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 BUG_ON(whch >= (oreq)->r_num_ops); \
126 &(oreq)->r_ops[whch].typ.fld; \
129 static struct ceph_osd_data *
130 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
132 BUG_ON(which >= osd_req->r_num_ops);
134 return &osd_req->r_ops[which].raw_data_in;
137 struct ceph_osd_data *
138 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
141 return osd_req_op_data(osd_req, which, extent, osd_data);
143 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
145 struct ceph_osd_data *
146 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
149 return osd_req_op_data(osd_req, which, cls, response_data);
151 EXPORT_SYMBOL(osd_req_op_cls_response_data); /* ??? */
153 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
154 unsigned int which, struct page **pages,
155 u64 length, u32 alignment,
156 bool pages_from_pool, bool own_pages)
158 struct ceph_osd_data *osd_data;
160 osd_data = osd_req_op_raw_data_in(osd_req, which);
161 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
162 pages_from_pool, own_pages);
164 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
166 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
167 unsigned int which, struct page **pages,
168 u64 length, u32 alignment,
169 bool pages_from_pool, bool own_pages)
171 struct ceph_osd_data *osd_data;
173 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
174 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
175 pages_from_pool, own_pages);
177 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
179 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
180 unsigned int which, struct ceph_pagelist *pagelist)
182 struct ceph_osd_data *osd_data;
184 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
185 ceph_osd_data_pagelist_init(osd_data, pagelist);
187 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
190 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
191 unsigned int which, struct bio *bio, size_t bio_length)
193 struct ceph_osd_data *osd_data;
195 osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
196 ceph_osd_data_bio_init(osd_data, bio, bio_length);
198 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
199 #endif /* CONFIG_BLOCK */
201 static void osd_req_op_cls_request_info_pagelist(
202 struct ceph_osd_request *osd_req,
203 unsigned int which, struct ceph_pagelist *pagelist)
205 struct ceph_osd_data *osd_data;
207 osd_data = osd_req_op_data(osd_req, which, cls, request_info);
208 ceph_osd_data_pagelist_init(osd_data, pagelist);
211 void osd_req_op_cls_request_data_pagelist(
212 struct ceph_osd_request *osd_req,
213 unsigned int which, struct ceph_pagelist *pagelist)
215 struct ceph_osd_data *osd_data;
217 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
218 ceph_osd_data_pagelist_init(osd_data, pagelist);
220 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
222 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
223 unsigned int which, struct page **pages, u64 length,
224 u32 alignment, bool pages_from_pool, bool own_pages)
226 struct ceph_osd_data *osd_data;
228 osd_data = osd_req_op_data(osd_req, which, cls, request_data);
229 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
230 pages_from_pool, own_pages);
232 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
234 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
235 unsigned int which, struct page **pages, u64 length,
236 u32 alignment, bool pages_from_pool, bool own_pages)
238 struct ceph_osd_data *osd_data;
240 osd_data = osd_req_op_data(osd_req, which, cls, response_data);
241 ceph_osd_data_pages_init(osd_data, pages, length, alignment,
242 pages_from_pool, own_pages);
244 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
246 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
248 switch (osd_data->type) {
249 case CEPH_OSD_DATA_TYPE_NONE:
251 case CEPH_OSD_DATA_TYPE_PAGES:
252 return osd_data->length;
253 case CEPH_OSD_DATA_TYPE_PAGELIST:
254 return (u64)osd_data->pagelist->length;
256 case CEPH_OSD_DATA_TYPE_BIO:
257 return (u64)osd_data->bio_length;
258 #endif /* CONFIG_BLOCK */
260 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
265 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
267 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
270 num_pages = calc_pages_for((u64)osd_data->alignment,
271 (u64)osd_data->length);
272 ceph_release_page_vector(osd_data->pages, num_pages);
274 ceph_osd_data_init(osd_data);
277 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
280 struct ceph_osd_req_op *op;
282 BUG_ON(which >= osd_req->r_num_ops);
283 op = &osd_req->r_ops[which];
286 case CEPH_OSD_OP_READ:
287 case CEPH_OSD_OP_WRITE:
288 ceph_osd_data_release(&op->extent.osd_data);
290 case CEPH_OSD_OP_CALL:
291 ceph_osd_data_release(&op->cls.request_info);
292 ceph_osd_data_release(&op->cls.request_data);
293 ceph_osd_data_release(&op->cls.response_data);
303 static void ceph_osdc_release_request(struct kref *kref)
305 struct ceph_osd_request *req = container_of(kref,
306 struct ceph_osd_request, r_kref);
309 dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
310 req->r_request, req->r_reply);
311 WARN_ON(!RB_EMPTY_NODE(&req->r_node));
312 WARN_ON(!list_empty(&req->r_req_lru_item));
313 WARN_ON(!list_empty(&req->r_osd_item));
314 WARN_ON(!list_empty(&req->r_linger_item));
315 WARN_ON(!list_empty(&req->r_linger_osd_item));
319 ceph_msg_put(req->r_request);
321 ceph_msg_revoke_incoming(req->r_reply);
322 ceph_msg_put(req->r_reply);
325 for (which = 0; which < req->r_num_ops; which++)
326 osd_req_op_data_release(req, which);
328 ceph_put_snap_context(req->r_snapc);
330 mempool_free(req, req->r_osdc->req_mempool);
332 kmem_cache_free(ceph_osd_request_cache, req);
336 void ceph_osdc_get_request(struct ceph_osd_request *req)
338 dout("%s %p (was %d)\n", __func__, req,
339 atomic_read(&req->r_kref.refcount));
340 kref_get(&req->r_kref);
342 EXPORT_SYMBOL(ceph_osdc_get_request);
344 void ceph_osdc_put_request(struct ceph_osd_request *req)
346 dout("%s %p (was %d)\n", __func__, req,
347 atomic_read(&req->r_kref.refcount));
348 kref_put(&req->r_kref, ceph_osdc_release_request);
350 EXPORT_SYMBOL(ceph_osdc_put_request);
352 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
353 struct ceph_snap_context *snapc,
354 unsigned int num_ops,
358 struct ceph_osd_request *req;
359 struct ceph_msg *msg;
362 BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
363 BUG_ON(num_ops > CEPH_OSD_MAX_OP);
365 msg_size = 4 + 4 + 8 + 8 + 4+8;
366 msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
367 msg_size += 1 + 8 + 4 + 4; /* pg_t */
368 msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
369 msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
370 msg_size += 8; /* snapid */
371 msg_size += 8; /* snap_seq */
372 msg_size += 8 * (snapc ? snapc->num_snaps : 0); /* snaps */
376 req = mempool_alloc(osdc->req_mempool, gfp_flags);
377 memset(req, 0, sizeof(*req));
379 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
385 req->r_mempool = use_mempool;
386 req->r_num_ops = num_ops;
388 kref_init(&req->r_kref);
389 init_completion(&req->r_completion);
390 init_completion(&req->r_safe_completion);
391 RB_CLEAR_NODE(&req->r_node);
392 INIT_LIST_HEAD(&req->r_unsafe_item);
393 INIT_LIST_HEAD(&req->r_linger_item);
394 INIT_LIST_HEAD(&req->r_linger_osd_item);
395 INIT_LIST_HEAD(&req->r_req_lru_item);
396 INIT_LIST_HEAD(&req->r_osd_item);
398 req->r_base_oloc.pool = -1;
399 req->r_target_oloc.pool = -1;
401 /* create reply message */
403 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
405 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
406 OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
408 ceph_osdc_put_request(req);
413 /* create request message; allow space for oid */
415 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
417 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
419 ceph_osdc_put_request(req);
423 memset(msg->front.iov_base, 0, msg->front.iov_len);
425 req->r_request = msg;
429 EXPORT_SYMBOL(ceph_osdc_alloc_request);
431 static bool osd_req_opcode_valid(u16 opcode)
434 #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true;
435 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
443 * This is an osd op init function for opcodes that have no data or
444 * other information associated with them. It also serves as a
445 * common init routine for all the other init functions, below.
