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);
91 ceph_release_page_vector(req->r_pages,
93 ceph_put_snap_context(req->r_snapc);
95 mempool_free(req, req->r_osdc->req_mempool);
101 * build new request AND message, calculate layout, and adjust file
104 * if the file was recently truncated, we include information about its
105 * old and new size so that the object can be updated appropriately. (we
106 * avoid synchronously deleting truncated objects because it's slow.)
108 * if @do_sync, include a 'startsync' command so that the osd will flush
111 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
112 struct ceph_file_layout *layout,
113 struct ceph_vino vino,
115 int opcode, int flags,
116 struct ceph_snap_context *snapc,
120 struct timespec *mtime,
121 bool use_mempool, int num_reply)
123 struct ceph_osd_request *req;
124 struct ceph_msg *msg;
125 struct ceph_osd_request_head *head;
126 struct ceph_osd_op *op;
128 int do_trunc = truncate_seq && (off + *plen > truncate_size);
129 int num_op = 1 + do_sync + do_trunc;
130 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
135 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
136 memset(req, 0, sizeof(*req));
138 req = kzalloc(sizeof(*req), GFP_NOFS);
141 return ERR_PTR(-ENOMEM);
143 err = ceph_msgpool_resv(&osdc->msgpool_op_reply, num_reply);
145 ceph_osdc_put_request(req);
146 return ERR_PTR(-ENOMEM);
148 req->r_num_prealloc_reply = num_reply;
151 req->r_mempool = use_mempool;
152 kref_init(&req->r_kref);
153 init_completion(&req->r_completion);
154 init_completion(&req->r_safe_completion);
155 INIT_LIST_HEAD(&req->r_unsafe_item);
156 req->r_flags = flags;
158 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
160 /* create message; allow space for oid */
163 msg_size += sizeof(u64) * snapc->num_snaps;
165 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
167 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
169 ceph_msgpool_resv(&osdc->msgpool_op_reply, -num_reply);
170 ceph_osdc_put_request(req);
171 return ERR_PTR(PTR_ERR(msg));
173 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
174 memset(msg->front.iov_base, 0, msg->front.iov_len);
175 head = msg->front.iov_base;
176 op = (void *)(head + 1);
177 p = (void *)(op + num_op);
179 req->r_request = msg;
180 req->r_snapc = ceph_get_snap_context(snapc);
182 head->client_inc = cpu_to_le32(1); /* always, for now. */
183 head->flags = cpu_to_le32(flags);
184 if (flags & CEPH_OSD_FLAG_WRITE)
185 ceph_encode_timespec(&head->mtime, mtime);
186 head->num_ops = cpu_to_le16(num_op);
187 op->op = cpu_to_le16(opcode);
189 /* calculate max write size */
190 calc_layout(osdc, vino, layout, off, plen, req);
191 req->r_file_layout = *layout; /* keep a copy */
193 if (flags & CEPH_OSD_FLAG_WRITE) {
194 req->r_request->hdr.data_off = cpu_to_le16(off);
195 req->r_request->hdr.data_len = cpu_to_le32(*plen);
196 op->payload_len = cpu_to_le32(*plen);
200 head->object_len = cpu_to_le32(req->r_oid_len);
201 memcpy(p, req->r_oid, req->r_oid_len);
207 op->op = cpu_to_le16(opcode == CEPH_OSD_OP_READ ?
208 CEPH_OSD_OP_MASKTRUNC : CEPH_OSD_OP_SETTRUNC);
209 op->trunc.truncate_seq = cpu_to_le32(truncate_seq);
210 prevofs = le64_to_cpu((op-1)->extent.offset);
211 op->trunc.truncate_size = cpu_to_le64(truncate_size -
216 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
219 head->snap_seq = cpu_to_le64(snapc->seq);
220 head->num_snaps = cpu_to_le32(snapc->num_snaps);
221 for (i = 0; i < snapc->num_snaps; i++) {
222 put_unaligned_le64(snapc->snaps[i], p);
227 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
232 * We keep osd requests in an rbtree, sorted by ->r_tid.
