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);
36 * Implement client access to distributed object storage cluster.
38 * All data objects are stored within a cluster/cloud of OSDs, or
39 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
40 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
41 * remote daemons serving up and coordinating consistent and safe
44 * Cluster membership and the mapping of data objects onto storage devices
45 * are described by the osd map.
47 * We keep track of pending OSD requests (read, write), resubmit
48 * requests to different OSDs when the cluster topology/data layout
49 * change, or retry the affected requests when the communications
50 * channel with an OSD is reset.
54 * calculate the mapping of a file extent onto an object, and fill out the
55 * request accordingly. shorten extent as necessary if it crosses an
58 * fill osd op in request message.
60 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
61 u64 *objnum, u64 *objoff, u64 *objlen)
67 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
71 if (*objlen < orig_len) {
73 dout(" skipping last %llu, final file extent %llu~%llu\n",
74 orig_len - *plen, off, *plen);
77 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
85 void ceph_osdc_release_request(struct kref *kref)
88 struct ceph_osd_request *req = container_of(kref,
89 struct ceph_osd_request,
93 ceph_msg_put(req->r_request);
94 if (req->r_con_filling_msg) {
95 dout("%s revoking msg %p from con %p\n", __func__,
96 req->r_reply, req->r_con_filling_msg);
97 ceph_msg_revoke_incoming(req->r_reply);
98 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
99 req->r_con_filling_msg = NULL;
102 ceph_msg_put(req->r_reply);
104 if (req->r_data_in.type == CEPH_OSD_DATA_TYPE_PAGES &&
105 req->r_data_in.own_pages) {
106 num_pages = calc_pages_for((u64)req->r_data_in.alignment,
107 (u64)req->r_data_in.length);
108 ceph_release_page_vector(req->r_data_in.pages, num_pages);
110 if (req->r_data_out.type == CEPH_OSD_DATA_TYPE_PAGES &&
111 req->r_data_out.own_pages) {
112 num_pages = calc_pages_for((u64)req->r_data_out.alignment,
113 (u64)req->r_data_out.length);
114 ceph_release_page_vector(req->r_data_out.pages, num_pages);
117 ceph_put_snap_context(req->r_snapc);
119 mempool_free(req, req->r_osdc->req_mempool);
123 EXPORT_SYMBOL(ceph_osdc_release_request);
125 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
126 struct ceph_snap_context *snapc,
127 unsigned int num_ops,
131 struct ceph_osd_request *req;
132 struct ceph_msg *msg;
135 msg_size = 4 + 4 + 8 + 8 + 4+8;
136 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
137 msg_size += 1 + 8 + 4 + 4; /* pg_t */
138 msg_size += 4 + MAX_OBJ_NAME_SIZE;
139 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
140 msg_size += 8; /* snapid */
141 msg_size += 8; /* snap_seq */
142 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
146 req = mempool_alloc(osdc->req_mempool, gfp_flags);
147 memset(req, 0, sizeof(*req));
149 req = kzalloc(sizeof(*req), gfp_flags);
155 req->r_mempool = use_mempool;
157 kref_init(&req->r_kref);
158 init_completion(&req->r_completion);
159 init_completion(&req->r_safe_completion);
160 RB_CLEAR_NODE(&req->r_node);
161 INIT_LIST_HEAD(&req->r_unsafe_item);
162 INIT_LIST_HEAD(&req->r_linger_item);
163 INIT_LIST_HEAD(&req->r_linger_osd);
164 INIT_LIST_HEAD(&req->r_req_lru_item);
165 INIT_LIST_HEAD(&req->r_osd_item);
167 /* create reply message */
169 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
171 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
172 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
174 ceph_osdc_put_request(req);
179 req->r_data_in.type = CEPH_OSD_DATA_TYPE_NONE;
180 req->r_data_out.type = CEPH_OSD_DATA_TYPE_NONE;
182 /* create request message; allow space for oid */
184 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
186 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
188 ceph_osdc_put_request(req);
192 memset(msg->front.iov_base, 0, msg->front.iov_len);
194 req->r_request = msg;
198 EXPORT_SYMBOL(ceph_osdc_alloc_request);
200 static bool osd_req_opcode_valid(u16 opcode)
203 case CEPH_OSD_OP_READ:
204 case CEPH_OSD_OP_STAT:
205 case CEPH_OSD_OP_MAPEXT:
206 case CEPH_OSD_OP_MASKTRUNC:
207 case CEPH_OSD_OP_SPARSE_READ:
208 case CEPH_OSD_OP_NOTIFY:
209 case CEPH_OSD_OP_NOTIFY_ACK:
210 case CEPH_OSD_OP_ASSERT_VER:
211 case CEPH_OSD_OP_WRITE:
212 case CEPH_OSD_OP_WRITEFULL:
213 case CEPH_OSD_OP_TRUNCATE:
214 case CEPH_OSD_OP_ZERO:
215 case CEPH_OSD_OP_DELETE:
216 case CEPH_OSD_OP_APPEND:
217 case CEPH_OSD_OP_STARTSYNC:
218 case CEPH_OSD_OP_SETTRUNC:
219 case CEPH_OSD_OP_TRIMTRUNC:
220 case CEPH_OSD_OP_TMAPUP:
221 case CEPH_OSD_OP_TMAPPUT:
222 case CEPH_OSD_OP_TMAPGET:
223 case CEPH_OSD_OP_CREATE:
224 case CEPH_OSD_OP_ROLLBACK:
225 case CEPH_OSD_OP_WATCH:
226 case CEPH_OSD_OP_OMAPGETKEYS:
227 case CEPH_OSD_OP_OMAPGETVALS:
228 case CEPH_OSD_OP_OMAPGETHEADER:
229 case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
230 case CEPH_OSD_OP_OMAPSETVALS:
231 case CEPH_OSD_OP_OMAPSETHEADER:
232 case CEPH_OSD_OP_OMAPCLEAR:
233 case CEPH_OSD_OP_OMAPRMKEYS:
234 case CEPH_OSD_OP_OMAP_CMP:
235 case CEPH_OSD_OP_CLONERANGE:
236 case CEPH_OSD_OP_ASSERT_SRC_VERSION:
237 case CEPH_OSD_OP_SRC_CMPXATTR:
238 case CEPH_OSD_OP_GETXATTR:
239 case CEPH_OSD_OP_GETXATTRS:
240 case CEPH_OSD_OP_CMPXATTR:
241 case CEPH_OSD_OP_SETXATTR:
242 case CEPH_OSD_OP_SETXATTRS:
243 case CEPH_OSD_OP_RESETXATTRS:
244 case CEPH_OSD_OP_RMXATTR:
245 case CEPH_OSD_OP_PULL:
246 case CEPH_OSD_OP_PUSH:
247 case CEPH_OSD_OP_BALANCEREADS:
248 case CEPH_OSD_OP_UNBALANCEREADS:
249 case CEPH_OSD_OP_SCRUB:
250 case CEPH_OSD_OP_SCRUB_RESERVE:
251 case CEPH_OSD_OP_SCRUB_UNRESERVE:
252 case CEPH_OSD_OP_SCRUB_STOP:
253 case CEPH_OSD_OP_SCRUB_MAP:
254 case CEPH_OSD_OP_WRLOCK:
255 case CEPH_OSD_OP_WRUNLOCK:
256 case CEPH_OSD_OP_RDLOCK:
257 case CEPH_OSD_OP_RDUNLOCK:
258 case CEPH_OSD_OP_UPLOCK:
259 case CEPH_OSD_OP_DNLOCK:
260 case CEPH_OSD_OP_CALL:
261 case CEPH_OSD_OP_PGLS:
262 case CEPH_OSD_OP_PGLS_FILTER:
270 * This is an osd op init function for opcodes that have no data or
271 * other information associated with them. It also serves as a
272 * common init routine for all the other init functions, below.
