xref: /openbmc/linux/net/ceph/osd_client.c (revision 95e9fd10)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/module.h>
4 #include <linux/err.h>
5 #include <linux/highmem.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/slab.h>
9 #include <linux/uaccess.h>
10 #ifdef CONFIG_BLOCK
11 #include <linux/bio.h>
12 #endif
13 
14 #include <linux/ceph/libceph.h>
15 #include <linux/ceph/osd_client.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/auth.h>
19 #include <linux/ceph/pagelist.h>
20 
21 #define OSD_OP_FRONT_LEN	4096
22 #define OSD_OPREPLY_FRONT_LEN	512
23 
24 static const struct ceph_connection_operations osd_con_ops;
25 
26 static void send_queued(struct ceph_osd_client *osdc);
27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
28 static void __register_request(struct ceph_osd_client *osdc,
29 			       struct ceph_osd_request *req);
30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
31 					struct ceph_osd_request *req);
32 static void __send_request(struct ceph_osd_client *osdc,
33 			   struct ceph_osd_request *req);
34 
35 static int op_needs_trail(int op)
36 {
37 	switch (op) {
38 	case CEPH_OSD_OP_GETXATTR:
39 	case CEPH_OSD_OP_SETXATTR:
40 	case CEPH_OSD_OP_CMPXATTR:
41 	case CEPH_OSD_OP_CALL:
42 	case CEPH_OSD_OP_NOTIFY:
43 		return 1;
44 	default:
45 		return 0;
46 	}
47 }
48 
49 static int op_has_extent(int op)
50 {
51 	return (op == CEPH_OSD_OP_READ ||
52 		op == CEPH_OSD_OP_WRITE);
53 }
54 
55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
56 			struct ceph_file_layout *layout,
57 			u64 snapid,
58 			u64 off, u64 *plen, u64 *bno,
59 			struct ceph_osd_request *req,
60 			struct ceph_osd_req_op *op)
61 {
62 	struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
63 	u64 orig_len = *plen;
64 	u64 objoff, objlen;    /* extent in object */
65 
66 	reqhead->snapid = cpu_to_le64(snapid);
67 
68 	/* object extent? */
69 	ceph_calc_file_object_mapping(layout, off, plen, bno,
70 				      &objoff, &objlen);
71 	if (*plen < orig_len)
72 		dout(" skipping last %llu, final file extent %llu~%llu\n",
73 		     orig_len - *plen, off, *plen);
74 
75 	if (op_has_extent(op->op)) {
76 		op->extent.offset = objoff;
77 		op->extent.length = objlen;
78 	}
79 	req->r_num_pages = calc_pages_for(off, *plen);
80 	req->r_page_alignment = off & ~PAGE_MASK;
81 	if (op->op == CEPH_OSD_OP_WRITE)
82 		op->payload_len = *plen;
83 
84 	dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
85 	     *bno, objoff, objlen, req->r_num_pages);
86 
87 }
88 EXPORT_SYMBOL(ceph_calc_raw_layout);
89 
90 /*
91  * Implement client access to distributed object storage cluster.
92  *
93  * All data objects are stored within a cluster/cloud of OSDs, or
94  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
95  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
96  * remote daemons serving up and coordinating consistent and safe
97  * access to storage.
98  *
99  * Cluster membership and the mapping of data objects onto storage devices
100  * are described by the osd map.
101  *
102  * We keep track of pending OSD requests (read, write), resubmit
103  * requests to different OSDs when the cluster topology/data layout
104  * change, or retry the affected requests when the communications
105  * channel with an OSD is reset.
106  */
107 
108 /*
109  * calculate the mapping of a file extent onto an object, and fill out the
110  * request accordingly.  shorten extent as necessary if it crosses an
111  * object boundary.
112  *
113  * fill osd op in request message.
114  */
115 static void calc_layout(struct ceph_osd_client *osdc,
116 			struct ceph_vino vino,
117 			struct ceph_file_layout *layout,
118 			u64 off, u64 *plen,
119 			struct ceph_osd_request *req,
120 			struct ceph_osd_req_op *op)
121 {
122 	u64 bno;
123 
124 	ceph_calc_raw_layout(osdc, layout, vino.snap, off,
125 			     plen, &bno, req, op);
126 
127 	snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
128 	req->r_oid_len = strlen(req->r_oid);
129 }
130 
131 /*
132  * requests
133  */
134 void ceph_osdc_release_request(struct kref *kref)
135 {
136 	struct ceph_osd_request *req = container_of(kref,
137 						    struct ceph_osd_request,
138 						    r_kref);
139 
140 	if (req->r_request)
141 		ceph_msg_put(req->r_request);
142 	if (req->r_con_filling_msg) {
143 		dout("%s revoking pages %p from con %p\n", __func__,
144 		     req->r_pages, req->r_con_filling_msg);
145 		ceph_msg_revoke_incoming(req->r_reply);
146 		req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
147 	}
148 	if (req->r_reply)
149 		ceph_msg_put(req->r_reply);
150 	if (req->r_own_pages)
151 		ceph_release_page_vector(req->r_pages,
152 					 req->r_num_pages);
153 #ifdef CONFIG_BLOCK
154 	if (req->r_bio)
155 		bio_put(req->r_bio);
156 #endif
157 	ceph_put_snap_context(req->r_snapc);
158 	if (req->r_trail) {
159 		ceph_pagelist_release(req->r_trail);
160 		kfree(req->r_trail);
161 	}
162 	if (req->r_mempool)
163 		mempool_free(req, req->r_osdc->req_mempool);
164 	else
165 		kfree(req);
166 }
167 EXPORT_SYMBOL(ceph_osdc_release_request);
168 
169 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
170 {
171 	int i = 0;
172 
173 	if (needs_trail)
174 		*needs_trail = 0;
175 	while (ops[i].op) {
176 		if (needs_trail && op_needs_trail(ops[i].op))
177 			*needs_trail = 1;
178 		i++;
179 	}
180 
181 	return i;
182 }
183 
184 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
185 					       int flags,
186 					       struct ceph_snap_context *snapc,
187 					       struct ceph_osd_req_op *ops,
188 					       bool use_mempool,
189 					       gfp_t gfp_flags,
190 					       struct page **pages,
191 					       struct bio *bio)
192 {
193 	struct ceph_osd_request *req;
194 	struct ceph_msg *msg;
195 	int needs_trail;
196 	int num_op = get_num_ops(ops, &needs_trail);
197 	size_t msg_size = sizeof(struct ceph_osd_request_head);
198 
199 	msg_size += num_op*sizeof(struct ceph_osd_op);
200 
201 	if (use_mempool) {
202 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
203 		memset(req, 0, sizeof(*req));
204 	} else {
205 		req = kzalloc(sizeof(*req), gfp_flags);
206 	}
207 	if (req == NULL)
208 		return NULL;
209 
210 	req->r_osdc = osdc;
211 	req->r_mempool = use_mempool;
212 
213 	kref_init(&req->r_kref);
214 	init_completion(&req->r_completion);
215 	init_completion(&req->r_safe_completion);
216 	rb_init_node(&req->r_node);
217 	INIT_LIST_HEAD(&req->r_unsafe_item);
218 	INIT_LIST_HEAD(&req->r_linger_item);
