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_has_extent(int op)
37 return (op == CEPH_OSD_OP_READ ||
38 op == CEPH_OSD_OP_WRITE);
42 * Implement client access to distributed object storage cluster.
44 * All data objects are stored within a cluster/cloud of OSDs, or
45 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
46 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
47 * remote daemons serving up and coordinating consistent and safe
50 * Cluster membership and the mapping of data objects onto storage devices
51 * are described by the osd map.
53 * We keep track of pending OSD requests (read, write), resubmit
54 * requests to different OSDs when the cluster topology/data layout
55 * change, or retry the affected requests when the communications
56 * channel with an OSD is reset.
60 * calculate the mapping of a file extent onto an object, and fill out the
61 * request accordingly. shorten extent as necessary if it crosses an
64 * fill osd op in request message.
66 static int calc_layout(struct ceph_vino vino,
67 struct ceph_file_layout *layout,
69 struct ceph_osd_request *req,
70 struct ceph_osd_req_op *op)
79 r = ceph_calc_file_object_mapping(layout, off, orig_len, &bno,
83 if (objlen < orig_len) {
85 dout(" skipping last %llu, final file extent %llu~%llu\n",
86 orig_len - *plen, off, *plen);
89 if (op_has_extent(op->op)) {
90 u32 osize = le32_to_cpu(layout->fl_object_size);
91 op->extent.offset = objoff;
92 op->extent.length = objlen;
93 if (op->extent.truncate_size <= off - objoff) {
94 op->extent.truncate_size = 0;
96 op->extent.truncate_size -= off - objoff;
97 if (op->extent.truncate_size > osize)
98 op->extent.truncate_size = osize;
101 req->r_num_pages = calc_pages_for(off, *plen);
102 req->r_page_alignment = off & ~PAGE_MASK;
103 if (op->op == CEPH_OSD_OP_WRITE)
104 op->payload_len = *plen;
106 dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
107 bno, objoff, objlen, req->r_num_pages);
109 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
110 req->r_oid_len = strlen(req->r_oid);
118 void ceph_osdc_release_request(struct kref *kref)
120 struct ceph_osd_request *req = container_of(kref,
121 struct ceph_osd_request,
125 ceph_msg_put(req->r_request);
126 if (req->r_con_filling_msg) {
127 dout("%s revoking msg %p from con %p\n", __func__,
128 req->r_reply, req->r_con_filling_msg);
129 ceph_msg_revoke_incoming(req->r_reply);
130 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
131 req->r_con_filling_msg = NULL;
134 ceph_msg_put(req->r_reply);
135 if (req->r_own_pages)
136 ceph_release_page_vector(req->r_pages,
138 ceph_put_snap_context(req->r_snapc);
139 ceph_pagelist_release(&req->r_trail);
141 mempool_free(req, req->r_osdc->req_mempool);
145 EXPORT_SYMBOL(ceph_osdc_release_request);
147 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
148 struct ceph_snap_context *snapc,
153 struct ceph_osd_request *req;
154 struct ceph_msg *msg;
155 size_t msg_size = sizeof(struct ceph_osd_request_head);
157 msg_size += num_op*sizeof(struct ceph_osd_op);
160 req = mempool_alloc(osdc->req_mempool, gfp_flags);
161 memset(req, 0, sizeof(*req));
163 req = kzalloc(sizeof(*req), gfp_flags);
169 req->r_mempool = use_mempool;
171 kref_init(&req->r_kref);
172 init_completion(&req->r_completion);
173 init_completion(&req->r_safe_completion);
174 RB_CLEAR_NODE(&req->r_node);
175 INIT_LIST_HEAD(&req->r_unsafe_item);
176 INIT_LIST_HEAD(&req->r_linger_item);
177 INIT_LIST_HEAD(&req->r_linger_osd);
178 INIT_LIST_HEAD(&req->r_req_lru_item);
179 INIT_LIST_HEAD(&req->r_osd_item);
181 /* create reply message */
183 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
185 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
186 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
188 ceph_osdc_put_request(req);
193 ceph_pagelist_init(&req->r_trail);
195 /* create request message; allow space for oid */
196 msg_size += MAX_OBJ_NAME_SIZE;
198 msg_size += sizeof(u64) * snapc->num_snaps;
200 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
202 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
204 ceph_osdc_put_request(req);
208 memset(msg->front.iov_base, 0, msg->front.iov_len);
210 req->r_request = msg;
214 EXPORT_SYMBOL(ceph_osdc_alloc_request);
216 static void osd_req_encode_op(struct ceph_osd_request *req,
217 struct ceph_osd_op *dst,
218 struct ceph_osd_req_op *src)
220 dst->op = cpu_to_le16(src->op);
223 case CEPH_OSD_OP_STAT:
225 case CEPH_OSD_OP_READ:
226 case CEPH_OSD_OP_WRITE:
228 cpu_to_le64(src->extent.offset);
230 cpu_to_le64(src->extent.length);
231 dst->extent.truncate_size =
232 cpu_to_le64(src->extent.truncate_size);
233 dst->extent.truncate_seq =
234 cpu_to_le32(src->extent.truncate_seq);
236 case CEPH_OSD_OP_CALL:
237 dst->cls.class_len = src->cls.class_len;
238 dst->cls.method_len = src->cls.method_len;
239 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
241 ceph_pagelist_append(&req->r_trail, src->cls.class_name,
243 ceph_pagelist_append(&req->r_trail, src->cls.method_name,
244 src->cls.method_len);
245 ceph_pagelist_append(&req->r_trail, src->cls.indata,
246 src->cls.indata_len);
248 case CEPH_OSD_OP_STARTSYNC:
250 case CEPH_OSD_OP_NOTIFY_ACK:
251 case CEPH_OSD_OP_WATCH:
252 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
253 dst->watch.ver = cpu_to_le64(src->watch.ver);
254 dst->watch.flag = src->watch.flag;
257 pr_err("unrecognized osd opcode %d\n", dst->op);
260 case CEPH_OSD_OP_MAPEXT:
261 case CEPH_OSD_OP_MASKTRUNC:
262 case CEPH_OSD_OP_SPARSE_READ:
263 case CEPH_OSD_OP_NOTIFY:
264 case CEPH_OSD_OP_ASSERT_VER:
265 case CEPH_OSD_OP_WRITEFULL:
266 case CEPH_OSD_OP_TRUNCATE:
267 case CEPH_OSD_OP_ZERO:
268 case CEPH_OSD_OP_DELETE:
269 case CEPH_OSD_OP_APPEND:
270 case CEPH_OSD_OP_SETTRUNC:
271 case CEPH_OSD_OP_TRIMTRUNC:
272 case CEPH_OSD_OP_TMAPUP:
