1 #include <linux/ceph/ceph_debug.h>
3 #include <linux/module.h>
5 #include <linux/highmem.h>
7 #include <linux/pagemap.h>
8 #include <linux/slab.h>
9 #include <linux/uaccess.h>
11 #include <linux/bio.h>
14 #include <linux/ceph/libceph.h>
15 #include <linux/ceph/osd_client.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/auth.h>
19 #include <linux/ceph/pagelist.h>
21 #define OSD_OP_FRONT_LEN 4096
22 #define OSD_OPREPLY_FRONT_LEN 512
24 static const struct ceph_connection_operations osd_con_ops;
26 static void send_queued(struct ceph_osd_client *osdc);
27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
28 static void __register_request(struct ceph_osd_client *osdc,
29 struct ceph_osd_request *req);
30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
31 struct ceph_osd_request *req);
32 static void __send_request(struct ceph_osd_client *osdc,
33 struct ceph_osd_request *req);
35 static int op_needs_trail(int op)
38 case CEPH_OSD_OP_GETXATTR:
39 case CEPH_OSD_OP_SETXATTR:
40 case CEPH_OSD_OP_CMPXATTR:
41 case CEPH_OSD_OP_CALL:
42 case CEPH_OSD_OP_NOTIFY:
49 static int op_has_extent(int op)
51 return (op == CEPH_OSD_OP_READ ||
52 op == CEPH_OSD_OP_WRITE);
55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
56 struct ceph_file_layout *layout,
58 u64 off, u64 *plen, u64 *bno,
59 struct ceph_osd_request *req,
60 struct ceph_osd_req_op *op)
62 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
64 u64 objoff, objlen; /* extent in object */
66 reqhead->snapid = cpu_to_le64(snapid);
69 ceph_calc_file_object_mapping(layout, off, plen, bno,
72 dout(" skipping last %llu, final file extent %llu~%llu\n",
73 orig_len - *plen, off, *plen);
75 if (op_has_extent(op->op)) {
76 op->extent.offset = objoff;
77 op->extent.length = objlen;
79 req->r_num_pages = calc_pages_for(off, *plen);
80 req->r_page_alignment = off & ~PAGE_MASK;
81 if (op->op == CEPH_OSD_OP_WRITE)
82 op->payload_len = *plen;
84 dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
85 *bno, objoff, objlen, req->r_num_pages);
88 EXPORT_SYMBOL(ceph_calc_raw_layout);
91 * Implement client access to distributed object storage cluster.
93 * All data objects are stored within a cluster/cloud of OSDs, or
94 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
95 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
96 * remote daemons serving up and coordinating consistent and safe
99 * Cluster membership and the mapping of data objects onto storage devices
100 * are described by the osd map.
102 * We keep track of pending OSD requests (read, write), resubmit
103 * requests to different OSDs when the cluster topology/data layout
104 * change, or retry the affected requests when the communications
105 * channel with an OSD is reset.
109 * calculate the mapping of a file extent onto an object, and fill out the
110 * request accordingly. shorten extent as necessary if it crosses an
113 * fill osd op in request message.
115 static void calc_layout(struct ceph_osd_client *osdc,
116 struct ceph_vino vino,
117 struct ceph_file_layout *layout,
119 struct ceph_osd_request *req,
120 struct ceph_osd_req_op *op)
124 ceph_calc_raw_layout(osdc, layout, vino.snap, off,
125 plen, &bno, req, op);
127 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
128 req->r_oid_len = strlen(req->r_oid);
134 void ceph_osdc_release_request(struct kref *kref)
136 struct ceph_osd_request *req = container_of(kref,
137 struct ceph_osd_request,
141 ceph_msg_put(req->r_request);
142 if (req->r_con_filling_msg) {
143 dout("%s revoking pages %p from con %p\n", __func__,
144 req->r_pages, req->r_con_filling_msg);
145 ceph_msg_revoke_incoming(req->r_reply);
146 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
149 ceph_msg_put(req->r_reply);
150 if (req->r_own_pages)
151 ceph_release_page_vector(req->r_pages,
157 ceph_put_snap_context(req->r_snapc);
159 ceph_pagelist_release(req->r_trail);
163 mempool_free(req, req->r_osdc->req_mempool);
167 EXPORT_SYMBOL(ceph_osdc_release_request);
169 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
176 if (needs_trail && op_needs_trail(ops[i].op))
184 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
186 struct ceph_snap_context *snapc,
187 struct ceph_osd_req_op *ops,
193 struct ceph_osd_request *req;
194 struct ceph_msg *msg;
196 int num_op = get_num_ops(ops, &needs_trail);
197 size_t msg_size = sizeof(struct ceph_osd_request_head);
199 msg_size += num_op*sizeof(struct ceph_osd_op);
202 req = mempool_alloc(osdc->req_mempool, gfp_flags);
203 memset(req, 0, sizeof(*req));
205 req = kzalloc(sizeof(*req), gfp_flags);
211 req->r_mempool = use_mempool;
213 kref_init(&req->r_kref);
214 init_completion(&req->r_completion);
215 init_completion(&req->r_safe_completion);
216 INIT_LIST_HEAD(&req->r_unsafe_item);
217 INIT_LIST_HEAD(&req->r_linger_item);
218 INIT_LIST_HEAD(&req->r_linger_osd);
219 INIT_LIST_HEAD(&req->r_req_lru_item);
220 req->r_flags = flags;
222 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
224 /* create reply message */
226 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
228 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
229 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
231 ceph_osdc_put_request(req);
236 /* allocate space for the trailing data */
238 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
240 ceph_osdc_put_request(req);
243 ceph_pagelist_init(req->r_trail);
245 /* create request message; allow space for oid */
246 msg_size += MAX_OBJ_NAME_SIZE;
248 msg_size += sizeof(u64) * snapc->num_snaps;
250 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
252 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
254 ceph_osdc_put_request(req);
258 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
259 memset(msg->front.iov_base, 0, msg->front.iov_len);
261 req->r_request = msg;
262 req->r_pages = pages;
272 EXPORT_SYMBOL(ceph_osdc_alloc_request);
274 static void osd_req_encode_op(struct ceph_osd_request *req,
275 struct ceph_osd_op *dst,
276 struct ceph_osd_req_op *src)
278 dst->op = cpu_to_le16(src->op);
281 case CEPH_OSD_OP_READ:
282 case CEPH_OSD_OP_WRITE:
284 cpu_to_le64(src->extent.offset);