447 static struct ceph_osd_req_op *
448 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
451 struct ceph_osd_req_op *op;
453 BUG_ON(which >= osd_req->r_num_ops);
454 BUG_ON(!osd_req_opcode_valid(opcode));
456 op = &osd_req->r_ops[which];
457 memset(op, 0, sizeof (*op));
463 void osd_req_op_init(struct ceph_osd_request *osd_req,
464 unsigned int which, u16 opcode)
466 (void)_osd_req_op_init(osd_req, which, opcode);
468 EXPORT_SYMBOL(osd_req_op_init);
470 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
471 unsigned int which, u16 opcode,
472 u64 offset, u64 length,
473 u64 truncate_size, u32 truncate_seq)
475 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
476 size_t payload_len = 0;
478 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
479 opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
480 opcode != CEPH_OSD_OP_TRUNCATE);
482 op->extent.offset = offset;
483 op->extent.length = length;
484 op->extent.truncate_size = truncate_size;
485 op->extent.truncate_seq = truncate_seq;
486 if (opcode == CEPH_OSD_OP_WRITE)
487 payload_len += length;
489 op->payload_len = payload_len;
491 EXPORT_SYMBOL(osd_req_op_extent_init);
493 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
494 unsigned int which, u64 length)
496 struct ceph_osd_req_op *op;
499 BUG_ON(which >= osd_req->r_num_ops);
500 op = &osd_req->r_ops[which];
501 previous = op->extent.length;
503 if (length == previous)
504 return; /* Nothing to do */
505 BUG_ON(length > previous);
507 op->extent.length = length;
508 op->payload_len -= previous - length;
510 EXPORT_SYMBOL(osd_req_op_extent_update);
512 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
513 u16 opcode, const char *class, const char *method)
515 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
516 struct ceph_pagelist *pagelist;
517 size_t payload_len = 0;
520 BUG_ON(opcode != CEPH_OSD_OP_CALL);
522 pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
524 ceph_pagelist_init(pagelist);
526 op->cls.class_name = class;
527 size = strlen(class);
528 BUG_ON(size > (size_t) U8_MAX);
529 op->cls.class_len = size;
530 ceph_pagelist_append(pagelist, class, size);
533 op->cls.method_name = method;
534 size = strlen(method);
535 BUG_ON(size > (size_t) U8_MAX);
536 op->cls.method_len = size;
537 ceph_pagelist_append(pagelist, method, size);
540 osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
542 op->cls.argc = 0; /* currently unused */
544 op->payload_len = payload_len;
546 EXPORT_SYMBOL(osd_req_op_cls_init);
548 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
549 unsigned int which, u16 opcode,
550 u64 cookie, u64 version, int flag)
552 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
554 BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
556 op->watch.cookie = cookie;
557 op->watch.ver = version;
558 if (opcode == CEPH_OSD_OP_WATCH && flag)
559 op->watch.flag = (u8)1;
561 EXPORT_SYMBOL(osd_req_op_watch_init);
563 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
565 u64 expected_object_size,
566 u64 expected_write_size)
568 struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
569 CEPH_OSD_OP_SETALLOCHINT);
571 op->alloc_hint.expected_object_size = expected_object_size;
572 op->alloc_hint.expected_write_size = expected_write_size;
575 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
576 * not worth a feature bit. Set FAILOK per-op flag to make
577 * sure older osds don't trip over an unsupported opcode.
579 op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
581 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
583 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
584 struct ceph_osd_data *osd_data)
586 u64 length = ceph_osd_data_length(osd_data);
588 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
589 BUG_ON(length > (u64) SIZE_MAX);
591 ceph_msg_data_add_pages(msg, osd_data->pages,
592 length, osd_data->alignment);
593 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
595 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
597 } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
598 ceph_msg_data_add_bio(msg, osd_data->bio, length);
601 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
605 static u64 osd_req_encode_op(struct ceph_osd_request *req,
606 struct ceph_osd_op *dst, unsigned int which)
608 struct ceph_osd_req_op *src;
609 struct ceph_osd_data *osd_data;
610 u64 request_data_len = 0;
613 BUG_ON(which >= req->r_num_ops);
614 src = &req->r_ops[which];
615 if (WARN_ON(!osd_req_opcode_valid(src->op))) {
616 pr_err("unrecognized osd opcode %d\n", src->op);
622 case CEPH_OSD_OP_STAT:
623 osd_data = &src->raw_data_in;
624 ceph_osdc_msg_data_add(req->r_reply, osd_data);
626 case CEPH_OSD_OP_READ:
627 case CEPH_OSD_OP_WRITE:
628 case CEPH_OSD_OP_ZERO:
629 case CEPH_OSD_OP_DELETE:
630 case CEPH_OSD_OP_TRUNCATE:
631 if (src->op == CEPH_OSD_OP_WRITE)
632 request_data_len = src->extent.length;
633 dst->extent.offset = cpu_to_le64(src->extent.offset);
634 dst->extent.length = cpu_to_le64(src->extent.length);
635 dst->extent.truncate_size =
636 cpu_to_le64(src->extent.truncate_size);
637 dst->extent.truncate_seq =
638 cpu_to_le32(src->extent.truncate_seq);
639 osd_data = &src->extent.osd_data;
640 if (src->op == CEPH_OSD_OP_WRITE)
641 ceph_osdc_msg_data_add(req->r_request, osd_data);
643 ceph_osdc_msg_data_add(req->r_reply, osd_data);
645 case CEPH_OSD_OP_CALL:
646 dst->cls.class_len = src->cls.class_len;
647 dst->cls.method_len = src->cls.method_len;
648 osd_data = &src->cls.request_info;
649 ceph_osdc_msg_data_add(req->r_request, osd_data);
650 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
651 request_data_len = osd_data->pagelist->length;
653 osd_data = &src->cls.request_data;
654 data_length = ceph_osd_data_length(osd_data);
656 BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
657 dst->cls.indata_len = cpu_to_le32(data_length);
658 ceph_osdc_msg_data_add(req->r_request, osd_data);
659 src->payload_len += data_length;
660 request_data_len += data_length;
662 osd_data = &src->cls.response_data;
663 ceph_osdc_msg_data_add(req->r_reply, osd_data);
665 case CEPH_OSD_OP_STARTSYNC:
667 case CEPH_OSD_OP_NOTIFY_ACK:
668 case CEPH_OSD_OP_WATCH:
669 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
670 dst->watch.ver = cpu_to_le64(src->watch.ver);
671 dst->watch.flag = src->watch.flag;
673 case CEPH_OSD_OP_SETALLOCHINT:
674 dst->alloc_hint.expected_object_size =
675 cpu_to_le64(src->alloc_hint.expected_object_size);
676 dst->alloc_hint.expected_write_size =
677 cpu_to_le64(src->alloc_hint.expected_write_size);
680 pr_err("unsupported osd opcode %s\n",
681 ceph_osd_op_name(src->op));
687 dst->op = cpu_to_le16(src->op);
688 dst->flags = cpu_to_le32(src->flags);
689 dst->payload_len = cpu_to_le32(src->payload_len);
691 return request_data_len;
695 * build new request AND message, calculate layout, and adjust file
698 * if the file was recently truncated, we include information about its
699 * old and new size so that the object can be updated appropriately. (we
700 * avoid synchronously deleting truncated objects because it's slow.)
702 * if @do_sync, include a 'startsync' command so that the osd will flush
705 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
706 struct ceph_file_layout *layout,
707 struct ceph_vino vino,
708 u64 off, u64 *plen, int num_ops,
709 int opcode, int flags,
710 struct ceph_snap_context *snapc,
715 struct ceph_osd_request *req;
723 BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
724 opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
725 opcode != CEPH_OSD_OP_TRUNCATE);
727 req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
730 return ERR_PTR(-ENOMEM);
732 req->r_flags = flags;
734 /* calculate max write size */
735 r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
737 ceph_osdc_put_request(req);
741 object_size = le32_to_cpu(layout->fl_object_size);
742 object_base = off - objoff;
743 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
744 if (truncate_size <= object_base) {
747 truncate_size -= object_base;
748 if (truncate_size > object_size)
749 truncate_size = object_size;
753 osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
754 truncate_size, truncate_seq);
757 * A second op in the ops array means the caller wants to
758 * also issue a include a 'startsync' command so that the
759 * osd will flush data quickly.
762 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
764 req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
766 snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
767 "%llx.%08llx", vino.ino, objnum);
768 req->r_base_oid.name_len = strlen(req->r_base_oid.name);
772 EXPORT_SYMBOL(ceph_osdc_new_request);
775 * We keep osd requests in an rbtree, sorted by ->r_tid.
