2 #include <linux/ceph/ceph_debug.h>
4 #include <linux/module.h>
6 #include <linux/highmem.h>
8 #include <linux/pagemap.h>
9 #include <linux/slab.h>
10 #include <linux/uaccess.h>
12 #include <linux/bio.h>
15 #include <linux/ceph/libceph.h>
16 #include <linux/ceph/osd_client.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/pagelist.h>
22 #define OSD_OP_FRONT_LEN 4096
23 #define OSD_OPREPLY_FRONT_LEN 512
25 static struct kmem_cache *ceph_osd_request_cache;
27 static const struct ceph_connection_operations osd_con_ops;
29 static void __send_queued(struct ceph_osd_client *osdc);
30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
31 static void __register_request(struct ceph_osd_client *osdc,
32 struct ceph_osd_request *req);
33 static void __unregister_linger_request(struct ceph_osd_client *osdc,
34 struct ceph_osd_request *req);
35 static void __send_request(struct ceph_osd_client *osdc,
36 struct ceph_osd_request *req);
39 * Implement client access to distributed object storage cluster.
41 * All data objects are stored within a cluster/cloud of OSDs, or
42 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
43 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
44 * remote daemons serving up and coordinating consistent and safe
47 * Cluster membership and the mapping of data objects onto storage devices
48 * are described by the osd map.
50 * We keep track of pending OSD requests (read, write), resubmit
51 * requests to different OSDs when the cluster topology/data layout
52 * change, or retry the affected requests when the communications
53 * channel with an OSD is reset.
57 * calculate the mapping of a file extent onto an object, and fill out the
58 * request accordingly. shorten extent as necessary if it crosses an
61 * fill osd op in request message.
63 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
64 u64 *objnum, u64 *objoff, u64 *objlen)
70 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
74 if (*objlen < orig_len) {
76 dout(" skipping last %llu, final file extent %llu~%llu\n",
77 orig_len - *plen, off, *plen);
80 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
85 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
87 memset(osd_data, 0, sizeof (*osd_data));
88 osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
91 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
92 struct page **pages, u64 length, u32 alignment,
93 bool pages_from_pool, bool own_pages)
95 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
96 osd_data->pages = pages;
97 osd_data->length = length;
98 osd_data->alignment = alignment;
99 osd_data->pages_from_pool = pages_from_pool;
100 osd_data->own_pages = own_pages;
103 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
104 struct ceph_pagelist *pagelist)
106 osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
107 osd_data->pagelist = pagelist;
111 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
112 struct bio *bio, size_t bio_length)
114 osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
116 osd_data->bio_length = bio_length;
118 #endif /* CONFIG_BLOCK */
120 #define osd_req_op_data(oreq, whch, typ, fld) \
122 BUG_ON(whch >= (oreq)->r_num_ops); \
123 &(oreq)->r_ops[whch].typ.fld; \
126 static struct ceph_osd_data *
127 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
129 BUG_ON(which >= osd_req->r_num_ops);
131 return &osd_req->r_ops[which].raw_data_in;
134 struct ceph_osd_data *
135 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
138 return osd_req_op_data(osd_req, which, extent, osd_data);
140 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
142 struct ceph_osd_data *
143 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
146 return osd_req_op_data(osd_req, which, cls, response_data);
148 EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */
150 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
151 unsigned int which, struct page **pages,
152 u64 length, u32 alignment,
153 bool pages_from_pool, bool own_pages)
155 struct ceph_osd_data *osd_data;
157 osd_data = osd_req_op_raw_data_in(osd_req, which);
158 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
159 pages_from_pool, own_pages);
161 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
163 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
164 unsigned int which, struct page **pages,
165 u64 length, u32 alignment,
166 bool pages_from_pool, bool own_pages)
168 struct ceph_osd_data *osd_data;
170 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
171 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
172 pages_from_pool, own_pages);
174 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
176 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
177 unsigned int which, struct ceph_pagelist *pagelist)
179 struct ceph_osd_data *osd_data;
181 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
182 ceph_osd_data_pagelist_init(osd_data, pagelist);
184 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
187 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
188 unsigned int which, struct bio *bio, size_t bio_length)
190 struct ceph_osd_data *osd_data;
192 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
193 ceph_osd_data_bio_init(osd_data, bio, bio_length);
195 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
196 #endif /* CONFIG_BLOCK */
198 static void osd_req_op_cls_request_info_pagelist(
199 struct ceph_osd_request *osd_req,
200 unsigned int which, struct ceph_pagelist *pagelist)
202 struct ceph_osd_data *osd_data;
204 osd_data = osd_req_op_data(osd_req, which, cls, request_info);
205 ceph_osd_data_pagelist_init(osd_data, pagelist);
208 void osd_req_op_cls_request_data_pagelist(
209 struct ceph_osd_request *osd_req,
210 unsigned int which, struct ceph_pagelist *pagelist)
212 struct ceph_osd_data *osd_data;
214 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
215 ceph_osd_data_pagelist_init(osd_data, pagelist);
217 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
219 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
220 unsigned int which, struct page **pages, u64 length,
221 u32 alignment, bool pages_from_pool, bool own_pages)
223 struct ceph_osd_data *osd_data;
225 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
226 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
227 pages_from_pool, own_pages);
229 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
231 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
232 unsigned int which, struct page **pages, u64 length,
233 u32 alignment, bool pages_from_pool, bool own_pages)
235 struct ceph_osd_data *osd_data;
237 osd_data = osd_req_op_data(osd_req, which, cls, response_data);
238 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
239 pages_from_pool, own_pages);
241 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
243 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
245 switch (osd_data->type) {
246 case CEPH_OSD_DATA_TYPE_NONE:
248 case CEPH_OSD_DATA_TYPE_PAGES:
249 return osd_data->length;
250 case CEPH_OSD_DATA_TYPE_PAGELIST:
251 return (u64)osd_data->pagelist->length;
253 case CEPH_OSD_DATA_TYPE_BIO:
254 return (u64)osd_data->bio_length;
255 #endif /* CONFIG_BLOCK */
257 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
262 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
264 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
267 num_pages = calc_pages_for((u64)osd_data->alignment,
268 (u64)osd_data->length);
269 ceph_release_page_vector(osd_data->pages, num_pages);
271 ceph_osd_data_init(osd_data);
274 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
277 struct ceph_osd_req_op *op;
279 BUG_ON(which >= osd_req->r_num_ops);
280 op = &osd_req->r_ops[which];
283 case CEPH_OSD_OP_READ:
284 case CEPH_OSD_OP_WRITE:
285 ceph_osd_data_release(&op->extent.osd_data);
287 case CEPH_OSD_OP_CALL:
288 ceph_osd_data_release(&op->cls.request_info);
289 ceph_osd_data_release(&op->cls.request_data);
290 ceph_osd_data_release(&op->cls.response_data);
300 static void ceph_osdc_release_request(struct kref *kref)
302 struct ceph_osd_request *req = container_of(kref,
303 struct ceph_osd_request, r_kref);
306 dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
307 req->r_request, req->r_reply);
308 WARN_ON(!RB_EMPTY_NODE(&req->r_node));
309 WARN_ON(!list_empty(&req->r_req_lru_item));
310 WARN_ON(!list_empty(&req->r_osd_item));