286 cpu_to_le64(src->extent.length);
287 dst->extent.truncate_size =
288 cpu_to_le64(src->extent.truncate_size);
289 dst->extent.truncate_seq =
290 cpu_to_le32(src->extent.truncate_seq);
293 case CEPH_OSD_OP_GETXATTR:
294 case CEPH_OSD_OP_SETXATTR:
295 case CEPH_OSD_OP_CMPXATTR:
296 BUG_ON(!req->r_trail);
298 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
299 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
300 dst->xattr.cmp_op = src->xattr.cmp_op;
301 dst->xattr.cmp_mode = src->xattr.cmp_mode;
302 ceph_pagelist_append(req->r_trail, src->xattr.name,
303 src->xattr.name_len);
304 ceph_pagelist_append(req->r_trail, src->xattr.val,
305 src->xattr.value_len);
307 case CEPH_OSD_OP_CALL:
308 BUG_ON(!req->r_trail);
310 dst->cls.class_len = src->cls.class_len;
311 dst->cls.method_len = src->cls.method_len;
312 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
314 ceph_pagelist_append(req->r_trail, src->cls.class_name,
316 ceph_pagelist_append(req->r_trail, src->cls.method_name,
317 src->cls.method_len);
318 ceph_pagelist_append(req->r_trail, src->cls.indata,
319 src->cls.indata_len);
321 case CEPH_OSD_OP_ROLLBACK:
322 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
324 case CEPH_OSD_OP_STARTSYNC:
326 case CEPH_OSD_OP_NOTIFY:
328 __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
329 __le32 timeout = cpu_to_le32(src->watch.timeout);
331 BUG_ON(!req->r_trail);
333 ceph_pagelist_append(req->r_trail,
334 &prot_ver, sizeof(prot_ver));
335 ceph_pagelist_append(req->r_trail,
336 &timeout, sizeof(timeout));
338 case CEPH_OSD_OP_NOTIFY_ACK:
339 case CEPH_OSD_OP_WATCH:
340 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
341 dst->watch.ver = cpu_to_le64(src->watch.ver);
342 dst->watch.flag = src->watch.flag;
345 pr_err("unrecognized osd opcode %d\n", dst->op);
349 dst->payload_len = cpu_to_le32(src->payload_len);
353 * build new request AND message
356 void ceph_osdc_build_request(struct ceph_osd_request *req,
358 struct ceph_osd_req_op *src_ops,
359 struct ceph_snap_context *snapc,
360 struct timespec *mtime,
364 struct ceph_msg *msg = req->r_request;
365 struct ceph_osd_request_head *head;
366 struct ceph_osd_req_op *src_op;
367 struct ceph_osd_op *op;
369 int num_op = get_num_ops(src_ops, NULL);
370 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
371 int flags = req->r_flags;
375 head = msg->front.iov_base;
376 op = (void *)(head + 1);
377 p = (void *)(op + num_op);
379 req->r_snapc = ceph_get_snap_context(snapc);
381 head->client_inc = cpu_to_le32(1); /* always, for now. */
382 head->flags = cpu_to_le32(flags);
383 if (flags & CEPH_OSD_FLAG_WRITE)
384 ceph_encode_timespec(&head->mtime, mtime);
385 head->num_ops = cpu_to_le16(num_op);
389 head->object_len = cpu_to_le32(oid_len);
390 memcpy(p, oid, oid_len);
395 osd_req_encode_op(req, op, src_op);
401 data_len += req->r_trail->length;
404 head->snap_seq = cpu_to_le64(snapc->seq);
405 head->num_snaps = cpu_to_le32(snapc->num_snaps);
406 for (i = 0; i < snapc->num_snaps; i++) {
407 put_unaligned_le64(snapc->snaps[i], p);
412 if (flags & CEPH_OSD_FLAG_WRITE) {
413 req->r_request->hdr.data_off = cpu_to_le16(off);
414 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
415 } else if (data_len) {
416 req->r_request->hdr.data_off = 0;
417 req->r_request->hdr.data_len = cpu_to_le32(data_len);
420 req->r_request->page_alignment = req->r_page_alignment;
422 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
423 msg_size = p - msg->front.iov_base;
424 msg->front.iov_len = msg_size;
425 msg->hdr.front_len = cpu_to_le32(msg_size);
428 EXPORT_SYMBOL(ceph_osdc_build_request);
431 * build new request AND message, calculate layout, and adjust file
434 * if the file was recently truncated, we include information about its
435 * old and new size so that the object can be updated appropriately. (we
436 * avoid synchronously deleting truncated objects because it's slow.)
438 * if @do_sync, include a 'startsync' command so that the osd will flush
441 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
442 struct ceph_file_layout *layout,
443 struct ceph_vino vino,
445 int opcode, int flags,
446 struct ceph_snap_context *snapc,
450 struct timespec *mtime,
451 bool use_mempool, int num_reply,
454 struct ceph_osd_req_op ops[3];
455 struct ceph_osd_request *req;
458 ops[0].extent.truncate_seq = truncate_seq;
459 ops[0].extent.truncate_size = truncate_size;
460 ops[0].payload_len = 0;
463 ops[1].op = CEPH_OSD_OP_STARTSYNC;
464 ops[1].payload_len = 0;
469 req = ceph_osdc_alloc_request(osdc, flags,
472 GFP_NOFS, NULL, NULL);
476 /* calculate max write size */
477 calc_layout(osdc, vino, layout, off, plen, req, ops);
478 req->r_file_layout = *layout; /* keep a copy */
480 /* in case it differs from natural (file) alignment that
481 calc_layout filled in for us */
482 req->r_num_pages = calc_pages_for(page_align, *plen);
483 req->r_page_alignment = page_align;
485 ceph_osdc_build_request(req, off, plen, ops,
488 req->r_oid, req->r_oid_len);
492 EXPORT_SYMBOL(ceph_osdc_new_request);
495 * We keep osd requests in an rbtree, sorted by ->r_tid.
497 static void __insert_request(struct ceph_osd_client *osdc,
498 struct ceph_osd_request *new)
500 struct rb_node **p = &osdc->requests.rb_node;
501 struct rb_node *parent = NULL;
502 struct ceph_osd_request *req = NULL;
506 req = rb_entry(parent, struct ceph_osd_request, r_node);
507 if (new->r_tid < req->r_tid)
509 else if (new->r_tid > req->r_tid)
515 rb_link_node(&new->r_node, parent, p);
516 rb_insert_color(&new->r_node, &osdc->requests);
519 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
522 struct ceph_osd_request *req;
523 struct rb_node *n = osdc->requests.rb_node;
526 req = rb_entry(n, struct ceph_osd_request, r_node);
527 if (tid < req->r_tid)
529 else if (tid > req->r_tid)
537 static struct ceph_osd_request *
538 __lookup_request_ge(struct ceph_osd_client *osdc,
541 struct ceph_osd_request *req;
542 struct rb_node *n = osdc->requests.rb_node;
545 req = rb_entry(n, struct ceph_osd_request, r_node);
546 if (tid < req->r_tid) {
550 } else if (tid > req->r_tid) {
560 * Resubmit requests pending on the given osd.
