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 #define OSD_REPLY_RESERVE_FRONT_LEN 512
18 const static struct ceph_connection_operations osd_con_ops;
20 static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
23 * Implement client access to distributed object storage cluster.
25 * All data objects are stored within a cluster/cloud of OSDs, or
26 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
27 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
28 * remote daemons serving up and coordinating consistent and safe
31 * Cluster membership and the mapping of data objects onto storage devices
32 * are described by the osd map.
34 * We keep track of pending OSD requests (read, write), resubmit
35 * requests to different OSDs when the cluster topology/data layout
36 * change, or retry the affected requests when the communications
37 * channel with an OSD is reset.
41 * calculate the mapping of a file extent onto an object, and fill out the
42 * request accordingly. shorten extent as necessary if it crosses an
45 * fill osd op in request message.
47 static void calc_layout(struct ceph_osd_client *osdc,
48 struct ceph_vino vino, struct ceph_file_layout *layout,
50 struct ceph_osd_request *req)
52 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
53 struct ceph_osd_op *op = (void *)(reqhead + 1);
55 u64 objoff, objlen; /* extent in object */
58 reqhead->snapid = cpu_to_le64(vino.snap);
61 ceph_calc_file_object_mapping(layout, off, plen, &bno,
64 dout(" skipping last %llu, final file extent %llu~%llu\n",
65 orig_len - *plen, off, *plen);
67 sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
68 req->r_oid_len = strlen(req->r_oid);
70 op->extent.offset = cpu_to_le64(objoff);
71 op->extent.length = cpu_to_le64(objlen);
72 req->r_num_pages = calc_pages_for(off, *plen);
74 dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
75 req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
78 static void remove_replies(struct ceph_osd_request *req)
81 int max = ARRAY_SIZE(req->replies);
83 for (i=0; i<max; i++) {
85 ceph_msg_put(req->replies[i]);
92 void ceph_osdc_release_request(struct kref *kref)
94 struct ceph_osd_request *req = container_of(kref,
95 struct ceph_osd_request,
99 ceph_msg_put(req->r_request);
101 ceph_msg_put(req->r_reply);
103 if (req->r_con_filling_msg) {
104 dout("release_request revoking pages %p from con %p\n",
105 req->r_pages, req->r_con_filling_msg);
106 ceph_con_revoke_message(req->r_con_filling_msg,
108 ceph_con_put(req->r_con_filling_msg);
110 if (req->r_own_pages)
111 ceph_release_page_vector(req->r_pages,
113 ceph_put_snap_context(req->r_snapc);
115 mempool_free(req, req->r_osdc->req_mempool);
120 static int alloc_replies(struct ceph_osd_request *req, int num_reply)
123 int max = ARRAY_SIZE(req->replies);
125 BUG_ON(num_reply > max);
127 for (i=0; i<num_reply; i++) {
128 req->replies[i] = ceph_msg_new(0, OSD_REPLY_RESERVE_FRONT_LEN, 0, 0, NULL);
129 if (IS_ERR(req->replies[i])) {
131 int err = PTR_ERR(req->replies[i]);
132 for (j = 0; j<=i; j++) {
133 ceph_msg_put(req->replies[j]);
140 req->replies[i] = NULL;
148 static struct ceph_msg *__get_next_reply(struct ceph_connection *con,
149 struct ceph_osd_request *req,
152 struct ceph_msg *reply;
153 if (req->r_con_filling_msg) {
154 dout("revoking reply msg %p from old con %p\n", req->r_reply,
155 req->r_con_filling_msg);
156 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
157 ceph_con_put(req->r_con_filling_msg);
160 reply = req->replies[req->cur_reply];
161 if (!reply || front_len > OSD_REPLY_RESERVE_FRONT_LEN) {
162 /* maybe we can allocate it now? */
163 reply = ceph_msg_new(0, front_len, 0, 0, NULL);
164 if (!reply || IS_ERR(reply)) {
165 pr_err(" reply alloc failed, front_len=%d\n", front_len);
166 return ERR_PTR(-ENOMEM);
169 req->r_con_filling_msg = ceph_con_get(con);
170 req->r_reply = ceph_msg_get(reply); /* for duration of read over socket */
171 return ceph_msg_get(reply);
175 * build new request AND message, calculate layout, and adjust file
178 * if the file was recently truncated, we include information about its
179 * old and new size so that the object can be updated appropriately. (we
180 * avoid synchronously deleting truncated objects because it's slow.)