274 void osd_req_op_init(struct ceph_osd_req_op *op, u16 opcode)
276 BUG_ON(!osd_req_opcode_valid(opcode));
278 memset(op, 0, sizeof (*op));
283 void osd_req_op_extent_init(struct ceph_osd_req_op *op, u16 opcode,
284 u64 offset, u64 length,
285 u64 truncate_size, u32 truncate_seq)
287 size_t payload_len = 0;
289 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE);
291 osd_req_op_init(op, opcode);
293 op->extent.offset = offset;
294 op->extent.length = length;
295 op->extent.truncate_size = truncate_size;
296 op->extent.truncate_seq = truncate_seq;
297 if (opcode == CEPH_OSD_OP_WRITE)
298 payload_len += length;
300 op->payload_len = payload_len;
302 EXPORT_SYMBOL(osd_req_op_extent_init);
304 void osd_req_op_cls_init(struct ceph_osd_req_op *op, u16 opcode,
305 const char *class, const char *method,
306 const void *request_data, size_t request_data_size)
308 size_t payload_len = 0;
311 BUG_ON(opcode != CEPH_OSD_OP_CALL);
313 osd_req_op_init(op, opcode);
315 op->cls.class_name = class;
316 size = strlen(class);
317 BUG_ON(size > (size_t) U8_MAX);
318 op->cls.class_len = size;
321 op->cls.method_name = method;
322 size = strlen(method);
323 BUG_ON(size > (size_t) U8_MAX);
324 op->cls.method_len = size;
327 op->cls.indata = request_data;
328 BUG_ON(request_data_size > (size_t) U32_MAX);
329 op->cls.indata_len = (u32) request_data_size;
330 payload_len += request_data_size;
332 op->cls.argc = 0; /* currently unused */
334 op->payload_len = payload_len;
336 EXPORT_SYMBOL(osd_req_op_cls_init);
338 void osd_req_op_watch_init(struct ceph_osd_req_op *op, u16 opcode,
339 u64 cookie, u64 version, int flag)
341 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
343 osd_req_op_init(op, opcode);
345 op->watch.cookie = cookie;
346 /* op->watch.ver = version; */ /* XXX 3847 */
347 op->watch.ver = cpu_to_le64(version);
348 if (opcode == CEPH_OSD_OP_WATCH && flag)
349 op->watch.flag = (u8) 1;
351 EXPORT_SYMBOL(osd_req_op_watch_init);
353 static u64 osd_req_encode_op(struct ceph_osd_request *req,
354 struct ceph_osd_op *dst,
355 struct ceph_osd_req_op *src)
357 u64 out_data_len = 0;
358 struct ceph_pagelist *pagelist;
360 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
361 pr_err("unrecognized osd opcode %d\n", src->op);
367 case CEPH_OSD_OP_STAT:
369 case CEPH_OSD_OP_READ:
370 case CEPH_OSD_OP_WRITE:
371 if (src->op == CEPH_OSD_OP_WRITE)
372 out_data_len = src->extent.length;
373 dst->extent.offset = cpu_to_le64(src->extent.offset);
374 dst->extent.length = cpu_to_le64(src->extent.length);
375 dst->extent.truncate_size =
376 cpu_to_le64(src->extent.truncate_size);
377 dst->extent.truncate_seq =
378 cpu_to_le32(src->extent.truncate_seq);
380 case CEPH_OSD_OP_CALL:
381 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
383 ceph_pagelist_init(pagelist);
385 dst->cls.class_len = src->cls.class_len;
386 dst->cls.method_len = src->cls.method_len;
387 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
388 ceph_pagelist_append(pagelist, src->cls.class_name,
390 ceph_pagelist_append(pagelist, src->cls.method_name,
391 src->cls.method_len);
392 ceph_pagelist_append(pagelist, src->cls.indata,
393 src->cls.indata_len);
395 req->r_data_out.type = CEPH_OSD_DATA_TYPE_PAGELIST;
396 req->r_data_out.pagelist = pagelist;
397 out_data_len = pagelist->length;
399 case CEPH_OSD_OP_STARTSYNC:
401 case CEPH_OSD_OP_NOTIFY_ACK:
402 case CEPH_OSD_OP_WATCH:
403 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
404 dst->watch.ver = cpu_to_le64(src->watch.ver);
405 dst->watch.flag = src->watch.flag;
408 pr_err("unsupported osd opcode %s\n",
409 ceph_osd_op_name(src->op));
414 dst->op = cpu_to_le16(src->op);
415 dst->payload_len = cpu_to_le32(src->payload_len);
421 * build new request AND message
424 void ceph_osdc_build_request(struct ceph_osd_request *req,
425 u64 off, unsigned int num_ops,
426 struct ceph_osd_req_op *src_ops,
427 struct ceph_snap_context *snapc, u64 snap_id,
428 struct timespec *mtime)
430 struct ceph_msg *msg = req->r_request;
431 struct ceph_osd_req_op *src_op;
434 int flags = req->r_flags;
438 req->r_num_ops = num_ops;
439 req->r_snapid = snap_id;
440 req->r_snapc = ceph_get_snap_context(snapc);
443 msg->hdr.version = cpu_to_le16(4);
445 p = msg->front.iov_base;
446 ceph_encode_32(&p, 1); /* client_inc is always 1 */
447 req->r_request_osdmap_epoch = p;
449 req->r_request_flags = p;
451 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
452 ceph_encode_timespec(p, mtime);
453 p += sizeof(struct ceph_timespec);
454 req->r_request_reassert_version = p;
455 p += sizeof(struct ceph_eversion); /* will get filled in */
458 ceph_encode_8(&p, 4);
459 ceph_encode_8(&p, 4);
460 ceph_encode_32(&p, 8 + 4 + 4);
461 req->r_request_pool = p;
463 ceph_encode_32(&p, -1); /* preferred */
464 ceph_encode_32(&p, 0); /* key len */
466 ceph_encode_8(&p, 1);
467 req->r_request_pgid = p;
469 ceph_encode_32(&p, -1); /* preferred */
472 ceph_encode_32(&p, req->r_oid_len);
473 memcpy(p, req->r_oid, req->r_oid_len);
474 dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len);
477 /* ops--can imply data */
478 ceph_encode_16(&p, num_ops);
480 req->r_request_ops = p;
482 for (i = 0; i < num_ops; i++, src_op++) {
483 data_len += osd_req_encode_op(req, p, src_op);
484 p += sizeof(struct ceph_osd_op);
488 ceph_encode_64(&p, req->r_snapid);
489 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
490 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
492 for (i = 0; i < snapc->num_snaps; i++) {
493 ceph_encode_64(&p, req->r_snapc->snaps[i]);
497 req->r_request_attempts = p;
501 if (flags & CEPH_OSD_FLAG_WRITE) {
505 * The header "data_off" is a hint to the receiver
506 * allowing it to align received data into its
507 * buffers such that there's no need to re-copy
508 * it before writing it to disk (direct I/O).
510 data_off = (u16) (off & 0xffff);
511 req->r_request->hdr.data_off = cpu_to_le16(data_off);
513 req->r_request->hdr.data_len = cpu_to_le32(data_len);
515 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
516 msg_size = p - msg->front.iov_base;
517 msg->front.iov_len = msg_size;
518 msg->hdr.front_len = cpu_to_le32(msg_size);
520 dout("build_request msg_size was %d num_ops %d\n", (int)msg_size,
524 EXPORT_SYMBOL(ceph_osdc_build_request);
527 * build new request AND message, calculate layout, and adjust file
530 * if the file was recently truncated, we include information about its
531 * old and new size so that the object can be updated appropriately. (we
532 * avoid synchronously deleting truncated objects because it's slow.)