234 static void __insert_request(struct ceph_osd_client *osdc,
235 struct ceph_osd_request *new)
237 struct rb_node **p = &osdc->requests.rb_node;
238 struct rb_node *parent = NULL;
239 struct ceph_osd_request *req = NULL;
243 req = rb_entry(parent, struct ceph_osd_request, r_node);
244 if (new->r_tid < req->r_tid)
246 else if (new->r_tid > req->r_tid)
252 rb_link_node(&new->r_node, parent, p);
253 rb_insert_color(&new->r_node, &osdc->requests);
256 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
259 struct ceph_osd_request *req;
260 struct rb_node *n = osdc->requests.rb_node;
263 req = rb_entry(n, struct ceph_osd_request, r_node);
264 if (tid < req->r_tid)
266 else if (tid > req->r_tid)
274 static struct ceph_osd_request *
275 __lookup_request_ge(struct ceph_osd_client *osdc,
278 struct ceph_osd_request *req;
279 struct rb_node *n = osdc->requests.rb_node;
282 req = rb_entry(n, struct ceph_osd_request, r_node);
283 if (tid < req->r_tid) {
287 } else if (tid > req->r_tid) {
298 * If the osd connection drops, we need to resubmit all requests.
300 static void osd_reset(struct ceph_connection *con)
302 struct ceph_osd *osd = con->private;
303 struct ceph_osd_client *osdc;
307 dout("osd_reset osd%d\n", osd->o_osd);
309 osd->o_incarnation++;
310 down_read(&osdc->map_sem);
311 kick_requests(osdc, osd);
312 up_read(&osdc->map_sem);
316 * Track open sessions with osds.
318 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
320 struct ceph_osd *osd;
322 osd = kzalloc(sizeof(*osd), GFP_NOFS);
326 atomic_set(&osd->o_ref, 1);
328 INIT_LIST_HEAD(&osd->o_requests);
329 osd->o_incarnation = 1;
331 ceph_con_init(osdc->client->msgr, &osd->o_con);
332 osd->o_con.private = osd;
333 osd->o_con.ops = &osd_con_ops;
334 osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
339 static struct ceph_osd *get_osd(struct ceph_osd *osd)
341 if (atomic_inc_not_zero(&osd->o_ref)) {
342 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
343 atomic_read(&osd->o_ref));
346 dout("get_osd %p FAIL\n", osd);
351 static void put_osd(struct ceph_osd *osd)
353 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
354 atomic_read(&osd->o_ref) - 1);
355 if (atomic_dec_and_test(&osd->o_ref))
360 * remove an osd from our map
362 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
364 dout("remove_osd %p\n", osd);
365 BUG_ON(!list_empty(&osd->o_requests));
366 rb_erase(&osd->o_node, &osdc->osds);
367 ceph_con_close(&osd->o_con);
374 static int reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
378 dout("reset_osd %p osd%d\n", osd, osd->o_osd);
379 if (list_empty(&osd->o_requests)) {
380 remove_osd(osdc, osd);
382 ceph_con_close(&osd->o_con);
383 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
384 osd->o_incarnation++;
389 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
391 struct rb_node **p = &osdc->osds.rb_node;
392 struct rb_node *parent = NULL;
393 struct ceph_osd *osd = NULL;
397 osd = rb_entry(parent, struct ceph_osd, o_node);
398 if (new->o_osd < osd->o_osd)
400 else if (new->o_osd > osd->o_osd)
406 rb_link_node(&new->o_node, parent, p);
407 rb_insert_color(&new->o_node, &osdc->osds);
410 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
412 struct ceph_osd *osd;
413 struct rb_node *n = osdc->osds.rb_node;
416 osd = rb_entry(n, struct ceph_osd, o_node);
419 else if (o > osd->o_osd)
