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 void ceph_osdc_release_request(struct kref *kref)
302 struct ceph_osd_request *req;
305 req = container_of(kref, struct ceph_osd_request, r_kref);
307 ceph_msg_put(req->r_request);
309 ceph_msg_revoke_incoming(req->r_reply);
310 ceph_msg_put(req->r_reply);
313 for (which = 0; which < req->r_num_ops; which++)
314 osd_req_op_data_release(req, which);
316 ceph_put_snap_context(req->r_snapc);
318 mempool_free(req, req->r_osdc->req_mempool);
320 kmem_cache_free(ceph_osd_request_cache, req);
323 EXPORT_SYMBOL(ceph_osdc_release_request);
325 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
326 struct ceph_snap_context *snapc,
327 unsigned int num_ops,
331 struct ceph_osd_request *req;
332 struct ceph_msg *msg;
335 BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
336 BUG_ON(num_ops > CEPH_OSD_MAX_OP);
338 msg_size = 4 + 4 + 8 + 8 + 4+8;
339 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
340 msg_size += 1 + 8 + 4 + 4; /* pg_t */
341 msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
342 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
343 msg_size += 8; /* snapid */
344 msg_size += 8; /* snap_seq */
345 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
349 req = mempool_alloc(osdc->req_mempool, gfp_flags);
350 memset(req, 0, sizeof(*req));
352 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
358 req->r_mempool = use_mempool;
359 req->r_num_ops = num_ops;
361 kref_init(&req->r_kref);
362 init_completion(&req->r_completion);
363 init_completion(&req->r_safe_completion);
364 RB_CLEAR_NODE(&req->r_node);
365 INIT_LIST_HEAD(&req->r_unsafe_item);
366 INIT_LIST_HEAD(&req->r_linger_item);
367 INIT_LIST_HEAD(&req->r_linger_osd);
368 INIT_LIST_HEAD(&req->r_req_lru_item);
369 INIT_LIST_HEAD(&req->r_osd_item);
371 req->r_base_oloc.pool = -1;
372 req->r_target_oloc.pool = -1;
374 /* create reply message */
376 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
378 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
379 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
381 ceph_osdc_put_request(req);
386 /* create request message; allow space for oid */
388 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
390 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
392 ceph_osdc_put_request(req);
396 memset(msg->front.iov_base, 0, msg->front.iov_len);
398 req->r_request = msg;
402 EXPORT_SYMBOL(ceph_osdc_alloc_request);
404 static bool osd_req_opcode_valid(u16 opcode)
407 case CEPH_OSD_OP_READ:
408 case CEPH_OSD_OP_STAT:
409 case CEPH_OSD_OP_MAPEXT:
410 case CEPH_OSD_OP_MASKTRUNC:
411 case CEPH_OSD_OP_SPARSE_READ:
412 case CEPH_OSD_OP_NOTIFY:
413 case CEPH_OSD_OP_NOTIFY_ACK:
414 case CEPH_OSD_OP_ASSERT_VER:
415 case CEPH_OSD_OP_WRITE:
416 case CEPH_OSD_OP_WRITEFULL:
417 case CEPH_OSD_OP_TRUNCATE:
418 case CEPH_OSD_OP_ZERO:
419 case CEPH_OSD_OP_DELETE:
420 case CEPH_OSD_OP_APPEND:
421 case CEPH_OSD_OP_STARTSYNC:
422 case CEPH_OSD_OP_SETTRUNC:
423 case CEPH_OSD_OP_TRIMTRUNC:
424 case CEPH_OSD_OP_TMAPUP:
425 case CEPH_OSD_OP_TMAPPUT:
426 case CEPH_OSD_OP_TMAPGET:
427 case CEPH_OSD_OP_CREATE:
428 case CEPH_OSD_OP_ROLLBACK:
429 case CEPH_OSD_OP_WATCH:
430 case CEPH_OSD_OP_OMAPGETKEYS:
431 case CEPH_OSD_OP_OMAPGETVALS:
432 case CEPH_OSD_OP_OMAPGETHEADER:
433 case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
434 case CEPH_OSD_OP_OMAPSETVALS:
435 case CEPH_OSD_OP_OMAPSETHEADER:
436 case CEPH_OSD_OP_OMAPCLEAR:
437 case CEPH_OSD_OP_OMAPRMKEYS:
438 case CEPH_OSD_OP_OMAP_CMP:
439 case CEPH_OSD_OP_CLONERANGE:
440 case CEPH_OSD_OP_ASSERT_SRC_VERSION:
441 case CEPH_OSD_OP_SRC_CMPXATTR:
442 case CEPH_OSD_OP_GETXATTR:
443 case CEPH_OSD_OP_GETXATTRS:
444 case CEPH_OSD_OP_CMPXATTR:
445 case CEPH_OSD_OP_SETXATTR:
446 case CEPH_OSD_OP_SETXATTRS:
447 case CEPH_OSD_OP_RESETXATTRS:
448 case CEPH_OSD_OP_RMXATTR:
449 case CEPH_OSD_OP_PULL:
450 case CEPH_OSD_OP_PUSH:
451 case CEPH_OSD_OP_BALANCEREADS:
452 case CEPH_OSD_OP_UNBALANCEREADS:
453 case CEPH_OSD_OP_SCRUB:
454 case CEPH_OSD_OP_SCRUB_RESERVE:
455 case CEPH_OSD_OP_SCRUB_UNRESERVE:
456 case CEPH_OSD_OP_SCRUB_STOP:
457 case CEPH_OSD_OP_SCRUB_MAP:
458 case CEPH_OSD_OP_WRLOCK:
459 case CEPH_OSD_OP_WRUNLOCK:
460 case CEPH_OSD_OP_RDLOCK:
461 case CEPH_OSD_OP_RDUNLOCK:
462 case CEPH_OSD_OP_UPLOCK:
463 case CEPH_OSD_OP_DNLOCK:
464 case CEPH_OSD_OP_CALL:
465 case CEPH_OSD_OP_PGLS:
466 case CEPH_OSD_OP_PGLS_FILTER:
474 * This is an osd op init function for opcodes that have no data or
475 * other information associated with them. It also serves as a
476 * common init routine for all the other init functions, below.