273 case CEPH_OSD_OP_TMAPPUT:
274 case CEPH_OSD_OP_TMAPGET:
275 case CEPH_OSD_OP_CREATE:
276 case CEPH_OSD_OP_ROLLBACK:
277 case CEPH_OSD_OP_OMAPGETKEYS:
278 case CEPH_OSD_OP_OMAPGETVALS:
279 case CEPH_OSD_OP_OMAPGETHEADER:
280 case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
281 case CEPH_OSD_OP_MODE_RD:
282 case CEPH_OSD_OP_OMAPSETVALS:
283 case CEPH_OSD_OP_OMAPSETHEADER:
284 case CEPH_OSD_OP_OMAPCLEAR:
285 case CEPH_OSD_OP_OMAPRMKEYS:
286 case CEPH_OSD_OP_OMAP_CMP:
287 case CEPH_OSD_OP_CLONERANGE:
288 case CEPH_OSD_OP_ASSERT_SRC_VERSION:
289 case CEPH_OSD_OP_SRC_CMPXATTR:
290 case CEPH_OSD_OP_GETXATTR:
291 case CEPH_OSD_OP_GETXATTRS:
292 case CEPH_OSD_OP_CMPXATTR:
293 case CEPH_OSD_OP_SETXATTR:
294 case CEPH_OSD_OP_SETXATTRS:
295 case CEPH_OSD_OP_RESETXATTRS:
296 case CEPH_OSD_OP_RMXATTR:
297 case CEPH_OSD_OP_PULL:
298 case CEPH_OSD_OP_PUSH:
299 case CEPH_OSD_OP_BALANCEREADS:
300 case CEPH_OSD_OP_UNBALANCEREADS:
301 case CEPH_OSD_OP_SCRUB:
302 case CEPH_OSD_OP_SCRUB_RESERVE:
303 case CEPH_OSD_OP_SCRUB_UNRESERVE:
304 case CEPH_OSD_OP_SCRUB_STOP:
305 case CEPH_OSD_OP_SCRUB_MAP:
306 case CEPH_OSD_OP_WRLOCK:
307 case CEPH_OSD_OP_WRUNLOCK:
308 case CEPH_OSD_OP_RDLOCK:
309 case CEPH_OSD_OP_RDUNLOCK:
310 case CEPH_OSD_OP_UPLOCK:
311 case CEPH_OSD_OP_DNLOCK:
312 case CEPH_OSD_OP_PGLS:
313 case CEPH_OSD_OP_PGLS_FILTER:
314 pr_err("unsupported osd opcode %s\n",
315 ceph_osd_op_name(dst->op));
319 dst->payload_len = cpu_to_le32(src->payload_len);
323 * build new request AND message
326 void ceph_osdc_build_request(struct ceph_osd_request *req,
327 u64 off, u64 len, unsigned int num_op,
328 struct ceph_osd_req_op *src_ops,
329 struct ceph_snap_context *snapc, u64 snap_id,
330 struct timespec *mtime)
332 struct ceph_msg *msg = req->r_request;
333 struct ceph_osd_request_head *head;
334 struct ceph_osd_req_op *src_op;
335 struct ceph_osd_op *op;
337 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
338 int flags = req->r_flags;
342 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
344 head = msg->front.iov_base;
345 head->snapid = cpu_to_le64(snap_id);
346 op = (void *)(head + 1);
347 p = (void *)(op + num_op);
349 req->r_snapc = ceph_get_snap_context(snapc);
351 head->client_inc = cpu_to_le32(1); /* always, for now. */
352 head->flags = cpu_to_le32(flags);
353 if (flags & CEPH_OSD_FLAG_WRITE)
354 ceph_encode_timespec(&head->mtime, mtime);
355 BUG_ON(num_op > (unsigned int) ((u16) -1));
356 head->num_ops = cpu_to_le16(num_op);
359 head->object_len = cpu_to_le32(req->r_oid_len);
360 memcpy(p, req->r_oid, req->r_oid_len);
365 osd_req_encode_op(req, op++, src_op++);
368 head->snap_seq = cpu_to_le64(snapc->seq);
369 head->num_snaps = cpu_to_le32(snapc->num_snaps);
370 for (i = 0; i < snapc->num_snaps; i++) {
371 put_unaligned_le64(snapc->snaps[i], p);
376 data_len = req->r_trail.length;
377 if (flags & CEPH_OSD_FLAG_WRITE) {
378 req->r_request->hdr.data_off = cpu_to_le16(off);
381 req->r_request->hdr.data_len = cpu_to_le32(data_len);
382 req->r_request->page_alignment = req->r_page_alignment;
384 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
385 msg_size = p - msg->front.iov_base;
386 msg->front.iov_len = msg_size;
387 msg->hdr.front_len = cpu_to_le32(msg_size);
390 EXPORT_SYMBOL(ceph_osdc_build_request);
393 * build new request AND message, calculate layout, and adjust file
396 * if the file was recently truncated, we include information about its
397 * old and new size so that the object can be updated appropriately. (we
398 * avoid synchronously deleting truncated objects because it's slow.)
400 * if @do_sync, include a 'startsync' command so that the osd will flush
403 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
404 struct ceph_file_layout *layout,
405 struct ceph_vino vino,
407 int opcode, int flags,
408 struct ceph_snap_context *snapc,
412 struct timespec *mtime,
416 struct ceph_osd_req_op ops[2];
417 struct ceph_osd_request *req;
418 unsigned int num_op = 1;
421 memset(&ops, 0, sizeof ops);
424 ops[0].extent.truncate_seq = truncate_seq;
425 ops[0].extent.truncate_size = truncate_size;
428 ops[1].op = CEPH_OSD_OP_STARTSYNC;
432 req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
435 return ERR_PTR(-ENOMEM);
436 req->r_flags = flags;
438 /* calculate max write size */
439 r = calc_layout(vino, layout, off, plen, req, ops);
442 req->r_file_layout = *layout; /* keep a copy */
444 /* in case it differs from natural (file) alignment that
445 calc_layout filled in for us */
446 req->r_num_pages = calc_pages_for(page_align, *plen);
447 req->r_page_alignment = page_align;
449 ceph_osdc_build_request(req, off, *plen, num_op, ops,
450 snapc, vino.snap, mtime);
454 EXPORT_SYMBOL(ceph_osdc_new_request);
457 * We keep osd requests in an rbtree, sorted by ->r_tid.
459 static void __insert_request(struct ceph_osd_client *osdc,
460 struct ceph_osd_request *new)
462 struct rb_node **p = &osdc->requests.rb_node;
463 struct rb_node *parent = NULL;
464 struct ceph_osd_request *req = NULL;
468 req = rb_entry(parent, struct ceph_osd_request, r_node);
469 if (new->r_tid < req->r_tid)
471 else if (new->r_tid > req->r_tid)
477 rb_link_node(&new->r_node, parent, p);
478 rb_insert_color(&new->r_node, &osdc->requests);
481 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
484 struct ceph_osd_request *req;
485 struct rb_node *n = osdc->requests.rb_node;
488 req = rb_entry(n, struct ceph_osd_request, r_node);
489 if (tid < req->r_tid)
491 else if (tid > req->r_tid)
499 static struct ceph_osd_request *
500 __lookup_request_ge(struct ceph_osd_client *osdc,
503 struct ceph_osd_request *req;
504 struct rb_node *n = osdc->requests.rb_node;
507 req = rb_entry(n, struct ceph_osd_request, r_node);
508 if (tid < req->r_tid) {
512 } else if (tid > req->r_tid) {
522 * Resubmit requests pending on the given osd.