429 * Register request, assign tid. If this is the first request, set up
432 static void register_request(struct ceph_osd_client *osdc,
433 struct ceph_osd_request *req)
435 struct ceph_osd_request_head *head = req->r_request->front.iov_base;
437 mutex_lock(&osdc->request_mutex);
438 req->r_tid = ++osdc->last_tid;
439 head->tid = cpu_to_le64(req->r_tid);
441 dout("register_request %p tid %lld\n", req, req->r_tid);
442 __insert_request(osdc, req);
443 ceph_osdc_get_request(req);
444 osdc->num_requests++;
446 req->r_timeout_stamp =
447 jiffies + osdc->client->mount_args->osd_timeout*HZ;
449 if (osdc->num_requests == 1) {
450 osdc->timeout_tid = req->r_tid;
451 dout(" timeout on tid %llu at %lu\n", req->r_tid,
452 req->r_timeout_stamp);
453 schedule_delayed_work(&osdc->timeout_work,
454 round_jiffies_relative(req->r_timeout_stamp - jiffies));
456 mutex_unlock(&osdc->request_mutex);
460 * called under osdc->request_mutex
462 static void __unregister_request(struct ceph_osd_client *osdc,
463 struct ceph_osd_request *req)
465 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
466 rb_erase(&req->r_node, &osdc->requests);
467 osdc->num_requests--;
469 ceph_msgpool_resv(&osdc->msgpool_op_reply, -req->r_num_prealloc_reply);
472 /* make sure the original request isn't in flight. */
473 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
475 list_del_init(&req->r_osd_item);
476 if (list_empty(&req->r_osd->o_requests))
477 remove_osd(osdc, req->r_osd);
481 ceph_osdc_put_request(req);
483 if (req->r_tid == osdc->timeout_tid) {
484 if (osdc->num_requests == 0) {
485 dout("no requests, canceling timeout\n");
486 osdc->timeout_tid = 0;
487 cancel_delayed_work(&osdc->timeout_work);
489 req = rb_entry(rb_first(&osdc->requests),
490 struct ceph_osd_request, r_node);
491 osdc->timeout_tid = req->r_tid;
492 dout("rescheduled timeout on tid %llu at %lu\n",
493 req->r_tid, req->r_timeout_stamp);
494 schedule_delayed_work(&osdc->timeout_work,
495 round_jiffies_relative(req->r_timeout_stamp -
502 * Cancel a previously queued request message
504 static void __cancel_request(struct ceph_osd_request *req)
507 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
513 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
514 * (as needed), and set the request r_osd appropriately. If there is
515 * no up osd, set r_osd to NULL.
517 * Return 0 if unchanged, 1 if changed, or negative on error.
519 * Caller should hold map_sem for read and request_mutex.
521 static int __map_osds(struct ceph_osd_client *osdc,
522 struct ceph_osd_request *req)
524 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
528 struct ceph_osd *newosd = NULL;
530 dout("map_osds %p tid %lld\n", req, req->r_tid);
531 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
532 &req->r_file_layout, osdc->osdmap);
535 pgid = reqhead->layout.ol_pgid;
536 o = ceph_calc_pg_primary(osdc->osdmap, pgid);
538 if ((req->r_osd && req->r_osd->o_osd == o &&
539 req->r_sent >= req->r_osd->o_incarnation) ||
540 (req->r_osd == NULL && o == -1))
541 return 0; /* no change */
543 dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
544 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
545 req->r_osd ? req->r_osd->o_osd : -1);
548 __cancel_request(req);
549 list_del_init(&req->r_osd_item);
550 if (list_empty(&req->r_osd->o_requests)) {
551 /* try to re-use r_osd if possible */