182 * if @do_sync, include a 'startsync' command so that the osd will flush
185 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
186 struct ceph_file_layout *layout,
187 struct ceph_vino vino,
189 int opcode, int flags,
190 struct ceph_snap_context *snapc,
194 struct timespec *mtime,
195 bool use_mempool, int num_reply)
197 struct ceph_osd_request *req;
198 struct ceph_msg *msg;
199 struct ceph_osd_request_head *head;
200 struct ceph_osd_op *op;
202 int num_op = 1 + do_sync;
203 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
207 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
208 memset(req, 0, sizeof(*req));
210 req = kzalloc(sizeof(*req), GFP_NOFS);
213 return ERR_PTR(-ENOMEM);
215 err = alloc_replies(req, num_reply);
217 ceph_osdc_put_request(req);
218 return ERR_PTR(-ENOMEM);
220 req->r_num_prealloc_reply = num_reply;
223 req->r_mempool = use_mempool;
224 kref_init(&req->r_kref);
225 init_completion(&req->r_completion);
226 init_completion(&req->r_safe_completion);
227 INIT_LIST_HEAD(&req->r_unsafe_item);
228 req->r_flags = flags;
230 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
232 /* create message; allow space for oid */
235 msg_size += sizeof(u64) * snapc->num_snaps;
237 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
239 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
241 ceph_osdc_put_request(req);
242 return ERR_PTR(PTR_ERR(msg));
244 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
245 memset(msg->front.iov_base, 0, msg->front.iov_len);
246 head = msg->front.iov_base;
247 op = (void *)(head + 1);
248 p = (void *)(op + num_op);
250 req->r_request = msg;
251 req->r_snapc = ceph_get_snap_context(snapc);
253 head->client_inc = cpu_to_le32(1); /* always, for now. */
254 head->flags = cpu_to_le32(flags);
255 if (flags & CEPH_OSD_FLAG_WRITE)
256 ceph_encode_timespec(&head->mtime, mtime);
257 head->num_ops = cpu_to_le16(num_op);
258 op->op = cpu_to_le16(opcode);
260 /* calculate max write size */
261 calc_layout(osdc, vino, layout, off, plen, req);
262 req->r_file_layout = *layout; /* keep a copy */
264 if (flags & CEPH_OSD_FLAG_WRITE) {
265 req->r_request->hdr.data_off = cpu_to_le16(off);
266 req->r_request->hdr.data_len = cpu_to_le32(*plen);
267 op->payload_len = cpu_to_le32(*plen);
269 op->extent.truncate_size = cpu_to_le64(truncate_size);
270 op->extent.truncate_seq = cpu_to_le32(truncate_seq);
273 head->object_len = cpu_to_le32(req->r_oid_len);
274 memcpy(p, req->r_oid, req->r_oid_len);
279 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
282 head->snap_seq = cpu_to_le64(snapc->seq);
283 head->num_snaps = cpu_to_le32(snapc->num_snaps);
284 for (i = 0; i < snapc->num_snaps; i++) {
285 put_unaligned_le64(snapc->snaps[i], p);
290 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
295 * We keep osd requests in an rbtree, sorted by ->r_tid.
297 static void __insert_request(struct ceph_osd_client *osdc,
298 struct ceph_osd_request *new)
300 struct rb_node **p = &osdc->requests.rb_node;
301 struct rb_node *parent = NULL;
302 struct ceph_osd_request *req = NULL;
306 req = rb_entry(parent, struct ceph_osd_request, r_node);
307 if (new->r_tid < req->r_tid)
309 else if (new->r_tid > req->r_tid)
315 rb_link_node(&new->r_node, parent, p);
316 rb_insert_color(&new->r_node, &osdc->requests);
319 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
322 struct ceph_osd_request *req;
323 struct rb_node *n = osdc->requests.rb_node;
326 req = rb_entry(n, struct ceph_osd_request, r_node);
327 if (tid < req->r_tid)
329 else if (tid > req->r_tid)
337 static struct ceph_osd_request *
338 __lookup_request_ge(struct ceph_osd_client *osdc,
341 struct ceph_osd_request *req;
342 struct rb_node *n = osdc->requests.rb_node;
345 req = rb_entry(n, struct ceph_osd_request, r_node);
346 if (tid < req->r_tid) {
350 } else if (tid > req->r_tid) {
361 * If the osd connection drops, we need to resubmit all requests.
