FRV: Use generic show_interrupts()
[cris-mirror.git] / net / ceph / osd_client.c
blob02212ed50852eaa41b5d0c75c7f121b77f074ce3
1 #include <linux/ceph/ceph_debug.h>
3 #include <linux/module.h>
4 #include <linux/err.h>
5 #include <linux/highmem.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/slab.h>
9 #include <linux/uaccess.h>
10 #ifdef CONFIG_BLOCK
11 #include <linux/bio.h>
12 #endif
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 int __send_request(struct ceph_osd_client *osdc,
33 struct ceph_osd_request *req);
35 static int op_needs_trail(int op)
37 switch (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:
43 return 1;
44 default:
45 return 0;
49 static int op_has_extent(int op)
51 return (op == CEPH_OSD_OP_READ ||
52 op == CEPH_OSD_OP_WRITE);
55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
56 struct ceph_file_layout *layout,
57 u64 snapid,
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;
63 u64 orig_len = *plen;
64 u64 objoff, objlen; /* extent in object */
66 reqhead->snapid = cpu_to_le64(snapid);
68 /* object extent? */
69 ceph_calc_file_object_mapping(layout, off, plen, bno,
70 &objoff, &objlen);
71 if (*plen < orig_len)
72 dout(" skipping last %llu, final file extent %llu~%llu\n",
73 orig_len - *plen, off, *plen);
75 if (op_has_extent(op->op)) {
76 op->extent.offset = objoff;
77 op->extent.length = objlen;
79 req->r_num_pages = calc_pages_for(off, *plen);
80 req->r_page_alignment = off & ~PAGE_MASK;
81 if (op->op == CEPH_OSD_OP_WRITE)
82 op->payload_len = *plen;
84 dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
85 *bno, objoff, objlen, req->r_num_pages);
88 EXPORT_SYMBOL(ceph_calc_raw_layout);
91 * Implement client access to distributed object storage cluster.
93 * All data objects are stored within a cluster/cloud of OSDs, or
94 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
95 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
96 * remote daemons serving up and coordinating consistent and safe
97 * access to storage.
99 * Cluster membership and the mapping of data objects onto storage devices
100 * are described by the osd map.
102 * We keep track of pending OSD requests (read, write), resubmit
103 * requests to different OSDs when the cluster topology/data layout
104 * change, or retry the affected requests when the communications
105 * channel with an OSD is reset.
109 * calculate the mapping of a file extent onto an object, and fill out the
110 * request accordingly. shorten extent as necessary if it crosses an
111 * object boundary.
113 * fill osd op in request message.
115 static void calc_layout(struct ceph_osd_client *osdc,
116 struct ceph_vino vino,
117 struct ceph_file_layout *layout,
118 u64 off, u64 *plen,
119 struct ceph_osd_request *req,
120 struct ceph_osd_req_op *op)
122 u64 bno;
124 ceph_calc_raw_layout(osdc, layout, vino.snap, off,
125 plen, &bno, req, op);
127 sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
128 req->r_oid_len = strlen(req->r_oid);
132 * requests
134 void ceph_osdc_release_request(struct kref *kref)
136 struct ceph_osd_request *req = container_of(kref,
137 struct ceph_osd_request,
138 r_kref);
140 if (req->r_request)
141 ceph_msg_put(req->r_request);
142 if (req->r_reply)
143 ceph_msg_put(req->r_reply);
144 if (req->r_con_filling_msg) {
145 dout("release_request revoking pages %p from con %p\n",
146 req->r_pages, req->r_con_filling_msg);
147 ceph_con_revoke_message(req->r_con_filling_msg,
148 req->r_reply);
149 ceph_con_put(req->r_con_filling_msg);
151 if (req->r_own_pages)
152 ceph_release_page_vector(req->r_pages,
153 req->r_num_pages);
154 #ifdef CONFIG_BLOCK
155 if (req->r_bio)
156 bio_put(req->r_bio);
157 #endif
158 ceph_put_snap_context(req->r_snapc);
159 if (req->r_trail) {
160 ceph_pagelist_release(req->r_trail);
161 kfree(req->r_trail);
163 if (req->r_mempool)
164 mempool_free(req, req->r_osdc->req_mempool);
165 else
166 kfree(req);
168 EXPORT_SYMBOL(ceph_osdc_release_request);
170 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
172 int i = 0;
174 if (needs_trail)
175 *needs_trail = 0;
176 while (ops[i].op) {
177 if (needs_trail && op_needs_trail(ops[i].op))
178 *needs_trail = 1;
179 i++;
182 return i;
185 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
186 int flags,
187 struct ceph_snap_context *snapc,
188 struct ceph_osd_req_op *ops,
189 bool use_mempool,
190 gfp_t gfp_flags,
191 struct page **pages,
192 struct bio *bio)
194 struct ceph_osd_request *req;
195 struct ceph_msg *msg;
196 int needs_trail;
197 int num_op = get_num_ops(ops, &needs_trail);
198 size_t msg_size = sizeof(struct ceph_osd_request_head);
200 msg_size += num_op*sizeof(struct ceph_osd_op);
202 if (use_mempool) {
203 req = mempool_alloc(osdc->req_mempool, gfp_flags);
204 memset(req, 0, sizeof(*req));
205 } else {
206 req = kzalloc(sizeof(*req), gfp_flags);
208 if (req == NULL)
209 return NULL;
211 req->r_osdc = osdc;
212 req->r_mempool = use_mempool;
214 kref_init(&req->r_kref);
215 init_completion(&req->r_completion);
216 init_completion(&req->r_safe_completion);
217 INIT_LIST_HEAD(&req->r_unsafe_item);
218 INIT_LIST_HEAD(&req->r_linger_item);
219 INIT_LIST_HEAD(&req->r_linger_osd);
220 req->r_flags = flags;
222 WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
224 /* create reply message */
225 if (use_mempool)
226 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
227 else
228 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
229 OSD_OPREPLY_FRONT_LEN, gfp_flags);
230 if (!msg) {
231 ceph_osdc_put_request(req);
232 return NULL;
234 req->r_reply = msg;
236 /* allocate space for the trailing data */
237 if (needs_trail) {
238 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
239 if (!req->r_trail) {
240 ceph_osdc_put_request(req);
241 return NULL;
243 ceph_pagelist_init(req->r_trail);
245 /* create request message; allow space for oid */
246 msg_size += 40;
247 if (snapc)
248 msg_size += sizeof(u64) * snapc->num_snaps;
249 if (use_mempool)
250 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
251 else
252 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags);
253 if (!msg) {
254 ceph_osdc_put_request(req);
255 return NULL;
258 msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
259 memset(msg->front.iov_base, 0, msg->front.iov_len);
261 req->r_request = msg;
262 req->r_pages = pages;
263 #ifdef CONFIG_BLOCK
264 if (bio) {
265 req->r_bio = bio;
266 bio_get(req->r_bio);
268 #endif
270 return req;
272 EXPORT_SYMBOL(ceph_osdc_alloc_request);
274 static void osd_req_encode_op(struct ceph_osd_request *req,
275 struct ceph_osd_op *dst,
276 struct ceph_osd_req_op *src)
278 dst->op = cpu_to_le16(src->op);
280 switch (dst->op) {
281 case CEPH_OSD_OP_READ:
282 case CEPH_OSD_OP_WRITE:
283 dst->extent.offset =
284 cpu_to_le64(src->extent.offset);
285 dst->extent.length =
286 cpu_to_le64(src->extent.length);
287 dst->extent.truncate_size =
288 cpu_to_le64(src->extent.truncate_size);
289 dst->extent.truncate_seq =
290 cpu_to_le32(src->extent.truncate_seq);
291 break;
