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_request(struct ceph_osd_client *osdc,
34 struct ceph_osd_request *req);
35 static void __unregister_linger_request(struct ceph_osd_client *osdc,
36 struct ceph_osd_request *req);
37 static void __enqueue_request(struct ceph_osd_request *req);
38 static void __send_request(struct ceph_osd_client *osdc,
39 struct ceph_osd_request *req);
42 * Implement client access to distributed object storage cluster.
44 * All data objects are stored within a cluster/cloud of OSDs, or
45 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
46 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
47 * remote daemons serving up and coordinating consistent and safe
50 * Cluster membership and the mapping of data objects onto storage devices
51 * are described by the osd map.
53 * We keep track of pending OSD requests (read, write), resubmit
54 * requests to different OSDs when the cluster topology/data layout
55 * change, or retry the affected requests when the communications
56 * channel with an OSD is reset.
60 * calculate the mapping of a file extent onto an object, and fill out the
61 * request accordingly. shorten extent as necessary if it crosses an
64 * fill osd op in request message.
66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
67 u64 *objnum, u64 *objoff, u64 *objlen)
73 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
77 if (*objlen < orig_len) {
79 dout(" skipping last %llu, final file extent %llu~%llu\n",
80 orig_len - *plen, off, *plen);
83 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
90 memset(osd_data, 0, sizeof (*osd_data));
91 osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
95 struct page **pages, u64 length, u32 alignment,
96 bool pages_from_pool, bool own_pages)
98 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
99 osd_data->pages = pages;
100 osd_data->length = length;
101 osd_data->alignment = alignment;
102 osd_data->pages_from_pool = pages_from_pool;
103 osd_data->own_pages = own_pages;
106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
107 struct ceph_pagelist *pagelist)
109 osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
110 osd_data->pagelist = pagelist;
114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
115 struct bio *bio, size_t bio_length)
117 osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
119 osd_data->bio_length = bio_length;
121 #endif /* CONFIG_BLOCK */
123 #define osd_req_op_data(oreq, whch, typ, fld) \
125 struct ceph_osd_request *__oreq = (oreq); \
126 unsigned int __whch = (whch); \
127 BUG_ON(__whch >= __oreq->r_num_ops); \
128 &__oreq->r_ops[__whch].typ.fld; \
131 static struct ceph_osd_data *
132 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
134 BUG_ON(which >= osd_req->r_num_ops);
136 return &osd_req->r_ops[which].raw_data_in;
139 struct ceph_osd_data *
140 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
143 return osd_req_op_data(osd_req, which, extent, osd_data);
145 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
147 struct ceph_osd_data *
148 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
151 return osd_req_op_data(osd_req, which, cls, response_data);
153 EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */
155 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
156 unsigned int which, struct page **pages,
157 u64 length, u32 alignment,
158 bool pages_from_pool, bool own_pages)
160 struct ceph_osd_data *osd_data;
162 osd_data = osd_req_op_raw_data_in(osd_req, which);
163 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
164 pages_from_pool, own_pages);
166 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
168 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
169 unsigned int which, struct page **pages,
170 u64 length, u32 alignment,
171 bool pages_from_pool, bool own_pages)
173 struct ceph_osd_data *osd_data;
175 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
176 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
177 pages_from_pool, own_pages);
179 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
181 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
182 unsigned int which, struct ceph_pagelist *pagelist)
184 struct ceph_osd_data *osd_data;
186 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
187 ceph_osd_data_pagelist_init(osd_data, pagelist);
189 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
192 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
193 unsigned int which, struct bio *bio, size_t bio_length)
195 struct ceph_osd_data *osd_data;
197 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
198 ceph_osd_data_bio_init(osd_data, bio, bio_length);
200 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
201 #endif /* CONFIG_BLOCK */
203 static void osd_req_op_cls_request_info_pagelist(
204 struct ceph_osd_request *osd_req,
205 unsigned int which, struct ceph_pagelist *pagelist)
207 struct ceph_osd_data *osd_data;
209 osd_data = osd_req_op_data(osd_req, which, cls, request_info);
210 ceph_osd_data_pagelist_init(osd_data, pagelist);
213 void osd_req_op_cls_request_data_pagelist(
214 struct ceph_osd_request *osd_req,
215 unsigned int which, struct ceph_pagelist *pagelist)
217 struct ceph_osd_data *osd_data;
219 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
220 ceph_osd_data_pagelist_init(osd_data, pagelist);
222 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
224 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
225 unsigned int which, struct page **pages, u64 length,
226 u32 alignment, bool pages_from_pool, bool own_pages)
228 struct ceph_osd_data *osd_data;
230 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
231 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
232 pages_from_pool, own_pages);
234 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
236 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
237 unsigned int which, struct page **pages, u64 length,
238 u32 alignment, bool pages_from_pool, bool own_pages)
240 struct ceph_osd_data *osd_data;
242 osd_data = osd_req_op_data(osd_req, which, cls, response_data);
243 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
244 pages_from_pool, own_pages);
246 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
248 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
250 switch (osd_data->type) {
251 case CEPH_OSD_DATA_TYPE_NONE:
253 case CEPH_OSD_DATA_TYPE_PAGES:
254 return osd_data->length;
255 case CEPH_OSD_DATA_TYPE_PAGELIST:
256 return (u64)osd_data->pagelist->length;
258 case CEPH_OSD_DATA_TYPE_BIO:
259 return (u64)osd_data->bio_length;
260 #endif /* CONFIG_BLOCK */
262 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
267 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
269 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
272 num_pages = calc_pages_for((u64)osd_data->alignment,
273 (u64)osd_data->length);
274 ceph_release_page_vector(osd_data->pages, num_pages);
276 ceph_osd_data_init(osd_data);
279 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
282 struct ceph_osd_req_op *op;
284 BUG_ON(which >= osd_req->r_num_ops);
285 op = &osd_req->r_ops[which];
288 case CEPH_OSD_OP_READ:
289 case CEPH_OSD_OP_WRITE:
290 case CEPH_OSD_OP_WRITEFULL:
291 ceph_osd_data_release(&op->extent.osd_data);
293 case CEPH_OSD_OP_CALL:
294 ceph_osd_data_release(&op->cls.request_info);
295 ceph_osd_data_release(&op->cls.request_data);
296 ceph_osd_data_release(&op->cls.response_data);
298 case CEPH_OSD_OP_SETXATTR:
299 case CEPH_OSD_OP_CMPXATTR:
300 ceph_osd_data_release(&op->xattr.osd_data);
302 case CEPH_OSD_OP_STAT:
303 ceph_osd_data_release(&op->raw_data_in);
313 static void ceph_osdc_release_request(struct kref *kref)
315 struct ceph_osd_request *req = container_of(kref,
316 struct ceph_osd_request, r_kref);
319 dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
320 req->r_request, req->r_reply);
321 WARN_ON(!RB_EMPTY_NODE(&req->r_node));
322 WARN_ON(!list_empty(&req->r_req_lru_item));
323 WARN_ON(!list_empty(&req->r_osd_item));
324 WARN_ON(!list_empty(&req->r_linger_item));
325 WARN_ON(!list_empty(&req->r_linger_osd_item));
329 ceph_msg_put(req->r_request);
331 ceph_msg_revoke_incoming(req->r_reply);
332 ceph_msg_put(req->r_reply);
335 for (which = 0; which < req->r_num_ops; which++)
336 osd_req_op_data_release(req, which);
338 ceph_put_snap_context(req->r_snapc);
340 mempool_free(req, req->r_osdc->req_mempool);
341 else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS)
342 kmem_cache_free(ceph_osd_request_cache, req);
347 void ceph_osdc_get_request(struct ceph_osd_request *req)
349 dout("%s %p (was %d)\n", __func__, req,
350 atomic_read(&req->r_kref.refcount));
351 kref_get(&req->r_kref);
353 EXPORT_SYMBOL(ceph_osdc_get_request);
355 void ceph_osdc_put_request(struct ceph_osd_request *req)
357 dout("%s %p (was %d)\n", __func__, req,
358 atomic_read(&req->r_kref.refcount));
359 kref_put(&req->r_kref, ceph_osdc_release_request);
361 EXPORT_SYMBOL(ceph_osdc_put_request);
363 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
364 struct ceph_snap_context *snapc,
365 unsigned int num_ops,
369 struct ceph_osd_request *req;
370 struct ceph_msg *msg;
374 BUG_ON(num_ops > CEPH_OSD_SLAB_OPS);
375 req = mempool_alloc(osdc->req_mempool, gfp_flags);
376 } else if (num_ops <= CEPH_OSD_SLAB_OPS) {
377 req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags);
379 BUG_ON(num_ops > CEPH_OSD_MAX_OPS);
380 req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]),
386 /* req only, each op is zeroed in _osd_req_op_init() */
387 memset(req, 0, sizeof(*req));
390 req->r_mempool = use_mempool;
391 req->r_num_ops = num_ops;
393 kref_init(&req->r_kref);
394 init_completion(&req->r_completion);
395 init_completion(&req->r_safe_completion);
396 RB_CLEAR_NODE(&req->r_node);
397 INIT_LIST_HEAD(&req->r_unsafe_item);
398 INIT_LIST_HEAD(&req->r_linger_item);
399 INIT_LIST_HEAD(&req->r_linger_osd_item);
400 INIT_LIST_HEAD(&req->r_req_lru_item);
401 INIT_LIST_HEAD(&req->r_osd_item);
403 req->r_base_oloc.pool = -1;
404 req->r_target_oloc.pool = -1;
406 msg_size = OSD_OPREPLY_FRONT_LEN;
407 if (num_ops > CEPH_OSD_SLAB_OPS) {
408 /* ceph_osd_op and rval */
409 msg_size += (num_ops - CEPH_OSD_SLAB_OPS) *
410 (sizeof(struct ceph_osd_op) + 4);
413 /* create reply message */
415 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
417 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size,
420 ceph_osdc_put_request(req);
425 msg_size = 4 + 4 + 4; /* client_inc, osdmap_epoch, flags */
426 msg_size += 4 + 4 + 4 + 8; /* mtime, reassert_version */
427 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
428 msg_size += 1 + 8 + 4 + 4; /* pgid */
429 msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
430 msg_size += 2 + num_ops * sizeof(struct ceph_osd_op);
431 msg_size += 8; /* snapid */
432 msg_size += 8; /* snap_seq */
433 msg_size += 4 + 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
434 msg_size += 4; /* retry_attempt */
436 /* create request message; allow space for oid */
438 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
440 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
442 ceph_osdc_put_request(req);
446 memset(msg->front.iov_base, 0, msg->front.iov_len);
448 req->r_request = msg;
452 EXPORT_SYMBOL(ceph_osdc_alloc_request);
454 static bool osd_req_opcode_valid(u16 opcode)
457 #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true;
458 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
466 * This is an osd op init function for opcodes that have no data or
467 * other information associated with them. It also serves as a
468 * common init routine for all the other init functions, below.