478 static struct ceph_osd_req_op *
479 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
482 struct ceph_osd_req_op *op;
484 BUG_ON(which >= osd_req->r_num_ops);
485 BUG_ON(!osd_req_opcode_valid(opcode));
487 op = &osd_req->r_ops[which];
488 memset(op, 0, sizeof (*op));
494 void osd_req_op_init(struct ceph_osd_request *osd_req,
495 unsigned int which, u16 opcode)
497 (void)_osd_req_op_init(osd_req, which, opcode);
499 EXPORT_SYMBOL(osd_req_op_init);
501 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
502 unsigned int which, u16 opcode,
503 u64 offset, u64 length,
504 u64 truncate_size, u32 truncate_seq)
506 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
507 size_t payload_len = 0;
509 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
510 opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
511 opcode != CEPH_OSD_OP_TRUNCATE);
513 op->extent.offset = offset;
514 op->extent.length = length;
515 op->extent.truncate_size = truncate_size;
516 op->extent.truncate_seq = truncate_seq;
517 if (opcode == CEPH_OSD_OP_WRITE)
518 payload_len += length;
520 op->payload_len = payload_len;
522 EXPORT_SYMBOL(osd_req_op_extent_init);
524 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
525 unsigned int which, u64 length)
527 struct ceph_osd_req_op *op;
530 BUG_ON(which >= osd_req->r_num_ops);
531 op = &osd_req->r_ops[which];
532 previous = op->extent.length;
534 if (length == previous)
535 return; /* Nothing to do */
536 BUG_ON(length > previous);
538 op->extent.length = length;
539 op->payload_len -= previous - length;
541 EXPORT_SYMBOL(osd_req_op_extent_update);
543 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
544 u16 opcode, const char *class, const char *method)
546 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
547 struct ceph_pagelist *pagelist;
548 size_t payload_len = 0;
551 BUG_ON(opcode != CEPH_OSD_OP_CALL);
553 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
555 ceph_pagelist_init(pagelist);
557 op->cls.class_name = class;
558 size = strlen(class);
559 BUG_ON(size > (size_t) U8_MAX);
560 op->cls.class_len = size;
561 ceph_pagelist_append(pagelist, class, size);
564 op->cls.method_name = method;
565 size = strlen(method);
566 BUG_ON(size > (size_t) U8_MAX);
567 op->cls.method_len = size;
568 ceph_pagelist_append(pagelist, method, size);
571 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
573 op->cls.argc = 0; /* currently unused */
575 op->payload_len = payload_len;
577 EXPORT_SYMBOL(osd_req_op_cls_init);
579 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
580 unsigned int which, u16 opcode,
581 u64 cookie, u64 version, int flag)
583 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
585 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
587 op->watch.cookie = cookie;
588 op->watch.ver = version;
589 if (opcode == CEPH_OSD_OP_WATCH && flag)
590 op->watch.flag = (u8)1;
592 EXPORT_SYMBOL(osd_req_op_watch_init);
594 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
595 struct ceph_osd_data *osd_data)
597 u64 length = ceph_osd_data_length(osd_data);
599 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
600 BUG_ON(length > (u64) SIZE_MAX);
602 ceph_msg_data_add_pages(msg, osd_data->pages,
603 length, osd_data->alignment);
604 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
606 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
608 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
609 ceph_msg_data_add_bio(msg, osd_data->bio, length);
612 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
616 static u64 osd_req_encode_op(struct ceph_osd_request *req,
617 struct ceph_osd_op *dst, unsigned int which)
619 struct ceph_osd_req_op *src;
620 struct ceph_osd_data *osd_data;
621 u64 request_data_len = 0;
624 BUG_ON(which >= req->r_num_ops);
625 src = &req->r_ops[which];
626 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
627 pr_err("unrecognized osd opcode %d\n", src->op);
633 case CEPH_OSD_OP_STAT:
634 osd_data = &src->raw_data_in;
635 ceph_osdc_msg_data_add(req->r_reply, osd_data);
637 case CEPH_OSD_OP_READ:
638 case CEPH_OSD_OP_WRITE:
639 case CEPH_OSD_OP_ZERO:
640 case CEPH_OSD_OP_DELETE:
641 case CEPH_OSD_OP_TRUNCATE:
642 if (src->op == CEPH_OSD_OP_WRITE)
643 request_data_len = src->extent.length;
644 dst->extent.offset = cpu_to_le64(src->extent.offset);
645 dst->extent.length = cpu_to_le64(src->extent.length);
646 dst->extent.truncate_size =
647 cpu_to_le64(src->extent.truncate_size);
648 dst->extent.truncate_seq =
649 cpu_to_le32(src->extent.truncate_seq);
650 osd_data = &src->extent.osd_data;
651 if (src->op == CEPH_OSD_OP_WRITE)
652 ceph_osdc_msg_data_add(req->r_request, osd_data);
654 ceph_osdc_msg_data_add(req->r_reply, osd_data);
656 case CEPH_OSD_OP_CALL:
657 dst->cls.class_len = src->cls.class_len;
658 dst->cls.method_len = src->cls.method_len;
659 osd_data = &src->cls.request_info;
660 ceph_osdc_msg_data_add(req->r_request, osd_data);
661 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
662 request_data_len = osd_data->pagelist->length;
664 osd_data = &src->cls.request_data;
665 data_length = ceph_osd_data_length(osd_data);
667 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
668 dst->cls.indata_len = cpu_to_le32(data_length);
669 ceph_osdc_msg_data_add(req->r_request, osd_data);
670 src->payload_len += data_length;
671 request_data_len += data_length;
673 osd_data = &src->cls.response_data;
674 ceph_osdc_msg_data_add(req->r_reply, osd_data);
676 case CEPH_OSD_OP_STARTSYNC:
678 case CEPH_OSD_OP_NOTIFY_ACK:
679 case CEPH_OSD_OP_WATCH:
680 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
681 dst->watch.ver = cpu_to_le64(src->watch.ver);
682 dst->watch.flag = src->watch.flag;
685 pr_err("unsupported osd opcode %s\n",
686 ceph_osd_op_name(src->op));
691 dst->op = cpu_to_le16(src->op);
692 dst->payload_len = cpu_to_le32(src->payload_len);
694 return request_data_len;
698 * build new request AND message, calculate layout, and adjust file
701 * if the file was recently truncated, we include information about its
702 * old and new size so that the object can be updated appropriately. (we
703 * avoid synchronously deleting truncated objects because it's slow.)
705 * if @do_sync, include a 'startsync' command so that the osd will flush
708 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
709 struct ceph_file_layout *layout,
710 struct ceph_vino vino,
711 u64 off, u64 *plen, int num_ops,
712 int opcode, int flags,
713 struct ceph_snap_context *snapc,
718 struct ceph_osd_request *req;
726 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
727 opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
728 opcode != CEPH_OSD_OP_TRUNCATE);
730 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
733 return ERR_PTR(-ENOMEM);
735 req->r_flags = flags;
737 /* calculate max write size */
738 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
740 ceph_osdc_put_request(req);
744 object_size = le32_to_cpu(layout->fl_object_size);
745 object_base = off - objoff;
746 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
747 if (truncate_size <= object_base) {
750 truncate_size -= object_base;
751 if (truncate_size > object_size)
752 truncate_size = object_size;
756 osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
757 truncate_size, truncate_seq);
760 * A second op in the ops array means the caller wants to
761 * also issue a include a 'startsync' command so that the
762 * osd will flush data quickly.
765 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
767 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
769 snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
770 "%llx.%08llx", vino.ino, objnum);
771 req->r_base_oid.name_len = strlen(req->r_base_oid.name);
775 EXPORT_SYMBOL(ceph_osdc_new_request);
778 * We keep osd requests in an rbtree, sorted by ->r_tid.
780 static void __insert_request(struct ceph_osd_client *osdc,
781 struct ceph_osd_request *new)
783 struct rb_node **p = &osdc->requests.rb_node;
784 struct rb_node *parent = NULL;
785 struct ceph_osd_request *req = NULL;
789 req = rb_entry(parent, struct ceph_osd_request, r_node);
790 if (new->r_tid < req->r_tid)
792 else if (new->r_tid > req->r_tid)
798 rb_link_node(&new->r_node, parent, p);
799 rb_insert_color(&new->r_node, &osdc->requests);
802 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
805 struct ceph_osd_request *req;
806 struct rb_node *n = osdc->requests.rb_node;
809 req = rb_entry(n, struct ceph_osd_request, r_node);
810 if (tid < req->r_tid)
812 else if (tid > req->r_tid)
820 static struct ceph_osd_request *
821 __lookup_request_ge(struct ceph_osd_client *osdc,
824 struct ceph_osd_request *req;
825 struct rb_node *n = osdc->requests.rb_node;
828 req = rb_entry(n, struct ceph_osd_request, r_node);
829 if (tid < req->r_tid) {
833 } else if (tid > req->r_tid) {
843 * Resubmit requests pending on the given osd.
845 static void __kick_osd_requests(struct ceph_osd_client *osdc,
846 struct ceph_osd *osd)
848 struct ceph_osd_request *req, *nreq;
852 dout("__kick_osd_requests osd%d\n", osd->o_osd);
853 err = __reset_osd(osdc, osd);
857 * Build up a list of requests to resend by traversing the
858 * osd's list of requests. Requests for a given object are
859 * sent in tid order, and that is also the order they're
860 * kept on this list. Therefore all requests that are in
861 * flight will be found first, followed by all requests that
862 * have not yet been sent. And to resend requests while
863 * preserving this order we will want to put any sent
864 * requests back on the front of the osd client's unsent
867 * So we build a separate ordered list of already-sent
868 * requests for the affected osd and splice it onto the
869 * front of the osd client's unsent list. Once we've seen a
870 * request that has not yet been sent we're done. Those
871 * requests are already sitting right where they belong.