777 static void __insert_request(struct ceph_osd_client *osdc,
778 struct ceph_osd_request *new)
780 struct rb_node **p = &osdc->requests.rb_node;
781 struct rb_node *parent = NULL;
782 struct ceph_osd_request *req = NULL;
786 req = rb_entry(parent, struct ceph_osd_request, r_node);
787 if (new->r_tid < req->r_tid)
789 else if (new->r_tid > req->r_tid)
795 rb_link_node(&new->r_node, parent, p);
796 rb_insert_color(&new->r_node, &osdc->requests);
799 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
802 struct ceph_osd_request *req;
803 struct rb_node *n = osdc->requests.rb_node;
806 req = rb_entry(n, struct ceph_osd_request, r_node);
807 if (tid < req->r_tid)
809 else if (tid > req->r_tid)
817 static struct ceph_osd_request *
818 __lookup_request_ge(struct ceph_osd_client *osdc,
821 struct ceph_osd_request *req;
822 struct rb_node *n = osdc->requests.rb_node;
825 req = rb_entry(n, struct ceph_osd_request, r_node);
826 if (tid < req->r_tid) {
830 } else if (tid > req->r_tid) {
839 static void __kick_linger_request(struct ceph_osd_request *req)
841 struct ceph_osd_client *osdc = req->r_osdc;
842 struct ceph_osd *osd = req->r_osd;
845 * Linger requests need to be resent with a new tid to avoid
846 * the dup op detection logic on the OSDs. Achieve this with
847 * a re-register dance instead of open-coding.
849 ceph_osdc_get_request(req);
850 if (!list_empty(&req->r_linger_item))
851 __unregister_linger_request(osdc, req);
853 __unregister_request(osdc, req);
854 __register_request(osdc, req);
855 ceph_osdc_put_request(req);
858 * Unless request has been registered as both normal and
859 * lingering, __unregister{,_linger}_request clears r_osd.
860 * However, here we need to preserve r_osd to make sure we
861 * requeue on the same OSD.
863 WARN_ON(req->r_osd || !osd);
866 dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
867 __enqueue_request(req);
871 * Resubmit requests pending on the given osd.
873 static void __kick_osd_requests(struct ceph_osd_client *osdc,
874 struct ceph_osd *osd)
876 struct ceph_osd_request *req, *nreq;
878 LIST_HEAD(resend_linger);
881 dout("%s osd%d\n", __func__, osd->o_osd);
882 err = __reset_osd(osdc, osd);
887 * Build up a list of requests to resend by traversing the
888 * osd's list of requests. Requests for a given object are
889 * sent in tid order, and that is also the order they're
890 * kept on this list. Therefore all requests that are in
891 * flight will be found first, followed by all requests that
892 * have not yet been sent. And to resend requests while
893 * preserving this order we will want to put any sent
894 * requests back on the front of the osd client's unsent
897 * So we build a separate ordered list of already-sent
898 * requests for the affected osd and splice it onto the
899 * front of the osd client's unsent list. Once we've seen a
900 * request that has not yet been sent we're done. Those
901 * requests are already sitting right where they belong.
903 list_for_each_entry(req, &osd->o_requests, r_osd_item) {
907 if (!req->r_linger) {
908 dout("%s requeueing %p tid %llu\n", __func__, req,
910 list_move_tail(&req->r_req_lru_item, &resend);
911 req->r_flags |= CEPH_OSD_FLAG_RETRY;
913 list_move_tail(&req->r_req_lru_item, &resend_linger);
916 list_splice(&resend, &osdc->req_unsent);
919 * Both registered and not yet registered linger requests are
920 * enqueued with a new tid on the same OSD. We add/move them
921 * to req_unsent/o_requests at the end to keep things in tid
924 list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
926 WARN_ON(!list_empty(&req->r_req_lru_item));
927 __kick_linger_request(req);
930 list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
931 __kick_linger_request(req);
935 * If the osd connection drops, we need to resubmit all requests.
937 static void osd_reset(struct ceph_connection *con)
939 struct ceph_osd *osd = con->private;
940 struct ceph_osd_client *osdc;
944 dout("osd_reset osd%d\n", osd->o_osd);
946 down_read(&osdc->map_sem);
947 mutex_lock(&osdc->request_mutex);
948 __kick_osd_requests(osdc, osd);
950 mutex_unlock(&osdc->request_mutex);
951 up_read(&osdc->map_sem);
955 * Track open sessions with osds.
957 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
959 struct ceph_osd *osd;
961 osd = kzalloc(sizeof(*osd), GFP_NOFS);
965 atomic_set(&osd->o_ref, 1);
968 RB_CLEAR_NODE(&osd->o_node);
969 INIT_LIST_HEAD(&osd->o_requests);
970 INIT_LIST_HEAD(&osd->o_linger_requests);
971 INIT_LIST_HEAD(&osd->o_osd_lru);
972 osd->o_incarnation = 1;
974 ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
976 INIT_LIST_HEAD(&osd->o_keepalive_item);
980 static struct ceph_osd *get_osd(struct ceph_osd *osd)
982 if (atomic_inc_not_zero(&osd->o_ref)) {
983 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
984 atomic_read(&osd->o_ref));
987 dout("get_osd %p FAIL\n", osd);
992 static void put_osd(struct ceph_osd *osd)
994 dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
995 atomic_read(&osd->o_ref) - 1);
996 if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
997 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
999 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1005 * remove an osd from our map
1007 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1009 dout("__remove_osd %p\n", osd);
1010 BUG_ON(!list_empty(&osd->o_requests));
1011 BUG_ON(!list_empty(&osd->o_linger_requests));
1013 rb_erase(&osd->o_node, &osdc->osds);
1014 list_del_init(&osd->o_osd_lru);
1015 ceph_con_close(&osd->o_con);
1019 static void remove_all_osds(struct ceph_osd_client *osdc)
1021 dout("%s %p\n", __func__, osdc);
1022 mutex_lock(&osdc->request_mutex);
1023 while (!RB_EMPTY_ROOT(&osdc->osds)) {
1024 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1025 struct ceph_osd, o_node);
1026 __remove_osd(osdc, osd);
1028 mutex_unlock(&osdc->request_mutex);
1031 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1032 struct ceph_osd *osd)
1034 dout("%s %p\n", __func__, osd);
1035 BUG_ON(!list_empty(&osd->o_osd_lru));
1037 list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1038 osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1041 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1042 struct ceph_osd *osd)
1044 dout("%s %p\n", __func__, osd);
1046 if (list_empty(&osd->o_requests) &&
1047 list_empty(&osd->o_linger_requests))
1048 __move_osd_to_lru(osdc, osd);
1051 static void __remove_osd_from_lru(struct ceph_osd *osd)
1053 dout("__remove_osd_from_lru %p\n", osd);
1054 if (!list_empty(&osd->o_osd_lru))
1055 list_del_init(&osd->o_osd_lru);
1058 static void remove_old_osds(struct ceph_osd_client *osdc)
1060 struct ceph_osd *osd, *nosd;
1062 dout("__remove_old_osds %p\n", osdc);
1063 mutex_lock(&osdc->request_mutex);
1064 list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1065 if (time_before(jiffies, osd->lru_ttl))
1067 __remove_osd(osdc, osd);
1069 mutex_unlock(&osdc->request_mutex);
1075 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1077 struct ceph_entity_addr *peer_addr;
1079 dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1080 if (list_empty(&osd->o_requests) &&
1081 list_empty(&osd->o_linger_requests)) {
1082 __remove_osd(osdc, osd);
1087 peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1088 if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1089 !ceph_con_opened(&osd->o_con)) {
1090 struct ceph_osd_request *req;
1092 dout("osd addr hasn't changed and connection never opened, "
1093 "letting msgr retry\n");
1094 /* touch each r_stamp for handle_timeout()'s benfit */
1095 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1096 req->r_stamp = jiffies;
1101 ceph_con_close(&osd->o_con);
1102 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1103 osd->o_incarnation++;
1108 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1110 struct rb_node **p = &osdc->osds.rb_node;
1111 struct rb_node *parent = NULL;
1112 struct ceph_osd *osd = NULL;
1114 dout("__insert_osd %p osd%d\n", new, new->o_osd);
1117 osd = rb_entry(parent, struct ceph_osd, o_node);
1118 if (new->o_osd < osd->o_osd)
1120 else if (new->o_osd > osd->o_osd)
1121 p = &(*p)->rb_right;
1126 rb_link_node(&new->o_node, parent, p);
1127 rb_insert_color(&new->o_node, &osdc->osds);
1130 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1132 struct ceph_osd *osd;
1133 struct rb_node *n = osdc->osds.rb_node;
1136 osd = rb_entry(n, struct ceph_osd, o_node);
1139 else if (o > osd->o_osd)
1147 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1149 schedule_delayed_work(&osdc->timeout_work,
1150 osdc->client->options->osd_keepalive_timeout * HZ);
1153 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1155 cancel_delayed_work(&osdc->timeout_work);
1159 * Register request, assign tid. If this is the first request, set up
1160 * the timeout event.