219 	INIT_LIST_HEAD(&req->r_linger_osd);
220 	INIT_LIST_HEAD(&req->r_req_lru_item);
221 	INIT_LIST_HEAD(&req->r_osd_item);
222 
223 	req->r_flags = flags;
224 
225 	WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
226 
227 	/* create reply message */
228 	if (use_mempool)
229 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
230 	else
231 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
232 				   OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
233 	if (!msg) {
234 		ceph_osdc_put_request(req);
235 		return NULL;
236 	}
237 	req->r_reply = msg;
238 
239 	/* allocate space for the trailing data */
240 	if (needs_trail) {
241 		req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
242 		if (!req->r_trail) {
243 			ceph_osdc_put_request(req);
244 			return NULL;
245 		}
246 		ceph_pagelist_init(req->r_trail);
247 	}
248 
249 	/* create request message; allow space for oid */
250 	msg_size += MAX_OBJ_NAME_SIZE;
251 	if (snapc)
252 		msg_size += sizeof(u64) * snapc->num_snaps;
253 	if (use_mempool)
254 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
255 	else
256 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
257 	if (!msg) {
258 		ceph_osdc_put_request(req);
259 		return NULL;
260 	}
261 
262 	memset(msg->front.iov_base, 0, msg->front.iov_len);
263 
264 	req->r_request = msg;
265 	req->r_pages = pages;
266 #ifdef CONFIG_BLOCK
267 	if (bio) {
268 		req->r_bio = bio;
269 		bio_get(req->r_bio);
270 	}
271 #endif
272 
273 	return req;
274 }
275 EXPORT_SYMBOL(ceph_osdc_alloc_request);
276 
277 static void osd_req_encode_op(struct ceph_osd_request *req,
278 			      struct ceph_osd_op *dst,
279 			      struct ceph_osd_req_op *src)
280 {
281 	dst->op = cpu_to_le16(src->op);
282 
283 	switch (src->op) {
284 	case CEPH_OSD_OP_READ:
285 	case CEPH_OSD_OP_WRITE:
286 		dst->extent.offset =
287 			cpu_to_le64(src->extent.offset);
288 		dst->extent.length =
289 			cpu_to_le64(src->extent.length);
290 		dst->extent.truncate_size =
291 			cpu_to_le64(src->extent.truncate_size);
292 		dst->extent.truncate_seq =
293 			cpu_to_le32(src->extent.truncate_seq);
294 		break;
295 
296 	case CEPH_OSD_OP_GETXATTR:
297 	case CEPH_OSD_OP_SETXATTR:
298 	case CEPH_OSD_OP_CMPXATTR:
299 		BUG_ON(!req->r_trail);
300 
301 		dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
302 		dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
303 		dst->xattr.cmp_op = src->xattr.cmp_op;
304 		dst->xattr.cmp_mode = src->xattr.cmp_mode;
305 		ceph_pagelist_append(req->r_trail, src->xattr.name,
306 				     src->xattr.name_len);
307 		ceph_pagelist_append(req->r_trail, src->xattr.val,
308 				     src->xattr.value_len);
309 		break;
310 	case CEPH_OSD_OP_CALL:
311 		BUG_ON(!req->r_trail);
312 
313 		dst->cls.class_len = src->cls.class_len;
314 		dst->cls.method_len = src->cls.method_len;
315 		dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
316 
317 		ceph_pagelist_append(req->r_trail, src->cls.class_name,
318 				     src->cls.class_len);
319 		ceph_pagelist_append(req->r_trail, src->cls.method_name,
320 				     src->cls.method_len);
321 		ceph_pagelist_append(req->r_trail, src->cls.indata,
322 				     src->cls.indata_len);
323 		break;
324 	case CEPH_OSD_OP_ROLLBACK:
325 		dst->snap.snapid = cpu_to_le64(src->snap.snapid);
326 		break;
327 	case CEPH_OSD_OP_STARTSYNC:
328 		break;
329 	case CEPH_OSD_OP_NOTIFY:
330 		{
331 			__le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
332 			__le32 timeout = cpu_to_le32(src->watch.timeout);
333 
334 			BUG_ON(!req->r_trail);
335 
336 			ceph_pagelist_append(req->r_trail,
337 						&prot_ver, sizeof(prot_ver));
338 			ceph_pagelist_append(req->r_trail,
339 						&timeout, sizeof(timeout));
340 		}
341 	case CEPH_OSD_OP_NOTIFY_ACK:
342 	case CEPH_OSD_OP_WATCH:
343 		dst->watch.cookie = cpu_to_le64(src->watch.cookie);
344 		dst->watch.ver = cpu_to_le64(src->watch.ver);
345 		dst->watch.flag = src->watch.flag;
346 		break;
347 	default:
348 		pr_err("unrecognized osd opcode %d\n", dst->op);
349 		WARN_ON(1);
350 		break;
351 	}
352 	dst->payload_len = cpu_to_le32(src->payload_len);
353 }
354 
355 /*
356  * build new request AND message
357  *
358  */
359 void ceph_osdc_build_request(struct ceph_osd_request *req,
360 			     u64 off, u64 *plen,
361 			     struct ceph_osd_req_op *src_ops,
362 			     struct ceph_snap_context *snapc,
363 			     struct timespec *mtime,
364 			     const char *oid,
365 			     int oid_len)
366 {
367 	struct ceph_msg *msg = req->r_request;
368 	struct ceph_osd_request_head *head;
369 	struct ceph_osd_req_op *src_op;
370 	struct ceph_osd_op *op;
371 	void *p;
372 	int num_op = get_num_ops(src_ops, NULL);
373 	size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
374 	int flags = req->r_flags;
375 	u64 data_len = 0;
376 	int i;
377 
378 	head = msg->front.iov_base;
379 	op = (void *)(head + 1);
380 	p = (void *)(op + num_op);
381 
382 	req->r_snapc = ceph_get_snap_context(snapc);
383 
384 	head->client_inc = cpu_to_le32(1); /* always, for now. */
385 	head->flags = cpu_to_le32(flags);
386 	if (flags & CEPH_OSD_FLAG_WRITE)
387 		ceph_encode_timespec(&head->mtime, mtime);
388 	head->num_ops = cpu_to_le16(num_op);
389 
390 
391 	/* fill in oid */
392 	head->object_len = cpu_to_le32(oid_len);
393 	memcpy(p, oid, oid_len);
394 	p += oid_len;
395 
396 	src_op = src_ops;
397 	while (src_op->op) {
398 		osd_req_encode_op(req, op, src_op);
399 		src_op++;
400 		op++;
401 	}
402 
403 	if (req->r_trail)
404 		data_len += req->r_trail->length;
405 
406 	if (snapc) {
407 		head->snap_seq = cpu_to_le64(snapc->seq);
408 		head->num_snaps = cpu_to_le32(snapc->num_snaps);
409 		for (i = 0; i < snapc->num_snaps; i++) {
410 			put_unaligned_le64(snapc->snaps[i], p);
411 			p += sizeof(u64);
412 		}
413 	}
414 
415 	if (flags & CEPH_OSD_FLAG_WRITE) {
416 		req->r_request->hdr.data_off = cpu_to_le16(off);
417 		req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
418 	} else if (data_len) {
419 		req->r_request->hdr.data_off = 0;
420 		req->r_request->hdr.data_len = cpu_to_le32(data_len);
421 	}
422 
423 	req->r_request->page_alignment = req->r_page_alignment;
424 
425 	BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
426 	msg_size = p - msg->front.iov_base;
427 	msg->front.iov_len = msg_size;
428 	msg->hdr.front_len = cpu_to_le32(msg_size);
429 	return;
430 }
431 EXPORT_SYMBOL(ceph_osdc_build_request);
432 
433 /*
434  * build new request AND message, calculate layout, and adjust file
435  * extent as needed.