873 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
876 list_move_tail(&req->r_req_lru_item, &resend);
877 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
880 req->r_flags |= CEPH_OSD_FLAG_RETRY;
882 list_splice(&resend, &osdc->req_unsent);
885 * Linger requests are re-registered before sending, which
886 * sets up a new tid for each. We add them to the unsent
887 * list at the end to keep things in tid order.
889 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
892 * reregister request prior to unregistering linger so
893 * that r_osd is preserved.
895 BUG_ON(!list_empty(&req->r_req_lru_item));
896 __register_request(osdc, req);
897 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
898 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
899 __unregister_linger_request(osdc, req);
900 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
906 * If the osd connection drops, we need to resubmit all requests.
908 static void osd_reset(struct ceph_connection *con)
910 struct ceph_osd *osd = con->private;
911 struct ceph_osd_client *osdc;
915 dout("osd_reset osd%d\n", osd->o_osd);
917 down_read(&osdc->map_sem);
918 mutex_lock(&osdc->request_mutex);
919 __kick_osd_requests(osdc, osd);
921 mutex_unlock(&osdc->request_mutex);
922 up_read(&osdc->map_sem);
926 * Track open sessions with osds.
928 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
930 struct ceph_osd *osd;
932 osd = kzalloc(sizeof(*osd), GFP_NOFS);
936 atomic_set(&osd->o_ref, 1);
939 RB_CLEAR_NODE(&osd->o_node);
940 INIT_LIST_HEAD(&osd->o_requests);
941 INIT_LIST_HEAD(&osd->o_linger_requests);
942 INIT_LIST_HEAD(&osd->o_osd_lru);
943 osd->o_incarnation = 1;
945 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
947 INIT_LIST_HEAD(&osd->o_keepalive_item);
951 static struct ceph_osd *get_osd(struct ceph_osd *osd)
953 if (atomic_inc_not_zero(&osd->o_ref)) {
954 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
955 atomic_read(&osd->o_ref));
958 dout("get_osd %p FAIL\n", osd);
963 static void put_osd(struct ceph_osd *osd)
965 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
966 atomic_read(&osd->o_ref) - 1);
967 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
968 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
970 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
976 * remove an osd from our map
978 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
980 dout("__remove_osd %p\n", osd);
981 BUG_ON(!list_empty(&osd->o_requests));
982 rb_erase(&osd->o_node, &osdc->osds);
983 list_del_init(&osd->o_osd_lru);
984 ceph_con_close(&osd->o_con);
988 static void remove_all_osds(struct ceph_osd_client *osdc)
990 dout("%s %p\n", __func__, osdc);
991 mutex_lock(&osdc->request_mutex);
992 while (!RB_EMPTY_ROOT(&osdc->osds)) {
993 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
994 struct ceph_osd, o_node);
995 __remove_osd(osdc, osd);
997 mutex_unlock(&osdc->request_mutex);
1000 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1001 struct ceph_osd *osd)
1003 dout("__move_osd_to_lru %p\n", osd);
1004 BUG_ON(!list_empty(&osd->o_osd_lru));
1005 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1006 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1009 static void __remove_osd_from_lru(struct ceph_osd *osd)
1011 dout("__remove_osd_from_lru %p\n", osd);
1012 if (!list_empty(&osd->o_osd_lru))
1013 list_del_init(&osd->o_osd_lru);
1016 static void remove_old_osds(struct ceph_osd_client *osdc)
1018 struct ceph_osd *osd, *nosd;
1020 dout("__remove_old_osds %p\n", osdc);
1021 mutex_lock(&osdc->request_mutex);
1022 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1023 if (time_before(jiffies, osd->lru_ttl))
1025 __remove_osd(osdc, osd);
1027 mutex_unlock(&osdc->request_mutex);
1033 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1035 struct ceph_entity_addr *peer_addr;
1037 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1038 if (list_empty(&osd->o_requests) &&
1039 list_empty(&osd->o_linger_requests)) {
1040 __remove_osd(osdc, osd);
1045 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1046 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1047 !ceph_con_opened(&osd->o_con)) {
1048 struct ceph_osd_request *req;
1050 dout("osd addr hasn't changed and connection never opened, "
1051 "letting msgr retry\n");
1052 /* touch each r_stamp for handle_timeout()'s benfit */
1053 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1054 req->r_stamp = jiffies;
1059 ceph_con_close(&osd->o_con);
1060 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1061 osd->o_incarnation++;
1066 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1068 struct rb_node **p = &osdc->osds.rb_node;
1069 struct rb_node *parent = NULL;
1070 struct ceph_osd *osd = NULL;
1072 dout("__insert_osd %p osd%d\n", new, new->o_osd);
1075 osd = rb_entry(parent, struct ceph_osd, o_node);
1076 if (new->o_osd < osd->o_osd)
1078 else if (new->o_osd > osd->o_osd)
1079 p = &(*p)->rb_right;
1084 rb_link_node(&new->o_node, parent, p);
1085 rb_insert_color(&new->o_node, &osdc->osds);
1088 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1090 struct ceph_osd *osd;
1091 struct rb_node *n = osdc->osds.rb_node;
1094 osd = rb_entry(n, struct ceph_osd, o_node);
1097 else if (o > osd->o_osd)
1105 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1107 schedule_delayed_work(&osdc->timeout_work,
1108 osdc->client->options->osd_keepalive_timeout * HZ);
1111 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1113 cancel_delayed_work(&osdc->timeout_work);
1117 * Register request, assign tid. If this is the first request, set up
1118 * the timeout event.
1120 static void __register_request(struct ceph_osd_client *osdc,
1121 struct ceph_osd_request *req)
1123 req->r_tid = ++osdc->last_tid;
1124 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1125 dout("__register_request %p tid %lld\n", req, req->r_tid);
1126 __insert_request(osdc, req);
1127 ceph_osdc_get_request(req);
1128 osdc->num_requests++;
1129 if (osdc->num_requests == 1) {
1130 dout(" first request, scheduling timeout\n");
1131 __schedule_osd_timeout(osdc);
1136 * called under osdc->request_mutex
1138 static void __unregister_request(struct ceph_osd_client *osdc,
1139 struct ceph_osd_request *req)
1141 if (RB_EMPTY_NODE(&req->r_node)) {
1142 dout("__unregister_request %p tid %lld not registered\n",
1147 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1148 rb_erase(&req->r_node, &osdc->requests);
1149 osdc->num_requests--;
1152 /* make sure the original request isn't in flight. */
1153 ceph_msg_revoke(req->r_request);
1155 list_del_init(&req->r_osd_item);
1156 if (list_empty(&req->r_osd->o_requests) &&
1157 list_empty(&req->r_osd->o_linger_requests)) {
1158 dout("moving osd to %p lru\n", req->r_osd);
1159 __move_osd_to_lru(osdc, req->r_osd);
1161 if (list_empty(&req->r_linger_item))
1165 list_del_init(&req->r_req_lru_item);
1166 ceph_osdc_put_request(req);
1168 if (osdc->num_requests == 0) {
1169 dout(" no requests, canceling timeout\n");
1170 __cancel_osd_timeout(osdc);
1175 * Cancel a previously queued request message
1177 static void __cancel_request(struct ceph_osd_request *req)
1179 if (req->r_sent && req->r_osd) {
1180 ceph_msg_revoke(req->r_request);
1185 static void __register_linger_request(struct ceph_osd_client *osdc,
1186 struct ceph_osd_request *req)
1188 dout("__register_linger_request %p\n", req);
1189 ceph_osdc_get_request(req);
1190 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1192 list_add_tail(&req->r_linger_osd,
1193 &req->r_osd->o_linger_requests);
1196 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1197 struct ceph_osd_request *req)
1199 dout("__unregister_linger_request %p\n", req);
1200 list_del_init(&req->r_linger_item);
1202 list_del_init(&req->r_linger_osd);
1204 if (list_empty(&req->r_osd->o_requests) &&
1205 list_empty(&req->r_osd->o_linger_requests)) {
1206 dout("moving osd to %p lru\n", req->r_osd);
1207 __move_osd_to_lru(osdc, req->r_osd);
1209 if (list_empty(&req->r_osd_item))
1212 ceph_osdc_put_request(req);
1215 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1216 struct ceph_osd_request *req)
1218 mutex_lock(&osdc->request_mutex);
1219 if (req->r_linger) {
1221 __unregister_linger_request(osdc, req);
1223 mutex_unlock(&osdc->request_mutex);
1225 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1227 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1228 struct ceph_osd_request *req)
1230 if (!req->r_linger) {
1231 dout("set_request_linger %p\n", req);
1235 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1238 * Returns whether a request should be blocked from being sent
1239 * based on the current osdmap and osd_client settings.