363 static void osd_reset(struct ceph_connection *con)
365 struct ceph_osd *osd = con->private;
366 struct ceph_osd_client *osdc;
370 dout("osd_reset osd%d\n", osd->o_osd);
372 osd->o_incarnation++;
373 down_read(&osdc->map_sem);
374 kick_requests(osdc, osd);
375 up_read(&osdc->map_sem);
379 * Track open sessions with osds.
381 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
383 struct ceph_osd *osd;
385 osd = kzalloc(sizeof(*osd), GFP_NOFS);
389 atomic_set(&osd->o_ref, 1);
391 INIT_LIST_HEAD(&osd->o_requests);
392 osd->o_incarnation = 1;
394 ceph_con_init(osdc->client->msgr, &osd->o_con);
395 osd->o_con.private = osd;
396 osd->o_con.ops = &osd_con_ops;
397 osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
402 static struct ceph_osd *get_osd(struct ceph_osd *osd)
404 if (atomic_inc_not_zero(&osd->o_ref)) {
405 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
406 atomic_read(&osd->o_ref));
409 dout("get_osd %p FAIL\n", osd);
414 static void put_osd(struct ceph_osd *osd)
416 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
417 atomic_read(&osd->o_ref) - 1);
418 if (atomic_dec_and_test(&osd->o_ref))
423 * remove an osd from our map
425 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
427 dout("remove_osd %p\n", osd);
428 BUG_ON(!list_empty(&osd->o_requests));
429 rb_erase(&osd->o_node, &osdc->osds);
430 ceph_con_close(&osd->o_con);
437 static int reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
441 dout("reset_osd %p osd%d\n", osd, osd->o_osd);
442 if (list_empty(&osd->o_requests)) {
443 remove_osd(osdc, osd);
445 ceph_con_close(&osd->o_con);
446 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
447 osd->o_incarnation++;
452 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
454 struct rb_node **p = &osdc->osds.rb_node;
455 struct rb_node *parent = NULL;
456 struct ceph_osd *osd = NULL;
460 osd = rb_entry(parent, struct ceph_osd, o_node);
461 if (new->o_osd < osd->o_osd)
463 else if (new->o_osd > osd->o_osd)
469 rb_link_node(&new->o_node, parent, p);
470 rb_insert_color(&new->o_node, &osdc->osds);
473 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
475 struct ceph_osd *osd;
476 struct rb_node *n = osdc->osds.rb_node;
479 osd = rb_entry(n, struct ceph_osd, o_node);
482 else if (o > osd->o_osd)
492 * Register request, assign tid. If this is the first request, set up
495 static void register_request(struct ceph_osd_client *osdc,
496 struct ceph_osd_request *req)
498 mutex_lock(&osdc->request_mutex);
499 req->r_tid = ++osdc->last_tid;
500 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
502 dout("register_request %p tid %lld\n", req, req->r_tid);
503 __insert_request(osdc, req);
504 ceph_osdc_get_request(req);
505 osdc->num_requests++;
507 req->r_timeout_stamp =
508 jiffies + osdc->client->mount_args->osd_timeout*HZ;
510 if (osdc->num_requests == 1) {
511 osdc->timeout_tid = req->r_tid;
512 dout(" timeout on tid %llu at %lu\n", req->r_tid,
513 req->r_timeout_stamp);
514 schedule_delayed_work(&osdc->timeout_work,
515 round_jiffies_relative(req->r_timeout_stamp - jiffies));
517 mutex_unlock(&osdc->request_mutex);
521 * called under osdc->request_mutex
523 static void __unregister_request(struct ceph_osd_client *osdc,
524 struct ceph_osd_request *req)
526 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
527 rb_erase(&req->r_node, &osdc->requests);
528 osdc->num_requests--;
531 /* make sure the original request isn't in flight. */
532 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
534 list_del_init(&req->r_osd_item);
535 if (list_empty(&req->r_osd->o_requests))
536 remove_osd(osdc, req->r_osd);
540 ceph_osdc_put_request(req);
542 if (req->r_tid == osdc->timeout_tid) {
543 if (osdc->num_requests == 0) {
544 dout("no requests, canceling timeout\n");
545 osdc->timeout_tid = 0;
546 cancel_delayed_work(&osdc->timeout_work);
548 req = rb_entry(rb_first(&osdc->requests),
549 struct ceph_osd_request, r_node);
550 osdc->timeout_tid = req->r_tid;
551 dout("rescheduled timeout on tid %llu at %lu\n",
552 req->r_tid, req->r_timeout_stamp);
553 schedule_delayed_work(&osdc->timeout_work,
554 round_jiffies_relative(req->r_timeout_stamp -
561 * Cancel a previously queued request message
563 static void __cancel_request(struct ceph_osd_request *req)
566 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
572 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
573 * (as needed), and set the request r_osd appropriately. If there is
574 * no up osd, set r_osd to NULL.
