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_OPREPLY_FRONT_LEN 512
24 static struct kmem_cache *ceph_osd_request_cache;
26 static const struct ceph_connection_operations osd_con_ops;
28 static void __send_queued(struct ceph_osd_client *osdc);
29 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
30 static void __register_request(struct ceph_osd_client *osdc,
31 struct ceph_osd_request *req);
32 static void __unregister_request(struct ceph_osd_client *osdc,
33 struct ceph_osd_request *req);
34 static void __unregister_linger_request(struct ceph_osd_client *osdc,
35 struct ceph_osd_request *req);
36 static void __enqueue_request(struct ceph_osd_request *req);
37 static void __send_request(struct ceph_osd_client *osdc,
38 struct ceph_osd_request *req);
41 * Implement client access to distributed object storage cluster.
43 * All data objects are stored within a cluster/cloud of OSDs, or
44 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
45 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
46 * remote daemons serving up and coordinating consistent and safe
49 * Cluster membership and the mapping of data objects onto storage devices
50 * are described by the osd map.
52 * We keep track of pending OSD requests (read, write), resubmit
53 * requests to different OSDs when the cluster topology/data layout
54 * change, or retry the affected requests when the communications
55 * channel with an OSD is reset.
59 * calculate the mapping of a file extent onto an object, and fill out the
60 * request accordingly. shorten extent as necessary if it crosses an
63 * fill osd op in request message.
65 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
66 u64 *objnum, u64 *objoff, u64 *objlen)
72 r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
76 if (*objlen < orig_len) {
78 dout(" skipping last %llu, final file extent %llu~%llu\n",
79 orig_len - *plen, off, *plen);
82 dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
87 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
89 memset(osd_data, 0, sizeof (*osd_data));
90 osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
93 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
94 struct page **pages, u64 length, u32 alignment,
95 bool pages_from_pool, bool own_pages)
97 osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
98 osd_data->pages = pages;
99 osd_data->length = length;
100 osd_data->alignment = alignment;
101 osd_data->pages_from_pool = pages_from_pool;
102 osd_data->own_pages = own_pages;
105 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
106 struct ceph_pagelist *pagelist)
108 osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
109 osd_data->pagelist = pagelist;
113 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
114 struct bio *bio, size_t bio_length)
116 osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
118 osd_data->bio_length = bio_length;
120 #endif /* CONFIG_BLOCK */
122 #define osd_req_op_data(oreq, whch, typ, fld) \
124 struct ceph_osd_request *__oreq = (oreq); \
125 unsigned int __whch = (whch); \
126 BUG_ON(__whch >= __oreq->r_num_ops); \
127 &__oreq->r_ops[__whch].typ.fld; \
130 static struct ceph_osd_data *
131 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
133 BUG_ON(which >= osd_req->r_num_ops);
135 return &osd_req->r_ops[which].raw_data_in;
138 struct ceph_osd_data *
139 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
142 return osd_req_op_data(osd_req, which, extent, osd_data);
144 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
146 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
147 unsigned int which, struct page **pages,
148 u64 length, u32 alignment,
149 bool pages_from_pool, bool own_pages)
151 struct ceph_osd_data *osd_data;
153 osd_data = osd_req_op_raw_data_in(osd_req, which);
154 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
155 pages_from_pool, own_pages);
157 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
159 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
160 unsigned int which, struct page **pages,
161 u64 length, u32 alignment,
162 bool pages_from_pool, bool own_pages)
164 struct ceph_osd_data *osd_data;
166 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
167 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
168 pages_from_pool, own_pages);
170 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
172 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
173 unsigned int which, struct ceph_pagelist *pagelist)
175 struct ceph_osd_data *osd_data;
177 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
178 ceph_osd_data_pagelist_init(osd_data, pagelist);
180 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
183 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
184 unsigned int which, struct bio *bio, size_t bio_length)
186 struct ceph_osd_data *osd_data;
188 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
189 ceph_osd_data_bio_init(osd_data, bio, bio_length);
191 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
192 #endif /* CONFIG_BLOCK */
194 static void osd_req_op_cls_request_info_pagelist(
195 struct ceph_osd_request *osd_req,
196 unsigned int which, struct ceph_pagelist *pagelist)
198 struct ceph_osd_data *osd_data;
200 osd_data = osd_req_op_data(osd_req, which, cls, request_info);
201 ceph_osd_data_pagelist_init(osd_data, pagelist);
204 void osd_req_op_cls_request_data_pagelist(
205 struct ceph_osd_request *osd_req,
206 unsigned int which, struct ceph_pagelist *pagelist)
208 struct ceph_osd_data *osd_data;
210 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
211 ceph_osd_data_pagelist_init(osd_data, pagelist);
213 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
215 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
216 unsigned int which, struct page **pages, u64 length,
217 u32 alignment, bool pages_from_pool, bool own_pages)
219 struct ceph_osd_data *osd_data;
221 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
222 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
223 pages_from_pool, own_pages);
225 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
227 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
228 unsigned int which, struct page **pages, u64 length,
229 u32 alignment, bool pages_from_pool, bool own_pages)
231 struct ceph_osd_data *osd_data;
233 osd_data = osd_req_op_data(osd_req, which, cls, response_data);
234 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
235 pages_from_pool, own_pages);
237 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
239 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
241 switch (osd_data->type) {
242 case CEPH_OSD_DATA_TYPE_NONE:
244 case CEPH_OSD_DATA_TYPE_PAGES:
245 return osd_data->length;
246 case CEPH_OSD_DATA_TYPE_PAGELIST:
247 return (u64)osd_data->pagelist->length;
249 case CEPH_OSD_DATA_TYPE_BIO:
250 return (u64)osd_data->bio_length;
251 #endif /* CONFIG_BLOCK */
253 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
258 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
260 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
263 num_pages = calc_pages_for((u64)osd_data->alignment,
264 (u64)osd_data->length);
265 ceph_release_page_vector(osd_data->pages, num_pages);
267 ceph_osd_data_init(osd_data);
270 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
273 struct ceph_osd_req_op *op;
275 BUG_ON(which >= osd_req->r_num_ops);
276 op = &osd_req->r_ops[which];
279 case CEPH_OSD_OP_READ:
280 case CEPH_OSD_OP_WRITE:
281 case CEPH_OSD_OP_WRITEFULL:
282 ceph_osd_data_release(&op->extent.osd_data);
284 case CEPH_OSD_OP_CALL:
285 ceph_osd_data_release(&op->cls.request_info);
286 ceph_osd_data_release(&op->cls.request_data);
287 ceph_osd_data_release(&op->cls.response_data);
289 case CEPH_OSD_OP_SETXATTR:
290 case CEPH_OSD_OP_CMPXATTR:
291 ceph_osd_data_release(&op->xattr.osd_data);
293 case CEPH_OSD_OP_STAT:
294 ceph_osd_data_release(&op->raw_data_in);
304 static void ceph_osdc_release_request(struct kref *kref)
306 struct ceph_osd_request *req = container_of(kref,
307 struct ceph_osd_request, r_kref);
310 dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
311 req->r_request, req->r_reply);
312 WARN_ON(!RB_EMPTY_NODE(&req->r_node));
313 WARN_ON(!list_empty(&req->r_req_lru_item));
314 WARN_ON(!list_empty(&req->r_osd_item));
315 WARN_ON(!list_empty(&req->r_linger_item));
316 WARN_ON(!list_empty(&req->r_linger_osd_item));
320 ceph_msg_put(req->r_request);
322 ceph_msg_revoke_incoming(req->r_reply);
323 ceph_msg_put(req->r_reply);
326 for (which = 0; which < req->r_num_ops; which++)
327 osd_req_op_data_release(req, which);
329 ceph_oid_destroy(&req->r_base_oid);
330 ceph_oid_destroy(&req->r_target_oid);
331 ceph_put_snap_context(req->r_snapc);
334 mempool_free(req, req->r_osdc->req_mempool);
335 else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS)
336 kmem_cache_free(ceph_osd_request_cache, req);
341 void ceph_osdc_get_request(struct ceph_osd_request *req)
343 dout("%s %p (was %d)\n", __func__, req,
344 atomic_read(&req->r_kref.refcount));
345 kref_get(&req->r_kref);
347 EXPORT_SYMBOL(ceph_osdc_get_request);
349 void ceph_osdc_put_request(struct ceph_osd_request *req)
352 dout("%s %p (was %d)\n", __func__, req,
353 atomic_read(&req->r_kref.refcount));
354 kref_put(&req->r_kref, ceph_osdc_release_request);
357 EXPORT_SYMBOL(ceph_osdc_put_request);
359 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
360 struct ceph_snap_context *snapc,
361 unsigned int num_ops,
365 struct ceph_osd_request *req;
368 BUG_ON(num_ops > CEPH_OSD_SLAB_OPS);
369 req = mempool_alloc(osdc->req_mempool, gfp_flags);
370 } else if (num_ops <= CEPH_OSD_SLAB_OPS) {
371 req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags);
373 BUG_ON(num_ops > CEPH_OSD_MAX_OPS);
374 req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]),
380 /* req only, each op is zeroed in _osd_req_op_init() */
381 memset(req, 0, sizeof(*req));
384 req->r_mempool = use_mempool;
385 req->r_num_ops = num_ops;
386 req->r_snapid = CEPH_NOSNAP;
387 req->r_snapc = ceph_get_snap_context(snapc);
389 kref_init(&req->r_kref);
390 init_completion(&req->r_completion);
391 init_completion(&req->r_safe_completion);
392 RB_CLEAR_NODE(&req->r_node);
393 INIT_LIST_HEAD(&req->r_unsafe_item);
394 INIT_LIST_HEAD(&req->r_linger_item);
395 INIT_LIST_HEAD(&req->r_linger_osd_item);
396 INIT_LIST_HEAD(&req->r_req_lru_item);
397 INIT_LIST_HEAD(&req->r_osd_item);
399 ceph_oid_init(&req->r_base_oid);
400 req->r_base_oloc.pool = -1;
401 ceph_oid_init(&req->r_target_oid);
402 req->r_target_oloc.pool = -1;
404 dout("%s req %p\n", __func__, req);
407 EXPORT_SYMBOL(ceph_osdc_alloc_request);
409 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp)
411 struct ceph_osd_client *osdc = req->r_osdc;
412 struct ceph_msg *msg;
415 WARN_ON(ceph_oid_empty(&req->r_base_oid));
417 /* create request message */
418 msg_size = 4 + 4 + 4; /* client_inc, osdmap_epoch, flags */
419 msg_size += 4 + 4 + 4 + 8; /* mtime, reassert_version */
420 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
421 msg_size += 1 + 8 + 4 + 4; /* pgid */
422 msg_size += 4 + req->r_base_oid.name_len; /* oid */
423 msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op);
424 msg_size += 8; /* snapid */
425 msg_size += 8; /* snap_seq */
426 msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0);
427 msg_size += 4; /* retry_attempt */
430 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
432 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp, true);
436 memset(msg->front.iov_base, 0, msg->front.iov_len);
437 req->r_request = msg;
439 /* create reply message */
440 msg_size = OSD_OPREPLY_FRONT_LEN;
441 msg_size += req->r_base_oid.name_len;
442 msg_size += req->r_num_ops * sizeof(struct ceph_osd_op);
445 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
447 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size, gfp, true);
455 EXPORT_SYMBOL(ceph_osdc_alloc_messages);
457 static bool osd_req_opcode_valid(u16 opcode)
460 #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true;
461 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
469 * This is an osd op init function for opcodes that have no data or
470 * other information associated with them. It also serves as a
471 * common init routine for all the other init functions, below.
