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 int 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 */
67 reqhead->snapid = cpu_to_le64(snapid);
70 r = ceph_calc_file_object_mapping(layout, off, plen, bno,
75 dout(" skipping last %llu, final file extent %llu~%llu\n",
76 orig_len - *plen, off, *plen);
78 if (op_has_extent(op->op)) {
79 op->extent.offset = objoff;
80 op->extent.length = objlen;
82 req->r_num_pages = calc_pages_for(off, *plen);
83 req->r_page_alignment = off & ~PAGE_MASK;
84 if (op->op == CEPH_OSD_OP_WRITE)
85 op->payload_len = *plen;
87 dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
88 *bno, objoff, objlen, req->r_num_pages);
91 EXPORT_SYMBOL(ceph_calc_raw_layout);
94 * Implement client access to distributed object storage cluster.
96 * All data objects are stored within a cluster/cloud of OSDs, or
97 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
98 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
99 * remote daemons serving up and coordinating consistent and safe
102 * Cluster membership and the mapping of data objects onto storage devices
103 * are described by the osd map.
105 * We keep track of pending OSD requests (read, write), resubmit
106 * requests to different OSDs when the cluster topology/data layout
107 * change, or retry the affected requests when the communications
108 * channel with an OSD is reset.
112 * calculate the mapping of a file extent onto an object, and fill out the
113 * request accordingly. shorten extent as necessary if it crosses an
116 * fill osd op in request message.
118 static int calc_layout(struct ceph_osd_client *osdc,
119 struct ceph_vino vino,
120 struct ceph_file_layout *layout,
122 struct ceph_osd_request *req,
123 struct ceph_osd_req_op *op)
128 r = ceph_calc_raw_layout(osdc, layout, vino.snap, off,
129 plen, &bno, req, op);
133 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
134 req->r_oid_len = strlen(req->r_oid);
142 void ceph_osdc_release_request(struct kref *kref)
144 struct ceph_osd_request *req = container_of(kref,
145 struct ceph_osd_request,
149 ceph_msg_put(req->r_request);
150 if (req->r_con_filling_msg) {
151 dout("%s revoking pages %p from con %p\n", __func__,
152 req->r_pages, req->r_con_filling_msg);
153 ceph_msg_revoke_incoming(req->r_reply);
154 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
157 ceph_msg_put(req->r_reply);
158 if (req->r_own_pages)
159 ceph_release_page_vector(req->r_pages,
165 ceph_put_snap_context(req->r_snapc);
167 ceph_pagelist_release(req->r_trail);
171 mempool_free(req, req->r_osdc->req_mempool);
175 EXPORT_SYMBOL(ceph_osdc_release_request);
177 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
184 if (needs_trail && op_needs_trail(ops[i].op))
192 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
194 struct ceph_snap_context *snapc,
195 struct ceph_osd_req_op *ops,
201 struct ceph_osd_request *req;
202 struct ceph_msg *msg;
204 int num_op = get_num_ops(ops, &needs_trail);
205 size_t msg_size = sizeof(struct ceph_osd_request_head);
207 msg_size += num_op*sizeof(struct ceph_osd_op);
210 req = mempool_alloc(osdc->req_mempool, gfp_flags);
211 memset(req, 0, sizeof(*req));
213 req = kzalloc(sizeof(*req), gfp_flags);
219 req->r_mempool = use_mempool;
221 kref_init(&req->r_kref);
222 init_completion(&req->r_completion);
223 init_completion(&req->r_safe_completion);
224 rb_init_node(&req->r_node);
225 INIT_LIST_HEAD(&req->r_unsafe_item);
226 INIT_LIST_HEAD(&req->r_linger_item);
227 INIT_LIST_HEAD(&req->r_linger_osd);
228 INIT_LIST_HEAD(&req->r_req_lru_item);
229 INIT_LIST_HEAD(&req->r_osd_item);
231 req->r_flags = flags;
233 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
235 /* create reply message */
237 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
239 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
240 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
242 ceph_osdc_put_request(req);
247 /* allocate space for the trailing data */
249 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
251 ceph_osdc_put_request(req);
254 ceph_pagelist_init(req->r_trail);
257 /* create request message; allow space for oid */
258 msg_size += MAX_OBJ_NAME_SIZE;
260 msg_size += sizeof(u64) * snapc->num_snaps;
262 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
264 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
266 ceph_osdc_put_request(req);
270 memset(msg->front.iov_base, 0, msg->front.iov_len);
272 req->r_request = msg;
273 req->r_pages = pages;
283 EXPORT_SYMBOL(ceph_osdc_alloc_request);
285 static void osd_req_encode_op(struct ceph_osd_request *req,
286 struct ceph_osd_op *dst,
287 struct ceph_osd_req_op *src)
289 dst->op = cpu_to_le16(src->op);
292 case CEPH_OSD_OP_READ:
293 case CEPH_OSD_OP_WRITE:
295 cpu_to_le64(src->extent.offset);
297 cpu_to_le64(src->extent.length);
298 dst->extent.truncate_size =
299 cpu_to_le64(src->extent.truncate_size);
300 dst->extent.truncate_seq =
301 cpu_to_le32(src->extent.truncate_seq);
304 case CEPH_OSD_OP_GETXATTR:
305 case CEPH_OSD_OP_SETXATTR:
306 case CEPH_OSD_OP_CMPXATTR:
307 BUG_ON(!req->r_trail);
309 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
310 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
311 dst->xattr.cmp_op = src->xattr.cmp_op;
312 dst->xattr.cmp_mode = src->xattr.cmp_mode;
313 ceph_pagelist_append(req->r_trail, src->xattr.name,
314 src->xattr.name_len);
315 ceph_pagelist_append(req->r_trail, src->xattr.val,
316 src->xattr.value_len);
318 case CEPH_OSD_OP_CALL:
319 BUG_ON(!req->r_trail);
321 dst->cls.class_len = src->cls.class_len;
322 dst->cls.method_len = src->cls.method_len;
323 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
325 ceph_pagelist_append(req->r_trail, src->cls.class_name,
327 ceph_pagelist_append(req->r_trail, src->cls.method_name,
328 src->cls.method_len);
329 ceph_pagelist_append(req->r_trail, src->cls.indata,
330 src->cls.indata_len);
332 case CEPH_OSD_OP_ROLLBACK:
333 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
335 case CEPH_OSD_OP_STARTSYNC:
337 case CEPH_OSD_OP_NOTIFY:
339 __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
340 __le32 timeout = cpu_to_le32(src->watch.timeout);
342 BUG_ON(!req->r_trail);
344 ceph_pagelist_append(req->r_trail,
345 &prot_ver, sizeof(prot_ver));
346 ceph_pagelist_append(req->r_trail,
347 &timeout, sizeof(timeout));
349 case CEPH_OSD_OP_NOTIFY_ACK:
350 case CEPH_OSD_OP_WATCH:
351 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
352 dst->watch.ver = cpu_to_le64(src->watch.ver);
353 dst->watch.flag = src->watch.flag;
356 pr_err("unrecognized osd opcode %d\n", dst->op);
360 dst->payload_len = cpu_to_le32(src->payload_len);
364 * build new request AND message
367 void ceph_osdc_build_request(struct ceph_osd_request *req,
369 struct ceph_osd_req_op *src_ops,
370 struct ceph_snap_context *snapc,
371 struct timespec *mtime,
375 struct ceph_msg *msg = req->r_request;
376 struct ceph_osd_request_head *head;
377 struct ceph_osd_req_op *src_op;
378 struct ceph_osd_op *op;
380 int num_op = get_num_ops(src_ops, NULL);
381 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
382 int flags = req->r_flags;
386 head = msg->front.iov_base;
387 op = (void *)(head + 1);
388 p = (void *)(op + num_op);
390 req->r_snapc = ceph_get_snap_context(snapc);
392 head->client_inc = cpu_to_le32(1); /* always, for now. */
393 head->flags = cpu_to_le32(flags);
394 if (flags & CEPH_OSD_FLAG_WRITE)
395 ceph_encode_timespec(&head->mtime, mtime);
396 head->num_ops = cpu_to_le16(num_op);
400 head->object_len = cpu_to_le32(oid_len);
401 memcpy(p, oid, oid_len);
406 osd_req_encode_op(req, op, src_op);
412 data_len += req->r_trail->length;
415 head->snap_seq = cpu_to_le64(snapc->seq);
416 head->num_snaps = cpu_to_le32(snapc->num_snaps);
417 for (i = 0; i < snapc->num_snaps; i++) {
418 put_unaligned_le64(snapc->snaps[i], p);
423 if (flags & CEPH_OSD_FLAG_WRITE) {
424 req->r_request->hdr.data_off = cpu_to_le16(off);
425 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
426 } else if (data_len) {
427 req->r_request->hdr.data_off = 0;
428 req->r_request->hdr.data_len = cpu_to_le32(data_len);
431 req->r_request->page_alignment = req->r_page_alignment;
433 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
434 msg_size = p - msg->front.iov_base;
435 msg->front.iov_len = msg_size;
436 msg->hdr.front_len = cpu_to_le32(msg_size);
439 EXPORT_SYMBOL(ceph_osdc_build_request);
442 * build new request AND message, calculate layout, and adjust file
445 * if the file was recently truncated, we include information about its
446 * old and new size so that the object can be updated appropriately. (we
447 * avoid synchronously deleting truncated objects because it's slow.)
449 * if @do_sync, include a 'startsync' command so that the osd will flush
452 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
453 struct ceph_file_layout *layout,
454 struct ceph_vino vino,
456 int opcode, int flags,
457 struct ceph_snap_context *snapc,
461 struct timespec *mtime,
462 bool use_mempool, int num_reply,
465 struct ceph_osd_req_op ops[3];
466 struct ceph_osd_request *req;
470 ops[0].extent.truncate_seq = truncate_seq;
471 ops[0].extent.truncate_size = truncate_size;
472 ops[0].payload_len = 0;
475 ops[1].op = CEPH_OSD_OP_STARTSYNC;
476 ops[1].payload_len = 0;
481 req = ceph_osdc_alloc_request(osdc, flags,
484 GFP_NOFS, NULL, NULL);
486 return ERR_PTR(-ENOMEM);
488 /* calculate max write size */
489 r = calc_layout(osdc, vino, layout, off, plen, req, ops);
492 req->r_file_layout = *layout; /* keep a copy */
494 /* in case it differs from natural (file) alignment that
495 calc_layout filled in for us */
496 req->r_num_pages = calc_pages_for(page_align, *plen);
497 req->r_page_alignment = page_align;
499 ceph_osdc_build_request(req, off, plen, ops,
502 req->r_oid, req->r_oid_len);
506 EXPORT_SYMBOL(ceph_osdc_new_request);
509 * We keep osd requests in an rbtree, sorted by ->r_tid.
511 static void __insert_request(struct ceph_osd_client *osdc,
512 struct ceph_osd_request *new)
514 struct rb_node **p = &osdc->requests.rb_node;
515 struct rb_node *parent = NULL;
516 struct ceph_osd_request *req = NULL;
520 req = rb_entry(parent, struct ceph_osd_request, r_node);
521 if (new->r_tid < req->r_tid)
523 else if (new->r_tid > req->r_tid)
529 rb_link_node(&new->r_node, parent, p);
530 rb_insert_color(&new->r_node, &osdc->requests);
533 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
536 struct ceph_osd_request *req;
537 struct rb_node *n = osdc->requests.rb_node;
540 req = rb_entry(n, struct ceph_osd_request, r_node);
541 if (tid < req->r_tid)
543 else if (tid > req->r_tid)
551 static struct ceph_osd_request *
552 __lookup_request_ge(struct ceph_osd_client *osdc,
555 struct ceph_osd_request *req;
556 struct rb_node *n = osdc->requests.rb_node;
559 req = rb_entry(n, struct ceph_osd_request, r_node);
560 if (tid < req->r_tid) {
564 } else if (tid > req->r_tid) {
574 * Resubmit requests pending on the given osd.
576 static void __kick_osd_requests(struct ceph_osd_client *osdc,
577 struct ceph_osd *osd)
579 struct ceph_osd_request *req, *nreq;
582 dout("__kick_osd_requests osd%d\n", osd->o_osd);
583 err = __reset_osd(osdc, osd);
587 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
588 list_move(&req->r_req_lru_item, &osdc->req_unsent);
589 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
592 req->r_flags |= CEPH_OSD_FLAG_RETRY;
595 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
598 * reregister request prior to unregistering linger so
599 * that r_osd is preserved.
601 BUG_ON(!list_empty(&req->r_req_lru_item));
602 __register_request(osdc, req);
603 list_add(&req->r_req_lru_item, &osdc->req_unsent);
604 list_add(&req->r_osd_item, &req->r_osd->o_requests);
605 __unregister_linger_request(osdc, req);
606 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
612 * If the osd connection drops, we need to resubmit all requests.