1162 static void __register_request(struct ceph_osd_client *osdc,
1163 struct ceph_osd_request *req)
1165 req->r_tid = ++osdc->last_tid;
1166 req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1167 dout("__register_request %p tid %lld\n", req, req->r_tid);
1168 __insert_request(osdc, req);
1169 ceph_osdc_get_request(req);
1170 osdc->num_requests++;
1171 if (osdc->num_requests == 1) {
1172 dout(" first request, scheduling timeout\n");
1173 __schedule_osd_timeout(osdc);
1178 * called under osdc->request_mutex
1180 static void __unregister_request(struct ceph_osd_client *osdc,
1181 struct ceph_osd_request *req)
1183 if (RB_EMPTY_NODE(&req->r_node)) {
1184 dout("__unregister_request %p tid %lld not registered\n",
1189 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1190 rb_erase(&req->r_node, &osdc->requests);
1191 RB_CLEAR_NODE(&req->r_node);
1192 osdc->num_requests--;
1195 /* make sure the original request isn't in flight. */
1196 ceph_msg_revoke(req->r_request);
1198 list_del_init(&req->r_osd_item);
1199 maybe_move_osd_to_lru(osdc, req->r_osd);
1200 if (list_empty(&req->r_linger_osd_item))
1204 list_del_init(&req->r_req_lru_item);
1205 ceph_osdc_put_request(req);
1207 if (osdc->num_requests == 0) {
1208 dout(" no requests, canceling timeout\n");
1209 __cancel_osd_timeout(osdc);
1214 * Cancel a previously queued request message
1216 static void __cancel_request(struct ceph_osd_request *req)
1218 if (req->r_sent && req->r_osd) {
1219 ceph_msg_revoke(req->r_request);
1224 static void __register_linger_request(struct ceph_osd_client *osdc,
1225 struct ceph_osd_request *req)
1227 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1228 WARN_ON(!req->r_linger);
1230 ceph_osdc_get_request(req);
1231 list_add_tail(&req->r_linger_item, &osdc->req_linger);
1233 list_add_tail(&req->r_linger_osd_item,
1234 &req->r_osd->o_linger_requests);
1237 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1238 struct ceph_osd_request *req)
1240 WARN_ON(!req->r_linger);
1242 if (list_empty(&req->r_linger_item)) {
1243 dout("%s %p tid %llu not registered\n", __func__, req,
1248 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1249 list_del_init(&req->r_linger_item);
1252 list_del_init(&req->r_linger_osd_item);
1253 maybe_move_osd_to_lru(osdc, req->r_osd);
1254 if (list_empty(&req->r_osd_item))
1257 ceph_osdc_put_request(req);
1260 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1261 struct ceph_osd_request *req)
1263 if (!req->r_linger) {
1264 dout("set_request_linger %p\n", req);
1268 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1271 * Returns whether a request should be blocked from being sent
1272 * based on the current osdmap and osd_client settings.
1274 * Caller should hold map_sem for read.
1276 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1277 struct ceph_osd_request *req)
1279 bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1280 bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1281 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1282 return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1283 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1287 * Calculate mapping of a request to a PG. Takes tiering into account.
1289 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1290 struct ceph_osd_request *req,
1291 struct ceph_pg *pg_out)
1293 bool need_check_tiering;
1295 need_check_tiering = false;
1296 if (req->r_target_oloc.pool == -1) {
1297 req->r_target_oloc = req->r_base_oloc; /* struct */
1298 need_check_tiering = true;
1300 if (req->r_target_oid.name_len == 0) {
1301 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1302 need_check_tiering = true;
1305 if (need_check_tiering &&
1306 (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1307 struct ceph_pg_pool_info *pi;
1309 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1311 if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1313 req->r_target_oloc.pool = pi->read_tier;
1314 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1315 pi->write_tier >= 0)
1316 req->r_target_oloc.pool = pi->write_tier;
1318 /* !pi is caught in ceph_oloc_oid_to_pg() */
1321 return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1322 &req->r_target_oid, pg_out);
1325 static void __enqueue_request(struct ceph_osd_request *req)
1327 struct ceph_osd_client *osdc = req->r_osdc;
1329 dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1330 req->r_osd ? req->r_osd->o_osd : -1);
1333 __remove_osd_from_lru(req->r_osd);
1334 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1335 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1337 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1342 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1343 * (as needed), and set the request r_osd appropriately. If there is
1344 * no up osd, set r_osd to NULL. Move the request to the appropriate list
1345 * (unsent, homeless) or leave on in-flight lru.
1347 * Return 0 if unchanged, 1 if changed, or negative on error.
1349 * Caller should hold map_sem for read and request_mutex.
1351 static int __map_request(struct ceph_osd_client *osdc,
1352 struct ceph_osd_request *req, int force_resend)
1354 struct ceph_pg pgid;
1355 int acting[CEPH_PG_MAX_SIZE];
1360 dout("map_request %p tid %lld\n", req, req->r_tid);
1362 err = __calc_request_pg(osdc->osdmap, req, &pgid);
1364 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1369 num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1373 was_paused = req->r_paused;
1374 req->r_paused = __req_should_be_paused(osdc, req);
1375 if (was_paused && !req->r_paused)
1378 if ((!force_resend &&
1379 req->r_osd && req->r_osd->o_osd == o &&
1380 req->r_sent >= req->r_osd->o_incarnation &&
1381 req->r_num_pg_osds == num &&
1382 memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1383 (req->r_osd == NULL && o == -1) ||
1385 return 0; /* no change */
1387 dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1388 req->r_tid, pgid.pool, pgid.seed, o,
1389 req->r_osd ? req->r_osd->o_osd : -1);
1391 /* record full pg acting set */
1392 memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1393 req->r_num_pg_osds = num;
1396 __cancel_request(req);
1397 list_del_init(&req->r_osd_item);
1401 req->r_osd = __lookup_osd(osdc, o);
1402 if (!req->r_osd && o >= 0) {
1404 req->r_osd = create_osd(osdc, o);
1406 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1410 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1411 __insert_osd(osdc, req->r_osd);
1413 ceph_con_open(&req->r_osd->o_con,
1414 CEPH_ENTITY_TYPE_OSD, o,
1415 &osdc->osdmap->osd_addr[o]);
1418 __enqueue_request(req);
1419 err = 1; /* osd or pg changed */
1426 * caller should hold map_sem (for read) and request_mutex
1428 static void __send_request(struct ceph_osd_client *osdc,
1429 struct ceph_osd_request *req)
1433 dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1434 req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1435 (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1437 /* fill in message content that changes each time we send it */
1438 put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1439 put_unaligned_le32(req->r_flags, req->r_request_flags);
1440 put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1441 p = req->r_request_pgid;
1442 ceph_encode_64(&p, req->r_pgid.pool);
1443 ceph_encode_32(&p, req->r_pgid.seed);
1444 put_unaligned_le64(1, req->r_request_attempts); /* FIXME */
1445 memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1446 sizeof(req->r_reassert_version));
1448 req->r_stamp = jiffies;
1449 list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1451 ceph_msg_get(req->r_request); /* send consumes a ref */
1453 req->r_sent = req->r_osd->o_incarnation;
1455 ceph_con_send(&req->r_osd->o_con, req->r_request);
1459 * Send any requests in the queue (req_unsent).
1461 static void __send_queued(struct ceph_osd_client *osdc)
1463 struct ceph_osd_request *req, *tmp;
1465 dout("__send_queued\n");
1466 list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1467 __send_request(osdc, req);
1471 * Caller should hold map_sem for read and request_mutex.
1473 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1474 struct ceph_osd_request *req,
1479 __register_request(osdc, req);
1481 req->r_got_reply = 0;
1482 rc = __map_request(osdc, req, 0);
1485 dout("osdc_start_request failed map, "
1486 " will retry %lld\n", req->r_tid);
1489 __unregister_request(osdc, req);
1494 if (req->r_osd == NULL) {
1495 dout("send_request %p no up osds in pg\n", req);
1496 ceph_monc_request_next_osdmap(&osdc->client->monc);
1498 __send_queued(osdc);
1505 * Timeout callback, called every N seconds when 1 or more osd
1506 * requests has been active for more than N seconds. When this
1507 * happens, we ping all OSDs with requests who have timed out to
1508 * ensure any communications channel reset is detected. Reset the
1509 * request timeouts another N seconds in the future as we go.
1510 * Reschedule the timeout event another N seconds in future (unless
1511 * there are no open requests).
1513 static void handle_timeout(struct work_struct *work)
1515 struct ceph_osd_client *osdc =
1516 container_of(work, struct ceph_osd_client, timeout_work.work);
1517 struct ceph_osd_request *req;
1518 struct ceph_osd *osd;
1519 unsigned long keepalive =
1520 osdc->client->options->osd_keepalive_timeout * HZ;
1521 struct list_head slow_osds;
1523 down_read(&osdc->map_sem);
1525 ceph_monc_request_next_osdmap(&osdc->client->monc);
1527 mutex_lock(&osdc->request_mutex);
1530 * ping osds that are a bit slow. this ensures that if there
1531 * is a break in the TCP connection we will notice, and reopen
1532 * a connection with that osd (from the fault callback).