436  *
437  * if the file was recently truncated, we include information about its
438  * old and new size so that the object can be updated appropriately.  (we
439  * avoid synchronously deleting truncated objects because it's slow.)
440  *
441  * if @do_sync, include a 'startsync' command so that the osd will flush
442  * data quickly.
443  */
444 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
445 					       struct ceph_file_layout *layout,
446 					       struct ceph_vino vino,
447 					       u64 off, u64 *plen,
448 					       int opcode, int flags,
449 					       struct ceph_snap_context *snapc,
450 					       int do_sync,
451 					       u32 truncate_seq,
452 					       u64 truncate_size,
453 					       struct timespec *mtime,
454 					       bool use_mempool, int num_reply,
455 					       int page_align)
456 {
457 	struct ceph_osd_req_op ops[3];
458 	struct ceph_osd_request *req;
459 
460 	ops[0].op = opcode;
461 	ops[0].extent.truncate_seq = truncate_seq;
462 	ops[0].extent.truncate_size = truncate_size;
463 	ops[0].payload_len = 0;
464 
465 	if (do_sync) {
466 		ops[1].op = CEPH_OSD_OP_STARTSYNC;
467 		ops[1].payload_len = 0;
468 		ops[2].op = 0;
469 	} else
470 		ops[1].op = 0;
471 
472 	req = ceph_osdc_alloc_request(osdc, flags,
473 					 snapc, ops,
474 					 use_mempool,
475 					 GFP_NOFS, NULL, NULL);
476 	if (!req)
477 		return NULL;
478 
479 	/* calculate max write size */
480 	calc_layout(osdc, vino, layout, off, plen, req, ops);
481 	req->r_file_layout = *layout;  /* keep a copy */
482 
483 	/* in case it differs from natural (file) alignment that
484 	   calc_layout filled in for us */
485 	req->r_num_pages = calc_pages_for(page_align, *plen);
486 	req->r_page_alignment = page_align;
487 
488 	ceph_osdc_build_request(req, off, plen, ops,
489 				snapc,
490 				mtime,
491 				req->r_oid, req->r_oid_len);
492 
493 	return req;
494 }
495 EXPORT_SYMBOL(ceph_osdc_new_request);
496 
497 /*
498  * We keep osd requests in an rbtree, sorted by ->r_tid.
499  */
500 static void __insert_request(struct ceph_osd_client *osdc,
501 			     struct ceph_osd_request *new)
502 {
503 	struct rb_node **p = &osdc->requests.rb_node;
504 	struct rb_node *parent = NULL;
505 	struct ceph_osd_request *req = NULL;
506 
507 	while (*p) {
508 		parent = *p;
509 		req = rb_entry(parent, struct ceph_osd_request, r_node);
510 		if (new->r_tid < req->r_tid)
511 			p = &(*p)->rb_left;
512 		else if (new->r_tid > req->r_tid)
513 			p = &(*p)->rb_right;
514 		else
515 			BUG();
516 	}
517 
518 	rb_link_node(&new->r_node, parent, p);
519 	rb_insert_color(&new->r_node, &osdc->requests);
520 }
521 
522 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
523 						 u64 tid)
524 {
525 	struct ceph_osd_request *req;
526 	struct rb_node *n = osdc->requests.rb_node;
527 
528 	while (n) {
529 		req = rb_entry(n, struct ceph_osd_request, r_node);
530 		if (tid < req->r_tid)
531 			n = n->rb_left;
532 		else if (tid > req->r_tid)
533 			n = n->rb_right;
534 		else
535 			return req;
536 	}
537 	return NULL;
538 }
539 
540 static struct ceph_osd_request *
541 __lookup_request_ge(struct ceph_osd_client *osdc,
542 		    u64 tid)
543 {
544 	struct ceph_osd_request *req;
545 	struct rb_node *n = osdc->requests.rb_node;
546 
547 	while (n) {
548 		req = rb_entry(n, struct ceph_osd_request, r_node);
549 		if (tid < req->r_tid) {
550 			if (!n->rb_left)
551 				return req;
552 			n = n->rb_left;
553 		} else if (tid > req->r_tid) {
554 			n = n->rb_right;
555 		} else {
556 			return req;
557 		}
558 	}
559 	return NULL;
560 }
561 
562 /*
563  * Resubmit requests pending on the given osd.