473 static struct ceph_osd_req_op *
474 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
475 u16 opcode, u32 flags)
477 struct ceph_osd_req_op *op;
479 BUG_ON(which >= osd_req->r_num_ops);
480 BUG_ON(!osd_req_opcode_valid(opcode));
482 op = &osd_req->r_ops[which];
483 memset(op, 0, sizeof (*op));
490 void osd_req_op_init(struct ceph_osd_request *osd_req,
491 unsigned int which, u16 opcode, u32 flags)
493 (void)_osd_req_op_init(osd_req, which, opcode, flags);
495 EXPORT_SYMBOL(osd_req_op_init);
497 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
498 unsigned int which, u16 opcode,
499 u64 offset, u64 length,
500 u64 truncate_size, u32 truncate_seq)
502 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
504 size_t payload_len = 0;
506 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
507 opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO &&
508 opcode != CEPH_OSD_OP_TRUNCATE);
510 op->extent.offset = offset;
511 op->extent.length = length;
512 op->extent.truncate_size = truncate_size;
513 op->extent.truncate_seq = truncate_seq;
514 if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL)
515 payload_len += length;
517 op->indata_len = payload_len;
519 EXPORT_SYMBOL(osd_req_op_extent_init);
521 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
522 unsigned int which, u64 length)
524 struct ceph_osd_req_op *op;
527 BUG_ON(which >= osd_req->r_num_ops);
528 op = &osd_req->r_ops[which];
529 previous = op->extent.length;
531 if (length == previous)
532 return; /* Nothing to do */
533 BUG_ON(length > previous);
535 op->extent.length = length;
536 op->indata_len -= previous - length;
538 EXPORT_SYMBOL(osd_req_op_extent_update);
540 void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
541 unsigned int which, u64 offset_inc)
543 struct ceph_osd_req_op *op, *prev_op;
545 BUG_ON(which + 1 >= osd_req->r_num_ops);
547 prev_op = &osd_req->r_ops[which];
548 op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags);
549 /* dup previous one */
550 op->indata_len = prev_op->indata_len;
551 op->outdata_len = prev_op->outdata_len;
552 op->extent = prev_op->extent;
554 op->extent.offset += offset_inc;
555 op->extent.length -= offset_inc;
557 if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
558 op->indata_len -= offset_inc;
560 EXPORT_SYMBOL(osd_req_op_extent_dup_last);
562 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
563 u16 opcode, const char *class, const char *method)
565 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
567 struct ceph_pagelist *pagelist;
568 size_t payload_len = 0;
571 BUG_ON(opcode != CEPH_OSD_OP_CALL);
573 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
575 ceph_pagelist_init(pagelist);
577 op->cls.class_name = class;
578 size = strlen(class);
579 BUG_ON(size > (size_t) U8_MAX);
580 op->cls.class_len = size;
581 ceph_pagelist_append(pagelist, class, size);
584 op->cls.method_name = method;
585 size = strlen(method);
586 BUG_ON(size > (size_t) U8_MAX);
587 op->cls.method_len = size;
588 ceph_pagelist_append(pagelist, method, size);
591 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
593 op->cls.argc = 0; /* currently unused */
595 op->indata_len = payload_len;
597 EXPORT_SYMBOL(osd_req_op_cls_init);
599 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
600 u16 opcode, const char *name, const void *value,
601 size_t size, u8 cmp_op, u8 cmp_mode)
603 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
605 struct ceph_pagelist *pagelist;
608 BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
610 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
614 ceph_pagelist_init(pagelist);
616 payload_len = strlen(name);
617 op->xattr.name_len = payload_len;
618 ceph_pagelist_append(pagelist, name, payload_len);
620 op->xattr.value_len = size;
621 ceph_pagelist_append(pagelist, value, size);
624 op->xattr.cmp_op = cmp_op;
625 op->xattr.cmp_mode = cmp_mode;
627 ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
628 op->indata_len = payload_len;
631 EXPORT_SYMBOL(osd_req_op_xattr_init);
633 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
634 unsigned int which, u16 opcode,
635 u64 cookie, u64 version, int flag)
637 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
640 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
642 op->watch.cookie = cookie;
643 op->watch.ver = version;
644 if (opcode == CEPH_OSD_OP_WATCH && flag)
645 op->watch.flag = (u8)1;
647 EXPORT_SYMBOL(osd_req_op_watch_init);
649 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
651 u64 expected_object_size,
652 u64 expected_write_size)
654 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
655 CEPH_OSD_OP_SETALLOCHINT,
658 op->alloc_hint.expected_object_size = expected_object_size;
659 op->alloc_hint.expected_write_size = expected_write_size;
662 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
663 * not worth a feature bit. Set FAILOK per-op flag to make
664 * sure older osds don't trip over an unsupported opcode.
666 op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
668 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
670 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
671 struct ceph_osd_data *osd_data)
673 u64 length = ceph_osd_data_length(osd_data);
675 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
676 BUG_ON(length > (u64) SIZE_MAX);
678 ceph_msg_data_add_pages(msg, osd_data->pages,
679 length, osd_data->alignment);
680 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
682 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
684 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
685 ceph_msg_data_add_bio(msg, osd_data->bio, length);
688 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
692 static u64 osd_req_encode_op(struct ceph_osd_request *req,
693 struct ceph_osd_op *dst, unsigned int which)
695 struct ceph_osd_req_op *src;
696 struct ceph_osd_data *osd_data;
697 u64 request_data_len = 0;
700 BUG_ON(which >= req->r_num_ops);
701 src = &req->r_ops[which];
702 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
703 pr_err("unrecognized osd opcode %d\n", src->op);
709 case CEPH_OSD_OP_STAT:
710 osd_data = &src->raw_data_in;
711 ceph_osdc_msg_data_add(req->r_reply, osd_data);
713 case CEPH_OSD_OP_READ:
714 case CEPH_OSD_OP_WRITE:
715 case CEPH_OSD_OP_WRITEFULL:
716 case CEPH_OSD_OP_ZERO:
717 case CEPH_OSD_OP_TRUNCATE:
718 if (src->op == CEPH_OSD_OP_WRITE ||
719 src->op == CEPH_OSD_OP_WRITEFULL)
720 request_data_len = src->extent.length;
721 dst->extent.offset = cpu_to_le64(src->extent.offset);
722 dst->extent.length = cpu_to_le64(src->extent.length);
723 dst->extent.truncate_size =
724 cpu_to_le64(src->extent.truncate_size);
725 dst->extent.truncate_seq =
726 cpu_to_le32(src->extent.truncate_seq);
727 osd_data = &src->extent.osd_data;
728 if (src->op == CEPH_OSD_OP_WRITE ||
729 src->op == CEPH_OSD_OP_WRITEFULL)
730 ceph_osdc_msg_data_add(req->r_request, osd_data);
732 ceph_osdc_msg_data_add(req->r_reply, osd_data);
734 case CEPH_OSD_OP_CALL:
735 dst->cls.class_len = src->cls.class_len;
736 dst->cls.method_len = src->cls.method_len;
737 osd_data = &src->cls.request_info;
738 ceph_osdc_msg_data_add(req->r_request, osd_data);
739 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
740 request_data_len = osd_data->pagelist->length;
742 osd_data = &src->cls.request_data;
743 data_length = ceph_osd_data_length(osd_data);
745 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
746 dst->cls.indata_len = cpu_to_le32(data_length);
747 ceph_osdc_msg_data_add(req->r_request, osd_data);
748 src->indata_len += data_length;
749 request_data_len += data_length;
751 osd_data = &src->cls.response_data;
752 ceph_osdc_msg_data_add(req->r_reply, osd_data);
754 case CEPH_OSD_OP_STARTSYNC:
756 case CEPH_OSD_OP_NOTIFY_ACK:
757 case CEPH_OSD_OP_WATCH:
758 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
759 dst->watch.ver = cpu_to_le64(src->watch.ver);
760 dst->watch.flag = src->watch.flag;
762 case CEPH_OSD_OP_SETALLOCHINT:
763 dst->alloc_hint.expected_object_size =
764 cpu_to_le64(src->alloc_hint.expected_object_size);
765 dst->alloc_hint.expected_write_size =
766 cpu_to_le64(src->alloc_hint.expected_write_size);
768 case CEPH_OSD_OP_SETXATTR:
769 case CEPH_OSD_OP_CMPXATTR:
770 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
771 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
772 dst->xattr.cmp_op = src->xattr.cmp_op;
773 dst->xattr.cmp_mode = src->xattr.cmp_mode;
774 osd_data = &src->xattr.osd_data;
775 ceph_osdc_msg_data_add(req->r_request, osd_data);
776 request_data_len = osd_data->pagelist->length;
778 case CEPH_OSD_OP_CREATE:
779 case CEPH_OSD_OP_DELETE:
782 pr_err("unsupported osd opcode %s\n",
783 ceph_osd_op_name(src->op));
789 dst->op = cpu_to_le16(src->op);
790 dst->flags = cpu_to_le32(src->flags);
791 dst->payload_len = cpu_to_le32(src->indata_len);
793 return request_data_len;
797 * build new request AND message, calculate layout, and adjust file
800 * if the file was recently truncated, we include information about its
801 * old and new size so that the object can be updated appropriately. (we
802 * avoid synchronously deleting truncated objects because it's slow.)