311 WARN_ON(!list_empty(&req->r_linger_item));
312 WARN_ON(!list_empty(&req->r_linger_osd_item));
316 ceph_msg_put(req->r_request);
318 ceph_msg_revoke_incoming(req->r_reply);
319 ceph_msg_put(req->r_reply);
322 for (which = 0; which < req->r_num_ops; which++)
323 osd_req_op_data_release(req, which);
325 ceph_put_snap_context(req->r_snapc);
327 mempool_free(req, req->r_osdc->req_mempool);
329 kmem_cache_free(ceph_osd_request_cache, req);
333 void ceph_osdc_get_request(struct ceph_osd_request *req)
335 dout("%s %p (was %d)\n", __func__, req,
336 atomic_read(&req->r_kref.refcount));
337 kref_get(&req->r_kref);
339 EXPORT_SYMBOL(ceph_osdc_get_request);
341 void ceph_osdc_put_request(struct ceph_osd_request *req)
343 dout("%s %p (was %d)\n", __func__, req,
344 atomic_read(&req->r_kref.refcount));
345 kref_put(&req->r_kref, ceph_osdc_release_request);
347 EXPORT_SYMBOL(ceph_osdc_put_request);
349 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
350 struct ceph_snap_context *snapc,
351 unsigned int num_ops,
355 struct ceph_osd_request *req;
356 struct ceph_msg *msg;
359 BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
360 BUG_ON(num_ops > CEPH_OSD_MAX_OP);
362 msg_size = 4 + 4 + 8 + 8 + 4+8;
363 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
364 msg_size += 1 + 8 + 4 + 4; /* pg_t */
365 msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
366 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
367 msg_size += 8; /* snapid */
368 msg_size += 8; /* snap_seq */
369 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
373 req = mempool_alloc(osdc->req_mempool, gfp_flags);
374 memset(req, 0, sizeof(*req));
376 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
382 req->r_mempool = use_mempool;
383 req->r_num_ops = num_ops;
385 kref_init(&req->r_kref);
386 init_completion(&req->r_completion);
387 init_completion(&req->r_safe_completion);
388 RB_CLEAR_NODE(&req->r_node);
389 INIT_LIST_HEAD(&req->r_unsafe_item);
390 INIT_LIST_HEAD(&req->r_linger_item);
391 INIT_LIST_HEAD(&req->r_linger_osd_item);
392 INIT_LIST_HEAD(&req->r_req_lru_item);
393 INIT_LIST_HEAD(&req->r_osd_item);
395 req->r_base_oloc.pool = -1;
396 req->r_target_oloc.pool = -1;
398 /* create reply message */
400 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
402 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
403 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
405 ceph_osdc_put_request(req);
410 /* create request message; allow space for oid */
412 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
414 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
416 ceph_osdc_put_request(req);
420 memset(msg->front.iov_base, 0, msg->front.iov_len);
422 req->r_request = msg;
426 EXPORT_SYMBOL(ceph_osdc_alloc_request);
428 static bool osd_req_opcode_valid(u16 opcode)
431 case CEPH_OSD_OP_READ:
432 case CEPH_OSD_OP_STAT:
433 case CEPH_OSD_OP_MAPEXT:
434 case CEPH_OSD_OP_MASKTRUNC:
435 case CEPH_OSD_OP_SPARSE_READ:
436 case CEPH_OSD_OP_NOTIFY:
437 case CEPH_OSD_OP_NOTIFY_ACK:
438 case CEPH_OSD_OP_ASSERT_VER:
439 case CEPH_OSD_OP_WRITE:
440 case CEPH_OSD_OP_WRITEFULL:
441 case CEPH_OSD_OP_TRUNCATE:
442 case CEPH_OSD_OP_ZERO:
443 case CEPH_OSD_OP_DELETE:
444 case CEPH_OSD_OP_APPEND:
445 case CEPH_OSD_OP_STARTSYNC:
446 case CEPH_OSD_OP_SETTRUNC:
447 case CEPH_OSD_OP_TRIMTRUNC:
448 case CEPH_OSD_OP_TMAPUP:
449 case CEPH_OSD_OP_TMAPPUT:
450 case CEPH_OSD_OP_TMAPGET:
451 case CEPH_OSD_OP_CREATE:
452 case CEPH_OSD_OP_ROLLBACK:
453 case CEPH_OSD_OP_WATCH:
454 case CEPH_OSD_OP_OMAPGETKEYS:
455 case CEPH_OSD_OP_OMAPGETVALS:
456 case CEPH_OSD_OP_OMAPGETHEADER:
457 case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
458 case CEPH_OSD_OP_OMAPSETVALS:
459 case CEPH_OSD_OP_OMAPSETHEADER:
460 case CEPH_OSD_OP_OMAPCLEAR:
461 case CEPH_OSD_OP_OMAPRMKEYS:
462 case CEPH_OSD_OP_OMAP_CMP:
463 case CEPH_OSD_OP_SETALLOCHINT:
464 case CEPH_OSD_OP_CLONERANGE:
465 case CEPH_OSD_OP_ASSERT_SRC_VERSION:
466 case CEPH_OSD_OP_SRC_CMPXATTR:
467 case CEPH_OSD_OP_GETXATTR:
468 case CEPH_OSD_OP_GETXATTRS:
469 case CEPH_OSD_OP_CMPXATTR:
470 case CEPH_OSD_OP_SETXATTR:
471 case CEPH_OSD_OP_SETXATTRS:
472 case CEPH_OSD_OP_RESETXATTRS:
473 case CEPH_OSD_OP_RMXATTR:
474 case CEPH_OSD_OP_PULL:
475 case CEPH_OSD_OP_PUSH:
476 case CEPH_OSD_OP_BALANCEREADS:
477 case CEPH_OSD_OP_UNBALANCEREADS:
478 case CEPH_OSD_OP_SCRUB:
479 case CEPH_OSD_OP_SCRUB_RESERVE:
480 case CEPH_OSD_OP_SCRUB_UNRESERVE:
481 case CEPH_OSD_OP_SCRUB_STOP:
482 case CEPH_OSD_OP_SCRUB_MAP:
483 case CEPH_OSD_OP_WRLOCK:
484 case CEPH_OSD_OP_WRUNLOCK:
485 case CEPH_OSD_OP_RDLOCK:
486 case CEPH_OSD_OP_RDUNLOCK:
487 case CEPH_OSD_OP_UPLOCK:
488 case CEPH_OSD_OP_DNLOCK:
489 case CEPH_OSD_OP_CALL:
490 case CEPH_OSD_OP_PGLS:
491 case CEPH_OSD_OP_PGLS_FILTER:
499 * This is an osd op init function for opcodes that have no data or
500 * other information associated with them. It also serves as a
501 * common init routine for all the other init functions, below.
503 static struct ceph_osd_req_op *
504 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
507 struct ceph_osd_req_op *op;
509 BUG_ON(which >= osd_req->r_num_ops);
510 BUG_ON(!osd_req_opcode_valid(opcode));
512 op = &osd_req->r_ops[which];
513 memset(op, 0, sizeof (*op));
519 void osd_req_op_init(struct ceph_osd_request *osd_req,
520 unsigned int which, u16 opcode)
522 (void)_osd_req_op_init(osd_req, which, opcode);
524 EXPORT_SYMBOL(osd_req_op_init);
526 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
527 unsigned int which, u16 opcode,
528 u64 offset, u64 length,
529 u64 truncate_size, u32 truncate_seq)
531 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
532 size_t payload_len = 0;
534 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
535 opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
536 opcode != CEPH_OSD_OP_TRUNCATE);
538 op->extent.offset = offset;
539 op->extent.length = length;
540 op->extent.truncate_size = truncate_size;
541 op->extent.truncate_seq = truncate_seq;
542 if (opcode == CEPH_OSD_OP_WRITE)
543 payload_len += length;
545 op->payload_len = payload_len;
547 EXPORT_SYMBOL(osd_req_op_extent_init);
549 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
550 unsigned int which, u64 length)
552 struct ceph_osd_req_op *op;
555 BUG_ON(which >= osd_req->r_num_ops);
556 op = &osd_req->r_ops[which];
557 previous = op->extent.length;
559 if (length == previous)
560 return; /* Nothing to do */
561 BUG_ON(length > previous);
563 op->extent.length = length;
564 op->payload_len -= previous - length;
566 EXPORT_SYMBOL(osd_req_op_extent_update);
568 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
569 u16 opcode, const char *class, const char *method)
571 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
572 struct ceph_pagelist *pagelist;
573 size_t payload_len = 0;
576 BUG_ON(opcode != CEPH_OSD_OP_CALL);
578 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
580 ceph_pagelist_init(pagelist);
582 op->cls.class_name = class;
583 size = strlen(class);
584 BUG_ON(size > (size_t) U8_MAX);
585 op->cls.class_len = size;
586 ceph_pagelist_append(pagelist, class, size);
589 op->cls.method_name = method;
590 size = strlen(method);
591 BUG_ON(size > (size_t) U8_MAX);
592 op->cls.method_len = size;
593 ceph_pagelist_append(pagelist, method, size);
596 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
598 op->cls.argc = 0; /* currently unused */
600 op->payload_len = payload_len;
602 EXPORT_SYMBOL(osd_req_op_cls_init);
604 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
605 unsigned int which, u16 opcode,
606 u64 cookie, u64 version, int flag)
608 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
610 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
612 op->watch.cookie = cookie;
613 op->watch.ver = version;
614 if (opcode == CEPH_OSD_OP_WATCH && flag)
615 op->watch.flag = (u8)1;
617 EXPORT_SYMBOL(osd_req_op_watch_init);
619 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
621 u64 expected_object_size,
622 u64 expected_write_size)
624 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
625 CEPH_OSD_OP_SETALLOCHINT);
627 op->alloc_hint.expected_object_size = expected_object_size;
628 op->alloc_hint.expected_write_size = expected_write_size;
631 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
632 * not worth a feature bit. Set FAILOK per-op flag to make
633 * sure older osds don't trip over an unsupported opcode.
