_file_object_mapping(layout, off, plen, &bno,
62 dout(" skipping last %llu, final file extent %llu~%llu\n",
63 orig_len - *plen, off, *plen);
65 sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
66 req->r_oid_len = strlen(req->r_oid);
68 op->extent.offset = cpu_to_le64(objoff);
69 op->extent.length = cpu_to_le64(objlen);
70 req->r_num_pages = calc_pages_for(off, *plen);
72 dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
73 req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
80 void ceph_osdc_release_request(struct kref *kref)
82 struct ceph_osd_request *req = container_of(kref,
83 struct ceph_osd_request,
87 ceph_msg_put(req->r_request);
89 ceph_msg_put(req->r_reply);
90 if (req->r_con_filling_pages) {
91 dout("release_request revoking pages %p from con %p\n",
92 req->r_pages, req->r_con_filling_pages);
93 ceph_con_revoke_pages(req->r_con_filling_pages,
95 ceph_con_put(req->r_con_filling_pages);
98 ceph_release_page_vector(req->r_pages,
100 ceph_put_snap_context(req->r_snapc);
102 mempool_free(req, req->r_osdc->req_mempool);
108 * build new request AND message, calculate layout, and adjust file
111 * if the file was recently truncated, we include information about its
112 * old and new size so that the object can be updated appropriately. (we
113 * avoid synchronously deleting truncated objects because it's slow.)
115 * if @do_sync, include a 'startsync' command so that the osd will flush
118 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
119 struct ceph_file_layout *layout,
120 struct ceph_vino vino,
122 int opcode, int flags,
123 struct ceph_snap_context *snapc,
127 struct timespec *mtime,
128 bool use_mempool, int num_reply)
130 struct ceph_osd_request *req;
131 struct ceph_msg *msg;
132 struct ceph_osd_request_head *head;
133 struct ceph_osd_op *op;
135 int do_trunc = truncate_seq && (off + *plen > truncate_size);
136 int num_op = 1 + do_sync + do_trunc;
137 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
142 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
143 memset(req, 0, sizeof(*req));
145 req = kzalloc(sizeof(*req), GFP_NOFS);
148 return ERR_PTR(-ENOMEM);
150 err = ceph_msgpool_resv(&osdc->msgpool_op_reply, num_reply);
152 ceph_osdc_put_request(req);
153 return ERR_PTR(-ENOMEM);
155 req->r_num_prealloc_reply = num_reply;
158 req->r_mempool = use_mempool;
159 kref_init(&req->r_kref);
160 init_completion(&req->r_completion);
161 init_completion(&req->r_safe_completion);
162 INIT_LIST_HEAD(&req->r_unsafe_item);
163 req->r_flags = flags;
165 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
167 /* create message; allow space for oid */
170 msg_size += sizeof(u64) * snapc->num_snaps;
172 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
174 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
176 ceph_msgpool_resv(&osdc->msgpool_op_reply, -num_reply);
177 ceph_osdc_put_request(req);
178 return ERR_PTR(PTR_ERR(msg));
180 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
181 memset(msg->front.iov_base, 0, msg->front.iov_len);
182 head = msg->front.iov_base;
183 op = (void *)(head + 1);
184 p = (void *)(op + num_op);
186 req->r_request = msg;
187 req->r_snapc = ceph_get_snap_context(snapc);
189 head->client_inc = cpu_to_le32(1); /* always, for now. */
190 head->flags = cpu_to_le32(flags);
191 if (flags & CEPH_OSD_FLAG_WRITE)
192 ceph_encode_timespec(&head->mtime, mtime);
193 head->num_ops = cpu_to_le16(num_op);
194 op->op = cpu_to_le16(opcode);
196 /* calculate max write size */
197 calc_layout(osdc, vino, layout, off, plen, req);
198 req->r_file_layout = *layout; /* keep a copy */
200 if (flags & CEPH_OSD_FLAG_WRITE) {
201 req->r_request->hdr.data_off = cpu_to_le16(off);
202 req->r_request->hdr.data_len = cpu_to_le32(*plen);
203 op->payload_len = cpu_to_le32(*plen);
207 head->object_len = cpu_to_le32(req->r_oid_len);
208 memcpy(p, req->r_oid, req->r_oid_len);
214 op->op = cpu_to_le16(opcode == CEPH_OSD_OP_READ ?