534 * if @do_sync, include a 'startsync' command so that the osd will flush
537 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
538 struct ceph_file_layout *layout,
539 struct ceph_vino vino,
541 int opcode, int flags,
542 struct ceph_snap_context *snapc,
546 struct timespec *mtime,
549 struct ceph_osd_req_op ops[2];
550 struct ceph_osd_request *req;
551 unsigned int num_op = do_sync ? 2 : 1;
559 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE);
561 req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
564 return ERR_PTR(-ENOMEM);
565 req->r_flags = flags;
567 /* calculate max write size */
568 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
570 ceph_osdc_put_request(req);
574 object_size = le32_to_cpu(layout->fl_object_size);
575 object_base = off - objoff;
576 if (truncate_size <= object_base) {
579 truncate_size -= object_base;
580 if (truncate_size > object_size)
581 truncate_size = object_size;
584 osd_req_op_extent_init(&ops[0], opcode, objoff, objlen,
585 truncate_size, truncate_seq);
587 osd_req_op_init(&ops[1], CEPH_OSD_OP_STARTSYNC);
589 req->r_file_layout = *layout; /* keep a copy */
591 snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx",
593 req->r_oid_len = strlen(req->r_oid);
595 ceph_osdc_build_request(req, off, num_op, ops,
596 snapc, vino.snap, mtime);
600 EXPORT_SYMBOL(ceph_osdc_new_request);
603 * We keep osd requests in an rbtree, sorted by ->r_tid.
605 static void __insert_request(struct ceph_osd_client *osdc,
606 struct ceph_osd_request *new)
608 struct rb_node **p = &osdc->requests.rb_node;
609 struct rb_node *parent = NULL;
610 struct ceph_osd_request *req = NULL;
614 req = rb_entry(parent, struct ceph_osd_request, r_node);
615 if (new->r_tid < req->r_tid)
617 else if (new->r_tid > req->r_tid)
623 rb_link_node(&new->r_node, parent, p);
624 rb_insert_color(&new->r_node, &osdc->requests);
627 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
630 struct ceph_osd_request *req;
631 struct rb_node *n = osdc->requests.rb_node;
634 req = rb_entry(n, struct ceph_osd_request, r_node);
635 if (tid < req->r_tid)
637 else if (tid > req->r_tid)
645 static struct ceph_osd_request *
646 __lookup_request_ge(struct ceph_osd_client *osdc,
649 struct ceph_osd_request *req;
650 struct rb_node *n = osdc->requests.rb_node;
653 req = rb_entry(n, struct ceph_osd_request, r_node);
654 if (tid < req->r_tid) {
658 } else if (tid > req->r_tid) {
668 * Resubmit requests pending on the given osd.
670 static void __kick_osd_requests(struct ceph_osd_client *osdc,
671 struct ceph_osd *osd)
673 struct ceph_osd_request *req, *nreq;
677 dout("__kick_osd_requests osd%d\n", osd->o_osd);
678 err = __reset_osd(osdc, osd);
682 * Build up a list of requests to resend by traversing the
683 * osd's list of requests. Requests for a given object are
684 * sent in tid order, and that is also the order they're
685 * kept on this list. Therefore all requests that are in
686 * flight will be found first, followed by all requests that
687 * have not yet been sent. And to resend requests while
688 * preserving this order we will want to put any sent
689 * requests back on the front of the osd client's unsent
692 * So we build a separate ordered list of already-sent
693 * requests for the affected osd and splice it onto the
694 * front of the osd client's unsent list. Once we've seen a
695 * request that has not yet been sent we're done. Those
696 * requests are already sitting right where they belong.
698 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
701 list_move_tail(&req->r_req_lru_item, &resend);
702 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
705 req->r_flags |= CEPH_OSD_FLAG_RETRY;
707 list_splice(&resend, &osdc->req_unsent);
710 * Linger requests are re-registered before sending, which
711 * sets up a new tid for each. We add them to the unsent
712 * list at the end to keep things in tid order.
714 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
717 * reregister request prior to unregistering linger so
718 * that r_osd is preserved.
720 BUG_ON(!list_empty(&req->r_req_lru_item));
721 __register_request(osdc, req);
722 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
723 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
724 __unregister_linger_request(osdc, req);
725 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
731 * If the osd connection drops, we need to resubmit all requests.
733 static void osd_reset(struct ceph_connection *con)
735 struct ceph_osd *osd = con->private;
736 struct ceph_osd_client *osdc;
740 dout("osd_reset osd%d\n", osd->o_osd);
742 down_read(&osdc->map_sem);
743 mutex_lock(&osdc->request_mutex);
744 __kick_osd_requests(osdc, osd);
746 mutex_unlock(&osdc->request_mutex);
747 up_read(&osdc->map_sem);
751 * Track open sessions with osds.
753 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
755 struct ceph_osd *osd;
757 osd = kzalloc(sizeof(*osd), GFP_NOFS);
761 atomic_set(&osd->o_ref, 1);
764 RB_CLEAR_NODE(&osd->o_node);
765 INIT_LIST_HEAD(&osd->o_requests);
766 INIT_LIST_HEAD(&osd->o_linger_requests);
767 INIT_LIST_HEAD(&osd->o_osd_lru);
768 osd->o_incarnation = 1;
770 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
772 INIT_LIST_HEAD(&osd->o_keepalive_item);
776 static struct ceph_osd *get_osd(struct ceph_osd *osd)
778 if (atomic_inc_not_zero(&osd->o_ref)) {
779 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
780 atomic_read(&osd->o_ref));
783 dout("get_osd %p FAIL\n", osd);
788 static void put_osd(struct ceph_osd *osd)
790 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
791 atomic_read(&osd->o_ref) - 1);
792 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
793 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
795 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
801 * remove an osd from our map
803 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
805 dout("__remove_osd %p\n", osd);
806 BUG_ON(!list_empty(&osd->o_requests));
807 rb_erase(&osd->o_node, &osdc->osds);
808 list_del_init(&osd->o_osd_lru);
809 ceph_con_close(&osd->o_con);
813 static void remove_all_osds(struct ceph_osd_client *osdc)
815 dout("%s %p\n", __func__, osdc);
816 mutex_lock(&osdc->request_mutex);
817 while (!RB_EMPTY_ROOT(&osdc->osds)) {
818 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
819 struct ceph_osd, o_node);
820 __remove_osd(osdc, osd);
822 mutex_unlock(&osdc->request_mutex);
825 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
826 struct ceph_osd *osd)
828 dout("__move_osd_to_lru %p\n", osd);
829 BUG_ON(!list_empty(&osd->o_osd_lru));
830 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
831 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
834 static void __remove_osd_from_lru(struct ceph_osd *osd)
836 dout("__remove_osd_from_lru %p\n", osd);
837 if (!list_empty(&osd->o_osd_lru))
838 list_del_init(&osd->o_osd_lru);
841 static void remove_old_osds(struct ceph_osd_client *osdc)
843 struct ceph_osd *osd, *nosd;
845 dout("__remove_old_osds %p\n", osdc);
846 mutex_lock(&osdc->request_mutex);
847 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
848 if (time_before(jiffies, osd->lru_ttl))
850 __remove_osd(osdc, osd);
852 mutex_unlock(&osdc->request_mutex);
858 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
860 struct ceph_entity_addr *peer_addr;
862 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
863 if (list_empty(&osd->o_requests) &&
864 list_empty(&osd->o_linger_requests)) {
865 __remove_osd(osdc, osd);
870 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
871 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
872 !ceph_con_opened(&osd->o_con)) {