293 case CEPH_OSD_OP_GETXATTR:
294 case CEPH_OSD_OP_SETXATTR:
295 case CEPH_OSD_OP_CMPXATTR:
296 BUG_ON(!req->r_trail);
298 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
299 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
300 dst->xattr.cmp_op = src->xattr.cmp_op;
301 dst->xattr.cmp_mode = src->xattr.cmp_mode;
302 ceph_pagelist_append(req->r_trail, src->xattr.name,
303 src->xattr.name_len);
304 ceph_pagelist_append(req->r_trail, src->xattr.val,
305 src->xattr.value_len);
306 break;
307 case CEPH_OSD_OP_CALL:
308 BUG_ON(!req->r_trail);
310 dst->cls.class_len = src->cls.class_len;
311 dst->cls.method_len = src->cls.method_len;
312 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
314 ceph_pagelist_append(req->r_trail, src->cls.class_name,
315 src->cls.class_len);
316 ceph_pagelist_append(req->r_trail, src->cls.method_name,
317 src->cls.method_len);
318 ceph_pagelist_append(req->r_trail, src->cls.indata,
319 src->cls.indata_len);
320 break;
321 case CEPH_OSD_OP_ROLLBACK:
322 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
323 break;
324 case CEPH_OSD_OP_STARTSYNC:
325 break;
326 case CEPH_OSD_OP_NOTIFY:
328 __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
329 __le32 timeout = cpu_to_le32(src->watch.timeout);
331 BUG_ON(!req->r_trail);
333 ceph_pagelist_append(req->r_trail,
334 &prot_ver, sizeof(prot_ver));
335 ceph_pagelist_append(req->r_trail,
336 &timeout, sizeof(timeout));
338 case CEPH_OSD_OP_NOTIFY_ACK:
339 case CEPH_OSD_OP_WATCH:
340 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
341 dst->watch.ver = cpu_to_le64(src->watch.ver);
342 dst->watch.flag = src->watch.flag;
343 break;
344 default:
345 pr_err("unrecognized osd opcode %d\n", dst->op);
346 WARN_ON(1);
347 break;
349 dst->payload_len = cpu_to_le32(src->payload_len);
353 * build new request AND message
356 void ceph_osdc_build_request(struct ceph_osd_request *req,
357 u64 off, u64 *plen,
358 struct ceph_osd_req_op *src_ops,
359 struct ceph_snap_context *snapc,
360 struct timespec *mtime,
361 const char *oid,
362 int oid_len)
364 struct ceph_msg *msg = req->r_request;
365 struct ceph_osd_request_head *head;
366 struct ceph_osd_req_op *src_op;
367 struct ceph_osd_op *op;
368 void *p;
369 int num_op = get_num_ops(src_ops, NULL);
370 size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
371 int flags = req->r_flags;
372 u64 data_len = 0;
373 int i;
375 head = msg->front.iov_base;
376 op = (void *)(head + 1);
377 p = (void *)(op + num_op);
379 req->r_snapc = ceph_get_snap_context(snapc);
381 head->client_inc = cpu_to_le32(1); /* always, for now. */
382 head->flags = cpu_to_le32(flags);
383 if (flags & CEPH_OSD_FLAG_WRITE)
384 ceph_encode_timespec(&head->mtime, mtime);
385 head->num_ops = cpu_to_le16(num_op);
388 /* fill in oid */
389 head->object_len = cpu_to_le32(oid_len);
390 memcpy(p, oid, oid_len);
391 p += oid_len;
393 src_op = src_ops;
394 while (src_op->op) {
395 osd_req_encode_op(req, op, src_op);
396 src_op++;
397 op++;
400 if (req->r_trail)
401 data_len += req->r_trail->length;
403 if (snapc) {
404 head->snap_seq = cpu_to_le64(snapc->seq);
405 head->num_snaps = cpu_to_le32(snapc->num_snaps);
406 for (i = 0; i < snapc->num_snaps; i++) {
407 put_unaligned_le64(snapc->snaps[i], p);
408 p += sizeof(u64);
412 if (flags & CEPH_OSD_FLAG_WRITE) {
413 req->r_request->hdr.data_off = cpu_to_le16(off);
414 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
415 } else if (data_len) {
416 req->r_request->hdr.data_off = 0;
417 req->r_request->hdr.data_len = cpu_to_le32(data_len);
420 req->r_request->page_alignment = req->r_page_alignment;
422 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
423 msg_size = p - msg->front.iov_base;
424 msg->front.iov_len = msg_size;
425 msg->hdr.front_len = cpu_to_le32(msg_size);
426 return;
428 EXPORT_SYMBOL(ceph_osdc_build_request);
431 * build new request AND message, calculate layout, and adjust file
432 * extent as needed.
434 * if the file was recently truncated, we include information about its
435 * old and new size so that the object can be updated appropriately. (we
436 * avoid synchronously deleting truncated objects because it's slow.)
438 * if @do_sync, include a 'startsync' command so that the osd will flush
439 * data quickly.
441 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
442 struct ceph_file_layout *layout,
443 struct ceph_vino vino,
444 u64 off, u64 *plen,
445 int opcode, int flags,
446 struct ceph_snap_context *snapc,
447 int do_sync,
448 u32 truncate_seq,
449 u64 truncate_size,
450 struct timespec *mtime,
451 bool use_mempool, int num_reply,
452 int page_align)
454 struct ceph_osd_req_op ops[3];
455 struct ceph_osd_request *req;
457 ops[0].op = opcode;
458 ops[0].extent.truncate_seq = truncate_seq;
459 ops[0].extent.truncate_size = truncate_size;
460 ops[0].payload_len = 0;
462 if (do_sync) {
463 ops[1].op = CEPH_OSD_OP_STARTSYNC;
464 ops[1].payload_len = 0;
465 ops[2].op = 0;
466 } else
467 ops[1].op = 0;
469 req = ceph_osdc_alloc_request(osdc, flags,
470 snapc, ops,
471 use_mempool,
472 GFP_NOFS, NULL, NULL);
473 if (IS_ERR(req))
474 return req;
476 /* calculate max write size */
477 calc_layout(osdc, vino, layout, off, plen, req, ops);
478 req->r_file_layout = *layout; /* keep a copy */
480 /* in case it differs from natural alignment that calc_layout
481 filled in for us */
482 req->r_page_alignment = page_align;
484 ceph_osdc_build_request(req, off, plen, ops,
485 snapc,
486 mtime,
487 req->r_oid, req->r_oid_len);
489 return req;
491 EXPORT_SYMBOL(ceph_osdc_new_request);
494 * We keep osd requests in an rbtree, sorted by ->r_tid.
496 static void __insert_request(struct ceph_osd_client *osdc,
497 struct ceph_osd_request *new)
499 struct rb_node **p = &osdc->requests.rb_node;
500 struct rb_node *parent = NULL;
501 struct ceph_osd_request *req = NULL;
503 while (*p) {
504 parent = *p;
505 req = rb_entry(parent, struct ceph_osd_request, r_node);
506 if (new->r_tid < req->r_tid)
507 p = &(*p)->rb_left;
508 else if (new->r_tid > req->r_tid)
509 p = &(*p)->rb_right;
510 else
511 BUG();
514 rb_link_node(&new->r_node, parent, p);
515 rb_insert_color(&new->r_node, &osdc->requests);
518 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
519 u64 tid)
521 struct ceph_osd_request *req;
522 struct rb_node *n = osdc->requests.rb_node;
524 while (n) {
525 req = rb_entry(n, struct ceph_osd_request, r_node);
526 if (tid < req->r_tid)
527 n = n->rb_left;
528 else if (tid > req->r_tid)
529 n = n->rb_right;
530 else
531 return req;
533 return NULL;
536 static struct ceph_osd_request *
537 __lookup_request_ge(struct ceph_osd_client *osdc,
538 u64 tid)
540 struct ceph_osd_request *req;
541 struct rb_node *n = osdc->requests.rb_node;
543 while (n) {
544 req = rb_entry(n, struct ceph_osd_request, r_node);
545 if (tid < req->r_tid) {
546 if (!n->rb_left)
547 return req;
548 n = n->rb_left;
549 } else if (tid > req->r_tid) {
550 n = n->rb_right;
551 } else {
552 return req;
555 return NULL;
559 * Resubmit requests pending on the given osd.