1534 INIT_LIST_HEAD(&slow_osds);
1535 list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1536 if (time_before(jiffies, req->r_stamp + keepalive))
1541 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1542 req->r_tid, osd->o_osd);
1543 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1545 while (!list_empty(&slow_osds)) {
1546 osd = list_entry(slow_osds.next, struct ceph_osd,
1548 list_del_init(&osd->o_keepalive_item);
1549 ceph_con_keepalive(&osd->o_con);
1552 __schedule_osd_timeout(osdc);
1553 __send_queued(osdc);
1554 mutex_unlock(&osdc->request_mutex);
1555 up_read(&osdc->map_sem);
1558 static void handle_osds_timeout(struct work_struct *work)
1560 struct ceph_osd_client *osdc =
1561 container_of(work, struct ceph_osd_client,
1562 osds_timeout_work.work);
1563 unsigned long delay =
1564 osdc->client->options->osd_idle_ttl * HZ >> 2;
1566 dout("osds timeout\n");
1567 down_read(&osdc->map_sem);
1568 remove_old_osds(osdc);
1569 up_read(&osdc->map_sem);
1571 schedule_delayed_work(&osdc->osds_timeout_work,
1572 round_jiffies_relative(delay));
1575 static int ceph_oloc_decode(void **p, void *end,
1576 struct ceph_object_locator *oloc)
1578 u8 struct_v, struct_cv;
1583 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1584 struct_v = ceph_decode_8(p);
1585 struct_cv = ceph_decode_8(p);
1587 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1588 struct_v, struct_cv);
1591 if (struct_cv > 6) {
1592 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1593 struct_v, struct_cv);
1596 len = ceph_decode_32(p);
1597 ceph_decode_need(p, end, len, e_inval);
1598 struct_end = *p + len;
1600 oloc->pool = ceph_decode_64(p);
1601 *p += 4; /* skip preferred */
1603 len = ceph_decode_32(p);
1605 pr_warn("ceph_object_locator::key is set\n");
1609 if (struct_v >= 5) {
1610 len = ceph_decode_32(p);
1612 pr_warn("ceph_object_locator::nspace is set\n");
1617 if (struct_v >= 6) {
1618 s64 hash = ceph_decode_64(p);
1620 pr_warn("ceph_object_locator::hash is set\n");
1635 static int ceph_redirect_decode(void **p, void *end,
1636 struct ceph_request_redirect *redir)
1638 u8 struct_v, struct_cv;
1643 ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1644 struct_v = ceph_decode_8(p);
1645 struct_cv = ceph_decode_8(p);
1646 if (struct_cv > 1) {
1647 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1648 struct_v, struct_cv);
1651 len = ceph_decode_32(p);
1652 ceph_decode_need(p, end, len, e_inval);
1653 struct_end = *p + len;
1655 ret = ceph_oloc_decode(p, end, &redir->oloc);
1659 len = ceph_decode_32(p);
1661 pr_warn("ceph_request_redirect::object_name is set\n");
1665 len = ceph_decode_32(p);
1666 *p += len; /* skip osd_instructions */
1678 static void complete_request(struct ceph_osd_request *req)
1680 complete_all(&req->r_safe_completion); /* fsync waiter */
1684 * handle osd op reply. either call the callback if it is specified,
1685 * or do the completion to wake up the waiting thread.
1687 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1688 struct ceph_connection *con)
1691 struct ceph_osd_request *req;
1692 struct ceph_request_redirect redir;
1695 unsigned int numops;
1696 int payload_len, flags;
1702 u64 reassert_version;
1704 int already_completed;
1708 tid = le64_to_cpu(msg->hdr.tid);
1709 dout("handle_reply %p tid %llu\n", msg, tid);
1711 p = msg->front.iov_base;
1712 end = p + msg->front.iov_len;
1714 ceph_decode_need(&p, end, 4, bad);
1715 object_len = ceph_decode_32(&p);
1716 ceph_decode_need(&p, end, object_len, bad);
1719 err = ceph_decode_pgid(&p, end, &pg);
1723 ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1724 flags = ceph_decode_64(&p);
1725 result = ceph_decode_32(&p);
1726 reassert_epoch = ceph_decode_32(&p);
1727 reassert_version = ceph_decode_64(&p);
1728 osdmap_epoch = ceph_decode_32(&p);
1731 down_read(&osdc->map_sem);
1732 mutex_lock(&osdc->request_mutex);
1733 req = __lookup_request(osdc, tid);
1735 dout("handle_reply tid %llu dne\n", tid);
1738 ceph_osdc_get_request(req);
1740 dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1743 ceph_decode_need(&p, end, 4, bad_put);
1744 numops = ceph_decode_32(&p);
1745 if (numops > CEPH_OSD_MAX_OP)
1747 if (numops != req->r_num_ops)
1750 ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1751 for (i = 0; i < numops; i++) {
1752 struct ceph_osd_op *op = p;
1755 len = le32_to_cpu(op->payload_len);
1756 req->r_reply_op_len[i] = len;
1757 dout(" op %d has %d bytes\n", i, len);
1761 bytes = le32_to_cpu(msg->hdr.data_len);
1762 if (payload_len != bytes) {
1763 pr_warn("sum of op payload lens %d != data_len %d\n",
1764 payload_len, bytes);
1768 ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1769 retry_attempt = ceph_decode_32(&p);
1770 for (i = 0; i < numops; i++)
1771 req->r_reply_op_result[i] = ceph_decode_32(&p);
1773 if (le16_to_cpu(msg->hdr.version) >= 6) {
1774 p += 8 + 4; /* skip replay_version */
1775 p += 8; /* skip user_version */
1777 err = ceph_redirect_decode(&p, end, &redir);
1781 redir.oloc.pool = -1;
1784 if (redir.oloc.pool != -1) {
1785 dout("redirect pool %lld\n", redir.oloc.pool);
1787 __unregister_request(osdc, req);
1789 req->r_target_oloc = redir.oloc; /* struct */
1792 * Start redirect requests with nofail=true. If
1793 * mapping fails, request will end up on the notarget
1794 * list, waiting for the new osdmap (which can take
1795 * a while), even though the original request mapped
1796 * successfully. In the future we might want to follow
1797 * original request's nofail setting here.
1799 err = __ceph_osdc_start_request(osdc, req, true);
1805 already_completed = req->r_got_reply;
1806 if (!req->r_got_reply) {
1807 req->r_result = result;
1808 dout("handle_reply result %d bytes %d\n", req->r_result,
1810 if (req->r_result == 0)
1811 req->r_result = bytes;
1813 /* in case this is a write and we need to replay, */
1814 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1815 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1817 req->r_got_reply = 1;
1818 } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1819 dout("handle_reply tid %llu dup ack\n", tid);
1823 dout("handle_reply tid %llu flags %d\n", tid, flags);
1825 if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1826 __register_linger_request(osdc, req);
1828 /* either this is a read, or we got the safe response */
1830 (flags & CEPH_OSD_FLAG_ONDISK) ||
1831 ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1832 __unregister_request(osdc, req);
1834 mutex_unlock(&osdc->request_mutex);
1835 up_read(&osdc->map_sem);
1837 if (!already_completed) {
1838 if (req->r_unsafe_callback &&
1839 result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1840 req->r_unsafe_callback(req, true);
1841 if (req->r_callback)
1842 req->r_callback(req, msg);
1844 complete_all(&req->r_completion);
1847 if (flags & CEPH_OSD_FLAG_ONDISK) {
1848 if (req->r_unsafe_callback && already_completed)
1849 req->r_unsafe_callback(req, false);
1850 complete_request(req);
1854 dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1855 ceph_osdc_put_request(req);
1858 mutex_unlock(&osdc->request_mutex);
1859 up_read(&osdc->map_sem);
1863 req->r_result = -EIO;
1864 __unregister_request(osdc, req);
1865 if (req->r_callback)
1866 req->r_callback(req, msg);
1868 complete_all(&req->r_completion);
1869 complete_request(req);
1870 ceph_osdc_put_request(req);
1872 mutex_unlock(&osdc->request_mutex);
1873 up_read(&osdc->map_sem);
1875 pr_err("corrupt osd_op_reply got %d %d\n",
1876 (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1880 static void reset_changed_osds(struct ceph_osd_client *osdc)
1882 struct rb_node *p, *n;
1884 for (p = rb_first(&osdc->osds); p; p = n) {
1885 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1888 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1889 memcmp(&osd->o_con.peer_addr,
1890 ceph_osd_addr(osdc->osdmap,
1892 sizeof(struct ceph_entity_addr)) != 0)
1893 __reset_osd(osdc, osd);
1898 * Requeue requests whose mapping to an OSD has changed. If requests map to
1899 * no osd, request a new map.