215 CEPH_OSD_OP_MASKTRUNC : CEPH_OSD_OP_SETTRUNC);
216 op->trunc.truncate_seq = cpu_to_le32(truncate_seq);
217 prevofs = le64_to_cpu((op-1)->extent.offset);
218 op->trunc.truncate_size = cpu_to_le64(truncate_size -
223 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
226 head->snap_seq = cpu_to_le64(snapc->seq);
227 head->num_snaps = cpu_to_le32(snapc->num_snaps);
228 for (i = 0; i < snapc->num_snaps; i++) {
229 put_unaligned_le64(snapc->snaps[i], p);
234 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
239 * We keep osd requests in an rbtree, sorted by ->r_tid.
241 static void __insert_request(struct ceph_osd_client *osdc,
242 struct ceph_osd_request *new)
244 struct rb_node **p = &osdc->requests.rb_node;
245 struct rb_node *parent = NULL;
246 struct ceph_osd_request *req = NULL;
250 req = rb_entry(parent, struct ceph_osd_request, r_node);
251 if (new->r_tid < req->r_tid)
253 else if (new->r_tid > req->r_tid)
259 rb_link_node(&new->r_node, parent, p);
260 rb_insert_color(&new->r_node, &osdc->requests);
263 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
266 struct ceph_osd_request *req;
267 struct rb_node *n = osdc->requests.rb_node;
270 req = rb_entry(n, struct ceph_osd_request, r_node);
271 if (tid < req->r_tid)
273 else if (tid > req->r_tid)
281 static struct ceph_osd_request *
282 __lookup_request_ge(struct ceph_osd_client *osdc,
285 struct ceph_osd_request *req;
286 struct rb_node *n = osdc->requests.rb_node;
289 req = rb_entry(n, struct ceph_osd_request, r_node);
290 if (tid < req->r_tid) {
294 } else if (tid > req->r_tid) {
305 * If the osd connection drops, we need to resubmit all requests.
307 static void osd_reset(struct ceph_connection *con)
309 struct ceph_osd *osd = con->private;
310 struct ceph_osd_client *osdc;
314 dout("osd_reset osd%d\n", osd->o_osd);
316 osd->o_incarnation++;
317 down_read(&osdc->map_sem);
318 kick_requests(osdc, osd);
319 up_read(&osdc->map_sem);
323 * Track open sessions with osds.
325 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
327 struct ceph_osd *osd;
329 osd = kzalloc(sizeof(*osd), GFP_NOFS);
333 atomic_set(&osd->o_ref, 1);
335 INIT_LIST_HEAD(&osd->o_requests);
336 osd->o_incarnation = 1;
338 ceph_con_init(osdc->client->msgr, &osd->o_con);
339 osd->o_con.private = osd;
340 osd->o_con.ops = &osd_con_ops;
341 osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
346 static struct ceph_osd *get_osd(struct ceph_osd *osd)
348 if (atomic_inc_not_zero(&osd->o_ref)) {
349 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
350 atomic_read(&osd->o_ref));
353 dout("get_osd %p FAIL\n", osd);
358 static void put_osd(struct ceph_osd *osd)
360 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
361 atomic_read(&osd->o_ref) - 1);
362 if (atomic_dec_and_test(&osd->o_ref))
367 * remove an osd from our map
369 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
371 dout("remove_osd %p\n", osd);
372 BUG_ON(!list_empty(&osd->o_requests));
373 rb_erase(&osd->o_node, &osdc->osds);
374 ceph_con_close(&osd->o_con);
381 static int reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
385 dout("reset_osd %p osd%d\n", osd, osd->o_osd);
386 if (list_empty(&osd->o_requests)) {