564  */
565 static void __kick_osd_requests(struct ceph_osd_client *osdc,
566 				struct ceph_osd *osd)
567 {
568 	struct ceph_osd_request *req, *nreq;
569 	int err;
570 
571 	dout("__kick_osd_requests osd%d\n", osd->o_osd);
572 	err = __reset_osd(osdc, osd);
573 	if (err == -EAGAIN)
574 		return;
575 
576 	list_for_each_entry(req, &osd->o_requests, r_osd_item) {
577 		list_move(&req->r_req_lru_item, &osdc->req_unsent);
578 		dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
579 		     osd->o_osd);
580 		if (!req->r_linger)
581 			req->r_flags |= CEPH_OSD_FLAG_RETRY;
582 	}
583 
584 	list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
585 				 r_linger_osd) {
586 		/*
587 		 * reregister request prior to unregistering linger so
588 		 * that r_osd is preserved.
589 		 */
590 		BUG_ON(!list_empty(&req->r_req_lru_item));
591 		__register_request(osdc, req);
592 		list_add(&req->r_req_lru_item, &osdc->req_unsent);
593 		list_add(&req->r_osd_item, &req->r_osd->o_requests);
594 		__unregister_linger_request(osdc, req);
595 		dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
596 		     osd->o_osd);
597 	}
598 }
599 
600 static void kick_osd_requests(struct ceph_osd_client *osdc,
601 			      struct ceph_osd *kickosd)
602 {
603 	mutex_lock(&osdc->request_mutex);
604 	__kick_osd_requests(osdc, kickosd);
605 	mutex_unlock(&osdc->request_mutex);
606 }
607 
608 /*
609  * If the osd connection drops, we need to resubmit all requests.
610  */
611 static void osd_reset(struct ceph_connection *con)
612 {
613 	struct ceph_osd *osd = con->private;
614 	struct ceph_osd_client *osdc;
615 
616 	if (!osd)
617 		return;
618 	dout("osd_reset osd%d\n", osd->o_osd);
619 	osdc = osd->o_osdc;
620 	down_read(&osdc->map_sem);
621 	kick_osd_requests(osdc, osd);
622 	send_queued(osdc);
623 	up_read(&osdc->map_sem);
624 }
625 
626 /*
627  * Track open sessions with osds.
628  */
629 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
630 {
631 	struct ceph_osd *osd;
632 
633 	osd = kzalloc(sizeof(*osd), GFP_NOFS);
634 	if (!osd)
635 		return NULL;
636 
637 	atomic_set(&osd->o_ref, 1);
638 	osd->o_osdc = osdc;
639 	osd->o_osd = onum;
640 	INIT_LIST_HEAD(&osd->o_requests);
641 	INIT_LIST_HEAD(&osd->o_linger_requests);
642 	INIT_LIST_HEAD(&osd->o_osd_lru);
643 	osd->o_incarnation = 1;
644 
645 	ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
646 
647 	INIT_LIST_HEAD(&osd->o_keepalive_item);
648 	return osd;
649 }
650 
651 static struct ceph_osd *get_osd(struct ceph_osd *osd)
652 {
653 	if (atomic_inc_not_zero(&osd->o_ref)) {
654 		dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
655 		     atomic_read(&osd->o_ref));
656 		return osd;
657 	} else {
658 		dout("get_osd %p FAIL\n", osd);
659 		return NULL;
660 	}
661 }
662 
663 static void put_osd(struct ceph_osd *osd)
664 {
665 	dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
666 	     atomic_read(&osd->o_ref) - 1);
667 	if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
668 		struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
669 
670 		if (ac->ops && ac->ops->destroy_authorizer)
671 			ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
672 		kfree(osd);
673 	}
674 }
675 
676 /*
677  * remove an osd from our map
678  */
679 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
680 {
681 	dout("__remove_osd %p\n", osd);
682 	BUG_ON(!list_empty(&osd->o_requests));
683 	rb_erase(&osd->o_node, &osdc->osds);
684 	list_del_init(&osd->o_osd_lru);
685 	ceph_con_close(&osd->o_con);
686 	put_osd(osd);
687 }
688 
689 static void remove_all_osds(struct ceph_osd_client *osdc)
690 {
691 	dout("%s %p\n", __func__, osdc);
692 	mutex_lock(&osdc->request_mutex);
693 	while (!RB_EMPTY_ROOT(&osdc->osds)) {
694 		struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
695 						struct ceph_osd, o_node);
696 		__remove_osd(osdc, osd);
697 	}
698 	mutex_unlock(&osdc->request_mutex);
699 }
700 
701 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
702 			      struct ceph_osd *osd)
703 {
704 	dout("__move_osd_to_lru %p\n", osd);
705 	BUG_ON(!list_empty(&osd->o_osd_lru));
706 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
707 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
708 }
709 
710 static void __remove_osd_from_lru(struct ceph_osd *osd)
711 {
712 	dout("__remove_osd_from_lru %p\n", osd);
713 	if (!list_empty(&osd->o_osd_lru))
714 		list_del_init(&osd->o_osd_lru);
715 }
716 
717 static void remove_old_osds(struct ceph_osd_client *osdc)
718 {
719 	struct ceph_osd *osd, *nosd;
720 
721 	dout("__remove_old_osds %p\n", osdc);
722 	mutex_lock(&osdc->request_mutex);
723 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
724 		if (time_before(jiffies, osd->lru_ttl))
725 			break;
726 		__remove_osd(osdc, osd);
727 	}
728 	mutex_unlock(&osdc->request_mutex);
729 }
730 
731 /*
732  * reset osd connect
733  */
734 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
735 {
736 	struct ceph_osd_request *req;
737 	int ret = 0;
738 
739 	dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
740 	if (list_empty(&osd->o_requests) &&
741 	    list_empty(&osd->o_linger_requests)) {
742 		__remove_osd(osdc, osd);
743 	} else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
744 			  &osd->o_con.peer_addr,
745 			  sizeof(osd->o_con.peer_addr)) == 0 &&
746 		   !ceph_con_opened(&osd->o_con)) {
747 		dout(" osd addr hasn't changed and connection never opened,"
748 		     " letting msgr retry");
749 		/* touch each r_stamp for handle_timeout()'s benfit */
750 		list_for_each_entry(req, &osd->o_requests, r_osd_item)
751 			req->r_stamp = jiffies;
752 		ret = -EAGAIN;
753 	} else {
754 		ceph_con_close(&osd->o_con);
755 		ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
756 			      &osdc->osdmap->osd_addr[osd->o_osd]);
757 		osd->o_incarnation++;
758 	}
759 	return ret;
760 }
761 
762 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
763 {
764 	struct rb_node **p = &osdc->osds.rb_node;
765 	struct rb_node *parent = NULL;
766 	struct ceph_osd *osd = NULL;
767 
768 	dout("__insert_osd %p osd%d\n", new, new->o_osd);
769 	while (*p) {
770 		parent = *p;
771 		osd = rb_entry(parent, struct ceph_osd, o_node);
772 		if (new->o_osd < osd->o_osd)
773 			p = &(*p)->rb_left;
774 		else if (new->o_osd > osd->o_osd)
775 			p = &(*p)->rb_right;
776 		else
777 			BUG();
778 	}
779 
780 	rb_link_node(&new->o_node, parent, p);
781 	rb_insert_color(&new->o_node, &osdc->osds);
782 }
783 
784 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
785 {
786 	struct ceph_osd *osd;
787 	struct rb_node *n = osdc->osds.rb_node;
788 
789 	while (n) {
790 		osd = rb_entry(n, struct ceph_osd, o_node);
791 		if (o < osd->o_osd)
792 			n = n->rb_left;
793 		else if (o > osd->o_osd)
794 			n = n->rb_right;
795 		else
796 			return osd;
797 	}
798 	return NULL;
799 }
800 
801 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
802 {
803 	schedule_delayed_work(&osdc->timeout_work,
804 			osdc->client->options->osd_keepalive_timeout * HZ);
805 }
806 
807 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
808 {
809 	cancel_delayed_work(&osdc->timeout_work);
810 }
811 
812 /*
813  * Register request, assign tid.  If this is the first request, set up
814  * the timeout event.