552 newosd = get_osd(req->r_osd);
553 remove_osd(osdc, newosd);
558 req->r_osd = __lookup_osd(osdc, o);
559 if (!req->r_osd && o >= 0) {
565 req->r_osd = create_osd(osdc);
570 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
571 req->r_osd->o_osd = o;
572 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
573 __insert_osd(osdc, req->r_osd);
575 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
579 list_add(&req->r_osd_item, &req->r_osd->o_requests);
580 err = 1; /* osd changed */
589 * caller should hold map_sem (for read) and request_mutex
591 static int __send_request(struct ceph_osd_client *osdc,
592 struct ceph_osd_request *req)
594 struct ceph_osd_request_head *reqhead;
597 err = __map_osds(osdc, req);
600 if (req->r_osd == NULL) {
601 dout("send_request %p no up osds in pg\n", req);
602 ceph_monc_request_next_osdmap(&osdc->client->monc);
606 dout("send_request %p tid %llu to osd%d flags %d\n",
607 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
609 reqhead = req->r_request->front.iov_base;
610 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
611 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
612 reqhead->reassert_version = req->r_reassert_version;
614 req->r_timeout_stamp = jiffies+osdc->client->mount_args->osd_timeout*HZ;
616 ceph_msg_get(req->r_request); /* send consumes a ref */
617 ceph_con_send(&req->r_osd->o_con, req->r_request);
618 req->r_sent = req->r_osd->o_incarnation;
623 * Timeout callback, called every N seconds when 1 or more osd
624 * requests has been active for more than N seconds. When this
625 * happens, we ping all OSDs with requests who have timed out to
626 * ensure any communications channel reset is detected. Reset the
627 * request timeouts another N seconds in the future as we go.
628 * Reschedule the timeout event another N seconds in future (unless
629 * there are no open requests).
631 static void handle_timeout(struct work_struct *work)
633 struct ceph_osd_client *osdc =
634 container_of(work, struct ceph_osd_client, timeout_work.work);
635 struct ceph_osd_request *req;
636 struct ceph_osd *osd;
637 unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
638 unsigned long next_timeout = timeout + jiffies;
642 down_read(&osdc->map_sem);
644 ceph_monc_request_next_osdmap(&osdc->client->monc);
646 mutex_lock(&osdc->request_mutex);
647 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
648 req = rb_entry(p, struct ceph_osd_request, r_node);
653 dout("osdc resending prev failed %lld\n", req->r_tid);
654 err = __send_request(osdc, req);
656 dout("osdc failed again on %lld\n", req->r_tid);
658 req->r_resend = false;
662 for (p = rb_first(&osdc->osds); p; p = rb_next(p)) {
663 osd = rb_entry(p, struct ceph_osd, o_node);
664 if (list_empty(&osd->o_requests))
666 req = list_first_entry(&osd->o_requests,
667 struct ceph_osd_request, r_osd_item);
668 if (time_before(jiffies, req->r_timeout_stamp))
671 dout(" tid %llu (at least) timed out on osd%d\n",
672 req->r_tid, osd->o_osd);
673 req->r_timeout_stamp = next_timeout;
674 ceph_con_keepalive(&osd->o_con);
677 if (osdc->timeout_tid)
678 schedule_delayed_work(&osdc->timeout_work,
679 round_jiffies_relative(timeout));
681 mutex_unlock(&osdc->request_mutex);
683 up_read(&osdc->map_sem);
687 * handle osd op reply. either call the callback if it is specified,
688 * or do the completion to wake up the waiting thread.