470 static struct ceph_osd_req_op *
471 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
472 u16 opcode, u32 flags)
474 struct ceph_osd_req_op *op;
476 BUG_ON(which >= osd_req->r_num_ops);
477 BUG_ON(!osd_req_opcode_valid(opcode));
479 op = &osd_req->r_ops[which];
480 memset(op, 0, sizeof (*op));
487 void osd_req_op_init(struct ceph_osd_request *osd_req,
488 unsigned int which, u16 opcode, u32 flags)
490 (void)_osd_req_op_init(osd_req, which, opcode, flags);
492 EXPORT_SYMBOL(osd_req_op_init);
494 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
495 unsigned int which, u16 opcode,
496 u64 offset, u64 length,
497 u64 truncate_size, u32 truncate_seq)
499 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
501 size_t payload_len = 0;
503 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
504 opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO &&
505 opcode != CEPH_OSD_OP_TRUNCATE);
507 op->extent.offset = offset;
508 op->extent.length = length;
509 op->extent.truncate_size = truncate_size;
510 op->extent.truncate_seq = truncate_seq;
511 if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL)
512 payload_len += length;
514 op->indata_len = payload_len;
516 EXPORT_SYMBOL(osd_req_op_extent_init);
518 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
519 unsigned int which, u64 length)
521 struct ceph_osd_req_op *op;
524 BUG_ON(which >= osd_req->r_num_ops);
525 op = &osd_req->r_ops[which];
526 previous = op->extent.length;
528 if (length == previous)
529 return; /* Nothing to do */
530 BUG_ON(length > previous);
532 op->extent.length = length;
533 op->indata_len -= previous - length;
535 EXPORT_SYMBOL(osd_req_op_extent_update);
537 void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
538 unsigned int which, u64 offset_inc)
540 struct ceph_osd_req_op *op, *prev_op;
542 BUG_ON(which + 1 >= osd_req->r_num_ops);
544 prev_op = &osd_req->r_ops[which];
545 op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags);
546 /* dup previous one */
547 op->indata_len = prev_op->indata_len;
548 op->outdata_len = prev_op->outdata_len;
549 op->extent = prev_op->extent;
551 op->extent.offset += offset_inc;
552 op->extent.length -= offset_inc;
554 if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
555 op->indata_len -= offset_inc;
557 EXPORT_SYMBOL(osd_req_op_extent_dup_last);
559 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
560 u16 opcode, const char *class, const char *method)
562 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
564 struct ceph_pagelist *pagelist;
565 size_t payload_len = 0;
568 BUG_ON(opcode != CEPH_OSD_OP_CALL);
570 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
572 ceph_pagelist_init(pagelist);
574 op->cls.class_name = class;
575 size = strlen(class);
576 BUG_ON(size > (size_t) U8_MAX);
577 op->cls.class_len = size;
578 ceph_pagelist_append(pagelist, class, size);
581 op->cls.method_name = method;
582 size = strlen(method);
583 BUG_ON(size > (size_t) U8_MAX);
584 op->cls.method_len = size;
585 ceph_pagelist_append(pagelist, method, size);
588 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
590 op->cls.argc = 0; /* currently unused */
592 op->indata_len = payload_len;
594 EXPORT_SYMBOL(osd_req_op_cls_init);
596 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
597 u16 opcode, const char *name, const void *value,
598 size_t size, u8 cmp_op, u8 cmp_mode)
600 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
602 struct ceph_pagelist *pagelist;
605 BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
607 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
611 ceph_pagelist_init(pagelist);
613 payload_len = strlen(name);
614 op->xattr.name_len = payload_len;
615 ceph_pagelist_append(pagelist, name, payload_len);
617 op->xattr.value_len = size;
618 ceph_pagelist_append(pagelist, value, size);
621 op->xattr.cmp_op = cmp_op;
622 op->xattr.cmp_mode = cmp_mode;
624 ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
625 op->indata_len = payload_len;
628 EXPORT_SYMBOL(osd_req_op_xattr_init);
630 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
631 unsigned int which, u16 opcode,
632 u64 cookie, u64 version, int flag)
634 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
637 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
639 op->watch.cookie = cookie;
640 op->watch.ver = version;
641 if (opcode == CEPH_OSD_OP_WATCH && flag)
642 op->watch.flag = (u8)1;
644 EXPORT_SYMBOL(osd_req_op_watch_init);
646 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
648 u64 expected_object_size,
649 u64 expected_write_size)
651 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
652 CEPH_OSD_OP_SETALLOCHINT,
655 op->alloc_hint.expected_object_size = expected_object_size;
656 op->alloc_hint.expected_write_size = expected_write_size;
659 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
660 * not worth a feature bit. Set FAILOK per-op flag to make
661 * sure older osds don't trip over an unsupported opcode.
663 op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
665 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
667 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
668 struct ceph_osd_data *osd_data)
670 u64 length = ceph_osd_data_length(osd_data);
672 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
673 BUG_ON(length > (u64) SIZE_MAX);
675 ceph_msg_data_add_pages(msg, osd_data->pages,
676 length, osd_data->alignment);
677 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
679 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
681 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
682 ceph_msg_data_add_bio(msg, osd_data->bio, length);
685 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
689 static u64 osd_req_encode_op(struct ceph_osd_request *req,
690 struct ceph_osd_op *dst, unsigned int which)
692 struct ceph_osd_req_op *src;
693 struct ceph_osd_data *osd_data;
694 u64 request_data_len = 0;
697 BUG_ON(which >= req->r_num_ops);
698 src = &req->r_ops[which];
699 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
700 pr_err("unrecognized osd opcode %d\n", src->op);
706 case CEPH_OSD_OP_STAT:
707 osd_data = &src->raw_data_in;
708 ceph_osdc_msg_data_add(req->r_reply, osd_data);
710 case CEPH_OSD_OP_READ:
711 case CEPH_OSD_OP_WRITE:
712 case CEPH_OSD_OP_WRITEFULL:
713 case CEPH_OSD_OP_ZERO:
714 case CEPH_OSD_OP_TRUNCATE:
715 if (src->op == CEPH_OSD_OP_WRITE ||
716 src->op == CEPH_OSD_OP_WRITEFULL)
717 request_data_len = src->extent.length;
718 dst->extent.offset = cpu_to_le64(src->extent.offset);
719 dst->extent.length = cpu_to_le64(src->extent.length);
720 dst->extent.truncate_size =
721 cpu_to_le64(src->extent.truncate_size);
722 dst->extent.truncate_seq =
723 cpu_to_le32(src->extent.truncate_seq);
724 osd_data = &src->extent.osd_data;
725 if (src->op == CEPH_OSD_OP_WRITE ||
726 src->op == CEPH_OSD_OP_WRITEFULL)
727 ceph_osdc_msg_data_add(req->r_request, osd_data);
729 ceph_osdc_msg_data_add(req->r_reply, osd_data);
731 case CEPH_OSD_OP_CALL:
732 dst->cls.class_len = src->cls.class_len;
733 dst->cls.method_len = src->cls.method_len;
734 osd_data = &src->cls.request_info;
735 ceph_osdc_msg_data_add(req->r_request, osd_data);
736 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
737 request_data_len = osd_data->pagelist->length;
739 osd_data = &src->cls.request_data;
740 data_length = ceph_osd_data_length(osd_data);
742 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
743 dst->cls.indata_len = cpu_to_le32(data_length);
744 ceph_osdc_msg_data_add(req->r_request, osd_data);
745 src->indata_len += data_length;
746 request_data_len += data_length;
748 osd_data = &src->cls.response_data;
749 ceph_osdc_msg_data_add(req->r_reply, osd_data);
751 case CEPH_OSD_OP_STARTSYNC:
753 case CEPH_OSD_OP_NOTIFY_ACK:
754 case CEPH_OSD_OP_WATCH:
755 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
756 dst->watch.ver = cpu_to_le64(src->watch.ver);
757 dst->watch.flag = src->watch.flag;
759 case CEPH_OSD_OP_SETALLOCHINT:
760 dst->alloc_hint.expected_object_size =
761 cpu_to_le64(src->alloc_hint.expected_object_size);
762 dst->alloc_hint.expected_write_size =
763 cpu_to_le64(src->alloc_hint.expected_write_size);
765 case CEPH_OSD_OP_SETXATTR:
766 case CEPH_OSD_OP_CMPXATTR:
767 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
768 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
769 dst->xattr.cmp_op = src->xattr.cmp_op;
770 dst->xattr.cmp_mode = src->xattr.cmp_mode;
771 osd_data = &src->xattr.osd_data;
772 ceph_osdc_msg_data_add(req->r_request, osd_data);
773 request_data_len = osd_data->pagelist->length;
775 case CEPH_OSD_OP_CREATE:
776 case CEPH_OSD_OP_DELETE:
779 pr_err("unsupported osd opcode %s\n",
780 ceph_osd_op_name(src->op));
786 dst->op = cpu_to_le16(src->op);
787 dst->flags = cpu_to_le32(src->flags);
788 dst->payload_len = cpu_to_le32(src->indata_len);
790 return request_data_len;
794 * build new request AND message, calculate layout, and adjust file
797 * if the file was recently truncated, we include information about its
798 * old and new size so that the object can be updated appropriately. (we
799 * avoid synchronously deleting truncated objects because it's slow.)