815  */
816 static void __register_request(struct ceph_osd_client *osdc,
817 			       struct ceph_osd_request *req)
818 {
819 	req->r_tid = ++osdc->last_tid;
820 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
821 	dout("__register_request %p tid %lld\n", req, req->r_tid);
822 	__insert_request(osdc, req);
823 	ceph_osdc_get_request(req);
824 	osdc->num_requests++;
825 	if (osdc->num_requests == 1) {
826 		dout(" first request, scheduling timeout\n");
827 		__schedule_osd_timeout(osdc);
828 	}
829 }
830 
831 static void register_request(struct ceph_osd_client *osdc,
832 			     struct ceph_osd_request *req)
833 {
834 	mutex_lock(&osdc->request_mutex);
835 	__register_request(osdc, req);
836 	mutex_unlock(&osdc->request_mutex);
837 }
838 
839 /*
840  * called under osdc->request_mutex
841  */
842 static void __unregister_request(struct ceph_osd_client *osdc,
843 				 struct ceph_osd_request *req)
844 {
845 	if (RB_EMPTY_NODE(&req->r_node)) {
846 		dout("__unregister_request %p tid %lld not registered\n",
847 			req, req->r_tid);
848 		return;
849 	}
850 
851 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
852 	rb_erase(&req->r_node, &osdc->requests);
853 	osdc->num_requests--;
854 
855 	if (req->r_osd) {
856 		/* make sure the original request isn't in flight. */
857 		ceph_msg_revoke(req->r_request);
858 
859 		list_del_init(&req->r_osd_item);
860 		if (list_empty(&req->r_osd->o_requests) &&
861 		    list_empty(&req->r_osd->o_linger_requests)) {
862 			dout("moving osd to %p lru\n", req->r_osd);
863 			__move_osd_to_lru(osdc, req->r_osd);
864 		}
865 		if (list_empty(&req->r_linger_item))
866 			req->r_osd = NULL;
867 	}
868 
869 	ceph_osdc_put_request(req);
870 
871 	list_del_init(&req->r_req_lru_item);
872 	if (osdc->num_requests == 0) {
873 		dout(" no requests, canceling timeout\n");
874 		__cancel_osd_timeout(osdc);
875 	}
876 }
877 
878 /*
879  * Cancel a previously queued request message
880  */
881 static void __cancel_request(struct ceph_osd_request *req)
882 {
883 	if (req->r_sent && req->r_osd) {
884 		ceph_msg_revoke(req->r_request);
885 		req->r_sent = 0;
886 	}
887 }
888 
889 static void __register_linger_request(struct ceph_osd_client *osdc,
890 				    struct ceph_osd_request *req)
891 {
892 	dout("__register_linger_request %p\n", req);
893 	list_add_tail(&req->r_linger_item, &osdc->req_linger);
894 	if (req->r_osd)
895 		list_add_tail(&req->r_linger_osd,
896 			      &req->r_osd->o_linger_requests);
897 }
898 
899 static void __unregister_linger_request(struct ceph_osd_client *osdc,
900 					struct ceph_osd_request *req)
901 {
902 	dout("__unregister_linger_request %p\n", req);
903 	if (req->r_osd) {
904 		list_del_init(&req->r_linger_item);
905 		list_del_init(&req->r_linger_osd);
906 
907 		if (list_empty(&req->r_osd->o_requests) &&
908 		    list_empty(&req->r_osd->o_linger_requests)) {
909 			dout("moving osd to %p lru\n", req->r_osd);
910 			__move_osd_to_lru(osdc, req->r_osd);
911 		}
912 		if (list_empty(&req->r_osd_item))
913 			req->r_osd = NULL;
914 	}
915 }
916 
917 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
918 					 struct ceph_osd_request *req)
919 {
920 	mutex_lock(&osdc->request_mutex);
921 	if (req->r_linger) {
922 		__unregister_linger_request(osdc, req);
923 		ceph_osdc_put_request(req);
924 	}
925 	mutex_unlock(&osdc->request_mutex);
926 }
927 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
928 
929 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
930 				  struct ceph_osd_request *req)
931 {
932 	if (!req->r_linger) {
933 		dout("set_request_linger %p\n", req);
934 		req->r_linger = 1;
935 		/*
936 		 * caller is now responsible for calling
937 		 * unregister_linger_request
938 		 */
939 		ceph_osdc_get_request(req);
940 	}
941 }
942 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
943 
944 /*
945  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
946  * (as needed), and set the request r_osd appropriately.  If there is
947  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
948  * (unsent, homeless) or leave on in-flight lru.
949  *
950  * Return 0 if unchanged, 1 if changed, or negative on error.
951  *
952  * Caller should hold map_sem for read and request_mutex.