873 struct ceph_osd_request *req;
875 dout(" osd addr hasn't changed and connection never opened,"
876 " letting msgr retry");
877 /* touch each r_stamp for handle_timeout()'s benfit */
878 list_for_each_entry(req, &osd->o_requests, r_osd_item)
879 req->r_stamp = jiffies;
884 ceph_con_close(&osd->o_con);
885 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
886 osd->o_incarnation++;
891 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
893 struct rb_node **p = &osdc->osds.rb_node;
894 struct rb_node *parent = NULL;
895 struct ceph_osd *osd = NULL;
897 dout("__insert_osd %p osd%d\n", new, new->o_osd);
900 osd = rb_entry(parent, struct ceph_osd, o_node);
901 if (new->o_osd < osd->o_osd)
903 else if (new->o_osd > osd->o_osd)
909 rb_link_node(&new->o_node, parent, p);
910 rb_insert_color(&new->o_node, &osdc->osds);
913 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
915 struct ceph_osd *osd;
916 struct rb_node *n = osdc->osds.rb_node;
919 osd = rb_entry(n, struct ceph_osd, o_node);
922 else if (o > osd->o_osd)
930 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
932 schedule_delayed_work(&osdc->timeout_work,
933 osdc->client->options->osd_keepalive_timeout * HZ);
936 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
938 cancel_delayed_work(&osdc->timeout_work);
942 * Register request, assign tid. If this is the first request, set up
945 static void __register_request(struct ceph_osd_client *osdc,
946 struct ceph_osd_request *req)
948 req->r_tid = ++osdc->last_tid;
949 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
950 dout("__register_request %p tid %lld\n", req, req->r_tid);
951 __insert_request(osdc, req);
952 ceph_osdc_get_request(req);
953 osdc->num_requests++;
954 if (osdc->num_requests == 1) {
955 dout(" first request, scheduling timeout\n");
956 __schedule_osd_timeout(osdc);
961 * called under osdc->request_mutex
963 static void __unregister_request(struct ceph_osd_client *osdc,
964 struct ceph_osd_request *req)
966 if (RB_EMPTY_NODE(&req->r_node)) {
967 dout("__unregister_request %p tid %lld not registered\n",
972 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
973 rb_erase(&req->r_node, &osdc->requests);
974 osdc->num_requests--;
977 /* make sure the original request isn't in flight. */
978 ceph_msg_revoke(req->r_request);
980 list_del_init(&req->r_osd_item);
981 if (list_empty(&req->r_osd->o_requests) &&
982 list_empty(&req->r_osd->o_linger_requests)) {
983 dout("moving osd to %p lru\n", req->r_osd);
984 __move_osd_to_lru(osdc, req->r_osd);
986 if (list_empty(&req->r_linger_item))
990 list_del_init(&req->r_req_lru_item);
991 ceph_osdc_put_request(req);
993 if (osdc->num_requests == 0) {
994 dout(" no requests, canceling timeout\n");
995 __cancel_osd_timeout(osdc);
1000 * Cancel a previously queued request message
1002 static void __cancel_request(struct ceph_osd_request *req)
1004 if (req->r_sent && req->r_osd) {
1005 ceph_msg_revoke(req->r_request);
1010 static void __register_linger_request(struct ceph_osd_client *osdc,
1011 struct ceph_osd_request *req)
1013 dout("__register_linger_request %p\n", req);
1014 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1016 list_add_tail(&req->r_linger_osd,
1017 &req->r_osd->o_linger_requests);
1020 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1021 struct ceph_osd_request *req)
1023 dout("__unregister_linger_request %p\n", req);
1024 list_del_init(&req->r_linger_item);
1026 list_del_init(&req->r_linger_osd);
1028 if (list_empty(&req->r_osd->o_requests) &&
1029 list_empty(&req->r_osd->o_linger_requests)) {
1030 dout("moving osd to %p lru\n", req->r_osd);
1031 __move_osd_to_lru(osdc, req->r_osd);
1033 if (list_empty(&req->r_osd_item))
1038 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1039 struct ceph_osd_request *req)
1041 mutex_lock(&osdc->request_mutex);
1042 if (req->r_linger) {
1043 __unregister_linger_request(osdc, req);
1044 ceph_osdc_put_request(req);
1046 mutex_unlock(&osdc->request_mutex);
1048 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1050 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1051 struct ceph_osd_request *req)
1053 if (!req->r_linger) {
1054 dout("set_request_linger %p\n", req);
1057 * caller is now responsible for calling
1058 * unregister_linger_request
1060 ceph_osdc_get_request(req);
1063 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1066 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1067 * (as needed), and set the request r_osd appropriately. If there is
1068 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1069 * (unsent, homeless) or leave on in-flight lru.
1071 * Return 0 if unchanged, 1 if changed, or negative on error.
1073 * Caller should hold map_sem for read and request_mutex.
1075 static int __map_request(struct ceph_osd_client *osdc,
1076 struct ceph_osd_request *req, int force_resend)
1078 struct ceph_pg pgid;
1079 int acting[CEPH_PG_MAX_SIZE];
1080 int o = -1, num = 0;
1083 dout("map_request %p tid %lld\n", req, req->r_tid);
1084 err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap,
1085 ceph_file_layout_pg_pool(req->r_file_layout));
1087 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1092 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
1098 if ((!force_resend &&
1099 req->r_osd && req->r_osd->o_osd == o &&
1100 req->r_sent >= req->r_osd->o_incarnation &&
1101 req->r_num_pg_osds == num &&
1102 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1103 (req->r_osd == NULL && o == -1))
1104 return 0; /* no change */
1106 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1107 req->r_tid, pgid.pool, pgid.seed, o,
1108 req->r_osd ? req->r_osd->o_osd : -1);
1110 /* record full pg acting set */
1111 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1112 req->r_num_pg_osds = num;
1115 __cancel_request(req);
1116 list_del_init(&req->r_osd_item);
1120 req->r_osd = __lookup_osd(osdc, o);
1121 if (!req->r_osd && o >= 0) {
1123 req->r_osd = create_osd(osdc, o);
1125 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1129 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1130 __insert_osd(osdc, req->r_osd);
1132 ceph_con_open(&req->r_osd->o_con,
1133 CEPH_ENTITY_TYPE_OSD, o,
1134 &osdc->osdmap->osd_addr[o]);
1138 __remove_osd_from_lru(req->r_osd);
1139 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1140 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1142 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1144 err = 1; /* osd or pg changed */
1151 * caller should hold map_sem (for read) and request_mutex
1153 static void __send_request(struct ceph_osd_client *osdc,
1154 struct ceph_osd_request *req)
1158 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1159 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1160 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1162 /* fill in message content that changes each time we send it */
1163 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1164 put_unaligned_le32(req->r_flags, req->r_request_flags);
1165 put_unaligned_le64(req->r_pgid.pool, req->r_request_pool);
1166 p = req->r_request_pgid;
1167 ceph_encode_64(&p, req->r_pgid.pool);
1168 ceph_encode_32(&p, req->r_pgid.seed);
1169 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1170 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1171 sizeof(req->r_reassert_version));
1173 req->r_stamp = jiffies;
1174 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1176 ceph_msg_get(req->r_request); /* send consumes a ref */
1177 ceph_con_send(&req->r_osd->o_con, req->r_request);
1178 req->r_sent = req->r_osd->o_incarnation;
1182 * Send any requests in the queue (req_unsent).