801 * if @do_sync, include a 'startsync' command so that the osd will flush
804 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
805 struct ceph_file_layout *layout,
806 struct ceph_vino vino,
808 unsigned int which, int num_ops,
809 int opcode, int flags,
810 struct ceph_snap_context *snapc,
815 struct ceph_osd_request *req;
821 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
822 opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
823 opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
825 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
828 return ERR_PTR(-ENOMEM);
830 req->r_flags = flags;
832 /* calculate max write size */
833 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
835 ceph_osdc_put_request(req);
839 if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
840 osd_req_op_init(req, which, opcode, 0);
842 u32 object_size = le32_to_cpu(layout->fl_object_size);
843 u32 object_base = off - objoff;
844 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
845 if (truncate_size <= object_base) {
848 truncate_size -= object_base;
849 if (truncate_size > object_size)
850 truncate_size = object_size;
853 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
854 truncate_size, truncate_seq);
857 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
859 snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
860 "%llx.%08llx", vino.ino, objnum);
861 req->r_base_oid.name_len = strlen(req->r_base_oid.name);
865 EXPORT_SYMBOL(ceph_osdc_new_request);
868 * We keep osd requests in an rbtree, sorted by ->r_tid.
870 static void __insert_request(struct ceph_osd_client *osdc,
871 struct ceph_osd_request *new)
873 struct rb_node **p = &osdc->requests.rb_node;
874 struct rb_node *parent = NULL;
875 struct ceph_osd_request *req = NULL;
879 req = rb_entry(parent, struct ceph_osd_request, r_node);
880 if (new->r_tid < req->r_tid)
882 else if (new->r_tid > req->r_tid)
888 rb_link_node(&new->r_node, parent, p);
889 rb_insert_color(&new->r_node, &osdc->requests);
892 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
895 struct ceph_osd_request *req;
896 struct rb_node *n = osdc->requests.rb_node;
899 req = rb_entry(n, struct ceph_osd_request, r_node);
900 if (tid < req->r_tid)
902 else if (tid > req->r_tid)
910 static struct ceph_osd_request *
911 __lookup_request_ge(struct ceph_osd_client *osdc,
914 struct ceph_osd_request *req;
915 struct rb_node *n = osdc->requests.rb_node;
918 req = rb_entry(n, struct ceph_osd_request, r_node);
919 if (tid < req->r_tid) {
923 } else if (tid > req->r_tid) {
932 static void __kick_linger_request(struct ceph_osd_request *req)
934 struct ceph_osd_client *osdc = req->r_osdc;
935 struct ceph_osd *osd = req->r_osd;
938 * Linger requests need to be resent with a new tid to avoid
939 * the dup op detection logic on the OSDs. Achieve this with
940 * a re-register dance instead of open-coding.
942 ceph_osdc_get_request(req);
943 if (!list_empty(&req->r_linger_item))
944 __unregister_linger_request(osdc, req);
946 __unregister_request(osdc, req);
947 __register_request(osdc, req);
948 ceph_osdc_put_request(req);
951 * Unless request has been registered as both normal and
952 * lingering, __unregister{,_linger}_request clears r_osd.
953 * However, here we need to preserve r_osd to make sure we
954 * requeue on the same OSD.
956 WARN_ON(req->r_osd || !osd);
959 dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
960 __enqueue_request(req);
964 * Resubmit requests pending on the given osd.
966 static void __kick_osd_requests(struct ceph_osd_client *osdc,
967 struct ceph_osd *osd)
969 struct ceph_osd_request *req, *nreq;
971 LIST_HEAD(resend_linger);
974 dout("%s osd%d\n", __func__, osd->o_osd);
975 err = __reset_osd(osdc, osd);
980 * Build up a list of requests to resend by traversing the
981 * osd's list of requests. Requests for a given object are
982 * sent in tid order, and that is also the order they're
983 * kept on this list. Therefore all requests that are in
984 * flight will be found first, followed by all requests that
985 * have not yet been sent. And to resend requests while
986 * preserving this order we will want to put any sent
987 * requests back on the front of the osd client's unsent
990 * So we build a separate ordered list of already-sent
991 * requests for the affected osd and splice it onto the
992 * front of the osd client's unsent list. Once we've seen a
993 * request that has not yet been sent we're done. Those
994 * requests are already sitting right where they belong.
996 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
1000 if (!req->r_linger) {
1001 dout("%s requeueing %p tid %llu\n", __func__, req,
1003 list_move_tail(&req->r_req_lru_item, &resend);
1004 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1006 list_move_tail(&req->r_req_lru_item, &resend_linger);
1009 list_splice(&resend, &osdc->req_unsent);
1012 * Both registered and not yet registered linger requests are
1013 * enqueued with a new tid on the same OSD. We add/move them
1014 * to req_unsent/o_requests at the end to keep things in tid
1017 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
1018 r_linger_osd_item) {
1019 WARN_ON(!list_empty(&req->r_req_lru_item));
1020 __kick_linger_request(req);
1023 list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
1024 __kick_linger_request(req);
1028 * If the osd connection drops, we need to resubmit all requests.
1030 static void osd_reset(struct ceph_connection *con)
1032 struct ceph_osd *osd = con->private;
1033 struct ceph_osd_client *osdc;
1037 dout("osd_reset osd%d\n", osd->o_osd);
1039 down_read(&osdc->map_sem);
1040 mutex_lock(&osdc->request_mutex);
1041 __kick_osd_requests(osdc, osd);
1042 __send_queued(osdc);
1043 mutex_unlock(&osdc->request_mutex);
1044 up_read(&osdc->map_sem);
1048 * Track open sessions with osds.