635 op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
637 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
639 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
640 struct ceph_osd_data *osd_data)
642 u64 length = ceph_osd_data_length(osd_data);
644 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
645 BUG_ON(length > (u64) SIZE_MAX);
647 ceph_msg_data_add_pages(msg, osd_data->pages,
648 length, osd_data->alignment);
649 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
651 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
653 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
654 ceph_msg_data_add_bio(msg, osd_data->bio, length);
657 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
661 static u64 osd_req_encode_op(struct ceph_osd_request *req,
662 struct ceph_osd_op *dst, unsigned int which)
664 struct ceph_osd_req_op *src;
665 struct ceph_osd_data *osd_data;
666 u64 request_data_len = 0;
669 BUG_ON(which >= req->r_num_ops);
670 src = &req->r_ops[which];
671 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
672 pr_err("unrecognized osd opcode %d\n", src->op);
678 case CEPH_OSD_OP_STAT:
679 osd_data = &src->raw_data_in;
680 ceph_osdc_msg_data_add(req->r_reply, osd_data);
682 case CEPH_OSD_OP_READ:
683 case CEPH_OSD_OP_WRITE:
684 case CEPH_OSD_OP_ZERO:
685 case CEPH_OSD_OP_DELETE:
686 case CEPH_OSD_OP_TRUNCATE:
687 if (src->op == CEPH_OSD_OP_WRITE)
688 request_data_len = src->extent.length;
689 dst->extent.offset = cpu_to_le64(src->extent.offset);
690 dst->extent.length = cpu_to_le64(src->extent.length);
691 dst->extent.truncate_size =
692 cpu_to_le64(src->extent.truncate_size);
693 dst->extent.truncate_seq =
694 cpu_to_le32(src->extent.truncate_seq);
695 osd_data = &src->extent.osd_data;
696 if (src->op == CEPH_OSD_OP_WRITE)
697 ceph_osdc_msg_data_add(req->r_request, osd_data);
699 ceph_osdc_msg_data_add(req->r_reply, osd_data);
701 case CEPH_OSD_OP_CALL:
702 dst->cls.class_len = src->cls.class_len;
703 dst->cls.method_len = src->cls.method_len;
704 osd_data = &src->cls.request_info;
705 ceph_osdc_msg_data_add(req->r_request, osd_data);
706 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
707 request_data_len = osd_data->pagelist->length;
709 osd_data = &src->cls.request_data;
710 data_length = ceph_osd_data_length(osd_data);
712 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
713 dst->cls.indata_len = cpu_to_le32(data_length);
714 ceph_osdc_msg_data_add(req->r_request, osd_data);
715 src->payload_len += data_length;
716 request_data_len += data_length;
718 osd_data = &src->cls.response_data;
719 ceph_osdc_msg_data_add(req->r_reply, osd_data);
721 case CEPH_OSD_OP_STARTSYNC:
723 case CEPH_OSD_OP_NOTIFY_ACK:
724 case CEPH_OSD_OP_WATCH:
725 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
726 dst->watch.ver = cpu_to_le64(src->watch.ver);
727 dst->watch.flag = src->watch.flag;
729 case CEPH_OSD_OP_SETALLOCHINT:
730 dst->alloc_hint.expected_object_size =
731 cpu_to_le64(src->alloc_hint.expected_object_size);
732 dst->alloc_hint.expected_write_size =
733 cpu_to_le64(src->alloc_hint.expected_write_size);
736 pr_err("unsupported osd opcode %s\n",
737 ceph_osd_op_name(src->op));
743 dst->op = cpu_to_le16(src->op);
744 dst->flags = cpu_to_le32(src->flags);
745 dst->payload_len = cpu_to_le32(src->payload_len);
747 return request_data_len;
751 * build new request AND message, calculate layout, and adjust file
754 * if the file was recently truncated, we include information about its
755 * old and new size so that the object can be updated appropriately. (we
756 * avoid synchronously deleting truncated objects because it's slow.)
758 * if @do_sync, include a 'startsync' command so that the osd will flush
761 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
762 struct ceph_file_layout *layout,
763 struct ceph_vino vino,
764 u64 off, u64 *plen, int num_ops,
765 int opcode, int flags,
766 struct ceph_snap_context *snapc,
771 struct ceph_osd_request *req;
779 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
780 opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
781 opcode != CEPH_OSD_OP_TRUNCATE);
783 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
786 return ERR_PTR(-ENOMEM);
788 req->r_flags = flags;
790 /* calculate max write size */
791 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
793 ceph_osdc_put_request(req);
797 object_size = le32_to_cpu(layout->fl_object_size);
798 object_base = off - objoff;
799 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
800 if (truncate_size <= object_base) {
803 truncate_size -= object_base;
804 if (truncate_size > object_size)
805 truncate_size = object_size;
809 osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
810 truncate_size, truncate_seq);
813 * A second op in the ops array means the caller wants to
814 * also issue a include a 'startsync' command so that the
815 * osd will flush data quickly.
818 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
820 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
822 snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
823 "%llx.%08llx", vino.ino, objnum);
824 req->r_base_oid.name_len = strlen(req->r_base_oid.name);
828 EXPORT_SYMBOL(ceph_osdc_new_request);
831 * We keep osd requests in an rbtree, sorted by ->r_tid.
833 static void __insert_request(struct ceph_osd_client *osdc,
834 struct ceph_osd_request *new)
836 struct rb_node **p = &osdc->requests.rb_node;
837 struct rb_node *parent = NULL;
838 struct ceph_osd_request *req = NULL;
842 req = rb_entry(parent, struct ceph_osd_request, r_node);
843 if (new->r_tid < req->r_tid)
845 else if (new->r_tid > req->r_tid)
851 rb_link_node(&new->r_node, parent, p);
852 rb_insert_color(&new->r_node, &osdc->requests);
855 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
858 struct ceph_osd_request *req;
859 struct rb_node *n = osdc->requests.rb_node;
862 req = rb_entry(n, struct ceph_osd_request, r_node);
863 if (tid < req->r_tid)
865 else if (tid > req->r_tid)
873 static struct ceph_osd_request *
874 __lookup_request_ge(struct ceph_osd_client *osdc,
877 struct ceph_osd_request *req;
878 struct rb_node *n = osdc->requests.rb_node;
881 req = rb_entry(n, struct ceph_osd_request, r_node);
882 if (tid < req->r_tid) {
886 } else if (tid > req->r_tid) {
896 * Resubmit requests pending on the given osd.
898 static void __kick_osd_requests(struct ceph_osd_client *osdc,
899 struct ceph_osd *osd)
901 struct ceph_osd_request *req, *nreq;
905 dout("__kick_osd_requests osd%d\n", osd->o_osd);
906 err = __reset_osd(osdc, osd);
910 * Build up a list of requests to resend by traversing the
911 * osd's list of requests. Requests for a given object are
912 * sent in tid order, and that is also the order they're
913 * kept on this list. Therefore all requests that are in
914 * flight will be found first, followed by all requests that
915 * have not yet been sent. And to resend requests while
916 * preserving this order we will want to put any sent
917 * requests back on the front of the osd client's unsent
920 * So we build a separate ordered list of already-sent
921 * requests for the affected osd and splice it onto the
922 * front of the osd client's unsent list. Once we've seen a
923 * request that has not yet been sent we're done. Those
924 * requests are already sitting right where they belong.
926 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
929 list_move_tail(&req->r_req_lru_item, &resend);
930 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
933 req->r_flags |= CEPH_OSD_FLAG_RETRY;
935 list_splice(&resend, &osdc->req_unsent);
938 * Linger requests are re-registered before sending, which
939 * sets up a new tid for each. We add them to the unsent
940 * list at the end to keep things in tid order.
942 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
945 * reregister request prior to unregistering linger so
946 * that r_osd is preserved.
948 BUG_ON(!list_empty(&req->r_req_lru_item));
949 __register_request(osdc, req);
950 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
951 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
952 __unregister_linger_request(osdc, req);
953 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
959 * If the osd connection drops, we need to resubmit all requests.
961 static void osd_reset(struct ceph_connection *con)
963 struct ceph_osd *osd = con->private;
964 struct ceph_osd_client *osdc;
968 dout("osd_reset osd%d\n", osd->o_osd);
970 down_read(&osdc->map_sem);
971 mutex_lock(&osdc->request_mutex);
972 __kick_osd_requests(osdc, osd);
974 mutex_unlock(&osdc->request_mutex);
975 up_read(&osdc->map_sem);
979 * Track open sessions with osds.