524 static void __kick_osd_requests(struct ceph_osd_client *osdc,
525 struct ceph_osd *osd)
527 struct ceph_osd_request *req, *nreq;
530 dout("__kick_osd_requests osd%d\n", osd->o_osd);
531 err = __reset_osd(osdc, osd);
535 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
536 list_move(&req->r_req_lru_item, &osdc->req_unsent);
537 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
540 req->r_flags |= CEPH_OSD_FLAG_RETRY;
543 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
546 * reregister request prior to unregistering linger so
547 * that r_osd is preserved.
549 BUG_ON(!list_empty(&req->r_req_lru_item));
550 __register_request(osdc, req);
551 list_add(&req->r_req_lru_item, &osdc->req_unsent);
552 list_add(&req->r_osd_item, &req->r_osd->o_requests);
553 __unregister_linger_request(osdc, req);
554 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
560 * If the osd connection drops, we need to resubmit all requests.
562 static void osd_reset(struct ceph_connection *con)
564 struct ceph_osd *osd = con->private;
565 struct ceph_osd_client *osdc;
569 dout("osd_reset osd%d\n", osd->o_osd);
571 down_read(&osdc->map_sem);
572 mutex_lock(&osdc->request_mutex);
573 __kick_osd_requests(osdc, osd);
575 mutex_unlock(&osdc->request_mutex);
576 up_read(&osdc->map_sem);
580 * Track open sessions with osds.
582 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
584 struct ceph_osd *osd;
586 osd = kzalloc(sizeof(*osd), GFP_NOFS);
590 atomic_set(&osd->o_ref, 1);
593 RB_CLEAR_NODE(&osd->o_node);
594 INIT_LIST_HEAD(&osd->o_requests);
595 INIT_LIST_HEAD(&osd->o_linger_requests);
596 INIT_LIST_HEAD(&osd->o_osd_lru);
597 osd->o_incarnation = 1;
599 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
601 INIT_LIST_HEAD(&osd->o_keepalive_item);
605 static struct ceph_osd *get_osd(struct ceph_osd *osd)
607 if (atomic_inc_not_zero(&osd->o_ref)) {
608 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
609 atomic_read(&osd->o_ref));
612 dout("get_osd %p FAIL\n", osd);
617 static void put_osd(struct ceph_osd *osd)
619 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
620 atomic_read(&osd->o_ref) - 1);
621 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
622 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
624 if (ac->ops && ac->ops->destroy_authorizer)
625 ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
631 * remove an osd from our map
633 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
635 dout("__remove_osd %p\n", osd);
636 BUG_ON(!list_empty(&osd->o_requests));
637 rb_erase(&osd->o_node, &osdc->osds);
638 list_del_init(&osd->o_osd_lru);
639 ceph_con_close(&osd->o_con);
643 static void remove_all_osds(struct ceph_osd_client *osdc)
645 dout("%s %p\n", __func__, osdc);
646 mutex_lock(&osdc->request_mutex);
647 while (!RB_EMPTY_ROOT(&osdc->osds)) {
648 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
649 struct ceph_osd, o_node);
650 __remove_osd(osdc, osd);
652 mutex_unlock(&osdc->request_mutex);
655 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
656 struct ceph_osd *osd)
658 dout("__move_osd_to_lru %p\n", osd);
659 BUG_ON(!list_empty(&osd->o_osd_lru));
660 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
661 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
664 static void __remove_osd_from_lru(struct ceph_osd *osd)
666 dout("__remove_osd_from_lru %p\n", osd);
667 if (!list_empty(&osd->o_osd_lru))
668 list_del_init(&osd->o_osd_lru);
671 static void remove_old_osds(struct ceph_osd_client *osdc)
673 struct ceph_osd *osd, *nosd;
675 dout("__remove_old_osds %p\n", osdc);
676 mutex_lock(&osdc->request_mutex);
677 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
678 if (time_before(jiffies, osd->lru_ttl))
680 __remove_osd(osdc, osd);
682 mutex_unlock(&osdc->request_mutex);
688 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
690 struct ceph_entity_addr *peer_addr;
692 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
693 if (list_empty(&osd->o_requests) &&
694 list_empty(&osd->o_linger_requests)) {
695 __remove_osd(osdc, osd);
700 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
701 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
702 !ceph_con_opened(&osd->o_con)) {
703 struct ceph_osd_request *req;
705 dout(" osd addr hasn't changed and connection never opened,"
706 " letting msgr retry");
707 /* touch each r_stamp for handle_timeout()'s benfit */
708 list_for_each_entry(req, &osd->o_requests, r_osd_item)
709 req->r_stamp = jiffies;
714 ceph_con_close(&osd->o_con);
715 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
716 osd->o_incarnation++;
721 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
723 struct rb_node **p = &osdc->osds.rb_node;
724 struct rb_node *parent = NULL;
725 struct ceph_osd *osd = NULL;
727 dout("__insert_osd %p osd%d\n", new, new->o_osd);
730 osd = rb_entry(parent, struct ceph_osd, o_node);
731 if (new->o_osd < osd->o_osd)
733 else if (new->o_osd > osd->o_osd)
739 rb_link_node(&new->o_node, parent, p);
740 rb_insert_color(&new->o_node, &osdc->osds);
743 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
745 struct ceph_osd *osd;
746 struct rb_node *n = osdc->osds.rb_node;
749 osd = rb_entry(n, struct ceph_osd, o_node);
752 else if (o > osd->o_osd)
760 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
762 schedule_delayed_work(&osdc->timeout_work,
763 osdc->client->options->osd_keepalive_timeout * HZ);
766 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
768 cancel_delayed_work(&osdc->timeout_work);
772 * Register request, assign tid. If this is the first request, set up
775 static void __register_request(struct ceph_osd_client *osdc,
776 struct ceph_osd_request *req)
778 req->r_tid = ++osdc->last_tid;
779 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
780 dout("__register_request %p tid %lld\n", req, req->r_tid);
781 __insert_request(osdc, req);
782 ceph_osdc_get_request(req);
783 osdc->num_requests++;
784 if (osdc->num_requests == 1) {
785 dout(" first request, scheduling timeout\n");
786 __schedule_osd_timeout(osdc);
790 static void register_request(struct ceph_osd_client *osdc,
791 struct ceph_osd_request *req)
793 mutex_lock(&osdc->request_mutex);
794 __register_request(osdc, req);
795 mutex_unlock(&osdc->request_mutex);
799 * called under osdc->request_mutex
801 static void __unregister_request(struct ceph_osd_client *osdc,
802 struct ceph_osd_request *req)
804 if (RB_EMPTY_NODE(&req->r_node)) {
805 dout("__unregister_request %p tid %lld not registered\n",
810 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