562 static void __kick_osd_requests(struct ceph_osd_client *osdc,
563 struct ceph_osd *osd)
565 struct ceph_osd_request *req, *nreq;
568 dout("__kick_osd_requests osd%d\n", osd->o_osd);
569 err = __reset_osd(osdc, osd);
573 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
574 list_move(&req->r_req_lru_item, &osdc->req_unsent);
575 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
578 req->r_flags |= CEPH_OSD_FLAG_RETRY;
581 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
584 * reregister request prior to unregistering linger so
585 * that r_osd is preserved.
587 BUG_ON(!list_empty(&req->r_req_lru_item));
588 __register_request(osdc, req);
589 list_add(&req->r_req_lru_item, &osdc->req_unsent);
590 list_add(&req->r_osd_item, &req->r_osd->o_requests);
591 __unregister_linger_request(osdc, req);
592 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
597 static void kick_osd_requests(struct ceph_osd_client *osdc,
598 struct ceph_osd *kickosd)
600 mutex_lock(&osdc->request_mutex);
601 __kick_osd_requests(osdc, kickosd);
602 mutex_unlock(&osdc->request_mutex);
606 * If the osd connection drops, we need to resubmit all requests.
608 static void osd_reset(struct ceph_connection *con)
610 struct ceph_osd *osd = con->private;
611 struct ceph_osd_client *osdc;
615 dout("osd_reset osd%d\n", osd->o_osd);
617 down_read(&osdc->map_sem);
618 kick_osd_requests(osdc, osd);
620 up_read(&osdc->map_sem);
624 * Track open sessions with osds.
626 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
628 struct ceph_osd *osd;
630 osd = kzalloc(sizeof(*osd), GFP_NOFS);
634 atomic_set(&osd->o_ref, 1);
637 INIT_LIST_HEAD(&osd->o_requests);
638 INIT_LIST_HEAD(&osd->o_linger_requests);
639 INIT_LIST_HEAD(&osd->o_osd_lru);
640 osd->o_incarnation = 1;
642 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr,
643 CEPH_ENTITY_TYPE_OSD, onum);
645 INIT_LIST_HEAD(&osd->o_keepalive_item);
649 static struct ceph_osd *get_osd(struct ceph_osd *osd)
651 if (atomic_inc_not_zero(&osd->o_ref)) {
652 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
653 atomic_read(&osd->o_ref));
656 dout("get_osd %p FAIL\n", osd);
661 static void put_osd(struct ceph_osd *osd)
663 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
664 atomic_read(&osd->o_ref) - 1);
665 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
666 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
668 if (ac->ops && ac->ops->destroy_authorizer)
669 ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
675 * remove an osd from our map
677 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
679 dout("__remove_osd %p\n", osd);
680 BUG_ON(!list_empty(&osd->o_requests));
681 rb_erase(&osd->o_node, &osdc->osds);
682 list_del_init(&osd->o_osd_lru);
683 ceph_con_close(&osd->o_con);
687 static void remove_all_osds(struct ceph_osd_client *osdc)
689 dout("__remove_old_osds %p\n", osdc);
690 mutex_lock(&osdc->request_mutex);
691 while (!RB_EMPTY_ROOT(&osdc->osds)) {
692 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
693 struct ceph_osd, o_node);
694 __remove_osd(osdc, osd);
696 mutex_unlock(&osdc->request_mutex);
699 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
700 struct ceph_osd *osd)
702 dout("__move_osd_to_lru %p\n", osd);
703 BUG_ON(!list_empty(&osd->o_osd_lru));
704 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
705 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
708 static void __remove_osd_from_lru(struct ceph_osd *osd)
710 dout("__remove_osd_from_lru %p\n", osd);
711 if (!list_empty(&osd->o_osd_lru))
712 list_del_init(&osd->o_osd_lru);
715 static void remove_old_osds(struct ceph_osd_client *osdc)
717 struct ceph_osd *osd, *nosd;
719 dout("__remove_old_osds %p\n", osdc);
720 mutex_lock(&osdc->request_mutex);
721 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
722 if (time_before(jiffies, osd->lru_ttl))
724 __remove_osd(osdc, osd);
726 mutex_unlock(&osdc->request_mutex);
732 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
734 struct ceph_osd_request *req;
737 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
738 if (list_empty(&osd->o_requests) &&
739 list_empty(&osd->o_linger_requests)) {
740 __remove_osd(osdc, osd);
741 } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
742 &osd->o_con.peer_addr,
743 sizeof(osd->o_con.peer_addr)) == 0 &&
744 !ceph_con_opened(&osd->o_con)) {
745 dout(" osd addr hasn't changed and connection never opened,"
746 " letting msgr retry");
747 /* touch each r_stamp for handle_timeout()'s benfit */
748 list_for_each_entry(req, &osd->o_requests, r_osd_item)
749 req->r_stamp = jiffies;
752 ceph_con_close(&osd->o_con);
753 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
754 osd->o_incarnation++;
759 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
761 struct rb_node **p = &osdc->osds.rb_node;
762 struct rb_node *parent = NULL;
763 struct ceph_osd *osd = NULL;
765 dout("__insert_osd %p osd%d\n", new, new->o_osd);
768 osd = rb_entry(parent, struct ceph_osd, o_node);
769 if (new->o_osd < osd->o_osd)
771 else if (new->o_osd > osd->o_osd)
777 rb_link_node(&new->o_node, parent, p);
778 rb_insert_color(&new->o_node, &osdc->osds);
781 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
783 struct ceph_osd *osd;
784 struct rb_node *n = osdc->osds.rb_node;
787 osd = rb_entry(n, struct ceph_osd, o_node);
790 else if (o > osd->o_osd)
798 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
800 schedule_delayed_work(&osdc->timeout_work,
801 osdc->client->options->osd_keepalive_timeout * HZ);
804 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
806 cancel_delayed_work(&osdc->timeout_work);
810 * Register request, assign tid. If this is the first request, set up
813 static void __register_request(struct ceph_osd_client *osdc,
814 struct ceph_osd_request *req)
816 req->r_tid = ++osdc->last_tid;
817 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
818 dout("__register_request %p tid %lld\n", req, req->r_tid);
819 __insert_request(osdc, req);
820 ceph_osdc_get_request(req);
821 osdc->num_requests++;
822 if (osdc->num_requests == 1) {
823 dout(" first request, scheduling timeout\n");
824 __schedule_osd_timeout(osdc);
828 static void register_request(struct ceph_osd_client *osdc,
829 struct ceph_osd_request *req)
831 mutex_lock(&osdc->request_mutex);
832 __register_request(osdc, req);
833 mutex_unlock(&osdc->request_mutex);
837 * called under osdc->request_mutex
839 static void __unregister_request(struct ceph_osd_client *osdc,
840 struct ceph_osd_request *req)
842 if (RB_EMPTY_NODE(&req->r_node)) {
843 dout("__unregister_request %p tid %lld not registered\n",
848 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
849 rb_erase(&req->r_node, &osdc->requests);