1050 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1052 struct ceph_osd *osd;
1054 osd = kzalloc(sizeof(*osd), GFP_NOFS);
1058 atomic_set(&osd->o_ref, 1);
1061 RB_CLEAR_NODE(&osd->o_node);
1062 INIT_LIST_HEAD(&osd->o_requests);
1063 INIT_LIST_HEAD(&osd->o_linger_requests);
1064 INIT_LIST_HEAD(&osd->o_osd_lru);
1065 osd->o_incarnation = 1;
1067 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1069 INIT_LIST_HEAD(&osd->o_keepalive_item);
1073 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1075 if (atomic_inc_not_zero(&osd->o_ref)) {
1076 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1077 atomic_read(&osd->o_ref));
1080 dout("get_osd %p FAIL\n", osd);
1085 static void put_osd(struct ceph_osd *osd)
1087 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1088 atomic_read(&osd->o_ref) - 1);
1089 if (atomic_dec_and_test(&osd->o_ref)) {
1090 if (osd->o_auth.authorizer)
1091 ceph_auth_destroy_authorizer(osd->o_auth.authorizer);
1097 * remove an osd from our map
1099 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1101 dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1102 WARN_ON(!list_empty(&osd->o_requests));
1103 WARN_ON(!list_empty(&osd->o_linger_requests));
1105 list_del_init(&osd->o_osd_lru);
1106 rb_erase(&osd->o_node, &osdc->osds);
1107 RB_CLEAR_NODE(&osd->o_node);
1110 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1112 dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1114 if (!RB_EMPTY_NODE(&osd->o_node)) {
1115 ceph_con_close(&osd->o_con);
1116 __remove_osd(osdc, osd);
1121 static void remove_all_osds(struct ceph_osd_client *osdc)
1123 dout("%s %p\n", __func__, osdc);
1124 mutex_lock(&osdc->request_mutex);
1125 while (!RB_EMPTY_ROOT(&osdc->osds)) {
1126 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1127 struct ceph_osd, o_node);
1128 remove_osd(osdc, osd);
1130 mutex_unlock(&osdc->request_mutex);
1133 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1134 struct ceph_osd *osd)
1136 dout("%s %p\n", __func__, osd);
1137 BUG_ON(!list_empty(&osd->o_osd_lru));
1139 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1140 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1143 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1144 struct ceph_osd *osd)
1146 dout("%s %p\n", __func__, osd);
1148 if (list_empty(&osd->o_requests) &&
1149 list_empty(&osd->o_linger_requests))
1150 __move_osd_to_lru(osdc, osd);
1153 static void __remove_osd_from_lru(struct ceph_osd *osd)
1155 dout("__remove_osd_from_lru %p\n", osd);
1156 if (!list_empty(&osd->o_osd_lru))
1157 list_del_init(&osd->o_osd_lru);
1160 static void remove_old_osds(struct ceph_osd_client *osdc)
1162 struct ceph_osd *osd, *nosd;
1164 dout("__remove_old_osds %p\n", osdc);
1165 mutex_lock(&osdc->request_mutex);
1166 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1167 if (time_before(jiffies, osd->lru_ttl))
1169 remove_osd(osdc, osd);
1171 mutex_unlock(&osdc->request_mutex);
1177 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1179 struct ceph_entity_addr *peer_addr;
1181 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1182 if (list_empty(&osd->o_requests) &&
1183 list_empty(&osd->o_linger_requests)) {
1184 remove_osd(osdc, osd);
1188 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1189 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1190 !ceph_con_opened(&osd->o_con)) {
1191 struct ceph_osd_request *req;
1193 dout("osd addr hasn't changed and connection never opened, "
1194 "letting msgr retry\n");
1195 /* touch each r_stamp for handle_timeout()'s benfit */
1196 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1197 req->r_stamp = jiffies;
1202 ceph_con_close(&osd->o_con);
1203 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1204 osd->o_incarnation++;
1209 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1211 struct rb_node **p = &osdc->osds.rb_node;
1212 struct rb_node *parent = NULL;
1213 struct ceph_osd *osd = NULL;
1215 dout("__insert_osd %p osd%d\n", new, new->o_osd);
1218 osd = rb_entry(parent, struct ceph_osd, o_node);
1219 if (new->o_osd < osd->o_osd)
1221 else if (new->o_osd > osd->o_osd)
1222 p = &(*p)->rb_right;
1227 rb_link_node(&new->o_node, parent, p);
1228 rb_insert_color(&new->o_node, &osdc->osds);
1231 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1233 struct ceph_osd *osd;
1234 struct rb_node *n = osdc->osds.rb_node;
1237 osd = rb_entry(n, struct ceph_osd, o_node);
1240 else if (o > osd->o_osd)
1248 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1250 schedule_delayed_work(&osdc->timeout_work,
1251 osdc->client->options->osd_keepalive_timeout);
1254 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1256 cancel_delayed_work(&osdc->timeout_work);
1260 * Register request, assign tid. If this is the first request, set up
1261 * the timeout event.
1263 static void __register_request(struct ceph_osd_client *osdc,
1264 struct ceph_osd_request *req)
1266 req->r_tid = ++osdc->last_tid;
1267 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1268 dout("__register_request %p tid %lld\n", req, req->r_tid);
1269 __insert_request(osdc, req);
1270 ceph_osdc_get_request(req);
1271 osdc->num_requests++;
1272 if (osdc->num_requests == 1) {
1273 dout(" first request, scheduling timeout\n");
1274 __schedule_osd_timeout(osdc);
1279 * called under osdc->request_mutex
1281 static void __unregister_request(struct ceph_osd_client *osdc,
1282 struct ceph_osd_request *req)
1284 if (RB_EMPTY_NODE(&req->r_node)) {
1285 dout("__unregister_request %p tid %lld not registered\n",
1290 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1291 rb_erase(&req->r_node, &osdc->requests);
1292 RB_CLEAR_NODE(&req->r_node);
1293 osdc->num_requests--;
1296 /* make sure the original request isn't in flight. */
1297 ceph_msg_revoke(req->r_request);
1299 list_del_init(&req->r_osd_item);
1300 maybe_move_osd_to_lru(osdc, req->r_osd);
1301 if (list_empty(&req->r_linger_osd_item))
1305 list_del_init(&req->r_req_lru_item);
1306 ceph_osdc_put_request(req);
1308 if (osdc->num_requests == 0) {
1309 dout(" no requests, canceling timeout\n");
1310 __cancel_osd_timeout(osdc);
1315 * Cancel a previously queued request message
1317 static void __cancel_request(struct ceph_osd_request *req)
1319 if (req->r_sent && req->r_osd) {
1320 ceph_msg_revoke(req->r_request);
1325 static void __register_linger_request(struct ceph_osd_client *osdc,
1326 struct ceph_osd_request *req)
1328 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1329 WARN_ON(!req->r_linger);
1331 ceph_osdc_get_request(req);
1332 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1334 list_add_tail(&req->r_linger_osd_item,
1335 &req->r_osd->o_linger_requests);
1338 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1339 struct ceph_osd_request *req)
1341 WARN_ON(!req->r_linger);
1343 if (list_empty(&req->r_linger_item)) {
1344 dout("%s %p tid %llu not registered\n", __func__, req,
1349 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1350 list_del_init(&req->r_linger_item);
1353 list_del_init(&req->r_linger_osd_item);
1354 maybe_move_osd_to_lru(osdc, req->r_osd);
1355 if (list_empty(&req->r_osd_item))
1358 ceph_osdc_put_request(req);
1361 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1362 struct ceph_osd_request *req)
1364 if (!req->r_linger) {
1365 dout("set_request_linger %p\n", req);
1369 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1372 * Returns whether a request should be blocked from being sent
1373 * based on the current osdmap and osd_client settings.
1375 * Caller should hold map_sem for read.
1377 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1378 struct ceph_osd_request *req)
1380 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1381 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1382 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1383 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1384 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1388 * Calculate mapping of a request to a PG. Takes tiering into account.
1390 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1391 struct ceph_osd_request *req,
1392 struct ceph_pg *pg_out)
1394 bool need_check_tiering;
1396 need_check_tiering = false;
1397 if (req->r_target_oloc.pool == -1) {
1398 req->r_target_oloc = req->r_base_oloc; /* struct */
1399 need_check_tiering = true;
1401 if (req->r_target_oid.name_len == 0) {
1402 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1403 need_check_tiering = true;
1406 if (need_check_tiering &&
1407 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1408 struct ceph_pg_pool_info *pi;
1410 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1412 if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1414 req->r_target_oloc.pool = pi->read_tier;
1415 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1416 pi->write_tier >= 0)
1417 req->r_target_oloc.pool = pi->write_tier;
1419 /* !pi is caught in ceph_oloc_oid_to_pg() */
1422 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1423 &req->r_target_oid, pg_out);
1426 static void __enqueue_request(struct ceph_osd_request *req)
1428 struct ceph_osd_client *osdc = req->r_osdc;
1430 dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1431 req->r_osd ? req->r_osd->o_osd : -1);
1434 __remove_osd_from_lru(req->r_osd);
1435 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1436 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1438 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1443 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1444 * (as needed), and set the request r_osd appropriately. If there is
1445 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1446 * (unsent, homeless) or leave on in-flight lru.
1448 * Return 0 if unchanged, 1 if changed, or negative on error.
1450 * Caller should hold map_sem for read and request_mutex.
1452 static int __map_request(struct ceph_osd_client *osdc,
1453 struct ceph_osd_request *req, int force_resend)
1455 struct ceph_pg pgid;
1456 int acting[CEPH_PG_MAX_SIZE];
1461 dout("map_request %p tid %lld\n", req, req->r_tid);
1463 err = __calc_request_pg(osdc->osdmap, req, &pgid);
1465 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1470 num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1474 was_paused = req->r_paused;
1475 req->r_paused = __req_should_be_paused(osdc, req);
1476 if (was_paused && !req->r_paused)
1479 if ((!force_resend &&
1480 req->r_osd && req->r_osd->o_osd == o &&
1481 req->r_sent >= req->r_osd->o_incarnation &&
1482 req->r_num_pg_osds == num &&
1483 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1484 (req->r_osd == NULL && o == -1) ||
1486 return 0; /* no change */
1488 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1489 req->r_tid, pgid.pool, pgid.seed, o,
1490 req->r_osd ? req->r_osd->o_osd : -1);
1492 /* record full pg acting set */
1493 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1494 req->r_num_pg_osds = num;
1497 __cancel_request(req);
1498 list_del_init(&req->r_osd_item);
1499 list_del_init(&req->r_linger_osd_item);
1503 req->r_osd = __lookup_osd(osdc, o);
1504 if (!req->r_osd && o >= 0) {
1506 req->r_osd = create_osd(osdc, o);
1508 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1512 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1513 __insert_osd(osdc, req->r_osd);
1515 ceph_con_open(&req->r_osd->o_con,
1516 CEPH_ENTITY_TYPE_OSD, o,
1517 &osdc->osdmap->osd_addr[o]);
1520 __enqueue_request(req);
1521 err = 1; /* osd or pg changed */
1528 * caller should hold map_sem (for read) and request_mutex
1530 static void __send_request(struct ceph_osd_client *osdc,
1531 struct ceph_osd_request *req)
1535 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1536 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1537 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1539 /* fill in message content that changes each time we send it */
1540 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1541 put_unaligned_le32(req->r_flags, req->r_request_flags);
1542 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1543 p = req->r_request_pgid;
1544 ceph_encode_64(&p, req->r_pgid.pool);
1545 ceph_encode_32(&p, req->r_pgid.seed);
1546 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1547 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1548 sizeof(req->r_reassert_version));
1550 req->r_stamp = jiffies;
1551 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1553 ceph_msg_get(req->r_request); /* send consumes a ref */
1555 req->r_sent = req->r_osd->o_incarnation;
1557 ceph_con_send(&req->r_osd->o_con, req->r_request);
1561 * Send any requests in the queue (req_unsent).