561 static void __kick_osd_requests(struct ceph_osd_client *osdc,
562 struct ceph_osd *osd)
564 struct ceph_osd_request *req, *nreq;
565 int err;
567 dout("__kick_osd_requests osd%d\n", osd->o_osd);
568 err = __reset_osd(osdc, osd);
569 if (err == -EAGAIN)
570 return;
572 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
573 list_move(&req->r_req_lru_item, &osdc->req_unsent);
574 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
575 osd->o_osd);
576 if (!req->r_linger)
577 req->r_flags |= CEPH_OSD_FLAG_RETRY;
580 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
581 r_linger_osd) {
582 __unregister_linger_request(osdc, req);
583 __register_request(osdc, req);
584 list_move(&req->r_req_lru_item, &osdc->req_unsent);
585 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
586 osd->o_osd);
590 static void kick_osd_requests(struct ceph_osd_client *osdc,
591 struct ceph_osd *kickosd)
593 mutex_lock(&osdc->request_mutex);
594 __kick_osd_requests(osdc, kickosd);
595 mutex_unlock(&osdc->request_mutex);
599 * If the osd connection drops, we need to resubmit all requests.
601 static void osd_reset(struct ceph_connection *con)
603 struct ceph_osd *osd = con->private;
604 struct ceph_osd_client *osdc;
606 if (!osd)
607 return;
608 dout("osd_reset osd%d\n", osd->o_osd);
609 osdc = osd->o_osdc;
610 down_read(&osdc->map_sem);
611 kick_osd_requests(osdc, osd);
612 send_queued(osdc);
613 up_read(&osdc->map_sem);
617 * Track open sessions with osds.
619 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
621 struct ceph_osd *osd;
623 osd = kzalloc(sizeof(*osd), GFP_NOFS);
624 if (!osd)
625 return NULL;
627 atomic_set(&osd->o_ref, 1);
628 osd->o_osdc = osdc;
629 INIT_LIST_HEAD(&osd->o_requests);
630 INIT_LIST_HEAD(&osd->o_linger_requests);
631 INIT_LIST_HEAD(&osd->o_osd_lru);
632 osd->o_incarnation = 1;
634 ceph_con_init(osdc->client->msgr, &osd->o_con);
635 osd->o_con.private = osd;
636 osd->o_con.ops = &osd_con_ops;
637 osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
639 INIT_LIST_HEAD(&osd->o_keepalive_item);
640 return osd;
643 static struct ceph_osd *get_osd(struct ceph_osd *osd)
645 if (atomic_inc_not_zero(&osd->o_ref)) {
646 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
647 atomic_read(&osd->o_ref));
648 return osd;
649 } else {
650 dout("get_osd %p FAIL\n", osd);
651 return NULL;
655 static void put_osd(struct ceph_osd *osd)
657 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
658 atomic_read(&osd->o_ref) - 1);
659 if (atomic_dec_and_test(&osd->o_ref)) {
660 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
662 if (osd->o_authorizer)
663 ac->ops->destroy_authorizer(ac, osd->o_authorizer);
664 kfree(osd);
669 * remove an osd from our map
671 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
673 dout("__remove_osd %p\n", osd);
674 BUG_ON(!list_empty(&osd->o_requests));
675 rb_erase(&osd->o_node, &osdc->osds);
676 list_del_init(&osd->o_osd_lru);
677 ceph_con_close(&osd->o_con);
678 put_osd(osd);
681 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
682 struct ceph_osd *osd)
684 dout("__move_osd_to_lru %p\n", osd);
685 BUG_ON(!list_empty(&osd->o_osd_lru));
686 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
687 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
690 static void __remove_osd_from_lru(struct ceph_osd *osd)
692 dout("__remove_osd_from_lru %p\n", osd);
693 if (!list_empty(&osd->o_osd_lru))
694 list_del_init(&osd->o_osd_lru);
697 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
699 struct ceph_osd *osd, *nosd;
701 dout("__remove_old_osds %p\n", osdc);
702 mutex_lock(&osdc->request_mutex);
703 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
704 if (!remove_all && time_before(jiffies, osd->lru_ttl))
705 break;
706 __remove_osd(osdc, osd);
708 mutex_unlock(&osdc->request_mutex);
712 * reset osd connect
714 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
716 struct ceph_osd_request *req;
717 int ret = 0;
719 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
720 if (list_empty(&osd->o_requests) &&
721 list_empty(&osd->o_linger_requests)) {
722 __remove_osd(osdc, osd);
723 } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
724 &osd->o_con.peer_addr,
725 sizeof(osd->o_con.peer_addr)) == 0 &&
726 !ceph_con_opened(&osd->o_con)) {
727 dout(" osd addr hasn't changed and connection never opened,"
728 " letting msgr retry");
729 /* touch each r_stamp for handle_timeout()'s benfit */
730 list_for_each_entry(req, &osd->o_requests, r_osd_item)
731 req->r_stamp = jiffies;
732 ret = -EAGAIN;
733 } else {
734 ceph_con_close(&osd->o_con);
735 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
736 osd->o_incarnation++;
738 return ret;
741 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
743 struct rb_node **p = &osdc->osds.rb_node;
744 struct rb_node *parent = NULL;
745 struct ceph_osd *osd = NULL;
747 while (*p) {
748 parent = *p;
749 osd = rb_entry(parent, struct ceph_osd, o_node);
750 if (new->o_osd < osd->o_osd)
751 p = &(*p)->rb_left;
752 else if (new->o_osd > osd->o_osd)
753 p = &(*p)->rb_right;
754 else
755 BUG();
758 rb_link_node(&new->o_node, parent, p);
759 rb_insert_color(&new->o_node, &osdc->osds);
762 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
764 struct ceph_osd *osd;
765 struct rb_node *n = osdc->osds.rb_node;
767 while (n) {
768 osd = rb_entry(n, struct ceph_osd, o_node);
769 if (o < osd->o_osd)
770 n = n->rb_left;
771 else if (o > osd->o_osd)
772 n = n->rb_right;
773 else
774 return osd;
776 return NULL;
779 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
781 schedule_delayed_work(&osdc->timeout_work,
782 osdc->client->options->osd_keepalive_timeout * HZ);
785 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
787 cancel_delayed_work(&osdc->timeout_work);
791 * Register request, assign tid. If this is the first request, set up
792 * the timeout event.