804 * if @do_sync, include a 'startsync' command so that the osd will flush
807 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
808 struct ceph_file_layout *layout,
809 struct ceph_vino vino,
811 unsigned int which, int num_ops,
812 int opcode, int flags,
813 struct ceph_snap_context *snapc,
818 struct ceph_osd_request *req;
824 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
825 opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
826 opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
828 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
835 req->r_flags = flags;
837 /* calculate max write size */
838 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
842 if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
843 osd_req_op_init(req, which, opcode, 0);
845 u32 object_size = le32_to_cpu(layout->fl_object_size);
846 u32 object_base = off - objoff;
847 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
848 if (truncate_size <= object_base) {
851 truncate_size -= object_base;
852 if (truncate_size > object_size)
853 truncate_size = object_size;
856 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
857 truncate_size, truncate_seq);
860 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
861 ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum);
863 r = ceph_osdc_alloc_messages(req, GFP_NOFS);
870 ceph_osdc_put_request(req);
873 EXPORT_SYMBOL(ceph_osdc_new_request);
876 * We keep osd requests in an rbtree, sorted by ->r_tid.
878 static void __insert_request(struct ceph_osd_client *osdc,
879 struct ceph_osd_request *new)
881 struct rb_node **p = &osdc->requests.rb_node;
882 struct rb_node *parent = NULL;
883 struct ceph_osd_request *req = NULL;
887 req = rb_entry(parent, struct ceph_osd_request, r_node);
888 if (new->r_tid < req->r_tid)
890 else if (new->r_tid > req->r_tid)
896 rb_link_node(&new->r_node, parent, p);
897 rb_insert_color(&new->r_node, &osdc->requests);
900 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
903 struct ceph_osd_request *req;
904 struct rb_node *n = osdc->requests.rb_node;
907 req = rb_entry(n, struct ceph_osd_request, r_node);
908 if (tid < req->r_tid)
910 else if (tid > req->r_tid)
918 static struct ceph_osd_request *
919 __lookup_request_ge(struct ceph_osd_client *osdc,
922 struct ceph_osd_request *req;
923 struct rb_node *n = osdc->requests.rb_node;
926 req = rb_entry(n, struct ceph_osd_request, r_node);
927 if (tid < req->r_tid) {
931 } else if (tid > req->r_tid) {
940 static void __kick_linger_request(struct ceph_osd_request *req)
942 struct ceph_osd_client *osdc = req->r_osdc;
943 struct ceph_osd *osd = req->r_osd;
946 * Linger requests need to be resent with a new tid to avoid
947 * the dup op detection logic on the OSDs. Achieve this with
948 * a re-register dance instead of open-coding.
950 ceph_osdc_get_request(req);
951 if (!list_empty(&req->r_linger_item))
952 __unregister_linger_request(osdc, req);
954 __unregister_request(osdc, req);
955 __register_request(osdc, req);
956 ceph_osdc_put_request(req);
959 * Unless request has been registered as both normal and
960 * lingering, __unregister{,_linger}_request clears r_osd.
961 * However, here we need to preserve r_osd to make sure we
962 * requeue on the same OSD.
964 WARN_ON(req->r_osd || !osd);
967 dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
968 __enqueue_request(req);
972 * Resubmit requests pending on the given osd.
974 static void __kick_osd_requests(struct ceph_osd_client *osdc,
975 struct ceph_osd *osd)
977 struct ceph_osd_request *req, *nreq;
979 LIST_HEAD(resend_linger);
982 dout("%s osd%d\n", __func__, osd->o_osd);
983 err = __reset_osd(osdc, osd);
988 * Build up a list of requests to resend by traversing the
989 * osd's list of requests. Requests for a given object are
990 * sent in tid order, and that is also the order they're
991 * kept on this list. Therefore all requests that are in
992 * flight will be found first, followed by all requests that
993 * have not yet been sent. And to resend requests while
994 * preserving this order we will want to put any sent
995 * requests back on the front of the osd client's unsent
998 * So we build a separate ordered list of already-sent
999 * requests for the affected osd and splice it onto the
1000 * front of the osd client's unsent list. Once we've seen a
1001 * request that has not yet been sent we're done. Those
1002 * requests are already sitting right where they belong.
1004 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
1008 if (!req->r_linger) {
1009 dout("%s requeueing %p tid %llu\n", __func__, req,
1011 list_move_tail(&req->r_req_lru_item, &resend);
1012 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1014 list_move_tail(&req->r_req_lru_item, &resend_linger);
1017 list_splice(&resend, &osdc->req_unsent);
1020 * Both registered and not yet registered linger requests are
1021 * enqueued with a new tid on the same OSD. We add/move them
1022 * to req_unsent/o_requests at the end to keep things in tid
1025 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
1026 r_linger_osd_item) {
1027 WARN_ON(!list_empty(&req->r_req_lru_item));
1028 __kick_linger_request(req);
1031 list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
1032 __kick_linger_request(req);
1036 * If the osd connection drops, we need to resubmit all requests.
1038 static void osd_reset(struct ceph_connection *con)
1040 struct ceph_osd *osd = con->private;
1041 struct ceph_osd_client *osdc;
1045 dout("osd_reset osd%d\n", osd->o_osd);
1047 down_read(&osdc->map_sem);
1048 mutex_lock(&osdc->request_mutex);
1049 __kick_osd_requests(osdc, osd);
1050 __send_queued(osdc);
1051 mutex_unlock(&osdc->request_mutex);
1052 up_read(&osdc->map_sem);
1056 * Track open sessions with osds.
