1 #include "ceph_debug.h"
4 #include <linux/highmem.h>
6 #include <linux/pagemap.h>
7 #include <linux/slab.h>
8 #include <linux/uaccess.h>
11 #include "osd_client.h"
12 #include "messenger.h"
16 const static struct ceph_connection_operations osd_con_ops
;
18 static void kick_requests(struct ceph_osd_client
*osdc
, struct ceph_osd
*osd
);
21 * Implement client access to distributed object storage cluster.
23 * All data objects are stored within a cluster/cloud of OSDs, or
24 * "object storage devices." (Note that Ceph OSDs have _nothing_ to
25 * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply
26 * remote daemons serving up and coordinating consistent and safe
29 * Cluster membership and the mapping of data objects onto storage devices
30 * are described by the osd map.
32 * We keep track of pending OSD requests (read, write), resubmit
33 * requests to different OSDs when the cluster topology/data layout
34 * change, or retry the affected requests when the communications
35 * channel with an OSD is reset.
39 * calculate the mapping of a file extent onto an object, and fill out the
40 * request accordingly. shorten extent as necessary if it crosses an
43 * fill osd op in request message.
45 static void calc_layout(struct ceph_osd_client
*osdc
,
46 struct ceph_vino vino
, struct ceph_file_layout
*layout
,
48 struct ceph_osd_request
*req
)
50 struct ceph_osd_request_head
*reqhead
= req
->r_request
->front
.iov_base
;
51 struct ceph_osd_op
*op
= (void *)(reqhead
+ 1);
53 u64 objoff
, objlen
; /* extent in object */
56 reqhead
->snapid
= cpu_to_le64(vino
.snap
);
59 ceph_calc_file_object_mapping(layout
, off
, plen
, &bno
,
62 dout(" skipping last %llu, final file extent %llu~%llu\n",
63 orig_len
- *plen
, off
, *plen
);
65 sprintf(req
->r_oid
, "%llx.%08llx", vino
.ino
, bno
);
66 req
->r_oid_len
= strlen(req
->r_oid
);
68 op
->extent
.offset
= cpu_to_le64(objoff
);
69 op
->extent
.length
= cpu_to_le64(objlen
);
70 req
->r_num_pages
= calc_pages_for(off
, *plen
);
72 dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
73 req
->r_oid
, req
->r_oid_len
, objoff
, objlen
, req
->r_num_pages
);
80 void ceph_osdc_release_request(struct kref
*kref
)
82 struct ceph_osd_request
*req
= container_of(kref
,
83 struct ceph_osd_request
,
87 ceph_msg_put(req
->r_request
);
89 ceph_msg_put(req
->r_reply
);
90 if (req
->r_con_filling_pages
) {
91 dout("release_request revoking pages %p from con %p\n",
92 req
->r_pages
, req
->r_con_filling_pages
);
93 ceph_con_revoke_pages(req
->r_con_filling_pages
,
95 ceph_con_put(req
->r_con_filling_pages
);
98 ceph_release_page_vector(req
->r_pages
,
100 ceph_put_snap_context(req
->r_snapc
);
102 mempool_free(req
, req
->r_osdc
->req_mempool
);
108 * build new request AND message, calculate layout, and adjust file
111 * if the file was recently truncated, we include information about its
112 * old and new size so that the object can be updated appropriately. (we
113 * avoid synchronously deleting truncated objects because it's slow.)
115 * if @do_sync, include a 'startsync' command so that the osd will flush
118 struct ceph_osd_request
*ceph_osdc_new_request(struct ceph_osd_client
*osdc
,
119 struct ceph_file_layout
*layout
,
120 struct ceph_vino vino
,
122 int opcode
, int flags
,
123 struct ceph_snap_context
*snapc
,
127 struct timespec
*mtime
,
128 bool use_mempool
, int num_reply
)
130 struct ceph_osd_request
*req
;
131 struct ceph_msg
*msg
;
132 struct ceph_osd_request_head
*head
;
133 struct ceph_osd_op
*op
;
135 int do_trunc
= truncate_seq
&& (off
+ *plen
> truncate_size
);
136 int num_op
= 1 + do_sync
+ do_trunc
;
137 size_t msg_size
= sizeof(*head
) + num_op
*sizeof(*op
);
142 req
= mempool_alloc(osdc
->req_mempool
, GFP_NOFS
);
143 memset(req
, 0, sizeof(*req
));
145 req
= kzalloc(sizeof(*req
), GFP_NOFS
);
148 return ERR_PTR(-ENOMEM
);
150 err
= ceph_msgpool_resv(&osdc
->msgpool_op_reply