981 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
983 struct ceph_osd *osd;
985 osd = kzalloc(sizeof(*osd), GFP_NOFS);
989 atomic_set(&osd->o_ref, 1);
992 RB_CLEAR_NODE(&osd->o_node);
993 INIT_LIST_HEAD(&osd->o_requests);
994 INIT_LIST_HEAD(&osd->o_linger_requests);
995 INIT_LIST_HEAD(&osd->o_osd_lru);
996 osd->o_incarnation = 1;
998 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1000 INIT_LIST_HEAD(&osd->o_keepalive_item);
1004 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1006 if (atomic_inc_not_zero(&osd->o_ref)) {
1007 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1008 atomic_read(&osd->o_ref));
1011 dout("get_osd %p FAIL\n", osd);
1016 static void put_osd(struct ceph_osd *osd)
1018 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1019 atomic_read(&osd->o_ref) - 1);
1020 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
1021 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1023 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1029 * remove an osd from our map
1031 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1033 dout("__remove_osd %p\n", osd);
1034 BUG_ON(!list_empty(&osd->o_requests));
1035 BUG_ON(!list_empty(&osd->o_linger_requests));
1037 rb_erase(&osd->o_node, &osdc->osds);
1038 list_del_init(&osd->o_osd_lru);
1039 ceph_con_close(&osd->o_con);
1043 static void remove_all_osds(struct ceph_osd_client *osdc)
1045 dout("%s %p\n", __func__, osdc);
1046 mutex_lock(&osdc->request_mutex);
1047 while (!RB_EMPTY_ROOT(&osdc->osds)) {
1048 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1049 struct ceph_osd, o_node);
1050 __remove_osd(osdc, osd);
1052 mutex_unlock(&osdc->request_mutex);
1055 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1056 struct ceph_osd *osd)
1058 dout("%s %p\n", __func__, osd);
1059 BUG_ON(!list_empty(&osd->o_osd_lru));
1061 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1062 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1065 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1066 struct ceph_osd *osd)
1068 dout("%s %p\n", __func__, osd);
1070 if (list_empty(&osd->o_requests) &&
1071 list_empty(&osd->o_linger_requests))
1072 __move_osd_to_lru(osdc, osd);
1075 static void __remove_osd_from_lru(struct ceph_osd *osd)
1077 dout("__remove_osd_from_lru %p\n", osd);
1078 if (!list_empty(&osd->o_osd_lru))
1079 list_del_init(&osd->o_osd_lru);
1082 static void remove_old_osds(struct ceph_osd_client *osdc)
1084 struct ceph_osd *osd, *nosd;
1086 dout("__remove_old_osds %p\n", osdc);
1087 mutex_lock(&osdc->request_mutex);
1088 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1089 if (time_before(jiffies, osd->lru_ttl))
1091 __remove_osd(osdc, osd);
1093 mutex_unlock(&osdc->request_mutex);
1099 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1101 struct ceph_entity_addr *peer_addr;
1103 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1104 if (list_empty(&osd->o_requests) &&
1105 list_empty(&osd->o_linger_requests)) {
1106 __remove_osd(osdc, osd);
1111 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1112 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1113 !ceph_con_opened(&osd->o_con)) {
1114 struct ceph_osd_request *req;
1116 dout("osd addr hasn't changed and connection never opened, "
1117 "letting msgr retry\n");
1118 /* touch each r_stamp for handle_timeout()'s benfit */
1119 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1120 req->r_stamp = jiffies;
1125 ceph_con_close(&osd->o_con);
1126 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1127 osd->o_incarnation++;
1132 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1134 struct rb_node **p = &osdc->osds.rb_node;
1135 struct rb_node *parent = NULL;
1136 struct ceph_osd *osd = NULL;
1138 dout("__insert_osd %p osd%d\n", new, new->o_osd);
1141 osd = rb_entry(parent, struct ceph_osd, o_node);
1142 if (new->o_osd < osd->o_osd)
1144 else if (new->o_osd > osd->o_osd)
1145 p = &(*p)->rb_right;
1150 rb_link_node(&new->o_node, parent, p);
1151 rb_insert_color(&new->o_node, &osdc->osds);
1154 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1156 struct ceph_osd *osd;
1157 struct rb_node *n = osdc->osds.rb_node;
1160 osd = rb_entry(n, struct ceph_osd, o_node);
1163 else if (o > osd->o_osd)
1171 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1173 schedule_delayed_work(&osdc->timeout_work,
1174 osdc->client->options->osd_keepalive_timeout * HZ);
1177 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1179 cancel_delayed_work(&osdc->timeout_work);
1183 * Register request, assign tid. If this is the first request, set up
1184 * the timeout event.
1186 static void __register_request(struct ceph_osd_client *osdc,
1187 struct ceph_osd_request *req)
1189 req->r_tid = ++osdc->last_tid;
1190 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1191 dout("__register_request %p tid %lld\n", req, req->r_tid);
1192 __insert_request(osdc, req);
1193 ceph_osdc_get_request(req);
1194 osdc->num_requests++;
1195 if (osdc->num_requests == 1) {
1196 dout(" first request, scheduling timeout\n");
1197 __schedule_osd_timeout(osdc);
1202 * called under osdc->request_mutex
1204 static void __unregister_request(struct ceph_osd_client *osdc,
1205 struct ceph_osd_request *req)
1207 if (RB_EMPTY_NODE(&req->r_node)) {
1208 dout("__unregister_request %p tid %lld not registered\n",
1213 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1214 rb_erase(&req->r_node, &osdc->requests);
1215 RB_CLEAR_NODE(&req->r_node);
1216 osdc->num_requests--;
1219 /* make sure the original request isn't in flight. */
1220 ceph_msg_revoke(req->r_request);
1222 list_del_init(&req->r_osd_item);
1223 maybe_move_osd_to_lru(osdc, req->r_osd);
1224 if (list_empty(&req->r_linger_osd_item))
1228 list_del_init(&req->r_req_lru_item);
1229 ceph_osdc_put_request(req);
1231 if (osdc->num_requests == 0) {
1232 dout(" no requests, canceling timeout\n");
1233 __cancel_osd_timeout(osdc);
1238 * Cancel a previously queued request message
1240 static void __cancel_request(struct ceph_osd_request *req)
1242 if (req->r_sent && req->r_osd) {
1243 ceph_msg_revoke(req->r_request);
1248 static void __register_linger_request(struct ceph_osd_client *osdc,
1249 struct ceph_osd_request *req)
1251 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1252 WARN_ON(!req->r_linger);
1254 ceph_osdc_get_request(req);
1255 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1257 list_add_tail(&req->r_linger_osd_item,
1258 &req->r_osd->o_linger_requests);
1261 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1262 struct ceph_osd_request *req)
1264 WARN_ON(!req->r_linger);
1266 if (list_empty(&req->r_linger_item)) {
1267 dout("%s %p tid %llu not registered\n", __func__, req,
1272 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1273 list_del_init(&req->r_linger_item);
1276 list_del_init(&req->r_linger_osd_item);
1277 maybe_move_osd_to_lru(osdc, req->r_osd);
1278 if (list_empty(&req->r_osd_item))
1281 ceph_osdc_put_request(req);
1284 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1285 struct ceph_osd_request *req)
1287 mutex_lock(&osdc->request_mutex);
1289 __unregister_linger_request(osdc, req);
1290 mutex_unlock(&osdc->request_mutex);
1292 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1294 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1295 struct ceph_osd_request *req)
1297 if (!req->r_linger) {
1298 dout("set_request_linger %p\n", req);
1302 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1305 * Returns whether a request should be blocked from being sent
1306 * based on the current osdmap and osd_client settings.
1308 * Caller should hold map_sem for read.
1310 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1311 struct ceph_osd_request *req)
1313 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1314 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1315 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1316 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1317 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1321 * Calculate mapping of a request to a PG. Takes tiering into account.
1323 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1324 struct ceph_osd_request *req,
1325 struct ceph_pg *pg_out)
1327 bool need_check_tiering;
1329 need_check_tiering = false;
1330 if (req->r_target_oloc.pool == -1) {
1331 req->r_target_oloc = req->r_base_oloc; /* struct */
1332 need_check_tiering = true;
1334 if (req->r_target_oid.name_len == 0) {
1335 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1336 need_check_tiering = true;
1339 if (need_check_tiering &&
1340 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1341 struct ceph_pg_pool_info *pi;
1343 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1345 if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1347 req->r_target_oloc.pool = pi->read_tier;
1348 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1349 pi->write_tier >= 0)
1350 req->r_target_oloc.pool = pi->write_tier;
1352 /* !pi is caught in ceph_oloc_oid_to_pg() */
1355 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1356 &req->r_target_oid, pg_out);
1360 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1361 * (as needed), and set the request r_osd appropriately. If there is
1362 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1363 * (unsent, homeless) or leave on in-flight lru.
1365 * Return 0 if unchanged, 1 if changed, or negative on error.
1367 * Caller should hold map_sem for read and request_mutex.
1369 static int __map_request(struct ceph_osd_client *osdc,
1370 struct ceph_osd_request *req, int force_resend)
1372 struct ceph_pg pgid;
1373 int acting[CEPH_PG_MAX_SIZE];
1378 dout("map_request %p tid %lld\n", req, req->r_tid);
1380 err = __calc_request_pg(osdc->osdmap, req, &pgid);
1382 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1387 num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1391 was_paused = req->r_paused;
1392 req->r_paused = __req_should_be_paused(osdc, req);
1393 if (was_paused && !req->r_paused)
1396 if ((!force_resend &&
1397 req->r_osd && req->r_osd->o_osd == o &&
1398 req->r_sent >= req->r_osd->o_incarnation &&
1399 req->r_num_pg_osds == num &&
1400 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1401 (req->r_osd == NULL && o == -1) ||
1403 return 0; /* no change */
1405 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1406 req->r_tid, pgid.pool, pgid.seed, o,
1407 req->r_osd ? req->r_osd->o_osd : -1);
1409 /* record full pg acting set */
1410 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1411 req->r_num_pg_osds = num;
1414 __cancel_request(req);
1415 list_del_init(&req->r_osd_item);
1419 req->r_osd = __lookup_osd(osdc, o);
1420 if (!req->r_osd && o >= 0) {
1422 req->r_osd = create_osd(osdc, o);
1424 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1428 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1429 __insert_osd(osdc, req->r_osd);
1431 ceph_con_open(&req->r_osd->o_con,
1432 CEPH_ENTITY_TYPE_OSD, o,
1433 &osdc->osdmap->osd_addr[o]);
1437 __remove_osd_from_lru(req->r_osd);
1438 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1439 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1441 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1443 err = 1; /* osd or pg changed */
1450 * caller should hold map_sem (for read) and request_mutex
1452 static void __send_request(struct ceph_osd_client *osdc,
1453 struct ceph_osd_request *req)
1457 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1458 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1459 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1461 /* fill in message content that changes each time we send it */
1462 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1463 put_unaligned_le32(req->r_flags, req->r_request_flags);
1464 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1465 p = req->r_request_pgid;
1466 ceph_encode_64(&p, req->r_pgid.pool);
1467 ceph_encode_32(&p, req->r_pgid.seed);
1468 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1469 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1470 sizeof(req->r_reassert_version));
1472 req->r_stamp = jiffies;
1473 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1475 ceph_msg_get(req->r_request); /* send consumes a ref */
1477 req->r_sent = req->r_osd->o_incarnation;
1479 ceph_con_send(&req->r_osd->o_con, req->r_request);
1483 * Send any requests in the queue (req_unsent).