850 osdc->num_requests--;
853 /* make sure the original request isn't in flight. */
854 ceph_msg_revoke(req->r_request);
856 list_del_init(&req->r_osd_item);
857 if (list_empty(&req->r_osd->o_requests) &&
858 list_empty(&req->r_osd->o_linger_requests)) {
859 dout("moving osd to %p lru\n", req->r_osd);
860 __move_osd_to_lru(osdc, req->r_osd);
862 if (list_empty(&req->r_linger_item))
866 ceph_osdc_put_request(req);
868 list_del_init(&req->r_req_lru_item);
869 if (osdc->num_requests == 0) {
870 dout(" no requests, canceling timeout\n");
871 __cancel_osd_timeout(osdc);
876 * Cancel a previously queued request message
878 static void __cancel_request(struct ceph_osd_request *req)
880 if (req->r_sent && req->r_osd) {
881 ceph_msg_revoke(req->r_request);
886 static void __register_linger_request(struct ceph_osd_client *osdc,
887 struct ceph_osd_request *req)
889 dout("__register_linger_request %p\n", req);
890 list_add_tail(&req->r_linger_item, &osdc->req_linger);
891 list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests);
894 static void __unregister_linger_request(struct ceph_osd_client *osdc,
895 struct ceph_osd_request *req)
897 dout("__unregister_linger_request %p\n", req);
899 list_del_init(&req->r_linger_item);
900 list_del_init(&req->r_linger_osd);
902 if (list_empty(&req->r_osd->o_requests) &&
903 list_empty(&req->r_osd->o_linger_requests)) {
904 dout("moving osd to %p lru\n", req->r_osd);
905 __move_osd_to_lru(osdc, req->r_osd);
907 if (list_empty(&req->r_osd_item))
912 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
913 struct ceph_osd_request *req)
915 mutex_lock(&osdc->request_mutex);
917 __unregister_linger_request(osdc, req);
918 ceph_osdc_put_request(req);
920 mutex_unlock(&osdc->request_mutex);
922 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
924 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
925 struct ceph_osd_request *req)
927 if (!req->r_linger) {
928 dout("set_request_linger %p\n", req);
931 * caller is now responsible for calling
932 * unregister_linger_request
934 ceph_osdc_get_request(req);
937 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
940 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
941 * (as needed), and set the request r_osd appropriately. If there is
942 * no up osd, set r_osd to NULL. Move the request to the appropriate list
943 * (unsent, homeless) or leave on in-flight lru.
945 * Return 0 if unchanged, 1 if changed, or negative on error.
947 * Caller should hold map_sem for read and request_mutex.
949 static int __map_request(struct ceph_osd_client *osdc,
950 struct ceph_osd_request *req, int force_resend)
952 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
954 int acting[CEPH_PG_MAX_SIZE];
958 dout("map_request %p tid %lld\n", req, req->r_tid);
959 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
960 &req->r_file_layout, osdc->osdmap);
962 list_move(&req->r_req_lru_item, &osdc->req_notarget);
965 pgid = reqhead->layout.ol_pgid;
968 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
974 if ((!force_resend &&
975 req->r_osd && req->r_osd->o_osd == o &&
976 req->r_sent >= req->r_osd->o_incarnation &&
977 req->r_num_pg_osds == num &&
978 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
979 (req->r_osd == NULL && o == -1))
980 return 0; /* no change */
982 dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
983 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
984 req->r_osd ? req->r_osd->o_osd : -1);
986 /* record full pg acting set */
987 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
988 req->r_num_pg_osds = num;
991 __cancel_request(req);
992 list_del_init(&req->r_osd_item);
996 req->r_osd = __lookup_osd(osdc, o);
997 if (!req->r_osd && o >= 0) {
999 req->r_osd = create_osd(osdc, o);
1001 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1005 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1006 __insert_osd(osdc, req->r_osd);
1008 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
1012 __remove_osd_from_lru(req->r_osd);
1013 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1014 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1016 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1018 err = 1; /* osd or pg changed */
1025 * caller should hold map_sem (for read) and request_mutex
1027 static void __send_request(struct ceph_osd_client *osdc,
1028 struct ceph_osd_request *req)
1030 struct ceph_osd_request_head *reqhead;
1032 dout("send_request %p tid %llu to osd%d flags %d\n",
1033 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1035 reqhead = req->r_request->front.iov_base;
1036 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1037 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
1038 reqhead->reassert_version = req->r_reassert_version;
1040 req->r_stamp = jiffies;
1041 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1043 ceph_msg_get(req->r_request); /* send consumes a ref */
1044 ceph_con_send(&req->r_osd->o_con, req->r_request);
1045 req->r_sent = req->r_osd->o_incarnation;
1049 * Send any requests in the queue (req_unsent).
1051 static void send_queued(struct ceph_osd_client *osdc)
1053 struct ceph_osd_request *req, *tmp;
1055 dout("send_queued\n");
1056 mutex_lock(&osdc->request_mutex);
1057 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1058 __send_request(osdc, req);
1060 mutex_unlock(&osdc->request_mutex);
1064 * Timeout callback, called every N seconds when 1 or more osd
1065 * requests has been active for more than N seconds. When this
1066 * happens, we ping all OSDs with requests who have timed out to
1067 * ensure any communications channel reset is detected. Reset the
1068 * request timeouts another N seconds in the future as we go.
1069 * Reschedule the timeout event another N seconds in future (unless
1070 * there are no open requests).
1072 static void handle_timeout(struct work_struct *work)
1074 struct ceph_osd_client *osdc =
1075 container_of(work, struct ceph_osd_client, timeout_work.work);
1076 struct ceph_osd_request *req, *last_req = NULL;
1077 struct ceph_osd *osd;
1078 unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1079 unsigned long keepalive =
1080 osdc->client->options->osd_keepalive_timeout * HZ;
1081 unsigned long last_stamp = 0;
1082 struct list_head slow_osds;
1084 down_read(&osdc->map_sem);
1086 ceph_monc_request_next_osdmap(&osdc->client->monc);
1088 mutex_lock(&osdc->request_mutex);
1091 * reset osds that appear to be _really_ unresponsive. this
1092 * is a failsafe measure.. we really shouldn't be getting to
1093 * this point if the system is working properly. the monitors
1094 * should mark the osd as failed and we should find out about
1095 * it from an updated osd map.