1058 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1060 struct ceph_osd *osd;
1062 osd = kzalloc(sizeof(*osd), GFP_NOFS);
1066 atomic_set(&osd->o_ref, 1);
1069 RB_CLEAR_NODE(&osd->o_node);
1070 INIT_LIST_HEAD(&osd->o_requests);
1071 INIT_LIST_HEAD(&osd->o_linger_requests);
1072 INIT_LIST_HEAD(&osd->o_osd_lru);
1073 osd->o_incarnation = 1;
1075 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1077 INIT_LIST_HEAD(&osd->o_keepalive_item);
1081 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1083 if (atomic_inc_not_zero(&osd->o_ref)) {
1084 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1085 atomic_read(&osd->o_ref));
1088 dout("get_osd %p FAIL\n", osd);
1093 static void put_osd(struct ceph_osd *osd)
1095 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1096 atomic_read(&osd->o_ref) - 1);
1097 if (atomic_dec_and_test(&osd->o_ref)) {
1098 if (osd->o_auth.authorizer)
1099 ceph_auth_destroy_authorizer(osd->o_auth.authorizer);
1105 * remove an osd from our map
1107 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1109 dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1110 WARN_ON(!list_empty(&osd->o_requests));
1111 WARN_ON(!list_empty(&osd->o_linger_requests));
1113 list_del_init(&osd->o_osd_lru);
1114 rb_erase(&osd->o_node, &osdc->osds);
1115 RB_CLEAR_NODE(&osd->o_node);
1118 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1120 dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1122 if (!RB_EMPTY_NODE(&osd->o_node)) {
1123 ceph_con_close(&osd->o_con);
1124 __remove_osd(osdc, osd);
1129 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1130 struct ceph_osd *osd)
1132 dout("%s %p\n", __func__, osd);
1133 BUG_ON(!list_empty(&osd->o_osd_lru));
1135 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1136 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1139 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1140 struct ceph_osd *osd)
1142 dout("%s %p\n", __func__, osd);
1144 if (list_empty(&osd->o_requests) &&
1145 list_empty(&osd->o_linger_requests))
1146 __move_osd_to_lru(osdc, osd);
1149 static void __remove_osd_from_lru(struct ceph_osd *osd)
1151 dout("__remove_osd_from_lru %p\n", osd);
1152 if (!list_empty(&osd->o_osd_lru))
1153 list_del_init(&osd->o_osd_lru);
1159 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1161 struct ceph_entity_addr *peer_addr;
1163 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1164 if (list_empty(&osd->o_requests) &&
1165 list_empty(&osd->o_linger_requests)) {
1166 remove_osd(osdc, osd);
1170 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1171 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1172 !ceph_con_opened(&osd->o_con)) {
1173 struct ceph_osd_request *req;
1175 dout("osd addr hasn't changed and connection never opened, "
1176 "letting msgr retry\n");
1177 /* touch each r_stamp for handle_timeout()'s benfit */
1178 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1179 req->r_stamp = jiffies;
1184 ceph_con_close(&osd->o_con);
1185 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1186 osd->o_incarnation++;
1191 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1193 struct rb_node **p = &osdc->osds.rb_node;
1194 struct rb_node *parent = NULL;
1195 struct ceph_osd *osd = NULL;
1197 dout("__insert_osd %p osd%d\n", new, new->o_osd);
1200 osd = rb_entry(parent, struct ceph_osd, o_node);
1201 if (new->o_osd < osd->o_osd)
1203 else if (new->o_osd > osd->o_osd)
1204 p = &(*p)->rb_right;
1209 rb_link_node(&new->o_node, parent, p);
1210 rb_insert_color(&new->o_node, &osdc->osds);
1213 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1215 struct ceph_osd *osd;
1216 struct rb_node *n = osdc->osds.rb_node;
1219 osd = rb_entry(n, struct ceph_osd, o_node);
1222 else if (o > osd->o_osd)
1230 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1232 schedule_delayed_work(&osdc->timeout_work,
1233 osdc->client->options->osd_keepalive_timeout);
1236 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1238 cancel_delayed_work(&osdc->timeout_work);
1242 * Register request, assign tid. If this is the first request, set up
1243 * the timeout event.
1245 static void __register_request(struct ceph_osd_client *osdc,
1246 struct ceph_osd_request *req)
1248 req->r_tid = ++osdc->last_tid;
1249 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1250 dout("__register_request %p tid %lld\n", req, req->r_tid);
1251 __insert_request(osdc, req);
1252 ceph_osdc_get_request(req);
1253 osdc->num_requests++;
1254 if (osdc->num_requests == 1) {
1255 dout(" first request, scheduling timeout\n");
1256 __schedule_osd_timeout(osdc);
1261 * called under osdc->request_mutex
1263 static void __unregister_request(struct ceph_osd_client *osdc,
1264 struct ceph_osd_request *req)
1266 if (RB_EMPTY_NODE(&req->r_node)) {
1267 dout("__unregister_request %p tid %lld not registered\n",
1272 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1273 rb_erase(&req->r_node, &osdc->requests);
1274 RB_CLEAR_NODE(&req->r_node);
1275 osdc->num_requests--;
1278 /* make sure the original request isn't in flight. */
1279 ceph_msg_revoke(req->r_request);
1281 list_del_init(&req->r_osd_item);
1282 maybe_move_osd_to_lru(osdc, req->r_osd);
1283 if (list_empty(&req->r_linger_osd_item))
1287 list_del_init(&req->r_req_lru_item);
1288 ceph_osdc_put_request(req);
1290 if (osdc->num_requests == 0) {
1291 dout(" no requests, canceling timeout\n");
1292 __cancel_osd_timeout(osdc);
1297 * Cancel a previously queued request message
1299 static void __cancel_request(struct ceph_osd_request *req)
1301 if (req->r_sent && req->r_osd) {
1302 ceph_msg_revoke(req->r_request);
1307 static void __register_linger_request(struct ceph_osd_client *osdc,
1308 struct ceph_osd_request *req)
1310 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1311 WARN_ON(!req->r_linger);
1313 ceph_osdc_get_request(req);
1314 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1316 list_add_tail(&req->r_linger_osd_item,
1317 &req->r_osd->o_linger_requests);
1320 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1321 struct ceph_osd_request *req)
1323 WARN_ON(!req->r_linger);
1325 if (list_empty(&req->r_linger_item)) {
1326 dout("%s %p tid %llu not registered\n", __func__, req,
1331 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1332 list_del_init(&req->r_linger_item);
1335 list_del_init(&req->r_linger_osd_item);
1336 maybe_move_osd_to_lru(osdc, req->r_osd);
1337 if (list_empty(&req->r_osd_item))
1340 ceph_osdc_put_request(req);
1343 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1344 struct ceph_osd_request *req)
1346 if (!req->r_linger) {
1347 dout("set_request_linger %p\n", req);
1351 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1354 * Returns whether a request should be blocked from being sent
1355 * based on the current osdmap and osd_client settings.
1357 * Caller should hold map_sem for read.
1359 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1360 struct ceph_osd_request *req)
1362 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1363 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1364 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1365 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1366 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1370 * Calculate mapping of a request to a PG. Takes tiering into account.
1372 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1373 struct ceph_osd_request *req,
1374 struct ceph_pg *pg_out)
1376 bool need_check_tiering;
1378 need_check_tiering = false;
1379 if (req->r_target_oloc.pool == -1) {
1380 req->r_target_oloc = req->r_base_oloc; /* struct */
1381 need_check_tiering = true;
1383 if (ceph_oid_empty(&req->r_target_oid)) {
1384 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1385 need_check_tiering = true;
1388 if (need_check_tiering &&
1389 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1390 struct ceph_pg_pool_info *pi;
1392 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1394 if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1396 req->r_target_oloc.pool = pi->read_tier;
1397 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1398 pi->write_tier >= 0)
1399 req->r_target_oloc.pool = pi->write_tier;
1401 /* !pi is caught in ceph_oloc_oid_to_pg() */
1404 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1405 &req->r_target_oid, pg_out);
1408 static void __enqueue_request(struct ceph_osd_request *req)
1410 struct ceph_osd_client *osdc = req->r_osdc;
1412 dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1413 req->r_osd ? req->r_osd->o_osd : -1);
1416 __remove_osd_from_lru(req->r_osd);
1417 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1418 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1420 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1425 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1426 * (as needed), and set the request r_osd appropriately. If there is
1427 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1428 * (unsent, homeless) or leave on in-flight lru.
1430 * Return 0 if unchanged, 1 if changed, or negative on error.
1432 * Caller should hold map_sem for read and request_mutex.
1434 static int __map_request(struct ceph_osd_client *osdc,
1435 struct ceph_osd_request *req, int force_resend)
1437 struct ceph_pg pgid;
1438 int acting[CEPH_PG_MAX_SIZE];
1443 dout("map_request %p tid %lld\n", req, req->r_tid);
1445 err = __calc_request_pg(osdc->osdmap, req, &pgid);
1447 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1452 num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1456 was_paused = req->r_paused;
1457 req->r_paused = __req_should_be_paused(osdc, req);
1458 if (was_paused && !req->r_paused)
1461 if ((!force_resend &&
1462 req->r_osd && req->r_osd->o_osd == o &&
1463 req->r_sent >= req->r_osd->o_incarnation &&
1464 req->r_num_pg_osds == num &&
1465 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1466 (req->r_osd == NULL && o == -1) ||
1468 return 0; /* no change */
1470 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1471 req->r_tid, pgid.pool, pgid.seed, o,
1472 req->r_osd ? req->r_osd->o_osd : -1);
1474 /* record full pg acting set */
1475 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1476 req->r_num_pg_osds = num;
1479 __cancel_request(req);
1480 list_del_init(&req->r_osd_item);
1481 list_del_init(&req->r_linger_osd_item);
1485 req->r_osd = __lookup_osd(osdc, o);
1486 if (!req->r_osd && o >= 0) {
1488 req->r_osd = create_osd(osdc, o);
1490 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1494 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1495 __insert_osd(osdc, req->r_osd);
1497 ceph_con_open(&req->r_osd->o_con,
1498 CEPH_ENTITY_TYPE_OSD, o,
1499 &osdc->osdmap->osd_addr[o]);
1502 __enqueue_request(req);
1503 err = 1; /* osd or pg changed */
1510 * caller should hold map_sem (for read) and request_mutex
1512 static void __send_request(struct ceph_osd_client *osdc,
1513 struct ceph_osd_request *req)
1517 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1518 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1519 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1521 /* fill in message content that changes each time we send it */
1522 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1523 put_unaligned_le32(req->r_flags, req->r_request_flags);
1524 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1525 p = req->r_request_pgid;
1526 ceph_encode_64(&p, req->r_pgid.pool);
1527 ceph_encode_32(&p, req->r_pgid.seed);
1528 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1529 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1530 sizeof(req->r_reassert_version));
1532 req->r_stamp = jiffies;
1533 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1535 ceph_msg_get(req->r_request); /* send consumes a ref */
1537 req->r_sent = req->r_osd->o_incarnation;
1539 ceph_con_send(&req->r_osd->o_con, req->r_request);
1543 * Send any requests in the queue (req_unsent).