, num_reply
);
152 ceph_osdc_put_request(req
);
153 return ERR_PTR(-ENOMEM
);
155 req
->r_num_prealloc_reply
= num_reply
;
158 req
->r_mempool
= use_mempool
;
159 kref_init(&req
->r_kref
);
160 init_completion(&req
->r_completion
);
161 init_completion(&req
->r_safe_completion
);
162 INIT_LIST_HEAD(&req
->r_unsafe_item
);
163 req
->r_flags
= flags
;
165 WARN_ON((flags
& (CEPH_OSD_FLAG_READ
|CEPH_OSD_FLAG_WRITE
)) == 0);
167 /* create message; allow space for oid */
170 msg_size
+= sizeof(u64
) * snapc
->num_snaps
;
172 msg
= ceph_msgpool_get(&osdc
->msgpool_op
, 0);
174 msg
= ceph_msg_new(CEPH_MSG_OSD_OP
, msg_size
, 0, 0, NULL
);
176 ceph_msgpool_resv(&osdc
->msgpool_op_reply
, -num_reply
);
177 ceph_osdc_put_request(req
);
178 return ERR_PTR(PTR_ERR(msg
));
180 msg
->hdr
.type
= cpu_to_le16(CEPH_MSG_OSD_OP
);
181 memset(msg
->front
.iov_base
, 0, msg
->front
.iov_len
);
182 head
= msg
->front
.iov_base
;
183 op
= (void *)(head
+ 1);
184 p
= (void *)(op
+ num_op
);
186 req
->r_request
= msg
;
187 req
->r_snapc
= ceph_get_snap_context(snapc
);
189 head
->client_inc
= cpu_to_le32(1); /* always, for now. */
190 head
->flags
= cpu_to_le32(flags
);
191 if (flags
& CEPH_OSD_FLAG_WRITE
)
192 ceph_encode_timespec(&head
->mtime
, mtime
);
193 head
->num_ops
= cpu_to_le16(num_op
);
194 op
->op
= cpu_to_le16(opcode
);
196 /* calculate max write size */
197 calc_layout(osdc
, vino
, layout
, off
, plen
, req
);
198 req
->r_file_layout
= *layout
; /* keep a copy */
200 if (flags
& CEPH_OSD_FLAG_WRITE
) {
201 req
->r_request
->hdr
.data_off
= cpu_to_le16(off
);
202 req
->r_request
->hdr
.data_len
= cpu_to_le32(*plen
);
203 op
->payload_len
= cpu_to_le32(*plen
);
207 head
->object_len
= cpu_to_le32(req
->r_oid_len
);
208 memcpy(p
, req
->r_oid
, req
->r_oid_len
);
214 op
->op
= cpu_to_le16(opcode
== CEPH_OSD_OP_READ
?
215 CEPH_OSD_OP_MASKTRUNC
: CEPH_OSD_OP_SETTRUNC
);
216 op
->trunc
.truncate_seq
= cpu_to_le32(truncate_seq
);
217 prevofs
= le64_to_cpu((op
-1)->extent
.offset
);
218 op
->trunc
.truncate_size
= cpu_to_le64(truncate_size
-
223 op
->op
= cpu_to_le16(CEPH_OSD_OP_STARTSYNC
);
226 head
->snap_seq
= cpu_to_le64(snapc
->seq
);
227 head
->num_snaps
= cpu_to_le32(snapc
->num_snaps
);
228 for (i
= 0; i
< snapc
->num_snaps
; i
++) {
229 put_unaligned_le64(snapc
->snaps
[i
], p
);
234 BUG_ON(p
> msg
->front
.iov_base
+ msg
->front
.iov_len
);
239 * We keep osd requests in an rbtree, sorted by ->r_tid.
241 static void __insert_request(struct ceph_osd_client
*osdc
,
242 struct ceph_osd_request
*new)
244 struct rb_node
**p
= &osdc
->requests
.rb_node
;
245 struct rb_node
*parent
= NULL
;
246 struct ceph_osd_request
*req
= NULL
;
250 req
= rb_entry(parent
, struct ceph_osd_request
, r_node
);
251 if (new->r_tid
< req
->r_tid
)
253 else if (new->r_tid
> req
->r_tid
)
259 rb_link_node(&new->r_node
, parent
, p
);
260 rb_insert_color(&new->r_node
, &osdc
->requests
);
263 static struct ceph_osd_request
*__lookup_request(struct ceph_osd_client
*osdc
,
266 struct ceph_osd_request
*req
;
267 struct rb_node
*n
= osdc
->requests
.rb_node
;
270 req
= rb_entry(n
, struct ceph_osd_request
, r_node
);
271 if (tid
< req
->r_tid
)
273 else if (tid
> req
->r_tid
)
281 static struct ceph_osd_request
*
282 __lookup_request_ge(struct ceph_osd_client
*osdc
,
285 struct ceph_osd_request
*req
;
286 struct rb_node
*n
= osdc
->requests
.rb_node
;
289 req
= rb_entry(n
, struct ceph_osd_request
, r_node
);
290 if (tid
< req
->r_tid
) {
294 } else if (tid
> req
->r_tid
) {
305 * If the osd connection drops, we need to resubmit all requests.
307 static void osd_reset(struct ceph_connection
*con
)
309 struct ceph_osd
*osd
= con
->private;
310 struct ceph_osd_client
*osdc
;
314 dout("osd_reset osd%d\n", osd
->o_osd
);
316 osd
->o_incarnation
++;
317 down_read(&osdc
->map_sem
);
318 kick_requests(osdc
, osd
);
319 up_read(&osdc
->map_sem
);
323 * Track open sessions with osds.