387 remove_osd(osdc, osd);
389 ceph_con_close(&osd->o_con);
390 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
391 osd->o_incarnation++;
396 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
398 struct rb_node **p = &osdc->osds.rb_node;
399 struct rb_node *parent = NULL;
400 struct ceph_osd *osd = NULL;
404 osd = rb_entry(parent, struct ceph_osd, o_node);
405 if (new->o_osd < osd->o_osd)
407 else if (new->o_osd > osd->o_osd)
413 rb_link_node(&new->o_node, parent, p);
414 rb_insert_color(&new->o_node, &osdc->osds);
417 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
419 struct ceph_osd *osd;
420 struct rb_node *n = osdc->osds.rb_node;
423 osd = rb_entry(n, struct ceph_osd, o_node);
426 else if (o > osd->o_osd)
436 * Register request, assign tid. If this is the first request, set up
439 static void register_request(struct ceph_osd_client *osdc,
440 struct ceph_osd_request *req)
442 mutex_lock(&osdc->request_mutex);
443 req->r_tid = ++osdc->last_tid;
444 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
446 dout("register_request %p tid %lld\n", req, req->r_tid);
447 __insert_request(osdc, req);
448 ceph_osdc_get_request(req);
449 osdc->num_requests++;
451 req->r_timeout_stamp =
452 jiffies + osdc->client->mount_args->osd_timeout*HZ;
454 if (osdc->num_requests == 1) {
455 osdc->timeout_tid = req->r_tid;
456 dout(" timeout on tid %llu at %lu\n", req->r_tid,
457 req->r_timeout_stamp);
458 schedule_delayed_work(&osdc->timeout_work,
459 round_jiffies_relative(req->r_timeout_stamp - jiffies));
461 mutex_unlock(&osdc->request_mutex);
465 * called under osdc->request_mutex
467 static void __unregister_request(struct ceph_osd_client *osdc,
468 struct ceph_osd_request *req)
470 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
471 rb_erase(&req->r_node, &osdc->requests);
472 osdc->num_requests--;
474 ceph_msgpool_resv(&osdc->msgpool_op_reply, -req->r_num_prealloc_reply);
477 /* make sure the original request isn't in flight. */
478 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
480 list_del_init(&req->r_osd_item);
481 if (list_empty(&req->r_osd->o_requests))
482 remove_osd(osdc, req->r_osd);
486 ceph_osdc_put_request(req);
488 if (req->r_tid == osdc->timeout_tid) {
489 if (osdc->num_requests == 0) {
490 dout("no requests, canceling timeout\n");
491 osdc->timeout_tid = 0;
492 cancel_delayed_work(&osdc->timeout_work);
494 req = rb_entry(rb_first(&osdc->requests),
495 struct ceph_osd_request, r_node);
496 osdc->timeout_tid = req->r_tid;
497 dout("rescheduled timeout on tid %llu at %lu\n",
498 req->r_tid, req->r_timeout_stamp);
499 schedule_delayed_work(&osdc->timeout_work,
500 round_jiffies_relative(req->r_timeout_stamp -
507 * Cancel a previously queued request message
509 static void __cancel_request(struct ceph_osd_request *req)
512 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
518 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
519 * (as needed), and set the request r_osd appropriately. If there is
520 * no up osd, set r_osd to NULL.
522 * Return 0 if unchanged, 1 if changed, or negative on error.
524 * Caller should hold map_sem for read and request_mutex.