1545 static void __send_queued(struct ceph_osd_client *osdc)
1547 struct ceph_osd_request *req, *tmp;
1549 dout("__send_queued\n");
1550 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1551 __send_request(osdc, req);
1555 * Caller should hold map_sem for read and request_mutex.
1557 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1558 struct ceph_osd_request *req,
1563 __register_request(osdc, req);
1565 req->r_got_reply = 0;
1566 rc = __map_request(osdc, req, 0);
1569 dout("osdc_start_request failed map, "
1570 " will retry %lld\n", req->r_tid);
1573 __unregister_request(osdc, req);
1578 if (req->r_osd == NULL) {
1579 dout("send_request %p no up osds in pg\n", req);
1580 ceph_monc_request_next_osdmap(&osdc->client->monc);
1582 __send_queued(osdc);
1589 * Timeout callback, called every N seconds when 1 or more osd
1590 * requests has been active for more than N seconds. When this
1591 * happens, we ping all OSDs with requests who have timed out to
1592 * ensure any communications channel reset is detected. Reset the
1593 * request timeouts another N seconds in the future as we go.
1594 * Reschedule the timeout event another N seconds in future (unless
1595 * there are no open requests).
1597 static void handle_timeout(struct work_struct *work)
1599 struct ceph_osd_client *osdc =
1600 container_of(work, struct ceph_osd_client, timeout_work.work);
1601 struct ceph_options *opts = osdc->client->options;
1602 struct ceph_osd_request *req;
1603 struct ceph_osd *osd;
1604 struct list_head slow_osds;
1606 down_read(&osdc->map_sem);
1608 ceph_monc_request_next_osdmap(&osdc->client->monc);
1610 mutex_lock(&osdc->request_mutex);
1613 * ping osds that are a bit slow. this ensures that if there
1614 * is a break in the TCP connection we will notice, and reopen
1615 * a connection with that osd (from the fault callback).
1617 INIT_LIST_HEAD(&slow_osds);
1618 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1619 if (time_before(jiffies,
1620 req->r_stamp + opts->osd_keepalive_timeout))
1625 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1626 req->r_tid, osd->o_osd);
1627 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1629 while (!list_empty(&slow_osds)) {
1630 osd = list_entry(slow_osds.next, struct ceph_osd,
1632 list_del_init(&osd->o_keepalive_item);
1633 ceph_con_keepalive(&osd->o_con);
1636 __schedule_osd_timeout(osdc);
1637 __send_queued(osdc);
1638 mutex_unlock(&osdc->request_mutex);
1639 up_read(&osdc->map_sem);
1642 static void handle_osds_timeout(struct work_struct *work)
1644 struct ceph_osd_client *osdc =
1645 container_of(work, struct ceph_osd_client,
1646 osds_timeout_work.work);
1647 unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
1648 struct ceph_osd *osd, *nosd;
1650 dout("%s osdc %p\n", __func__, osdc);
1651 down_read(&osdc->map_sem);
1652 mutex_lock(&osdc->request_mutex);
1654 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1655 if (time_before(jiffies, osd->lru_ttl))
1658 remove_osd(osdc, osd);
1661 mutex_unlock(&osdc->request_mutex);
1662 up_read(&osdc->map_sem);
1663 schedule_delayed_work(&osdc->osds_timeout_work,
1664 round_jiffies_relative(delay));
1667 static int ceph_oloc_decode(void **p, void *end,
1668 struct ceph_object_locator *oloc)
1670 u8 struct_v, struct_cv;
1675 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1676 struct_v = ceph_decode_8(p);
1677 struct_cv = ceph_decode_8(p);
1679 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1680 struct_v, struct_cv);
1683 if (struct_cv > 6) {
1684 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1685 struct_v, struct_cv);
1688 len = ceph_decode_32(p);
1689 ceph_decode_need(p, end, len, e_inval);
1690 struct_end = *p + len;
1692 oloc->pool = ceph_decode_64(p);
1693 *p += 4; /* skip preferred */
1695 len = ceph_decode_32(p);
1697 pr_warn("ceph_object_locator::key is set\n");
1701 if (struct_v >= 5) {
1702 len = ceph_decode_32(p);
1704 pr_warn("ceph_object_locator::nspace is set\n");
1709 if (struct_v >= 6) {
1710 s64 hash = ceph_decode_64(p);
1712 pr_warn("ceph_object_locator::hash is set\n");
1727 static int ceph_redirect_decode(void **p, void *end,
1728 struct ceph_request_redirect *redir)
1730 u8 struct_v, struct_cv;
1735 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1736 struct_v = ceph_decode_8(p);
1737 struct_cv = ceph_decode_8(p);
1738 if (struct_cv > 1) {
1739 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1740 struct_v, struct_cv);
1743 len = ceph_decode_32(p);
1744 ceph_decode_need(p, end, len, e_inval);
1745 struct_end = *p + len;
1747 ret = ceph_oloc_decode(p, end, &redir->oloc);
1751 len = ceph_decode_32(p);
1753 pr_warn("ceph_request_redirect::object_name is set\n");
1757 len = ceph_decode_32(p);
1758 *p += len; /* skip osd_instructions */
1770 static void complete_request(struct ceph_osd_request *req)
1772 complete_all(&req->r_safe_completion); /* fsync waiter */
1776 * handle osd op reply. either call the callback if it is specified,
1777 * or do the completion to wake up the waiting thread.
1779 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1782 struct ceph_osd_request *req;
1783 struct ceph_request_redirect redir;
1786 unsigned int numops;
1787 int payload_len, flags;
1793 u64 reassert_version;
1795 int already_completed;
1800 tid = le64_to_cpu(msg->hdr.tid);
1801 dout("handle_reply %p tid %llu\n", msg, tid);
1803 p = msg->front.iov_base;
1804 end = p + msg->front.iov_len;
1806 ceph_decode_need(&p, end, 4, bad);
1807 object_len = ceph_decode_32(&p);
1808 ceph_decode_need(&p, end, object_len, bad);
1811 err = ceph_decode_pgid(&p, end, &pg);
1815 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1816 flags = ceph_decode_64(&p);
1817 result = ceph_decode_32(&p);
1818 reassert_epoch = ceph_decode_32(&p);
1819 reassert_version = ceph_decode_64(&p);
1820 osdmap_epoch = ceph_decode_32(&p);
1823 down_read(&osdc->map_sem);
1824 mutex_lock(&osdc->request_mutex);
1825 req = __lookup_request(osdc, tid);
1827 dout("handle_reply tid %llu dne\n", tid);
1830 ceph_osdc_get_request(req);
1832 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1835 ceph_decode_need(&p, end, 4, bad_put);
1836 numops = ceph_decode_32(&p);
1837 if (numops > CEPH_OSD_MAX_OPS)
1839 if (numops != req->r_num_ops)
1842 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1843 for (i = 0; i < numops; i++) {
1844 struct ceph_osd_op *op = p;
1847 len = le32_to_cpu(op->payload_len);
1848 req->r_ops[i].outdata_len = len;
1849 dout(" op %d has %d bytes\n", i, len);
1853 bytes = le32_to_cpu(msg->hdr.data_len);
1854 if (payload_len != bytes) {
1855 pr_warn("sum of op payload lens %d != data_len %d\n",
1856 payload_len, bytes);
1860 ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1861 retry_attempt = ceph_decode_32(&p);
1862 for (i = 0; i < numops; i++)
1863 req->r_ops[i].rval = ceph_decode_32(&p);
1865 if (le16_to_cpu(msg->hdr.version) >= 6) {
1866 p += 8 + 4; /* skip replay_version */
1867 p += 8; /* skip user_version */
1869 if (le16_to_cpu(msg->hdr.version) >= 7)
1870 ceph_decode_8_safe(&p, end, decode_redir, bad_put);
1878 err = ceph_redirect_decode(&p, end, &redir);
1882 redir.oloc.pool = -1;
1885 if (redir.oloc.pool != -1) {
1886 dout("redirect pool %lld\n", redir.oloc.pool);
1888 __unregister_request(osdc, req);
1890 req->r_target_oloc = redir.oloc; /* struct */
1893 * Start redirect requests with nofail=true. If
1894 * mapping fails, request will end up on the notarget
1895 * list, waiting for the new osdmap (which can take
1896 * a while), even though the original request mapped
1897 * successfully. In the future we might want to follow
1898 * original request's nofail setting here.