576 * Return 0 if unchanged, 1 if changed, or negative on error.
578 * Caller should hold map_sem for read and request_mutex.
580 static int __map_osds(struct ceph_osd_client *osdc,
581 struct ceph_osd_request *req)
583 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
587 struct ceph_osd *newosd = NULL;
589 dout("map_osds %p tid %lld\n", req, req->r_tid);
590 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
591 &req->r_file_layout, osdc->osdmap);
594 pgid = reqhead->layout.ol_pgid;
597 o = ceph_calc_pg_primary(osdc->osdmap, pgid);
599 if ((req->r_osd && req->r_osd->o_osd == o &&
600 req->r_sent >= req->r_osd->o_incarnation) ||
601 (req->r_osd == NULL && o == -1))
602 return 0; /* no change */
604 dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
605 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
606 req->r_osd ? req->r_osd->o_osd : -1);
609 __cancel_request(req);
610 list_del_init(&req->r_osd_item);
611 if (list_empty(&req->r_osd->o_requests)) {
612 /* try to re-use r_osd if possible */
613 newosd = get_osd(req->r_osd);
614 remove_osd(osdc, newosd);
619 req->r_osd = __lookup_osd(osdc, o);
620 if (!req->r_osd && o >= 0) {
626 req->r_osd = create_osd(osdc);
631 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
632 req->r_osd->o_osd = o;
633 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
634 __insert_osd(osdc, req->r_osd);
636 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
640 list_add(&req->r_osd_item, &req->r_osd->o_requests);
641 err = 1; /* osd changed */
650 * caller should hold map_sem (for read) and request_mutex
652 static int __send_request(struct ceph_osd_client *osdc,
653 struct ceph_osd_request *req)
655 struct ceph_osd_request_head *reqhead;
658 err = __map_osds(osdc, req);
661 if (req->r_osd == NULL) {
662 dout("send_request %p no up osds in pg\n", req);
663 ceph_monc_request_next_osdmap(&osdc->client->monc);
667 dout("send_request %p tid %llu to osd%d flags %d\n",
668 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
670 reqhead = req->r_request->front.iov_base;
671 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
672 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
673 reqhead->reassert_version = req->r_reassert_version;
675 req->r_timeout_stamp = jiffies+osdc->client->mount_args->osd_timeout*HZ;
677 ceph_msg_get(req->r_request); /* send consumes a ref */
678 ceph_con_send(&req->r_osd->o_con, req->r_request);
679 req->r_sent = req->r_osd->o_incarnation;
684 * Timeout callback, called every N seconds when 1 or more osd
685 * requests has been active for more than N seconds. When this
686 * happens, we ping all OSDs with requests who have timed out to
687 * ensure any communications channel reset is detected. Reset the
688 * request timeouts another N seconds in the future as we go.
689 * Reschedule the timeout event another N seconds in future (unless
690 * there are no open requests).
692 static void handle_timeout(struct work_struct *work)
694 struct ceph_osd_client *osdc =
695 container_of(work, struct ceph_osd_client, timeout_work.work);
696 struct ceph_osd_request *req;
697 struct ceph_osd *osd;
698 unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
699 unsigned long next_timeout = timeout + jiffies;
703 down_read(&osdc->map_sem);
705 ceph_monc_request_next_osdmap(&osdc->client->monc);
707 mutex_lock(&osdc->request_mutex);
708 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
709 req = rb_entry(p, struct ceph_osd_request, r_node);
714 dout("osdc resending prev failed %lld\n", req->r_tid);
715 err = __send_request(osdc, req);
717 dout("osdc failed again on %lld\n", req->r_tid);
719 req->r_resend = false;
723 for (p = rb_first(&osdc->osds); p; p = rb_next(p)) {
724 osd = rb_entry(p, struct ceph_osd, o_node);
725 if (list_empty(&osd->o_requests))
727 req = list_first_entry(&osd->o_requests,
728 struct ceph_osd_request, r_osd_item);
729 if (time_before(jiffies, req->r_timeout_stamp))
732 dout(" tid %llu (at least) timed out on osd%d\n",
733 req->r_tid, osd->o_osd);
734 req->r_timeout_stamp = next_timeout;
735 ceph_con_keepalive(&osd->o_con);
738 if (osdc->timeout_tid)
739 schedule_delayed_work(&osdc->timeout_work,
740 round_jiffies_relative(timeout));
742 mutex_unlock(&osdc->request_mutex);
744 up_read(&osdc->map_sem);
748 * handle osd op reply. either call the callback if it is specified,
749 * or do the completion to wake up the waiting thread.