811 rb_erase(&req->r_node, &osdc->requests);
812 osdc->num_requests--;
815 /* make sure the original request isn't in flight. */
816 ceph_msg_revoke(req->r_request);
818 list_del_init(&req->r_osd_item);
819 if (list_empty(&req->r_osd->o_requests) &&
820 list_empty(&req->r_osd->o_linger_requests)) {
821 dout("moving osd to %p lru\n", req->r_osd);
822 __move_osd_to_lru(osdc, req->r_osd);
824 if (list_empty(&req->r_linger_item))
828 list_del_init(&req->r_req_lru_item);
829 ceph_osdc_put_request(req);
831 if (osdc->num_requests == 0) {
832 dout(" no requests, canceling timeout\n");
833 __cancel_osd_timeout(osdc);
838 * Cancel a previously queued request message
840 static void __cancel_request(struct ceph_osd_request *req)
842 if (req->r_sent && req->r_osd) {
843 ceph_msg_revoke(req->r_request);
848 static void __register_linger_request(struct ceph_osd_client *osdc,
849 struct ceph_osd_request *req)
851 dout("__register_linger_request %p\n", req);
852 list_add_tail(&req->r_linger_item, &osdc->req_linger);
854 list_add_tail(&req->r_linger_osd,
855 &req->r_osd->o_linger_requests);
858 static void __unregister_linger_request(struct ceph_osd_client *osdc,
859 struct ceph_osd_request *req)
861 dout("__unregister_linger_request %p\n", req);
862 list_del_init(&req->r_linger_item);
864 list_del_init(&req->r_linger_osd);
866 if (list_empty(&req->r_osd->o_requests) &&
867 list_empty(&req->r_osd->o_linger_requests)) {
868 dout("moving osd to %p lru\n", req->r_osd);
869 __move_osd_to_lru(osdc, req->r_osd);
871 if (list_empty(&req->r_osd_item))
876 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
877 struct ceph_osd_request *req)
879 mutex_lock(&osdc->request_mutex);
881 __unregister_linger_request(osdc, req);
882 ceph_osdc_put_request(req);
884 mutex_unlock(&osdc->request_mutex);
886 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
888 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
889 struct ceph_osd_request *req)
891 if (!req->r_linger) {
892 dout("set_request_linger %p\n", req);
895 * caller is now responsible for calling
896 * unregister_linger_request
898 ceph_osdc_get_request(req);
901 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
904 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
905 * (as needed), and set the request r_osd appropriately. If there is
906 * no up osd, set r_osd to NULL. Move the request to the appropriate list
907 * (unsent, homeless) or leave on in-flight lru.
909 * Return 0 if unchanged, 1 if changed, or negative on error.
911 * Caller should hold map_sem for read and request_mutex.
913 static int __map_request(struct ceph_osd_client *osdc,
914 struct ceph_osd_request *req, int force_resend)
916 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
917 struct ceph_pg_v1 pgid;
918 int acting[CEPH_PG_MAX_SIZE];
922 dout("map_request %p tid %lld\n", req, req->r_tid);
923 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
924 &req->r_file_layout, osdc->osdmap);
926 list_move(&req->r_req_lru_item, &osdc->req_notarget);
929 pgid = reqhead->layout.ol_pgid;
932 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
938 if ((!force_resend &&
939 req->r_osd && req->r_osd->o_osd == o &&
940 req->r_sent >= req->r_osd->o_incarnation &&
941 req->r_num_pg_osds == num &&
942 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
943 (req->r_osd == NULL && o == -1))
944 return 0; /* no change */
946 dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
947 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
948 req->r_osd ? req->r_osd->o_osd : -1);
950 /* record full pg acting set */
951 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
952 req->r_num_pg_osds = num;
955 __cancel_request(req);
956 list_del_init(&req->r_osd_item);
960 req->r_osd = __lookup_osd(osdc, o);
961 if (!req->r_osd && o >= 0) {
963 req->r_osd = create_osd(osdc, o);
965 list_move(&req->r_req_lru_item, &osdc->req_notarget);
969 dout("map_request osd %p is osd%d\n", req->r_osd, o);
970 __insert_osd(osdc, req->r_osd);
972 ceph_con_open(&req->r_osd->o_con,
973 CEPH_ENTITY_TYPE_OSD, o,
974 &osdc->osdmap->osd_addr[o]);
978 __remove_osd_from_lru(req->r_osd);
979 list_add(&req->r_osd_item, &req->r_osd->o_requests);
980 list_move(&req->r_req_lru_item, &osdc->req_unsent);
982 list_move(&req->r_req_lru_item, &osdc->req_notarget);
984 err = 1; /* osd or pg changed */
991 * caller should hold map_sem (for read) and request_mutex
993 static void __send_request(struct ceph_osd_client *osdc,
994 struct ceph_osd_request *req)
996 struct ceph_osd_request_head *reqhead;
998 dout("send_request %p tid %llu to osd%d flags %d\n",
999 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1001 reqhead = req->r_request->front.iov_base;
1002 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1003 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
1004 reqhead->reassert_version = req->r_reassert_version;
1006 req->r_stamp = jiffies;
1007 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1009 ceph_msg_get(req->r_request); /* send consumes a ref */
1010 ceph_con_send(&req->r_osd->o_con, req->r_request);
1011 req->r_sent = req->r_osd->o_incarnation;
1015 * Send any requests in the queue (req_unsent).
1017 static void __send_queued(struct ceph_osd_client *osdc)
1019 struct ceph_osd_request *req, *tmp;
1021 dout("__send_queued\n");
1022 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1023 __send_request(osdc, req);
1027 * Timeout callback, called every N seconds when 1 or more osd
1028 * requests has been active for more than N seconds. When this
1029 * happens, we ping all OSDs with requests who have timed out to
1030 * ensure any communications channel reset is detected. Reset the
1031 * request timeouts another N seconds in the future as we go.
1032 * Reschedule the timeout event another N seconds in future (unless
1033 * there are no open requests).
1035 static void handle_timeout(struct work_struct *work)
1037 struct ceph_osd_client *osdc =
1038 container_of(work, struct ceph_osd_client, timeout_work.work);
1039 struct ceph_osd_request *req;
1040 struct ceph_osd *osd;
1041 unsigned long keepalive =
1042 osdc->client->options->osd_keepalive_timeout * HZ;
1043 struct list_head slow_osds;
1045 down_read(&osdc->map_sem);
1047 ceph_monc_request_next_osdmap(&osdc->client->monc);
1049 mutex_lock(&osdc->request_mutex);
1052 * ping osds that are a bit slow. this ensures that if there
1053 * is a break in the TCP connection we will notice, and reopen
1054 * a connection with that osd (from the fault callback).