614 static void osd_reset(struct ceph_connection *con)
616 struct ceph_osd *osd = con->private;
617 struct ceph_osd_client *osdc;
621 dout("osd_reset osd%d\n", osd->o_osd);
623 down_read(&osdc->map_sem);
624 mutex_lock(&osdc->request_mutex);
625 __kick_osd_requests(osdc, osd);
626 mutex_unlock(&osdc->request_mutex);
628 up_read(&osdc->map_sem);
632 * Track open sessions with osds.
634 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
636 struct ceph_osd *osd;
638 osd = kzalloc(sizeof(*osd), GFP_NOFS);
642 atomic_set(&osd->o_ref, 1);
645 INIT_LIST_HEAD(&osd->o_requests);
646 INIT_LIST_HEAD(&osd->o_linger_requests);
647 INIT_LIST_HEAD(&osd->o_osd_lru);
648 osd->o_incarnation = 1;
650 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
652 INIT_LIST_HEAD(&osd->o_keepalive_item);
656 static struct ceph_osd *get_osd(struct ceph_osd *osd)
658 if (atomic_inc_not_zero(&osd->o_ref)) {
659 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
660 atomic_read(&osd->o_ref));
663 dout("get_osd %p FAIL\n", osd);
668 static void put_osd(struct ceph_osd *osd)
670 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
671 atomic_read(&osd->o_ref) - 1);
672 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
673 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
675 if (ac->ops && ac->ops->destroy_authorizer)
676 ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
682 * remove an osd from our map
684 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
686 dout("__remove_osd %p\n", osd);
687 BUG_ON(!list_empty(&osd->o_requests));
688 rb_erase(&osd->o_node, &osdc->osds);
689 list_del_init(&osd->o_osd_lru);
690 ceph_con_close(&osd->o_con);
694 static void remove_all_osds(struct ceph_osd_client *osdc)
696 dout("%s %p\n", __func__, osdc);
697 mutex_lock(&osdc->request_mutex);
698 while (!RB_EMPTY_ROOT(&osdc->osds)) {
699 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
700 struct ceph_osd, o_node);
701 __remove_osd(osdc, osd);
703 mutex_unlock(&osdc->request_mutex);
706 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
707 struct ceph_osd *osd)
709 dout("__move_osd_to_lru %p\n", osd);
710 BUG_ON(!list_empty(&osd->o_osd_lru));
711 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
712 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
715 static void __remove_osd_from_lru(struct ceph_osd *osd)
717 dout("__remove_osd_from_lru %p\n", osd);
718 if (!list_empty(&osd->o_osd_lru))
719 list_del_init(&osd->o_osd_lru);
722 static void remove_old_osds(struct ceph_osd_client *osdc)
724 struct ceph_osd *osd, *nosd;
726 dout("__remove_old_osds %p\n", osdc);
727 mutex_lock(&osdc->request_mutex);
728 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
729 if (time_before(jiffies, osd->lru_ttl))
731 __remove_osd(osdc, osd);
733 mutex_unlock(&osdc->request_mutex);
739 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
741 struct ceph_osd_request *req;
744 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
745 if (list_empty(&osd->o_requests) &&
746 list_empty(&osd->o_linger_requests)) {
747 __remove_osd(osdc, osd);
748 } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
749 &osd->o_con.peer_addr,
750 sizeof(osd->o_con.peer_addr)) == 0 &&
751 !ceph_con_opened(&osd->o_con)) {
752 dout(" osd addr hasn't changed and connection never opened,"
753 " letting msgr retry");
754 /* touch each r_stamp for handle_timeout()'s benfit */
755 list_for_each_entry(req, &osd->o_requests, r_osd_item)
756 req->r_stamp = jiffies;
759 ceph_con_close(&osd->o_con);
760 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
761 &osdc->osdmap->osd_addr[osd->o_osd]);
762 osd->o_incarnation++;
767 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
769 struct rb_node **p = &osdc->osds.rb_node;
770 struct rb_node *parent = NULL;
771 struct ceph_osd *osd = NULL;
773 dout("__insert_osd %p osd%d\n", new, new->o_osd);
776 osd = rb_entry(parent, struct ceph_osd, o_node);
777 if (new->o_osd < osd->o_osd)
779 else if (new->o_osd > osd->o_osd)
785 rb_link_node(&new->o_node, parent, p);
786 rb_insert_color(&new->o_node, &osdc->osds);
789 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
791 struct ceph_osd *osd;
792 struct rb_node *n = osdc->osds.rb_node;
795 osd = rb_entry(n, struct ceph_osd, o_node);
798 else if (o > osd->o_osd)
806 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
808 schedule_delayed_work(&osdc->timeout_work,
809 osdc->client->options->osd_keepalive_timeout * HZ);
812 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
814 cancel_delayed_work(&osdc->timeout_work);
818 * Register request, assign tid. If this is the first request, set up
821 static void __register_request(struct ceph_osd_client *osdc,
822 struct ceph_osd_request *req)
824 req->r_tid = ++osdc->last_tid;
825 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
826 dout("__register_request %p tid %lld\n", req, req->r_tid);
827 __insert_request(osdc, req);
828 ceph_osdc_get_request(req);
829 osdc->num_requests++;
830 if (osdc->num_requests == 1) {
831 dout(" first request, scheduling timeout\n");
832 __schedule_osd_timeout(osdc);
836 static void register_request(struct ceph_osd_client *osdc,
837 struct ceph_osd_request *req)
839 mutex_lock(&osdc->request_mutex);
840 __register_request(osdc, req);
841 mutex_unlock(&osdc->request_mutex);
845 * called under osdc->request_mutex
847 static void __unregister_request(struct ceph_osd_client *osdc,
848 struct ceph_osd_request *req)
850 if (RB_EMPTY_NODE(&req->r_node)) {
851 dout("__unregister_request %p tid %lld not registered\n",
856 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
857 rb_erase(&req->r_node, &osdc->requests);
858 osdc->num_requests--;
861 /* make sure the original request isn't in flight. */
862 ceph_msg_revoke(req->r_request);
864 list_del_init(&req->r_osd_item);
865 if (list_empty(&req->r_osd->o_requests) &&
866 list_empty(&req->r_osd->o_linger_requests)) {
867 dout("moving osd to %p lru\n", req->r_osd);
868 __move_osd_to_lru(osdc, req->r_osd);
870 if (list_empty(&req->r_linger_item))
874 ceph_osdc_put_request(req);
876 list_del_init(&req->r_req_lru_item);
877 if (osdc->num_requests == 0) {
878 dout(" no requests, canceling timeout\n");
879 __cancel_osd_timeout(osdc);
884 * Cancel a previously queued request message
886 static void __cancel_request(struct ceph_osd_request *req)
888 if (req->r_sent && req->r_osd) {
889 ceph_msg_revoke(req->r_request);
894 static void __register_linger_request(struct ceph_osd_client *osdc,
895 struct ceph_osd_request *req)
897 dout("__register_linger_request %p\n", req);
898 list_add_tail(&req->r_linger_item, &osdc->req_linger);
900 list_add_tail(&req->r_linger_osd,
901 &req->r_osd->o_linger_requests);
904 static void __unregister_linger_request(struct ceph_osd_client *osdc,
905 struct ceph_osd_request *req)
907 dout("__unregister_linger_request %p\n", req);
909 list_del_init(&req->r_linger_item);
910 list_del_init(&req->r_linger_osd);
912 if (list_empty(&req->r_osd->o_requests) &&
913 list_empty(&req->r_osd->o_linger_requests)) {
914 dout("moving osd to %p lru\n", req->r_osd);
915 __move_osd_to_lru(osdc, req->r_osd);
917 if (list_empty(&req->r_osd_item))
922 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
923 struct ceph_osd_request *req)
925 mutex_lock(&osdc->request_mutex);
927 __unregister_linger_request(osdc, req);
928 ceph_osdc_put_request(req);
930 mutex_unlock(&osdc->request_mutex);
932 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
934 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
935 struct ceph_osd_request *req)
937 if (!req->r_linger) {
938 dout("set_request_linger %p\n", req);
941 * caller is now responsible for calling
942 * unregister_linger_request
944 ceph_osdc_get_request(req);
947 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
950 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
951 * (as needed), and set the request r_osd appropriately. If there is
952 * no up osd, set r_osd to NULL. Move the request to the appropriate list
953 * (unsent, homeless) or leave on in-flight lru.