794 static void __register_request(struct ceph_osd_client *osdc,
795 struct ceph_osd_request *req)
797 req->r_tid = ++osdc->last_tid;
798 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
799 INIT_LIST_HEAD(&req->r_req_lru_item);
801 dout("register_request %p tid %lld\n", req, req->r_tid);
802 __insert_request(osdc, req);
803 ceph_osdc_get_request(req);
804 osdc->num_requests++;
806 if (osdc->num_requests == 1) {
807 dout(" first request, scheduling timeout\n");
808 __schedule_osd_timeout(osdc);
812 static void register_request(struct ceph_osd_client *osdc,
813 struct ceph_osd_request *req)
815 mutex_lock(&osdc->request_mutex);
816 __register_request(osdc, req);
817 mutex_unlock(&osdc->request_mutex);
821 * called under osdc->request_mutex
823 static void __unregister_request(struct ceph_osd_client *osdc,
824 struct ceph_osd_request *req)
826 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
827 rb_erase(&req->r_node, &osdc->requests);
828 osdc->num_requests--;
830 if (req->r_osd) {
831 /* make sure the original request isn't in flight. */
832 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
834 list_del_init(&req->r_osd_item);
835 if (list_empty(&req->r_osd->o_requests) &&
836 list_empty(&req->r_osd->o_linger_requests)) {
837 dout("moving osd to %p lru\n", req->r_osd);
838 __move_osd_to_lru(osdc, req->r_osd);
840 if (list_empty(&req->r_osd_item) &&
841 list_empty(&req->r_linger_item))
842 req->r_osd = NULL;
845 ceph_osdc_put_request(req);
847 list_del_init(&req->r_req_lru_item);
848 if (osdc->num_requests == 0) {
849 dout(" no requests, canceling timeout\n");
850 __cancel_osd_timeout(osdc);
855 * Cancel a previously queued request message
857 static void __cancel_request(struct ceph_osd_request *req)
859 if (req->r_sent && req->r_osd) {
860 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
861 req->r_sent = 0;
865 static void __register_linger_request(struct ceph_osd_client *osdc,
866 struct ceph_osd_request *req)
868 dout("__register_linger_request %p\n", req);
869 list_add_tail(&req->r_linger_item, &osdc->req_linger);
870 list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests);
873 static void __unregister_linger_request(struct ceph_osd_client *osdc,
874 struct ceph_osd_request *req)
876 dout("__unregister_linger_request %p\n", req);
877 if (req->r_osd) {
878 list_del_init(&req->r_linger_item);
879 list_del_init(&req->r_linger_osd);
881 if (list_empty(&req->r_osd->o_requests) &&
882 list_empty(&req->r_osd->o_linger_requests)) {
883 dout("moving osd to %p lru\n", req->r_osd);
884 __move_osd_to_lru(osdc, req->r_osd);
886 req->r_osd = NULL;
890 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
891 struct ceph_osd_request *req)
893 mutex_lock(&osdc->request_mutex);
894 if (req->r_linger) {
895 __unregister_linger_request(osdc, req);
896 ceph_osdc_put_request(req);
898 mutex_unlock(&osdc->request_mutex);
900 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
902 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
903 struct ceph_osd_request *req)
905 if (!req->r_linger) {
906 dout("set_request_linger %p\n", req);
907 req->r_linger = 1;
909 * caller is now responsible for calling
910 * unregister_linger_request
912 ceph_osdc_get_request(req);
915 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
918 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
919 * (as needed), and set the request r_osd appropriately. If there is
920 * no up osd, set r_osd to NULL. Move the request to the appropiate list
921 * (unsent, homeless) or leave on in-flight lru.
923 * Return 0 if unchanged, 1 if changed, or negative on error.
925 * Caller should hold map_sem for read and request_mutex.
927 static int __map_request(struct ceph_osd_client *osdc,
928 struct ceph_osd_request *req)
930 struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
931 struct ceph_pg pgid;
932 int acting[CEPH_PG_MAX_SIZE];
933 int o = -1, num = 0;
934 int err;
936 dout("map_request %p tid %lld\n", req, req->r_tid);
937 err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
938 &req->r_file_layout, osdc->osdmap);
939 if (err) {
940 list_move(&req->r_req_lru_item, &osdc->req_notarget);
941 return err;
943 pgid = reqhead->layout.ol_pgid;
944 req->r_pgid = pgid;
946 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
947 if (err > 0) {
948 o = acting[0];
949 num = err;
952 if ((req->r_osd && req->r_osd->o_osd == o &&
953 req->r_sent >= req->r_osd->o_incarnation &&
954 req->r_num_pg_osds == num &&
955 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
956 (req->r_osd == NULL && o == -1))
957 return 0; /* no change */
959 dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
960 req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
961 req->r_osd ? req->r_osd->o_osd : -1);
963 /* record full pg acting set */
964 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
965 req->r_num_pg_osds = num;
967 if (req->r_osd) {
968 __cancel_request(req);
969 list_del_init(&req->r_osd_item);
970 req->r_osd = NULL;
973 req->r_osd = __lookup_osd(osdc, o);
974 if (!req->r_osd && o >= 0) {
975 err = -ENOMEM;
976 req->r_osd = create_osd(osdc);
977 if (!req->r_osd) {
978 list_move(&req->r_req_lru_item, &osdc->req_notarget);
979 goto out;
982 dout("map_request osd %p is osd%d\n", req->r_osd, o);
983 req->r_osd->o_osd = o;
984 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
985 __insert_osd(osdc, req->r_osd);
987 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
990 if (req->r_osd) {
991 __remove_osd_from_lru(req->r_osd);
992 list_add(&req->r_osd_item, &req->r_osd->o_requests);
993 list_move(&req->r_req_lru_item, &osdc->req_unsent);
994 } else {
995 list_move(&req->r_req_lru_item, &osdc->req_notarget);
997 err = 1; /* osd or pg changed */
999 out:
1000 return err;
1004 * caller should hold map_sem (for read) and request_mutex
1006 static int __send_request(struct ceph_osd_client *osdc,
1007 struct ceph_osd_request *req)
1009 struct ceph_osd_request_head *reqhead;
1011 dout("send_request %p tid %llu to osd%d flags %d\n",
1012 req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1014 reqhead = req->r_request->front.iov_base;
1015 reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1016 reqhead->flags |= cpu_to_le32(req->r_flags); /* e.g., RETRY */
1017 reqhead->reassert_version = req->r_reassert_version;
1019 req->r_stamp = jiffies;
1020 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1022 ceph_msg_get(req->r_request); /* send consumes a ref */
1023 ceph_con_send(&req->r_osd->o_con, req->r_request);
1024 req->r_sent = req->r_osd->o_incarnation;
1025 return 0;
1029 * Send any requests in the queue (req_unsent).
1031 static void send_queued(struct ceph_osd_client *osdc)
1033 struct ceph_osd_request *req, *tmp;
1035 dout("send_queued\n");
1036 mutex_lock(&osdc->request_mutex);
1037 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1038 __send_request(osdc, req);
1040 mutex_unlock(&osdc->request_mutex);
1044 * Timeout callback, called every N seconds when 1 or more osd
1045 * requests has been active for more than N seconds. When this
1046 * happens, we ping all OSDs with requests who have timed out to
1047 * ensure any communications channel reset is detected. Reset the
1048 * request timeouts another N seconds in the future as we go.
1049 * Reschedule the timeout event another N seconds in future (unless
1050 * there are no open requests).
1052 static void handle_timeout(struct work_struct *work)
1054 struct ceph_osd_client *osdc =
1055 container_of(work, struct ceph_osd_client, timeout_work.work);
1056 struct ceph_osd_request *req, *last_req = NULL;
1057 struct ceph_osd *osd;
1058 unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1059 unsigned long keepalive =
1060 osdc->client->options->osd_keepalive_timeout * HZ;
1061 unsigned long last_stamp = 0;
1062 struct list_head slow_osds;
1063 dout("timeout\n");
1064 down_read(&osdc->map_sem);
1066 ceph_monc_request_next_osdmap(&osdc->client->monc);
1068 mutex_lock(&osdc->request_mutex);
1071 * reset osds that appear to be _really_ unresponsive. this
1072 * is a failsafe measure.. we really shouldn't be getting to
1073 * this point if the system is working properly. the monitors
1074 * should mark the osd as failed and we should find out about
1075 * it from an updated osd map.
1077 while (timeout && !list_empty(&osdc->req_lru)) {
1078 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1079 r_req_lru_item);
1081 if (time_before(jiffies, req->r_stamp + timeout))
1082 break;
1084 BUG_ON(req == last_req && req->r_stamp == last_stamp);
1085 last_req = req;
1086 last_stamp = req->r_stamp;
1088 osd = req->r_osd;
1089 BUG_ON(!osd);
1090 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1091 req->r_tid, osd->o_osd);
1092 __kick_osd_requests(osdc, osd);
1096 * ping osds that are a bit slow. this ensures that if there
1097 * is a break in the TCP connection we will notice, and reopen
1098 * a connection with that osd (from the fault callback).