1056 INIT_LIST_HEAD(&slow_osds);
1057 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1058 if (time_before(jiffies, req->r_stamp + keepalive))
1063 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1064 req->r_tid, osd->o_osd);
1065 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1067 while (!list_empty(&slow_osds)) {
1068 osd = list_entry(slow_osds.next, struct ceph_osd,
1070 list_del_init(&osd->o_keepalive_item);
1071 ceph_con_keepalive(&osd->o_con);
1074 __schedule_osd_timeout(osdc);
1075 __send_queued(osdc);
1076 mutex_unlock(&osdc->request_mutex);
1077 up_read(&osdc->map_sem);
1080 static void handle_osds_timeout(struct work_struct *work)
1082 struct ceph_osd_client *osdc =
1083 container_of(work, struct ceph_osd_client,
1084 osds_timeout_work.work);
1085 unsigned long delay =
1086 osdc->client->options->osd_idle_ttl * HZ >> 2;
1088 dout("osds timeout\n");
1089 down_read(&osdc->map_sem);
1090 remove_old_osds(osdc);
1091 up_read(&osdc->map_sem);
1093 schedule_delayed_work(&osdc->osds_timeout_work,
1094 round_jiffies_relative(delay));
1097 static void complete_request(struct ceph_osd_request *req)
1099 if (req->r_safe_callback)
1100 req->r_safe_callback(req, NULL);
1101 complete_all(&req->r_safe_completion); /* fsync waiter */
1105 * handle osd op reply. either call the callback if it is specified,
1106 * or do the completion to wake up the waiting thread.
1108 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1109 struct ceph_connection *con)
1111 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1112 struct ceph_osd_request *req;
1114 int numops, object_len, flags;
1117 tid = le64_to_cpu(msg->hdr.tid);
1118 if (msg->front.iov_len < sizeof(*rhead))
1120 numops = le32_to_cpu(rhead->num_ops);
1121 object_len = le32_to_cpu(rhead->object_len);
1122 result = le32_to_cpu(rhead->result);
1123 if (msg->front.iov_len != sizeof(*rhead) + object_len +
1124 numops * sizeof(struct ceph_osd_op))
1126 dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1128 mutex_lock(&osdc->request_mutex);
1129 req = __lookup_request(osdc, tid);
1131 dout("handle_reply tid %llu dne\n", tid);
1132 mutex_unlock(&osdc->request_mutex);
1135 ceph_osdc_get_request(req);
1136 flags = le32_to_cpu(rhead->flags);
1139 * if this connection filled our message, drop our reference now, to
1140 * avoid a (safe but slower) revoke later.
1142 if (req->r_con_filling_msg == con && req->r_reply == msg) {
1143 dout(" dropping con_filling_msg ref %p\n", con);
1144 req->r_con_filling_msg = NULL;
1148 if (!req->r_got_reply) {
1151 req->r_result = le32_to_cpu(rhead->result);
1152 bytes = le32_to_cpu(msg->hdr.data_len);
1153 dout("handle_reply result %d bytes %d\n", req->r_result,
1155 if (req->r_result == 0)
1156 req->r_result = bytes;
1158 /* in case this is a write and we need to replay, */
1159 req->r_reassert_version = rhead->reassert_version;
1161 req->r_got_reply = 1;
1162 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1163 dout("handle_reply tid %llu dup ack\n", tid);
1164 mutex_unlock(&osdc->request_mutex);
1168 dout("handle_reply tid %llu flags %d\n", tid, flags);
1170 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1171 __register_linger_request(osdc, req);
1173 /* either this is a read, or we got the safe response */
1175 (flags & CEPH_OSD_FLAG_ONDISK) ||
1176 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1177 __unregister_request(osdc, req);
1179 mutex_unlock(&osdc->request_mutex);
1181 if (req->r_callback)
1182 req->r_callback(req, msg);
1184 complete_all(&req->r_completion);
1186 if (flags & CEPH_OSD_FLAG_ONDISK)
1187 complete_request(req);
1190 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1191 ceph_osdc_put_request(req);
1195 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1196 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1197 (int)sizeof(*rhead));
1201 static void reset_changed_osds(struct ceph_osd_client *osdc)
1203 struct rb_node *p, *n;
1205 for (p = rb_first(&osdc->osds); p; p = n) {
1206 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1209 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1210 memcmp(&osd->o_con.peer_addr,
1211 ceph_osd_addr(osdc->osdmap,
1213 sizeof(struct ceph_entity_addr)) != 0)
1214 __reset_osd(osdc, osd);
1219 * Requeue requests whose mapping to an OSD has changed. If requests map to
1220 * no osd, request a new map.
1222 * Caller should hold map_sem for read.
1224 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1226 struct ceph_osd_request *req, *nreq;
1231 dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1232 mutex_lock(&osdc->request_mutex);
1233 for (p = rb_first(&osdc->requests); p; ) {
1234 req = rb_entry(p, struct ceph_osd_request, r_node);
1238 * For linger requests that have not yet been
1239 * registered, move them to the linger list; they'll
1240 * be sent to the osd in the loop below. Unregister
1241 * the request before re-registering it as a linger
1242 * request to ensure the __map_request() below
1243 * will decide it needs to be sent.
1245 if (req->r_linger && list_empty(&req->r_linger_item)) {
1246 dout("%p tid %llu restart on osd%d\n",
1248 req->r_osd ? req->r_osd->o_osd : -1);
1249 __unregister_request(osdc, req);
1250 __register_linger_request(osdc, req);
1254 err = __map_request(osdc, req, force_resend);
1256 continue; /* error */
1257 if (req->r_osd == NULL) {
1258 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1259 needmap++; /* request a newer map */
1260 } else if (err > 0) {
1261 if (!req->r_linger) {
1262 dout("%p tid %llu requeued on osd%d\n", req,
1264 req->r_osd ? req->r_osd->o_osd : -1);
1265 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1270 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1272 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1274 err = __map_request(osdc, req, force_resend);
1275 dout("__map_request returned %d\n", err);
1277 continue; /* no change and no osd was specified */
1279 continue; /* hrm! */
1280 if (req->r_osd == NULL) {
1281 dout("tid %llu maps to no valid osd\n", req->r_tid);
1282 needmap++; /* request a newer map */
1286 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1287 req->r_osd ? req->r_osd->o_osd : -1);
1288 __register_request(osdc, req);
1289 __unregister_linger_request(osdc, req);
1291 mutex_unlock(&osdc->request_mutex);
1294 dout("%d requests for down osds, need new map\n", needmap);
1295 ceph_monc_request_next_osdmap(&osdc->client->monc);
1297 reset_changed_osds(osdc);
1302 * Process updated osd map.