1184 static void __send_queued(struct ceph_osd_client *osdc)
1186 struct ceph_osd_request *req, *tmp;
1188 dout("__send_queued\n");
1189 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1190 __send_request(osdc, req);
1194 * Timeout callback, called every N seconds when 1 or more osd
1195 * requests has been active for more than N seconds. When this
1196 * happens, we ping all OSDs with requests who have timed out to
1197 * ensure any communications channel reset is detected. Reset the
1198 * request timeouts another N seconds in the future as we go.
1199 * Reschedule the timeout event another N seconds in future (unless
1200 * there are no open requests).
1202 static void handle_timeout(struct work_struct *work)
1204 struct ceph_osd_client *osdc =
1205 container_of(work, struct ceph_osd_client, timeout_work.work);
1206 struct ceph_osd_request *req;
1207 struct ceph_osd *osd;
1208 unsigned long keepalive =
1209 osdc->client->options->osd_keepalive_timeout * HZ;
1210 struct list_head slow_osds;
1212 down_read(&osdc->map_sem);
1214 ceph_monc_request_next_osdmap(&osdc->client->monc);
1216 mutex_lock(&osdc->request_mutex);
1219 * ping osds that are a bit slow. this ensures that if there
1220 * is a break in the TCP connection we will notice, and reopen
1221 * a connection with that osd (from the fault callback).
1223 INIT_LIST_HEAD(&slow_osds);
1224 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1225 if (time_before(jiffies, req->r_stamp + keepalive))
1230 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1231 req->r_tid, osd->o_osd);
1232 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1234 while (!list_empty(&slow_osds)) {
1235 osd = list_entry(slow_osds.next, struct ceph_osd,
1237 list_del_init(&osd->o_keepalive_item);
1238 ceph_con_keepalive(&osd->o_con);
1241 __schedule_osd_timeout(osdc);
1242 __send_queued(osdc);
1243 mutex_unlock(&osdc->request_mutex);
1244 up_read(&osdc->map_sem);
1247 static void handle_osds_timeout(struct work_struct *work)
1249 struct ceph_osd_client *osdc =
1250 container_of(work, struct ceph_osd_client,
1251 osds_timeout_work.work);
1252 unsigned long delay =
1253 osdc->client->options->osd_idle_ttl * HZ >> 2;
1255 dout("osds timeout\n");
1256 down_read(&osdc->map_sem);
1257 remove_old_osds(osdc);
1258 up_read(&osdc->map_sem);
1260 schedule_delayed_work(&osdc->osds_timeout_work,
1261 round_jiffies_relative(delay));
1264 static void complete_request(struct ceph_osd_request *req)
1266 if (req->r_safe_callback)
1267 req->r_safe_callback(req, NULL);
1268 complete_all(&req->r_safe_completion); /* fsync waiter */
1271 static int __decode_pgid(void **p, void *end, struct ceph_pg *pgid)
1275 ceph_decode_need(p, end, 1 + 8 + 4 + 4, bad);
1276 v = ceph_decode_8(p);
1278 pr_warning("do not understand pg encoding %d > 1", v);
1281 pgid->pool = ceph_decode_64(p);
1282 pgid->seed = ceph_decode_32(p);
1287 pr_warning("incomplete pg encoding");
1292 * handle osd op reply. either call the callback if it is specified,
1293 * or do the completion to wake up the waiting thread.
1295 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1296 struct ceph_connection *con)
1299 struct ceph_osd_request *req;
1302 int numops, payload_len, flags;
1308 u64 reassert_version;
1310 int already_completed;
1313 tid = le64_to_cpu(msg->hdr.tid);
1314 dout("handle_reply %p tid %llu\n", msg, tid);
1316 p = msg->front.iov_base;
1317 end = p + msg->front.iov_len;
1319 ceph_decode_need(&p, end, 4, bad);
1320 object_len = ceph_decode_32(&p);
1321 ceph_decode_need(&p, end, object_len, bad);
1324 err = __decode_pgid(&p, end, &pg);
1328 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1329 flags = ceph_decode_64(&p);
1330 result = ceph_decode_32(&p);
1331 reassert_epoch = ceph_decode_32(&p);
1332 reassert_version = ceph_decode_64(&p);
1333 osdmap_epoch = ceph_decode_32(&p);
1336 mutex_lock(&osdc->request_mutex);
1337 req = __lookup_request(osdc, tid);
1339 dout("handle_reply tid %llu dne\n", tid);
1340 mutex_unlock(&osdc->request_mutex);
1343 ceph_osdc_get_request(req);
1345 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1348 ceph_decode_need(&p, end, 4, bad);
1349 numops = ceph_decode_32(&p);
1350 if (numops > CEPH_OSD_MAX_OP)
1352 if (numops != req->r_num_ops)
1355 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad);
1356 for (i = 0; i < numops; i++) {
1357 struct ceph_osd_op *op = p;
1360 len = le32_to_cpu(op->payload_len);
1361 req->r_reply_op_len[i] = len;
1362 dout(" op %d has %d bytes\n", i, len);
1366 if (payload_len != le32_to_cpu(msg->hdr.data_len)) {
1367 pr_warning("sum of op payload lens %d != data_len %d",
1368 payload_len, le32_to_cpu(msg->hdr.data_len));
1372 ceph_decode_need(&p, end, 4 + numops * 4, bad);
1373 retry_attempt = ceph_decode_32(&p);
1374 for (i = 0; i < numops; i++)
1375 req->r_reply_op_result[i] = ceph_decode_32(&p);
1378 * if this connection filled our message, drop our reference now, to
1379 * avoid a (safe but slower) revoke later.
1381 if (req->r_con_filling_msg == con && req->r_reply == msg) {
1382 dout(" dropping con_filling_msg ref %p\n", con);
1383 req->r_con_filling_msg = NULL;
1387 if (!req->r_got_reply) {
1390 req->r_result = result;
1391 bytes = le32_to_cpu(msg->hdr.data_len);
1392 dout("handle_reply result %d bytes %d\n", req->r_result,
1394 if (req->r_result == 0)
1395 req->r_result = bytes;
1397 /* in case this is a write and we need to replay, */
1398 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1399 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1401 req->r_got_reply = 1;
1402 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1403 dout("handle_reply tid %llu dup ack\n", tid);
1404 mutex_unlock(&osdc->request_mutex);
1408 dout("handle_reply tid %llu flags %d\n", tid, flags);
1410 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1411 __register_linger_request(osdc, req);
1413 /* either this is a read, or we got the safe response */
1415 (flags & CEPH_OSD_FLAG_ONDISK) ||
1416 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1417 __unregister_request(osdc, req);
1419 already_completed = req->r_completed;
1420 req->r_completed = 1;
1421 mutex_unlock(&osdc->request_mutex);
1422 if (already_completed)
1425 if (req->r_callback)
1426 req->r_callback(req, msg);
1428 complete_all(&req->r_completion);
1430 if (flags & CEPH_OSD_FLAG_ONDISK)
1431 complete_request(req);
1434 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1435 ceph_osdc_put_request(req);
1439 ceph_osdc_put_request(req);
1441 pr_err("corrupt osd_op_reply got %d %d\n",
1442 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1446 static void reset_changed_osds(struct ceph_osd_client *osdc)
1448 struct rb_node *p, *n;
1450 for (p = rb_first(&osdc->osds); p; p = n) {
1451 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1454 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1455 memcmp(&osd->o_con.peer_addr,
1456 ceph_osd_addr(osdc->osdmap,
1458 sizeof(struct ceph_entity_addr)) != 0)
1459 __reset_osd(osdc, osd);
1464 * Requeue requests whose mapping to an OSD has changed. If requests map to
1465 * no osd, request a new map.