1100 INIT_LIST_HEAD(&slow_osds);
1101 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1102 if (time_before(jiffies, req->r_stamp + keepalive))
1103 break;
1105 osd = req->r_osd;
1106 BUG_ON(!osd);
1107 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1108 req->r_tid, osd->o_osd);
1109 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1111 while (!list_empty(&slow_osds)) {
1112 osd = list_entry(slow_osds.next, struct ceph_osd,
1113 o_keepalive_item);
1114 list_del_init(&osd->o_keepalive_item);
1115 ceph_con_keepalive(&osd->o_con);
1118 __schedule_osd_timeout(osdc);
1119 mutex_unlock(&osdc->request_mutex);
1120 send_queued(osdc);
1121 up_read(&osdc->map_sem);
1124 static void handle_osds_timeout(struct work_struct *work)
1126 struct ceph_osd_client *osdc =
1127 container_of(work, struct ceph_osd_client,
1128 osds_timeout_work.work);
1129 unsigned long delay =
1130 osdc->client->options->osd_idle_ttl * HZ >> 2;
1132 dout("osds timeout\n");
1133 down_read(&osdc->map_sem);
1134 remove_old_osds(osdc, 0);
1135 up_read(&osdc->map_sem);
1137 schedule_delayed_work(&osdc->osds_timeout_work,
1138 round_jiffies_relative(delay));
1142 * handle osd op reply. either call the callback if it is specified,
1143 * or do the completion to wake up the waiting thread.
1145 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1146 struct ceph_connection *con)
1148 struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1149 struct ceph_osd_request *req;
1150 u64 tid;
1151 int numops, object_len, flags;
1152 s32 result;
1154 tid = le64_to_cpu(msg->hdr.tid);
1155 if (msg->front.iov_len < sizeof(*rhead))
1156 goto bad;
1157 numops = le32_to_cpu(rhead->num_ops);
1158 object_len = le32_to_cpu(rhead->object_len);
1159 result = le32_to_cpu(rhead->result);
1160 if (msg->front.iov_len != sizeof(*rhead) + object_len +
1161 numops * sizeof(struct ceph_osd_op))
1162 goto bad;
1163 dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1164 /* lookup */
1165 mutex_lock(&osdc->request_mutex);
1166 req = __lookup_request(osdc, tid);
1167 if (req == NULL) {
1168 dout("handle_reply tid %llu dne\n", tid);
1169 mutex_unlock(&osdc->request_mutex);
1170 return;
1172 ceph_osdc_get_request(req);
1173 flags = le32_to_cpu(rhead->flags);
1176 * if this connection filled our message, drop our reference now, to
1177 * avoid a (safe but slower) revoke later.
1179 if (req->r_con_filling_msg == con && req->r_reply == msg) {
1180 dout(" dropping con_filling_msg ref %p\n", con);
1181 req->r_con_filling_msg = NULL;
1182 ceph_con_put(con);
1185 if (!req->r_got_reply) {
1186 unsigned bytes;
1188 req->r_result = le32_to_cpu(rhead->result);
1189 bytes = le32_to_cpu(msg->hdr.data_len);
1190 dout("handle_reply result %d bytes %d\n", req->r_result,
1191 bytes);
1192 if (req->r_result == 0)
1193 req->r_result = bytes;
1195 /* in case this is a write and we need to replay, */
1196 req->r_reassert_version = rhead->reassert_version;
1198 req->r_got_reply = 1;
1199 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1200 dout("handle_reply tid %llu dup ack\n", tid);
1201 mutex_unlock(&osdc->request_mutex);
1202 goto done;
1205 dout("handle_reply tid %llu flags %d\n", tid, flags);
1207 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1208 __register_linger_request(osdc, req);
1210 /* either this is a read, or we got the safe response */
1211 if (result < 0 ||
1212 (flags & CEPH_OSD_FLAG_ONDISK) ||
1213 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1214 __unregister_request(osdc, req);
1216 mutex_unlock(&osdc->request_mutex);
1218 if (req->r_callback)
1219 req->r_callback(req, msg);
1220 else
1221 complete_all(&req->r_completion);
1223 if (flags & CEPH_OSD_FLAG_ONDISK) {
1224 if (req->r_safe_callback)
1225 req->r_safe_callback(req, msg);
1226 complete_all(&req->r_safe_completion); /* fsync waiter */
1229 done:
1230 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1231 ceph_osdc_put_request(req);
1232 return;
1234 bad:
1235 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1236 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1237 (int)sizeof(*rhead));
1238 ceph_msg_dump(msg);
1241 static void reset_changed_osds(struct ceph_osd_client *osdc)
1243 struct rb_node *p, *n;
1245 for (p = rb_first(&osdc->osds); p; p = n) {
1246 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1248 n = rb_next(p);
1249 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1250 memcmp(&osd->o_con.peer_addr,
1251 ceph_osd_addr(osdc->osdmap,
1252 osd->o_osd),
1253 sizeof(struct ceph_entity_addr)) != 0)
1254 __reset_osd(osdc, osd);
1259 * Requeue requests whose mapping to an OSD has changed. If requests map to
1260 * no osd, request a new map.
1262 * Caller should hold map_sem for read and request_mutex.
1264 static void kick_requests(struct ceph_osd_client *osdc)
1266 struct ceph_osd_request *req, *nreq;
1267 struct rb_node *p;
1268 int needmap = 0;
1269 int err;
1271 dout("kick_requests\n");
1272 mutex_lock(&osdc->request_mutex);
1273 for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1274 req = rb_entry(p, struct ceph_osd_request, r_node);
1275 err = __map_request(osdc, req);
1276 if (err < 0)
1277 continue; /* error */
1278 if (req->r_osd == NULL) {
1279 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1280 needmap++; /* request a newer map */
1281 } else if (err > 0) {
1282 dout("%p tid %llu requeued on osd%d\n", req, req->r_tid,
1283 req->r_osd ? req->r_osd->o_osd : -1);
1284 if (!req->r_linger)
1285 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1289 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1290 r_linger_item) {
1291 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1293 err = __map_request(osdc, req);
1294 if (err == 0)
1295 continue; /* no change and no osd was specified */
1296 if (err < 0)
1297 continue; /* hrm! */
1298 if (req->r_osd == NULL) {
1299 dout("tid %llu maps to no valid osd\n", req->r_tid);
1300 needmap++; /* request a newer map */
1301 continue;
1304 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1305 req->r_osd ? req->r_osd->o_osd : -1);
1306 __unregister_linger_request(osdc, req);
1307 __register_request(osdc, req);
1309 mutex_unlock(&osdc->request_mutex);
1311 if (needmap) {
1312 dout("%d requests for down osds, need new map\n", needmap);
1313 ceph_monc_request_next_osdmap(&osdc->client->monc);
1319 * Process updated osd map.
1321 * The message contains any number of incremental and full maps, normally
1322 * indicating some sort of topology change in the cluster. Kick requests
1323 * off to different OSDs as needed.