955 * Return 0 if unchanged, 1 if changed, or negative on error.
957 * Caller should hold map_sem for read and request_mutex.
959 static int __map_request(struct ceph_osd_client *osdc,
960 struct ceph_osd_request *req, int force_resend)
962 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
964 int acting[CEPH_PG_MAX_SIZE];
968 dout("map_request %p tid %lld\n", req, req->r_tid);
969 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
970 &req->r_file_layout, osdc->osdmap);
972 list_move(&req->r_req_lru_item, &osdc->req_notarget);
975 pgid = reqhead->layout.ol_pgid;
978 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
984 if ((!force_resend &&
985 req->r_osd && req->r_osd->o_osd == o &&
986 req->r_sent >= req->r_osd->o_incarnation &&
987 req->r_num_pg_osds == num &&
988 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
989 (req->r_osd == NULL && o == -1))
990 return 0; /* no change */
992 dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
993 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
994 req->r_osd ? req->r_osd->o_osd : -1);
996 /* record full pg acting set */
997 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
998 req->r_num_pg_osds = num;
1001 __cancel_request(req);
1002 list_del_init(&req->r_osd_item);
1006 req->r_osd = __lookup_osd(osdc, o);
1007 if (!req->r_osd && o >= 0) {
1009 req->r_osd = create_osd(osdc, o);
1011 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1015 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1016 __insert_osd(osdc, req->r_osd);
1018 ceph_con_open(&req->r_osd->o_con,
1019 CEPH_ENTITY_TYPE_OSD, o,
1020 &osdc->osdmap->osd_addr[o]);
1024 __remove_osd_from_lru(req->r_osd);
1025 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1026 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1028 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1030 err = 1; /* osd or pg changed */
1037 * caller should hold map_sem (for read) and request_mutex
1039 static void __send_request(struct ceph_osd_client *osdc,
1040 struct ceph_osd_request *req)
1042 struct ceph_osd_request_head *reqhead;
1044 dout("send_request %p tid %llu to osd%d flags %d\n",
1045 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1047 reqhead = req->r_request->front.iov_base;
1048 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1049 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
1050 reqhead->reassert_version = req->r_reassert_version;
1052 req->r_stamp = jiffies;
1053 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1055 ceph_msg_get(req->r_request); /* send consumes a ref */
1056 ceph_con_send(&req->r_osd->o_con, req->r_request);
1057 req->r_sent = req->r_osd->o_incarnation;
1061 * Send any requests in the queue (req_unsent).
1063 static void send_queued(struct ceph_osd_client *osdc)
1065 struct ceph_osd_request *req, *tmp;
1067 dout("send_queued\n");
1068 mutex_lock(&osdc->request_mutex);
1069 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1070 __send_request(osdc, req);
1072 mutex_unlock(&osdc->request_mutex);
1076 * Timeout callback, called every N seconds when 1 or more osd
1077 * requests has been active for more than N seconds. When this
1078 * happens, we ping all OSDs with requests who have timed out to
1079 * ensure any communications channel reset is detected. Reset the
1080 * request timeouts another N seconds in the future as we go.
1081 * Reschedule the timeout event another N seconds in future (unless
1082 * there are no open requests).
1084 static void handle_timeout(struct work_struct *work)
1086 struct ceph_osd_client *osdc =
1087 container_of(work, struct ceph_osd_client, timeout_work.work);
1088 struct ceph_osd_request *req;
1089 struct ceph_osd *osd;
1090 unsigned long keepalive =
1091 osdc->client->options->osd_keepalive_timeout * HZ;
1092 struct list_head slow_osds;
1094 down_read(&osdc->map_sem);
1096 ceph_monc_request_next_osdmap(&osdc->client->monc);
1098 mutex_lock(&osdc->request_mutex);
1101 * ping osds that are a bit slow. this ensures that if there
1102 * is a break in the TCP connection we will notice, and reopen
1103 * a connection with that osd (from the fault callback).
1105 INIT_LIST_HEAD(&slow_osds);
1106 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1107 if (time_before(jiffies, req->r_stamp + keepalive))
1112 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1113 req->r_tid, osd->o_osd);
1114 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1116 while (!list_empty(&slow_osds)) {
1117 osd = list_entry(slow_osds.next, struct ceph_osd,
1119 list_del_init(&osd->o_keepalive_item);
1120 ceph_con_keepalive(&osd->o_con);
1123 __schedule_osd_timeout(osdc);
1124 mutex_unlock(&osdc->request_mutex);
1126 up_read(&osdc->map_sem);
1129 static void handle_osds_timeout(struct work_struct *work)
1131 struct ceph_osd_client *osdc =
1132 container_of(work, struct ceph_osd_client,
1133 osds_timeout_work.work);
1134 unsigned long delay =
1135 osdc->client->options->osd_idle_ttl * HZ >> 2;
1137 dout("osds timeout\n");
1138 down_read(&osdc->map_sem);
1139 remove_old_osds(osdc);
1140 up_read(&osdc->map_sem);
1142 schedule_delayed_work(&osdc->osds_timeout_work,
1143 round_jiffies_relative(delay));
1146 static void complete_request(struct ceph_osd_request *req)
1148 if (req->r_safe_callback)
1149 req->r_safe_callback(req, NULL);
1150 complete_all(&req->r_safe_completion); /* fsync waiter */
1154 * handle osd op reply. either call the callback if it is specified,
1155 * or do the completion to wake up the waiting thread.