526 static int __map_osds(struct ceph_osd_client *osdc,
527 struct ceph_osd_request *req)
529 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
533 struct ceph_osd *newosd = NULL;
535 dout("map_osds %p tid %lld\n", req, req->r_tid);
536 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
537 &req->r_file_layout, osdc->osdmap);
540 pgid = reqhead->layout.ol_pgid;
543 o = ceph_calc_pg_primary(osdc->osdmap, pgid);
545 if ((req->r_osd && req->r_osd->o_osd == o &&
546 req->r_sent >= req->r_osd->o_incarnation) ||
547 (req->r_osd == NULL && o == -1))
548 return 0; /* no change */
550 dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
551 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
552 req->r_osd ? req->r_osd->o_osd : -1);
555 __cancel_request(req);
556 list_del_init(&req->r_osd_item);
557 if (list_empty(&req->r_osd->o_requests)) {
558 /* try to re-use r_osd if possible */
559 newosd = get_osd(req->r_osd);
560 remove_osd(osdc, newosd);
565 req->r_osd = __lookup_osd(osdc, o);
566 if (!req->r_osd && o >= 0) {
572 req->r_osd = create_osd(osdc);
577 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
578 req->r_osd->o_osd = o;
579 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
580 __insert_osd(osdc, req->r_osd);
582 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
586 list_add(&req->r_osd_item, &req->r_osd->o_requests);
587 err = 1; /* osd changed */
596 * caller should hold map_sem (for read) and request_mutex
598 static int __send_request(struct ceph_osd_client *osdc,
599 struct ceph_osd_request *req)
601 struct ceph_osd_request_head *reqhead;
604 err = __map_osds(osdc, req);
607 if (req->r_osd == NULL) {
608 dout("send_request %p no up osds in pg\n", req);
609 ceph_monc_request_next_osdmap(&osdc->client->monc);
613 dout("send_request %p tid %llu to osd%d flags %d\n",
614 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
616 reqhead = req->r_request->front.iov_base;
617 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
618 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
619 reqhead->reassert_version = req->r_reassert_version;
621 req->r_timeout_stamp = jiffies+osdc->client->mount_args->osd_timeout*HZ;
623 ceph_msg_get(req->r_request); /* send consumes a ref */
624 ceph_con_send(&req->r_osd->o_con, req->r_request);
625 req->r_sent = req->r_osd->o_incarnation;
630 * Timeout callback, called every N seconds when 1 or more osd
631 * requests has been active for more than N seconds. When this
632 * happens, we ping all OSDs with requests who have timed out to
633 * ensure any communications channel reset is detected. Reset the
634 * request timeouts another N seconds in the future as we go.
635 * Reschedule the timeout event another N seconds in future (unless
636 * there are no open requests).
638 static void handle_timeout(struct work_struct *work)
640 struct ceph_osd_client *osdc =
641 container_of(work, struct ceph_osd_client, timeout_work.work);
642 struct ceph_osd_request *req;
643 struct ceph_osd *osd;
644 unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
645 unsigned long next_timeout = timeout + jiffies;
649 down_read(&osdc->map_sem);
651 ceph_monc_request_next_osdmap(&osdc->client->monc);
653 mutex_lock(&osdc->request_mutex);
654 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
655 req = rb_entry(p, struct ceph_osd_request, r_node);
660 dout("osdc resending prev failed %lld\n", req->r_tid);
661 err = __send_request(osdc, req);
663 dout("osdc failed again on %lld\n", req->r_tid);
665 req->r_resend = false;
669 for (p = rb_first(&osdc->osds); p; p = rb_next(p)) {
670 osd = rb_entry(p, struct ceph_osd, o_node);
671 if (list_empty(&osd->o_requests))
673 req = list_first_entry(&osd->o_requests,
674 struct ceph_osd_request, r_osd_item);
675 if (time_before(jiffies, req->r_timeout_stamp))
678 dout(" tid %llu (at least) timed out on osd%d\n",
679 req->r_tid, osd->o_osd);
680 req->r_timeout_stamp = next_timeout;
681 ceph_con_keepalive(&osd->o_con);
684 if (osdc->timeout_tid)
685 schedule_delayed_work(&osdc->timeout_work,
686 round_jiffies_relative(timeout));
688 mutex_unlock(&osdc->request_mutex);
690 up_read(&osdc->map_sem);
694 * handle osd op reply. either call the callback if it is specified,
695 * or do the completion to wake up the waiting thread.
697 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
698 struct ceph_connection *con)
700 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
701 struct ceph_osd_request *req;
703 int numops, object_len, flags;
705 tid = le64_to_cpu(msg->hdr.tid);
706 if (msg->front.iov_len < sizeof(*rhead))
708 numops = le32_to_cpu(rhead->num_ops);
709 object_len = le32_to_cpu(rhead->object_len);
710 if (msg->front.iov_len != sizeof(*rhead) + object_len +
711 numops * sizeof(struct ceph_osd_op))
713 dout("handle_reply %p tid %llu\n", msg, tid);
716 mutex_lock(&osdc->request_mutex);
717 req = __lookup_request(osdc, tid);
719 dout("handle_reply tid %llu dne\n", tid);
720 mutex_unlock(&osdc->request_mutex);
723 ceph_osdc_get_request(req);
724 flags = le32_to_cpu(rhead->flags);
727 * if this connection filled our pages, drop our reference now, to
728 * avoid a (safe but slower) revoke later.