690 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
692 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
693 struct ceph_osd_request *req;
695 int numops, object_len, flags;
697 if (msg->front.iov_len < sizeof(*rhead))
699 tid = le64_to_cpu(rhead->tid);
700 numops = le32_to_cpu(rhead->num_ops);
701 object_len = le32_to_cpu(rhead->object_len);
702 if (msg->front.iov_len != sizeof(*rhead) + object_len +
703 numops * sizeof(struct ceph_osd_op))
705 dout("handle_reply %p tid %llu\n", msg, tid);
708 mutex_lock(&osdc->request_mutex);
709 req = __lookup_request(osdc, tid);
711 dout("handle_reply tid %llu dne\n", tid);
712 mutex_unlock(&osdc->request_mutex);
715 ceph_osdc_get_request(req);
716 flags = le32_to_cpu(rhead->flags);
720 * once we see the message has been received, we don't
721 * need a ref (which is only needed for revoking
724 ceph_msg_put(req->r_reply);
728 if (!req->r_got_reply) {
731 req->r_result = le32_to_cpu(rhead->result);
732 bytes = le32_to_cpu(msg->hdr.data_len);
733 dout("handle_reply result %d bytes %d\n", req->r_result,
735 if (req->r_result == 0)
736 req->r_result = bytes;
738 /* in case this is a write and we need to replay, */
739 req->r_reassert_version = rhead->reassert_version;
741 req->r_got_reply = 1;
742 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
743 dout("handle_reply tid %llu dup ack\n", tid);
744 mutex_unlock(&osdc->request_mutex);
748 dout("handle_reply tid %llu flags %d\n", tid, flags);
750 /* either this is a read, or we got the safe response */
751 if ((flags & CEPH_OSD_FLAG_ONDISK) ||
752 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
753 __unregister_request(osdc, req);
755 mutex_unlock(&osdc->request_mutex);
758 req->r_callback(req, msg);
760 complete(&req->r_completion);
762 if (flags & CEPH_OSD_FLAG_ONDISK) {
763 if (req->r_safe_callback)
764 req->r_safe_callback(req, msg);
765 complete(&req->r_safe_completion); /* fsync waiter */
769 ceph_osdc_put_request(req);
773 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
774 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
775 (int)sizeof(*rhead));
781 * Resubmit osd requests whose osd or osd address has changed. Request
782 * a new osd map if osds are down, or we are otherwise unable to determine
783 * how to direct a request.
785 * Close connections to down osds.
787 * If @who is specified, resubmit requests for that specific osd.
789 * Caller should hold map_sem for read and request_mutex.
791 static void kick_requests(struct ceph_osd_client *osdc,
792 struct ceph_osd *kickosd)
794 struct ceph_osd_request *req;
795 struct rb_node *p, *n;
799 dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
800 mutex_lock(&osdc->request_mutex);
802 for (p = rb_first(&osdc->osds); p; p = n) {
803 struct ceph_osd *osd =
804 rb_entry(p, struct ceph_osd, o_node);
807 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
808 !ceph_entity_addr_equal(&osd->o_con.peer_addr,
809 ceph_osd_addr(osdc->osdmap,
811 reset_osd(osdc, osd);
815 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
816 req = rb_entry(p, struct ceph_osd_request, r_node);
819 dout(" r_resend set on tid %llu\n", req->r_tid);
820 __cancel_request(req);
823 if (req->r_osd && kickosd == req->r_osd) {
824 __cancel_request(req);
828 err = __map_osds(osdc, req);
830 continue; /* no change */
833 * FIXME: really, we should set the request
834 * error and fail if this isn't a 'nofail'
835 * request, but that's a fair bit more
836 * complicated to do. So retry!
838 dout(" setting r_resend on %llu\n", req->r_tid);
839 req->r_resend = true;
842 if (req->r_osd == NULL) {
843 dout("tid %llu maps to no valid osd\n", req->r_tid);
844 needmap++; /* request a newer map */
849 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
851 req->r_flags |= CEPH_OSD_FLAG_RETRY;
852 err = __send_request(osdc, req);
854 dout(" setting r_resend on %llu\n", req->r_tid);
855 req->r_resend = true;
858 mutex_unlock(&osdc->request_mutex);
861 dout("%d requests for down osds, need new map\n", needmap);
862 ceph_monc_request_next_osdmap(&osdc->client->monc);
867 * Process updated osd map.
869 * The message contains any number of incremental and full maps, normally
870 * indicating some sort of topology change in the cluster. Kick requests
871 * off to different OSDs as needed.