1467 * Caller should hold map_sem for read.
1469 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1471 struct ceph_osd_request *req, *nreq;
1476 dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1477 mutex_lock(&osdc->request_mutex);
1478 for (p = rb_first(&osdc->requests); p; ) {
1479 req = rb_entry(p, struct ceph_osd_request, r_node);
1483 * For linger requests that have not yet been
1484 * registered, move them to the linger list; they'll
1485 * be sent to the osd in the loop below. Unregister
1486 * the request before re-registering it as a linger
1487 * request to ensure the __map_request() below
1488 * will decide it needs to be sent.
1490 if (req->r_linger && list_empty(&req->r_linger_item)) {
1491 dout("%p tid %llu restart on osd%d\n",
1493 req->r_osd ? req->r_osd->o_osd : -1);
1494 __unregister_request(osdc, req);
1495 __register_linger_request(osdc, req);
1499 err = __map_request(osdc, req, force_resend);
1501 continue; /* error */
1502 if (req->r_osd == NULL) {
1503 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1504 needmap++; /* request a newer map */
1505 } else if (err > 0) {
1506 if (!req->r_linger) {
1507 dout("%p tid %llu requeued on osd%d\n", req,
1509 req->r_osd ? req->r_osd->o_osd : -1);
1510 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1515 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1517 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1519 err = __map_request(osdc, req, force_resend);
1520 dout("__map_request returned %d\n", err);
1522 continue; /* no change and no osd was specified */
1524 continue; /* hrm! */
1525 if (req->r_osd == NULL) {
1526 dout("tid %llu maps to no valid osd\n", req->r_tid);
1527 needmap++; /* request a newer map */
1531 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1532 req->r_osd ? req->r_osd->o_osd : -1);
1533 __register_request(osdc, req);
1534 __unregister_linger_request(osdc, req);
1536 mutex_unlock(&osdc->request_mutex);
1539 dout("%d requests for down osds, need new map\n", needmap);
1540 ceph_monc_request_next_osdmap(&osdc->client->monc);
1542 reset_changed_osds(osdc);
1547 * Process updated osd map.
1549 * The message contains any number of incremental and full maps, normally
1550 * indicating some sort of topology change in the cluster. Kick requests
1551 * off to different OSDs as needed.
1553 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1555 void *p, *end, *next;
1556 u32 nr_maps, maplen;
1558 struct ceph_osdmap *newmap = NULL, *oldmap;
1560 struct ceph_fsid fsid;
1562 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1563 p = msg->front.iov_base;
1564 end = p + msg->front.iov_len;
1567 ceph_decode_need(&p, end, sizeof(fsid), bad);
1568 ceph_decode_copy(&p, &fsid, sizeof(fsid));
1569 if (ceph_check_fsid(osdc->client, &fsid) < 0)
1572 down_write(&osdc->map_sem);
1574 /* incremental maps */
1575 ceph_decode_32_safe(&p, end, nr_maps, bad);
1576 dout(" %d inc maps\n", nr_maps);
1577 while (nr_maps > 0) {
1578 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1579 epoch = ceph_decode_32(&p);
1580 maplen = ceph_decode_32(&p);
1581 ceph_decode_need(&p, end, maplen, bad);
1583 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1584 dout("applying incremental map %u len %d\n",
1586 newmap = osdmap_apply_incremental(&p, next,
1588 &osdc->client->msgr);
1589 if (IS_ERR(newmap)) {
1590 err = PTR_ERR(newmap);
1594 if (newmap != osdc->osdmap) {
1595 ceph_osdmap_destroy(osdc->osdmap);
1596 osdc->osdmap = newmap;
1598 kick_requests(osdc, 0);
1600 dout("ignoring incremental map %u len %d\n",
1610 ceph_decode_32_safe(&p, end, nr_maps, bad);
1611 dout(" %d full maps\n", nr_maps);
1613 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1614 epoch = ceph_decode_32(&p);
1615 maplen = ceph_decode_32(&p);
1616 ceph_decode_need(&p, end, maplen, bad);
1618 dout("skipping non-latest full map %u len %d\n",
1620 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1621 dout("skipping full map %u len %d, "
1622 "older than our %u\n", epoch, maplen,
1623 osdc->osdmap->epoch);
1625 int skipped_map = 0;
1627 dout("taking full map %u len %d\n", epoch, maplen);
1628 newmap = osdmap_decode(&p, p+maplen);
1629 if (IS_ERR(newmap)) {
1630 err = PTR_ERR(newmap);
1634 oldmap = osdc->osdmap;
1635 osdc->osdmap = newmap;
1637 if (oldmap->epoch + 1 < newmap->epoch)
1639 ceph_osdmap_destroy(oldmap);
1641 kick_requests(osdc, skipped_map);
1648 downgrade_write(&osdc->map_sem);
1649 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1652 * subscribe to subsequent osdmap updates if full to ensure
1653 * we find out when we are no longer full and stop returning
1656 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1657 ceph_monc_request_next_osdmap(&osdc->client->monc);
1659 mutex_lock(&osdc->request_mutex);
1660 __send_queued(osdc);
1661 mutex_unlock(&osdc->request_mutex);
1662 up_read(&osdc->map_sem);
1663 wake_up_all(&osdc->client->auth_wq);
1667 pr_err("osdc handle_map corrupt msg\n");
1669 up_write(&osdc->map_sem);
1674 * watch/notify callback event infrastructure
1676 * These callbacks are used both for watch and notify operations.