1325 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1327 void *p, *end, *next;
1328 u32 nr_maps, maplen;
1329 u32 epoch;
1330 struct ceph_osdmap *newmap = NULL, *oldmap;
1331 int err;
1332 struct ceph_fsid fsid;
1334 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1335 p = msg->front.iov_base;
1336 end = p + msg->front.iov_len;
1338 /* verify fsid */
1339 ceph_decode_need(&p, end, sizeof(fsid), bad);
1340 ceph_decode_copy(&p, &fsid, sizeof(fsid));
1341 if (ceph_check_fsid(osdc->client, &fsid) < 0)
1342 return;
1344 down_write(&osdc->map_sem);
1346 /* incremental maps */
1347 ceph_decode_32_safe(&p, end, nr_maps, bad);
1348 dout(" %d inc maps\n", nr_maps);
1349 while (nr_maps > 0) {
1350 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1351 epoch = ceph_decode_32(&p);
1352 maplen = ceph_decode_32(&p);
1353 ceph_decode_need(&p, end, maplen, bad);
1354 next = p + maplen;
1355 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1356 dout("applying incremental map %u len %d\n",
1357 epoch, maplen);
1358 newmap = osdmap_apply_incremental(&p, next,
1359 osdc->osdmap,
1360 osdc->client->msgr);
1361 if (IS_ERR(newmap)) {
1362 err = PTR_ERR(newmap);
1363 goto bad;
1365 BUG_ON(!newmap);
1366 if (newmap != osdc->osdmap) {
1367 ceph_osdmap_destroy(osdc->osdmap);
1368 osdc->osdmap = newmap;
1370 kick_requests(osdc);
1371 reset_changed_osds(osdc);
1372 } else {
1373 dout("ignoring incremental map %u len %d\n",
1374 epoch, maplen);
1376 p = next;
1377 nr_maps--;
1379 if (newmap)
1380 goto done;
1382 /* full maps */
1383 ceph_decode_32_safe(&p, end, nr_maps, bad);
1384 dout(" %d full maps\n", nr_maps);
1385 while (nr_maps) {
1386 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1387 epoch = ceph_decode_32(&p);
1388 maplen = ceph_decode_32(&p);
1389 ceph_decode_need(&p, end, maplen, bad);
1390 if (nr_maps > 1) {
1391 dout("skipping non-latest full map %u len %d\n",
1392 epoch, maplen);
1393 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1394 dout("skipping full map %u len %d, "
1395 "older than our %u\n", epoch, maplen,
1396 osdc->osdmap->epoch);
1397 } else {
1398 dout("taking full map %u len %d\n", epoch, maplen);
1399 newmap = osdmap_decode(&p, p+maplen);
1400 if (IS_ERR(newmap)) {
1401 err = PTR_ERR(newmap);
1402 goto bad;
1404 BUG_ON(!newmap);
1405 oldmap = osdc->osdmap;
1406 osdc->osdmap = newmap;
1407 if (oldmap)
1408 ceph_osdmap_destroy(oldmap);
1409 kick_requests(osdc);
1411 p += maplen;
1412 nr_maps--;
1415 done:
1416 downgrade_write(&osdc->map_sem);
1417 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1418 send_queued(osdc);
1419 up_read(&osdc->map_sem);
1420 wake_up_all(&osdc->client->auth_wq);
1421 return;
1423 bad:
1424 pr_err("osdc handle_map corrupt msg\n");
1425 ceph_msg_dump(msg);
1426 up_write(&osdc->map_sem);
1427 return;
1431 * watch/notify callback event infrastructure
1433 * These callbacks are used both for watch and notify operations.
1435 static void __release_event(struct kref *kref)
1437 struct ceph_osd_event *event =
1438 container_of(kref, struct ceph_osd_event, kref);
1440 dout("__release_event %p\n", event);
1441 kfree(event);
1444 static void get_event(struct ceph_osd_event *event)
1446 kref_get(&event->kref);
1449 void ceph_osdc_put_event(struct ceph_osd_event *event)
1451 kref_put(&event->kref, __release_event);
1453 EXPORT_SYMBOL(ceph_osdc_put_event);
1455 static void __insert_event(struct ceph_osd_client *osdc,
1456 struct ceph_osd_event *new)
1458 struct rb_node **p = &osdc->event_tree.rb_node;
1459 struct rb_node *parent = NULL;
1460 struct ceph_osd_event *event = NULL;
1462 while (*p) {
1463 parent = *p;
1464 event = rb_entry(parent, struct ceph_osd_event, node);
1465 if (new->cookie < event->cookie)
1466 p = &(*p)->rb_left;
1467 else if (new->cookie > event->cookie)
1468 p = &(*p)->rb_right;
1469 else
1470 BUG();
1473 rb_link_node(&new->node, parent, p);
1474 rb_insert_color(&new->node, &osdc->event_tree);
1477 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1478 u64 cookie)
1480 struct rb_node **p = &osdc->event_tree.rb_node;
1481 struct rb_node *parent = NULL;
1482 struct ceph_osd_event *event = NULL;
1484 while (*p) {
1485 parent = *p;
1486 event = rb_entry(parent, struct ceph_osd_event, node);
1487 if (cookie < event->cookie)
1488 p = &(*p)->rb_left;
1489 else if (cookie > event->cookie)
1490 p = &(*p)->rb_right;
1491 else
1492 return event;
1494 return NULL;
1497 static void __remove_event(struct ceph_osd_event *event)
1499 struct ceph_osd_client *osdc = event->osdc;
1501 if (!RB_EMPTY_NODE(&event->node)) {
1502 dout("__remove_event removed %p\n", event);
1503 rb_erase(&event->node, &osdc->event_tree);
1504 ceph_osdc_put_event(event);
1505 } else {
1506 dout("__remove_event didn't remove %p\n", event);
1510 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1511 void (*event_cb)(u64, u64, u8, void *),
1512 int one_shot, void *data,
1513 struct ceph_osd_event **pevent)
1515 struct ceph_osd_event *event;
1517 event = kmalloc(sizeof(*event), GFP_NOIO);
1518 if (!event)
1519 return -ENOMEM;
1521 dout("create_event %p\n", event);
1522 event->cb = event_cb;
1523 event->one_shot = one_shot;
1524 event->data = data;
1525 event->osdc = osdc;
1526 INIT_LIST_HEAD(&event->osd_node);
1527 kref_init(&event->kref); /* one ref for us */
1528 kref_get(&event->kref); /* one ref for the caller */
1529 init_completion(&event->completion);
1531 spin_lock(&osdc->event_lock);
1532 event->cookie = ++osdc->event_count;
1533 __insert_event(osdc, event);
1534 spin_unlock(&osdc->event_lock);
1536 *pevent = event;
1537 return 0;
1539 EXPORT_SYMBOL(ceph_osdc_create_event);
1541 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1543 struct ceph_osd_client *osdc = event->osdc;
1545 dout("cancel_event %p\n", event);
1546 spin_lock(&osdc->event_lock);
1547 __remove_event(event);
1548 spin_unlock(&osdc->event_lock);
1549 ceph_osdc_put_event(event); /* caller's */
1551 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1554 static void do_event_work(struct work_struct *work)
1556 struct ceph_osd_event_work *event_work =
1557 container_of(work, struct ceph_osd_event_work, work);
1558 struct ceph_osd_event *event = event_work->event;
1559 u64 ver = event_work->ver;
1560 u64 notify_id = event_work->notify_id;
1561 u8 opcode = event_work->opcode;
1563 dout("do_event_work completing %p\n", event);
1564 event->cb(ver, notify_id, opcode, event->data);
1565 complete(&event->completion);
1566 dout("do_event_work completed %p\n", event);
1567 ceph_osdc_put_event(event);
1568 kfree(event_work);
1573 * Process osd watch notifications
1575 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1577 void *p, *end;
1578 u8 proto_ver;
1579 u64 cookie, ver, notify_id;
1580 u8 opcode;
1581 struct ceph_osd_event *event;
1582 struct ceph_osd_event_work *event_work;
1584 p = msg->front.iov_base;
1585 end = p + msg->front.iov_len;
1587 ceph_decode_8_safe(&p, end, proto_ver, bad);
1588 ceph_decode_8_safe(&p, end, opcode, bad);
1589 ceph_decode_64_safe(&p, end, cookie, bad);
1590 ceph_decode_64_safe(&p, end, ver, bad);
1591 ceph_decode_64_safe(&p, end, notify_id, bad);
1593 spin_lock(&osdc->event_lock);
1594 event = __find_event(osdc, cookie);
1595 if (event) {
1596 get_event(event);
1597 if (event->one_shot)
1598 __remove_event(event);
1600 spin_unlock(&osdc->event_lock);
1601 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1602 cookie, ver, event);
1603 if (event) {
1604 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1605 INIT_WORK(&event_work->work, do_event_work);
1606 if (!event_work) {
1607 dout("ERROR: could not allocate event_work\n");
1608 goto done_err;
1610 event_work->event = event;
1611 event_work->ver = ver;
1612 event_work->notify_id = notify_id;
1613 event_work->opcode = opcode;
1614 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1615 dout("WARNING: failed to queue notify event work\n");
1616 goto done_err;
1620 return;
1622 done_err:
1623 complete(&event->completion);
1624 ceph_osdc_put_event(event);
1625 return;
1627 bad:
1628 pr_err("osdc handle_watch_notify corrupt msg\n");
1629 return;
1632 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1634 int err;
1636 dout("wait_event %p\n", event);
1637 err = wait_for_completion_interruptible_timeout(&event->completion,
1638 timeout * HZ);
1639 ceph_osdc_put_event(event);
1640 if (err > 0)
1641 err = 0;
1642 dout("wait_event %p returns %d\n", event, err);
1643 return err;
1645 EXPORT_SYMBOL(ceph_osdc_wait_event);
1648 * Register request, send initial attempt.