873 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
875 void *p, *end, *next;
878 struct ceph_osdmap *newmap = NULL, *oldmap;
880 struct ceph_fsid fsid;
882 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
883 p = msg->front.iov_base;
884 end = p + msg->front.iov_len;
887 ceph_decode_need(&p, end, sizeof(fsid), bad);
888 ceph_decode_copy(&p, &fsid, sizeof(fsid));
889 if (ceph_check_fsid(osdc->client, &fsid) < 0)
892 down_write(&osdc->map_sem);
894 /* incremental maps */
895 ceph_decode_32_safe(&p, end, nr_maps, bad);
896 dout(" %d inc maps\n", nr_maps);
897 while (nr_maps > 0) {
898 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
899 epoch = ceph_decode_32(&p);
900 maplen = ceph_decode_32(&p);
901 ceph_decode_need(&p, end, maplen, bad);
903 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
904 dout("applying incremental map %u len %d\n",
906 newmap = osdmap_apply_incremental(&p, next,
909 if (IS_ERR(newmap)) {
910 err = PTR_ERR(newmap);
914 if (newmap != osdc->osdmap) {
915 ceph_osdmap_destroy(osdc->osdmap);
916 osdc->osdmap = newmap;
919 dout("ignoring incremental map %u len %d\n",
929 ceph_decode_32_safe(&p, end, nr_maps, bad);
930 dout(" %d full maps\n", nr_maps);
932 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
933 epoch = ceph_decode_32(&p);
934 maplen = ceph_decode_32(&p);
935 ceph_decode_need(&p, end, maplen, bad);
937 dout("skipping non-latest full map %u len %d\n",
939 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
940 dout("skipping full map %u len %d, "
941 "older than our %u\n", epoch, maplen,
942 osdc->osdmap->epoch);
944 dout("taking full map %u len %d\n", epoch, maplen);
945 newmap = osdmap_decode(&p, p+maplen);
946 if (IS_ERR(newmap)) {
947 err = PTR_ERR(newmap);
951 oldmap = osdc->osdmap;
952 osdc->osdmap = newmap;
954 ceph_osdmap_destroy(oldmap);
961 downgrade_write(&osdc->map_sem);
962 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
964 kick_requests(osdc, NULL);
965 up_read(&osdc->map_sem);
969 pr_err("osdc handle_map corrupt msg\n");
971 up_write(&osdc->map_sem);
977 * A read request prepares specific pages that data is to be read into.
978 * When a message is being read off the wire, we call prepare_pages to
980 * 0 = success, -1 failure.
982 static int prepare_pages(struct ceph_connection *con, struct ceph_msg *m,
985 struct ceph_osd *osd = con->private;
986 struct ceph_osd_client *osdc;
987 struct ceph_osd_reply_head *rhead = m->front.iov_base;
988 struct ceph_osd_request *req;
991 int type = le16_to_cpu(m->hdr.type);
997 dout("prepare_pages on msg %p want %d\n", m, want);
998 if (unlikely(type != CEPH_MSG_OSD_OPREPLY))
999 return -1; /* hmm! */
1001 tid = le64_to_cpu(rhead->tid);
1002 mutex_lock(&osdc->request_mutex);
1003 req = __lookup_request(osdc, tid);
1005 dout("prepare_pages unknown tid %llu\n", tid);
1008 dout("prepare_pages tid %llu has %d pages, want %d\n",
1009 tid, req->r_num_pages, want);
1010 if (likely(req->r_num_pages >= want && !req->r_prepared_pages)) {
1011 m->pages = req->r_pages;
1012 m->nr_pages = req->r_num_pages;
1013 req->r_reply = m; /* only for duration of read over socket */
1015 req->r_prepared_pages = 1;
1016 ret = 0; /* success */
1019 mutex_unlock(&osdc->request_mutex);
1024 * Register request, send initial attempt.
1026 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1027 struct ceph_osd_request *req,
1032 req->r_request->pages = req->r_pages;
1033 req->r_request->nr_pages = req->r_num_pages;
1035 register_request(osdc, req);
1037 down_read(&osdc->map_sem);
1038 mutex_lock(&osdc->request_mutex);
1040 * a racing kick_requests() may have sent the message for us
1041 * while we dropped request_mutex above, so only send now if
1042 * the request still han't been touched yet.