1097 while (timeout && !list_empty(&osdc->req_lru)) {
1098 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1101 /* hasn't been long enough since we sent it? */
1102 if (time_before(jiffies, req->r_stamp + timeout))
1105 /* hasn't been long enough since it was acked? */
1106 if (req->r_request->ack_stamp == 0 ||
1107 time_before(jiffies, req->r_request->ack_stamp + timeout))
1110 BUG_ON(req == last_req && req->r_stamp == last_stamp);
1112 last_stamp = req->r_stamp;
1116 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1117 req->r_tid, osd->o_osd);
1118 __kick_osd_requests(osdc, osd);
1122 * ping osds that are a bit slow. this ensures that if there
1123 * is a break in the TCP connection we will notice, and reopen
1124 * a connection with that osd (from the fault callback).
1126 INIT_LIST_HEAD(&slow_osds);
1127 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1128 if (time_before(jiffies, req->r_stamp + keepalive))
1133 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1134 req->r_tid, osd->o_osd);
1135 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1137 while (!list_empty(&slow_osds)) {
1138 osd = list_entry(slow_osds.next, struct ceph_osd,
1140 list_del_init(&osd->o_keepalive_item);
1141 ceph_con_keepalive(&osd->o_con);
1144 __schedule_osd_timeout(osdc);
1145 mutex_unlock(&osdc->request_mutex);
1147 up_read(&osdc->map_sem);
1150 static void handle_osds_timeout(struct work_struct *work)
1152 struct ceph_osd_client *osdc =
1153 container_of(work, struct ceph_osd_client,
1154 osds_timeout_work.work);
1155 unsigned long delay =
1156 osdc->client->options->osd_idle_ttl * HZ >> 2;
1158 dout("osds timeout\n");
1159 down_read(&osdc->map_sem);
1160 remove_old_osds(osdc);
1161 up_read(&osdc->map_sem);
1163 schedule_delayed_work(&osdc->osds_timeout_work,
1164 round_jiffies_relative(delay));
1167 static void complete_request(struct ceph_osd_request *req)
1169 if (req->r_safe_callback)
1170 req->r_safe_callback(req, NULL);
1171 complete_all(&req->r_safe_completion); /* fsync waiter */
1175 * handle osd op reply. either call the callback if it is specified,
1176 * or do the completion to wake up the waiting thread.
1178 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1179 struct ceph_connection *con)
1181 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1182 struct ceph_osd_request *req;
1184 int numops, object_len, flags;
1187 tid = le64_to_cpu(msg->hdr.tid);
1188 if (msg->front.iov_len < sizeof(*rhead))
1190 numops = le32_to_cpu(rhead->num_ops);
1191 object_len = le32_to_cpu(rhead->object_len);
1192 result = le32_to_cpu(rhead->result);
1193 if (msg->front.iov_len != sizeof(*rhead) + object_len +
1194 numops * sizeof(struct ceph_osd_op))
1196 dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1198 mutex_lock(&osdc->request_mutex);
1199 req = __lookup_request(osdc, tid);
1201 dout("handle_reply tid %llu dne\n", tid);
1202 mutex_unlock(&osdc->request_mutex);
1205 ceph_osdc_get_request(req);
1206 flags = le32_to_cpu(rhead->flags);
1209 * if this connection filled our message, drop our reference now, to
1210 * avoid a (safe but slower) revoke later.
1212 if (req->r_con_filling_msg == con && req->r_reply == msg) {
1213 dout(" dropping con_filling_msg ref %p\n", con);
1214 req->r_con_filling_msg = NULL;
1218 if (!req->r_got_reply) {
1221 req->r_result = le32_to_cpu(rhead->result);
1222 bytes = le32_to_cpu(msg->hdr.data_len);
1223 dout("handle_reply result %d bytes %d\n", req->r_result,
1225 if (req->r_result == 0)
1226 req->r_result = bytes;
1228 /* in case this is a write and we need to replay, */
1229 req->r_reassert_version = rhead->reassert_version;
1231 req->r_got_reply = 1;
1232 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1233 dout("handle_reply tid %llu dup ack\n", tid);
1234 mutex_unlock(&osdc->request_mutex);
1238 dout("handle_reply tid %llu flags %d\n", tid, flags);
1240 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1241 __register_linger_request(osdc, req);
1243 /* either this is a read, or we got the safe response */
1245 (flags & CEPH_OSD_FLAG_ONDISK) ||
1246 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1247 __unregister_request(osdc, req);
1249 mutex_unlock(&osdc->request_mutex);
1251 if (req->r_callback)
1252 req->r_callback(req, msg);
1254 complete_all(&req->r_completion);
1256 if (flags & CEPH_OSD_FLAG_ONDISK)
1257 complete_request(req);
1260 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1261 ceph_osdc_put_request(req);
1265 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1266 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1267 (int)sizeof(*rhead));
1271 static void reset_changed_osds(struct ceph_osd_client *osdc)
1273 struct rb_node *p, *n;
1275 for (p = rb_first(&osdc->osds); p; p = n) {
1276 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1279 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1280 memcmp(&osd->o_con.peer_addr,
1281 ceph_osd_addr(osdc->osdmap,
1283 sizeof(struct ceph_entity_addr)) != 0)
1284 __reset_osd(osdc, osd);
1289 * Requeue requests whose mapping to an OSD has changed. If requests map to
1290 * no osd, request a new map.
1292 * Caller should hold map_sem for read and request_mutex.
1294 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1296 struct ceph_osd_request *req, *nreq;
1301 dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1302 mutex_lock(&osdc->request_mutex);
1303 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1304 req = rb_entry(p, struct ceph_osd_request, r_node);
1305 err = __map_request(osdc, req, force_resend);
1307 continue; /* error */
1308 if (req->r_osd == NULL) {
1309 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1310 needmap++; /* request a newer map */
1311 } else if (err > 0) {
1312 dout("%p tid %llu requeued on osd%d\n", req, req->r_tid,
1313 req->r_osd ? req->r_osd->o_osd : -1);
1315 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1319 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1321 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1323 err = __map_request(osdc, req, force_resend);
1325 continue; /* no change and no osd was specified */
1327 continue; /* hrm! */
1328 if (req->r_osd == NULL) {
1329 dout("tid %llu maps to no valid osd\n", req->r_tid);
1330 needmap++; /* request a newer map */
1334 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1335 req->r_osd ? req->r_osd->o_osd : -1);
1336 __unregister_linger_request(osdc, req);
1337 __register_request(osdc, req);
1339 mutex_unlock(&osdc->request_mutex);
1342 dout("%d requests for down osds, need new map\n", needmap);
1343 ceph_monc_request_next_osdmap(&osdc->client->monc);
1349 * Process updated osd map.
1351 * The message contains any number of incremental and full maps, normally
1352 * indicating some sort of topology change in the cluster. Kick requests
1353 * off to different OSDs as needed.