1157 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1158 struct ceph_connection *con)
1160 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1161 struct ceph_osd_request *req;
1163 int numops, object_len, flags;
1166 tid = le64_to_cpu(msg->hdr.tid);
1167 if (msg->front.iov_len < sizeof(*rhead))
1169 numops = le32_to_cpu(rhead->num_ops);
1170 object_len = le32_to_cpu(rhead->object_len);
1171 result = le32_to_cpu(rhead->result);
1172 if (msg->front.iov_len != sizeof(*rhead) + object_len +
1173 numops * sizeof(struct ceph_osd_op))
1175 dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1177 mutex_lock(&osdc->request_mutex);
1178 req = __lookup_request(osdc, tid);
1180 dout("handle_reply tid %llu dne\n", tid);
1181 mutex_unlock(&osdc->request_mutex);
1184 ceph_osdc_get_request(req);
1185 flags = le32_to_cpu(rhead->flags);
1188 * if this connection filled our message, drop our reference now, to
1189 * avoid a (safe but slower) revoke later.
1191 if (req->r_con_filling_msg == con && req->r_reply == msg) {
1192 dout(" dropping con_filling_msg ref %p\n", con);
1193 req->r_con_filling_msg = NULL;
1197 if (!req->r_got_reply) {
1200 req->r_result = le32_to_cpu(rhead->result);
1201 bytes = le32_to_cpu(msg->hdr.data_len);
1202 dout("handle_reply result %d bytes %d\n", req->r_result,
1204 if (req->r_result == 0)
1205 req->r_result = bytes;
1207 /* in case this is a write and we need to replay, */
1208 req->r_reassert_version = rhead->reassert_version;
1210 req->r_got_reply = 1;
1211 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1212 dout("handle_reply tid %llu dup ack\n", tid);
1213 mutex_unlock(&osdc->request_mutex);
1217 dout("handle_reply tid %llu flags %d\n", tid, flags);
1219 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1220 __register_linger_request(osdc, req);
1222 /* either this is a read, or we got the safe response */
1224 (flags & CEPH_OSD_FLAG_ONDISK) ||
1225 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1226 __unregister_request(osdc, req);
1228 mutex_unlock(&osdc->request_mutex);
1230 if (req->r_callback)
1231 req->r_callback(req, msg);
1233 complete_all(&req->r_completion);
1235 if (flags & CEPH_OSD_FLAG_ONDISK)
1236 complete_request(req);
1239 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1240 ceph_osdc_put_request(req);
1244 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1245 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1246 (int)sizeof(*rhead));
1250 static void reset_changed_osds(struct ceph_osd_client *osdc)
1252 struct rb_node *p, *n;
1254 for (p = rb_first(&osdc->osds); p; p = n) {
1255 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1258 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1259 memcmp(&osd->o_con.peer_addr,
1260 ceph_osd_addr(osdc->osdmap,
1262 sizeof(struct ceph_entity_addr)) != 0)
1263 __reset_osd(osdc, osd);
1268 * Requeue requests whose mapping to an OSD has changed. If requests map to
1269 * no osd, request a new map.
1271 * Caller should hold map_sem for read and request_mutex.
1273 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1275 struct ceph_osd_request *req, *nreq;
1280 dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1281 mutex_lock(&osdc->request_mutex);
1282 for (p = rb_first(&osdc->requests); p; ) {
1283 req = rb_entry(p, struct ceph_osd_request, r_node);
1285 err = __map_request(osdc, req, force_resend);
1287 continue; /* error */
1288 if (req->r_osd == NULL) {
1289 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1290 needmap++; /* request a newer map */
1291 } else if (err > 0) {
1292 if (!req->r_linger) {
1293 dout("%p tid %llu requeued on osd%d\n", req,
1295 req->r_osd ? req->r_osd->o_osd : -1);
1296 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1299 if (req->r_linger && list_empty(&req->r_linger_item)) {
1301 * register as a linger so that we will
1302 * re-submit below and get a new tid
1304 dout("%p tid %llu restart on osd%d\n",
1306 req->r_osd ? req->r_osd->o_osd : -1);
1307 __register_linger_request(osdc, req);
1308 __unregister_request(osdc, req);
1312 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1314 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1316 err = __map_request(osdc, req, force_resend);
1318 continue; /* no change and no osd was specified */
1320 continue; /* hrm! */
1321 if (req->r_osd == NULL) {
1322 dout("tid %llu maps to no valid osd\n", req->r_tid);
1323 needmap++; /* request a newer map */
1327 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1328 req->r_osd ? req->r_osd->o_osd : -1);
1329 __unregister_linger_request(osdc, req);
1330 __register_request(osdc, req);
1332 mutex_unlock(&osdc->request_mutex);
1335 dout("%d requests for down osds, need new map\n", needmap);
1336 ceph_monc_request_next_osdmap(&osdc->client->monc);
1342 * Process updated osd map.
1344 * The message contains any number of incremental and full maps, normally
1345 * indicating some sort of topology change in the cluster. Kick requests
1346 * off to different OSDs as needed.