1044 if (req->r_sent == 0) {
1045 rc = __send_request(osdc, req);
1048 dout("osdc_start_request failed send, "
1049 " marking %lld\n", req->r_tid);
1050 req->r_resend = true;
1053 __unregister_request(osdc, req);
1057 mutex_unlock(&osdc->request_mutex);
1058 up_read(&osdc->map_sem);
1063 * wait for a request to complete
1065 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1066 struct ceph_osd_request *req)
1070 rc = wait_for_completion_interruptible(&req->r_completion);
1072 mutex_lock(&osdc->request_mutex);
1073 __cancel_request(req);
1074 __unregister_request(osdc, req);
1075 mutex_unlock(&osdc->request_mutex);
1076 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1080 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1081 return req->r_result;
1085 * sync - wait for all in-flight requests to flush. avoid starvation.
1087 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1089 struct ceph_osd_request *req;
1090 u64 last_tid, next_tid = 0;
1092 mutex_lock(&osdc->request_mutex);
1093 last_tid = osdc->last_tid;
1095 req = __lookup_request_ge(osdc, next_tid);
1098 if (req->r_tid > last_tid)
1101 next_tid = req->r_tid + 1;
1102 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1105 ceph_osdc_get_request(req);
1106 mutex_unlock(&osdc->request_mutex);
1107 dout("sync waiting on tid %llu (last is %llu)\n",
1108 req->r_tid, last_tid);
1109 wait_for_completion(&req->r_safe_completion);
1110 mutex_lock(&osdc->request_mutex);
1111 ceph_osdc_put_request(req);
1113 mutex_unlock(&osdc->request_mutex);
1114 dout("sync done (thru tid %llu)\n", last_tid);
1120 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1125 osdc->client = client;
1126 osdc->osdmap = NULL;
1127 init_rwsem(&osdc->map_sem);
1128 init_completion(&osdc->map_waiters);
1129 osdc->last_requested_map = 0;
1130 mutex_init(&osdc->request_mutex);
1131 osdc->timeout_tid = 0;
1133 osdc->osds = RB_ROOT;
1134 osdc->requests = RB_ROOT;
1135 osdc->num_requests = 0;
1136 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1139 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1140 sizeof(struct ceph_osd_request));
1141 if (!osdc->req_mempool)
1144 err = ceph_msgpool_init(&osdc->msgpool_op, 4096, 10, true);
1147 err = ceph_msgpool_init(&osdc->msgpool_op_reply, 512, 0, false);
1153 ceph_msgpool_destroy(&osdc->msgpool_op);
1155 mempool_destroy(osdc->req_mempool);
1160 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1162 cancel_delayed_work_sync(&osdc->timeout_work);
1164 ceph_osdmap_destroy(osdc->osdmap);
1165 osdc->osdmap = NULL;
1167 mempool_destroy(osdc->req_mempool);
1168 ceph_msgpool_destroy(&osdc->msgpool_op);
1169 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1173 * Read some contiguous pages. If we cross a stripe boundary, shorten
1174 * *plen. Return number of bytes read, or error.