1563 static void __send_queued(struct ceph_osd_client *osdc)
1565 struct ceph_osd_request *req, *tmp;
1567 dout("__send_queued\n");
1568 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1569 __send_request(osdc, req);
1573 * Caller should hold map_sem for read and request_mutex.
1575 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1576 struct ceph_osd_request *req,
1581 __register_request(osdc, req);
1583 req->r_got_reply = 0;
1584 rc = __map_request(osdc, req, 0);
1587 dout("osdc_start_request failed map, "
1588 " will retry %lld\n", req->r_tid);
1591 __unregister_request(osdc, req);
1596 if (req->r_osd == NULL) {
1597 dout("send_request %p no up osds in pg\n", req);
1598 ceph_monc_request_next_osdmap(&osdc->client->monc);
1600 __send_queued(osdc);
1607 * Timeout callback, called every N seconds when 1 or more osd
1608 * requests has been active for more than N seconds. When this
1609 * happens, we ping all OSDs with requests who have timed out to
1610 * ensure any communications channel reset is detected. Reset the
1611 * request timeouts another N seconds in the future as we go.
1612 * Reschedule the timeout event another N seconds in future (unless
1613 * there are no open requests).
1615 static void handle_timeout(struct work_struct *work)
1617 struct ceph_osd_client *osdc =
1618 container_of(work, struct ceph_osd_client, timeout_work.work);
1619 struct ceph_options *opts = osdc->client->options;
1620 struct ceph_osd_request *req;
1621 struct ceph_osd *osd;
1622 struct list_head slow_osds;
1624 down_read(&osdc->map_sem);
1626 ceph_monc_request_next_osdmap(&osdc->client->monc);
1628 mutex_lock(&osdc->request_mutex);
1631 * ping osds that are a bit slow. this ensures that if there
1632 * is a break in the TCP connection we will notice, and reopen
1633 * a connection with that osd (from the fault callback).
1635 INIT_LIST_HEAD(&slow_osds);
1636 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1637 if (time_before(jiffies,
1638 req->r_stamp + opts->osd_keepalive_timeout))
1643 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1644 req->r_tid, osd->o_osd);
1645 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1647 while (!list_empty(&slow_osds)) {
1648 osd = list_entry(slow_osds.next, struct ceph_osd,
1650 list_del_init(&osd->o_keepalive_item);
1651 ceph_con_keepalive(&osd->o_con);
1654 __schedule_osd_timeout(osdc);
1655 __send_queued(osdc);
1656 mutex_unlock(&osdc->request_mutex);
1657 up_read(&osdc->map_sem);
1660 static void handle_osds_timeout(struct work_struct *work)
1662 struct ceph_osd_client *osdc =
1663 container_of(work, struct ceph_osd_client,
1664 osds_timeout_work.work);
1665 unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
1667 dout("osds timeout\n");
1668 down_read(&osdc->map_sem);
1669 remove_old_osds(osdc);
1670 up_read(&osdc->map_sem);
1672 schedule_delayed_work(&osdc->osds_timeout_work,
1673 round_jiffies_relative(delay));
1676 static int ceph_oloc_decode(void **p, void *end,
1677 struct ceph_object_locator *oloc)
1679 u8 struct_v, struct_cv;
1684 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1685 struct_v = ceph_decode_8(p);
1686 struct_cv = ceph_decode_8(p);
1688 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1689 struct_v, struct_cv);
1692 if (struct_cv > 6) {
1693 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1694 struct_v, struct_cv);
1697 len = ceph_decode_32(p);
1698 ceph_decode_need(p, end, len, e_inval);
1699 struct_end = *p + len;
1701 oloc->pool = ceph_decode_64(p);
1702 *p += 4; /* skip preferred */
1704 len = ceph_decode_32(p);
1706 pr_warn("ceph_object_locator::key is set\n");
1710 if (struct_v >= 5) {
1711 len = ceph_decode_32(p);
1713 pr_warn("ceph_object_locator::nspace is set\n");
1718 if (struct_v >= 6) {
1719 s64 hash = ceph_decode_64(p);
1721 pr_warn("ceph_object_locator::hash is set\n");
1736 static int ceph_redirect_decode(void **p, void *end,
1737 struct ceph_request_redirect *redir)
1739 u8 struct_v, struct_cv;
1744 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1745 struct_v = ceph_decode_8(p);
1746 struct_cv = ceph_decode_8(p);
1747 if (struct_cv > 1) {
1748 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1749 struct_v, struct_cv);
1752 len = ceph_decode_32(p);
1753 ceph_decode_need(p, end, len, e_inval);
1754 struct_end = *p + len;
1756 ret = ceph_oloc_decode(p, end, &redir->oloc);
1760 len = ceph_decode_32(p);
1762 pr_warn("ceph_request_redirect::object_name is set\n");
1766 len = ceph_decode_32(p);
1767 *p += len; /* skip osd_instructions */
1779 static void complete_request(struct ceph_osd_request *req)
1781 complete_all(&req->r_safe_completion); /* fsync waiter */
1785 * handle osd op reply. either call the callback if it is specified,
1786 * or do the completion to wake up the waiting thread.
1788 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1791 struct ceph_osd_request *req;
1792 struct ceph_request_redirect redir;
1795 unsigned int numops;
1796 int payload_len, flags;
1802 u64 reassert_version;
1804 int already_completed;
1809 tid = le64_to_cpu(msg->hdr.tid);
1810 dout("handle_reply %p tid %llu\n", msg, tid);
1812 p = msg->front.iov_base;
1813 end = p + msg->front.iov_len;
1815 ceph_decode_need(&p, end, 4, bad);
1816 object_len = ceph_decode_32(&p);
1817 ceph_decode_need(&p, end, object_len, bad);
1820 err = ceph_decode_pgid(&p, end, &pg);
1824 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1825 flags = ceph_decode_64(&p);
1826 result = ceph_decode_32(&p);
1827 reassert_epoch = ceph_decode_32(&p);
1828 reassert_version = ceph_decode_64(&p);
1829 osdmap_epoch = ceph_decode_32(&p);
1832 down_read(&osdc->map_sem);
1833 mutex_lock(&osdc->request_mutex);
1834 req = __lookup_request(osdc, tid);
1836 dout("handle_reply tid %llu dne\n", tid);
1839 ceph_osdc_get_request(req);
1841 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1844 ceph_decode_need(&p, end, 4, bad_put);
1845 numops = ceph_decode_32(&p);
1846 if (numops > CEPH_OSD_MAX_OPS)
1848 if (numops != req->r_num_ops)
1851 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1852 for (i = 0; i < numops; i++) {
1853 struct ceph_osd_op *op = p;
1856 len = le32_to_cpu(op->payload_len);
1857 req->r_ops[i].outdata_len = len;
1858 dout(" op %d has %d bytes\n", i, len);
1862 bytes = le32_to_cpu(msg->hdr.data_len);
1863 if (payload_len != bytes) {
1864 pr_warn("sum of op payload lens %d != data_len %d\n",
1865 payload_len, bytes);
1869 ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1870 retry_attempt = ceph_decode_32(&p);
1871 for (i = 0; i < numops; i++)
1872 req->r_ops[i].rval = ceph_decode_32(&p);
1874 if (le16_to_cpu(msg->hdr.version) >= 6) {
1875 p += 8 + 4; /* skip replay_version */
1876 p += 8; /* skip user_version */
1878 if (le16_to_cpu(msg->hdr.version) >= 7)
1879 ceph_decode_8_safe(&p, end, decode_redir, bad_put);
1887 err = ceph_redirect_decode(&p, end, &redir);
1891 redir.oloc.pool = -1;
1894 if (redir.oloc.pool != -1) {
1895 dout("redirect pool %lld\n", redir.oloc.pool);
1897 __unregister_request(osdc, req);
1899 req->r_target_oloc = redir.oloc; /* struct */
1902 * Start redirect requests with nofail=true. If
1903 * mapping fails, request will end up on the notarget
1904 * list, waiting for the new osdmap (which can take
1905 * a while), even though the original request mapped
1906 * successfully. In the future we might want to follow
1907 * original request's nofail setting here.