1485 static void __send_queued(struct ceph_osd_client *osdc)
1487 struct ceph_osd_request *req, *tmp;
1489 dout("__send_queued\n");
1490 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1491 __send_request(osdc, req);
1495 * Caller should hold map_sem for read and request_mutex.
1497 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1498 struct ceph_osd_request *req,
1503 __register_request(osdc, req);
1505 req->r_got_reply = 0;
1506 rc = __map_request(osdc, req, 0);
1509 dout("osdc_start_request failed map, "
1510 " will retry %lld\n", req->r_tid);
1513 __unregister_request(osdc, req);
1518 if (req->r_osd == NULL) {
1519 dout("send_request %p no up osds in pg\n", req);
1520 ceph_monc_request_next_osdmap(&osdc->client->monc);
1522 __send_queued(osdc);
1529 * Timeout callback, called every N seconds when 1 or more osd
1530 * requests has been active for more than N seconds. When this
1531 * happens, we ping all OSDs with requests who have timed out to
1532 * ensure any communications channel reset is detected. Reset the
1533 * request timeouts another N seconds in the future as we go.
1534 * Reschedule the timeout event another N seconds in future (unless
1535 * there are no open requests).
1537 static void handle_timeout(struct work_struct *work)
1539 struct ceph_osd_client *osdc =
1540 container_of(work, struct ceph_osd_client, timeout_work.work);
1541 struct ceph_osd_request *req;
1542 struct ceph_osd *osd;
1543 unsigned long keepalive =
1544 osdc->client->options->osd_keepalive_timeout * HZ;
1545 struct list_head slow_osds;
1547 down_read(&osdc->map_sem);
1549 ceph_monc_request_next_osdmap(&osdc->client->monc);
1551 mutex_lock(&osdc->request_mutex);
1554 * ping osds that are a bit slow. this ensures that if there
1555 * is a break in the TCP connection we will notice, and reopen
1556 * a connection with that osd (from the fault callback).
1558 INIT_LIST_HEAD(&slow_osds);
1559 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1560 if (time_before(jiffies, req->r_stamp + keepalive))
1565 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1566 req->r_tid, osd->o_osd);
1567 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1569 while (!list_empty(&slow_osds)) {
1570 osd = list_entry(slow_osds.next, struct ceph_osd,
1572 list_del_init(&osd->o_keepalive_item);
1573 ceph_con_keepalive(&osd->o_con);
1576 __schedule_osd_timeout(osdc);
1577 __send_queued(osdc);
1578 mutex_unlock(&osdc->request_mutex);
1579 up_read(&osdc->map_sem);
1582 static void handle_osds_timeout(struct work_struct *work)
1584 struct ceph_osd_client *osdc =
1585 container_of(work, struct ceph_osd_client,
1586 osds_timeout_work.work);
1587 unsigned long delay =
1588 osdc->client->options->osd_idle_ttl * HZ >> 2;
1590 dout("osds timeout\n");
1591 down_read(&osdc->map_sem);
1592 remove_old_osds(osdc);
1593 up_read(&osdc->map_sem);
1595 schedule_delayed_work(&osdc->osds_timeout_work,
1596 round_jiffies_relative(delay));
1599 static int ceph_oloc_decode(void **p, void *end,
1600 struct ceph_object_locator *oloc)
1602 u8 struct_v, struct_cv;
1607 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1608 struct_v = ceph_decode_8(p);
1609 struct_cv = ceph_decode_8(p);
1611 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1612 struct_v, struct_cv);
1615 if (struct_cv > 6) {
1616 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1617 struct_v, struct_cv);
1620 len = ceph_decode_32(p);
1621 ceph_decode_need(p, end, len, e_inval);
1622 struct_end = *p + len;
1624 oloc->pool = ceph_decode_64(p);
1625 *p += 4; /* skip preferred */
1627 len = ceph_decode_32(p);
1629 pr_warn("ceph_object_locator::key is set\n");
1633 if (struct_v >= 5) {
1634 len = ceph_decode_32(p);
1636 pr_warn("ceph_object_locator::nspace is set\n");
1641 if (struct_v >= 6) {
1642 s64 hash = ceph_decode_64(p);
1644 pr_warn("ceph_object_locator::hash is set\n");
1659 static int ceph_redirect_decode(void **p, void *end,
1660 struct ceph_request_redirect *redir)
1662 u8 struct_v, struct_cv;
1667 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1668 struct_v = ceph_decode_8(p);
1669 struct_cv = ceph_decode_8(p);
1670 if (struct_cv > 1) {
1671 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1672 struct_v, struct_cv);
1675 len = ceph_decode_32(p);
1676 ceph_decode_need(p, end, len, e_inval);
1677 struct_end = *p + len;
1679 ret = ceph_oloc_decode(p, end, &redir->oloc);
1683 len = ceph_decode_32(p);
1685 pr_warn("ceph_request_redirect::object_name is set\n");
1689 len = ceph_decode_32(p);
1690 *p += len; /* skip osd_instructions */
1702 static void complete_request(struct ceph_osd_request *req)
1704 complete_all(&req->r_safe_completion); /* fsync waiter */
1708 * handle osd op reply. either call the callback if it is specified,
1709 * or do the completion to wake up the waiting thread.
1711 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1712 struct ceph_connection *con)
1715 struct ceph_osd_request *req;
1716 struct ceph_request_redirect redir;
1719 unsigned int numops;
1720 int payload_len, flags;
1726 u64 reassert_version;
1728 int already_completed;
1732 tid = le64_to_cpu(msg->hdr.tid);
1733 dout("handle_reply %p tid %llu\n", msg, tid);
1735 p = msg->front.iov_base;
1736 end = p + msg->front.iov_len;
1738 ceph_decode_need(&p, end, 4, bad);
1739 object_len = ceph_decode_32(&p);
1740 ceph_decode_need(&p, end, object_len, bad);
1743 err = ceph_decode_pgid(&p, end, &pg);
1747 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1748 flags = ceph_decode_64(&p);
1749 result = ceph_decode_32(&p);
1750 reassert_epoch = ceph_decode_32(&p);
1751 reassert_version = ceph_decode_64(&p);
1752 osdmap_epoch = ceph_decode_32(&p);
1755 down_read(&osdc->map_sem);
1756 mutex_lock(&osdc->request_mutex);
1757 req = __lookup_request(osdc, tid);
1759 dout("handle_reply tid %llu dne\n", tid);
1762 ceph_osdc_get_request(req);
1764 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1767 ceph_decode_need(&p, end, 4, bad_put);
1768 numops = ceph_decode_32(&p);
1769 if (numops > CEPH_OSD_MAX_OP)
1771 if (numops != req->r_num_ops)
1774 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1775 for (i = 0; i < numops; i++) {
1776 struct ceph_osd_op *op = p;
1779 len = le32_to_cpu(op->payload_len);
1780 req->r_reply_op_len[i] = len;
1781 dout(" op %d has %d bytes\n", i, len);
1785 bytes = le32_to_cpu(msg->hdr.data_len);
1786 if (payload_len != bytes) {
1787 pr_warning("sum of op payload lens %d != data_len %d",
1788 payload_len, bytes);
1792 ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1793 retry_attempt = ceph_decode_32(&p);
1794 for (i = 0; i < numops; i++)
1795 req->r_reply_op_result[i] = ceph_decode_32(&p);
1797 if (le16_to_cpu(msg->hdr.version) >= 6) {
1798 p += 8 + 4; /* skip replay_version */
1799 p += 8; /* skip user_version */
1801 err = ceph_redirect_decode(&p, end, &redir);
1805 redir.oloc.pool = -1;
1808 if (redir.oloc.pool != -1) {
1809 dout("redirect pool %lld\n", redir.oloc.pool);
1811 __unregister_request(osdc, req);
1813 req->r_target_oloc = redir.oloc; /* struct */
1816 * Start redirect requests with nofail=true. If
1817 * mapping fails, request will end up on the notarget
1818 * list, waiting for the new osdmap (which can take
1819 * a while), even though the original request mapped
1820 * successfully. In the future we might want to follow
1821 * original request's nofail setting here.