325 static struct ceph_osd
*create_osd(struct ceph_osd_client
*osdc
)
327 struct ceph_osd
*osd
;
329 osd
= kzalloc(sizeof(*osd
), GFP_NOFS
);
333 atomic_set(&osd
->o_ref
, 1);
335 INIT_LIST_HEAD(&osd
->o_requests
);
336 osd
->o_incarnation
= 1;
338 ceph_con_init(osdc
->client
->msgr
, &osd
->o_con
);
339 osd
->o_con
.private = osd
;
340 osd
->o_con
.ops
= &osd_con_ops
;
341 osd
->o_con
.peer_name
.type
= CEPH_ENTITY_TYPE_OSD
;
346 static struct ceph_osd
*get_osd(struct ceph_osd
*osd
)
348 if (atomic_inc_not_zero(&osd
->o_ref
)) {
349 dout("get_osd %p %d -> %d\n", osd
, atomic_read(&osd
->o_ref
)-1,
350 atomic_read(&osd
->o_ref
));
353 dout("get_osd %p FAIL\n", osd
);
358 static void put_osd(struct ceph_osd
*osd
)
360 dout("put_osd %p %d -> %d\n", osd
, atomic_read(&osd
->o_ref
),
361 atomic_read(&osd
->o_ref
) - 1);
362 if (atomic_dec_and_test(&osd
->o_ref
))
367 * remove an osd from our map
369 static void remove_osd(struct ceph_osd_client
*osdc
, struct ceph_osd
*osd
)
371 dout("remove_osd %p\n", osd
);
372 BUG_ON(!list_empty(&osd
->o_requests
));
373 rb_erase(&osd
->o_node
, &osdc
->osds
);
374 ceph_con_close(&osd
->o_con
);
381 static int reset_osd(struct ceph_osd_client
*osdc
, struct ceph_osd
*osd
)
385 dout("reset_osd %p osd%d\n", osd
, osd
->o_osd
);
386 if (list_empty(&osd
->o_requests
)) {
387 remove_osd(osdc
, osd
);
389 ceph_con_close(&osd
->o_con
);
390 ceph_con_open(&osd
->o_con
, &osdc
->osdmap
->osd_addr
[osd
->o_osd
]);
391 osd
->o_incarnation
++;
396 static void __insert_osd(struct ceph_osd_client
*osdc
, struct ceph_osd
*new)
398 struct rb_node
**p
= &osdc
->osds
.rb_node
;
399 struct rb_node
*parent
= NULL
;
400 struct ceph_osd
*osd
= NULL
;
404 osd
= rb_entry(parent
, struct ceph_osd
, o_node
);
405 if (new->o_osd
< osd
->o_osd
)
407 else if (new->o_osd
> osd
->o_osd
)
413 rb_link_node(&new->o_node
, parent
, p
);
414 rb_insert_color(&new->o_node
, &osdc
->osds
);
417 static struct ceph_osd
*__lookup_osd(struct ceph_osd_client
*osdc
, int o
)
419 struct ceph_osd
*osd
;
420 struct rb_node
*n
= osdc
->osds
.rb_node
;
423 osd
= rb_entry(n
, struct ceph_osd
, o_node
);
426 else if (o
> osd
->o_osd
)
436 * Register request, assign tid. If this is the first request, set up
439 static void register_request(struct ceph_osd_client
*osdc
,
440 struct ceph_osd_request
*req
)
442 mutex_lock(&osdc
->request_mutex
);
443 req
->r_tid
= ++osdc
->last_tid
;
444 req
->r_request
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
446 dout("register_request %p tid %lld\n", req
, req
->r_tid
);
447 __insert_request(osdc
, req
);
448 ceph_osdc_get_request(req
);
449 osdc
->num_requests
++;
451 req
->r_timeout_stamp
=
452 jiffies
+ osdc
->client
->mount_args
->osd_timeout
*HZ
;
454 if (osdc
->num_requests
== 1) {
455 osdc
->timeout_tid
= req
->r_tid
;
456 dout(" timeout on tid %llu at %lu\n", req
->r_tid
,
457 req
->r_timeout_stamp
);
458 schedule_delayed_work(&osdc
->timeout_work
,
459 round_jiffies_relative(req
->r_timeout_stamp
- jiffies
));
461 mutex_unlock(&osdc
->request_mutex
);
465 * called under osdc->request_mutex
467 static void __unregister_request(struct ceph_osd_client
*osdc
,
468 struct ceph_osd_request
*req
)
470 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
471 rb_erase(&req
->r_node
, &osdc
->requests
);
472 osdc
->num_requests
--;
474 ceph_msgpool_resv(&osdc
->msgpool_op_reply
, -req
->r_num_prealloc_reply
);
477 /* make sure the original request isn't in flight. */
478 ceph_con_revoke(&req
->r_osd
->o_con
, req
->r_request
);
480 list_del_init(&req
->r_osd_item
);
481 if (list_empty(&req
->r_osd
->o_requests
))
482 remove_osd(osdc
, req
->r_osd
);
486 ceph_osdc_put_request(req
);
488 if (req
->r_tid
== osdc
->timeout_tid
) {
489 if (osdc
->num_requests
== 0) {
490 dout("no requests, canceling timeout\n");
491 osdc
->timeout_tid
= 0;
492 cancel_delayed_work(&osdc
->timeout_work
);
494 req
= rb_entry(rb_first(&osdc
->requests
),
495 struct ceph_osd_request
, r_node
);
496 osdc
->timeout_tid
= req
->r_tid
;
497 dout("rescheduled timeout on tid %llu at %lu\n",
498 req
->r_tid
, req
->r_timeout_stamp
);
499 schedule_delayed_work(&osdc
->timeout_work
,
500 round_jiffies_relative(req
->r_timeout_stamp
-
507 * Cancel a previously queued request message
509 static void __cancel_request(struct ceph_osd_request
*req
)
512 ceph_con_revoke(&req
->r_osd
->o_con
, req
->r_request
);
518 * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
519 * (as needed), and set the request r_osd appropriately. If there is
520 * no up osd, set r_osd to NULL.
522 * Return 0 if unchanged, 1 if changed, or negative on error.
524 * Caller should hold map_sem for read and request_mutex.