953  */
954 static int __map_request(struct ceph_osd_client *osdc,
955 			 struct ceph_osd_request *req, int force_resend)
956 {
957 	struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
958 	struct ceph_pg pgid;
959 	int acting[CEPH_PG_MAX_SIZE];
960 	int o = -1, num = 0;
961 	int err;
962 
963 	dout("map_request %p tid %lld\n", req, req->r_tid);
964 	err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
965 				      &req->r_file_layout, osdc->osdmap);
966 	if (err) {
967 		list_move(&req->r_req_lru_item, &osdc->req_notarget);
968 		return err;
969 	}
970 	pgid = reqhead->layout.ol_pgid;
971 	req->r_pgid = pgid;
972 
973 	err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
974 	if (err > 0) {
975 		o = acting[0];
976 		num = err;
977 	}
978 
979 	if ((!force_resend &&
980 	     req->r_osd && req->r_osd->o_osd == o &&
981 	     req->r_sent >= req->r_osd->o_incarnation &&
982 	     req->r_num_pg_osds == num &&
983 	     memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
984 	    (req->r_osd == NULL && o == -1))
985 		return 0;  /* no change */
986 
987 	dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
988 	     req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
989 	     req->r_osd ? req->r_osd->o_osd : -1);
990 
991 	/* record full pg acting set */
992 	memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
993 	req->r_num_pg_osds = num;
994 
995 	if (req->r_osd) {
996 		__cancel_request(req);
997 		list_del_init(&req->r_osd_item);
998 		req->r_osd = NULL;
999 	}
1000 
1001 	req->r_osd = __lookup_osd(osdc, o);
1002 	if (!req->r_osd && o >= 0) {
1003 		err = -ENOMEM;
1004 		req->r_osd = create_osd(osdc, o);
1005 		if (!req->r_osd) {
1006 			list_move(&req->r_req_lru_item, &osdc->req_notarget);
1007 			goto out;
1008 		}
1009 
1010 		dout("map_request osd %p is osd%d\n", req->r_osd, o);
1011 		__insert_osd(osdc, req->r_osd);
1012 
1013 		ceph_con_open(&req->r_osd->o_con,
1014 			      CEPH_ENTITY_TYPE_OSD, o,
1015 			      &osdc->osdmap->osd_addr[o]);
1016 	}
1017 
1018 	if (req->r_osd) {
1019 		__remove_osd_from_lru(req->r_osd);
1020 		list_add(&req->r_osd_item, &req->r_osd->o_requests);
1021 		list_move(&req->r_req_lru_item, &osdc->req_unsent);
1022 	} else {
1023 		list_move(&req->r_req_lru_item, &osdc->req_notarget);
1024 	}
1025 	err = 1;   /* osd or pg changed */
1026 
1027 out:
1028 	return err;
1029 }
1030 
1031 /*
1032  * caller should hold map_sem (for read) and request_mutex
1033  */
1034 static void __send_request(struct ceph_osd_client *osdc,
1035 			   struct ceph_osd_request *req)
1036 {
1037 	struct ceph_osd_request_head *reqhead;
1038 
1039 	dout("send_request %p tid %llu to osd%d flags %d\n",
1040 	     req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1041 
1042 	reqhead = req->r_request->front.iov_base;
1043 	reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1044 	reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1045 	reqhead->reassert_version = req->r_reassert_version;
1046 
1047 	req->r_stamp = jiffies;
1048 	list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1049 
1050 	ceph_msg_get(req->r_request); /* send consumes a ref */
1051 	ceph_con_send(&req->r_osd->o_con, req->r_request);
1052 	req->r_sent = req->r_osd->o_incarnation;
1053 }
1054 
1055 /*
1056  * Send any requests in the queue (req_unsent).
1057  */
1058 static void send_queued(struct ceph_osd_client *osdc)
1059 {
1060 	struct ceph_osd_request *req, *tmp;
1061 
1062 	dout("send_queued\n");
1063 	mutex_lock(&osdc->request_mutex);
1064 	list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1065 		__send_request(osdc, req);
1066 	}
1067 	mutex_unlock(&osdc->request_mutex);
1068 }
1069 
1070 /*
1071  * Timeout callback, called every N seconds when 1 or more osd
1072  * requests has been active for more than N seconds.  When this
1073  * happens, we ping all OSDs with requests who have timed out to
1074  * ensure any communications channel reset is detected.  Reset the
1075  * request timeouts another N seconds in the future as we go.
1076  * Reschedule the timeout event another N seconds in future (unless
1077  * there are no open requests).
1078  */
1079 static void handle_timeout(struct work_struct *work)
1080 {
1081 	struct ceph_osd_client *osdc =
1082 		container_of(work, struct ceph_osd_client, timeout_work.work);
1083 	struct ceph_osd_request *req, *last_req = NULL;
1084 	struct ceph_osd *osd;
1085 	unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1086 	unsigned long keepalive =
1087 		osdc->client->options->osd_keepalive_timeout * HZ;
1088 	unsigned long last_stamp = 0;
1089 	struct list_head slow_osds;
1090 	dout("timeout\n");
1091 	down_read(&osdc->map_sem);
1092 
1093 	ceph_monc_request_next_osdmap(&osdc->client->monc);
1094 
1095 	mutex_lock(&osdc->request_mutex);
1096 
1097 	/*
1098 	 * reset osds that appear to be _really_ unresponsive.  this
1099 	 * is a failsafe measure.. we really shouldn't be getting to
1100 	 * this point if the system is working properly.  the monitors
1101 	 * should mark the osd as failed and we should find out about
1102 	 * it from an updated osd map.
1103 	 */
1104 	while (timeout && !list_empty(&osdc->req_lru)) {
1105 		req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1106 				 r_req_lru_item);
1107 
1108 		/* hasn't been long enough since we sent it? */
1109 		if (time_before(jiffies, req->r_stamp + timeout))
1110 			break;
1111 
1112 		/* hasn't been long enough since it was acked? */
1113 		if (req->r_request->ack_stamp == 0 ||
1114 		    time_before(jiffies, req->r_request->ack_stamp + timeout))
1115 			break;
1116 
1117 		BUG_ON(req == last_req && req->r_stamp == last_stamp);
1118 		last_req = req;
1119 		last_stamp = req->r_stamp;
1120 
1121 		osd = req->r_osd;
1122 		BUG_ON(!osd);
1123 		pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1124 			   req->r_tid, osd->o_osd);
1125 		__kick_osd_requests(osdc, osd);
1126 	}
1127 
1128 	/*
1129 	 * ping osds that are a bit slow.  this ensures that if there
1130 	 * is a break in the TCP connection we will notice, and reopen
1131 	 * a connection with that osd (from the fault callback).