1901 * Caller should hold map_sem for read.
1903 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1904 bool force_resend_writes)
1906 struct ceph_osd_request *req, *nreq;
1910 bool force_resend_req;
1912 dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1913 force_resend_writes ? " (force resend writes)" : "");
1914 mutex_lock(&osdc->request_mutex);
1915 for (p = rb_first(&osdc->requests); p; ) {
1916 req = rb_entry(p, struct ceph_osd_request, r_node);
1920 * For linger requests that have not yet been
1921 * registered, move them to the linger list; they'll
1922 * be sent to the osd in the loop below. Unregister
1923 * the request before re-registering it as a linger
1924 * request to ensure the __map_request() below
1925 * will decide it needs to be sent.
1927 if (req->r_linger && list_empty(&req->r_linger_item)) {
1928 dout("%p tid %llu restart on osd%d\n",
1930 req->r_osd ? req->r_osd->o_osd : -1);
1931 ceph_osdc_get_request(req);
1932 __unregister_request(osdc, req);
1933 __register_linger_request(osdc, req);
1934 ceph_osdc_put_request(req);
1938 force_resend_req = force_resend ||
1939 (force_resend_writes &&
1940 req->r_flags & CEPH_OSD_FLAG_WRITE);
1941 err = __map_request(osdc, req, force_resend_req);
1943 continue; /* error */
1944 if (req->r_osd == NULL) {
1945 dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1946 needmap++; /* request a newer map */
1947 } else if (err > 0) {
1948 if (!req->r_linger) {
1949 dout("%p tid %llu requeued on osd%d\n", req,
1951 req->r_osd ? req->r_osd->o_osd : -1);
1952 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1957 list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1959 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1961 err = __map_request(osdc, req,
1962 force_resend || force_resend_writes);
1963 dout("__map_request returned %d\n", err);
1965 continue; /* no change and no osd was specified */
1967 continue; /* hrm! */
1968 if (req->r_osd == NULL) {
1969 dout("tid %llu maps to no valid osd\n", req->r_tid);
1970 needmap++; /* request a newer map */
1974 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1975 req->r_osd ? req->r_osd->o_osd : -1);
1976 __register_request(osdc, req);
1977 __unregister_linger_request(osdc, req);
1979 reset_changed_osds(osdc);
1980 mutex_unlock(&osdc->request_mutex);
1983 dout("%d requests for down osds, need new map\n", needmap);
1984 ceph_monc_request_next_osdmap(&osdc->client->monc);
1990 * Process updated osd map.
1992 * The message contains any number of incremental and full maps, normally
1993 * indicating some sort of topology change in the cluster. Kick requests
1994 * off to different OSDs as needed.
1996 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1998 void *p, *end, *next;
1999 u32 nr_maps, maplen;
2001 struct ceph_osdmap *newmap = NULL, *oldmap;
2003 struct ceph_fsid fsid;
2006 dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2007 p = msg->front.iov_base;
2008 end = p + msg->front.iov_len;
2011 ceph_decode_need(&p, end, sizeof(fsid), bad);
2012 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2013 if (ceph_check_fsid(osdc->client, &fsid) < 0)
2016 down_write(&osdc->map_sem);
2018 was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2020 /* incremental maps */
2021 ceph_decode_32_safe(&p, end, nr_maps, bad);
2022 dout(" %d inc maps\n", nr_maps);
2023 while (nr_maps > 0) {
2024 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2025 epoch = ceph_decode_32(&p);
2026 maplen = ceph_decode_32(&p);
2027 ceph_decode_need(&p, end, maplen, bad);
2029 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2030 dout("applying incremental map %u len %d\n",
2032 newmap = osdmap_apply_incremental(&p, next,
2034 &osdc->client->msgr);
2035 if (IS_ERR(newmap)) {
2036 err = PTR_ERR(newmap);
2040 if (newmap != osdc->osdmap) {
2041 ceph_osdmap_destroy(osdc->osdmap);
2042 osdc->osdmap = newmap;
2044 was_full = was_full ||
2045 ceph_osdmap_flag(osdc->osdmap,
2047 kick_requests(osdc, 0, was_full);
2049 dout("ignoring incremental map %u len %d\n",
2059 ceph_decode_32_safe(&p, end, nr_maps, bad);
2060 dout(" %d full maps\n", nr_maps);
2062 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2063 epoch = ceph_decode_32(&p);
2064 maplen = ceph_decode_32(&p);
2065 ceph_decode_need(&p, end, maplen, bad);
2067 dout("skipping non-latest full map %u len %d\n",
2069 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2070 dout("skipping full map %u len %d, "
2071 "older than our %u\n", epoch, maplen,
2072 osdc->osdmap->epoch);
2074 int skipped_map = 0;
2076 dout("taking full map %u len %d\n", epoch, maplen);
2077 newmap = ceph_osdmap_decode(&p, p+maplen);
2078 if (IS_ERR(newmap)) {
2079 err = PTR_ERR(newmap);
2083 oldmap = osdc->osdmap;
2084 osdc->osdmap = newmap;
2086 if (oldmap->epoch + 1 < newmap->epoch)
2088 ceph_osdmap_destroy(oldmap);
2090 was_full = was_full ||
2091 ceph_osdmap_flag(osdc->osdmap,
2093 kick_requests(osdc, skipped_map, was_full);
2102 downgrade_write(&osdc->map_sem);
2103 ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2106 * subscribe to subsequent osdmap updates if full to ensure
2107 * we find out when we are no longer full and stop returning
2110 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2111 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2112 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2113 ceph_monc_request_next_osdmap(&osdc->client->monc);
2115 mutex_lock(&osdc->request_mutex);
2116 __send_queued(osdc);
2117 mutex_unlock(&osdc->request_mutex);
2118 up_read(&osdc->map_sem);
2119 wake_up_all(&osdc->client->auth_wq);
2123 pr_err("osdc handle_map corrupt msg\n");
2125 up_write(&osdc->map_sem);
2129 * watch/notify callback event infrastructure
2131 * These callbacks are used both for watch and notify operations.