1304 * The message contains any number of incremental and full maps, normally
1305 * indicating some sort of topology change in the cluster. Kick requests
1306 * off to different OSDs as needed.
1308 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1310 void *p, *end, *next;
1311 u32 nr_maps, maplen;
1313 struct ceph_osdmap *newmap = NULL, *oldmap;
1315 struct ceph_fsid fsid;
1317 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1318 p = msg->front.iov_base;
1319 end = p + msg->front.iov_len;
1322 ceph_decode_need(&p, end, sizeof(fsid), bad);
1323 ceph_decode_copy(&p, &fsid, sizeof(fsid));
1324 if (ceph_check_fsid(osdc->client, &fsid) < 0)
1327 down_write(&osdc->map_sem);
1329 /* incremental maps */
1330 ceph_decode_32_safe(&p, end, nr_maps, bad);
1331 dout(" %d inc maps\n", nr_maps);
1332 while (nr_maps > 0) {
1333 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1334 epoch = ceph_decode_32(&p);
1335 maplen = ceph_decode_32(&p);
1336 ceph_decode_need(&p, end, maplen, bad);
1338 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1339 dout("applying incremental map %u len %d\n",
1341 newmap = osdmap_apply_incremental(&p, next,
1343 &osdc->client->msgr);
1344 if (IS_ERR(newmap)) {
1345 err = PTR_ERR(newmap);
1349 if (newmap != osdc->osdmap) {
1350 ceph_osdmap_destroy(osdc->osdmap);
1351 osdc->osdmap = newmap;
1353 kick_requests(osdc, 0);
1355 dout("ignoring incremental map %u len %d\n",
1365 ceph_decode_32_safe(&p, end, nr_maps, bad);
1366 dout(" %d full maps\n", nr_maps);
1368 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1369 epoch = ceph_decode_32(&p);
1370 maplen = ceph_decode_32(&p);
1371 ceph_decode_need(&p, end, maplen, bad);
1373 dout("skipping non-latest full map %u len %d\n",
1375 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1376 dout("skipping full map %u len %d, "
1377 "older than our %u\n", epoch, maplen,
1378 osdc->osdmap->epoch);
1380 int skipped_map = 0;
1382 dout("taking full map %u len %d\n", epoch, maplen);
1383 newmap = osdmap_decode(&p, p+maplen);
1384 if (IS_ERR(newmap)) {
1385 err = PTR_ERR(newmap);
1389 oldmap = osdc->osdmap;
1390 osdc->osdmap = newmap;
1392 if (oldmap->epoch + 1 < newmap->epoch)
1394 ceph_osdmap_destroy(oldmap);
1396 kick_requests(osdc, skipped_map);
1403 downgrade_write(&osdc->map_sem);
1404 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1407 * subscribe to subsequent osdmap updates if full to ensure
1408 * we find out when we are no longer full and stop returning
1411 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1412 ceph_monc_request_next_osdmap(&osdc->client->monc);
1414 mutex_lock(&osdc->request_mutex);
1415 __send_queued(osdc);
1416 mutex_unlock(&osdc->request_mutex);
1417 up_read(&osdc->map_sem);
1418 wake_up_all(&osdc->client->auth_wq);
1422 pr_err("osdc handle_map corrupt msg\n");
1424 up_write(&osdc->map_sem);
1429 * watch/notify callback event infrastructure
1431 * These callbacks are used both for watch and notify operations.
1433 static void __release_event(struct kref *kref)
1435 struct ceph_osd_event *event =
1436 container_of(kref, struct ceph_osd_event, kref);
1438 dout("__release_event %p\n", event);
1442 static void get_event(struct ceph_osd_event *event)
1444 kref_get(&event->kref);
1447 void ceph_osdc_put_event(struct ceph_osd_event *event)
1449 kref_put(&event->kref, __release_event);
1451 EXPORT_SYMBOL(ceph_osdc_put_event);
1453 static void __insert_event(struct ceph_osd_client *osdc,
1454 struct ceph_osd_event *new)
1456 struct rb_node **p = &osdc->event_tree.rb_node;
1457 struct rb_node *parent = NULL;
1458 struct ceph_osd_event *event = NULL;
1462 event = rb_entry(parent, struct ceph_osd_event, node);
1463 if (new->cookie < event->cookie)
1465 else if (new->cookie > event->cookie)
1466 p = &(*p)->rb_right;
1471 rb_link_node(&new->node, parent, p);
1472 rb_insert_color(&new->node, &osdc->event_tree);
1475 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1478 struct rb_node **p = &osdc->event_tree.rb_node;
1479 struct rb_node *parent = NULL;
1480 struct ceph_osd_event *event = NULL;
1484 event = rb_entry(parent, struct ceph_osd_event, node);
1485 if (cookie < event->cookie)
1487 else if (cookie > event->cookie)
1488 p = &(*p)->rb_right;
1495 static void __remove_event(struct ceph_osd_event *event)
1497 struct ceph_osd_client *osdc = event->osdc;
1499 if (!RB_EMPTY_NODE(&event->node)) {
1500 dout("__remove_event removed %p\n", event);
1501 rb_erase(&event->node, &osdc->event_tree);
1502 ceph_osdc_put_event(event);
1504 dout("__remove_event didn't remove %p\n", event);
1508 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1509 void (*event_cb)(u64, u64, u8, void *),
1510 void *data, struct ceph_osd_event **pevent)
1512 struct ceph_osd_event *event;
1514 event = kmalloc(sizeof(*event), GFP_NOIO);
1518 dout("create_event %p\n", event);
1519 event->cb = event_cb;
1520 event->one_shot = 0;
1523 INIT_LIST_HEAD(&event->osd_node);
1524 RB_CLEAR_NODE(&event->node);
1525 kref_init(&event->kref); /* one ref for us */
1526 kref_get(&event->kref); /* one ref for the caller */
1528 spin_lock(&osdc->event_lock);
1529 event->cookie = ++osdc->event_count;
1530 __insert_event(osdc, event);
1531 spin_unlock(&osdc->event_lock);
1536 EXPORT_SYMBOL(ceph_osdc_create_event);
1538 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1540 struct ceph_osd_client *osdc = event->osdc;
1542 dout("cancel_event %p\n", event);
1543 spin_lock(&osdc->event_lock);
1544 __remove_event(event);
1545 spin_unlock(&osdc->event_lock);
1546 ceph_osdc_put_event(event); /* caller's */
1548 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1551 static void do_event_work(struct work_struct *work)
1553 struct ceph_osd_event_work *event_work =
1554 container_of(work, struct ceph_osd_event_work, work);
1555 struct ceph_osd_event *event = event_work->event;
1556 u64 ver = event_work->ver;
1557 u64 notify_id = event_work->notify_id;
1558 u8 opcode = event_work->opcode;
1560 dout("do_event_work completing %p\n", event);
1561 event->cb(ver, notify_id, opcode, event->data);
1562 dout("do_event_work completed %p\n", event);
1563 ceph_osdc_put_event(event);
1569 * Process osd watch notifications
1571 static void handle_watch_notify(struct ceph_osd_client *osdc,
1572 struct ceph_msg *msg)
1576 u64 cookie, ver, notify_id;
1578 struct ceph_osd_event *event;
1579 struct ceph_osd_event_work *event_work;
1581 p = msg->front.iov_base;
1582 end = p + msg->front.iov_len;
1584 ceph_decode_8_safe(&p, end, proto_ver, bad);
1585 ceph_decode_8_safe(&p, end, opcode, bad);
1586 ceph_decode_64_safe(&p, end, cookie, bad);
1587 ceph_decode_64_safe(&p, end, ver, bad);
1588 ceph_decode_64_safe(&p, end, notify_id, bad);
1590 spin_lock(&osdc->event_lock);
1591 event = __find_event(osdc, cookie);
1593 BUG_ON(event->one_shot);
1596 spin_unlock(&osdc->event_lock);
1597 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1598 cookie, ver, event);
1600 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1602 dout("ERROR: could not allocate event_work\n");
1605 INIT_WORK(&event_work->work, do_event_work);
1606 event_work->event = event;
1607 event_work->ver = ver;
1608 event_work->notify_id = notify_id;
1609 event_work->opcode = opcode;
1610 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1611 dout("WARNING: failed to queue notify event work\n");
1619 ceph_osdc_put_event(event);
1623 pr_err("osdc handle_watch_notify corrupt msg\n");
1628 * Register request, send initial attempt.