526 static int __map_osds(struct ceph_osd_client
*osdc
,
527 struct ceph_osd_request
*req
)
529 struct ceph_osd_request_head
*reqhead
= req
->r_request
->front
.iov_base
;
533 struct ceph_osd
*newosd
= NULL
;
535 dout("map_osds %p tid %lld\n", req
, req
->r_tid
);
536 err
= ceph_calc_object_layout(&reqhead
->layout
, req
->r_oid
,
537 &req
->r_file_layout
, osdc
->osdmap
);
540 pgid
= reqhead
->layout
.ol_pgid
;
541 o
= ceph_calc_pg_primary(osdc
->osdmap
, pgid
);
543 if ((req
->r_osd
&& req
->r_osd
->o_osd
== o
&&
544 req
->r_sent
>= req
->r_osd
->o_incarnation
) ||
545 (req
->r_osd
== NULL
&& o
== -1))
546 return 0; /* no change */
548 dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
549 req
->r_tid
, le32_to_cpu(pgid
.pool
), le16_to_cpu(pgid
.ps
), o
,
550 req
->r_osd
? req
->r_osd
->o_osd
: -1);
553 __cancel_request(req
);
554 list_del_init(&req
->r_osd_item
);
555 if (list_empty(&req
->r_osd
->o_requests
)) {
556 /* try to re-use r_osd if possible */
557 newosd
= get_osd(req
->r_osd
);
558 remove_osd(osdc
, newosd
);
563 req
->r_osd
= __lookup_osd(osdc
, o
);
564 if (!req
->r_osd
&& o
>= 0) {
570 req
->r_osd
= create_osd(osdc
);
575 dout("map_osds osd %p is osd%d\n", req
->r_osd
, o
);
576 req
->r_osd
->o_osd
= o
;
577 req
->r_osd
->o_con
.peer_name
.num
= cpu_to_le64(o
);
578 __insert_osd(osdc
, req
->r_osd
);
580 ceph_con_open(&req
->r_osd
->o_con
, &osdc
->osdmap
->osd_addr
[o
]);
584 list_add(&req
->r_osd_item
, &req
->r_osd
->o_requests
);
585 err
= 1; /* osd changed */
594 * caller should hold map_sem (for read) and request_mutex
596 static int __send_request(struct ceph_osd_client
*osdc
,
597 struct ceph_osd_request
*req
)
599 struct ceph_osd_request_head
*reqhead
;
602 err
= __map_osds(osdc
, req
);
605 if (req
->r_osd
== NULL
) {
606 dout("send_request %p no up osds in pg\n", req
);
607 ceph_monc_request_next_osdmap(&osdc
->client
->monc
);
611 dout("send_request %p tid %llu to osd%d flags %d\n",
612 req
, req
->r_tid
, req
->r_osd
->o_osd
, req
->r_flags
);
614 reqhead
= req
->r_request
->front
.iov_base
;
615 reqhead
->osdmap_epoch
= cpu_to_le32(osdc
->osdmap
->epoch
);
616 reqhead
->flags
|= cpu_to_le32(req
->r_flags
); /* e.g., RETRY */
617 reqhead
->reassert_version
= req
->r_reassert_version
;
619 req
->r_timeout_stamp
= jiffies
+osdc
->client
->mount_args
->osd_timeout
*HZ
;
621 ceph_msg_get(req
->r_request
); /* send consumes a ref */
622 ceph_con_send(&req
->r_osd
->o_con
, req
->r_request
);
623 req
->r_sent
= req
->r_osd
->o_incarnation
;
628 * Timeout callback, called every N seconds when 1 or more osd
629 * requests has been active for more than N seconds. When this
630 * happens, we ping all OSDs with requests who have timed out to
631 * ensure any communications channel reset is detected. Reset the
632 * request timeouts another N seconds in the future as we go.
633 * Reschedule the timeout event another N seconds in future (unless
634 * there are no open requests).
636 static void handle_timeout(struct work_struct
*work
)
638 struct ceph_osd_client
*osdc
=
639 container_of(work
, struct ceph_osd_client
, timeout_work
.work
);
640 struct ceph_osd_request
*req
;
641 struct ceph_osd
*osd
;
642 unsigned long timeout
= osdc
->client
->mount_args
->osd_timeout
* HZ
;
643 unsigned long next_timeout
= timeout
+ jiffies
;
647 down_read(&osdc
->map_sem
);
649 ceph_monc_request_next_osdmap(&osdc
->client
->monc
);
651 mutex_lock(&osdc
->request_mutex
);
652 for (p
= rb_first(&osdc
->requests
); p
; p
= rb_next(p
)) {
653 req
= rb_entry(p
, struct ceph_osd_request
, r_node
);
658 dout("osdc resending prev failed %lld\n", req
->r_tid
);
659 err
= __send_request(osdc
, req
);
661 dout("osdc failed again on %lld\n", req
->r_tid
);
663 req
->r_resend
= false;
667 for (p
= rb_first(&osdc
->osds
); p
; p
= rb_next(p
)) {
668 osd
= rb_entry(p
, struct ceph_osd
, o_node
);
669 if (list_empty(&osd
->o_requests
))
671 req
= list_first_entry(&osd
->o_requests
,
672 struct ceph_osd_request
, r_osd_item
);
673 if (time_before(jiffies
, req
->r_timeout_stamp
))
676 dout(" tid %llu (at least) timed out on osd%d\n",
677 req
->r_tid
, osd
->o_osd
);
678 req
->r_timeout_stamp
= next_timeout
;
679 ceph_con_keepalive(&osd
->o_con
);
682 if (osdc
->timeout_tid
)
683 schedule_delayed_work(&osdc
->timeout_work
,
684 round_jiffies_relative(timeout
));
686 mutex_unlock(&osdc
->request_mutex
);
688 up_read(&osdc
->map_sem
);
692 * handle osd op reply. either call the callback if it is specified,
693 * or do the completion to wake up the waiting thread.