1241 * Caller should hold map_sem for read.
1243 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1244 struct ceph_osd_request *req)
1246 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1247 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1248 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1249 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1250 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1254 * Calculate mapping of a request to a PG. Takes tiering into account.
1256 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1257 struct ceph_osd_request *req,
1258 struct ceph_pg *pg_out)
1260 bool need_check_tiering;
1262 need_check_tiering = false;
1263 if (req->r_target_oloc.pool == -1) {
1264 req->r_target_oloc = req->r_base_oloc; /* struct */
1265 need_check_tiering = true;
1267 if (req->r_target_oid.name_len == 0) {
1268 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1269 need_check_tiering = true;
1272 if (need_check_tiering &&
1273 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1274 struct ceph_pg_pool_info *pi;
1276 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1278 if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1280 req->r_target_oloc.pool = pi->read_tier;
1281 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1282 pi->write_tier >= 0)
1283 req->r_target_oloc.pool = pi->write_tier;
1285 /* !pi is caught in ceph_oloc_oid_to_pg() */
1288 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1289 &req->r_target_oid, pg_out);
1293 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1294 * (as needed), and set the request r_osd appropriately. If there is
1295 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1296 * (unsent, homeless) or leave on in-flight lru.
1298 * Return 0 if unchanged, 1 if changed, or negative on error.
1300 * Caller should hold map_sem for read and request_mutex.
1302 static int __map_request(struct ceph_osd_client *osdc,
1303 struct ceph_osd_request *req, int force_resend)
1305 struct ceph_pg pgid;
1306 int acting[CEPH_PG_MAX_SIZE];
1307 int o = -1, num = 0;
1311 dout("map_request %p tid %lld\n", req, req->r_tid);
1313 err = __calc_request_pg(osdc->osdmap, req, &pgid);
1315 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1320 err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
1326 was_paused = req->r_paused;
1327 req->r_paused = __req_should_be_paused(osdc, req);
1328 if (was_paused && !req->r_paused)
1331 if ((!force_resend &&
1332 req->r_osd && req->r_osd->o_osd == o &&
1333 req->r_sent >= req->r_osd->o_incarnation &&
1334 req->r_num_pg_osds == num &&
1335 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1336 (req->r_osd == NULL && o == -1) ||
1338 return 0; /* no change */
1340 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1341 req->r_tid, pgid.pool, pgid.seed, o,
1342 req->r_osd ? req->r_osd->o_osd : -1);
1344 /* record full pg acting set */
1345 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1346 req->r_num_pg_osds = num;
1349 __cancel_request(req);
1350 list_del_init(&req->r_osd_item);
1354 req->r_osd = __lookup_osd(osdc, o);
1355 if (!req->r_osd && o >= 0) {
1357 req->r_osd = create_osd(osdc, o);
1359 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1363 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1364 __insert_osd(osdc, req->r_osd);
1366 ceph_con_open(&req->r_osd->o_con,
1367 CEPH_ENTITY_TYPE_OSD, o,
1368 &osdc->osdmap->osd_addr[o]);
1372 __remove_osd_from_lru(req->r_osd);
1373 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1374 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1376 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1378 err = 1; /* osd or pg changed */
1385 * caller should hold map_sem (for read) and request_mutex
1387 static void __send_request(struct ceph_osd_client *osdc,
1388 struct ceph_osd_request *req)
1392 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1393 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1394 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1396 /* fill in message content that changes each time we send it */
1397 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1398 put_unaligned_le32(req->r_flags, req->r_request_flags);
1399 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1400 p = req->r_request_pgid;
1401 ceph_encode_64(&p, req->r_pgid.pool);
1402 ceph_encode_32(&p, req->r_pgid.seed);
1403 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1404 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1405 sizeof(req->r_reassert_version));
1407 req->r_stamp = jiffies;
1408 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1410 ceph_msg_get(req->r_request); /* send consumes a ref */
1412 req->r_sent = req->r_osd->o_incarnation;
1414 ceph_con_send(&req->r_osd->o_con, req->r_request);
1418 * Send any requests in the queue (req_unsent).
1420 static void __send_queued(struct ceph_osd_client *osdc)
1422 struct ceph_osd_request *req, *tmp;
1424 dout("__send_queued\n");
1425 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1426 __send_request(osdc, req);
1430 * Timeout callback, called every N seconds when 1 or more osd
1431 * requests has been active for more than N seconds. When this
1432 * happens, we ping all OSDs with requests who have timed out to
1433 * ensure any communications channel reset is detected. Reset the
1434 * request timeouts another N seconds in the future as we go.
1435 * Reschedule the timeout event another N seconds in future (unless
1436 * there are no open requests).
1438 static void handle_timeout(struct work_struct *work)
1440 struct ceph_osd_client *osdc =
1441 container_of(work, struct ceph_osd_client, timeout_work.work);
1442 struct ceph_osd_request *req;
1443 struct ceph_osd *osd;
1444 unsigned long keepalive =
1445 osdc->client->options->osd_keepalive_timeout * HZ;
1446 struct list_head slow_osds;
1448 down_read(&osdc->map_sem);
1450 ceph_monc_request_next_osdmap(&osdc->client->monc);
1452 mutex_lock(&osdc->request_mutex);
1455 * ping osds that are a bit slow. this ensures that if there
1456 * is a break in the TCP connection we will notice, and reopen
1457 * a connection with that osd (from the fault callback).
1459 INIT_LIST_HEAD(&slow_osds);
1460 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1461 if (time_before(jiffies, req->r_stamp + keepalive))
1466 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1467 req->r_tid, osd->o_osd);
1468 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1470 while (!list_empty(&slow_osds)) {
1471 osd = list_entry(slow_osds.next, struct ceph_osd,
1473 list_del_init(&osd->o_keepalive_item);
1474 ceph_con_keepalive(&osd->o_con);
1477 __schedule_osd_timeout(osdc);
1478 __send_queued(osdc);
1479 mutex_unlock(&osdc->request_mutex);
1480 up_read(&osdc->map_sem);
1483 static void handle_osds_timeout(struct work_struct *work)
1485 struct ceph_osd_client *osdc =
1486 container_of(work, struct ceph_osd_client,
1487 osds_timeout_work.work);
1488 unsigned long delay =
1489 osdc->client->options->osd_idle_ttl * HZ >> 2;
1491 dout("osds timeout\n");
1492 down_read(&osdc->map_sem);
1493 remove_old_osds(osdc);
1494 up_read(&osdc->map_sem);
1496 schedule_delayed_work(&osdc->osds_timeout_work,
1497 round_jiffies_relative(delay));
1500 static int ceph_oloc_decode(void **p, void *end,
1501 struct ceph_object_locator *oloc)
1503 u8 struct_v, struct_cv;
1508 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1509 struct_v = ceph_decode_8(p);
1510 struct_cv = ceph_decode_8(p);
1512 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1513 struct_v, struct_cv);
1516 if (struct_cv > 6) {
1517 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1518 struct_v, struct_cv);
1521 len = ceph_decode_32(p);
1522 ceph_decode_need(p, end, len, e_inval);
1523 struct_end = *p + len;
1525 oloc->pool = ceph_decode_64(p);
1526 *p += 4; /* skip preferred */
1528 len = ceph_decode_32(p);
1530 pr_warn("ceph_object_locator::key is set\n");
1534 if (struct_v >= 5) {
1535 len = ceph_decode_32(p);
1537 pr_warn("ceph_object_locator::nspace is set\n");
1542 if (struct_v >= 6) {
1543 s64 hash = ceph_decode_64(p);
1545 pr_warn("ceph_object_locator::hash is set\n");
1560 static int ceph_redirect_decode(void **p, void *end,
1561 struct ceph_request_redirect *redir)
1563 u8 struct_v, struct_cv;
1568 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1569 struct_v = ceph_decode_8(p);
1570 struct_cv = ceph_decode_8(p);
1571 if (struct_cv > 1) {
1572 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1573 struct_v, struct_cv);
1576 len = ceph_decode_32(p);
1577 ceph_decode_need(p, end, len, e_inval);
1578 struct_end = *p + len;
1580 ret = ceph_oloc_decode(p, end, &redir->oloc);
1584 len = ceph_decode_32(p);
1586 pr_warn("ceph_request_redirect::object_name is set\n");
1590 len = ceph_decode_32(p);
1591 *p += len; /* skip osd_instructions */
1603 static void complete_request(struct ceph_osd_request *req)
1605 complete_all(&req->r_safe_completion); /* fsync waiter */
1609 * handle osd op reply. either call the callback if it is specified,
1610 * or do the completion to wake up the waiting thread.