751 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
752 struct ceph_connection *con)
754 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
755 struct ceph_osd_request *req;
757 int numops, object_len, flags;
759 tid = le64_to_cpu(msg->hdr.tid);
760 if (msg->front.iov_len < sizeof(*rhead))
762 numops = le32_to_cpu(rhead->num_ops);
763 object_len = le32_to_cpu(rhead->object_len);
764 if (msg->front.iov_len != sizeof(*rhead) + object_len +
765 numops * sizeof(struct ceph_osd_op))
767 dout("handle_reply %p tid %llu\n", msg, tid);
770 mutex_lock(&osdc->request_mutex);
771 req = __lookup_request(osdc, tid);
773 dout("handle_reply tid %llu dne\n", tid);
774 mutex_unlock(&osdc->request_mutex);
777 ceph_osdc_get_request(req);
778 flags = le32_to_cpu(rhead->flags);
781 * if this connection filled our message, drop our reference now, to
782 * avoid a (safe but slower) revoke later.
784 if (req->r_con_filling_msg == con && req->r_reply == msg) {
785 dout(" got pages, dropping con_filling_msg ref %p\n", con);
786 req->r_con_filling_msg = NULL;
792 * once we see the message has been received, we don't
793 * need a ref (which is only needed for revoking
796 ceph_msg_put(req->r_reply);
800 if (!req->r_got_reply) {
803 req->r_result = le32_to_cpu(rhead->result);
804 bytes = le32_to_cpu(msg->hdr.data_len);
805 dout("handle_reply result %d bytes %d\n", req->r_result,
807 if (req->r_result == 0)
808 req->r_result = bytes;
810 /* in case this is a write and we need to replay, */
811 req->r_reassert_version = rhead->reassert_version;
813 req->r_got_reply = 1;
814 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
815 dout("handle_reply tid %llu dup ack\n", tid);
816 mutex_unlock(&osdc->request_mutex);
820 dout("handle_reply tid %llu flags %d\n", tid, flags);
822 /* either this is a read, or we got the safe response */
823 if ((flags & CEPH_OSD_FLAG_ONDISK) ||
824 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
825 __unregister_request(osdc, req);
827 mutex_unlock(&osdc->request_mutex);
830 req->r_callback(req, msg);
832 complete(&req->r_completion);
834 if (flags & CEPH_OSD_FLAG_ONDISK) {
835 if (req->r_safe_callback)
836 req->r_safe_callback(req, msg);
837 complete(&req->r_safe_completion); /* fsync waiter */
841 ceph_osdc_put_request(req);
845 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
846 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
847 (int)sizeof(*rhead));
853 * Resubmit osd requests whose osd or osd address has changed. Request
854 * a new osd map if osds are down, or we are otherwise unable to determine
855 * how to direct a request.
857 * Close connections to down osds.
859 * If @who is specified, resubmit requests for that specific osd.
861 * Caller should hold map_sem for read and request_mutex.
863 static void kick_requests(struct ceph_osd_client *osdc,
864 struct ceph_osd *kickosd)
866 struct ceph_osd_request *req;
867 struct rb_node *p, *n;
871 dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
872 mutex_lock(&osdc->request_mutex);
874 for (p = rb_first(&osdc->osds); p; p = n) {
875 struct ceph_osd *osd =
876 rb_entry(p, struct ceph_osd, o_node);
879 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
880 memcmp(&osd->o_con.peer_addr,
881 ceph_osd_addr(osdc->osdmap,
883 sizeof(struct ceph_entity_addr)) != 0)
884 reset_osd(osdc, osd);
888 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
889 req = rb_entry(p, struct ceph_osd_request, r_node);
892 dout(" r_resend set on tid %llu\n", req->r_tid);
893 __cancel_request(req);
896 if (req->r_osd && kickosd == req->r_osd) {
897 __cancel_request(req);
901 err = __map_osds(osdc, req);
903 continue; /* no change */
906 * FIXME: really, we should set the request
907 * error and fail if this isn't a 'nofail'
908 * request, but that's a fair bit more
909 * complicated to do. So retry!