1355 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1357 void *p, *end, *next;
1358 u32 nr_maps, maplen;
1360 struct ceph_osdmap *newmap = NULL, *oldmap;
1362 struct ceph_fsid fsid;
1364 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1365 p = msg->front.iov_base;
1366 end = p + msg->front.iov_len;
1369 ceph_decode_need(&p, end, sizeof(fsid), bad);
1370 ceph_decode_copy(&p, &fsid, sizeof(fsid));
1371 if (ceph_check_fsid(osdc->client, &fsid) < 0)
1374 down_write(&osdc->map_sem);
1376 /* incremental maps */
1377 ceph_decode_32_safe(&p, end, nr_maps, bad);
1378 dout(" %d inc maps\n", nr_maps);
1379 while (nr_maps > 0) {
1380 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1381 epoch = ceph_decode_32(&p);
1382 maplen = ceph_decode_32(&p);
1383 ceph_decode_need(&p, end, maplen, bad);
1385 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1386 dout("applying incremental map %u len %d\n",
1388 newmap = osdmap_apply_incremental(&p, next,
1390 &osdc->client->msgr);
1391 if (IS_ERR(newmap)) {
1392 err = PTR_ERR(newmap);
1396 if (newmap != osdc->osdmap) {
1397 ceph_osdmap_destroy(osdc->osdmap);
1398 osdc->osdmap = newmap;
1400 kick_requests(osdc, 0);
1401 reset_changed_osds(osdc);
1403 dout("ignoring incremental map %u len %d\n",
1413 ceph_decode_32_safe(&p, end, nr_maps, bad);
1414 dout(" %d full maps\n", nr_maps);
1416 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1417 epoch = ceph_decode_32(&p);
1418 maplen = ceph_decode_32(&p);
1419 ceph_decode_need(&p, end, maplen, bad);
1421 dout("skipping non-latest full map %u len %d\n",
1423 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1424 dout("skipping full map %u len %d, "
1425 "older than our %u\n", epoch, maplen,
1426 osdc->osdmap->epoch);
1428 int skipped_map = 0;
1430 dout("taking full map %u len %d\n", epoch, maplen);
1431 newmap = osdmap_decode(&p, p+maplen);
1432 if (IS_ERR(newmap)) {
1433 err = PTR_ERR(newmap);
1437 oldmap = osdc->osdmap;
1438 osdc->osdmap = newmap;
1440 if (oldmap->epoch + 1 < newmap->epoch)
1442 ceph_osdmap_destroy(oldmap);
1444 kick_requests(osdc, skipped_map);
1451 downgrade_write(&osdc->map_sem);
1452 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1455 * subscribe to subsequent osdmap updates if full to ensure
1456 * we find out when we are no longer full and stop returning
1459 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1460 ceph_monc_request_next_osdmap(&osdc->client->monc);
1463 up_read(&osdc->map_sem);
1464 wake_up_all(&osdc->client->auth_wq);
1468 pr_err("osdc handle_map corrupt msg\n");
1470 up_write(&osdc->map_sem);
1475 * watch/notify callback event infrastructure
1477 * These callbacks are used both for watch and notify operations.
1479 static void __release_event(struct kref *kref)
1481 struct ceph_osd_event *event =
1482 container_of(kref, struct ceph_osd_event, kref);
1484 dout("__release_event %p\n", event);
1488 static void get_event(struct ceph_osd_event *event)
1490 kref_get(&event->kref);
1493 void ceph_osdc_put_event(struct ceph_osd_event *event)
1495 kref_put(&event->kref, __release_event);
1497 EXPORT_SYMBOL(ceph_osdc_put_event);
1499 static void __insert_event(struct ceph_osd_client *osdc,
1500 struct ceph_osd_event *new)
1502 struct rb_node **p = &osdc->event_tree.rb_node;
1503 struct rb_node *parent = NULL;
1504 struct ceph_osd_event *event = NULL;
1508 event = rb_entry(parent, struct ceph_osd_event, node);
1509 if (new->cookie < event->cookie)
1511 else if (new->cookie > event->cookie)
1512 p = &(*p)->rb_right;
1517 rb_link_node(&new->node, parent, p);
1518 rb_insert_color(&new->node, &osdc->event_tree);
1521 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1524 struct rb_node **p = &osdc->event_tree.rb_node;
1525 struct rb_node *parent = NULL;
1526 struct ceph_osd_event *event = NULL;
1530 event = rb_entry(parent, struct ceph_osd_event, node);
1531 if (cookie < event->cookie)
1533 else if (cookie > event->cookie)
1534 p = &(*p)->rb_right;
1541 static void __remove_event(struct ceph_osd_event *event)
1543 struct ceph_osd_client *osdc = event->osdc;
1545 if (!RB_EMPTY_NODE(&event->node)) {
1546 dout("__remove_event removed %p\n", event);
1547 rb_erase(&event->node, &osdc->event_tree);
1548 ceph_osdc_put_event(event);
1550 dout("__remove_event didn't remove %p\n", event);
1554 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1555 void (*event_cb)(u64, u64, u8, void *),
1556 int one_shot, void *data,
1557 struct ceph_osd_event **pevent)
1559 struct ceph_osd_event *event;
1561 event = kmalloc(sizeof(*event), GFP_NOIO);
1565 dout("create_event %p\n", event);
1566 event->cb = event_cb;
1567 event->one_shot = one_shot;
1570 INIT_LIST_HEAD(&event->osd_node);
1571 kref_init(&event->kref); /* one ref for us */
1572 kref_get(&event->kref); /* one ref for the caller */
1573 init_completion(&event->completion);
1575 spin_lock(&osdc->event_lock);
1576 event->cookie = ++osdc->event_count;
1577 __insert_event(osdc, event);
1578 spin_unlock(&osdc->event_lock);
1583 EXPORT_SYMBOL(ceph_osdc_create_event);
1585 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1587 struct ceph_osd_client *osdc = event->osdc;
1589 dout("cancel_event %p\n", event);
1590 spin_lock(&osdc->event_lock);
1591 __remove_event(event);
1592 spin_unlock(&osdc->event_lock);
1593 ceph_osdc_put_event(event); /* caller's */
1595 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1598 static void do_event_work(struct work_struct *work)
1600 struct ceph_osd_event_work *event_work =
1601 container_of(work, struct ceph_osd_event_work, work);
1602 struct ceph_osd_event *event = event_work->event;
1603 u64 ver = event_work->ver;
1604 u64 notify_id = event_work->notify_id;
1605 u8 opcode = event_work->opcode;
1607 dout("do_event_work completing %p\n", event);
1608 event->cb(ver, notify_id, opcode, event->data);
1609 complete(&event->completion);
1610 dout("do_event_work completed %p\n", event);
1611 ceph_osdc_put_event(event);
1617 * Process osd watch notifications
1619 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1623 u64 cookie, ver, notify_id;
1625 struct ceph_osd_event *event;
1626 struct ceph_osd_event_work *event_work;
1628 p = msg->front.iov_base;
1629 end = p + msg->front.iov_len;
1631 ceph_decode_8_safe(&p, end, proto_ver, bad);
1632 ceph_decode_8_safe(&p, end, opcode, bad);
1633 ceph_decode_64_safe(&p, end, cookie, bad);
1634 ceph_decode_64_safe(&p, end, ver, bad);
1635 ceph_decode_64_safe(&p, end, notify_id, bad);
1637 spin_lock(&osdc->event_lock);
1638 event = __find_event(osdc, cookie);
1641 if (event->one_shot)
1642 __remove_event(event);
1644 spin_unlock(&osdc->event_lock);
1645 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1646 cookie, ver, event);
1648 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1650 dout("ERROR: could not allocate event_work\n");
1653 INIT_WORK(&event_work->work, do_event_work);
1654 event_work->event = event;
1655 event_work->ver = ver;
1656 event_work->notify_id = notify_id;
1657 event_work->opcode = opcode;
1658 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1659 dout("WARNING: failed to queue notify event work\n");
1667 complete(&event->completion);
1668 ceph_osdc_put_event(event);
1672 pr_err("osdc handle_watch_notify corrupt msg\n");
1676 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1680 dout("wait_event %p\n", event);
1681 err = wait_for_completion_interruptible_timeout(&event->completion,
1683 ceph_osdc_put_event(event);
1686 dout("wait_event %p returns %d\n", event, err);
1689 EXPORT_SYMBOL(ceph_osdc_wait_event);
1692 * Register request, send initial attempt.