1612 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1613 struct ceph_connection *con)
1616 struct ceph_osd_request *req;
1617 struct ceph_request_redirect redir;
1620 unsigned int numops;
1621 int payload_len, flags;
1627 u64 reassert_version;
1629 int already_completed;
1633 tid = le64_to_cpu(msg->hdr.tid);
1634 dout("handle_reply %p tid %llu\n", msg, tid);
1636 p = msg->front.iov_base;
1637 end = p + msg->front.iov_len;
1639 ceph_decode_need(&p, end, 4, bad);
1640 object_len = ceph_decode_32(&p);
1641 ceph_decode_need(&p, end, object_len, bad);
1644 err = ceph_decode_pgid(&p, end, &pg);
1648 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1649 flags = ceph_decode_64(&p);
1650 result = ceph_decode_32(&p);
1651 reassert_epoch = ceph_decode_32(&p);
1652 reassert_version = ceph_decode_64(&p);
1653 osdmap_epoch = ceph_decode_32(&p);
1656 mutex_lock(&osdc->request_mutex);
1657 req = __lookup_request(osdc, tid);
1659 dout("handle_reply tid %llu dne\n", tid);
1662 ceph_osdc_get_request(req);
1664 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1667 ceph_decode_need(&p, end, 4, bad_put);
1668 numops = ceph_decode_32(&p);
1669 if (numops > CEPH_OSD_MAX_OP)
1671 if (numops != req->r_num_ops)
1674 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1675 for (i = 0; i < numops; i++) {
1676 struct ceph_osd_op *op = p;
1679 len = le32_to_cpu(op->payload_len);
1680 req->r_reply_op_len[i] = len;
1681 dout(" op %d has %d bytes\n", i, len);
1685 bytes = le32_to_cpu(msg->hdr.data_len);
1686 if (payload_len != bytes) {
1687 pr_warning("sum of op payload lens %d != data_len %d",
1688 payload_len, bytes);
1692 ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1693 retry_attempt = ceph_decode_32(&p);
1694 for (i = 0; i < numops; i++)
1695 req->r_reply_op_result[i] = ceph_decode_32(&p);
1697 if (le16_to_cpu(msg->hdr.version) >= 6) {
1698 p += 8 + 4; /* skip replay_version */
1699 p += 8; /* skip user_version */
1701 err = ceph_redirect_decode(&p, end, &redir);
1705 redir.oloc.pool = -1;
1708 if (redir.oloc.pool != -1) {
1709 dout("redirect pool %lld\n", redir.oloc.pool);
1711 __unregister_request(osdc, req);
1712 mutex_unlock(&osdc->request_mutex);
1714 req->r_target_oloc = redir.oloc; /* struct */
1717 * Start redirect requests with nofail=true. If
1718 * mapping fails, request will end up on the notarget
1719 * list, waiting for the new osdmap (which can take
1720 * a while), even though the original request mapped
1721 * successfully. In the future we might want to follow
1722 * original request's nofail setting here.
1724 err = ceph_osdc_start_request(osdc, req, true);
1730 already_completed = req->r_got_reply;
1731 if (!req->r_got_reply) {
1732 req->r_result = result;
1733 dout("handle_reply result %d bytes %d\n", req->r_result,
1735 if (req->r_result == 0)
1736 req->r_result = bytes;
1738 /* in case this is a write and we need to replay, */
1739 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1740 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1742 req->r_got_reply = 1;
1743 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1744 dout("handle_reply tid %llu dup ack\n", tid);
1745 mutex_unlock(&osdc->request_mutex);
1749 dout("handle_reply tid %llu flags %d\n", tid, flags);
1751 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1752 __register_linger_request(osdc, req);
1754 /* either this is a read, or we got the safe response */
1756 (flags & CEPH_OSD_FLAG_ONDISK) ||
1757 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1758 __unregister_request(osdc, req);
1760 mutex_unlock(&osdc->request_mutex);
1762 if (!already_completed) {
1763 if (req->r_unsafe_callback &&
1764 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1765 req->r_unsafe_callback(req, true);
1766 if (req->r_callback)
1767 req->r_callback(req, msg);
1769 complete_all(&req->r_completion);
1772 if (flags & CEPH_OSD_FLAG_ONDISK) {
1773 if (req->r_unsafe_callback && already_completed)
1774 req->r_unsafe_callback(req, false);
1775 complete_request(req);
1779 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1780 ceph_osdc_put_request(req);
1784 req->r_result = -EIO;
1785 __unregister_request(osdc, req);
1786 if (req->r_callback)
1787 req->r_callback(req, msg);
1789 complete_all(&req->r_completion);
1790 complete_request(req);
1791 ceph_osdc_put_request(req);
1793 mutex_unlock(&osdc->request_mutex);
1795 pr_err("corrupt osd_op_reply got %d %d\n",
1796 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1800 static void reset_changed_osds(struct ceph_osd_client *osdc)
1802 struct rb_node *p, *n;
1804 for (p = rb_first(&osdc->osds); p; p = n) {
1805 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1808 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1809 memcmp(&osd->o_con.peer_addr,
1810 ceph_osd_addr(osdc->osdmap,
1812 sizeof(struct ceph_entity_addr)) != 0)
1813 __reset_osd(osdc, osd);
1818 * Requeue requests whose mapping to an OSD has changed. If requests map to
1819 * no osd, request a new map.
1821 * Caller should hold map_sem for read.
1823 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1824 bool force_resend_writes)
1826 struct ceph_osd_request *req, *nreq;
1830 bool force_resend_req;
1832 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1833 force_resend_writes ? " (force resend writes)" : "");
1834 mutex_lock(&osdc->request_mutex);
1835 for (p = rb_first(&osdc->requests); p; ) {
1836 req = rb_entry(p, struct ceph_osd_request, r_node);
1840 * For linger requests that have not yet been
1841 * registered, move them to the linger list; they'll
1842 * be sent to the osd in the loop below. Unregister
1843 * the request before re-registering it as a linger
1844 * request to ensure the __map_request() below
1845 * will decide it needs to be sent.
1847 if (req->r_linger && list_empty(&req->r_linger_item)) {
1848 dout("%p tid %llu restart on osd%d\n",
1850 req->r_osd ? req->r_osd->o_osd : -1);
1851 ceph_osdc_get_request(req);
1852 __unregister_request(osdc, req);
1853 __register_linger_request(osdc, req);
1854 ceph_osdc_put_request(req);
1858 force_resend_req = force_resend ||
1859 (force_resend_writes &&
1860 req->r_flags & CEPH_OSD_FLAG_WRITE);
1861 err = __map_request(osdc, req, force_resend_req);
1863 continue; /* error */
1864 if (req->r_osd == NULL) {
1865 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1866 needmap++; /* request a newer map */
1867 } else if (err > 0) {
1868 if (!req->r_linger) {
1869 dout("%p tid %llu requeued on osd%d\n", req,
1871 req->r_osd ? req->r_osd->o_osd : -1);
1872 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1877 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1879 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1881 err = __map_request(osdc, req,
1882 force_resend || force_resend_writes);
1883 dout("__map_request returned %d\n", err);
1885 continue; /* no change and no osd was specified */
1887 continue; /* hrm! */
1888 if (req->r_osd == NULL) {
1889 dout("tid %llu maps to no valid osd\n", req->r_tid);
1890 needmap++; /* request a newer map */
1894 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1895 req->r_osd ? req->r_osd->o_osd : -1);
1896 __register_request(osdc, req);
1897 __unregister_linger_request(osdc, req);
1899 reset_changed_osds(osdc);
1900 mutex_unlock(&osdc->request_mutex);
1903 dout("%d requests for down osds, need new map\n", needmap);
1904 ceph_monc_request_next_osdmap(&osdc->client->monc);
1910 * Process updated osd map.