1630 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1631 struct ceph_osd_request *req,
1636 req->r_request->pages = req->r_pages;
1637 req->r_request->nr_pages = req->r_num_pages;
1639 req->r_request->bio = req->r_bio;
1641 req->r_request->trail = &req->r_trail;
1643 register_request(osdc, req);
1645 down_read(&osdc->map_sem);
1646 mutex_lock(&osdc->request_mutex);
1648 * a racing kick_requests() may have sent the message for us
1649 * while we dropped request_mutex above, so only send now if
1650 * the request still han't been touched yet.
1652 if (req->r_sent == 0) {
1653 rc = __map_request(osdc, req, 0);
1656 dout("osdc_start_request failed map, "
1657 " will retry %lld\n", req->r_tid);
1662 if (req->r_osd == NULL) {
1663 dout("send_request %p no up osds in pg\n", req);
1664 ceph_monc_request_next_osdmap(&osdc->client->monc);
1666 __send_request(osdc, req);
1672 mutex_unlock(&osdc->request_mutex);
1673 up_read(&osdc->map_sem);
1676 EXPORT_SYMBOL(ceph_osdc_start_request);
1679 * wait for a request to complete
1681 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1682 struct ceph_osd_request *req)
1686 rc = wait_for_completion_interruptible(&req->r_completion);
1688 mutex_lock(&osdc->request_mutex);
1689 __cancel_request(req);
1690 __unregister_request(osdc, req);
1691 mutex_unlock(&osdc->request_mutex);
1692 complete_request(req);
1693 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1697 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1698 return req->r_result;
1700 EXPORT_SYMBOL(ceph_osdc_wait_request);
1703 * sync - wait for all in-flight requests to flush. avoid starvation.
1705 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1707 struct ceph_osd_request *req;
1708 u64 last_tid, next_tid = 0;
1710 mutex_lock(&osdc->request_mutex);
1711 last_tid = osdc->last_tid;
1713 req = __lookup_request_ge(osdc, next_tid);
1716 if (req->r_tid > last_tid)
1719 next_tid = req->r_tid + 1;
1720 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1723 ceph_osdc_get_request(req);
1724 mutex_unlock(&osdc->request_mutex);
1725 dout("sync waiting on tid %llu (last is %llu)\n",
1726 req->r_tid, last_tid);
1727 wait_for_completion(&req->r_safe_completion);
1728 mutex_lock(&osdc->request_mutex);
1729 ceph_osdc_put_request(req);
1731 mutex_unlock(&osdc->request_mutex);
1732 dout("sync done (thru tid %llu)\n", last_tid);
1734 EXPORT_SYMBOL(ceph_osdc_sync);
1739 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1744 osdc->client = client;
1745 osdc->osdmap = NULL;
1746 init_rwsem(&osdc->map_sem);
1747 init_completion(&osdc->map_waiters);
1748 osdc->last_requested_map = 0;
1749 mutex_init(&osdc->request_mutex);
1751 osdc->osds = RB_ROOT;
1752 INIT_LIST_HEAD(&osdc->osd_lru);
1753 osdc->requests = RB_ROOT;
1754 INIT_LIST_HEAD(&osdc->req_lru);
1755 INIT_LIST_HEAD(&osdc->req_unsent);
1756 INIT_LIST_HEAD(&osdc->req_notarget);
1757 INIT_LIST_HEAD(&osdc->req_linger);
1758 osdc->num_requests = 0;
1759 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1760 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1761 spin_lock_init(&osdc->event_lock);
1762 osdc->event_tree = RB_ROOT;
1763 osdc->event_count = 0;
1765 schedule_delayed_work(&osdc->osds_timeout_work,
1766 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1769 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1770 sizeof(struct ceph_osd_request));
1771 if (!osdc->req_mempool)
1774 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1775 OSD_OP_FRONT_LEN, 10, true,
1779 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1780 OSD_OPREPLY_FRONT_LEN, 10, true,
1785 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1786 if (IS_ERR(osdc->notify_wq)) {
1787 err = PTR_ERR(osdc->notify_wq);
1788 osdc->notify_wq = NULL;
1794 ceph_msgpool_destroy(&osdc->msgpool_op);
1796 mempool_destroy(osdc->req_mempool);
1801 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1803 flush_workqueue(osdc->notify_wq);
1804 destroy_workqueue(osdc->notify_wq);
1805 cancel_delayed_work_sync(&osdc->timeout_work);
1806 cancel_delayed_work_sync(&osdc->osds_timeout_work);
1808 ceph_osdmap_destroy(osdc->osdmap);
1809 osdc->osdmap = NULL;
1811 remove_all_osds(osdc);
1812 mempool_destroy(osdc->req_mempool);
1813 ceph_msgpool_destroy(&osdc->msgpool_op);
1814 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1818 * Read some contiguous pages. If we cross a stripe boundary, shorten
1819 * *plen. Return number of bytes read, or error.