730 if (req->r_con_filling_pages == con && req->r_pages == msg->pages) {
731 dout(" got pages, dropping con_filling_pages ref %p\n", con);
732 req->r_con_filling_pages = NULL;
738 * once we see the message has been received, we don't
739 * need a ref (which is only needed for revoking
742 ceph_msg_put(req->r_reply);
746 if (!req->r_got_reply) {
749 req->r_result = le32_to_cpu(rhead->result);
750 bytes = le32_to_cpu(msg->hdr.data_len);
751 dout("handle_reply result %d bytes %d\n", req->r_result,
753 if (req->r_result == 0)
754 req->r_result = bytes;
756 /* in case this is a write and we need to replay, */
757 req->r_reassert_version = rhead->reassert_version;
759 req->r_got_reply = 1;
760 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
761 dout("handle_reply tid %llu dup ack\n", tid);
762 mutex_unlock(&osdc->request_mutex);
766 dout("handle_reply tid %llu flags %d\n", tid, flags);
768 /* either this is a read, or we got the safe response */
769 if ((flags & CEPH_OSD_FLAG_ONDISK) ||
770 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
771 __unregister_request(osdc, req);
773 mutex_unlock(&osdc->request_mutex);
776 req->r_callback(req, msg);
778 complete(&req->r_completion);
780 if (flags & CEPH_OSD_FLAG_ONDISK) {
781 if (req->r_safe_callback)
782 req->r_safe_callback(req, msg);
783 complete(&req->r_safe_completion); /* fsync waiter */
787 ceph_osdc_put_request(req);
791 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
792 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
793 (int)sizeof(*rhead));
799 * Resubmit osd requests whose osd or osd address has changed. Request
800 * a new osd map if osds are down, or we are otherwise unable to determine
801 * how to direct a request.
803 * Close connections to down osds.
805 * If @who is specified, resubmit requests for that specific osd.
807 * Caller should hold map_sem for read and request_mutex.
809 static void kick_requests(struct ceph_osd_client *osdc,
810 struct ceph_osd *kickosd)
812 struct ceph_osd_request *req;
813 struct rb_node *p, *n;
817 dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
818 mutex_lock(&osdc->request_mutex);
820 for (p = rb_first(&osdc->osds); p; p = n) {
821 struct ceph_osd *osd =
822 rb_entry(p, struct ceph_osd, o_node);
825 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
826 memcmp(&osd->o_con.peer_addr,
827 ceph_osd_addr(osdc->osdmap,
829 sizeof(struct ceph_entity_addr)) != 0)
830 reset_osd(osdc, osd);
834 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
835 req = rb_entry(p, struct ceph_osd_request, r_node);
838 dout(" r_resend set on tid %llu\n", req->r_tid);
839 __cancel_request(req);
842 if (req->r_osd && kickosd == req->r_osd) {
843 __cancel_request(req);
847 err = __map_osds(osdc, req);
849 continue; /* no change */
852 * FIXME: really, we should set the request
853 * error and fail if this isn't a 'nofail'
854 * request, but that's a fair bit more
855 * complicated to do. So retry!
857 dout(" setting r_resend on %llu\n", req->r_tid);
858 req->r_resend = true;
861 if (req->r_osd == NULL) {
862 dout("tid %llu maps to no valid osd\n", req->r_tid);
863 needmap++; /* request a newer map */
868 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
870 req->r_flags |= CEPH_OSD_FLAG_RETRY;
871 err = __send_request(osdc, req);
873 dout(" setting r_resend on %llu\n", req->r_tid);
874 req->r_resend = true;
877 mutex_unlock(&osdc->request_mutex);
880 dout("%d requests for down osds, need new map\n", needmap);
881 ceph_monc_request_next_osdmap(&osdc->client->monc);
886 * Process updated osd map.
888 * The message contains any number of incremental and full maps, normally
889 * indicating some sort of topology change in the cluster. Kick requests
890 * off to different OSDs as needed.