1912 * The message contains any number of incremental and full maps, normally
1913 * indicating some sort of topology change in the cluster. Kick requests
1914 * off to different OSDs as needed.
1916 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1918 void *p, *end, *next;
1919 u32 nr_maps, maplen;
1921 struct ceph_osdmap *newmap = NULL, *oldmap;
1923 struct ceph_fsid fsid;
1926 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1927 p = msg->front.iov_base;
1928 end = p + msg->front.iov_len;
1931 ceph_decode_need(&p, end, sizeof(fsid), bad);
1932 ceph_decode_copy(&p, &fsid, sizeof(fsid));
1933 if (ceph_check_fsid(osdc->client, &fsid) < 0)
1936 down_write(&osdc->map_sem);
1938 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1940 /* incremental maps */
1941 ceph_decode_32_safe(&p, end, nr_maps, bad);
1942 dout(" %d inc maps\n", nr_maps);
1943 while (nr_maps > 0) {
1944 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1945 epoch = ceph_decode_32(&p);
1946 maplen = ceph_decode_32(&p);
1947 ceph_decode_need(&p, end, maplen, bad);
1949 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1950 dout("applying incremental map %u len %d\n",
1952 newmap = osdmap_apply_incremental(&p, next,
1954 &osdc->client->msgr);
1955 if (IS_ERR(newmap)) {
1956 err = PTR_ERR(newmap);
1960 if (newmap != osdc->osdmap) {
1961 ceph_osdmap_destroy(osdc->osdmap);
1962 osdc->osdmap = newmap;
1964 was_full = was_full ||
1965 ceph_osdmap_flag(osdc->osdmap,
1967 kick_requests(osdc, 0, was_full);
1969 dout("ignoring incremental map %u len %d\n",
1979 ceph_decode_32_safe(&p, end, nr_maps, bad);
1980 dout(" %d full maps\n", nr_maps);
1982 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1983 epoch = ceph_decode_32(&p);
1984 maplen = ceph_decode_32(&p);
1985 ceph_decode_need(&p, end, maplen, bad);
1987 dout("skipping non-latest full map %u len %d\n",
1989 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1990 dout("skipping full map %u len %d, "
1991 "older than our %u\n", epoch, maplen,
1992 osdc->osdmap->epoch);
1994 int skipped_map = 0;
1996 dout("taking full map %u len %d\n", epoch, maplen);
1997 newmap = osdmap_decode(&p, p+maplen);
1998 if (IS_ERR(newmap)) {
1999 err = PTR_ERR(newmap);
2003 oldmap = osdc->osdmap;
2004 osdc->osdmap = newmap;
2006 if (oldmap->epoch + 1 < newmap->epoch)
2008 ceph_osdmap_destroy(oldmap);
2010 was_full = was_full ||
2011 ceph_osdmap_flag(osdc->osdmap,
2013 kick_requests(osdc, skipped_map, was_full);
2022 downgrade_write(&osdc->map_sem);
2023 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2026 * subscribe to subsequent osdmap updates if full to ensure
2027 * we find out when we are no longer full and stop returning
2030 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2031 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2032 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2033 ceph_monc_request_next_osdmap(&osdc->client->monc);
2035 mutex_lock(&osdc->request_mutex);
2036 __send_queued(osdc);
2037 mutex_unlock(&osdc->request_mutex);
2038 up_read(&osdc->map_sem);
2039 wake_up_all(&osdc->client->auth_wq);
2043 pr_err("osdc handle_map corrupt msg\n");
2045 up_write(&osdc->map_sem);
2050 * watch/notify callback event infrastructure
2052 * These callbacks are used both for watch and notify operations.
2054 static void __release_event(struct kref *kref)
2056 struct ceph_osd_event *event =
2057 container_of(kref, struct ceph_osd_event, kref);
2059 dout("__release_event %p\n", event);
2063 static void get_event(struct ceph_osd_event *event)
2065 kref_get(&event->kref);
2068 void ceph_osdc_put_event(struct ceph_osd_event *event)
2070 kref_put(&event->kref, __release_event);
2072 EXPORT_SYMBOL(ceph_osdc_put_event);
2074 static void __insert_event(struct ceph_osd_client *osdc,
2075 struct ceph_osd_event *new)
2077 struct rb_node **p = &osdc->event_tree.rb_node;
2078 struct rb_node *parent = NULL;
2079 struct ceph_osd_event *event = NULL;
2083 event = rb_entry(parent, struct ceph_osd_event, node);
2084 if (new->cookie < event->cookie)
2086 else if (new->cookie > event->cookie)
2087 p = &(*p)->rb_right;
2092 rb_link_node(&new->node, parent, p);
2093 rb_insert_color(&new->node, &osdc->event_tree);
2096 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2099 struct rb_node **p = &osdc->event_tree.rb_node;
2100 struct rb_node *parent = NULL;
2101 struct ceph_osd_event *event = NULL;
2105 event = rb_entry(parent, struct ceph_osd_event, node);
2106 if (cookie < event->cookie)
2108 else if (cookie > event->cookie)
2109 p = &(*p)->rb_right;
2116 static void __remove_event(struct ceph_osd_event *event)
2118 struct ceph_osd_client *osdc = event->osdc;
2120 if (!RB_EMPTY_NODE(&event->node)) {
2121 dout("__remove_event removed %p\n", event);
2122 rb_erase(&event->node, &osdc->event_tree);
2123 ceph_osdc_put_event(event);
2125 dout("__remove_event didn't remove %p\n", event);
2129 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2130 void (*event_cb)(u64, u64, u8, void *),
2131 void *data, struct ceph_osd_event **pevent)
2133 struct ceph_osd_event *event;
2135 event = kmalloc(sizeof(*event), GFP_NOIO);
2139 dout("create_event %p\n", event);
2140 event->cb = event_cb;
2141 event->one_shot = 0;
2144 INIT_LIST_HEAD(&event->osd_node);
2145 RB_CLEAR_NODE(&event->node);
2146 kref_init(&event->kref); /* one ref for us */
2147 kref_get(&event->kref); /* one ref for the caller */
2149 spin_lock(&osdc->event_lock);
2150 event->cookie = ++osdc->event_count;
2151 __insert_event(osdc, event);
2152 spin_unlock(&osdc->event_lock);
2157 EXPORT_SYMBOL(ceph_osdc_create_event);
2159 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2161 struct ceph_osd_client *osdc = event->osdc;
2163 dout("cancel_event %p\n", event);
2164 spin_lock(&osdc->event_lock);
2165 __remove_event(event);
2166 spin_unlock(&osdc->event_lock);
2167 ceph_osdc_put_event(event); /* caller's */
2169 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2172 static void do_event_work(struct work_struct *work)
2174 struct ceph_osd_event_work *event_work =
2175 container_of(work, struct ceph_osd_event_work, work);
2176 struct ceph_osd_event *event = event_work->event;
2177 u64 ver = event_work->ver;
2178 u64 notify_id = event_work->notify_id;
2179 u8 opcode = event_work->opcode;
2181 dout("do_event_work completing %p\n", event);
2182 event->cb(ver, notify_id, opcode, event->data);
2183 dout("do_event_work completed %p\n", event);
2184 ceph_osdc_put_event(event);
2190 * Process osd watch notifications
2192 static void handle_watch_notify(struct ceph_osd_client *osdc,
2193 struct ceph_msg *msg)
2197 u64 cookie, ver, notify_id;
2199 struct ceph_osd_event *event;
2200 struct ceph_osd_event_work *event_work;
2202 p = msg->front.iov_base;
2203 end = p + msg->front.iov_len;
2205 ceph_decode_8_safe(&p, end, proto_ver, bad);
2206 ceph_decode_8_safe(&p, end, opcode, bad);
2207 ceph_decode_64_safe(&p, end, cookie, bad);
2208 ceph_decode_64_safe(&p, end, ver, bad);
2209 ceph_decode_64_safe(&p, end, notify_id, bad);
2211 spin_lock(&osdc->event_lock);
2212 event = __find_event(osdc, cookie);
2214 BUG_ON(event->one_shot);
2217 spin_unlock(&osdc->event_lock);
2218 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2219 cookie, ver, event);
2221 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2223 dout("ERROR: could not allocate event_work\n");