1694 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1695 struct ceph_osd_request *req,
1700 req->r_request->pages = req->r_pages;
1701 req->r_request->nr_pages = req->r_num_pages;
1703 req->r_request->bio = req->r_bio;
1705 req->r_request->trail = req->r_trail;
1707 register_request(osdc, req);
1709 down_read(&osdc->map_sem);
1710 mutex_lock(&osdc->request_mutex);
1712 * a racing kick_requests() may have sent the message for us
1713 * while we dropped request_mutex above, so only send now if
1714 * the request still han't been touched yet.
1716 if (req->r_sent == 0) {
1717 rc = __map_request(osdc, req, 0);
1720 dout("osdc_start_request failed map, "
1721 " will retry %lld\n", req->r_tid);
1726 if (req->r_osd == NULL) {
1727 dout("send_request %p no up osds in pg\n", req);
1728 ceph_monc_request_next_osdmap(&osdc->client->monc);
1730 __send_request(osdc, req);
1736 mutex_unlock(&osdc->request_mutex);
1737 up_read(&osdc->map_sem);
1740 EXPORT_SYMBOL(ceph_osdc_start_request);
1743 * wait for a request to complete
1745 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1746 struct ceph_osd_request *req)
1750 rc = wait_for_completion_interruptible(&req->r_completion);
1752 mutex_lock(&osdc->request_mutex);
1753 __cancel_request(req);
1754 __unregister_request(osdc, req);
1755 mutex_unlock(&osdc->request_mutex);
1756 complete_request(req);
1757 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1761 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1762 return req->r_result;
1764 EXPORT_SYMBOL(ceph_osdc_wait_request);
1767 * sync - wait for all in-flight requests to flush. avoid starvation.
1769 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1771 struct ceph_osd_request *req;
1772 u64 last_tid, next_tid = 0;
1774 mutex_lock(&osdc->request_mutex);
1775 last_tid = osdc->last_tid;
1777 req = __lookup_request_ge(osdc, next_tid);
1780 if (req->r_tid > last_tid)
1783 next_tid = req->r_tid + 1;
1784 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1787 ceph_osdc_get_request(req);
1788 mutex_unlock(&osdc->request_mutex);
1789 dout("sync waiting on tid %llu (last is %llu)\n",
1790 req->r_tid, last_tid);
1791 wait_for_completion(&req->r_safe_completion);
1792 mutex_lock(&osdc->request_mutex);
1793 ceph_osdc_put_request(req);
1795 mutex_unlock(&osdc->request_mutex);
1796 dout("sync done (thru tid %llu)\n", last_tid);
1798 EXPORT_SYMBOL(ceph_osdc_sync);
1803 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1808 osdc->client = client;
1809 osdc->osdmap = NULL;
1810 init_rwsem(&osdc->map_sem);
1811 init_completion(&osdc->map_waiters);
1812 osdc->last_requested_map = 0;
1813 mutex_init(&osdc->request_mutex);
1815 osdc->osds = RB_ROOT;
1816 INIT_LIST_HEAD(&osdc->osd_lru);
1817 osdc->requests = RB_ROOT;
1818 INIT_LIST_HEAD(&osdc->req_lru);
1819 INIT_LIST_HEAD(&osdc->req_unsent);
1820 INIT_LIST_HEAD(&osdc->req_notarget);
1821 INIT_LIST_HEAD(&osdc->req_linger);
1822 osdc->num_requests = 0;
1823 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1824 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1825 spin_lock_init(&osdc->event_lock);
1826 osdc->event_tree = RB_ROOT;
1827 osdc->event_count = 0;
1829 schedule_delayed_work(&osdc->osds_timeout_work,
1830 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1833 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1834 sizeof(struct ceph_osd_request));
1835 if (!osdc->req_mempool)
1838 err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1842 err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1843 OSD_OPREPLY_FRONT_LEN, 10, true,
1848 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1849 if (IS_ERR(osdc->notify_wq)) {
1850 err = PTR_ERR(osdc->notify_wq);
1851 osdc->notify_wq = NULL;
1857 ceph_msgpool_destroy(&osdc->msgpool_op);
1859 mempool_destroy(osdc->req_mempool);
1863 EXPORT_SYMBOL(ceph_osdc_init);
1865 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1867 flush_workqueue(osdc->notify_wq);
1868 destroy_workqueue(osdc->notify_wq);
1869 cancel_delayed_work_sync(&osdc->timeout_work);
1870 cancel_delayed_work_sync(&osdc->osds_timeout_work);
1872 ceph_osdmap_destroy(osdc->osdmap);
1873 osdc->osdmap = NULL;
1875 remove_all_osds(osdc);
1876 mempool_destroy(osdc->req_mempool);
1877 ceph_msgpool_destroy(&osdc->msgpool_op);
1878 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1880 EXPORT_SYMBOL(ceph_osdc_stop);
1883 * Read some contiguous pages. If we cross a stripe boundary, shorten
1884 * *plen. Return number of bytes read, or error.