695 static void handle_reply(struct ceph_osd_client
*osdc
, struct ceph_msg
*msg
,
696 struct ceph_connection
*con
)
698 struct ceph_osd_reply_head
*rhead
= msg
->front
.iov_base
;
699 struct ceph_osd_request
*req
;
701 int numops
, object_len
, flags
;
703 tid
= le64_to_cpu(msg
->hdr
.tid
);
704 if (msg
->front
.iov_len
< sizeof(*rhead
))
706 numops
= le32_to_cpu(rhead
->num_ops
);
707 object_len
= le32_to_cpu(rhead
->object_len
);
708 if (msg
->front
.iov_len
!= sizeof(*rhead
) + object_len
+
709 numops
* sizeof(struct ceph_osd_op
))
711 dout("handle_reply %p tid %llu\n", msg
, tid
);
714 mutex_lock(&osdc
->request_mutex
);
715 req
= __lookup_request(osdc
, tid
);
717 dout("handle_reply tid %llu dne\n", tid
);
718 mutex_unlock(&osdc
->request_mutex
);
721 ceph_osdc_get_request(req
);
722 flags
= le32_to_cpu(rhead
->flags
);
725 * if this connection filled our pages, drop our reference now, to
726 * avoid a (safe but slower) revoke later.
728 if (req
->r_con_filling_pages
== con
&& req
->r_pages
== msg
->pages
) {
729 dout(" got pages, dropping con_filling_pages ref %p\n", con
);
730 req
->r_con_filling_pages
= NULL
;
736 * once we see the message has been received, we don't
737 * need a ref (which is only needed for revoking
740 ceph_msg_put(req
->r_reply
);
744 if (!req
->r_got_reply
) {
747 req
->r_result
= le32_to_cpu(rhead
->result
);
748 bytes
= le32_to_cpu(msg
->hdr
.data_len
);
749 dout("handle_reply result %d bytes %d\n", req
->r_result
,
751 if (req
->r_result
== 0)
752 req
->r_result
= bytes
;
754 /* in case this is a write and we need to replay, */
755 req
->r_reassert_version
= rhead
->reassert_version
;
757 req
->r_got_reply
= 1;
758 } else if ((flags
& CEPH_OSD_FLAG_ONDISK
) == 0) {
759 dout("handle_reply tid %llu dup ack\n", tid
);
760 mutex_unlock(&osdc
->request_mutex
);
764 dout("handle_reply tid %llu flags %d\n", tid
, flags
);
766 /* either this is a read, or we got the safe response */
767 if ((flags
& CEPH_OSD_FLAG_ONDISK
) ||
768 ((flags
& CEPH_OSD_FLAG_WRITE
) == 0))
769 __unregister_request(osdc
, req
);
771 mutex_unlock(&osdc
->request_mutex
);
774 req
->r_callback(req
, msg
);
776 complete(&req
->r_completion
);
778 if (flags
& CEPH_OSD_FLAG_ONDISK
) {
779 if (req
->r_safe_callback
)
780 req
->r_safe_callback(req
, msg
);
781 complete(&req
->r_safe_completion
); /* fsync waiter */
785 ceph_osdc_put_request(req
);
789 pr_err("corrupt osd_op_reply got %d %d expected %d\n",
790 (int)msg
->front
.iov_len
, le32_to_cpu(msg
->hdr
.front_len
),
791 (int)sizeof(*rhead
));
797 * Resubmit osd requests whose osd or osd address has changed. Request
798 * a new osd map if osds are down, or we are otherwise unable to determine
799 * how to direct a request.
801 * Close connections to down osds.
803 * If @who is specified, resubmit requests for that specific osd.
805 * Caller should hold map_sem for read and request_mutex.
807 static void kick_requests(struct ceph_osd_client
*osdc
,
808 struct ceph_osd
*kickosd
)
810 struct ceph_osd_request
*req
;
811 struct rb_node
*p
, *n
;
815 dout("kick_requests osd%d\n", kickosd
? kickosd
->o_osd
: -1);
816 mutex_lock(&osdc
->request_mutex
);
818 for (p
= rb_first(&osdc
->osds
); p
; p
= n
) {
819 struct ceph_osd
*osd
=
820 rb_entry(p
, struct ceph_osd
, o_node
);
823 if (!ceph_osd_is_up(osdc
->osdmap
, osd
->o_osd
) ||
824 !ceph_entity_addr_equal(&osd
->o_con
.peer_addr
,
825 ceph_osd_addr(osdc
->osdmap
,
827 reset_osd(osdc
, osd
);
831 for (p
= rb_first(&osdc
->requests
); p
; p
= rb_next(p
)) {
832 req
= rb_entry(p
, struct ceph_osd_request
, r_node
);
835 dout(" r_resend set on tid %llu\n", req
->r_tid
);
836 __cancel_request(req
);
839 if (req
->r_osd
&& kickosd
== req
->r_osd
) {
840 __cancel_request(req
);
844 err
= __map_osds(osdc
, req
);
846 continue; /* no change */
849 * FIXME: really, we should set the request
850 * error and fail if this isn't a 'nofail'
851 * request, but that's a fair bit more
852 * complicated to do. So retry!