1900 err = __ceph_osdc_start_request(osdc, req, true);
1906 already_completed = req->r_got_reply;
1907 if (!req->r_got_reply) {
1908 req->r_result = result;
1909 dout("handle_reply result %d bytes %d\n", req->r_result,
1911 if (req->r_result == 0)
1912 req->r_result = bytes;
1914 /* in case this is a write and we need to replay, */
1915 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1916 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1918 req->r_got_reply = 1;
1919 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1920 dout("handle_reply tid %llu dup ack\n", tid);
1924 dout("handle_reply tid %llu flags %d\n", tid, flags);
1926 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1927 __register_linger_request(osdc, req);
1929 /* either this is a read, or we got the safe response */
1931 (flags & CEPH_OSD_FLAG_ONDISK) ||
1932 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1933 __unregister_request(osdc, req);
1935 mutex_unlock(&osdc->request_mutex);
1936 up_read(&osdc->map_sem);
1938 if (!already_completed) {
1939 if (req->r_unsafe_callback &&
1940 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1941 req->r_unsafe_callback(req, true);
1942 if (req->r_callback)
1943 req->r_callback(req, msg);
1945 complete_all(&req->r_completion);
1948 if (flags & CEPH_OSD_FLAG_ONDISK) {
1949 if (req->r_unsafe_callback && already_completed)
1950 req->r_unsafe_callback(req, false);
1951 complete_request(req);
1955 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1956 ceph_osdc_put_request(req);
1959 mutex_unlock(&osdc->request_mutex);
1960 up_read(&osdc->map_sem);
1964 req->r_result = -EIO;
1965 __unregister_request(osdc, req);
1966 if (req->r_callback)
1967 req->r_callback(req, msg);
1969 complete_all(&req->r_completion);
1970 complete_request(req);
1971 ceph_osdc_put_request(req);
1973 mutex_unlock(&osdc->request_mutex);
1974 up_read(&osdc->map_sem);
1976 pr_err("corrupt osd_op_reply got %d %d\n",
1977 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1981 static void reset_changed_osds(struct ceph_osd_client *osdc)
1983 struct rb_node *p, *n;
1985 dout("%s %p\n", __func__, osdc);
1986 for (p = rb_first(&osdc->osds); p; p = n) {
1987 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1990 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1991 memcmp(&osd->o_con.peer_addr,
1992 ceph_osd_addr(osdc->osdmap,
1994 sizeof(struct ceph_entity_addr)) != 0)
1995 __reset_osd(osdc, osd);
2000 * Requeue requests whose mapping to an OSD has changed. If requests map to
2001 * no osd, request a new map.
2003 * Caller should hold map_sem for read.
2005 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
2006 bool force_resend_writes)
2008 struct ceph_osd_request *req, *nreq;
2012 bool force_resend_req;
2014 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
2015 force_resend_writes ? " (force resend writes)" : "");
2016 mutex_lock(&osdc->request_mutex);
2017 for (p = rb_first(&osdc->requests); p; ) {
2018 req = rb_entry(p, struct ceph_osd_request, r_node);
2022 * For linger requests that have not yet been
2023 * registered, move them to the linger list; they'll
2024 * be sent to the osd in the loop below. Unregister
2025 * the request before re-registering it as a linger
2026 * request to ensure the __map_request() below
2027 * will decide it needs to be sent.
2029 if (req->r_linger && list_empty(&req->r_linger_item)) {
2030 dout("%p tid %llu restart on osd%d\n",
2032 req->r_osd ? req->r_osd->o_osd : -1);
2033 ceph_osdc_get_request(req);
2034 __unregister_request(osdc, req);
2035 __register_linger_request(osdc, req);
2036 ceph_osdc_put_request(req);
2040 force_resend_req = force_resend ||
2041 (force_resend_writes &&
2042 req->r_flags & CEPH_OSD_FLAG_WRITE);
2043 err = __map_request(osdc, req, force_resend_req);
2045 continue; /* error */
2046 if (req->r_osd == NULL) {
2047 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2048 needmap++; /* request a newer map */
2049 } else if (err > 0) {
2050 if (!req->r_linger) {
2051 dout("%p tid %llu requeued on osd%d\n", req,
2053 req->r_osd ? req->r_osd->o_osd : -1);
2054 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2059 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2061 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2063 err = __map_request(osdc, req,
2064 force_resend || force_resend_writes);
2065 dout("__map_request returned %d\n", err);
2067 continue; /* hrm! */
2068 if (req->r_osd == NULL || err > 0) {
2069 if (req->r_osd == NULL) {
2070 dout("lingering %p tid %llu maps to no osd\n",
2073 * A homeless lingering request makes
2074 * no sense, as it's job is to keep
2075 * a particular OSD connection open.
2076 * Request a newer map and kick the
2077 * request, knowing that it won't be
2078 * resent until we actually get a map
2079 * that can tell us where to send it.
2084 dout("kicking lingering %p tid %llu osd%d\n", req,
2085 req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2086 __register_request(osdc, req);
2087 __unregister_linger_request(osdc, req);
2090 reset_changed_osds(osdc);
2091 mutex_unlock(&osdc->request_mutex);
2094 dout("%d requests for down osds, need new map\n", needmap);
2095 ceph_monc_request_next_osdmap(&osdc->client->monc);
2101 * Process updated osd map.
2103 * The message contains any number of incremental and full maps, normally
2104 * indicating some sort of topology change in the cluster. Kick requests
2105 * off to different OSDs as needed.
2107 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2109 void *p, *end, *next;
2110 u32 nr_maps, maplen;
2112 struct ceph_osdmap *newmap = NULL, *oldmap;
2114 struct ceph_fsid fsid;
2117 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2118 p = msg->front.iov_base;
2119 end = p + msg->front.iov_len;
2122 ceph_decode_need(&p, end, sizeof(fsid), bad);
2123 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2124 if (ceph_check_fsid(osdc->client, &fsid) < 0)
2127 down_write(&osdc->map_sem);
2129 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2131 /* incremental maps */
2132 ceph_decode_32_safe(&p, end, nr_maps, bad);
2133 dout(" %d inc maps\n", nr_maps);
2134 while (nr_maps > 0) {
2135 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2136 epoch = ceph_decode_32(&p);
2137 maplen = ceph_decode_32(&p);
2138 ceph_decode_need(&p, end, maplen, bad);
2140 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2141 dout("applying incremental map %u len %d\n",
2143 newmap = osdmap_apply_incremental(&p, next,
2145 if (IS_ERR(newmap)) {
2146 err = PTR_ERR(newmap);
2150 if (newmap != osdc->osdmap) {
2151 ceph_osdmap_destroy(osdc->osdmap);
2152 osdc->osdmap = newmap;
2154 was_full = was_full ||
2155 ceph_osdmap_flag(osdc->osdmap,
2157 kick_requests(osdc, 0, was_full);
2159 dout("ignoring incremental map %u len %d\n",
2169 ceph_decode_32_safe(&p, end, nr_maps, bad);
2170 dout(" %d full maps\n", nr_maps);
2172 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2173 epoch = ceph_decode_32(&p);
2174 maplen = ceph_decode_32(&p);
2175 ceph_decode_need(&p, end, maplen, bad);
2177 dout("skipping non-latest full map %u len %d\n",
2179 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2180 dout("skipping full map %u len %d, "
2181 "older than our %u\n", epoch, maplen,
2182 osdc->osdmap->epoch);
2184 int skipped_map = 0;
2186 dout("taking full map %u len %d\n", epoch, maplen);
2187 newmap = ceph_osdmap_decode(&p, p+maplen);
2188 if (IS_ERR(newmap)) {
2189 err = PTR_ERR(newmap);
2193 oldmap = osdc->osdmap;
2194 osdc->osdmap = newmap;
2196 if (oldmap->epoch + 1 < newmap->epoch)
2198 ceph_osdmap_destroy(oldmap);
2200 was_full = was_full ||
2201 ceph_osdmap_flag(osdc->osdmap,
2203 kick_requests(osdc, skipped_map, was_full);
2212 downgrade_write(&osdc->map_sem);
2213 ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
2214 osdc->osdmap->epoch);
2217 * subscribe to subsequent osdmap updates if full to ensure
2218 * we find out when we are no longer full and stop returning
2221 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2222 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2223 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2224 ceph_monc_request_next_osdmap(&osdc->client->monc);
2226 mutex_lock(&osdc->request_mutex);
2227 __send_queued(osdc);
2228 mutex_unlock(&osdc->request_mutex);
2229 up_read(&osdc->map_sem);
2230 wake_up_all(&osdc->client->auth_wq);
2234 pr_err("osdc handle_map corrupt msg\n");
2236 up_write(&osdc->map_sem);
2240 * watch/notify callback event infrastructure
2242 * These callbacks are used both for watch and notify operations.