1823 err = __ceph_osdc_start_request(osdc, req, true);
1829 already_completed = req->r_got_reply;
1830 if (!req->r_got_reply) {
1831 req->r_result = result;
1832 dout("handle_reply result %d bytes %d\n", req->r_result,
1834 if (req->r_result == 0)
1835 req->r_result = bytes;
1837 /* in case this is a write and we need to replay, */
1838 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1839 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1841 req->r_got_reply = 1;
1842 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1843 dout("handle_reply tid %llu dup ack\n", tid);
1847 dout("handle_reply tid %llu flags %d\n", tid, flags);
1849 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1850 __register_linger_request(osdc, req);
1852 /* either this is a read, or we got the safe response */
1854 (flags & CEPH_OSD_FLAG_ONDISK) ||
1855 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1856 __unregister_request(osdc, req);
1858 mutex_unlock(&osdc->request_mutex);
1859 up_read(&osdc->map_sem);
1861 if (!already_completed) {
1862 if (req->r_unsafe_callback &&
1863 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1864 req->r_unsafe_callback(req, true);
1865 if (req->r_callback)
1866 req->r_callback(req, msg);
1868 complete_all(&req->r_completion);
1871 if (flags & CEPH_OSD_FLAG_ONDISK) {
1872 if (req->r_unsafe_callback && already_completed)
1873 req->r_unsafe_callback(req, false);
1874 complete_request(req);
1878 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1879 ceph_osdc_put_request(req);
1882 mutex_unlock(&osdc->request_mutex);
1883 up_read(&osdc->map_sem);
1887 req->r_result = -EIO;
1888 __unregister_request(osdc, req);
1889 if (req->r_callback)
1890 req->r_callback(req, msg);
1892 complete_all(&req->r_completion);
1893 complete_request(req);
1894 ceph_osdc_put_request(req);
1896 mutex_unlock(&osdc->request_mutex);
1897 up_read(&osdc->map_sem);
1899 pr_err("corrupt osd_op_reply got %d %d\n",
1900 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1904 static void reset_changed_osds(struct ceph_osd_client *osdc)
1906 struct rb_node *p, *n;
1908 for (p = rb_first(&osdc->osds); p; p = n) {
1909 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1912 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1913 memcmp(&osd->o_con.peer_addr,
1914 ceph_osd_addr(osdc->osdmap,
1916 sizeof(struct ceph_entity_addr)) != 0)
1917 __reset_osd(osdc, osd);
1922 * Requeue requests whose mapping to an OSD has changed. If requests map to
1923 * no osd, request a new map.
1925 * Caller should hold map_sem for read.
1927 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1928 bool force_resend_writes)
1930 struct ceph_osd_request *req, *nreq;
1934 bool force_resend_req;
1936 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1937 force_resend_writes ? " (force resend writes)" : "");
1938 mutex_lock(&osdc->request_mutex);
1939 for (p = rb_first(&osdc->requests); p; ) {
1940 req = rb_entry(p, struct ceph_osd_request, r_node);
1944 * For linger requests that have not yet been
1945 * registered, move them to the linger list; they'll
1946 * be sent to the osd in the loop below. Unregister
1947 * the request before re-registering it as a linger
1948 * request to ensure the __map_request() below
1949 * will decide it needs to be sent.
1951 if (req->r_linger && list_empty(&req->r_linger_item)) {
1952 dout("%p tid %llu restart on osd%d\n",
1954 req->r_osd ? req->r_osd->o_osd : -1);
1955 ceph_osdc_get_request(req);
1956 __unregister_request(osdc, req);
1957 __register_linger_request(osdc, req);
1958 ceph_osdc_put_request(req);
1962 force_resend_req = force_resend ||
1963 (force_resend_writes &&
1964 req->r_flags & CEPH_OSD_FLAG_WRITE);
1965 err = __map_request(osdc, req, force_resend_req);
1967 continue; /* error */
1968 if (req->r_osd == NULL) {
1969 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1970 needmap++; /* request a newer map */
1971 } else if (err > 0) {
1972 if (!req->r_linger) {
1973 dout("%p tid %llu requeued on osd%d\n", req,
1975 req->r_osd ? req->r_osd->o_osd : -1);
1976 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1981 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1983 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1985 err = __map_request(osdc, req,
1986 force_resend || force_resend_writes);
1987 dout("__map_request returned %d\n", err);
1989 continue; /* no change and no osd was specified */
1991 continue; /* hrm! */
1992 if (req->r_osd == NULL) {
1993 dout("tid %llu maps to no valid osd\n", req->r_tid);
1994 needmap++; /* request a newer map */
1998 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1999 req->r_osd ? req->r_osd->o_osd : -1);
2000 __register_request(osdc, req);
2001 __unregister_linger_request(osdc, req);
2003 reset_changed_osds(osdc);
2004 mutex_unlock(&osdc->request_mutex);
2007 dout("%d requests for down osds, need new map\n", needmap);
2008 ceph_monc_request_next_osdmap(&osdc->client->monc);
2014 * Process updated osd map.
2016 * The message contains any number of incremental and full maps, normally
2017 * indicating some sort of topology change in the cluster. Kick requests
2018 * off to different OSDs as needed.
2020 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2022 void *p, *end, *next;
2023 u32 nr_maps, maplen;
2025 struct ceph_osdmap *newmap = NULL, *oldmap;
2027 struct ceph_fsid fsid;
2030 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2031 p = msg->front.iov_base;
2032 end = p + msg->front.iov_len;
2035 ceph_decode_need(&p, end, sizeof(fsid), bad);
2036 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2037 if (ceph_check_fsid(osdc->client, &fsid) < 0)
2040 down_write(&osdc->map_sem);
2042 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2044 /* incremental maps */
2045 ceph_decode_32_safe(&p, end, nr_maps, bad);
2046 dout(" %d inc maps\n", nr_maps);
2047 while (nr_maps > 0) {
2048 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2049 epoch = ceph_decode_32(&p);
2050 maplen = ceph_decode_32(&p);
2051 ceph_decode_need(&p, end, maplen, bad);
2053 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2054 dout("applying incremental map %u len %d\n",
2056 newmap = osdmap_apply_incremental(&p, next,
2058 &osdc->client->msgr);
2059 if (IS_ERR(newmap)) {
2060 err = PTR_ERR(newmap);
2064 if (newmap != osdc->osdmap) {
2065 ceph_osdmap_destroy(osdc->osdmap);
2066 osdc->osdmap = newmap;
2068 was_full = was_full ||
2069 ceph_osdmap_flag(osdc->osdmap,
2071 kick_requests(osdc, 0, was_full);
2073 dout("ignoring incremental map %u len %d\n",
2083 ceph_decode_32_safe(&p, end, nr_maps, bad);
2084 dout(" %d full maps\n", nr_maps);
2086 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2087 epoch = ceph_decode_32(&p);
2088 maplen = ceph_decode_32(&p);
2089 ceph_decode_need(&p, end, maplen, bad);
2091 dout("skipping non-latest full map %u len %d\n",
2093 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2094 dout("skipping full map %u len %d, "
2095 "older than our %u\n", epoch, maplen,
2096 osdc->osdmap->epoch);
2098 int skipped_map = 0;
2100 dout("taking full map %u len %d\n", epoch, maplen);
2101 newmap = ceph_osdmap_decode(&p, p+maplen);
2102 if (IS_ERR(newmap)) {
2103 err = PTR_ERR(newmap);
2107 oldmap = osdc->osdmap;
2108 osdc->osdmap = newmap;
2110 if (oldmap->epoch + 1 < newmap->epoch)
2112 ceph_osdmap_destroy(oldmap);
2114 was_full = was_full ||
2115 ceph_osdmap_flag(osdc->osdmap,
2117 kick_requests(osdc, skipped_map, was_full);
2126 downgrade_write(&osdc->map_sem);
2127 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2130 * subscribe to subsequent osdmap updates if full to ensure
2131 * we find out when we are no longer full and stop returning
2134 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2135 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2136 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2137 ceph_monc_request_next_osdmap(&osdc->client->monc);
2139 mutex_lock(&osdc->request_mutex);
2140 __send_queued(osdc);
2141 mutex_unlock(&osdc->request_mutex);
2142 up_read(&osdc->map_sem);
2143 wake_up_all(&osdc->client->auth_wq);
2147 pr_err("osdc handle_map corrupt msg\n");
2149 up_write(&osdc->map_sem);
2153 * watch/notify callback event infrastructure
2155 * These callbacks are used both for watch and notify operations.