1348 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1350 void *p, *end, *next;
1351 u32 nr_maps, maplen;
1353 struct ceph_osdmap *newmap = NULL, *oldmap;
1355 struct ceph_fsid fsid;
1357 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1358 p = msg->front.iov_base;
1359 end = p + msg->front.iov_len;
1362 ceph_decode_need(&p, end, sizeof(fsid), bad);
1363 ceph_decode_copy(&p, &fsid, sizeof(fsid));
1364 if (ceph_check_fsid(osdc->client, &fsid) < 0)
1367 down_write(&osdc->map_sem);
1369 /* incremental maps */
1370 ceph_decode_32_safe(&p, end, nr_maps, bad);
1371 dout(" %d inc maps\n", nr_maps);
1372 while (nr_maps > 0) {
1373 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1374 epoch = ceph_decode_32(&p);
1375 maplen = ceph_decode_32(&p);
1376 ceph_decode_need(&p, end, maplen, bad);
1378 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1379 dout("applying incremental map %u len %d\n",
1381 newmap = osdmap_apply_incremental(&p, next,
1383 &osdc->client->msgr);
1384 if (IS_ERR(newmap)) {
1385 err = PTR_ERR(newmap);
1389 if (newmap != osdc->osdmap) {
1390 ceph_osdmap_destroy(osdc->osdmap);
1391 osdc->osdmap = newmap;
1393 kick_requests(osdc, 0);
1394 reset_changed_osds(osdc);
1396 dout("ignoring incremental map %u len %d\n",
1406 ceph_decode_32_safe(&p, end, nr_maps, bad);
1407 dout(" %d full maps\n", nr_maps);
1409 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1410 epoch = ceph_decode_32(&p);
1411 maplen = ceph_decode_32(&p);
1412 ceph_decode_need(&p, end, maplen, bad);
1414 dout("skipping non-latest full map %u len %d\n",
1416 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1417 dout("skipping full map %u len %d, "
1418 "older than our %u\n", epoch, maplen,
1419 osdc->osdmap->epoch);
1421 int skipped_map = 0;
1423 dout("taking full map %u len %d\n", epoch, maplen);
1424 newmap = osdmap_decode(&p, p+maplen);
1425 if (IS_ERR(newmap)) {
1426 err = PTR_ERR(newmap);
1430 oldmap = osdc->osdmap;
1431 osdc->osdmap = newmap;
1433 if (oldmap->epoch + 1 < newmap->epoch)
1435 ceph_osdmap_destroy(oldmap);
1437 kick_requests(osdc, skipped_map);
1444 downgrade_write(&osdc->map_sem);
1445 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1448 * subscribe to subsequent osdmap updates if full to ensure
1449 * we find out when we are no longer full and stop returning
1452 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1453 ceph_monc_request_next_osdmap(&osdc->client->monc);
1456 up_read(&osdc->map_sem);
1457 wake_up_all(&osdc->client->auth_wq);
1461 pr_err("osdc handle_map corrupt msg\n");
1463 up_write(&osdc->map_sem);
1468 * watch/notify callback event infrastructure
1470 * These callbacks are used both for watch and notify operations.
1472 static void __release_event(struct kref *kref)
1474 struct ceph_osd_event *event =
1475 container_of(kref, struct ceph_osd_event, kref);
1477 dout("__release_event %p\n", event);
1481 static void get_event(struct ceph_osd_event *event)
1483 kref_get(&event->kref);
1486 void ceph_osdc_put_event(struct ceph_osd_event *event)
1488 kref_put(&event->kref, __release_event);
1490 EXPORT_SYMBOL(ceph_osdc_put_event);
1492 static void __insert_event(struct ceph_osd_client *osdc,
1493 struct ceph_osd_event *new)
1495 struct rb_node **p = &osdc->event_tree.rb_node;
1496 struct rb_node *parent = NULL;
1497 struct ceph_osd_event *event = NULL;
1501 event = rb_entry(parent, struct ceph_osd_event, node);
1502 if (new->cookie < event->cookie)
1504 else if (new->cookie > event->cookie)
1505 p = &(*p)->rb_right;
1510 rb_link_node(&new->node, parent, p);
1511 rb_insert_color(&new->node, &osdc->event_tree);
1514 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1517 struct rb_node **p = &osdc->event_tree.rb_node;
1518 struct rb_node *parent = NULL;
1519 struct ceph_osd_event *event = NULL;
1523 event = rb_entry(parent, struct ceph_osd_event, node);
1524 if (cookie < event->cookie)
1526 else if (cookie > event->cookie)
1527 p = &(*p)->rb_right;
1534 static void __remove_event(struct ceph_osd_event *event)
1536 struct ceph_osd_client *osdc = event->osdc;
1538 if (!RB_EMPTY_NODE(&event->node)) {
1539 dout("__remove_event removed %p\n", event);
1540 rb_erase(&event->node, &osdc->event_tree);
1541 ceph_osdc_put_event(event);
1543 dout("__remove_event didn't remove %p\n", event);
1547 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1548 void (*event_cb)(u64, u64, u8, void *),
1549 int one_shot, void *data,
1550 struct ceph_osd_event **pevent)
1552 struct ceph_osd_event *event;
1554 event = kmalloc(sizeof(*event), GFP_NOIO);
1558 dout("create_event %p\n", event);
1559 event->cb = event_cb;
1560 event->one_shot = one_shot;
1563 INIT_LIST_HEAD(&event->osd_node);
1564 kref_init(&event->kref); /* one ref for us */
1565 kref_get(&event->kref); /* one ref for the caller */
1566 init_completion(&event->completion);
1568 spin_lock(&osdc->event_lock);
1569 event->cookie = ++osdc->event_count;
1570 __insert_event(osdc, event);
1571 spin_unlock(&osdc->event_lock);
1576 EXPORT_SYMBOL(ceph_osdc_create_event);
1578 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1580 struct ceph_osd_client *osdc = event->osdc;
1582 dout("cancel_event %p\n", event);
1583 spin_lock(&osdc->event_lock);
1584 __remove_event(event);
1585 spin_unlock(&osdc->event_lock);
1586 ceph_osdc_put_event(event); /* caller's */
1588 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1591 static void do_event_work(struct work_struct *work)
1593 struct ceph_osd_event_work *event_work =
1594 container_of(work, struct ceph_osd_event_work, work);
1595 struct ceph_osd_event *event = event_work->event;
1596 u64 ver = event_work->ver;
1597 u64 notify_id = event_work->notify_id;
1598 u8 opcode = event_work->opcode;
1600 dout("do_event_work completing %p\n", event);
1601 event->cb(ver, notify_id, opcode, event->data);
1602 complete(&event->completion);
1603 dout("do_event_work completed %p\n", event);
1604 ceph_osdc_put_event(event);
1610 * Process osd watch notifications
1612 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1616 u64 cookie, ver, notify_id;
1618 struct ceph_osd_event *event;
1619 struct ceph_osd_event_work *event_work;
1621 p = msg->front.iov_base;
1622 end = p + msg->front.iov_len;
1624 ceph_decode_8_safe(&p, end, proto_ver, bad);
1625 ceph_decode_8_safe(&p, end, opcode, bad);
1626 ceph_decode_64_safe(&p, end, cookie, bad);
1627 ceph_decode_64_safe(&p, end, ver, bad);
1628 ceph_decode_64_safe(&p, end, notify_id, bad);
1630 spin_lock(&osdc->event_lock);
1631 event = __find_event(osdc, cookie);
1634 if (event->one_shot)
1635 __remove_event(event);
1637 spin_unlock(&osdc->event_lock);
1638 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1639 cookie, ver, event);
1641 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1643 dout("ERROR: could not allocate event_work\n");
1646 INIT_WORK(&event_work->work, do_event_work);
1647 event_work->event = event;
1648 event_work->ver = ver;
1649 event_work->notify_id = notify_id;
1650 event_work->opcode = opcode;
1651 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1652 dout("WARNING: failed to queue notify event work\n");
1660 complete(&event->completion);
1661 ceph_osdc_put_event(event);
1665 pr_err("osdc handle_watch_notify corrupt msg\n");
1669 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1673 dout("wait_event %p\n", event);
1674 err = wait_for_completion_interruptible_timeout(&event->completion,
1676 ceph_osdc_put_event(event);
1679 dout("wait_event %p returns %d\n", event, err);
1682 EXPORT_SYMBOL(ceph_osdc_wait_event);
1685 * Register request, send initial attempt.