911 dout(" setting r_resend on %llu\n", req->r_tid);
912 req->r_resend = true;
915 if (req->r_osd == NULL) {
916 dout("tid %llu maps to no valid osd\n", req->r_tid);
917 needmap++; /* request a newer map */
922 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
924 req->r_flags |= CEPH_OSD_FLAG_RETRY;
925 err = __send_request(osdc, req);
927 dout(" setting r_resend on %llu\n", req->r_tid);
928 req->r_resend = true;
931 mutex_unlock(&osdc->request_mutex);
934 dout("%d requests for down osds, need new map\n", needmap);
935 ceph_monc_request_next_osdmap(&osdc->client->monc);
940 * Process updated osd map.
942 * The message contains any number of incremental and full maps, normally
943 * indicating some sort of topology change in the cluster. Kick requests
944 * off to different OSDs as needed.
946 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
948 void *p, *end, *next;
951 struct ceph_osdmap *newmap = NULL, *oldmap;
953 struct ceph_fsid fsid;
955 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
956 p = msg->front.iov_base;
957 end = p + msg->front.iov_len;
960 ceph_decode_need(&p, end, sizeof(fsid), bad);
961 ceph_decode_copy(&p, &fsid, sizeof(fsid));
962 if (ceph_check_fsid(osdc->client, &fsid) < 0)
965 down_write(&osdc->map_sem);
967 /* incremental maps */
968 ceph_decode_32_safe(&p, end, nr_maps, bad);
969 dout(" %d inc maps\n", nr_maps);
970 while (nr_maps > 0) {
971 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
972 epoch = ceph_decode_32(&p);
973 maplen = ceph_decode_32(&p);
974 ceph_decode_need(&p, end, maplen, bad);
976 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
977 dout("applying incremental map %u len %d\n",
979 newmap = osdmap_apply_incremental(&p, next,
982 if (IS_ERR(newmap)) {
983 err = PTR_ERR(newmap);
987 if (newmap != osdc->osdmap) {
988 ceph_osdmap_destroy(osdc->osdmap);
989 osdc->osdmap = newmap;
992 dout("ignoring incremental map %u len %d\n",
1002 ceph_decode_32_safe(&p, end, nr_maps, bad);
1003 dout(" %d full maps\n", nr_maps);
1005 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1006 epoch = ceph_decode_32(&p);
1007 maplen = ceph_decode_32(&p);
1008 ceph_decode_need(&p, end, maplen, bad);
1010 dout("skipping non-latest full map %u len %d\n",
1012 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1013 dout("skipping full map %u len %d, "
1014 "older than our %u\n", epoch, maplen,
1015 osdc->osdmap->epoch);
1017 dout("taking full map %u len %d\n", epoch, maplen);
1018 newmap = osdmap_decode(&p, p+maplen);
1019 if (IS_ERR(newmap)) {
1020 err = PTR_ERR(newmap);
1024 oldmap = osdc->osdmap;
1025 osdc->osdmap = newmap;
1027 ceph_osdmap_destroy(oldmap);
1034 downgrade_write(&osdc->map_sem);
1035 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1037 kick_requests(osdc, NULL);
1038 up_read(&osdc->map_sem);
1042 pr_err("osdc handle_map corrupt msg\n");
1044 up_write(&osdc->map_sem);
1050 * A read request prepares specific pages that data is to be read into.
1051 * When a message is being read off the wire, we call prepare_pages to
1053 * 0 = success, -1 failure.
1055 static int __prepare_pages(struct ceph_connection *con,
1056 struct ceph_msg_header *hdr,
1057 struct ceph_osd_request *req,
1061 struct ceph_osd *osd = con->private;
1062 struct ceph_osd_client *osdc;
1064 int data_len = le32_to_cpu(hdr->data_len);
1065 unsigned data_off = le16_to_cpu(hdr->data_off);
1067 int want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1074 dout("__prepare_pages on msg %p tid %llu, has %d pages, want %d\n", m,
1075 tid, req->r_num_pages, want);
1076 if (unlikely(req->r_num_pages < want))
1078 m->pages = req->r_pages;
1079 m->nr_pages = req->r_num_pages;
1080 ret = 0; /* success */
1082 BUG_ON(ret < 0 || m->nr_pages < want);
1088 * Register request, send initial attempt.