2133 static void __release_event(struct kref *kref)
2135 struct ceph_osd_event *event =
2136 container_of(kref, struct ceph_osd_event, kref);
2138 dout("__release_event %p\n", event);
2142 static void get_event(struct ceph_osd_event *event)
2144 kref_get(&event->kref);
2147 void ceph_osdc_put_event(struct ceph_osd_event *event)
2149 kref_put(&event->kref, __release_event);
2151 EXPORT_SYMBOL(ceph_osdc_put_event);
2153 static void __insert_event(struct ceph_osd_client *osdc,
2154 struct ceph_osd_event *new)
2156 struct rb_node **p = &osdc->event_tree.rb_node;
2157 struct rb_node *parent = NULL;
2158 struct ceph_osd_event *event = NULL;
2162 event = rb_entry(parent, struct ceph_osd_event, node);
2163 if (new->cookie < event->cookie)
2165 else if (new->cookie > event->cookie)
2166 p = &(*p)->rb_right;
2171 rb_link_node(&new->node, parent, p);
2172 rb_insert_color(&new->node, &osdc->event_tree);
2175 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2178 struct rb_node **p = &osdc->event_tree.rb_node;
2179 struct rb_node *parent = NULL;
2180 struct ceph_osd_event *event = NULL;
2184 event = rb_entry(parent, struct ceph_osd_event, node);
2185 if (cookie < event->cookie)
2187 else if (cookie > event->cookie)
2188 p = &(*p)->rb_right;
2195 static void __remove_event(struct ceph_osd_event *event)
2197 struct ceph_osd_client *osdc = event->osdc;
2199 if (!RB_EMPTY_NODE(&event->node)) {
2200 dout("__remove_event removed %p\n", event);
2201 rb_erase(&event->node, &osdc->event_tree);
2202 ceph_osdc_put_event(event);
2204 dout("__remove_event didn't remove %p\n", event);
2208 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2209 void (*event_cb)(u64, u64, u8, void *),
2210 void *data, struct ceph_osd_event **pevent)
2212 struct ceph_osd_event *event;
2214 event = kmalloc(sizeof(*event), GFP_NOIO);
2218 dout("create_event %p\n", event);
2219 event->cb = event_cb;
2220 event->one_shot = 0;
2223 INIT_LIST_HEAD(&event->osd_node);
2224 RB_CLEAR_NODE(&event->node);
2225 kref_init(&event->kref); /* one ref for us */
2226 kref_get(&event->kref); /* one ref for the caller */
2228 spin_lock(&osdc->event_lock);
2229 event->cookie = ++osdc->event_count;
2230 __insert_event(osdc, event);
2231 spin_unlock(&osdc->event_lock);
2236 EXPORT_SYMBOL(ceph_osdc_create_event);
2238 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2240 struct ceph_osd_client *osdc = event->osdc;
2242 dout("cancel_event %p\n", event);
2243 spin_lock(&osdc->event_lock);
2244 __remove_event(event);
2245 spin_unlock(&osdc->event_lock);
2246 ceph_osdc_put_event(event); /* caller's */
2248 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2251 static void do_event_work(struct work_struct *work)
2253 struct ceph_osd_event_work *event_work =
2254 container_of(work, struct ceph_osd_event_work, work);
2255 struct ceph_osd_event *event = event_work->event;
2256 u64 ver = event_work->ver;
2257 u64 notify_id = event_work->notify_id;
2258 u8 opcode = event_work->opcode;
2260 dout("do_event_work completing %p\n", event);
2261 event->cb(ver, notify_id, opcode, event->data);
2262 dout("do_event_work completed %p\n", event);
2263 ceph_osdc_put_event(event);
2269 * Process osd watch notifications
2271 static void handle_watch_notify(struct ceph_osd_client *osdc,
2272 struct ceph_msg *msg)
2276 u64 cookie, ver, notify_id;
2278 struct ceph_osd_event *event;
2279 struct ceph_osd_event_work *event_work;
2281 p = msg->front.iov_base;
2282 end = p + msg->front.iov_len;
2284 ceph_decode_8_safe(&p, end, proto_ver, bad);
2285 ceph_decode_8_safe(&p, end, opcode, bad);
2286 ceph_decode_64_safe(&p, end, cookie, bad);
2287 ceph_decode_64_safe(&p, end, ver, bad);
2288 ceph_decode_64_safe(&p, end, notify_id, bad);
2290 spin_lock(&osdc->event_lock);
2291 event = __find_event(osdc, cookie);
2293 BUG_ON(event->one_shot);
2296 spin_unlock(&osdc->event_lock);
2297 dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2298 cookie, ver, event);
2300 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2302 pr_err("couldn't allocate event_work\n");
2303 ceph_osdc_put_event(event);
2306 INIT_WORK(&event_work->work, do_event_work);
2307 event_work->event = event;
2308 event_work->ver = ver;
2309 event_work->notify_id = notify_id;
2310 event_work->opcode = opcode;
2312 queue_work(osdc->notify_wq, &event_work->work);
2318 pr_err("osdc handle_watch_notify corrupt msg\n");
2322 * build new request AND message
2325 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2326 struct ceph_snap_context *snapc, u64 snap_id,
2327 struct timespec *mtime)
2329 struct ceph_msg *msg = req->r_request;
2332 int flags = req->r_flags;
2336 req->r_snapid = snap_id;
2337 req->r_snapc = ceph_get_snap_context(snapc);
2339 /* encode request */
2340 msg->hdr.version = cpu_to_le16(4);
2342 p = msg->front.iov_base;
2343 ceph_encode_32(&p, 1); /* client_inc is always 1 */
2344 req->r_request_osdmap_epoch = p;
2346 req->r_request_flags = p;
2348 if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2349 ceph_encode_timespec(p, mtime);
2350 p += sizeof(struct ceph_timespec);
2351 req->r_request_reassert_version = p;
2352 p += sizeof(struct ceph_eversion); /* will get filled in */
2355 ceph_encode_8(&p, 4);
2356 ceph_encode_8(&p, 4);
2357 ceph_encode_32(&p, 8 + 4 + 4);
2358 req->r_request_pool = p;
2360 ceph_encode_32(&p, -1); /* preferred */
2361 ceph_encode_32(&p, 0); /* key len */
2363 ceph_encode_8(&p, 1);
2364 req->r_request_pgid = p;
2366 ceph_encode_32(&p, -1); /* preferred */
2369 ceph_encode_32(&p, req->r_base_oid.name_len);
2370 memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2371 dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2372 req->r_base_oid.name, req->r_base_oid.name_len);
2373 p += req->r_base_oid.name_len;
2375 /* ops--can imply data */
2376 ceph_encode_16(&p, (u16)req->r_num_ops);
2378 for (i = 0; i < req->r_num_ops; i++) {
2379 data_len += osd_req_encode_op(req, p, i);
2380 p += sizeof(struct ceph_osd_op);
2384 ceph_encode_64(&p, req->r_snapid);
2385 ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2386 ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2388 for (i = 0; i < snapc->num_snaps; i++) {
2389 ceph_encode_64(&p, req->r_snapc->snaps[i]);
2393 req->r_request_attempts = p;
2397 if (flags & CEPH_OSD_FLAG_WRITE) {
2401 * The header "data_off" is a hint to the receiver
2402 * allowing it to align received data into its
2403 * buffers such that there's no need to re-copy
2404 * it before writing it to disk (direct I/O).
2406 data_off = (u16) (off & 0xffff);
2407 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2409 req->r_request->hdr.data_len = cpu_to_le32(data_len);
2411 BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2412 msg_size = p - msg->front.iov_base;
2413 msg->front.iov_len = msg_size;
2414 msg->hdr.front_len = cpu_to_le32(msg_size);
2416 dout("build_request msg_size was %d\n", (int)msg_size);
2418 EXPORT_SYMBOL(ceph_osdc_build_request);
2421 * Register request, send initial attempt.
2423 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2424 struct ceph_osd_request *req,
2429 down_read(&osdc->map_sem);
2430 mutex_lock(&osdc->request_mutex);
2432 rc = __ceph_osdc_start_request(osdc, req, nofail);
2434 mutex_unlock(&osdc->request_mutex);
2435 up_read(&osdc->map_sem);
2439 EXPORT_SYMBOL(ceph_osdc_start_request);
2442 * Unregister a registered request. The request is not completed (i.e.
2443 * no callbacks or wakeups) - higher layers are supposed to know what
2444 * they are canceling.
2446 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2448 struct ceph_osd_client *osdc = req->r_osdc;
2450 mutex_lock(&osdc->request_mutex);
2452 __unregister_linger_request(osdc, req);
2453 __unregister_request(osdc, req);
2454 mutex_unlock(&osdc->request_mutex);
2456 dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2458 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2461 * wait for a request to complete
2463 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2464 struct ceph_osd_request *req)
2468 dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2470 rc = wait_for_completion_interruptible(&req->r_completion);
2472 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2473 ceph_osdc_cancel_request(req);
2474 complete_request(req);
2478 dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2480 return req->r_result;
2482 EXPORT_SYMBOL(ceph_osdc_wait_request);
2485 * sync - wait for all in-flight requests to flush. avoid starvation.
2487 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2489 struct ceph_osd_request *req;
2490 u64 last_tid, next_tid = 0;
2492 mutex_lock(&osdc->request_mutex);
2493 last_tid = osdc->last_tid;
2495 req = __lookup_request_ge(osdc, next_tid);
2498 if (req->r_tid > last_tid)
2501 next_tid = req->r_tid + 1;
2502 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2505 ceph_osdc_get_request(req);
2506 mutex_unlock(&osdc->request_mutex);
2507 dout("sync waiting on tid %llu (last is %llu)\n",
2508 req->r_tid, last_tid);
2509 wait_for_completion(&req->r_safe_completion);
2510 mutex_lock(&osdc->request_mutex);
2511 ceph_osdc_put_request(req);
2513 mutex_unlock(&osdc->request_mutex);
2514 dout("sync done (thru tid %llu)\n", last_tid);
2516 EXPORT_SYMBOL(ceph_osdc_sync);
2519 * Call all pending notify callbacks - for use after a watch is
2520 * unregistered, to make sure no more callbacks for it will be invoked
2522 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2524 flush_workqueue(osdc->notify_wq);
2526 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2532 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2537 osdc->client = client;
2538 osdc->osdmap = NULL;
2539 init_rwsem(&osdc->map_sem);
2540 init_completion(&osdc->map_waiters);
2541 osdc->last_requested_map = 0;
2542 mutex_init(&osdc->request_mutex);
2544 osdc->osds = RB_ROOT;
2545 INIT_LIST_HEAD(&osdc->osd_lru);
2546 osdc->requests = RB_ROOT;
2547 INIT_LIST_HEAD(&osdc->req_lru);
2548 INIT_LIST_HEAD(&osdc->req_unsent);
2549 INIT_LIST_HEAD(&osdc->req_notarget);
2550 INIT_LIST_HEAD(&osdc->req_linger);
2551 osdc->num_requests = 0;
2552 INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2553 INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2554 spin_lock_init(&osdc->event_lock);
2555 osdc->event_tree = RB_ROOT;
2556 osdc->event_count = 0;
2558 schedule_delayed_work(&osdc->osds_timeout_work,
2559 round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2562 osdc->req_mempool = mempool_create_kmalloc_pool(10,
2563 sizeof(struct ceph_osd_request));
2564 if (!osdc->req_mempool)
2567 err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2568 OSD_OP_FRONT_LEN, 10, true,
2572 err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2573 OSD_OPREPLY_FRONT_LEN, 10, true,
2579 osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2580 if (!osdc->notify_wq)
2581 goto out_msgpool_reply;
2586 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2588 ceph_msgpool_destroy(&osdc->msgpool_op);
2590 mempool_destroy(osdc->req_mempool);
2595 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2597 flush_workqueue(osdc->notify_wq);
2598 destroy_workqueue(osdc->notify_wq);
2599 cancel_delayed_work_sync(&osdc->timeout_work);
2600 cancel_delayed_work_sync(&osdc->osds_timeout_work);
2602 ceph_osdmap_destroy(osdc->osdmap);
2603 osdc->osdmap = NULL;
2605 remove_all_osds(osdc);
2606 mempool_destroy(osdc->req_mempool);
2607 ceph_msgpool_destroy(&osdc->msgpool_op);
2608 ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2612 * Read some contiguous pages. If we cross a stripe boundary, shorten
2613 * *plen. Return number of bytes read, or error.