1909 err = __ceph_osdc_start_request(osdc, req, true);
1915 already_completed = req->r_got_reply;
1916 if (!req->r_got_reply) {
1917 req->r_result = result;
1918 dout("handle_reply result %d bytes %d\n", req->r_result,
1920 if (req->r_result == 0)
1921 req->r_result = bytes;
1923 /* in case this is a write and we need to replay, */
1924 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1925 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1927 req->r_got_reply = 1;
1928 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1929 dout("handle_reply tid %llu dup ack\n", tid);
1933 dout("handle_reply tid %llu flags %d\n", tid, flags);
1935 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1936 __register_linger_request(osdc, req);
1938 /* either this is a read, or we got the safe response */
1940 (flags & CEPH_OSD_FLAG_ONDISK) ||
1941 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1942 __unregister_request(osdc, req);
1944 mutex_unlock(&osdc->request_mutex);
1945 up_read(&osdc->map_sem);
1947 if (!already_completed) {
1948 if (req->r_unsafe_callback &&
1949 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1950 req->r_unsafe_callback(req, true);
1951 if (req->r_callback)
1952 req->r_callback(req, msg);
1954 complete_all(&req->r_completion);
1957 if (flags & CEPH_OSD_FLAG_ONDISK) {
1958 if (req->r_unsafe_callback && already_completed)
1959 req->r_unsafe_callback(req, false);
1960 complete_request(req);
1964 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1965 ceph_osdc_put_request(req);
1968 mutex_unlock(&osdc->request_mutex);
1969 up_read(&osdc->map_sem);
1973 req->r_result = -EIO;
1974 __unregister_request(osdc, req);
1975 if (req->r_callback)
1976 req->r_callback(req, msg);
1978 complete_all(&req->r_completion);
1979 complete_request(req);
1980 ceph_osdc_put_request(req);
1982 mutex_unlock(&osdc->request_mutex);
1983 up_read(&osdc->map_sem);
1985 pr_err("corrupt osd_op_reply got %d %d\n",
1986 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1990 static void reset_changed_osds(struct ceph_osd_client *osdc)
1992 struct rb_node *p, *n;
1994 dout("%s %p\n", __func__, osdc);
1995 for (p = rb_first(&osdc->osds); p; p = n) {
1996 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1999 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
2000 memcmp(&osd->o_con.peer_addr,
2001 ceph_osd_addr(osdc->osdmap,
2003 sizeof(struct ceph_entity_addr)) != 0)
2004 __reset_osd(osdc, osd);
2009 * Requeue requests whose mapping to an OSD has changed. If requests map to
2010 * no osd, request a new map.
2012 * Caller should hold map_sem for read.
2014 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
2015 bool force_resend_writes)
2017 struct ceph_osd_request *req, *nreq;
2021 bool force_resend_req;
2023 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
2024 force_resend_writes ? " (force resend writes)" : "");
2025 mutex_lock(&osdc->request_mutex);
2026 for (p = rb_first(&osdc->requests); p; ) {
2027 req = rb_entry(p, struct ceph_osd_request, r_node);
2031 * For linger requests that have not yet been
2032 * registered, move them to the linger list; they'll
2033 * be sent to the osd in the loop below. Unregister
2034 * the request before re-registering it as a linger
2035 * request to ensure the __map_request() below
2036 * will decide it needs to be sent.
2038 if (req->r_linger && list_empty(&req->r_linger_item)) {
2039 dout("%p tid %llu restart on osd%d\n",
2041 req->r_osd ? req->r_osd->o_osd : -1);
2042 ceph_osdc_get_request(req);
2043 __unregister_request(osdc, req);
2044 __register_linger_request(osdc, req);
2045 ceph_osdc_put_request(req);
2049 force_resend_req = force_resend ||
2050 (force_resend_writes &&
2051 req->r_flags & CEPH_OSD_FLAG_WRITE);
2052 err = __map_request(osdc, req, force_resend_req);
2054 continue; /* error */
2055 if (req->r_osd == NULL) {
2056 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2057 needmap++; /* request a newer map */
2058 } else if (err > 0) {
2059 if (!req->r_linger) {
2060 dout("%p tid %llu requeued on osd%d\n", req,
2062 req->r_osd ? req->r_osd->o_osd : -1);
2063 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2068 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2070 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2072 err = __map_request(osdc, req,
2073 force_resend || force_resend_writes);
2074 dout("__map_request returned %d\n", err);
2076 continue; /* hrm! */
2077 if (req->r_osd == NULL || err > 0) {
2078 if (req->r_osd == NULL) {
2079 dout("lingering %p tid %llu maps to no osd\n",
2082 * A homeless lingering request makes
2083 * no sense, as it's job is to keep
2084 * a particular OSD connection open.
2085 * Request a newer map and kick the
2086 * request, knowing that it won't be
2087 * resent until we actually get a map
2088 * that can tell us where to send it.
2093 dout("kicking lingering %p tid %llu osd%d\n", req,
2094 req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2095 __register_request(osdc, req);
2096 __unregister_linger_request(osdc, req);
2099 reset_changed_osds(osdc);
2100 mutex_unlock(&osdc->request_mutex);
2103 dout("%d requests for down osds, need new map\n", needmap);
2104 ceph_monc_request_next_osdmap(&osdc->client->monc);
2110 * Process updated osd map.
2112 * The message contains any number of incremental and full maps, normally
2113 * indicating some sort of topology change in the cluster. Kick requests
2114 * off to different OSDs as needed.
2116 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2118 void *p, *end, *next;
2119 u32 nr_maps, maplen;
2121 struct ceph_osdmap *newmap = NULL, *oldmap;
2123 struct ceph_fsid fsid;
2126 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2127 p = msg->front.iov_base;
2128 end = p + msg->front.iov_len;
2131 ceph_decode_need(&p, end, sizeof(fsid), bad);
2132 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2133 if (ceph_check_fsid(osdc->client, &fsid) < 0)
2136 down_write(&osdc->map_sem);
2138 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2140 /* incremental maps */
2141 ceph_decode_32_safe(&p, end, nr_maps, bad);
2142 dout(" %d inc maps\n", nr_maps);
2143 while (nr_maps > 0) {
2144 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2145 epoch = ceph_decode_32(&p);
2146 maplen = ceph_decode_32(&p);
2147 ceph_decode_need(&p, end, maplen, bad);
2149 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2150 dout("applying incremental map %u len %d\n",
2152 newmap = osdmap_apply_incremental(&p, next,
2154 &osdc->client->msgr);
2155 if (IS_ERR(newmap)) {
2156 err = PTR_ERR(newmap);
2160 if (newmap != osdc->osdmap) {
2161 ceph_osdmap_destroy(osdc->osdmap);
2162 osdc->osdmap = newmap;
2164 was_full = was_full ||
2165 ceph_osdmap_flag(osdc->osdmap,
2167 kick_requests(osdc, 0, was_full);
2169 dout("ignoring incremental map %u len %d\n",
2179 ceph_decode_32_safe(&p, end, nr_maps, bad);
2180 dout(" %d full maps\n", nr_maps);
2182 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2183 epoch = ceph_decode_32(&p);
2184 maplen = ceph_decode_32(&p);
2185 ceph_decode_need(&p, end, maplen, bad);
2187 dout("skipping non-latest full map %u len %d\n",
2189 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2190 dout("skipping full map %u len %d, "
2191 "older than our %u\n", epoch, maplen,
2192 osdc->osdmap->epoch);
2194 int skipped_map = 0;
2196 dout("taking full map %u len %d\n", epoch, maplen);
2197 newmap = ceph_osdmap_decode(&p, p+maplen);
2198 if (IS_ERR(newmap)) {
2199 err = PTR_ERR(newmap);
2203 oldmap = osdc->osdmap;
2204 osdc->osdmap = newmap;
2206 if (oldmap->epoch + 1 < newmap->epoch)
2208 ceph_osdmap_destroy(oldmap);
2210 was_full = was_full ||
2211 ceph_osdmap_flag(osdc->osdmap,
2213 kick_requests(osdc, skipped_map, was_full);
2222 downgrade_write(&osdc->map_sem);
2223 ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
2224 osdc->osdmap->epoch);
2227 * subscribe to subsequent osdmap updates if full to ensure
2228 * we find out when we are no longer full and stop returning
2231 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2232 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2233 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2234 ceph_monc_request_next_osdmap(&osdc->client->monc);
2236 mutex_lock(&osdc->request_mutex);
2237 __send_queued(osdc);
2238 mutex_unlock(&osdc->request_mutex);
2239 up_read(&osdc->map_sem);
2240 wake_up_all(&osdc->client->auth_wq);
2244 pr_err("osdc handle_map corrupt msg\n");
2246 up_write(&osdc->map_sem);
2250 * watch/notify callback event infrastructure
2252 * These callbacks are used both for watch and notify operations.