2226 INIT_WORK(&event_work->work, do_event_work);
2227 event_work->event = event;
2228 event_work->ver = ver;
2229 event_work->notify_id = notify_id;
2230 event_work->opcode = opcode;
2231 if (!queue_work(osdc->notify_wq, &event_work->work)) {
2232 dout("WARNING: failed to queue notify event work\n");
2240 ceph_osdc_put_event(event);
2244 pr_err("osdc handle_watch_notify corrupt msg\n");
2249 * build new request AND message
2252 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2253 struct ceph_snap_context *snapc, u64 snap_id,
2254 struct timespec *mtime)
2256 struct ceph_msg *msg = req->r_request;
2259 int flags = req->r_flags;
2263 req->r_snapid = snap_id;
2264 req->r_snapc = ceph_get_snap_context(snapc);
2266 /* encode request */
2267 msg->hdr.version = cpu_to_le16(4);
2269 p = msg->front.iov_base;
2270 ceph_encode_32(&p, 1); /* client_inc is always 1 */
2271 req->r_request_osdmap_epoch = p;
2273 req->r_request_flags = p;
2275 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2276 ceph_encode_timespec(p, mtime);
2277 p += sizeof(struct ceph_timespec);
2278 req->r_request_reassert_version = p;
2279 p += sizeof(struct ceph_eversion); /* will get filled in */
2282 ceph_encode_8(&p, 4);
2283 ceph_encode_8(&p, 4);
2284 ceph_encode_32(&p, 8 + 4 + 4);
2285 req->r_request_pool = p;
2287 ceph_encode_32(&p, -1); /* preferred */
2288 ceph_encode_32(&p, 0); /* key len */
2290 ceph_encode_8(&p, 1);
2291 req->r_request_pgid = p;
2293 ceph_encode_32(&p, -1); /* preferred */
2296 ceph_encode_32(&p, req->r_base_oid.name_len);
2297 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2298 dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2299 req->r_base_oid.name, req->r_base_oid.name_len);
2300 p += req->r_base_oid.name_len;
2302 /* ops--can imply data */
2303 ceph_encode_16(&p, (u16)req->r_num_ops);
2305 for (i = 0; i < req->r_num_ops; i++) {
2306 data_len += osd_req_encode_op(req, p, i);
2307 p += sizeof(struct ceph_osd_op);
2311 ceph_encode_64(&p, req->r_snapid);
2312 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2313 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2315 for (i = 0; i < snapc->num_snaps; i++) {
2316 ceph_encode_64(&p, req->r_snapc->snaps[i]);
2320 req->r_request_attempts = p;
2324 if (flags & CEPH_OSD_FLAG_WRITE) {
2328 * The header "data_off" is a hint to the receiver
2329 * allowing it to align received data into its
2330 * buffers such that there's no need to re-copy
2331 * it before writing it to disk (direct I/O).
2333 data_off = (u16) (off & 0xffff);
2334 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2336 req->r_request->hdr.data_len = cpu_to_le32(data_len);
2338 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2339 msg_size = p - msg->front.iov_base;
2340 msg->front.iov_len = msg_size;
2341 msg->hdr.front_len = cpu_to_le32(msg_size);
2343 dout("build_request msg_size was %d\n", (int)msg_size);
2345 EXPORT_SYMBOL(ceph_osdc_build_request);
2348 * Register request, send initial attempt.
2350 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2351 struct ceph_osd_request *req,
2356 down_read(&osdc->map_sem);
2357 mutex_lock(&osdc->request_mutex);
2358 __register_request(osdc, req);
2360 req->r_got_reply = 0;
2361 rc = __map_request(osdc, req, 0);
2364 dout("osdc_start_request failed map, "
2365 " will retry %lld\n", req->r_tid);
2368 __unregister_request(osdc, req);
2372 if (req->r_osd == NULL) {
2373 dout("send_request %p no up osds in pg\n", req);
2374 ceph_monc_request_next_osdmap(&osdc->client->monc);
2376 __send_queued(osdc);
2380 mutex_unlock(&osdc->request_mutex);
2381 up_read(&osdc->map_sem);
2384 EXPORT_SYMBOL(ceph_osdc_start_request);
2387 * wait for a request to complete
2389 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2390 struct ceph_osd_request *req)
2394 rc = wait_for_completion_interruptible(&req->r_completion);
2396 mutex_lock(&osdc->request_mutex);
2397 __cancel_request(req);
2398 __unregister_request(osdc, req);
2399 mutex_unlock(&osdc->request_mutex);
2400 complete_request(req);
2401 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
2405 dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
2406 return req->r_result;
2408 EXPORT_SYMBOL(ceph_osdc_wait_request);
2411 * sync - wait for all in-flight requests to flush. avoid starvation.
2413 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2415 struct ceph_osd_request *req;
2416 u64 last_tid, next_tid = 0;
2418 mutex_lock(&osdc->request_mutex);
2419 last_tid = osdc->last_tid;
2421 req = __lookup_request_ge(osdc, next_tid);
2424 if (req->r_tid > last_tid)
2427 next_tid = req->r_tid + 1;
2428 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2431 ceph_osdc_get_request(req);
2432 mutex_unlock(&osdc->request_mutex);
2433 dout("sync waiting on tid %llu (last is %llu)\n",
2434 req->r_tid, last_tid);
2435 wait_for_completion(&req->r_safe_completion);
2436 mutex_lock(&osdc->request_mutex);
2437 ceph_osdc_put_request(req);
2439 mutex_unlock(&osdc->request_mutex);
2440 dout("sync done (thru tid %llu)\n", last_tid);
2442 EXPORT_SYMBOL(ceph_osdc_sync);
2445 * Call all pending notify callbacks - for use after a watch is
2446 * unregistered, to make sure no more callbacks for it will be invoked
2448 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2450 flush_workqueue(osdc->notify_wq);
2452 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2458 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2463 osdc->client = client;
2464 osdc->osdmap = NULL;
2465 init_rwsem(&osdc->map_sem);
2466 init_completion(&osdc->map_waiters);
2467 osdc->last_requested_map = 0;
2468 mutex_init(&osdc->request_mutex);
2470 osdc->osds = RB_ROOT;
2471 INIT_LIST_HEAD(&osdc->osd_lru);
2472 osdc->requests = RB_ROOT;
2473 INIT_LIST_HEAD(&osdc->req_lru);
2474 INIT_LIST_HEAD(&osdc->req_unsent);
2475 INIT_LIST_HEAD(&osdc->req_notarget);
2476 INIT_LIST_HEAD(&osdc->req_linger);
2477 osdc->num_requests = 0;
2478 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2479 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2480 spin_lock_init(&osdc->event_lock);
2481 osdc->event_tree = RB_ROOT;
2482 osdc->event_count = 0;
2484 schedule_delayed_work(&osdc->osds_timeout_work,
2485 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2488 osdc->req_mempool = mempool_create_kmalloc_pool(10,
2489 sizeof(struct ceph_osd_request));
2490 if (!osdc->req_mempool)
2493 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2494 OSD_OP_FRONT_LEN, 10, true,
2498 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2499 OSD_OPREPLY_FRONT_LEN, 10, true,
2505 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2506 if (!osdc->notify_wq)
2511 ceph_msgpool_destroy(&osdc->msgpool_op);
2513 mempool_destroy(osdc->req_mempool);
2518 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2520 flush_workqueue(osdc->notify_wq);
2521 destroy_workqueue(osdc->notify_wq);
2522 cancel_delayed_work_sync(&osdc->timeout_work);
2523 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2525 ceph_osdmap_destroy(osdc->osdmap);
2526 osdc->osdmap = NULL;
2528 remove_all_osds(osdc);
2529 mempool_destroy(osdc->req_mempool);
2530 ceph_msgpool_destroy(&osdc->msgpool_op);
2531 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2535 * Read some contiguous pages. If we cross a stripe boundary, shorten
2536 * *plen. Return number of bytes read, or error.