2157 static void __release_event(struct kref *kref)
2159 struct ceph_osd_event *event =
2160 container_of(kref, struct ceph_osd_event, kref);
2162 dout("__release_event %p\n", event);
2166 static void get_event(struct ceph_osd_event *event)
2168 kref_get(&event->kref);
2171 void ceph_osdc_put_event(struct ceph_osd_event *event)
2173 kref_put(&event->kref, __release_event);
2175 EXPORT_SYMBOL(ceph_osdc_put_event);
2177 static void __insert_event(struct ceph_osd_client *osdc,
2178 struct ceph_osd_event *new)
2180 struct rb_node **p = &osdc->event_tree.rb_node;
2181 struct rb_node *parent = NULL;
2182 struct ceph_osd_event *event = NULL;
2186 event = rb_entry(parent, struct ceph_osd_event, node);
2187 if (new->cookie < event->cookie)
2189 else if (new->cookie > event->cookie)
2190 p = &(*p)->rb_right;
2195 rb_link_node(&new->node, parent, p);
2196 rb_insert_color(&new->node, &osdc->event_tree);
2199 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2202 struct rb_node **p = &osdc->event_tree.rb_node;
2203 struct rb_node *parent = NULL;
2204 struct ceph_osd_event *event = NULL;
2208 event = rb_entry(parent, struct ceph_osd_event, node);
2209 if (cookie < event->cookie)
2211 else if (cookie > event->cookie)
2212 p = &(*p)->rb_right;
2219 static void __remove_event(struct ceph_osd_event *event)
2221 struct ceph_osd_client *osdc = event->osdc;
2223 if (!RB_EMPTY_NODE(&event->node)) {
2224 dout("__remove_event removed %p\n", event);
2225 rb_erase(&event->node, &osdc->event_tree);
2226 ceph_osdc_put_event(event);
2228 dout("__remove_event didn't remove %p\n", event);
2232 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2233 void (*event_cb)(u64, u64, u8, void *),
2234 void *data, struct ceph_osd_event **pevent)
2236 struct ceph_osd_event *event;
2238 event = kmalloc(sizeof(*event), GFP_NOIO);
2242 dout("create_event %p\n", event);
2243 event->cb = event_cb;
2244 event->one_shot = 0;
2247 INIT_LIST_HEAD(&event->osd_node);
2248 RB_CLEAR_NODE(&event->node);
2249 kref_init(&event->kref); /* one ref for us */
2250 kref_get(&event->kref); /* one ref for the caller */
2252 spin_lock(&osdc->event_lock);
2253 event->cookie = ++osdc->event_count;
2254 __insert_event(osdc, event);
2255 spin_unlock(&osdc->event_lock);
2260 EXPORT_SYMBOL(ceph_osdc_create_event);
2262 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2264 struct ceph_osd_client *osdc = event->osdc;
2266 dout("cancel_event %p\n", event);
2267 spin_lock(&osdc->event_lock);
2268 __remove_event(event);
2269 spin_unlock(&osdc->event_lock);
2270 ceph_osdc_put_event(event); /* caller's */
2272 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2275 static void do_event_work(struct work_struct *work)
2277 struct ceph_osd_event_work *event_work =
2278 container_of(work, struct ceph_osd_event_work, work);
2279 struct ceph_osd_event *event = event_work->event;
2280 u64 ver = event_work->ver;
2281 u64 notify_id = event_work->notify_id;
2282 u8 opcode = event_work->opcode;
2284 dout("do_event_work completing %p\n", event);
2285 event->cb(ver, notify_id, opcode, event->data);
2286 dout("do_event_work completed %p\n", event);
2287 ceph_osdc_put_event(event);
2293 * Process osd watch notifications
2295 static void handle_watch_notify(struct ceph_osd_client *osdc,
2296 struct ceph_msg *msg)
2300 u64 cookie, ver, notify_id;
2302 struct ceph_osd_event *event;
2303 struct ceph_osd_event_work *event_work;
2305 p = msg->front.iov_base;
2306 end = p + msg->front.iov_len;
2308 ceph_decode_8_safe(&p, end, proto_ver, bad);
2309 ceph_decode_8_safe(&p, end, opcode, bad);
2310 ceph_decode_64_safe(&p, end, cookie, bad);
2311 ceph_decode_64_safe(&p, end, ver, bad);
2312 ceph_decode_64_safe(&p, end, notify_id, bad);
2314 spin_lock(&osdc->event_lock);
2315 event = __find_event(osdc, cookie);
2317 BUG_ON(event->one_shot);
2320 spin_unlock(&osdc->event_lock);
2321 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2322 cookie, ver, event);
2324 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2326 dout("ERROR: could not allocate event_work\n");
2329 INIT_WORK(&event_work->work, do_event_work);
2330 event_work->event = event;
2331 event_work->ver = ver;
2332 event_work->notify_id = notify_id;
2333 event_work->opcode = opcode;
2334 if (!queue_work(osdc->notify_wq, &event_work->work)) {
2335 dout("WARNING: failed to queue notify event work\n");
2343 ceph_osdc_put_event(event);
2347 pr_err("osdc handle_watch_notify corrupt msg\n");
2351 * build new request AND message
2354 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2355 struct ceph_snap_context *snapc, u64 snap_id,
2356 struct timespec *mtime)
2358 struct ceph_msg *msg = req->r_request;
2361 int flags = req->r_flags;
2365 req->r_snapid = snap_id;
2366 req->r_snapc = ceph_get_snap_context(snapc);
2368 /* encode request */
2369 msg->hdr.version = cpu_to_le16(4);
2371 p = msg->front.iov_base;
2372 ceph_encode_32(&p, 1); /* client_inc is always 1 */
2373 req->r_request_osdmap_epoch = p;
2375 req->r_request_flags = p;
2377 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2378 ceph_encode_timespec(p, mtime);
2379 p += sizeof(struct ceph_timespec);
2380 req->r_request_reassert_version = p;
2381 p += sizeof(struct ceph_eversion); /* will get filled in */
2384 ceph_encode_8(&p, 4);
2385 ceph_encode_8(&p, 4);
2386 ceph_encode_32(&p, 8 + 4 + 4);
2387 req->r_request_pool = p;
2389 ceph_encode_32(&p, -1); /* preferred */
2390 ceph_encode_32(&p, 0); /* key len */
2392 ceph_encode_8(&p, 1);
2393 req->r_request_pgid = p;
2395 ceph_encode_32(&p, -1); /* preferred */
2398 ceph_encode_32(&p, req->r_base_oid.name_len);
2399 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2400 dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2401 req->r_base_oid.name, req->r_base_oid.name_len);
2402 p += req->r_base_oid.name_len;
2404 /* ops--can imply data */
2405 ceph_encode_16(&p, (u16)req->r_num_ops);
2407 for (i = 0; i < req->r_num_ops; i++) {
2408 data_len += osd_req_encode_op(req, p, i);
2409 p += sizeof(struct ceph_osd_op);
2413 ceph_encode_64(&p, req->r_snapid);
2414 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2415 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2417 for (i = 0; i < snapc->num_snaps; i++) {
2418 ceph_encode_64(&p, req->r_snapc->snaps[i]);
2422 req->r_request_attempts = p;
2426 if (flags & CEPH_OSD_FLAG_WRITE) {
2430 * The header "data_off" is a hint to the receiver
2431 * allowing it to align received data into its
2432 * buffers such that there's no need to re-copy
2433 * it before writing it to disk (direct I/O).
2435 data_off = (u16) (off & 0xffff);
2436 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2438 req->r_request->hdr.data_len = cpu_to_le32(data_len);
2440 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2441 msg_size = p - msg->front.iov_base;
2442 msg->front.iov_len = msg_size;
2443 msg->hdr.front_len = cpu_to_le32(msg_size);
2445 dout("build_request msg_size was %d\n", (int)msg_size);
2447 EXPORT_SYMBOL(ceph_osdc_build_request);
2450 * Register request, send initial attempt.
2452 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2453 struct ceph_osd_request *req,
2458 down_read(&osdc->map_sem);
2459 mutex_lock(&osdc->request_mutex);
2461 rc = __ceph_osdc_start_request(osdc, req, nofail);
2463 mutex_unlock(&osdc->request_mutex);
2464 up_read(&osdc->map_sem);
2468 EXPORT_SYMBOL(ceph_osdc_start_request);
2471 * Unregister a registered request. The request is not completed (i.e.
2472 * no callbacks or wakeups) - higher layers are supposed to know what
2473 * they are canceling.
2475 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2477 struct ceph_osd_client *osdc = req->r_osdc;
2479 mutex_lock(&osdc->request_mutex);
2481 __unregister_linger_request(osdc, req);
2482 __unregister_request(osdc, req);
2483 mutex_unlock(&osdc->request_mutex);
2485 dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2487 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2490 * wait for a request to complete
2492 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2493 struct ceph_osd_request *req)
2497 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2499 rc = wait_for_completion_interruptible(&req->r_completion);
2501 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2502 ceph_osdc_cancel_request(req);
2503 complete_request(req);
2507 dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2509 return req->r_result;
2511 EXPORT_SYMBOL(ceph_osdc_wait_request);
2514 * sync - wait for all in-flight requests to flush. avoid starvation.