2254 static void __release_event(struct kref *kref)
2256 struct ceph_osd_event *event =
2257 container_of(kref, struct ceph_osd_event, kref);
2259 dout("__release_event %p\n", event);
2263 static void get_event(struct ceph_osd_event *event)
2265 kref_get(&event->kref);
2268 void ceph_osdc_put_event(struct ceph_osd_event *event)
2270 kref_put(&event->kref, __release_event);
2272 EXPORT_SYMBOL(ceph_osdc_put_event);
2274 static void __insert_event(struct ceph_osd_client *osdc,
2275 struct ceph_osd_event *new)
2277 struct rb_node **p = &osdc->event_tree.rb_node;
2278 struct rb_node *parent = NULL;
2279 struct ceph_osd_event *event = NULL;
2283 event = rb_entry(parent, struct ceph_osd_event, node);
2284 if (new->cookie < event->cookie)
2286 else if (new->cookie > event->cookie)
2287 p = &(*p)->rb_right;
2292 rb_link_node(&new->node, parent, p);
2293 rb_insert_color(&new->node, &osdc->event_tree);
2296 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2299 struct rb_node **p = &osdc->event_tree.rb_node;
2300 struct rb_node *parent = NULL;
2301 struct ceph_osd_event *event = NULL;
2305 event = rb_entry(parent, struct ceph_osd_event, node);
2306 if (cookie < event->cookie)
2308 else if (cookie > event->cookie)
2309 p = &(*p)->rb_right;
2316 static void __remove_event(struct ceph_osd_event *event)
2318 struct ceph_osd_client *osdc = event->osdc;
2320 if (!RB_EMPTY_NODE(&event->node)) {
2321 dout("__remove_event removed %p\n", event);
2322 rb_erase(&event->node, &osdc->event_tree);
2323 ceph_osdc_put_event(event);
2325 dout("__remove_event didn't remove %p\n", event);
2329 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2330 void (*event_cb)(u64, u64, u8, void *),
2331 void *data, struct ceph_osd_event **pevent)
2333 struct ceph_osd_event *event;
2335 event = kmalloc(sizeof(*event), GFP_NOIO);
2339 dout("create_event %p\n", event);
2340 event->cb = event_cb;
2341 event->one_shot = 0;
2344 INIT_LIST_HEAD(&event->osd_node);
2345 RB_CLEAR_NODE(&event->node);
2346 kref_init(&event->kref); /* one ref for us */
2347 kref_get(&event->kref); /* one ref for the caller */
2349 spin_lock(&osdc->event_lock);
2350 event->cookie = ++osdc->event_count;
2351 __insert_event(osdc, event);
2352 spin_unlock(&osdc->event_lock);
2357 EXPORT_SYMBOL(ceph_osdc_create_event);
2359 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2361 struct ceph_osd_client *osdc = event->osdc;
2363 dout("cancel_event %p\n", event);
2364 spin_lock(&osdc->event_lock);
2365 __remove_event(event);
2366 spin_unlock(&osdc->event_lock);
2367 ceph_osdc_put_event(event); /* caller's */
2369 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2372 static void do_event_work(struct work_struct *work)
2374 struct ceph_osd_event_work *event_work =
2375 container_of(work, struct ceph_osd_event_work, work);
2376 struct ceph_osd_event *event = event_work->event;
2377 u64 ver = event_work->ver;
2378 u64 notify_id = event_work->notify_id;
2379 u8 opcode = event_work->opcode;
2381 dout("do_event_work completing %p\n", event);
2382 event->cb(ver, notify_id, opcode, event->data);
2383 dout("do_event_work completed %p\n", event);
2384 ceph_osdc_put_event(event);
2390 * Process osd watch notifications
2392 static void handle_watch_notify(struct ceph_osd_client *osdc,
2393 struct ceph_msg *msg)
2397 u64 cookie, ver, notify_id;
2399 struct ceph_osd_event *event;
2400 struct ceph_osd_event_work *event_work;
2402 p = msg->front.iov_base;
2403 end = p + msg->front.iov_len;
2405 ceph_decode_8_safe(&p, end, proto_ver, bad);
2406 ceph_decode_8_safe(&p, end, opcode, bad);
2407 ceph_decode_64_safe(&p, end, cookie, bad);
2408 ceph_decode_64_safe(&p, end, ver, bad);
2409 ceph_decode_64_safe(&p, end, notify_id, bad);
2411 spin_lock(&osdc->event_lock);
2412 event = __find_event(osdc, cookie);
2414 BUG_ON(event->one_shot);
2417 spin_unlock(&osdc->event_lock);
2418 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2419 cookie, ver, event);
2421 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2423 pr_err("couldn't allocate event_work\n");
2424 ceph_osdc_put_event(event);
2427 INIT_WORK(&event_work->work, do_event_work);
2428 event_work->event = event;
2429 event_work->ver = ver;
2430 event_work->notify_id = notify_id;
2431 event_work->opcode = opcode;
2433 queue_work(osdc->notify_wq, &event_work->work);
2439 pr_err("osdc handle_watch_notify corrupt msg\n");
2443 * build new request AND message
2446 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2447 struct ceph_snap_context *snapc, u64 snap_id,
2448 struct timespec *mtime)
2450 struct ceph_msg *msg = req->r_request;
2453 int flags = req->r_flags;
2457 req->r_snapid = snap_id;
2458 req->r_snapc = ceph_get_snap_context(snapc);
2460 /* encode request */
2461 msg->hdr.version = cpu_to_le16(4);
2463 p = msg->front.iov_base;
2464 ceph_encode_32(&p, 1); /* client_inc is always 1 */
2465 req->r_request_osdmap_epoch = p;
2467 req->r_request_flags = p;
2469 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2470 ceph_encode_timespec(p, mtime);
2471 p += sizeof(struct ceph_timespec);
2472 req->r_request_reassert_version = p;
2473 p += sizeof(struct ceph_eversion); /* will get filled in */
2476 ceph_encode_8(&p, 4);
2477 ceph_encode_8(&p, 4);
2478 ceph_encode_32(&p, 8 + 4 + 4);
2479 req->r_request_pool = p;
2481 ceph_encode_32(&p, -1); /* preferred */
2482 ceph_encode_32(&p, 0); /* key len */
2484 ceph_encode_8(&p, 1);
2485 req->r_request_pgid = p;
2487 ceph_encode_32(&p, -1); /* preferred */
2490 ceph_encode_32(&p, req->r_base_oid.name_len);
2491 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2492 dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2493 req->r_base_oid.name, req->r_base_oid.name_len);
2494 p += req->r_base_oid.name_len;
2496 /* ops--can imply data */
2497 ceph_encode_16(&p, (u16)req->r_num_ops);
2499 for (i = 0; i < req->r_num_ops; i++) {
2500 data_len += osd_req_encode_op(req, p, i);
2501 p += sizeof(struct ceph_osd_op);
2505 ceph_encode_64(&p, req->r_snapid);
2506 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2507 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2509 for (i = 0; i < snapc->num_snaps; i++) {
2510 ceph_encode_64(&p, req->r_snapc->snaps[i]);
2514 req->r_request_attempts = p;
2518 if (flags & CEPH_OSD_FLAG_WRITE) {
2522 * The header "data_off" is a hint to the receiver
2523 * allowing it to align received data into its
2524 * buffers such that there's no need to re-copy
2525 * it before writing it to disk (direct I/O).
2527 data_off = (u16) (off & 0xffff);
2528 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2530 req->r_request->hdr.data_len = cpu_to_le32(data_len);
2532 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2533 msg_size = p - msg->front.iov_base;
2534 msg->front.iov_len = msg_size;
2535 msg->hdr.front_len = cpu_to_le32(msg_size);
2537 dout("build_request msg_size was %d\n", (int)msg_size);
2539 EXPORT_SYMBOL(ceph_osdc_build_request);
2542 * Register request, send initial attempt.
2544 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2545 struct ceph_osd_request *req,
2550 down_read(&osdc->map_sem);
2551 mutex_lock(&osdc->request_mutex);
2553 rc = __ceph_osdc_start_request(osdc, req, nofail);
2555 mutex_unlock(&osdc->request_mutex);
2556 up_read(&osdc->map_sem);
2560 EXPORT_SYMBOL(ceph_osdc_start_request);
2563 * Unregister a registered request. The request is not completed (i.e.
2564 * no callbacks or wakeups) - higher layers are supposed to know what
2565 * they are canceling.
2567 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2569 struct ceph_osd_client *osdc = req->r_osdc;
2571 mutex_lock(&osdc->request_mutex);
2573 __unregister_linger_request(osdc, req);
2574 __unregister_request(osdc, req);
2575 mutex_unlock(&osdc->request_mutex);
2577 dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2579 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2582 * wait for a request to complete
2584 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2585 struct ceph_osd_request *req)
2589 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2591 rc = wait_for_completion_interruptible(&req->r_completion);
2593 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2594 ceph_osdc_cancel_request(req);
2595 complete_request(req);
2599 dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2601 return req->r_result;
2603 EXPORT_SYMBOL(ceph_osdc_wait_request);
2606 * sync - wait for all in-flight requests to flush. avoid starvation.