1650 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1651 struct ceph_osd_request *req,
1652 bool nofail)
1654 int rc = 0;
1656 req->r_request->pages = req->r_pages;
1657 req->r_request->nr_pages = req->r_num_pages;
1658 #ifdef CONFIG_BLOCK
1659 req->r_request->bio = req->r_bio;
1660 #endif
1661 req->r_request->trail = req->r_trail;
1663 register_request(osdc, req);
1665 down_read(&osdc->map_sem);
1666 mutex_lock(&osdc->request_mutex);
1668 * a racing kick_requests() may have sent the message for us
1669 * while we dropped request_mutex above, so only send now if
1670 * the request still han't been touched yet.
1672 if (req->r_sent == 0) {
1673 rc = __map_request(osdc, req);
1674 if (rc < 0)
1675 return rc;
1676 if (req->r_osd == NULL) {
1677 dout("send_request %p no up osds in pg\n", req);
1678 ceph_monc_request_next_osdmap(&osdc->client->monc);
1679 } else {
1680 rc = __send_request(osdc, req);
1681 if (rc) {
1682 if (nofail) {
1683 dout("osdc_start_request failed send, "
1684 " will retry %lld\n", req->r_tid);
1685 rc = 0;
1686 } else {
1687 __unregister_request(osdc, req);
1692 mutex_unlock(&osdc->request_mutex);
1693 up_read(&osdc->map_sem);
1694 return rc;
1696 EXPORT_SYMBOL(ceph_osdc_start_request);
1699 * wait for a request to complete
1701 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1702 struct ceph_osd_request *req)
1704 int rc;
1706 rc = wait_for_completion_interruptible(&req->r_completion);
1707 if (rc < 0) {
1708 mutex_lock(&osdc->request_mutex);
1709 __cancel_request(req);
1710 __unregister_request(osdc, req);
1711 mutex_unlock(&osdc->request_mutex);
1712 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1713 return rc;
1716 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1717 return req->r_result;
1719 EXPORT_SYMBOL(ceph_osdc_wait_request);
1722 * sync - wait for all in-flight requests to flush. avoid starvation.
1724 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1726 struct ceph_osd_request *req;
1727 u64 last_tid, next_tid = 0;
1729 mutex_lock(&osdc->request_mutex);
1730 last_tid = osdc->last_tid;
1731 while (1) {
1732 req = __lookup_request_ge(osdc, next_tid);
1733 if (!req)
1734 break;
1735 if (req->r_tid > last_tid)
1736 break;
1738 next_tid = req->r_tid + 1;
1739 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1740 continue;
1742 ceph_osdc_get_request(req);
1743 mutex_unlock(&osdc->request_mutex);
1744 dout("sync waiting on tid %llu (last is %llu)\n",
1745 req->r_tid, last_tid);
1746 wait_for_completion(&req->r_safe_completion);
1747 mutex_lock(&osdc->request_mutex);
1748 ceph_osdc_put_request(req);
1750 mutex_unlock(&osdc->request_mutex);
1751 dout("sync done (thru tid %llu)\n", last_tid);
1753 EXPORT_SYMBOL(ceph_osdc_sync);
1756 * init, shutdown
1758 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1760 int err;
1762 dout("init\n");
1763 osdc->client = client;
1764 osdc->osdmap = NULL;
1765 init_rwsem(&osdc->map_sem);
1766 init_completion(&osdc->map_waiters);
1767 osdc->last_requested_map = 0;
1768 mutex_init(&osdc->request_mutex);
1769 osdc->last_tid = 0;
1770 osdc->osds = RB_ROOT;
1771 INIT_LIST_HEAD(&osdc->osd_lru);
1772 osdc->requests = RB_ROOT;
1773 INIT_LIST_HEAD(&osdc->req_lru);
1774 INIT_LIST_HEAD(&osdc->req_unsent);
1775 INIT_LIST_HEAD(&osdc->req_notarget);
1776 INIT_LIST_HEAD(&osdc->req_linger);
1777 osdc->num_requests = 0;
1778 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1779 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1780 spin_lock_init(&osdc->event_lock);
1781 osdc->event_tree = RB_ROOT;
1782 osdc->event_count = 0;
1784 schedule_delayed_work(&osdc->osds_timeout_work,
1785 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1787 err = -ENOMEM;
1788 osdc->req_mempool = mempool_create_kmalloc_pool(10,
1789 sizeof(struct ceph_osd_request));
1790 if (!osdc->req_mempool)
1791 goto out;
1793 err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1794 "osd_op");
1795 if (err < 0)
1796 goto out_mempool;
1797 err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1798 OSD_OPREPLY_FRONT_LEN, 10, true,
1799 "osd_op_reply");
1800 if (err < 0)
1801 goto out_msgpool;
1803 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1804 if (IS_ERR(osdc->notify_wq)) {
1805 err = PTR_ERR(osdc->notify_wq);
1806 osdc->notify_wq = NULL;
1807 goto out_msgpool;
1809 return 0;
1811 out_msgpool:
1812 ceph_msgpool_destroy(&osdc->msgpool_op);
1813 out_mempool:
1814 mempool_destroy(osdc->req_mempool);
1815 out:
1816 return err;
1818 EXPORT_SYMBOL(ceph_osdc_init);
1820 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1822 flush_workqueue(osdc->notify_wq);
1823 destroy_workqueue(osdc->notify_wq);
1824 cancel_delayed_work_sync(&osdc->timeout_work);
1825 cancel_delayed_work_sync(&osdc->osds_timeout_work);
1826 if (osdc->osdmap) {
1827 ceph_osdmap_destroy(osdc->osdmap);
1828 osdc->osdmap = NULL;
1830 remove_old_osds(osdc, 1);
1831 WARN_ON(!RB_EMPTY_ROOT(&osdc->osds));
1832 mempool_destroy(osdc->req_mempool);
1833 ceph_msgpool_destroy(&osdc->msgpool_op);
1834 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1836 EXPORT_SYMBOL(ceph_osdc_stop);
1839 * Read some contiguous pages. If we cross a stripe boundary, shorten
1840 * *plen. Return number of bytes read, or error.