2244 static void __release_event(struct kref *kref)
2246 struct ceph_osd_event *event =
2247 container_of(kref, struct ceph_osd_event, kref);
2249 dout("__release_event %p\n", event);
2253 static void get_event(struct ceph_osd_event *event)
2255 kref_get(&event->kref);
2258 void ceph_osdc_put_event(struct ceph_osd_event *event)
2260 kref_put(&event->kref, __release_event);
2262 EXPORT_SYMBOL(ceph_osdc_put_event);
2264 static void __insert_event(struct ceph_osd_client *osdc,
2265 struct ceph_osd_event *new)
2267 struct rb_node **p = &osdc->event_tree.rb_node;
2268 struct rb_node *parent = NULL;
2269 struct ceph_osd_event *event = NULL;
2273 event = rb_entry(parent, struct ceph_osd_event, node);
2274 if (new->cookie < event->cookie)
2276 else if (new->cookie > event->cookie)
2277 p = &(*p)->rb_right;
2282 rb_link_node(&new->node, parent, p);
2283 rb_insert_color(&new->node, &osdc->event_tree);
2286 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2289 struct rb_node **p = &osdc->event_tree.rb_node;
2290 struct rb_node *parent = NULL;
2291 struct ceph_osd_event *event = NULL;
2295 event = rb_entry(parent, struct ceph_osd_event, node);
2296 if (cookie < event->cookie)
2298 else if (cookie > event->cookie)
2299 p = &(*p)->rb_right;
2306 static void __remove_event(struct ceph_osd_event *event)
2308 struct ceph_osd_client *osdc = event->osdc;
2310 if (!RB_EMPTY_NODE(&event->node)) {
2311 dout("__remove_event removed %p\n", event);
2312 rb_erase(&event->node, &osdc->event_tree);
2313 ceph_osdc_put_event(event);
2315 dout("__remove_event didn't remove %p\n", event);
2319 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2320 void (*event_cb)(u64, u64, u8, void *),
2321 void *data, struct ceph_osd_event **pevent)
2323 struct ceph_osd_event *event;
2325 event = kmalloc(sizeof(*event), GFP_NOIO);
2329 dout("create_event %p\n", event);
2330 event->cb = event_cb;
2331 event->one_shot = 0;
2334 INIT_LIST_HEAD(&event->osd_node);
2335 RB_CLEAR_NODE(&event->node);
2336 kref_init(&event->kref); /* one ref for us */
2337 kref_get(&event->kref); /* one ref for the caller */
2339 spin_lock(&osdc->event_lock);
2340 event->cookie = ++osdc->event_count;
2341 __insert_event(osdc, event);
2342 spin_unlock(&osdc->event_lock);
2347 EXPORT_SYMBOL(ceph_osdc_create_event);
2349 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2351 struct ceph_osd_client *osdc = event->osdc;
2353 dout("cancel_event %p\n", event);
2354 spin_lock(&osdc->event_lock);
2355 __remove_event(event);
2356 spin_unlock(&osdc->event_lock);
2357 ceph_osdc_put_event(event); /* caller's */
2359 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2362 static void do_event_work(struct work_struct *work)
2364 struct ceph_osd_event_work *event_work =
2365 container_of(work, struct ceph_osd_event_work, work);
2366 struct ceph_osd_event *event = event_work->event;
2367 u64 ver = event_work->ver;
2368 u64 notify_id = event_work->notify_id;
2369 u8 opcode = event_work->opcode;
2371 dout("do_event_work completing %p\n", event);
2372 event->cb(ver, notify_id, opcode, event->data);
2373 dout("do_event_work completed %p\n", event);
2374 ceph_osdc_put_event(event);
2380 * Process osd watch notifications
2382 static void handle_watch_notify(struct ceph_osd_client *osdc,
2383 struct ceph_msg *msg)
2387 u64 cookie, ver, notify_id;
2389 struct ceph_osd_event *event;
2390 struct ceph_osd_event_work *event_work;
2392 p = msg->front.iov_base;
2393 end = p + msg->front.iov_len;
2395 ceph_decode_8_safe(&p, end, proto_ver, bad);
2396 ceph_decode_8_safe(&p, end, opcode, bad);
2397 ceph_decode_64_safe(&p, end, cookie, bad);
2398 ceph_decode_64_safe(&p, end, ver, bad);
2399 ceph_decode_64_safe(&p, end, notify_id, bad);
2401 spin_lock(&osdc->event_lock);
2402 event = __find_event(osdc, cookie);
2404 BUG_ON(event->one_shot);
2407 spin_unlock(&osdc->event_lock);
2408 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2409 cookie, ver, event);
2411 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2413 pr_err("couldn't allocate event_work\n");
2414 ceph_osdc_put_event(event);
2417 INIT_WORK(&event_work->work, do_event_work);
2418 event_work->event = event;
2419 event_work->ver = ver;
2420 event_work->notify_id = notify_id;
2421 event_work->opcode = opcode;
2423 queue_work(osdc->notify_wq, &event_work->work);
2429 pr_err("osdc handle_watch_notify corrupt msg\n");
2433 * build new request AND message
2436 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2437 struct ceph_snap_context *snapc, u64 snap_id,
2438 struct timespec *mtime)
2440 struct ceph_msg *msg = req->r_request;
2443 int flags = req->r_flags;
2447 req->r_snapid = snap_id;
2448 WARN_ON(snapc != req->r_snapc);
2450 /* encode request */
2451 msg->hdr.version = cpu_to_le16(4);
2453 p = msg->front.iov_base;
2454 ceph_encode_32(&p, 1); /* client_inc is always 1 */
2455 req->r_request_osdmap_epoch = p;
2457 req->r_request_flags = p;
2459 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2460 ceph_encode_timespec(p, mtime);
2461 p += sizeof(struct ceph_timespec);
2462 req->r_request_reassert_version = p;
2463 p += sizeof(struct ceph_eversion); /* will get filled in */
2466 ceph_encode_8(&p, 4);
2467 ceph_encode_8(&p, 4);
2468 ceph_encode_32(&p, 8 + 4 + 4);
2469 req->r_request_pool = p;
2471 ceph_encode_32(&p, -1); /* preferred */
2472 ceph_encode_32(&p, 0); /* key len */
2474 ceph_encode_8(&p, 1);
2475 req->r_request_pgid = p;
2477 ceph_encode_32(&p, -1); /* preferred */
2480 ceph_encode_32(&p, req->r_base_oid.name_len);
2481 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2482 dout("oid %*pE len %d\n", req->r_base_oid.name_len,
2483 req->r_base_oid.name, req->r_base_oid.name_len);
2484 p += req->r_base_oid.name_len;
2486 /* ops--can imply data */
2487 ceph_encode_16(&p, (u16)req->r_num_ops);
2489 for (i = 0; i < req->r_num_ops; i++) {
2490 data_len += osd_req_encode_op(req, p, i);
2491 p += sizeof(struct ceph_osd_op);
2495 ceph_encode_64(&p, req->r_snapid);
2496 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2497 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2499 for (i = 0; i < req->r_snapc->num_snaps; i++) {
2500 ceph_encode_64(&p, req->r_snapc->snaps[i]);
2504 req->r_request_attempts = p;
2508 if (flags & CEPH_OSD_FLAG_WRITE) {
2512 * The header "data_off" is a hint to the receiver
2513 * allowing it to align received data into its
2514 * buffers such that there's no need to re-copy
2515 * it before writing it to disk (direct I/O).
2517 data_off = (u16) (off & 0xffff);
2518 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2520 req->r_request->hdr.data_len = cpu_to_le32(data_len);
2522 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2523 msg_size = p - msg->front.iov_base;
2524 msg->front.iov_len = msg_size;
2525 msg->hdr.front_len = cpu_to_le32(msg_size);
2527 dout("build_request msg_size was %d\n", (int)msg_size);
2529 EXPORT_SYMBOL(ceph_osdc_build_request);
2532 * Register request, send initial attempt.
2534 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2535 struct ceph_osd_request *req,
2540 down_read(&osdc->map_sem);
2541 mutex_lock(&osdc->request_mutex);
2543 rc = __ceph_osdc_start_request(osdc, req, nofail);
2545 mutex_unlock(&osdc->request_mutex);
2546 up_read(&osdc->map_sem);
2550 EXPORT_SYMBOL(ceph_osdc_start_request);
2553 * Unregister a registered request. The request is not completed (i.e.
2554 * no callbacks or wakeups) - higher layers are supposed to know what
2555 * they are canceling.
2557 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2559 struct ceph_osd_client *osdc = req->r_osdc;
2561 mutex_lock(&osdc->request_mutex);
2563 __unregister_linger_request(osdc, req);
2564 __unregister_request(osdc, req);
2565 mutex_unlock(&osdc->request_mutex);
2567 dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2569 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2572 * wait for a request to complete
2574 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2575 struct ceph_osd_request *req)
2579 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2581 rc = wait_for_completion_interruptible(&req->r_completion);
2583 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2584 ceph_osdc_cancel_request(req);
2585 complete_request(req);
2589 dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2591 return req->r_result;
2593 EXPORT_SYMBOL(ceph_osdc_wait_request);
2596 * sync - wait for all in-flight requests to flush. avoid starvation.