854 dout(" setting r_resend on %llu\n", req
->r_tid
);
855 req
->r_resend
= true;
858 if (req
->r_osd
== NULL
) {
859 dout("tid %llu maps to no valid osd\n", req
->r_tid
);
860 needmap
++; /* request a newer map */
865 dout("kicking %p tid %llu osd%d\n", req
, req
->r_tid
,
867 req
->r_flags
|= CEPH_OSD_FLAG_RETRY
;
868 err
= __send_request(osdc
, req
);
870 dout(" setting r_resend on %llu\n", req
->r_tid
);
871 req
->r_resend
= true;
874 mutex_unlock(&osdc
->request_mutex
);
877 dout("%d requests for down osds, need new map\n", needmap
);
878 ceph_monc_request_next_osdmap(&osdc
->client
->monc
);
883 * Process updated osd map.
885 * The message contains any number of incremental and full maps, normally
886 * indicating some sort of topology change in the cluster. Kick requests
887 * off to different OSDs as needed.
889 void ceph_osdc_handle_map(struct ceph_osd_client
*osdc
, struct ceph_msg
*msg
)
891 void *p
, *end
, *next
;
894 struct ceph_osdmap
*newmap
= NULL
, *oldmap
;
896 struct ceph_fsid fsid
;
898 dout("handle_map have %u\n", osdc
->osdmap
? osdc
->osdmap
->epoch
: 0);
899 p
= msg
->front
.iov_base
;
900 end
= p
+ msg
->front
.iov_len
;
903 ceph_decode_need(&p
, end
, sizeof(fsid
), bad
);
904 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
905 if (ceph_check_fsid(osdc
->client
, &fsid
) < 0)
908 down_write(&osdc
->map_sem
);
910 /* incremental maps */
911 ceph_decode_32_safe(&p
, end
, nr_maps
, bad
);
912 dout(" %d inc maps\n", nr_maps
);
913 while (nr_maps
> 0) {
914 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
915 epoch
= ceph_decode_32(&p
);
916 maplen
= ceph_decode_32(&p
);
917 ceph_decode_need(&p
, end
, maplen
, bad
);
919 if (osdc
->osdmap
&& osdc
->osdmap
->epoch
+1 == epoch
) {
920 dout("applying incremental map %u len %d\n",
922 newmap
= osdmap_apply_incremental(&p
, next
,
925 if (IS_ERR(newmap
)) {
926 err
= PTR_ERR(newmap
);
930 if (newmap
!= osdc
->osdmap
) {
931 ceph_osdmap_destroy(osdc
->osdmap
);
932 osdc
->osdmap
= newmap
;
935 dout("ignoring incremental map %u len %d\n",
945 ceph_decode_32_safe(&p
, end
, nr_maps
, bad
);
946 dout(" %d full maps\n", nr_maps
);
948 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
949 epoch
= ceph_decode_32(&p
);
950 maplen
= ceph_decode_32(&p
);
951 ceph_decode_need(&p
, end
, maplen
, bad
);
953 dout("skipping non-latest full map %u len %d\n",
955 } else if (osdc
->osdmap
&& osdc
->osdmap
->epoch
>= epoch
) {
956 dout("skipping full map %u len %d, "
957 "older than our %u\n", epoch
, maplen
,
958 osdc
->osdmap
->epoch
);
960 dout("taking full map %u len %d\n", epoch
, maplen
);
961 newmap
= osdmap_decode(&p
, p
+maplen
);
962 if (IS_ERR(newmap
)) {
963 err
= PTR_ERR(newmap
);
967 oldmap
= osdc
->osdmap
;
968 osdc
->osdmap
= newmap
;
970 ceph_osdmap_destroy(oldmap
);
977 downgrade_write(&osdc
->map_sem
);
978 ceph_monc_got_osdmap(&osdc
->client
->monc
, osdc
->osdmap
->epoch
);
980 kick_requests(osdc
, NULL
);
981 up_read(&osdc
->map_sem
);
985 pr_err("osdc handle_map corrupt msg\n");
987 up_write(&osdc
->map_sem
);
993 * A read request prepares specific pages that data is to be read into.
994 * When a message is being read off the wire, we call prepare_pages to
996 * 0 = success, -1 failure.
998 static int prepare_pages(struct ceph_connection
*con
, struct ceph_msg
*m
,
1001 struct ceph_osd
*osd
= con
->private;
1002 struct ceph_osd_client
*osdc
;
1003 struct ceph_osd_request
*req
;
1006 int type
= le16_to_cpu(m
->hdr
.type
);
1012 dout("prepare_pages on msg %p want %d\n", m
, want
);
1013 if (unlikely(type
!= CEPH_MSG_OSD_OPREPLY
))
1014 return -1; /* hmm! */
1016 tid
= le64_to_cpu(m
->hdr
.tid
);
1017 mutex_lock(&osdc
->request_mutex
);
1018 req
= __lookup_request(osdc
, tid
);
1020 dout("prepare_pages unknown tid %llu\n", tid
);
1023 dout("prepare_pages tid %llu has %d pages, want %d\n",
1024 tid
, req
->r_num_pages
, want
);
1025 if (unlikely(req
->r_num_pages
< want
))
1028 if (req
->r_con_filling_pages
) {
1029 dout("revoking pages %p from old con %p\n", req
->r_pages
,
1030 req
->r_con_filling_pages
);
1031 ceph_con_revoke_pages(req
->r_con_filling_pages
, req
->r_pages
);
1032 ceph_con_put(req
->r_con_filling_pages
);
1034 req
->r_con_filling_pages
= ceph_con_get(con
);
1035 req
->r_reply
= ceph_msg_get(m
); /* for duration of read over socket */
1036 m
->pages
= req
->r_pages
;
1037 m
->nr_pages
= req
->r_num_pages
;
1038 ret
= 0; /* success */
1040 mutex_unlock(&osdc
->request_mutex
);
1045 * Register request, send initial attempt.