892 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
894 void *p, *end, *next;
897 struct ceph_osdmap *newmap = NULL, *oldmap;
899 struct ceph_fsid fsid;
901 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
902 p = msg->front.iov_base;
903 end = p + msg->front.iov_len;
906 ceph_decode_need(&p, end, sizeof(fsid), bad);
907 ceph_decode_copy(&p, &fsid, sizeof(fsid));
908 if (ceph_check_fsid(osdc->client, &fsid) < 0)
911 down_write(&osdc->map_sem);
913 /* incremental maps */
914 ceph_decode_32_safe(&p, end, nr_maps, bad);
915 dout(" %d inc maps\n", nr_maps);
916 while (nr_maps > 0) {
917 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
918 epoch = ceph_decode_32(&p);
919 maplen = ceph_decode_32(&p);
920 ceph_decode_need(&p, end, maplen, bad);
922 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
923 dout("applying incremental map %u len %d\n",
925 newmap = osdmap_apply_incremental(&p, next,
928 if (IS_ERR(newmap)) {
929 err = PTR_ERR(newmap);
933 if (newmap != osdc->osdmap) {
934 ceph_osdmap_destroy(osdc->osdmap);
935 osdc->osdmap = newmap;
938 dout("ignoring incremental map %u len %d\n",
948 ceph_decode_32_safe(&p, end, nr_maps, bad);
949 dout(" %d full maps\n", nr_maps);
951 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
952 epoch = ceph_decode_32(&p);
953 maplen = ceph_decode_32(&p);
954 ceph_decode_need(&p, end, maplen, bad);
956 dout("skipping non-latest full map %u len %d\n",
958 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
959 dout("skipping full map %u len %d, "
960 "older than our %u\n", epoch, maplen,
961 osdc->osdmap->epoch);
963 dout("taking full map %u len %d\n", epoch, maplen);
964 newmap = osdmap_decode(&p, p+maplen);
965 if (IS_ERR(newmap)) {
966 err = PTR_ERR(newmap);
970 oldmap = osdc->osdmap;
971 osdc->osdmap = newmap;
973 ceph_osdmap_destroy(oldmap);
980 downgrade_write(&osdc->map_sem);
981 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
983 kick_requests(osdc, NULL);
984 up_read(&osdc->map_sem);
988 pr_err("osdc handle_map corrupt msg\n");
990 up_write(&osdc->map_sem);
996 * A read request prepares specific pages that data is to be read into.
997 * When a message is being read off the wire, we call prepare_pages to
999 * 0 = success, -1 failure.
1001 static int prepare_pages(struct ceph_connection *con, struct ceph_msg *m,
1004 struct ceph_osd *osd = con->private;
1005 struct ceph_osd_client *osdc;
1006 struct ceph_osd_request *req;
1009 int type = le16_to_cpu(m->hdr.type);
1015 dout("prepare_pages on msg %p want %d\n", m, want);
1016 if (unlikely(type != CEPH_MSG_OSD_OPREPLY))
1017 return -1; /* hmm! */
1019 tid = le64_to_cpu(m->hdr.tid);
1020 mutex_lock(&osdc->request_mutex);
1021 req = __lookup_request(osdc, tid);
1023 dout("prepare_pages unknown tid %llu\n", tid);
1026 dout("prepare_pages tid %llu has %d pages, want %d\n",
1027 tid, req->r_num_pages, want);
1028 if (unlikely(req->r_num_pages < want))
1031 if (req->r_con_filling_pages) {
1032 dout("revoking pages %p from old con %p\n", req->r_pages,
1033 req->r_con_filling_pages);
1034 ceph_con_revoke_pages(req->r_con_filling_pages, req->r_pages);
1035 ceph_con_put(req->r_con_filling_pages);
1037 req->r_con_filling_pages = ceph_con_get(con);
1038 req->r_reply = ceph_msg_get(m); /* for duration of read over socket */
1039 m->pages = req->r_pages;
1040 m->nr_pages = req->r_num_pages;
1041 ret = 0; /* success */
1043 mutex_unlock(&osdc->request_mutex);
1048 * Register request, send initial attempt.