1132 	 */
1133 	INIT_LIST_HEAD(&slow_osds);
1134 	list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1135 		if (time_before(jiffies, req->r_stamp + keepalive))
1136 			break;
1137 
1138 		osd = req->r_osd;
1139 		BUG_ON(!osd);
1140 		dout(" tid %llu is slow, will send keepalive on osd%d\n",
1141 		     req->r_tid, osd->o_osd);
1142 		list_move_tail(&osd->o_keepalive_item, &slow_osds);
1143 	}
1144 	while (!list_empty(&slow_osds)) {
1145 		osd = list_entry(slow_osds.next, struct ceph_osd,
1146 				 o_keepalive_item);
1147 		list_del_init(&osd->o_keepalive_item);
1148 		ceph_con_keepalive(&osd->o_con);
1149 	}
1150 
1151 	__schedule_osd_timeout(osdc);
1152 	mutex_unlock(&osdc->request_mutex);
1153 	send_queued(osdc);
1154 	up_read(&osdc->map_sem);
1155 }
1156 
1157 static void handle_osds_timeout(struct work_struct *work)
1158 {
1159 	struct ceph_osd_client *osdc =
1160 		container_of(work, struct ceph_osd_client,
1161 			     osds_timeout_work.work);
1162 	unsigned long delay =
1163 		osdc->client->options->osd_idle_ttl * HZ >> 2;
1164 
1165 	dout("osds timeout\n");
1166 	down_read(&osdc->map_sem);
1167 	remove_old_osds(osdc);
1168 	up_read(&osdc->map_sem);
1169 
1170 	schedule_delayed_work(&osdc->osds_timeout_work,
1171 			      round_jiffies_relative(delay));
1172 }
1173 
1174 static void complete_request(struct ceph_osd_request *req)
1175 {
1176 	if (req->r_safe_callback)
1177 		req->r_safe_callback(req, NULL);
1178 	complete_all(&req->r_safe_completion);  /* fsync waiter */
1179 }
1180 
1181 /*
1182  * handle osd op reply.  either call the callback if it is specified,
1183  * or do the completion to wake up the waiting thread.
1184  */
1185 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1186 			 struct ceph_connection *con)
1187 {
1188 	struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1189 	struct ceph_osd_request *req;
1190 	u64 tid;
1191 	int numops, object_len, flags;
1192 	s32 result;
1193 
1194 	tid = le64_to_cpu(msg->hdr.tid);
1195 	if (msg->front.iov_len < sizeof(*rhead))
1196 		goto bad;
1197 	numops = le32_to_cpu(rhead->num_ops);
1198 	object_len = le32_to_cpu(rhead->object_len);
1199 	result = le32_to_cpu(rhead->result);
1200 	if (msg->front.iov_len != sizeof(*rhead) + object_len +
1201 	    numops * sizeof(struct ceph_osd_op))
1202 		goto bad;
1203 	dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1204 	/* lookup */
1205 	mutex_lock(&osdc->request_mutex);
1206 	req = __lookup_request(osdc, tid);
1207 	if (req == NULL) {
1208 		dout("handle_reply tid %llu dne\n", tid);
1209 		mutex_unlock(&osdc->request_mutex);
1210 		return;
1211 	}
1212 	ceph_osdc_get_request(req);
1213 	flags = le32_to_cpu(rhead->flags);
1214 
1215 	/*
1216 	 * if this connection filled our message, drop our reference now, to
1217 	 * avoid a (safe but slower) revoke later.
1218 	 */
1219 	if (req->r_con_filling_msg == con && req->r_reply == msg) {
1220 		dout(" dropping con_filling_msg ref %p\n", con);
1221 		req->r_con_filling_msg = NULL;
1222 		con->ops->put(con);
1223 	}
1224 
1225 	if (!req->r_got_reply) {
1226 		unsigned int bytes;
1227 
1228 		req->r_result = le32_to_cpu(rhead->result);
1229 		bytes = le32_to_cpu(msg->hdr.data_len);
1230 		dout("handle_reply result %d bytes %d\n", req->r_result,
1231 		     bytes);
1232 		if (req->r_result == 0)
1233 			req->r_result = bytes;
1234 
1235 		/* in case this is a write and we need to replay, */
1236 		req->r_reassert_version = rhead->reassert_version;
1237 
1238 		req->r_got_reply = 1;
1239 	} else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1240 		dout("handle_reply tid %llu dup ack\n", tid);
1241 		mutex_unlock(&osdc->request_mutex);
1242 		goto done;
1243 	}
1244 
1245 	dout("handle_reply tid %llu flags %d\n", tid, flags);
1246 
1247 	if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1248 		__register_linger_request(osdc, req);
1249 
1250 	/* either this is a read, or we got the safe response */
1251 	if (result < 0 ||
1252 	    (flags & CEPH_OSD_FLAG_ONDISK) ||
1253 	    ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1254 		__unregister_request(osdc, req);
1255 
1256 	mutex_unlock(&osdc->request_mutex);
1257 
1258 	if (req->r_callback)
1259 		req->r_callback(req, msg);
1260 	else
1261 		complete_all(&req->r_completion);
1262 
1263 	if (flags & CEPH_OSD_FLAG_ONDISK)
1264 		complete_request(req);
1265 
1266 done:
1267 	dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1268 	ceph_osdc_put_request(req);
1269 	return;
1270 
1271 bad:
1272 	pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1273 	       (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1274 	       (int)sizeof(*rhead));
1275 	ceph_msg_dump(msg);
1276 }
1277 
1278 static void reset_changed_osds(struct ceph_osd_client *osdc)
1279 {
1280 	struct rb_node *p, *n;
1281 
1282 	for (p = rb_first(&osdc->osds); p; p = n) {
1283 		struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1284 
1285 		n = rb_next(p);
1286 		if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1287 		    memcmp(&osd->o_con.peer_addr,
1288 			   ceph_osd_addr(osdc->osdmap,
1289 					 osd->o_osd),
1290 			   sizeof(struct ceph_entity_addr)) != 0)
1291 			__reset_osd(osdc, osd);
1292 	}
1293 }
1294 
1295 /*
1296  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1297  * no osd, request a new map.
1298  *
1299  * Caller should hold map_sem for read and request_mutex.