1678 static void __release_event(struct kref *kref)
1680 struct ceph_osd_event *event =
1681 container_of(kref, struct ceph_osd_event, kref);
1683 dout("__release_event %p\n", event);
1687 static void get_event(struct ceph_osd_event *event)
1689 kref_get(&event->kref);
1692 void ceph_osdc_put_event(struct ceph_osd_event *event)
1694 kref_put(&event->kref, __release_event);
1696 EXPORT_SYMBOL(ceph_osdc_put_event);
1698 static void __insert_event(struct ceph_osd_client *osdc,
1699 struct ceph_osd_event *new)
1701 struct rb_node **p = &osdc->event_tree.rb_node;
1702 struct rb_node *parent = NULL;
1703 struct ceph_osd_event *event = NULL;
1707 event = rb_entry(parent, struct ceph_osd_event, node);
1708 if (new->cookie < event->cookie)
1710 else if (new->cookie > event->cookie)
1711 p = &(*p)->rb_right;
1716 rb_link_node(&new->node, parent, p);
1717 rb_insert_color(&new->node, &osdc->event_tree);
1720 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1723 struct rb_node **p = &osdc->event_tree.rb_node;
1724 struct rb_node *parent = NULL;
1725 struct ceph_osd_event *event = NULL;
1729 event = rb_entry(parent, struct ceph_osd_event, node);
1730 if (cookie < event->cookie)
1732 else if (cookie > event->cookie)
1733 p = &(*p)->rb_right;
1740 static void __remove_event(struct ceph_osd_event *event)
1742 struct ceph_osd_client *osdc = event->osdc;
1744 if (!RB_EMPTY_NODE(&event->node)) {
1745 dout("__remove_event removed %p\n", event);
1746 rb_erase(&event->node, &osdc->event_tree);
1747 ceph_osdc_put_event(event);
1749 dout("__remove_event didn't remove %p\n", event);
1753 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1754 void (*event_cb)(u64, u64, u8, void *),
1755 void *data, struct ceph_osd_event **pevent)
1757 struct ceph_osd_event *event;
1759 event = kmalloc(sizeof(*event), GFP_NOIO);
1763 dout("create_event %p\n", event);
1764 event->cb = event_cb;
1765 event->one_shot = 0;
1768 INIT_LIST_HEAD(&event->osd_node);
1769 RB_CLEAR_NODE(&event->node);
1770 kref_init(&event->kref); /* one ref for us */
1771 kref_get(&event->kref); /* one ref for the caller */
1773 spin_lock(&osdc->event_lock);
1774 event->cookie = ++osdc->event_count;
1775 __insert_event(osdc, event);
1776 spin_unlock(&osdc->event_lock);
1781 EXPORT_SYMBOL(ceph_osdc_create_event);
1783 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1785 struct ceph_osd_client *osdc = event->osdc;
1787 dout("cancel_event %p\n", event);
1788 spin_lock(&osdc->event_lock);
1789 __remove_event(event);
1790 spin_unlock(&osdc->event_lock);
1791 ceph_osdc_put_event(event); /* caller's */
1793 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1796 static void do_event_work(struct work_struct *work)
1798 struct ceph_osd_event_work *event_work =
1799 container_of(work, struct ceph_osd_event_work, work);
1800 struct ceph_osd_event *event = event_work->event;
1801 u64 ver = event_work->ver;
1802 u64 notify_id = event_work->notify_id;
1803 u8 opcode = event_work->opcode;
1805 dout("do_event_work completing %p\n", event);
1806 event->cb(ver, notify_id, opcode, event->data);
1807 dout("do_event_work completed %p\n", event);
1808 ceph_osdc_put_event(event);
1814 * Process osd watch notifications
1816 static void handle_watch_notify(struct ceph_osd_client *osdc,
1817 struct ceph_msg *msg)
1821 u64 cookie, ver, notify_id;
1823 struct ceph_osd_event *event;
1824 struct ceph_osd_event_work *event_work;
1826 p = msg->front.iov_base;
1827 end = p + msg->front.iov_len;
1829 ceph_decode_8_safe(&p, end, proto_ver, bad);
1830 ceph_decode_8_safe(&p, end, opcode, bad);
1831 ceph_decode_64_safe(&p, end, cookie, bad);
1832 ceph_decode_64_safe(&p, end, ver, bad);
1833 ceph_decode_64_safe(&p, end, notify_id, bad);
1835 spin_lock(&osdc->event_lock);
1836 event = __find_event(osdc, cookie);
1838 BUG_ON(event->one_shot);
1841 spin_unlock(&osdc->event_lock);
1842 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1843 cookie, ver, event);
1845 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1847 dout("ERROR: could not allocate event_work\n");
1850 INIT_WORK(&event_work->work, do_event_work);
1851 event_work->event = event;
1852 event_work->ver = ver;
1853 event_work->notify_id = notify_id;
1854 event_work->opcode = opcode;
1855 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1856 dout("WARNING: failed to queue notify event work\n");
1864 ceph_osdc_put_event(event);
1868 pr_err("osdc handle_watch_notify corrupt msg\n");
1872 static void ceph_osdc_msg_data_set(struct ceph_msg *msg,
1873 struct ceph_osd_data *osd_data)
1875 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
1876 BUG_ON(osd_data->length > (u64) SIZE_MAX);
1877 if (osd_data->length)
1878 ceph_msg_data_set_pages(msg, osd_data->pages,
1879 osd_data->length, osd_data->alignment);
1880 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
1881 BUG_ON(!osd_data->pagelist->length);
1882 ceph_msg_data_set_pagelist(msg, osd_data->pagelist);
1884 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
1885 ceph_msg_data_set_bio(msg, osd_data->bio);
1888 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
1893 * Register request, send initial attempt.
1895 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1896 struct ceph_osd_request *req,
1901 /* Set up response incoming data and request outgoing data fields */
1903 ceph_osdc_msg_data_set(req->r_reply, &req->r_data_in);
1904 ceph_osdc_msg_data_set(req->r_request, &req->r_data_out);
1906 down_read(&osdc->map_sem);
1907 mutex_lock(&osdc->request_mutex);
1908 __register_request(osdc, req);
1909 WARN_ON(req->r_sent);
1910 rc = __map_request(osdc, req, 0);
1913 dout("osdc_start_request failed map, "
1914 " will retry %lld\n", req->r_tid);
1919 if (req->r_osd == NULL) {
1920 dout("send_request %p no up osds in pg\n", req);
1921 ceph_monc_request_next_osdmap(&osdc->client->monc);
1923 __send_queued(osdc);
1927 mutex_unlock(&osdc->request_mutex);
1928 up_read(&osdc->map_sem);
1931 EXPORT_SYMBOL(ceph_osdc_start_request);
1934 * wait for a request to complete
1936 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1937 struct ceph_osd_request *req)
1941 rc = wait_for_completion_interruptible(&req->r_completion);
1943 mutex_lock(&osdc->request_mutex);
1944 __cancel_request(req);
1945 __unregister_request(osdc, req);
1946 mutex_unlock(&osdc->request_mutex);
1947 complete_request(req);
1948 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1952 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1953 return req->r_result;
1955 EXPORT_SYMBOL(ceph_osdc_wait_request);
1958 * sync - wait for all in-flight requests to flush. avoid starvation.
1960 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1962 struct ceph_osd_request *req;
1963 u64 last_tid, next_tid = 0;
1965 mutex_lock(&osdc->request_mutex);
1966 last_tid = osdc->last_tid;
1968 req = __lookup_request_ge(osdc, next_tid);
1971 if (req->r_tid > last_tid)
1974 next_tid = req->r_tid + 1;
1975 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1978 ceph_osdc_get_request(req);
1979 mutex_unlock(&osdc->request_mutex);
1980 dout("sync waiting on tid %llu (last is %llu)\n",
1981 req->r_tid, last_tid);
1982 wait_for_completion(&req->r_safe_completion);
1983 mutex_lock(&osdc->request_mutex);
1984 ceph_osdc_put_request(req);
1986 mutex_unlock(&osdc->request_mutex);
1987 dout("sync done (thru tid %llu)\n", last_tid);
1989 EXPORT_SYMBOL(ceph_osdc_sync);
1994 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1999 osdc->client = client;
2000 osdc->osdmap = NULL;
2001 init_rwsem(&osdc->map_sem);
2002 init_completion(&osdc->map_waiters);
2003 osdc->last_requested_map = 0;
2004 mutex_init(&osdc->request_mutex);
2006 osdc->osds = RB_ROOT;
2007 INIT_LIST_HEAD(&osdc->osd_lru);
2008 osdc->requests = RB_ROOT;
2009 INIT_LIST_HEAD(&osdc->req_lru);
2010 INIT_LIST_HEAD(&osdc->req_unsent);
2011 INIT_LIST_HEAD(&osdc->req_notarget);
2012 INIT_LIST_HEAD(&osdc->req_linger);
2013 osdc->num_requests = 0;
2014 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2015 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2016 spin_lock_init(&osdc->event_lock);
2017 osdc->event_tree = RB_ROOT;
2018 osdc->event_count = 0;
2020 schedule_delayed_work(&osdc->osds_timeout_work,
2021 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2024 osdc->req_mempool = mempool_create_kmalloc_pool(10,
2025 sizeof(struct ceph_osd_request));
2026 if (!osdc->req_mempool)
2029 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2030 OSD_OP_FRONT_LEN, 10, true,
2034 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2035 OSD_OPREPLY_FRONT_LEN, 10, true,
2040 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2041 if (IS_ERR(osdc->notify_wq)) {
2042 err = PTR_ERR(osdc->notify_wq);
2043 osdc->notify_wq = NULL;
2049 ceph_msgpool_destroy(&osdc->msgpool_op);
2051 mempool_destroy(osdc->req_mempool);
2056 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2058 flush_workqueue(osdc->notify_wq);
2059 destroy_workqueue(osdc->notify_wq);
2060 cancel_delayed_work_sync(&osdc->timeout_work);
2061 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2063 ceph_osdmap_destroy(osdc->osdmap);
2064 osdc->osdmap = NULL;
2066 remove_all_osds(osdc);
2067 mempool_destroy(osdc->req_mempool);
2068 ceph_msgpool_destroy(&osdc->msgpool_op);
2069 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2073 * Read some contiguous pages. If we cross a stripe boundary, shorten
2074 * *plen. Return number of bytes read, or error.