1090 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1091 struct ceph_osd_request *req,
1096 req->r_request->pages = req->r_pages;
1097 req->r_request->nr_pages = req->r_num_pages;
1099 register_request(osdc, req);
1101 down_read(&osdc->map_sem);
1102 mutex_lock(&osdc->request_mutex);
1104 * a racing kick_requests() may have sent the message for us
1105 * while we dropped request_mutex above, so only send now if
1106 * the request still han't been touched yet.
1108 if (req->r_sent == 0) {
1109 rc = __send_request(osdc, req);
1112 dout("osdc_start_request failed send, "
1113 " marking %lld\n", req->r_tid);
1114 req->r_resend = true;
1117 __unregister_request(osdc, req);
1121 mutex_unlock(&osdc->request_mutex);
1122 up_read(&osdc->map_sem);
1127 * wait for a request to complete
1129 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1130 struct ceph_osd_request *req)
1134 rc = wait_for_completion_interruptible(&req->r_completion);
1136 mutex_lock(&osdc->request_mutex);
1137 __cancel_request(req);
1138 __unregister_request(osdc, req);
1139 mutex_unlock(&osdc->request_mutex);
1140 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1144 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1145 return req->r_result;
1149 * sync - wait for all in-flight requests to flush. avoid starvation.
1151 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1153 struct ceph_osd_request *req;
1154 u64 last_tid, next_tid = 0;
1156 mutex_lock(&osdc->request_mutex);
1157 last_tid = osdc->last_tid;
1159 req = __lookup_request_ge(osdc, next_tid);
1162 if (req->r_tid > last_tid)
1165 next_tid = req->r_tid + 1;
1166 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1169 ceph_osdc_get_request(req);
1170 mutex_unlock(&osdc->request_mutex);
1171 dout("sync waiting on tid %llu (last is %llu)\n",
1172 req->r_tid, last_tid);
1173 wait_for_completion(&req->r_safe_completion);
1174 mutex_lock(&osdc->request_mutex);
1175 ceph_osdc_put_request(req);
1177 mutex_unlock(&osdc->request_mutex);
1178 dout("sync done (thru tid %llu)\n", last_tid);
1184 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1189 osdc->client = client;
1190 osdc->osdmap = NULL;
1191 init_rwsem(&osdc->map_sem);
1192 init_completion(&osdc->map_waiters);
1193 osdc->last_requested_map = 0;
1194 mutex_init(&osdc->request_mutex);
1195 osdc->timeout_tid = 0;
1197 osdc->osds = RB_ROOT;
1198 osdc->requests = RB_ROOT;
1199 osdc->num_requests = 0;
1200 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1203 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1204 sizeof(struct ceph_osd_request));
1205 if (!osdc->req_mempool)
1208 err = ceph_msgpool_init(&osdc->msgpool_op, 4096, 10, true);
1214 mempool_destroy(osdc->req_mempool);
1219 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1221 cancel_delayed_work_sync(&osdc->timeout_work);
1223 ceph_osdmap_destroy(osdc->osdmap);
1224 osdc->osdmap = NULL;
1226 mempool_destroy(osdc->req_mempool);
1227 ceph_msgpool_destroy(&osdc->msgpool_op);
1231 * Read some contiguous pages. If we cross a stripe boundary, shorten
1232 * *plen. Return number of bytes read, or error.