1687 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1688 struct ceph_osd_request *req,
1693 req->r_request->pages = req->r_pages;
1694 req->r_request->nr_pages = req->r_num_pages;
1696 req->r_request->bio = req->r_bio;
1698 req->r_request->trail = req->r_trail;
1700 register_request(osdc, req);
1702 down_read(&osdc->map_sem);
1703 mutex_lock(&osdc->request_mutex);
1705 * a racing kick_requests() may have sent the message for us
1706 * while we dropped request_mutex above, so only send now if
1707 * the request still han't been touched yet.
1709 if (req->r_sent == 0) {
1710 rc = __map_request(osdc, req, 0);
1713 dout("osdc_start_request failed map, "
1714 " will retry %lld\n", req->r_tid);
1719 if (req->r_osd == NULL) {
1720 dout("send_request %p no up osds in pg\n", req);
1721 ceph_monc_request_next_osdmap(&osdc->client->monc);
1723 __send_request(osdc, req);
1729 mutex_unlock(&osdc->request_mutex);
1730 up_read(&osdc->map_sem);
1733 EXPORT_SYMBOL(ceph_osdc_start_request);
1736 * wait for a request to complete
1738 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1739 struct ceph_osd_request *req)
1743 rc = wait_for_completion_interruptible(&req->r_completion);
1745 mutex_lock(&osdc->request_mutex);
1746 __cancel_request(req);
1747 __unregister_request(osdc, req);
1748 mutex_unlock(&osdc->request_mutex);
1749 complete_request(req);
1750 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1754 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1755 return req->r_result;
1757 EXPORT_SYMBOL(ceph_osdc_wait_request);
1760 * sync - wait for all in-flight requests to flush. avoid starvation.
1762 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1764 struct ceph_osd_request *req;
1765 u64 last_tid, next_tid = 0;
1767 mutex_lock(&osdc->request_mutex);
1768 last_tid = osdc->last_tid;
1770 req = __lookup_request_ge(osdc, next_tid);
1773 if (req->r_tid > last_tid)
1776 next_tid = req->r_tid + 1;
1777 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1780 ceph_osdc_get_request(req);
1781 mutex_unlock(&osdc->request_mutex);
1782 dout("sync waiting on tid %llu (last is %llu)\n",
1783 req->r_tid, last_tid);
1784 wait_for_completion(&req->r_safe_completion);
1785 mutex_lock(&osdc->request_mutex);
1786 ceph_osdc_put_request(req);
1788 mutex_unlock(&osdc->request_mutex);
1789 dout("sync done (thru tid %llu)\n", last_tid);
1791 EXPORT_SYMBOL(ceph_osdc_sync);
1796 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1801 osdc->client = client;
1802 osdc->osdmap = NULL;
1803 init_rwsem(&osdc->map_sem);
1804 init_completion(&osdc->map_waiters);
1805 osdc->last_requested_map = 0;
1806 mutex_init(&osdc->request_mutex);
1808 osdc->osds = RB_ROOT;
1809 INIT_LIST_HEAD(&osdc->osd_lru);
1810 osdc->requests = RB_ROOT;
1811 INIT_LIST_HEAD(&osdc->req_lru);
1812 INIT_LIST_HEAD(&osdc->req_unsent);
1813 INIT_LIST_HEAD(&osdc->req_notarget);
1814 INIT_LIST_HEAD(&osdc->req_linger);
1815 osdc->num_requests = 0;
1816 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1817 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1818 spin_lock_init(&osdc->event_lock);
1819 osdc->event_tree = RB_ROOT;
1820 osdc->event_count = 0;
1822 schedule_delayed_work(&osdc->osds_timeout_work,
1823 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1826 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1827 sizeof(struct ceph_osd_request));
1828 if (!osdc->req_mempool)
1831 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1832 OSD_OP_FRONT_LEN, 10, true,
1836 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1837 OSD_OPREPLY_FRONT_LEN, 10, true,
1842 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1843 if (IS_ERR(osdc->notify_wq)) {
1844 err = PTR_ERR(osdc->notify_wq);
1845 osdc->notify_wq = NULL;
1851 ceph_msgpool_destroy(&osdc->msgpool_op);
1853 mempool_destroy(osdc->req_mempool);
1857 EXPORT_SYMBOL(ceph_osdc_init);
1859 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1861 flush_workqueue(osdc->notify_wq);
1862 destroy_workqueue(osdc->notify_wq);
1863 cancel_delayed_work_sync(&osdc->timeout_work);
1864 cancel_delayed_work_sync(&osdc->osds_timeout_work);
1866 ceph_osdmap_destroy(osdc->osdmap);
1867 osdc->osdmap = NULL;
1869 remove_all_osds(osdc);
1870 mempool_destroy(osdc->req_mempool);
1871 ceph_msgpool_destroy(&osdc->msgpool_op);
1872 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1874 EXPORT_SYMBOL(ceph_osdc_stop);
1877 * Read some contiguous pages. If we cross a stripe boundary, shorten
1878 * *plen. Return number of bytes read, or error.