1176 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1177 struct ceph_vino vino, struct ceph_file_layout *layout,
1179 u32 truncate_seq, u64 truncate_size,
1180 struct page **pages, int num_pages)
1182 struct ceph_osd_request *req;
1185 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1186 vino.snap, off, *plen);
1187 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1188 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1189 NULL, 0, truncate_seq, truncate_size, NULL,
1192 return PTR_ERR(req);
1194 /* it may be a short read due to an object boundary */
1195 req->r_pages = pages;
1196 num_pages = calc_pages_for(off, *plen);
1197 req->r_num_pages = num_pages;
1199 dout("readpages final extent is %llu~%llu (%d pages)\n",
1200 off, *plen, req->r_num_pages);
1202 rc = ceph_osdc_start_request(osdc, req, false);
1204 rc = ceph_osdc_wait_request(osdc, req);
1206 ceph_osdc_put_request(req);
1207 dout("readpages result %d\n", rc);
1212 * do a synchronous write on N pages
1214 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1215 struct ceph_file_layout *layout,
1216 struct ceph_snap_context *snapc,
1218 u32 truncate_seq, u64 truncate_size,
1219 struct timespec *mtime,
1220 struct page **pages, int num_pages,
1221 int flags, int do_sync, bool nofail)
1223 struct ceph_osd_request *req;
1226 BUG_ON(vino.snap != CEPH_NOSNAP);
1227 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1229 flags | CEPH_OSD_FLAG_ONDISK |
1230 CEPH_OSD_FLAG_WRITE,
1232 truncate_seq, truncate_size, mtime,
1235 return PTR_ERR(req);
1237 /* it may be a short write due to an object boundary */
1238 req->r_pages = pages;
1239 req->r_num_pages = calc_pages_for(off, len);
1240 dout("writepages %llu~%llu (%d pages)\n", off, len,
1243 rc = ceph_osdc_start_request(osdc, req, nofail);
1245 rc = ceph_osdc_wait_request(osdc, req);
1247 ceph_osdc_put_request(req);
1250 dout("writepages result %d\n", rc);
1255 * handle incoming message
1257 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1259 struct ceph_osd *osd = con->private;
1260 struct ceph_osd_client *osdc;
1261 int type = le16_to_cpu(msg->hdr.type);
1268 case CEPH_MSG_OSD_MAP:
1269 ceph_osdc_handle_map(osdc, msg);
1271 case CEPH_MSG_OSD_OPREPLY:
1272 handle_reply(osdc, msg);
1276 pr_err("received unknown message type %d %s\n", type,
1277 ceph_msg_type_name(type));
1282 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1283 struct ceph_msg_header *hdr)
1285 struct ceph_osd *osd = con->private;
1286 struct ceph_osd_client *osdc = osd->o_osdc;
1287 int type = le16_to_cpu(hdr->type);
1288 int front = le32_to_cpu(hdr->front_len);
1291 case CEPH_MSG_OSD_OPREPLY:
1292 return ceph_msgpool_get(&osdc->msgpool_op_reply, front);
1294 return ceph_alloc_msg(con, hdr);
1298 * Wrappers to refcount containing ceph_osd struct
1300 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1302 struct ceph_osd *osd = con->private;
1308 static void put_osd_con(struct ceph_connection *con)
1310 struct ceph_osd *osd = con->private;
1317 static int get_authorizer(struct ceph_connection *con,
1318 void **buf, int *len, int *proto,
1319 void **reply_buf, int *reply_len, int force_new)
1321 struct ceph_osd *o = con->private;
1322 struct ceph_osd_client *osdc = o->o_osdc;
1323 struct ceph_auth_client *ac = osdc->client->monc.auth;
1326 if (force_new && o->o_authorizer) {
1327 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1328 o->o_authorizer = NULL;
1330 if (o->o_authorizer == NULL) {
1331 ret = ac->ops->create_authorizer(
1332 ac, CEPH_ENTITY_TYPE_OSD,
1334 &o->o_authorizer_buf,
1335 &o->o_authorizer_buf_len,
1336 &o->o_authorizer_reply_buf,
1337 &o->o_authorizer_reply_buf_len);
1342 *proto = ac->protocol;
1343 *buf = o->o_authorizer_buf;
1344 *len = o->o_authorizer_buf_len;
1345 *reply_buf = o->o_authorizer_reply_buf;
1346 *reply_len = o->o_authorizer_reply_buf_len;
1351 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1353 struct ceph_osd *o = con->private;
1354 struct ceph_osd_client *osdc = o->o_osdc;
1355 struct ceph_auth_client *ac = osdc->client->monc.auth;
1357 return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1361 const static struct ceph_connection_operations osd_con_ops = {
1364 .dispatch = dispatch,
1365 .get_authorizer = get_authorizer,
1366 .verify_authorizer_reply = verify_authorizer_reply,
1367 .alloc_msg = alloc_msg,
1369 .alloc_middle = ceph_alloc_middle,
1370 .prepare_pages = prepare_pages,