1842 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1843 struct ceph_vino vino, struct ceph_file_layout *layout,
1844 u64 off, u64 *plen,
1845 u32 truncate_seq, u64 truncate_size,
1846 struct page **pages, int num_pages, int page_align)
1848 struct ceph_osd_request *req;
1849 int rc = 0;
1851 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1852 vino.snap, off, *plen);
1853 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1854 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1855 NULL, 0, truncate_seq, truncate_size, NULL,
1856 false, 1, page_align);
1857 if (!req)
1858 return -ENOMEM;
1860 /* it may be a short read due to an object boundary */
1861 req->r_pages = pages;
1863 dout("readpages final extent is %llu~%llu (%d pages align %d)\n",
1864 off, *plen, req->r_num_pages, page_align);
1866 rc = ceph_osdc_start_request(osdc, req, false);
1867 if (!rc)
1868 rc = ceph_osdc_wait_request(osdc, req);
1870 ceph_osdc_put_request(req);
1871 dout("readpages result %d\n", rc);
1872 return rc;
1874 EXPORT_SYMBOL(ceph_osdc_readpages);
1877 * do a synchronous write on N pages
1879 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1880 struct ceph_file_layout *layout,
1881 struct ceph_snap_context *snapc,
1882 u64 off, u64 len,
1883 u32 truncate_seq, u64 truncate_size,
1884 struct timespec *mtime,
1885 struct page **pages, int num_pages,
1886 int flags, int do_sync, bool nofail)
1888 struct ceph_osd_request *req;
1889 int rc = 0;
1890 int page_align = off & ~PAGE_MASK;
1892 BUG_ON(vino.snap != CEPH_NOSNAP);
1893 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1894 CEPH_OSD_OP_WRITE,
1895 flags | CEPH_OSD_FLAG_ONDISK |
1896 CEPH_OSD_FLAG_WRITE,
1897 snapc, do_sync,
1898 truncate_seq, truncate_size, mtime,
1899 nofail, 1, page_align);
1900 if (!req)
1901 return -ENOMEM;
1903 /* it may be a short write due to an object boundary */
1904 req->r_pages = pages;
1905 dout("writepages %llu~%llu (%d pages)\n", off, len,
1906 req->r_num_pages);
1908 rc = ceph_osdc_start_request(osdc, req, nofail);
1909 if (!rc)
1910 rc = ceph_osdc_wait_request(osdc, req);
1912 ceph_osdc_put_request(req);
1913 if (rc == 0)
1914 rc = len;
1915 dout("writepages result %d\n", rc);
1916 return rc;
1918 EXPORT_SYMBOL(ceph_osdc_writepages);
1921 * handle incoming message
1923 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1925 struct ceph_osd *osd = con->private;
1926 struct ceph_osd_client *osdc;
1927 int type = le16_to_cpu(msg->hdr.type);
1929 if (!osd)
1930 goto out;
1931 osdc = osd->o_osdc;
1933 switch (type) {
1934 case CEPH_MSG_OSD_MAP:
1935 ceph_osdc_handle_map(osdc, msg);
1936 break;
1937 case CEPH_MSG_OSD_OPREPLY:
1938 handle_reply(osdc, msg, con);
1939 break;
1940 case CEPH_MSG_WATCH_NOTIFY:
1941 handle_watch_notify(osdc, msg);
1942 break;
1944 default:
1945 pr_err("received unknown message type %d %s\n", type,
1946 ceph_msg_type_name(type));
1948 out:
1949 ceph_msg_put(msg);
1953 * lookup and return message for incoming reply. set up reply message
1954 * pages.
1956 static struct ceph_msg *get_reply(struct ceph_connection *con,
1957 struct ceph_msg_header *hdr,
1958 int *skip)
1960 struct ceph_osd *osd = con->private;
1961 struct ceph_osd_client *osdc = osd->o_osdc;
1962 struct ceph_msg *m;
1963 struct ceph_osd_request *req;
1964 int front = le32_to_cpu(hdr->front_len);
1965 int data_len = le32_to_cpu(hdr->data_len);
1966 u64 tid;
1968 tid = le64_to_cpu(hdr->tid);
1969 mutex_lock(&osdc->request_mutex);
1970 req = __lookup_request(osdc, tid);
1971 if (!req) {
1972 *skip = 1;
1973 m = NULL;
1974 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
1975 osd->o_osd);
1976 goto out;
1979 if (req->r_con_filling_msg) {
1980 dout("get_reply revoking msg %p from old con %p\n",
1981 req->r_reply, req->r_con_filling_msg);
1982 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
1983 ceph_con_put(req->r_con_filling_msg);
1984 req->r_con_filling_msg = NULL;
1987 if (front > req->r_reply->front.iov_len) {
1988 pr_warning("get_reply front %d > preallocated %d\n",
1989 front, (int)req->r_reply->front.iov_len);
1990 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS);
1991 if (!m)
1992 goto out;
1993 ceph_msg_put(req->r_reply);
1994 req->r_reply = m;
1996 m = ceph_msg_get(req->r_reply);
1998 if (data_len > 0) {
1999 int want = calc_pages_for(req->r_page_alignment, data_len);
2001 if (unlikely(req->r_num_pages < want)) {
2002 pr_warning("tid %lld reply %d > expected %d pages\n",
2003 tid, want, m->nr_pages);
2004 *skip = 1;
2005 ceph_msg_put(m);
2006 m = NULL;
2007 goto out;
2009 m->pages = req->r_pages;
2010 m->nr_pages = req->r_num_pages;
2011 m->page_alignment = req->r_page_alignment;
2012 #ifdef CONFIG_BLOCK
2013 m->bio = req->r_bio;
2014 #endif
2016 *skip = 0;
2017 req->r_con_filling_msg = ceph_con_get(con);
2018 dout("get_reply tid %lld %p\n", tid, m);
2020 out:
2021 mutex_unlock(&osdc->request_mutex);
2022 return m;
2026 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2027 struct ceph_msg_header *hdr,
2028 int *skip)
2030 struct ceph_osd *osd = con->private;
2031 int type = le16_to_cpu(hdr->type);
2032 int front = le32_to_cpu(hdr->front_len);
2034 switch (type) {
2035 case CEPH_MSG_OSD_MAP:
2036 case CEPH_MSG_WATCH_NOTIFY:
2037 return ceph_msg_new(type, front, GFP_NOFS);
2038 case CEPH_MSG_OSD_OPREPLY:
2039 return get_reply(con, hdr, skip);
2040 default:
2041 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2042 osd->o_osd);
2043 *skip = 1;
2044 return NULL;
2049 * Wrappers to refcount containing ceph_osd struct
2051 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2053 struct ceph_osd *osd = con->private;
2054 if (get_osd(osd))
2055 return con;
2056 return NULL;
2059 static void put_osd_con(struct ceph_connection *con)
2061 struct ceph_osd *osd = con->private;
2062 put_osd(osd);
2066 * authentication
2068 static int get_authorizer(struct ceph_connection *con,
2069 void **buf, int *len, int *proto,
2070 void **reply_buf, int *reply_len, int force_new)
2072 struct ceph_osd *o = con->private;
2073 struct ceph_osd_client *osdc = o->o_osdc;
2074 struct ceph_auth_client *ac = osdc->client->monc.auth;
2075 int ret = 0;
2077 if (force_new && o->o_authorizer) {
2078 ac->ops->destroy_authorizer(ac, o->o_authorizer);
2079 o->o_authorizer = NULL;
2081 if (o->o_authorizer == NULL) {
2082 ret = ac->ops->create_authorizer(
2083 ac, CEPH_ENTITY_TYPE_OSD,
2084 &o->o_authorizer,
2085 &o->o_authorizer_buf,
2086 &o->o_authorizer_buf_len,
2087 &o->o_authorizer_reply_buf,
2088 &o->o_authorizer_reply_buf_len);
2089 if (ret)
2090 return ret;
2093 *proto = ac->protocol;
2094 *buf = o->o_authorizer_buf;
2095 *len = o->o_authorizer_buf_len;
2096 *reply_buf = o->o_authorizer_reply_buf;
2097 *reply_len = o->o_authorizer_reply_buf_len;
2098 return 0;
2102 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2104 struct ceph_osd *o = con->private;
2105 struct ceph_osd_client *osdc = o->o_osdc;
2106 struct ceph_auth_client *ac = osdc->client->monc.auth;
2108 return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
2111 static int invalidate_authorizer(struct ceph_connection *con)
2113 struct ceph_osd *o = con->private;
2114 struct ceph_osd_client *osdc = o->o_osdc;
2115 struct ceph_auth_client *ac = osdc->client->monc.auth;
2117 if (ac->ops->invalidate_authorizer)
2118 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2120 return ceph_monc_validate_auth(&osdc->client->monc);
2123 static const struct ceph_connection_operations osd_con_ops = {
2124 .get = get_osd_con,
2125 .put = put_osd_con,
2126 .dispatch = dispatch,
2127 .get_authorizer = get_authorizer,
2128 .verify_authorizer_reply = verify_authorizer_reply,
2129 .invalidate_authorizer = invalidate_authorizer,
2130 .alloc_msg = alloc_msg,
2131 .fault = osd_reset,