1234 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1235 struct ceph_vino vino, struct ceph_file_layout *layout,
1237 u32 truncate_seq, u64 truncate_size,
1238 struct page **pages, int num_pages)
1240 struct ceph_osd_request *req;
1243 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1244 vino.snap, off, *plen);
1245 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1246 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1247 NULL, 0, truncate_seq, truncate_size, NULL,
1250 return PTR_ERR(req);
1252 /* it may be a short read due to an object boundary */
1253 req->r_pages = pages;
1254 num_pages = calc_pages_for(off, *plen);
1255 req->r_num_pages = num_pages;
1257 dout("readpages final extent is %llu~%llu (%d pages)\n",
1258 off, *plen, req->r_num_pages);
1260 rc = ceph_osdc_start_request(osdc, req, false);
1262 rc = ceph_osdc_wait_request(osdc, req);
1264 ceph_osdc_put_request(req);
1265 dout("readpages result %d\n", rc);
1270 * do a synchronous write on N pages
1272 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1273 struct ceph_file_layout *layout,
1274 struct ceph_snap_context *snapc,
1276 u32 truncate_seq, u64 truncate_size,
1277 struct timespec *mtime,
1278 struct page **pages, int num_pages,
1279 int flags, int do_sync, bool nofail)
1281 struct ceph_osd_request *req;
1284 BUG_ON(vino.snap != CEPH_NOSNAP);
1285 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1287 flags | CEPH_OSD_FLAG_ONDISK |
1288 CEPH_OSD_FLAG_WRITE,
1290 truncate_seq, truncate_size, mtime,
1293 return PTR_ERR(req);
1295 /* it may be a short write due to an object boundary */
1296 req->r_pages = pages;
1297 req->r_num_pages = calc_pages_for(off, len);
1298 dout("writepages %llu~%llu (%d pages)\n", off, len,
1301 rc = ceph_osdc_start_request(osdc, req, nofail);
1303 rc = ceph_osdc_wait_request(osdc, req);
1305 ceph_osdc_put_request(req);
1308 dout("writepages result %d\n", rc);
1313 * handle incoming message
1315 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1317 struct ceph_osd *osd = con->private;
1318 struct ceph_osd_client *osdc;
1319 int type = le16_to_cpu(msg->hdr.type);
1326 case CEPH_MSG_OSD_MAP:
1327 ceph_osdc_handle_map(osdc, msg);
1329 case CEPH_MSG_OSD_OPREPLY:
1330 handle_reply(osdc, msg, con);
1334 pr_err("received unknown message type %d %s\n", type,
1335 ceph_msg_type_name(type));
1340 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1341 struct ceph_msg_header *hdr,
1344 struct ceph_osd *osd = con->private;
1345 struct ceph_osd_client *osdc = osd->o_osdc;
1346 int type = le16_to_cpu(hdr->type);
1347 int front = le32_to_cpu(hdr->front_len);
1348 int data_len = le32_to_cpu(hdr->data_len);
1350 struct ceph_osd_request *req;
1355 if (type != CEPH_MSG_OSD_OPREPLY)
1358 tid = le64_to_cpu(hdr->tid);
1359 mutex_lock(&osdc->request_mutex);
1360 req = __lookup_request(osdc, tid);
1364 dout("alloc_msg unknown tid %llu\n", tid);
1367 m = __get_next_reply(con, req, front);
1368 if (!m || IS_ERR(m)) {
1374 err = __prepare_pages(con, hdr, req, tid, m);
1383 mutex_unlock(&osdc->request_mutex);
1389 * Wrappers to refcount containing ceph_osd struct
1391 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1393 struct ceph_osd *osd = con->private;
1399 static void put_osd_con(struct ceph_connection *con)
1401 struct ceph_osd *osd = con->private;
1408 static int get_authorizer(struct ceph_connection *con,
1409 void **buf, int *len, int *proto,
1410 void **reply_buf, int *reply_len, int force_new)
1412 struct ceph_osd *o = con->private;
1413 struct ceph_osd_client *osdc = o->o_osdc;
1414 struct ceph_auth_client *ac = osdc->client->monc.auth;
1417 if (force_new && o->o_authorizer) {
1418 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1419 o->o_authorizer = NULL;
1421 if (o->o_authorizer == NULL) {
1422 ret = ac->ops->create_authorizer(
1423 ac, CEPH_ENTITY_TYPE_OSD,
1425 &o->o_authorizer_buf,
1426 &o->o_authorizer_buf_len,
1427 &o->o_authorizer_reply_buf,
1428 &o->o_authorizer_reply_buf_len);
1433 *proto = ac->protocol;
1434 *buf = o->o_authorizer_buf;
1435 *len = o->o_authorizer_buf_len;
1436 *reply_buf = o->o_authorizer_reply_buf;
1437 *reply_len = o->o_authorizer_reply_buf_len;
1442 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1444 struct ceph_osd *o = con->private;
1445 struct ceph_osd_client *osdc = o->o_osdc;
1446 struct ceph_auth_client *ac = osdc->client->monc.auth;
1448 return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1452 const static struct ceph_connection_operations osd_con_ops = {
1455 .dispatch = dispatch,
1456 .get_authorizer = get_authorizer,
1457 .verify_authorizer_reply = verify_authorizer_reply,
1458 .alloc_msg = alloc_msg,