2076 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2077 struct ceph_vino vino, struct ceph_file_layout *layout,
2079 u32 truncate_seq, u64 truncate_size,
2080 struct page **pages, int num_pages, int page_align)
2082 struct ceph_osd_request *req;
2083 struct ceph_osd_data *osd_data;
2086 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2087 vino.snap, off, *plen);
2088 req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
2089 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2090 NULL, 0, truncate_seq, truncate_size, NULL,
2093 return PTR_ERR(req);
2095 /* it may be a short read due to an object boundary */
2097 osd_data = &req->r_data_in;
2098 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
2099 osd_data->pages = pages;
2100 osd_data->length = *plen;
2101 osd_data->alignment = page_align;
2103 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2104 off, *plen, osd_data->length, page_align);
2106 rc = ceph_osdc_start_request(osdc, req, false);
2108 rc = ceph_osdc_wait_request(osdc, req);
2110 ceph_osdc_put_request(req);
2111 dout("readpages result %d\n", rc);
2114 EXPORT_SYMBOL(ceph_osdc_readpages);
2117 * do a synchronous write on N pages
2119 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2120 struct ceph_file_layout *layout,
2121 struct ceph_snap_context *snapc,
2123 u32 truncate_seq, u64 truncate_size,
2124 struct timespec *mtime,
2125 struct page **pages, int num_pages)
2127 struct ceph_osd_request *req;
2128 struct ceph_osd_data *osd_data;
2130 int page_align = off & ~PAGE_MASK;
2132 BUG_ON(vino.snap != CEPH_NOSNAP);
2133 req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
2135 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2137 truncate_seq, truncate_size, mtime,
2140 return PTR_ERR(req);
2142 /* it may be a short write due to an object boundary */
2143 osd_data = &req->r_data_out;
2144 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
2145 osd_data->pages = pages;
2146 osd_data->length = len;
2147 osd_data->alignment = page_align;
2148 dout("writepages %llu~%llu (%llu bytes)\n", off, len, osd_data->length);
2150 rc = ceph_osdc_start_request(osdc, req, true);
2152 rc = ceph_osdc_wait_request(osdc, req);
2154 ceph_osdc_put_request(req);
2157 dout("writepages result %d\n", rc);
2160 EXPORT_SYMBOL(ceph_osdc_writepages);
2163 * handle incoming message
2165 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2167 struct ceph_osd *osd = con->private;
2168 struct ceph_osd_client *osdc;
2169 int type = le16_to_cpu(msg->hdr.type);
2176 case CEPH_MSG_OSD_MAP:
2177 ceph_osdc_handle_map(osdc, msg);
2179 case CEPH_MSG_OSD_OPREPLY:
2180 handle_reply(osdc, msg, con);
2182 case CEPH_MSG_WATCH_NOTIFY:
2183 handle_watch_notify(osdc, msg);
2187 pr_err("received unknown message type %d %s\n", type,
2188 ceph_msg_type_name(type));
2195 * lookup and return message for incoming reply. set up reply message
2198 static struct ceph_msg *get_reply(struct ceph_connection *con,
2199 struct ceph_msg_header *hdr,
2202 struct ceph_osd *osd = con->private;
2203 struct ceph_osd_client *osdc = osd->o_osdc;
2205 struct ceph_osd_request *req;
2206 int front = le32_to_cpu(hdr->front_len);
2207 int data_len = le32_to_cpu(hdr->data_len);
2210 tid = le64_to_cpu(hdr->tid);
2211 mutex_lock(&osdc->request_mutex);
2212 req = __lookup_request(osdc, tid);
2216 dout("get_reply unknown tid %llu from osd%d\n", tid,
2221 if (req->r_con_filling_msg) {
2222 dout("%s revoking msg %p from old con %p\n", __func__,
2223 req->r_reply, req->r_con_filling_msg);
2224 ceph_msg_revoke_incoming(req->r_reply);
2225 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2226 req->r_con_filling_msg = NULL;
2229 if (front > req->r_reply->front.iov_len) {
2230 pr_warning("get_reply front %d > preallocated %d\n",
2231 front, (int)req->r_reply->front.iov_len);
2232 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2235 ceph_msg_put(req->r_reply);
2238 m = ceph_msg_get(req->r_reply);
2241 struct ceph_osd_data *osd_data = &req->r_data_in;
2243 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2244 if (osd_data->pages &&
2245 unlikely(osd_data->length < data_len)) {
2247 pr_warning("tid %lld reply has %d bytes "
2248 "we had only %llu bytes ready\n",
2249 tid, data_len, osd_data->length);
2258 req->r_con_filling_msg = con->ops->get(con);
2259 dout("get_reply tid %lld %p\n", tid, m);
2262 mutex_unlock(&osdc->request_mutex);
2267 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2268 struct ceph_msg_header *hdr,
2271 struct ceph_osd *osd = con->private;
2272 int type = le16_to_cpu(hdr->type);
2273 int front = le32_to_cpu(hdr->front_len);
2277 case CEPH_MSG_OSD_MAP:
2278 case CEPH_MSG_WATCH_NOTIFY:
2279 return ceph_msg_new(type, front, GFP_NOFS, false);
2280 case CEPH_MSG_OSD_OPREPLY:
2281 return get_reply(con, hdr, skip);
2283 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2291 * Wrappers to refcount containing ceph_osd struct
2293 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2295 struct ceph_osd *osd = con->private;
2301 static void put_osd_con(struct ceph_connection *con)
2303 struct ceph_osd *osd = con->private;
2311 * Note: returned pointer is the address of a structure that's
2312 * managed separately. Caller must *not* attempt to free it.
2314 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2315 int *proto, int force_new)
2317 struct ceph_osd *o = con->private;
2318 struct ceph_osd_client *osdc = o->o_osdc;
2319 struct ceph_auth_client *ac = osdc->client->monc.auth;
2320 struct ceph_auth_handshake *auth = &o->o_auth;
2322 if (force_new && auth->authorizer) {
2323 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2324 auth->authorizer = NULL;
2326 if (!auth->authorizer) {
2327 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2330 return ERR_PTR(ret);
2332 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2335 return ERR_PTR(ret);
2337 *proto = ac->protocol;
2343 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2345 struct ceph_osd *o = con->private;
2346 struct ceph_osd_client *osdc = o->o_osdc;
2347 struct ceph_auth_client *ac = osdc->client->monc.auth;
2349 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2352 static int invalidate_authorizer(struct ceph_connection *con)
2354 struct ceph_osd *o = con->private;
2355 struct ceph_osd_client *osdc = o->o_osdc;
2356 struct ceph_auth_client *ac = osdc->client->monc.auth;
2358 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2359 return ceph_monc_validate_auth(&osdc->client->monc);
2362 static const struct ceph_connection_operations osd_con_ops = {
2365 .dispatch = dispatch,
2366 .get_authorizer = get_authorizer,
2367 .verify_authorizer_reply = verify_authorizer_reply,
2368 .invalidate_authorizer = invalidate_authorizer,
2369 .alloc_msg = alloc_msg,