2608 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2610 struct ceph_osd_request *req;
2611 u64 last_tid, next_tid = 0;
2613 mutex_lock(&osdc->request_mutex);
2614 last_tid = osdc->last_tid;
2616 req = __lookup_request_ge(osdc, next_tid);
2619 if (req->r_tid > last_tid)
2622 next_tid = req->r_tid + 1;
2623 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2626 ceph_osdc_get_request(req);
2627 mutex_unlock(&osdc->request_mutex);
2628 dout("sync waiting on tid %llu (last is %llu)\n",
2629 req->r_tid, last_tid);
2630 wait_for_completion(&req->r_safe_completion);
2631 mutex_lock(&osdc->request_mutex);
2632 ceph_osdc_put_request(req);
2634 mutex_unlock(&osdc->request_mutex);
2635 dout("sync done (thru tid %llu)\n", last_tid);
2637 EXPORT_SYMBOL(ceph_osdc_sync);
2640 * Call all pending notify callbacks - for use after a watch is
2641 * unregistered, to make sure no more callbacks for it will be invoked
2643 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2645 flush_workqueue(osdc->notify_wq);
2647 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2653 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2658 osdc->client = client;
2659 osdc->osdmap = NULL;
2660 init_rwsem(&osdc->map_sem);
2661 init_completion(&osdc->map_waiters);
2662 osdc->last_requested_map = 0;
2663 mutex_init(&osdc->request_mutex);
2665 osdc->osds = RB_ROOT;
2666 INIT_LIST_HEAD(&osdc->osd_lru);
2667 osdc->requests = RB_ROOT;
2668 INIT_LIST_HEAD(&osdc->req_lru);
2669 INIT_LIST_HEAD(&osdc->req_unsent);
2670 INIT_LIST_HEAD(&osdc->req_notarget);
2671 INIT_LIST_HEAD(&osdc->req_linger);
2672 osdc->num_requests = 0;
2673 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2674 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2675 spin_lock_init(&osdc->event_lock);
2676 osdc->event_tree = RB_ROOT;
2677 osdc->event_count = 0;
2679 schedule_delayed_work(&osdc->osds_timeout_work,
2680 round_jiffies_relative(osdc->client->options->osd_idle_ttl));
2683 osdc->req_mempool = mempool_create_slab_pool(10,
2684 ceph_osd_request_cache);
2685 if (!osdc->req_mempool)
2688 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2689 OSD_OP_FRONT_LEN, 10, true,
2693 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2694 OSD_OPREPLY_FRONT_LEN, 10, true,
2700 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2701 if (!osdc->notify_wq)
2702 goto out_msgpool_reply;
2707 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2709 ceph_msgpool_destroy(&osdc->msgpool_op);
2711 mempool_destroy(osdc->req_mempool);
2716 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2718 flush_workqueue(osdc->notify_wq);
2719 destroy_workqueue(osdc->notify_wq);
2720 cancel_delayed_work_sync(&osdc->timeout_work);
2721 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2723 ceph_osdmap_destroy(osdc->osdmap);
2724 osdc->osdmap = NULL;
2726 remove_all_osds(osdc);
2727 mempool_destroy(osdc->req_mempool);
2728 ceph_msgpool_destroy(&osdc->msgpool_op);
2729 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2733 * Read some contiguous pages. If we cross a stripe boundary, shorten
2734 * *plen. Return number of bytes read, or error.
2736 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2737 struct ceph_vino vino, struct ceph_file_layout *layout,
2739 u32 truncate_seq, u64 truncate_size,
2740 struct page **pages, int num_pages, int page_align)
2742 struct ceph_osd_request *req;
2745 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2746 vino.snap, off, *plen);
2747 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2748 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2749 NULL, truncate_seq, truncate_size,
2752 return PTR_ERR(req);
2754 /* it may be a short read due to an object boundary */
2756 osd_req_op_extent_osd_data_pages(req, 0,
2757 pages, *plen, page_align, false, false);
2759 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2760 off, *plen, *plen, page_align);
2762 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2764 rc = ceph_osdc_start_request(osdc, req, false);
2766 rc = ceph_osdc_wait_request(osdc, req);
2768 ceph_osdc_put_request(req);
2769 dout("readpages result %d\n", rc);
2772 EXPORT_SYMBOL(ceph_osdc_readpages);
2775 * do a synchronous write on N pages
2777 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2778 struct ceph_file_layout *layout,
2779 struct ceph_snap_context *snapc,
2781 u32 truncate_seq, u64 truncate_size,
2782 struct timespec *mtime,
2783 struct page **pages, int num_pages)
2785 struct ceph_osd_request *req;
2787 int page_align = off & ~PAGE_MASK;
2789 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */
2790 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2792 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2793 snapc, truncate_seq, truncate_size,
2796 return PTR_ERR(req);
2798 /* it may be a short write due to an object boundary */
2799 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2801 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2803 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2805 rc = ceph_osdc_start_request(osdc, req, true);
2807 rc = ceph_osdc_wait_request(osdc, req);
2809 ceph_osdc_put_request(req);
2812 dout("writepages result %d\n", rc);
2815 EXPORT_SYMBOL(ceph_osdc_writepages);
2817 int ceph_osdc_setup(void)
2819 size_t size = sizeof(struct ceph_osd_request) +
2820 CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op);
2822 BUG_ON(ceph_osd_request_cache);
2823 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size,
2826 return ceph_osd_request_cache ? 0 : -ENOMEM;
2828 EXPORT_SYMBOL(ceph_osdc_setup);
2830 void ceph_osdc_cleanup(void)
2832 BUG_ON(!ceph_osd_request_cache);
2833 kmem_cache_destroy(ceph_osd_request_cache);
2834 ceph_osd_request_cache = NULL;
2836 EXPORT_SYMBOL(ceph_osdc_cleanup);
2839 * handle incoming message
2841 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2843 struct ceph_osd *osd = con->private;
2844 struct ceph_osd_client *osdc;
2845 int type = le16_to_cpu(msg->hdr.type);
2852 case CEPH_MSG_OSD_MAP:
2853 ceph_osdc_handle_map(osdc, msg);
2855 case CEPH_MSG_OSD_OPREPLY:
2856 handle_reply(osdc, msg);
2858 case CEPH_MSG_WATCH_NOTIFY:
2859 handle_watch_notify(osdc, msg);
2863 pr_err("received unknown message type %d %s\n", type,
2864 ceph_msg_type_name(type));
2871 * Lookup and return message for incoming reply. Don't try to do
2872 * anything about a larger than preallocated data portion of the
2873 * message at the moment - for now, just skip the message.
2875 static struct ceph_msg *get_reply(struct ceph_connection *con,
2876 struct ceph_msg_header *hdr,
2879 struct ceph_osd *osd = con->private;
2880 struct ceph_osd_client *osdc = osd->o_osdc;
2882 struct ceph_osd_request *req;
2883 int front_len = le32_to_cpu(hdr->front_len);
2884 int data_len = le32_to_cpu(hdr->data_len);
2887 tid = le64_to_cpu(hdr->tid);
2888 mutex_lock(&osdc->request_mutex);
2889 req = __lookup_request(osdc, tid);
2891 dout("%s osd%d tid %llu unknown, skipping\n", __func__,
2898 ceph_msg_revoke_incoming(req->r_reply);
2900 if (front_len > req->r_reply->front_alloc_len) {
2901 pr_warn("%s osd%d tid %llu front %d > preallocated %d\n",
2902 __func__, osd->o_osd, req->r_tid, front_len,
2903 req->r_reply->front_alloc_len);
2904 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2908 ceph_msg_put(req->r_reply);
2912 if (data_len > req->r_reply->data_length) {
2913 pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n",
2914 __func__, osd->o_osd, req->r_tid, data_len,
2915 req->r_reply->data_length);
2921 m = ceph_msg_get(req->r_reply);
2922 dout("get_reply tid %lld %p\n", tid, m);
2925 mutex_unlock(&osdc->request_mutex);
2929 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2930 struct ceph_msg_header *hdr,
2933 struct ceph_osd *osd = con->private;
2934 int type = le16_to_cpu(hdr->type);
2935 int front = le32_to_cpu(hdr->front_len);
2939 case CEPH_MSG_OSD_MAP:
2940 case CEPH_MSG_WATCH_NOTIFY:
2941 return ceph_msg_new(type, front, GFP_NOFS, false);
2942 case CEPH_MSG_OSD_OPREPLY:
2943 return get_reply(con, hdr, skip);
2945 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2953 * Wrappers to refcount containing ceph_osd struct
2955 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2957 struct ceph_osd *osd = con->private;
2963 static void put_osd_con(struct ceph_connection *con)
2965 struct ceph_osd *osd = con->private;
2973 * Note: returned pointer is the address of a structure that's
2974 * managed separately. Caller must *not* attempt to free it.
2976 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2977 int *proto, int force_new)
2979 struct ceph_osd *o = con->private;
2980 struct ceph_osd_client *osdc = o->o_osdc;
2981 struct ceph_auth_client *ac = osdc->client->monc.auth;
2982 struct ceph_auth_handshake *auth = &o->o_auth;
2984 if (force_new && auth->authorizer) {
2985 ceph_auth_destroy_authorizer(auth->authorizer);
2986 auth->authorizer = NULL;
2988 if (!auth->authorizer) {
2989 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2992 return ERR_PTR(ret);
2994 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2997 return ERR_PTR(ret);
2999 *proto = ac->protocol;
3005 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3007 struct ceph_osd *o = con->private;
3008 struct ceph_osd_client *osdc = o->o_osdc;
3009 struct ceph_auth_client *ac = osdc->client->monc.auth;
3011 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
3014 static int invalidate_authorizer(struct ceph_connection *con)
3016 struct ceph_osd *o = con->private;
3017 struct ceph_osd_client *osdc = o->o_osdc;
3018 struct ceph_auth_client *ac = osdc->client->monc.auth;
3020 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
3021 return ceph_monc_validate_auth(&osdc->client->monc);
3024 static int osd_sign_message(struct ceph_msg *msg)
3026 struct ceph_osd *o = msg->con->private;
3027 struct ceph_auth_handshake *auth = &o->o_auth;
3029 return ceph_auth_sign_message(auth, msg);
3032 static int osd_check_message_signature(struct ceph_msg *msg)
3034 struct ceph_osd *o = msg->con->private;
3035 struct ceph_auth_handshake *auth = &o->o_auth;
3037 return ceph_auth_check_message_signature(auth, msg);
3040 static const struct ceph_connection_operations osd_con_ops = {
3043 .dispatch = dispatch,
3044 .get_authorizer = get_authorizer,
3045 .verify_authorizer_reply = verify_authorizer_reply,
3046 .invalidate_authorizer = invalidate_authorizer,
3047 .alloc_msg = alloc_msg,
3048 .sign_message = osd_sign_message,
3049 .check_message_signature = osd_check_message_signature,