2516 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2518 struct ceph_osd_request *req;
2519 u64 last_tid, next_tid = 0;
2521 mutex_lock(&osdc->request_mutex);
2522 last_tid = osdc->last_tid;
2524 req = __lookup_request_ge(osdc, next_tid);
2527 if (req->r_tid > last_tid)
2530 next_tid = req->r_tid + 1;
2531 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2534 ceph_osdc_get_request(req);
2535 mutex_unlock(&osdc->request_mutex);
2536 dout("sync waiting on tid %llu (last is %llu)\n",
2537 req->r_tid, last_tid);
2538 wait_for_completion(&req->r_safe_completion);
2539 mutex_lock(&osdc->request_mutex);
2540 ceph_osdc_put_request(req);
2542 mutex_unlock(&osdc->request_mutex);
2543 dout("sync done (thru tid %llu)\n", last_tid);
2545 EXPORT_SYMBOL(ceph_osdc_sync);
2548 * Call all pending notify callbacks - for use after a watch is
2549 * unregistered, to make sure no more callbacks for it will be invoked
2551 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2553 flush_workqueue(osdc->notify_wq);
2555 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2561 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2566 osdc->client = client;
2567 osdc->osdmap = NULL;
2568 init_rwsem(&osdc->map_sem);
2569 init_completion(&osdc->map_waiters);
2570 osdc->last_requested_map = 0;
2571 mutex_init(&osdc->request_mutex);
2573 osdc->osds = RB_ROOT;
2574 INIT_LIST_HEAD(&osdc->osd_lru);
2575 osdc->requests = RB_ROOT;
2576 INIT_LIST_HEAD(&osdc->req_lru);
2577 INIT_LIST_HEAD(&osdc->req_unsent);
2578 INIT_LIST_HEAD(&osdc->req_notarget);
2579 INIT_LIST_HEAD(&osdc->req_linger);
2580 osdc->num_requests = 0;
2581 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2582 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2583 spin_lock_init(&osdc->event_lock);
2584 osdc->event_tree = RB_ROOT;
2585 osdc->event_count = 0;
2587 schedule_delayed_work(&osdc->osds_timeout_work,
2588 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2591 osdc->req_mempool = mempool_create_kmalloc_pool(10,
2592 sizeof(struct ceph_osd_request));
2593 if (!osdc->req_mempool)
2596 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2597 OSD_OP_FRONT_LEN, 10, true,
2601 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2602 OSD_OPREPLY_FRONT_LEN, 10, true,
2608 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2609 if (!osdc->notify_wq)
2610 goto out_msgpool_reply;
2615 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2617 ceph_msgpool_destroy(&osdc->msgpool_op);
2619 mempool_destroy(osdc->req_mempool);
2624 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2626 flush_workqueue(osdc->notify_wq);
2627 destroy_workqueue(osdc->notify_wq);
2628 cancel_delayed_work_sync(&osdc->timeout_work);
2629 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2631 ceph_osdmap_destroy(osdc->osdmap);
2632 osdc->osdmap = NULL;
2634 remove_all_osds(osdc);
2635 mempool_destroy(osdc->req_mempool);
2636 ceph_msgpool_destroy(&osdc->msgpool_op);
2637 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2641 * Read some contiguous pages. If we cross a stripe boundary, shorten
2642 * *plen. Return number of bytes read, or error.
2644 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2645 struct ceph_vino vino, struct ceph_file_layout *layout,
2647 u32 truncate_seq, u64 truncate_size,
2648 struct page **pages, int num_pages, int page_align)
2650 struct ceph_osd_request *req;
2653 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2654 vino.snap, off, *plen);
2655 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2656 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2657 NULL, truncate_seq, truncate_size,
2660 return PTR_ERR(req);
2662 /* it may be a short read due to an object boundary */
2664 osd_req_op_extent_osd_data_pages(req, 0,
2665 pages, *plen, page_align, false, false);
2667 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2668 off, *plen, *plen, page_align);
2670 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2672 rc = ceph_osdc_start_request(osdc, req, false);
2674 rc = ceph_osdc_wait_request(osdc, req);
2676 ceph_osdc_put_request(req);
2677 dout("readpages result %d\n", rc);
2680 EXPORT_SYMBOL(ceph_osdc_readpages);
2683 * do a synchronous write on N pages
2685 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2686 struct ceph_file_layout *layout,
2687 struct ceph_snap_context *snapc,
2689 u32 truncate_seq, u64 truncate_size,
2690 struct timespec *mtime,
2691 struct page **pages, int num_pages)
2693 struct ceph_osd_request *req;
2695 int page_align = off & ~PAGE_MASK;
2697 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */
2698 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2700 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2701 snapc, truncate_seq, truncate_size,
2704 return PTR_ERR(req);
2706 /* it may be a short write due to an object boundary */
2707 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2709 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2711 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2713 rc = ceph_osdc_start_request(osdc, req, true);
2715 rc = ceph_osdc_wait_request(osdc, req);
2717 ceph_osdc_put_request(req);
2720 dout("writepages result %d\n", rc);
2723 EXPORT_SYMBOL(ceph_osdc_writepages);
2725 int ceph_osdc_setup(void)
2727 BUG_ON(ceph_osd_request_cache);
2728 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2729 sizeof (struct ceph_osd_request),
2730 __alignof__(struct ceph_osd_request),
2733 return ceph_osd_request_cache ? 0 : -ENOMEM;
2735 EXPORT_SYMBOL(ceph_osdc_setup);
2737 void ceph_osdc_cleanup(void)
2739 BUG_ON(!ceph_osd_request_cache);
2740 kmem_cache_destroy(ceph_osd_request_cache);
2741 ceph_osd_request_cache = NULL;
2743 EXPORT_SYMBOL(ceph_osdc_cleanup);
2746 * handle incoming message
2748 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2750 struct ceph_osd *osd = con->private;
2751 struct ceph_osd_client *osdc;
2752 int type = le16_to_cpu(msg->hdr.type);
2759 case CEPH_MSG_OSD_MAP:
2760 ceph_osdc_handle_map(osdc, msg);
2762 case CEPH_MSG_OSD_OPREPLY:
2763 handle_reply(osdc, msg, con);
2765 case CEPH_MSG_WATCH_NOTIFY:
2766 handle_watch_notify(osdc, msg);
2770 pr_err("received unknown message type %d %s\n", type,
2771 ceph_msg_type_name(type));
2778 * lookup and return message for incoming reply. set up reply message
2781 static struct ceph_msg *get_reply(struct ceph_connection *con,
2782 struct ceph_msg_header *hdr,
2785 struct ceph_osd *osd = con->private;
2786 struct ceph_osd_client *osdc = osd->o_osdc;
2788 struct ceph_osd_request *req;
2789 int front_len = le32_to_cpu(hdr->front_len);
2790 int data_len = le32_to_cpu(hdr->data_len);
2793 tid = le64_to_cpu(hdr->tid);
2794 mutex_lock(&osdc->request_mutex);
2795 req = __lookup_request(osdc, tid);
2799 dout("get_reply unknown tid %llu from osd%d\n", tid,
2804 if (req->r_reply->con)
2805 dout("%s revoking msg %p from old con %p\n", __func__,
2806 req->r_reply, req->r_reply->con);
2807 ceph_msg_revoke_incoming(req->r_reply);
2809 if (front_len > req->r_reply->front_alloc_len) {
2810 pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n",
2811 front_len, req->r_reply->front_alloc_len,
2812 (unsigned int)con->peer_name.type,
2813 le64_to_cpu(con->peer_name.num));
2814 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2818 ceph_msg_put(req->r_reply);
2821 m = ceph_msg_get(req->r_reply);
2824 struct ceph_osd_data *osd_data;
2827 * XXX This is assuming there is only one op containing
2828 * XXX page data. Probably OK for reads, but this
2829 * XXX ought to be done more generally.
2831 osd_data = osd_req_op_extent_osd_data(req, 0);
2832 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2833 if (osd_data->pages &&
2834 unlikely(osd_data->length < data_len)) {
2836 pr_warning("tid %lld reply has %d bytes "
2837 "we had only %llu bytes ready\n",
2838 tid, data_len, osd_data->length);
2847 dout("get_reply tid %lld %p\n", tid, m);
2850 mutex_unlock(&osdc->request_mutex);
2855 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2856 struct ceph_msg_header *hdr,
2859 struct ceph_osd *osd = con->private;
2860 int type = le16_to_cpu(hdr->type);
2861 int front = le32_to_cpu(hdr->front_len);
2865 case CEPH_MSG_OSD_MAP:
2866 case CEPH_MSG_WATCH_NOTIFY:
2867 return ceph_msg_new(type, front, GFP_NOFS, false);
2868 case CEPH_MSG_OSD_OPREPLY:
2869 return get_reply(con, hdr, skip);
2871 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2879 * Wrappers to refcount containing ceph_osd struct
2881 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2883 struct ceph_osd *osd = con->private;
2889 static void put_osd_con(struct ceph_connection *con)
2891 struct ceph_osd *osd = con->private;
2899 * Note: returned pointer is the address of a structure that's
2900 * managed separately. Caller must *not* attempt to free it.
2902 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2903 int *proto, int force_new)
2905 struct ceph_osd *o = con->private;
2906 struct ceph_osd_client *osdc = o->o_osdc;
2907 struct ceph_auth_client *ac = osdc->client->monc.auth;
2908 struct ceph_auth_handshake *auth = &o->o_auth;
2910 if (force_new && auth->authorizer) {
2911 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2912 auth->authorizer = NULL;
2914 if (!auth->authorizer) {
2915 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2918 return ERR_PTR(ret);
2920 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2923 return ERR_PTR(ret);
2925 *proto = ac->protocol;
2931 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2933 struct ceph_osd *o = con->private;
2934 struct ceph_osd_client *osdc = o->o_osdc;
2935 struct ceph_auth_client *ac = osdc->client->monc.auth;
2937 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2940 static int invalidate_authorizer(struct ceph_connection *con)
2942 struct ceph_osd *o = con->private;
2943 struct ceph_osd_client *osdc = o->o_osdc;
2944 struct ceph_auth_client *ac = osdc->client->monc.auth;
2946 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2947 return ceph_monc_validate_auth(&osdc->client->monc);
2950 static const struct ceph_connection_operations osd_con_ops = {
2953 .dispatch = dispatch,
2954 .get_authorizer = get_authorizer,
2955 .verify_authorizer_reply = verify_authorizer_reply,
2956 .invalidate_authorizer = invalidate_authorizer,
2957 .alloc_msg = alloc_msg,