1886 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1887 struct ceph_vino vino, struct ceph_file_layout *layout,
1889 u32 truncate_seq, u64 truncate_size,
1890 struct page **pages, int num_pages, int page_align)
1892 struct ceph_osd_request *req;
1895 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1896 vino.snap, off, *plen);
1897 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1898 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1899 NULL, 0, truncate_seq, truncate_size, NULL,
1900 false, 1, page_align);
1904 /* it may be a short read due to an object boundary */
1905 req->r_pages = pages;
1907 dout("readpages final extent is %llu~%llu (%d pages align %d)\n",
1908 off, *plen, req->r_num_pages, page_align);
1910 rc = ceph_osdc_start_request(osdc, req, false);
1912 rc = ceph_osdc_wait_request(osdc, req);
1914 ceph_osdc_put_request(req);
1915 dout("readpages result %d\n", rc);
1918 EXPORT_SYMBOL(ceph_osdc_readpages);
1921 * do a synchronous write on N pages
1923 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1924 struct ceph_file_layout *layout,
1925 struct ceph_snap_context *snapc,
1927 u32 truncate_seq, u64 truncate_size,
1928 struct timespec *mtime,
1929 struct page **pages, int num_pages,
1930 int flags, int do_sync, bool nofail)
1932 struct ceph_osd_request *req;
1934 int page_align = off & ~PAGE_MASK;
1936 BUG_ON(vino.snap != CEPH_NOSNAP);
1937 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1939 flags | CEPH_OSD_FLAG_ONDISK |
1940 CEPH_OSD_FLAG_WRITE,
1942 truncate_seq, truncate_size, mtime,
1943 nofail, 1, page_align);
1947 /* it may be a short write due to an object boundary */
1948 req->r_pages = pages;
1949 dout("writepages %llu~%llu (%d pages)\n", off, len,
1952 rc = ceph_osdc_start_request(osdc, req, nofail);
1954 rc = ceph_osdc_wait_request(osdc, req);
1956 ceph_osdc_put_request(req);
1959 dout("writepages result %d\n", rc);
1962 EXPORT_SYMBOL(ceph_osdc_writepages);
1965 * handle incoming message
1967 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1969 struct ceph_osd *osd = con->private;
1970 struct ceph_osd_client *osdc;
1971 int type = le16_to_cpu(msg->hdr.type);
1978 case CEPH_MSG_OSD_MAP:
1979 ceph_osdc_handle_map(osdc, msg);
1981 case CEPH_MSG_OSD_OPREPLY:
1982 handle_reply(osdc, msg, con);
1984 case CEPH_MSG_WATCH_NOTIFY:
1985 handle_watch_notify(osdc, msg);
1989 pr_err("received unknown message type %d %s\n", type,
1990 ceph_msg_type_name(type));
1997 * lookup and return message for incoming reply. set up reply message
2000 static struct ceph_msg *get_reply(struct ceph_connection *con,
2001 struct ceph_msg_header *hdr,
2004 struct ceph_osd *osd = con->private;
2005 struct ceph_osd_client *osdc = osd->o_osdc;
2007 struct ceph_osd_request *req;
2008 int front = le32_to_cpu(hdr->front_len);
2009 int data_len = le32_to_cpu(hdr->data_len);
2012 tid = le64_to_cpu(hdr->tid);
2013 mutex_lock(&osdc->request_mutex);
2014 req = __lookup_request(osdc, tid);
2018 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
2023 if (req->r_con_filling_msg) {
2024 dout("%s revoking msg %p from old con %p\n", __func__,
2025 req->r_reply, req->r_con_filling_msg);
2026 ceph_msg_revoke_incoming(req->r_reply);
2027 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2028 req->r_con_filling_msg = NULL;
2031 if (front > req->r_reply->front.iov_len) {
2032 pr_warning("get_reply front %d > preallocated %d\n",
2033 front, (int)req->r_reply->front.iov_len);
2034 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2037 ceph_msg_put(req->r_reply);
2040 m = ceph_msg_get(req->r_reply);
2043 int want = calc_pages_for(req->r_page_alignment, data_len);
2045 if (unlikely(req->r_num_pages < want)) {
2046 pr_warning("tid %lld reply has %d bytes %d pages, we"
2047 " had only %d pages ready\n", tid, data_len,
2048 want, req->r_num_pages);
2054 m->pages = req->r_pages;
2055 m->nr_pages = req->r_num_pages;
2056 m->page_alignment = req->r_page_alignment;
2058 m->bio = req->r_bio;
2062 req->r_con_filling_msg = con->ops->get(con);
2063 dout("get_reply tid %lld %p\n", tid, m);
2066 mutex_unlock(&osdc->request_mutex);
2071 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2072 struct ceph_msg_header *hdr,
2075 struct ceph_osd *osd = con->private;
2076 int type = le16_to_cpu(hdr->type);
2077 int front = le32_to_cpu(hdr->front_len);
2081 case CEPH_MSG_OSD_MAP:
2082 case CEPH_MSG_WATCH_NOTIFY:
2083 return ceph_msg_new(type, front, GFP_NOFS, false);
2084 case CEPH_MSG_OSD_OPREPLY:
2085 return get_reply(con, hdr, skip);
2087 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2095 * Wrappers to refcount containing ceph_osd struct
2097 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2099 struct ceph_osd *osd = con->private;
2105 static void put_osd_con(struct ceph_connection *con)
2107 struct ceph_osd *osd = con->private;
2115 * Note: returned pointer is the address of a structure that's
2116 * managed separately. Caller must *not* attempt to free it.
2118 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2119 int *proto, int force_new)
2121 struct ceph_osd *o = con->private;
2122 struct ceph_osd_client *osdc = o->o_osdc;
2123 struct ceph_auth_client *ac = osdc->client->monc.auth;
2124 struct ceph_auth_handshake *auth = &o->o_auth;
2126 if (force_new && auth->authorizer) {
2127 if (ac->ops && ac->ops->destroy_authorizer)
2128 ac->ops->destroy_authorizer(ac, auth->authorizer);
2129 auth->authorizer = NULL;
2131 if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2132 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2135 return ERR_PTR(ret);
2137 *proto = ac->protocol;
2143 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2145 struct ceph_osd *o = con->private;
2146 struct ceph_osd_client *osdc = o->o_osdc;
2147 struct ceph_auth_client *ac = osdc->client->monc.auth;
2150 * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2151 * XXX which do we do: succeed or fail?
2153 return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2156 static int invalidate_authorizer(struct ceph_connection *con)
2158 struct ceph_osd *o = con->private;
2159 struct ceph_osd_client *osdc = o->o_osdc;
2160 struct ceph_auth_client *ac = osdc->client->monc.auth;
2162 if (ac->ops && ac->ops->invalidate_authorizer)
2163 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2165 return ceph_monc_validate_auth(&osdc->client->monc);
2168 static const struct ceph_connection_operations osd_con_ops = {
2171 .dispatch = dispatch,
2172 .get_authorizer = get_authorizer,
2173 .verify_authorizer_reply = verify_authorizer_reply,
2174 .invalidate_authorizer = invalidate_authorizer,
2175 .alloc_msg = alloc_msg,