1050 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1051 struct ceph_osd_request *req,
1056 req->r_request->pages = req->r_pages;
1057 req->r_request->nr_pages = req->r_num_pages;
1059 register_request(osdc, req);
1061 down_read(&osdc->map_sem);
1062 mutex_lock(&osdc->request_mutex);
1064 * a racing kick_requests() may have sent the message for us
1065 * while we dropped request_mutex above, so only send now if
1066 * the request still han't been touched yet.
1068 if (req->r_sent == 0) {
1069 rc = __send_request(osdc, req);
1072 dout("osdc_start_request failed send, "
1073 " marking %lld\n", req->r_tid);
1074 req->r_resend = true;
1077 __unregister_request(osdc, req);
1081 mutex_unlock(&osdc->request_mutex);
1082 up_read(&osdc->map_sem);
1087 * wait for a request to complete
1089 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1090 struct ceph_osd_request *req)
1094 rc = wait_for_completion_interruptible(&req->r_completion);
1096 mutex_lock(&osdc->request_mutex);
1097 __cancel_request(req);
1098 __unregister_request(osdc, req);
1099 mutex_unlock(&osdc->request_mutex);
1100 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1104 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1105 return req->r_result;
1109 * sync - wait for all in-flight requests to flush. avoid starvation.
1111 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1113 struct ceph_osd_request *req;
1114 u64 last_tid, next_tid = 0;
1116 mutex_lock(&osdc->request_mutex);
1117 last_tid = osdc->last_tid;
1119 req = __lookup_request_ge(osdc, next_tid);
1122 if (req->r_tid > last_tid)
1125 next_tid = req->r_tid + 1;
1126 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1129 ceph_osdc_get_request(req);
1130 mutex_unlock(&osdc->request_mutex);
1131 dout("sync waiting on tid %llu (last is %llu)\n",
1132 req->r_tid, last_tid);
1133 wait_for_completion(&req->r_safe_completion);
1134 mutex_lock(&osdc->request_mutex);
1135 ceph_osdc_put_request(req);
1137 mutex_unlock(&osdc->request_mutex);
1138 dout("sync done (thru tid %llu)\n", last_tid);
1144 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1149 osdc->client = client;
1150 osdc->osdmap = NULL;
1151 init_rwsem(&osdc->map_sem);
1152 init_completion(&osdc->map_waiters);
1153 osdc->last_requested_map = 0;
1154 mutex_init(&osdc->request_mutex);
1155 osdc->timeout_tid = 0;
1157 osdc->osds = RB_ROOT;
1158 osdc->requests = RB_ROOT;
1159 osdc->num_requests = 0;
1160 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1163 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1164 sizeof(struct ceph_osd_request));
1165 if (!osdc->req_mempool)
1168 err = ceph_msgpool_init(&osdc->msgpool_op, 4096, 10, true);
1171 err = ceph_msgpool_init(&osdc->msgpool_op_reply, 512, 0, false);
1177 ceph_msgpool_destroy(&osdc->msgpool_op);
1179 mempool_destroy(osdc->req_mempool);
1184 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1186 cancel_delayed_work_sync(&osdc->timeout_work);
1188 ceph_osdmap_destroy(osdc->osdmap);
1189 osdc->osdmap = NULL;
1191 mempool_destroy(osdc->req_mempool);
1192 ceph_msgpool_destroy(&osdc->msgpool_op);
1193 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1197 * Read some contiguous pages. If we cross a stripe boundary, shorten
1198 * *plen. Return number of bytes read, or error.