2538 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2539 struct ceph_vino vino, struct ceph_file_layout *layout,
2541 u32 truncate_seq, u64 truncate_size,
2542 struct page **pages, int num_pages, int page_align)
2544 struct ceph_osd_request *req;
2547 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2548 vino.snap, off, *plen);
2549 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2550 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2551 NULL, truncate_seq, truncate_size,
2554 return PTR_ERR(req);
2556 /* it may be a short read due to an object boundary */
2558 osd_req_op_extent_osd_data_pages(req, 0,
2559 pages, *plen, page_align, false, false);
2561 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2562 off, *plen, *plen, page_align);
2564 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2566 rc = ceph_osdc_start_request(osdc, req, false);
2568 rc = ceph_osdc_wait_request(osdc, req);
2570 ceph_osdc_put_request(req);
2571 dout("readpages result %d\n", rc);
2574 EXPORT_SYMBOL(ceph_osdc_readpages);
2577 * do a synchronous write on N pages
2579 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2580 struct ceph_file_layout *layout,
2581 struct ceph_snap_context *snapc,
2583 u32 truncate_seq, u64 truncate_size,
2584 struct timespec *mtime,
2585 struct page **pages, int num_pages)
2587 struct ceph_osd_request *req;
2589 int page_align = off & ~PAGE_MASK;
2591 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */
2592 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2594 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2595 snapc, truncate_seq, truncate_size,
2598 return PTR_ERR(req);
2600 /* it may be a short write due to an object boundary */
2601 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2603 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2605 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2607 rc = ceph_osdc_start_request(osdc, req, true);
2609 rc = ceph_osdc_wait_request(osdc, req);
2611 ceph_osdc_put_request(req);
2614 dout("writepages result %d\n", rc);
2617 EXPORT_SYMBOL(ceph_osdc_writepages);
2619 int ceph_osdc_setup(void)
2621 BUG_ON(ceph_osd_request_cache);
2622 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2623 sizeof (struct ceph_osd_request),
2624 __alignof__(struct ceph_osd_request),
2627 return ceph_osd_request_cache ? 0 : -ENOMEM;
2629 EXPORT_SYMBOL(ceph_osdc_setup);
2631 void ceph_osdc_cleanup(void)
2633 BUG_ON(!ceph_osd_request_cache);
2634 kmem_cache_destroy(ceph_osd_request_cache);
2635 ceph_osd_request_cache = NULL;
2637 EXPORT_SYMBOL(ceph_osdc_cleanup);
2640 * handle incoming message
2642 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2644 struct ceph_osd *osd = con->private;
2645 struct ceph_osd_client *osdc;
2646 int type = le16_to_cpu(msg->hdr.type);
2653 case CEPH_MSG_OSD_MAP:
2654 ceph_osdc_handle_map(osdc, msg);
2656 case CEPH_MSG_OSD_OPREPLY:
2657 handle_reply(osdc, msg, con);
2659 case CEPH_MSG_WATCH_NOTIFY:
2660 handle_watch_notify(osdc, msg);
2664 pr_err("received unknown message type %d %s\n", type,
2665 ceph_msg_type_name(type));
2672 * lookup and return message for incoming reply. set up reply message
2675 static struct ceph_msg *get_reply(struct ceph_connection *con,
2676 struct ceph_msg_header *hdr,
2679 struct ceph_osd *osd = con->private;
2680 struct ceph_osd_client *osdc = osd->o_osdc;
2682 struct ceph_osd_request *req;
2683 int front_len = le32_to_cpu(hdr->front_len);
2684 int data_len = le32_to_cpu(hdr->data_len);
2687 tid = le64_to_cpu(hdr->tid);
2688 mutex_lock(&osdc->request_mutex);
2689 req = __lookup_request(osdc, tid);
2693 dout("get_reply unknown tid %llu from osd%d\n", tid,
2698 if (req->r_reply->con)
2699 dout("%s revoking msg %p from old con %p\n", __func__,
2700 req->r_reply, req->r_reply->con);
2701 ceph_msg_revoke_incoming(req->r_reply);
2703 if (front_len > req->r_reply->front_alloc_len) {
2704 pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n",
2705 front_len, req->r_reply->front_alloc_len,
2706 (unsigned int)con->peer_name.type,
2707 le64_to_cpu(con->peer_name.num));
2708 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2712 ceph_msg_put(req->r_reply);
2715 m = ceph_msg_get(req->r_reply);
2718 struct ceph_osd_data *osd_data;
2721 * XXX This is assuming there is only one op containing
2722 * XXX page data. Probably OK for reads, but this
2723 * XXX ought to be done more generally.
2725 osd_data = osd_req_op_extent_osd_data(req, 0);
2726 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2727 if (osd_data->pages &&
2728 unlikely(osd_data->length < data_len)) {
2730 pr_warning("tid %lld reply has %d bytes "
2731 "we had only %llu bytes ready\n",
2732 tid, data_len, osd_data->length);
2741 dout("get_reply tid %lld %p\n", tid, m);
2744 mutex_unlock(&osdc->request_mutex);
2749 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2750 struct ceph_msg_header *hdr,
2753 struct ceph_osd *osd = con->private;
2754 int type = le16_to_cpu(hdr->type);
2755 int front = le32_to_cpu(hdr->front_len);
2759 case CEPH_MSG_OSD_MAP:
2760 case CEPH_MSG_WATCH_NOTIFY:
2761 return ceph_msg_new(type, front, GFP_NOFS, false);
2762 case CEPH_MSG_OSD_OPREPLY:
2763 return get_reply(con, hdr, skip);
2765 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2773 * Wrappers to refcount containing ceph_osd struct
2775 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2777 struct ceph_osd *osd = con->private;
2783 static void put_osd_con(struct ceph_connection *con)
2785 struct ceph_osd *osd = con->private;
2793 * Note: returned pointer is the address of a structure that's
2794 * managed separately. Caller must *not* attempt to free it.
2796 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2797 int *proto, int force_new)
2799 struct ceph_osd *o = con->private;
2800 struct ceph_osd_client *osdc = o->o_osdc;
2801 struct ceph_auth_client *ac = osdc->client->monc.auth;
2802 struct ceph_auth_handshake *auth = &o->o_auth;
2804 if (force_new && auth->authorizer) {
2805 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2806 auth->authorizer = NULL;
2808 if (!auth->authorizer) {
2809 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2812 return ERR_PTR(ret);
2814 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2817 return ERR_PTR(ret);
2819 *proto = ac->protocol;
2825 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2827 struct ceph_osd *o = con->private;
2828 struct ceph_osd_client *osdc = o->o_osdc;
2829 struct ceph_auth_client *ac = osdc->client->monc.auth;
2831 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2834 static int invalidate_authorizer(struct ceph_connection *con)
2836 struct ceph_osd *o = con->private;
2837 struct ceph_osd_client *osdc = o->o_osdc;
2838 struct ceph_auth_client *ac = osdc->client->monc.auth;
2840 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2841 return ceph_monc_validate_auth(&osdc->client->monc);
2844 static const struct ceph_connection_operations osd_con_ops = {
2847 .dispatch = dispatch,
2848 .get_authorizer = get_authorizer,
2849 .verify_authorizer_reply = verify_authorizer_reply,
2850 .invalidate_authorizer = invalidate_authorizer,
2851 .alloc_msg = alloc_msg,