1880 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1881 struct ceph_vino vino, struct ceph_file_layout *layout,
1883 u32 truncate_seq, u64 truncate_size,
1884 struct page **pages, int num_pages, int page_align)
1886 struct ceph_osd_request *req;
1889 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1890 vino.snap, off, *plen);
1891 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1892 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1893 NULL, 0, truncate_seq, truncate_size, NULL,
1894 false, 1, page_align);
1896 return PTR_ERR(req);
1898 /* it may be a short read due to an object boundary */
1899 req->r_pages = pages;
1901 dout("readpages final extent is %llu~%llu (%d pages align %d)\n",
1902 off, *plen, req->r_num_pages, page_align);
1904 rc = ceph_osdc_start_request(osdc, req, false);
1906 rc = ceph_osdc_wait_request(osdc, req);
1908 ceph_osdc_put_request(req);
1909 dout("readpages result %d\n", rc);
1912 EXPORT_SYMBOL(ceph_osdc_readpages);
1915 * do a synchronous write on N pages
1917 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1918 struct ceph_file_layout *layout,
1919 struct ceph_snap_context *snapc,
1921 u32 truncate_seq, u64 truncate_size,
1922 struct timespec *mtime,
1923 struct page **pages, int num_pages,
1924 int flags, int do_sync, bool nofail)
1926 struct ceph_osd_request *req;
1928 int page_align = off & ~PAGE_MASK;
1930 BUG_ON(vino.snap != CEPH_NOSNAP);
1931 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1933 flags | CEPH_OSD_FLAG_ONDISK |
1934 CEPH_OSD_FLAG_WRITE,
1936 truncate_seq, truncate_size, mtime,
1937 nofail, 1, page_align);
1939 return PTR_ERR(req);
1941 /* it may be a short write due to an object boundary */
1942 req->r_pages = pages;
1943 dout("writepages %llu~%llu (%d pages)\n", off, len,
1946 rc = ceph_osdc_start_request(osdc, req, nofail);
1948 rc = ceph_osdc_wait_request(osdc, req);
1950 ceph_osdc_put_request(req);
1953 dout("writepages result %d\n", rc);
1956 EXPORT_SYMBOL(ceph_osdc_writepages);
1959 * handle incoming message
1961 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1963 struct ceph_osd *osd = con->private;
1964 struct ceph_osd_client *osdc;
1965 int type = le16_to_cpu(msg->hdr.type);
1972 case CEPH_MSG_OSD_MAP:
1973 ceph_osdc_handle_map(osdc, msg);
1975 case CEPH_MSG_OSD_OPREPLY:
1976 handle_reply(osdc, msg, con);
1978 case CEPH_MSG_WATCH_NOTIFY:
1979 handle_watch_notify(osdc, msg);
1983 pr_err("received unknown message type %d %s\n", type,
1984 ceph_msg_type_name(type));
1991 * lookup and return message for incoming reply. set up reply message
1994 static struct ceph_msg *get_reply(struct ceph_connection *con,
1995 struct ceph_msg_header *hdr,
1998 struct ceph_osd *osd = con->private;
1999 struct ceph_osd_client *osdc = osd->o_osdc;
2001 struct ceph_osd_request *req;
2002 int front = le32_to_cpu(hdr->front_len);
2003 int data_len = le32_to_cpu(hdr->data_len);
2006 tid = le64_to_cpu(hdr->tid);
2007 mutex_lock(&osdc->request_mutex);
2008 req = __lookup_request(osdc, tid);
2012 dout("get_reply unknown tid %llu from osd%d\n", tid,
2017 if (req->r_con_filling_msg) {
2018 dout("%s revoking msg %p from old con %p\n", __func__,
2019 req->r_reply, req->r_con_filling_msg);
2020 ceph_msg_revoke_incoming(req->r_reply);
2021 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2022 req->r_con_filling_msg = NULL;
2025 if (front > req->r_reply->front.iov_len) {
2026 pr_warning("get_reply front %d > preallocated %d\n",
2027 front, (int)req->r_reply->front.iov_len);
2028 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2031 ceph_msg_put(req->r_reply);
2034 m = ceph_msg_get(req->r_reply);
2037 int want = calc_pages_for(req->r_page_alignment, data_len);
2039 if (unlikely(req->r_num_pages < want)) {
2040 pr_warning("tid %lld reply has %d bytes %d pages, we"
2041 " had only %d pages ready\n", tid, data_len,
2042 want, req->r_num_pages);
2048 m->pages = req->r_pages;
2049 m->nr_pages = req->r_num_pages;
2050 m->page_alignment = req->r_page_alignment;
2052 m->bio = req->r_bio;
2056 req->r_con_filling_msg = con->ops->get(con);
2057 dout("get_reply tid %lld %p\n", tid, m);
2060 mutex_unlock(&osdc->request_mutex);
2065 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2066 struct ceph_msg_header *hdr,
2069 struct ceph_osd *osd = con->private;
2070 int type = le16_to_cpu(hdr->type);
2071 int front = le32_to_cpu(hdr->front_len);
2075 case CEPH_MSG_OSD_MAP:
2076 case CEPH_MSG_WATCH_NOTIFY:
2077 return ceph_msg_new(type, front, GFP_NOFS, false);
2078 case CEPH_MSG_OSD_OPREPLY:
2079 return get_reply(con, hdr, skip);
2081 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2089 * Wrappers to refcount containing ceph_osd struct
2091 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2093 struct ceph_osd *osd = con->private;
2099 static void put_osd_con(struct ceph_connection *con)
2101 struct ceph_osd *osd = con->private;
2109 * Note: returned pointer is the address of a structure that's
2110 * managed separately. Caller must *not* attempt to free it.
2112 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2113 int *proto, int force_new)
2115 struct ceph_osd *o = con->private;
2116 struct ceph_osd_client *osdc = o->o_osdc;
2117 struct ceph_auth_client *ac = osdc->client->monc.auth;
2118 struct ceph_auth_handshake *auth = &o->o_auth;
2120 if (force_new && auth->authorizer) {
2121 if (ac->ops && ac->ops->destroy_authorizer)
2122 ac->ops->destroy_authorizer(ac, auth->authorizer);
2123 auth->authorizer = NULL;
2125 if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2126 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2129 return ERR_PTR(ret);
2131 *proto = ac->protocol;
2137 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2139 struct ceph_osd *o = con->private;
2140 struct ceph_osd_client *osdc = o->o_osdc;
2141 struct ceph_auth_client *ac = osdc->client->monc.auth;
2144 * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2145 * XXX which do we do: succeed or fail?
2147 return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2150 static int invalidate_authorizer(struct ceph_connection *con)
2152 struct ceph_osd *o = con->private;
2153 struct ceph_osd_client *osdc = o->o_osdc;
2154 struct ceph_auth_client *ac = osdc->client->monc.auth;
2156 if (ac->ops && ac->ops->invalidate_authorizer)
2157 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2159 return ceph_monc_validate_auth(&osdc->client->monc);
2162 static const struct ceph_connection_operations osd_con_ops = {
2165 .dispatch = dispatch,
2166 .get_authorizer = get_authorizer,
2167 .verify_authorizer_reply = verify_authorizer_reply,
2168 .invalidate_authorizer = invalidate_authorizer,
2169 .alloc_msg = alloc_msg,