1200 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1201 struct ceph_vino vino, struct ceph_file_layout *layout,
1203 u32 truncate_seq, u64 truncate_size,
1204 struct page **pages, int num_pages)
1206 struct ceph_osd_request *req;
1209 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1210 vino.snap, off, *plen);
1211 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1212 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1213 NULL, 0, truncate_seq, truncate_size, NULL,
1216 return PTR_ERR(req);
1218 /* it may be a short read due to an object boundary */
1219 req->r_pages = pages;
1220 num_pages = calc_pages_for(off, *plen);
1221 req->r_num_pages = num_pages;
1223 dout("readpages final extent is %llu~%llu (%d pages)\n",
1224 off, *plen, req->r_num_pages);
1226 rc = ceph_osdc_start_request(osdc, req, false);
1228 rc = ceph_osdc_wait_request(osdc, req);
1230 ceph_osdc_put_request(req);
1231 dout("readpages result %d\n", rc);
1236 * do a synchronous write on N pages
1238 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1239 struct ceph_file_layout *layout,
1240 struct ceph_snap_context *snapc,
1242 u32 truncate_seq, u64 truncate_size,
1243 struct timespec *mtime,
1244 struct page **pages, int num_pages,
1245 int flags, int do_sync, bool nofail)
1247 struct ceph_osd_request *req;
1250 BUG_ON(vino.snap != CEPH_NOSNAP);
1251 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1253 flags | CEPH_OSD_FLAG_ONDISK |
1254 CEPH_OSD_FLAG_WRITE,
1256 truncate_seq, truncate_size, mtime,
1259 return PTR_ERR(req);
1261 /* it may be a short write due to an object boundary */
1262 req->r_pages = pages;
1263 req->r_num_pages = calc_pages_for(off, len);
1264 dout("writepages %llu~%llu (%d pages)\n", off, len,
1267 rc = ceph_osdc_start_request(osdc, req, nofail);
1269 rc = ceph_osdc_wait_request(osdc, req);
1271 ceph_osdc_put_request(req);
1274 dout("writepages result %d\n", rc);
1279 * handle incoming message
1281 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1283 struct ceph_osd *osd = con->private;
1284 struct ceph_osd_client *osdc;
1285 int type = le16_to_cpu(msg->hdr.type);
1292 case CEPH_MSG_OSD_MAP:
1293 ceph_osdc_handle_map(osdc, msg);
1295 case CEPH_MSG_OSD_OPREPLY:
1296 handle_reply(osdc, msg, con);
1300 pr_err("received unknown message type %d %s\n", type,
1301 ceph_msg_type_name(type));
1306 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1307 struct ceph_msg_header *hdr,
1310 struct ceph_osd *osd = con->private;
1311 struct ceph_osd_client *osdc = osd->o_osdc;
1312 int type = le16_to_cpu(hdr->type);
1313 int front = le32_to_cpu(hdr->front_len);
1318 case CEPH_MSG_OSD_OPREPLY:
1319 m = ceph_msgpool_get(&osdc->msgpool_op_reply, front);
1332 * Wrappers to refcount containing ceph_osd struct
1334 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1336 struct ceph_osd *osd = con->private;
1342 static void put_osd_con(struct ceph_connection *con)
1344 struct ceph_osd *osd = con->private;
1351 static int get_authorizer(struct ceph_connection *con,
1352 void **buf, int *len, int *proto,
1353 void **reply_buf, int *reply_len, int force_new)
1355 struct ceph_osd *o = con->private;
1356 struct ceph_osd_client *osdc = o->o_osdc;
1357 struct ceph_auth_client *ac = osdc->client->monc.auth;
1360 if (force_new && o->o_authorizer) {
1361 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1362 o->o_authorizer = NULL;
1364 if (o->o_authorizer == NULL) {
1365 ret = ac->ops->create_authorizer(
1366 ac, CEPH_ENTITY_TYPE_OSD,
1368 &o->o_authorizer_buf,
1369 &o->o_authorizer_buf_len,
1370 &o->o_authorizer_reply_buf,
1371 &o->o_authorizer_reply_buf_len);
1376 *proto = ac->protocol;
1377 *buf = o->o_authorizer_buf;
1378 *len = o->o_authorizer_buf_len;
1379 *reply_buf = o->o_authorizer_reply_buf;
1380 *reply_len = o->o_authorizer_reply_buf_len;
1385 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1387 struct ceph_osd *o = con->private;
1388 struct ceph_osd_client *osdc = o->o_osdc;
1389 struct ceph_auth_client *ac = osdc->client->monc.auth;
1391 return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1395 const static struct ceph_connection_operations osd_con_ops = {
1398 .dispatch = dispatch,
1399 .get_authorizer = get_authorizer,
1400 .verify_authorizer_reply = verify_authorizer_reply,
1401 .alloc_msg = alloc_msg,
1403 .prepare_pages = prepare_pages,