2615 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2616 struct ceph_vino vino, struct ceph_file_layout *layout,
2618 u32 truncate_seq, u64 truncate_size,
2619 struct page **pages, int num_pages, int page_align)
2621 struct ceph_osd_request *req;
2624 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2625 vino.snap, off, *plen);
2626 req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2627 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2628 NULL, truncate_seq, truncate_size,
2631 return PTR_ERR(req);
2633 /* it may be a short read due to an object boundary */
2635 osd_req_op_extent_osd_data_pages(req, 0,
2636 pages, *plen, page_align, false, false);
2638 dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n",
2639 off, *plen, *plen, page_align);
2641 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2643 rc = ceph_osdc_start_request(osdc, req, false);
2645 rc = ceph_osdc_wait_request(osdc, req);
2647 ceph_osdc_put_request(req);
2648 dout("readpages result %d\n", rc);
2651 EXPORT_SYMBOL(ceph_osdc_readpages);
2654 * do a synchronous write on N pages
2656 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2657 struct ceph_file_layout *layout,
2658 struct ceph_snap_context *snapc,
2660 u32 truncate_seq, u64 truncate_size,
2661 struct timespec *mtime,
2662 struct page **pages, int num_pages)
2664 struct ceph_osd_request *req;
2666 int page_align = off & ~PAGE_MASK;
2668 BUG_ON(vino.snap != CEPH_NOSNAP); /* snapshots aren't writeable */
2669 req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2671 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2672 snapc, truncate_seq, truncate_size,
2675 return PTR_ERR(req);
2677 /* it may be a short write due to an object boundary */
2678 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2680 dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2682 ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2684 rc = ceph_osdc_start_request(osdc, req, true);
2686 rc = ceph_osdc_wait_request(osdc, req);
2688 ceph_osdc_put_request(req);
2691 dout("writepages result %d\n", rc);
2694 EXPORT_SYMBOL(ceph_osdc_writepages);
2696 int ceph_osdc_setup(void)
2698 BUG_ON(ceph_osd_request_cache);
2699 ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2700 sizeof (struct ceph_osd_request),
2701 __alignof__(struct ceph_osd_request),
2704 return ceph_osd_request_cache ? 0 : -ENOMEM;
2706 EXPORT_SYMBOL(ceph_osdc_setup);
2708 void ceph_osdc_cleanup(void)
2710 BUG_ON(!ceph_osd_request_cache);
2711 kmem_cache_destroy(ceph_osd_request_cache);
2712 ceph_osd_request_cache = NULL;
2714 EXPORT_SYMBOL(ceph_osdc_cleanup);
2717 * handle incoming message
2719 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2721 struct ceph_osd *osd = con->private;
2722 struct ceph_osd_client *osdc;
2723 int type = le16_to_cpu(msg->hdr.type);
2730 case CEPH_MSG_OSD_MAP:
2731 ceph_osdc_handle_map(osdc, msg);
2733 case CEPH_MSG_OSD_OPREPLY:
2734 handle_reply(osdc, msg, con);
2736 case CEPH_MSG_WATCH_NOTIFY:
2737 handle_watch_notify(osdc, msg);
2741 pr_err("received unknown message type %d %s\n", type,
2742 ceph_msg_type_name(type));
2749 * lookup and return message for incoming reply. set up reply message
2752 static struct ceph_msg *get_reply(struct ceph_connection *con,
2753 struct ceph_msg_header *hdr,
2756 struct ceph_osd *osd = con->private;
2757 struct ceph_osd_client *osdc = osd->o_osdc;
2759 struct ceph_osd_request *req;
2760 int front_len = le32_to_cpu(hdr->front_len);
2761 int data_len = le32_to_cpu(hdr->data_len);
2764 tid = le64_to_cpu(hdr->tid);
2765 mutex_lock(&osdc->request_mutex);
2766 req = __lookup_request(osdc, tid);
2770 dout("get_reply unknown tid %llu from osd%d\n", tid,
2775 if (req->r_reply->con)
2776 dout("%s revoking msg %p from old con %p\n", __func__,
2777 req->r_reply, req->r_reply->con);
2778 ceph_msg_revoke_incoming(req->r_reply);
2780 if (front_len > req->r_reply->front_alloc_len) {
2781 pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2782 front_len, req->r_reply->front_alloc_len,
2783 (unsigned int)con->peer_name.type,
2784 le64_to_cpu(con->peer_name.num));
2785 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2789 ceph_msg_put(req->r_reply);
2792 m = ceph_msg_get(req->r_reply);
2795 struct ceph_osd_data *osd_data;
2798 * XXX This is assuming there is only one op containing
2799 * XXX page data. Probably OK for reads, but this
2800 * XXX ought to be done more generally.
2802 osd_data = osd_req_op_extent_osd_data(req, 0);
2803 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2804 if (osd_data->pages &&
2805 unlikely(osd_data->length < data_len)) {
2807 pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2808 tid, data_len, osd_data->length);
2817 dout("get_reply tid %lld %p\n", tid, m);
2820 mutex_unlock(&osdc->request_mutex);
2825 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2826 struct ceph_msg_header *hdr,
2829 struct ceph_osd *osd = con->private;
2830 int type = le16_to_cpu(hdr->type);
2831 int front = le32_to_cpu(hdr->front_len);
2835 case CEPH_MSG_OSD_MAP:
2836 case CEPH_MSG_WATCH_NOTIFY:
2837 return ceph_msg_new(type, front, GFP_NOFS, false);
2838 case CEPH_MSG_OSD_OPREPLY:
2839 return get_reply(con, hdr, skip);
2841 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2849 * Wrappers to refcount containing ceph_osd struct
2851 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2853 struct ceph_osd *osd = con->private;
2859 static void put_osd_con(struct ceph_connection *con)
2861 struct ceph_osd *osd = con->private;
2869 * Note: returned pointer is the address of a structure that's
2870 * managed separately. Caller must *not* attempt to free it.
2872 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2873 int *proto, int force_new)
2875 struct ceph_osd *o = con->private;
2876 struct ceph_osd_client *osdc = o->o_osdc;
2877 struct ceph_auth_client *ac = osdc->client->monc.auth;
2878 struct ceph_auth_handshake *auth = &o->o_auth;
2880 if (force_new && auth->authorizer) {
2881 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2882 auth->authorizer = NULL;
2884 if (!auth->authorizer) {
2885 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2888 return ERR_PTR(ret);
2890 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2893 return ERR_PTR(ret);
2895 *proto = ac->protocol;
2901 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2903 struct ceph_osd *o = con->private;
2904 struct ceph_osd_client *osdc = o->o_osdc;
2905 struct ceph_auth_client *ac = osdc->client->monc.auth;
2907 return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2910 static int invalidate_authorizer(struct ceph_connection *con)
2912 struct ceph_osd *o = con->private;
2913 struct ceph_osd_client *osdc = o->o_osdc;
2914 struct ceph_auth_client *ac = osdc->client->monc.auth;
2916 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2917 return ceph_monc_validate_auth(&osdc->client->monc);
2920 static const struct ceph_connection_operations osd_con_ops = {
2923 .dispatch = dispatch,
2924 .get_authorizer = get_authorizer,
2925 .verify_authorizer_reply = verify_authorizer_reply,
2926 .invalidate_authorizer = invalidate_authorizer,
2927 .alloc_msg = alloc_msg,