1821 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1822 struct ceph_vino vino, struct ceph_file_layout *layout,
1824 u32 truncate_seq, u64 truncate_size,
1825 struct page **pages, int num_pages, int page_align)
1827 struct ceph_osd_request *req;
1830 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1831 vino.snap, off, *plen);
1832 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1833 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1834 NULL, 0, truncate_seq, truncate_size, NULL,
1837 return PTR_ERR(req);
1839 /* it may be a short read due to an object boundary */
1840 req->r_pages = pages;
1842 dout("readpages final extent is %llu~%llu (%d pages align %d)\n",
1843 off, *plen, req->r_num_pages, page_align);
1845 rc = ceph_osdc_start_request(osdc, req, false);
1847 rc = ceph_osdc_wait_request(osdc, req);
1849 ceph_osdc_put_request(req);
1850 dout("readpages result %d\n", rc);
1853 EXPORT_SYMBOL(ceph_osdc_readpages);
1856 * do a synchronous write on N pages
1858 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1859 struct ceph_file_layout *layout,
1860 struct ceph_snap_context *snapc,
1862 u32 truncate_seq, u64 truncate_size,
1863 struct timespec *mtime,
1864 struct page **pages, int num_pages)
1866 struct ceph_osd_request *req;
1868 int page_align = off & ~PAGE_MASK;
1870 BUG_ON(vino.snap != CEPH_NOSNAP);
1871 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1873 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1875 truncate_seq, truncate_size, mtime,
1878 return PTR_ERR(req);
1880 /* it may be a short write due to an object boundary */
1881 req->r_pages = pages;
1882 dout("writepages %llu~%llu (%d pages)\n", off, len,
1885 rc = ceph_osdc_start_request(osdc, req, true);
1887 rc = ceph_osdc_wait_request(osdc, req);
1889 ceph_osdc_put_request(req);
1892 dout("writepages result %d\n", rc);
1895 EXPORT_SYMBOL(ceph_osdc_writepages);
1898 * handle incoming message
1900 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1902 struct ceph_osd *osd = con->private;
1903 struct ceph_osd_client *osdc;
1904 int type = le16_to_cpu(msg->hdr.type);
1911 case CEPH_MSG_OSD_MAP:
1912 ceph_osdc_handle_map(osdc, msg);
1914 case CEPH_MSG_OSD_OPREPLY:
1915 handle_reply(osdc, msg, con);
1917 case CEPH_MSG_WATCH_NOTIFY:
1918 handle_watch_notify(osdc, msg);
1922 pr_err("received unknown message type %d %s\n", type,
1923 ceph_msg_type_name(type));
1930 * lookup and return message for incoming reply. set up reply message
1933 static struct ceph_msg *get_reply(struct ceph_connection *con,
1934 struct ceph_msg_header *hdr,
1937 struct ceph_osd *osd = con->private;
1938 struct ceph_osd_client *osdc = osd->o_osdc;
1940 struct ceph_osd_request *req;
1941 int front = le32_to_cpu(hdr->front_len);
1942 int data_len = le32_to_cpu(hdr->data_len);
1945 tid = le64_to_cpu(hdr->tid);
1946 mutex_lock(&osdc->request_mutex);
1947 req = __lookup_request(osdc, tid);
1951 dout("get_reply unknown tid %llu from osd%d\n", tid,
1956 if (req->r_con_filling_msg) {
1957 dout("%s revoking msg %p from old con %p\n", __func__,
1958 req->r_reply, req->r_con_filling_msg);
1959 ceph_msg_revoke_incoming(req->r_reply);
1960 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
1961 req->r_con_filling_msg = NULL;
1964 if (front > req->r_reply->front.iov_len) {
1965 pr_warning("get_reply front %d > preallocated %d\n",
1966 front, (int)req->r_reply->front.iov_len);
1967 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
1970 ceph_msg_put(req->r_reply);
1973 m = ceph_msg_get(req->r_reply);
1976 int want = calc_pages_for(req->r_page_alignment, data_len);
1978 if (req->r_pages && unlikely(req->r_num_pages < want)) {
1979 pr_warning("tid %lld reply has %d bytes %d pages, we"
1980 " had only %d pages ready\n", tid, data_len,
1981 want, req->r_num_pages);
1987 m->pages = req->r_pages;
1988 m->nr_pages = req->r_num_pages;
1989 m->page_alignment = req->r_page_alignment;
1991 m->bio = req->r_bio;
1995 req->r_con_filling_msg = con->ops->get(con);
1996 dout("get_reply tid %lld %p\n", tid, m);
1999 mutex_unlock(&osdc->request_mutex);
2004 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2005 struct ceph_msg_header *hdr,
2008 struct ceph_osd *osd = con->private;
2009 int type = le16_to_cpu(hdr->type);
2010 int front = le32_to_cpu(hdr->front_len);
2014 case CEPH_MSG_OSD_MAP:
2015 case CEPH_MSG_WATCH_NOTIFY:
2016 return ceph_msg_new(type, front, GFP_NOFS, false);
2017 case CEPH_MSG_OSD_OPREPLY:
2018 return get_reply(con, hdr, skip);
2020 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2028 * Wrappers to refcount containing ceph_osd struct
2030 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2032 struct ceph_osd *osd = con->private;
2038 static void put_osd_con(struct ceph_connection *con)
2040 struct ceph_osd *osd = con->private;
2048 * Note: returned pointer is the address of a structure that's
2049 * managed separately. Caller must *not* attempt to free it.
2051 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2052 int *proto, int force_new)
2054 struct ceph_osd *o = con->private;
2055 struct ceph_osd_client *osdc = o->o_osdc;
2056 struct ceph_auth_client *ac = osdc->client->monc.auth;
2057 struct ceph_auth_handshake *auth = &o->o_auth;
2059 if (force_new && auth->authorizer) {
2060 if (ac->ops && ac->ops->destroy_authorizer)
2061 ac->ops->destroy_authorizer(ac, auth->authorizer);
2062 auth->authorizer = NULL;
2064 if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2065 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2068 return ERR_PTR(ret);
2070 *proto = ac->protocol;
2076 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2078 struct ceph_osd *o = con->private;
2079 struct ceph_osd_client *osdc = o->o_osdc;
2080 struct ceph_auth_client *ac = osdc->client->monc.auth;
2083 * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2084 * XXX which do we do: succeed or fail?
2086 return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2089 static int invalidate_authorizer(struct ceph_connection *con)
2091 struct ceph_osd *o = con->private;
2092 struct ceph_osd_client *osdc = o->o_osdc;
2093 struct ceph_auth_client *ac = osdc->client->monc.auth;
2095 if (ac->ops && ac->ops->invalidate_authorizer)
2096 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2098 return ceph_monc_validate_auth(&osdc->client->monc);
2101 static const struct ceph_connection_operations osd_con_ops = {
2104 .dispatch = dispatch,
2105 .get_authorizer = get_authorizer,
2106 .verify_authorizer_reply = verify_authorizer_reply,
2107 .invalidate_authorizer = invalidate_authorizer,
2108 .alloc_msg = alloc_msg,