1047 int ceph_osdc_start_request(struct ceph_osd_client
*osdc
,
1048 struct ceph_osd_request
*req
,
1053 req
->r_request
->pages
= req
->r_pages
;
1054 req
->r_request
->nr_pages
= req
->r_num_pages
;
1056 register_request(osdc
, req
);
1058 down_read(&osdc
->map_sem
);
1059 mutex_lock(&osdc
->request_mutex
);
1061 * a racing kick_requests() may have sent the message for us
1062 * while we dropped request_mutex above, so only send now if
1063 * the request still han't been touched yet.
1065 if (req
->r_sent
== 0) {
1066 rc
= __send_request(osdc
, req
);
1069 dout("osdc_start_request failed send, "
1070 " marking %lld\n", req
->r_tid
);
1071 req
->r_resend
= true;
1074 __unregister_request(osdc
, req
);
1078 mutex_unlock(&osdc
->request_mutex
);
1079 up_read(&osdc
->map_sem
);
1084 * wait for a request to complete
1086 int ceph_osdc_wait_request(struct ceph_osd_client
*osdc
,
1087 struct ceph_osd_request
*req
)
1091 rc
= wait_for_completion_interruptible(&req
->r_completion
);
1093 mutex_lock(&osdc
->request_mutex
);
1094 __cancel_request(req
);
1095 __unregister_request(osdc
, req
);
1096 mutex_unlock(&osdc
->request_mutex
);
1097 dout("wait_request tid %llu canceled/timed out\n", req
->r_tid
);
1101 dout("wait_request tid %llu result %d\n", req
->r_tid
, req
->r_result
);
1102 return req
->r_result
;
1106 * sync - wait for all in-flight requests to flush. avoid starvation.
1108 void ceph_osdc_sync(struct ceph_osd_client
*osdc
)
1110 struct ceph_osd_request
*req
;
1111 u64 last_tid
, next_tid
= 0;
1113 mutex_lock(&osdc
->request_mutex
);
1114 last_tid
= osdc
->last_tid
;
1116 req
= __lookup_request_ge(osdc
, next_tid
);
1119 if (req
->r_tid
> last_tid
)
1122 next_tid
= req
->r_tid
+ 1;
1123 if ((req
->r_flags
& CEPH_OSD_FLAG_WRITE
) == 0)
1126 ceph_osdc_get_request(req
);
1127 mutex_unlock(&osdc
->request_mutex
);
1128 dout("sync waiting on tid %llu (last is %llu)\n",
1129 req
->r_tid
, last_tid
);
1130 wait_for_completion(&req
->r_safe_completion
);
1131 mutex_lock(&osdc
->request_mutex
);
1132 ceph_osdc_put_request(req
);
1134 mutex_unlock(&osdc
->request_mutex
);
1135 dout("sync done (thru tid %llu)\n", last_tid
);
1141 int ceph_osdc_init(struct ceph_osd_client
*osdc
, struct ceph_client
*client
)
1146 osdc
->client
= client
;
1147 osdc
->osdmap
= NULL
;
1148 init_rwsem(&osdc
->map_sem
);
1149 init_completion(&osdc
->map_waiters
);
1150 osdc
->last_requested_map
= 0;
1151 mutex_init(&osdc
->request_mutex
);
1152 osdc
->timeout_tid
= 0;
1154 osdc
->osds
= RB_ROOT
;
1155 osdc
->requests
= RB_ROOT
;
1156 osdc
->num_requests
= 0;
1157 INIT_DELAYED_WORK(&osdc
->timeout_work
, handle_timeout
);
1160 osdc
->req_mempool
= mempool_create_kmalloc_pool(10,
1161 sizeof(struct ceph_osd_request
));
1162 if (!osdc
->req_mempool
)
1165 err
= ceph_msgpool_init(&osdc
->msgpool_op
, 4096, 10, true);
1168 err
= ceph_msgpool_init(&osdc
->msgpool_op_reply
, 512, 0, false);
1174 ceph_msgpool_destroy(&osdc
->msgpool_op
);
1176 mempool_destroy(osdc
->req_mempool
);
1181 void ceph_osdc_stop(struct ceph_osd_client
*osdc
)
1183 cancel_delayed_work_sync(&osdc
->timeout_work
);
1185 ceph_osdmap_destroy(osdc
->osdmap
);
1186 osdc
->osdmap
= NULL
;
1188 mempool_destroy(osdc
->req_mempool
);
1189 ceph_msgpool_destroy(&osdc
->msgpool_op
);
1190 ceph_msgpool_destroy(&osdc
->msgpool_op_reply
);
1194 * Read some contiguous pages. If we cross a stripe boundary, shorten
1195 * *plen. Return number of bytes read, or error.