2598 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2600 struct ceph_osd_request *req;
2601 u64 last_tid, next_tid = 0;
2603 mutex_lock(&osdc->request_mutex);
2604 last_tid = osdc->last_tid;
2606 req = __lookup_request_ge(osdc, next_tid);
2609 if (req->r_tid > last_tid)
2612 next_tid = req->r_tid + 1;
2613 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2616 ceph_osdc_get_request(req);
2617 mutex_unlock(&osdc->request_mutex);
2618 dout("sync waiting on tid %llu (last is %llu)\n",
2619 req->r_tid, last_tid);
2620 wait_for_completion(&req->r_safe_completion);
2621 mutex_lock(&osdc->request_mutex);
2622 ceph_osdc_put_request(req);
2624 mutex_unlock(&osdc->request_mutex);
2625 dout("sync done (thru tid %llu)\n", last_tid);
2627 EXPORT_SYMBOL(ceph_osdc_sync);
2630 * Call all pending notify callbacks - for use after a watch is
2631 * unregistered, to make sure no more callbacks for it will be invoked
2633 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2635 flush_workqueue(osdc->notify_wq);
2637 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2643 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2648 osdc->client = client;
2649 osdc->osdmap = NULL;
2650 init_rwsem(&osdc->map_sem);
2651 mutex_init(&osdc->request_mutex);
2653 osdc->osds = RB_ROOT;
2654 INIT_LIST_HEAD(&osdc->osd_lru);
2655 osdc->requests = RB_ROOT;
2656 INIT_LIST_HEAD(&osdc->req_lru);
2657 INIT_LIST_HEAD(&osdc->req_unsent);
2658 INIT_LIST_HEAD(&osdc->req_notarget);
2659 INIT_LIST_HEAD(&osdc->req_linger);
2660 osdc->num_requests = 0;
2661 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2662 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2663 spin_lock_init(&osdc->event_lock);
2664 osdc->event_tree = RB_ROOT;
2665 osdc->event_count = 0;
2667 schedule_delayed_work(&osdc->osds_timeout_work,
2668 round_jiffies_relative(osdc->client->options->osd_idle_ttl));
2671 osdc->req_mempool = mempool_create_slab_pool(10,
2672 ceph_osd_request_cache);
2673 if (!osdc->req_mempool)
2676 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2677 PAGE_SIZE, 10, true, "osd_op");
2680 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2681 PAGE_SIZE, 10, true, "osd_op_reply");
2686 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2687 if (!osdc->notify_wq)
2688 goto out_msgpool_reply;
2693 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2695 ceph_msgpool_destroy(&osdc->msgpool_op);
2697 mempool_destroy(osdc->req_mempool);
2702 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2704 flush_workqueue(osdc->notify_wq);
2705 destroy_workqueue(osdc->notify_wq);
2706 cancel_delayed_work_sync(&osdc->timeout_work);
2707 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2709 mutex_lock(&osdc->request_mutex);
2710 while (!RB_EMPTY_ROOT(&osdc->osds)) {
2711 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
2712 struct ceph_osd, o_node);
2713 remove_osd(osdc, osd);
2715 mutex_unlock(&osdc->request_mutex);
2718 ceph_osdmap_destroy(osdc->osdmap);
2719 osdc->osdmap = NULL;
2721 mempool_destroy(osdc->req_mempool);
2722 ceph_msgpool_destroy(&osdc->msgpool_op);
2723 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2727 * Read some contiguous pages. If we cross a stripe boundary, shorten
2728 * *plen. Return number of bytes read, or error.
2730 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2731 struct ceph_vino vino, struct ceph_file_layout *layout,
2733 u32 truncate_seq, u64 truncate_size,
2734 struct page **pages, int num_pages, int page_align)
2736 struct ceph_osd_request *req;
2739 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2740 vino.snap, off, *plen);
2741 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2742 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2743 NULL, truncate_seq, truncate_size,
2746 return PTR_ERR(req);
2748 /* it may be a short read due to an object boundary */
2750 osd_req_op_extent_osd_data_pages(req, 0,
2751 pages, *plen, page_align, false, false);
2753 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2754 off, *plen, *plen, page_align);
2756 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2758 rc = ceph_osdc_start_request(osdc, req, false);
2760 rc = ceph_osdc_wait_request(osdc, req);
2762 ceph_osdc_put_request(req);
2763 dout("readpages result %d\n", rc);
2766 EXPORT_SYMBOL(ceph_osdc_readpages);
2769 * do a synchronous write on N pages
2771 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2772 struct ceph_file_layout *layout,
2773 struct ceph_snap_context *snapc,
2775 u32 truncate_seq, u64 truncate_size,
2776 struct timespec *mtime,
2777 struct page **pages, int num_pages)
2779 struct ceph_osd_request *req;
2781 int page_align = off & ~PAGE_MASK;
2783 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */
2784 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2786 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2787 snapc, truncate_seq, truncate_size,
2790 return PTR_ERR(req);
2792 /* it may be a short write due to an object boundary */
2793 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2795 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2797 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2799 rc = ceph_osdc_start_request(osdc, req, true);
2801 rc = ceph_osdc_wait_request(osdc, req);
2803 ceph_osdc_put_request(req);
2806 dout("writepages result %d\n", rc);
2809 EXPORT_SYMBOL(ceph_osdc_writepages);
2811 int ceph_osdc_setup(void)
2813 size_t size = sizeof(struct ceph_osd_request) +
2814 CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op);
2816 BUG_ON(ceph_osd_request_cache);
2817 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size,
2820 return ceph_osd_request_cache ? 0 : -ENOMEM;
2822 EXPORT_SYMBOL(ceph_osdc_setup);
2824 void ceph_osdc_cleanup(void)
2826 BUG_ON(!ceph_osd_request_cache);
2827 kmem_cache_destroy(ceph_osd_request_cache);
2828 ceph_osd_request_cache = NULL;
2830 EXPORT_SYMBOL(ceph_osdc_cleanup);
2833 * handle incoming message
2835 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2837 struct ceph_osd *osd = con->private;
2838 struct ceph_osd_client *osdc;
2839 int type = le16_to_cpu(msg->hdr.type);
2846 case CEPH_MSG_OSD_MAP:
2847 ceph_osdc_handle_map(osdc, msg);
2849 case CEPH_MSG_OSD_OPREPLY:
2850 handle_reply(osdc, msg);
2852 case CEPH_MSG_WATCH_NOTIFY:
2853 handle_watch_notify(osdc, msg);
2857 pr_err("received unknown message type %d %s\n", type,
2858 ceph_msg_type_name(type));
2865 * Lookup and return message for incoming reply. Don't try to do
2866 * anything about a larger than preallocated data portion of the
2867 * message at the moment - for now, just skip the message.
2869 static struct ceph_msg *get_reply(struct ceph_connection *con,
2870 struct ceph_msg_header *hdr,
2873 struct ceph_osd *osd = con->private;
2874 struct ceph_osd_client *osdc = osd->o_osdc;
2876 struct ceph_osd_request *req;
2877 int front_len = le32_to_cpu(hdr->front_len);
2878 int data_len = le32_to_cpu(hdr->data_len);
2881 tid = le64_to_cpu(hdr->tid);
2882 mutex_lock(&osdc->request_mutex);
2883 req = __lookup_request(osdc, tid);
2885 dout("%s osd%d tid %llu unknown, skipping\n", __func__,
2892 ceph_msg_revoke_incoming(req->r_reply);
2894 if (front_len > req->r_reply->front_alloc_len) {
2895 pr_warn("%s osd%d tid %llu front %d > preallocated %d\n",
2896 __func__, osd->o_osd, req->r_tid, front_len,
2897 req->r_reply->front_alloc_len);
2898 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2902 ceph_msg_put(req->r_reply);
2906 if (data_len > req->r_reply->data_length) {
2907 pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n",
2908 __func__, osd->o_osd, req->r_tid, data_len,
2909 req->r_reply->data_length);
2915 m = ceph_msg_get(req->r_reply);
2916 dout("get_reply tid %lld %p\n", tid, m);
2919 mutex_unlock(&osdc->request_mutex);
2923 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2924 struct ceph_msg_header *hdr,
2927 struct ceph_osd *osd = con->private;
2928 int type = le16_to_cpu(hdr->type);
2929 int front = le32_to_cpu(hdr->front_len);
2933 case CEPH_MSG_OSD_MAP:
2934 case CEPH_MSG_WATCH_NOTIFY:
2935 return ceph_msg_new(type, front, GFP_NOFS, false);
2936 case CEPH_MSG_OSD_OPREPLY:
2937 return get_reply(con, hdr, skip);
2939 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2947 * Wrappers to refcount containing ceph_osd struct
2949 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2951 struct ceph_osd *osd = con->private;
2957 static void put_osd_con(struct ceph_connection *con)
2959 struct ceph_osd *osd = con->private;
2967 * Note: returned pointer is the address of a structure that's
2968 * managed separately. Caller must *not* attempt to free it.
2970 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2971 int *proto, int force_new)
2973 struct ceph_osd *o = con->private;
2974 struct ceph_osd_client *osdc = o->o_osdc;
2975 struct ceph_auth_client *ac = osdc->client->monc.auth;
2976 struct ceph_auth_handshake *auth = &o->o_auth;
2978 if (force_new && auth->authorizer) {
2979 ceph_auth_destroy_authorizer(auth->authorizer);
2980 auth->authorizer = NULL;
2982 if (!auth->authorizer) {
2983 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2986 return ERR_PTR(ret);
2988 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2991 return ERR_PTR(ret);
2993 *proto = ac->protocol;
2999 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3001 struct ceph_osd *o = con->private;
3002 struct ceph_osd_client *osdc = o->o_osdc;
3003 struct ceph_auth_client *ac = osdc->client->monc.auth;
3005 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
3008 static int invalidate_authorizer(struct ceph_connection *con)
3010 struct ceph_osd *o = con->private;
3011 struct ceph_osd_client *osdc = o->o_osdc;
3012 struct ceph_auth_client *ac = osdc->client->monc.auth;
3014 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
3015 return ceph_monc_validate_auth(&osdc->client->monc);
3018 static int osd_sign_message(struct ceph_msg *msg)
3020 struct ceph_osd *o = msg->con->private;
3021 struct ceph_auth_handshake *auth = &o->o_auth;
3023 return ceph_auth_sign_message(auth, msg);
3026 static int osd_check_message_signature(struct ceph_msg *msg)
3028 struct ceph_osd *o = msg->con->private;
3029 struct ceph_auth_handshake *auth = &o->o_auth;
3031 return ceph_auth_check_message_signature(auth, msg);
3034 static const struct ceph_connection_operations osd_con_ops = {
3037 .dispatch = dispatch,
3038 .get_authorizer = get_authorizer,
3039 .verify_authorizer_reply = verify_authorizer_reply,
3040 .invalidate_authorizer = invalidate_authorizer,
3041 .alloc_msg = alloc_msg,
3042 .sign_message = osd_sign_message,
3043 .check_message_signature = osd_check_message_signature,