1197 int ceph_osdc_readpages(struct ceph_osd_client
*osdc
,
1198 struct ceph_vino vino
, struct ceph_file_layout
*layout
,
1200 u32 truncate_seq
, u64 truncate_size
,
1201 struct page
**pages
, int num_pages
)
1203 struct ceph_osd_request
*req
;
1206 dout("readpages on ino %llx.%llx on %llu~%llu\n", vino
.ino
,
1207 vino
.snap
, off
, *plen
);
1208 req
= ceph_osdc_new_request(osdc
, layout
, vino
, off
, plen
,
1209 CEPH_OSD_OP_READ
, CEPH_OSD_FLAG_READ
,
1210 NULL
, 0, truncate_seq
, truncate_size
, NULL
,
1213 return PTR_ERR(req
);
1215 /* it may be a short read due to an object boundary */
1216 req
->r_pages
= pages
;
1217 num_pages
= calc_pages_for(off
, *plen
);
1218 req
->r_num_pages
= num_pages
;
1220 dout("readpages final extent is %llu~%llu (%d pages)\n",
1221 off
, *plen
, req
->r_num_pages
);
1223 rc
= ceph_osdc_start_request(osdc
, req
, false);
1225 rc
= ceph_osdc_wait_request(osdc
, req
);
1227 ceph_osdc_put_request(req
);
1228 dout("readpages result %d\n", rc
);
1233 * do a synchronous write on N pages
1235 int ceph_osdc_writepages(struct ceph_osd_client
*osdc
, struct ceph_vino vino
,
1236 struct ceph_file_layout
*layout
,
1237 struct ceph_snap_context
*snapc
,
1239 u32 truncate_seq
, u64 truncate_size
,
1240 struct timespec
*mtime
,
1241 struct page
**pages
, int num_pages
,
1242 int flags
, int do_sync
, bool nofail
)
1244 struct ceph_osd_request
*req
;
1247 BUG_ON(vino
.snap
!= CEPH_NOSNAP
);
1248 req
= ceph_osdc_new_request(osdc
, layout
, vino
, off
, &len
,
1250 flags
| CEPH_OSD_FLAG_ONDISK
|
1251 CEPH_OSD_FLAG_WRITE
,
1253 truncate_seq
, truncate_size
, mtime
,
1256 return PTR_ERR(req
);
1258 /* it may be a short write due to an object boundary */
1259 req
->r_pages
= pages
;
1260 req
->r_num_pages
= calc_pages_for(off
, len
);
1261 dout("writepages %llu~%llu (%d pages)\n", off
, len
,
1264 rc
= ceph_osdc_start_request(osdc
, req
, nofail
);
1266 rc
= ceph_osdc_wait_request(osdc
, req
);
1268 ceph_osdc_put_request(req
);
1271 dout("writepages result %d\n", rc
);
1276 * handle incoming message
1278 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
1280 struct ceph_osd
*osd
= con
->private;
1281 struct ceph_osd_client
*osdc
;
1282 int type
= le16_to_cpu(msg
->hdr
.type
);
1289 case CEPH_MSG_OSD_MAP
:
1290 ceph_osdc_handle_map(osdc
, msg
);
1292 case CEPH_MSG_OSD_OPREPLY
:
1293 handle_reply(osdc
, msg
, con
);
1297 pr_err("received unknown message type %d %s\n", type
,
1298 ceph_msg_type_name(type
));
1303 static struct ceph_msg
*alloc_msg(struct ceph_connection
*con
,
1304 struct ceph_msg_header
*hdr
)
1306 struct ceph_osd
*osd
= con
->private;
1307 struct ceph_osd_client
*osdc
= osd
->o_osdc
;
1308 int type
= le16_to_cpu(hdr
->type
);
1309 int front
= le32_to_cpu(hdr
->front_len
);
1312 case CEPH_MSG_OSD_OPREPLY
:
1313 return ceph_msgpool_get(&osdc
->msgpool_op_reply
, front
);
1315 return ceph_alloc_msg(con
, hdr
);
1319 * Wrappers to refcount containing ceph_osd struct
1321 static struct ceph_connection
*get_osd_con(struct ceph_connection
*con
)
1323 struct ceph_osd
*osd
= con
->private;
1329 static void put_osd_con(struct ceph_connection
*con
)
1331 struct ceph_osd
*osd
= con
->private;
1338 static int get_authorizer(struct ceph_connection
*con
,
1339 void **buf
, int *len
, int *proto
,
1340 void **reply_buf
, int *reply_len
, int force_new
)
1342 struct ceph_osd
*o
= con
->private;
1343 struct ceph_osd_client
*osdc
= o
->o_osdc
;
1344 struct ceph_auth_client
*ac
= osdc
->client
->monc
.auth
;
1347 if (force_new
&& o
->o_authorizer
) {
1348 ac
->ops
->destroy_authorizer(ac
, o
->o_authorizer
);
1349 o
->o_authorizer
= NULL
;
1351 if (o
->o_authorizer
== NULL
) {
1352 ret
= ac
->ops
->create_authorizer(
1353 ac
, CEPH_ENTITY_TYPE_OSD
,
1355 &o
->o_authorizer_buf
,
1356 &o
->o_authorizer_buf_len
,
1357 &o
->o_authorizer_reply_buf
,
1358 &o
->o_authorizer_reply_buf_len
);
1363 *proto
= ac
->protocol
;
1364 *buf
= o
->o_authorizer_buf
;
1365 *len
= o
->o_authorizer_buf_len
;
1366 *reply_buf
= o
->o_authorizer_reply_buf
;
1367 *reply_len
= o
->o_authorizer_reply_buf_len
;
1372 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
1374 struct ceph_osd
*o
= con
->private;
1375 struct ceph_osd_client
*osdc
= o
->o_osdc
;
1376 struct ceph_auth_client
*ac
= osdc
->client
->monc
.auth
;
1378 return ac
->ops
->verify_authorizer_reply(ac
, o
->o_authorizer
, len
);
1382 const static struct ceph_connection_operations osd_con_ops
= {
1385 .dispatch
= dispatch
,
1386 .get_authorizer
= get_authorizer
,
1387 .verify_authorizer_reply
= verify_authorizer_reply
,
1388 .alloc_msg
= alloc_msg
,
1390 .alloc_middle
= ceph_alloc_middle
,
1391 .prepare_pages
= prepare_pages
,