1300  */
1301 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1302 {
1303 	struct ceph_osd_request *req, *nreq;
1304 	struct rb_node *p;
1305 	int needmap = 0;
1306 	int err;
1307 
1308 	dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1309 	mutex_lock(&osdc->request_mutex);
1310 	for (p = rb_first(&osdc->requests); p; ) {
1311 		req = rb_entry(p, struct ceph_osd_request, r_node);
1312 		p = rb_next(p);
1313 		err = __map_request(osdc, req, force_resend);
1314 		if (err < 0)
1315 			continue;  /* error */
1316 		if (req->r_osd == NULL) {
1317 			dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1318 			needmap++;  /* request a newer map */
1319 		} else if (err > 0) {
1320 			if (!req->r_linger) {
1321 				dout("%p tid %llu requeued on osd%d\n", req,
1322 				     req->r_tid,
1323 				     req->r_osd ? req->r_osd->o_osd : -1);
1324 				req->r_flags |= CEPH_OSD_FLAG_RETRY;
1325 			}
1326 		}
1327 		if (req->r_linger && list_empty(&req->r_linger_item)) {
1328 			/*
1329 			 * register as a linger so that we will
1330 			 * re-submit below and get a new tid
1331 			 */
1332 			dout("%p tid %llu restart on osd%d\n",
1333 			     req, req->r_tid,
1334 			     req->r_osd ? req->r_osd->o_osd : -1);
1335 			__register_linger_request(osdc, req);
1336 			__unregister_request(osdc, req);
1337 		}
1338 	}
1339 
1340 	list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1341 				 r_linger_item) {
1342 		dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1343 
1344 		err = __map_request(osdc, req, force_resend);
1345 		if (err == 0)
1346 			continue;  /* no change and no osd was specified */
1347 		if (err < 0)
1348 			continue;  /* hrm! */
1349 		if (req->r_osd == NULL) {
1350 			dout("tid %llu maps to no valid osd\n", req->r_tid);
1351 			needmap++;  /* request a newer map */
1352 			continue;
1353 		}
1354 
1355 		dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1356 		     req->r_osd ? req->r_osd->o_osd : -1);
1357 		__unregister_linger_request(osdc, req);
1358 		__register_request(osdc, req);
1359 	}
1360 	mutex_unlock(&osdc->request_mutex);
1361 
1362 	if (needmap) {
1363 		dout("%d requests for down osds, need new map\n", needmap);
1364 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1365 	}
1366 }
1367 
1368 
1369 /*
1370  * Process updated osd map.
1371  *
1372  * The message contains any number of incremental and full maps, normally
1373  * indicating some sort of topology change in the cluster.  Kick requests
1374  * off to different OSDs as needed.
1375  */
1376 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1377 {
1378 	void *p, *end, *next;
1379 	u32 nr_maps, maplen;
1380 	u32 epoch;
1381 	struct ceph_osdmap *newmap = NULL, *oldmap;
1382 	int err;
1383 	struct ceph_fsid fsid;
1384 
1385 	dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1386 	p = msg->front.iov_base;
1387 	end = p + msg->front.iov_len;
1388 
1389 	/* verify fsid */
1390 	ceph_decode_need(&p, end, sizeof(fsid), bad);
1391 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
1392 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
1393 		return;
1394 
1395 	down_write(&osdc->map_sem);
1396 
1397 	/* incremental maps */
1398 	ceph_decode_32_safe(&p, end, nr_maps, bad);
1399 	dout(" %d inc maps\n", nr_maps);
1400 	while (nr_maps > 0) {
1401 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1402 		epoch = ceph_decode_32(&p);
1403 		maplen = ceph_decode_32(&p);
1404 		ceph_decode_need(&p, end, maplen, bad);
1405 		next = p + maplen;
1406 		if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1407 			dout("applying incremental map %u len %d\n",
1408 			     epoch, maplen);
1409 			newmap = osdmap_apply_incremental(&p, next,
1410 							  osdc->osdmap,
1411 							  &osdc->client->msgr);
1412 			if (IS_ERR(newmap)) {
1413 				err = PTR_ERR(newmap);
1414 				goto bad;
1415 			}
1416 			BUG_ON(!newmap);
1417 			if (newmap != osdc->osdmap) {
1418 				ceph_osdmap_destroy(osdc->osdmap);
1419 				osdc->osdmap = newmap;
1420 			}
1421 			kick_requests(osdc, 0);
1422 			reset_changed_osds(osdc);
1423 		} else {
1424 			dout("ignoring incremental map %u len %d\n",
1425 			     epoch, maplen);
1426 		}
1427 		p = next;
1428 		nr_maps--;
1429 	}
1430 	if (newmap)
1431 		goto done;
1432 
1433 	/* full maps */
1434 	ceph_decode_32_safe(&p, end, nr_maps, bad);
1435 	dout(" %d full maps\n", nr_maps);
1436 	while (nr_maps) {
1437 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1438 		epoch = ceph_decode_32(&p);
1439 		maplen = ceph_decode_32(&p);
1440 		ceph_decode_need(&p, end, maplen, bad);
1441 		if (nr_maps > 1) {
1442 			dout("skipping non-latest full map %u len %d\n",
1443 			     epoch, maplen);
1444 		} else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1445 			dout("skipping full map %u len %d, "
1446 			     "older than our %u\n", epoch, maplen,
1447 			     osdc->osdmap->epoch);
1448 		} else {
1449 			int skipped_map = 0;
1450 
1451 			dout("taking full map %u len %d\n", epoch, maplen);
1452 			newmap = osdmap_decode(&p, p+maplen);
1453 			if (IS_ERR(newmap)) {
1454 				err = PTR_ERR(newmap);
1455 				goto bad;
1456 			}
1457 			BUG_ON(!newmap);
1458 			oldmap = osdc->osdmap;
1459 			osdc->osdmap = newmap;
1460 			if (oldmap) {
1461 				if (oldmap->epoch + 1 < newmap->epoch)
1462 					skipped_map = 1;
1463 				ceph_osdmap_destroy(oldmap);
1464 			}
1465 			kick_requests(osdc, skipped_map);
1466 		}
1467 		p += maplen;
1468 		nr_maps--;
1469 	}
1470 
1471 done:
1472 	downgrade_write(&osdc->map_sem);
1473 	ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1474 
1475 	/*
1476 	 * subscribe to subsequent osdmap updates if full to ensure
1477 	 * we find out when we are no longer full and stop returning
1478 	 * ENOSPC.
1479 	 */
1480 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1481 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1482 
1483 	send_queued(osdc);
1484 	up_read(&osdc->map_sem);
1485 	wake_up_all(&osdc->client->auth_wq);
1486 	return;
1487 
1488 bad:
1489 	pr_err("osdc handle_map corrupt msg\n");
1490 	ceph_msg_dump(msg);
1491 	up_write(&osdc->map_sem);
1492 	return;
1493 }
1494 
1495 /*
1496  * watch/notify callback event infrastructure
1497  *
1498  * These callbacks are used both for watch and notify operations.
1499  */
1500 static void __release_event(struct kref *kref)
1501 {
1502 	struct ceph_osd_event *event =
1503 		container_of(kref, struct ceph_osd_event, kref);
1504 
1505 	dout("__release_event %p\n", event);
1506 	kfree(event);
1507 }
1508 
1509 static void get_event(struct ceph_osd_event *event)
1510 {
1511 	kref_get(&event->kref);
1512 }
1513 
1514 void ceph_osdc_put_event(struct ceph_osd_event *event)
1515 {
1516 	kref_put(&event->kref, __release_event);
1517 }
1518 EXPORT_SYMBOL(ceph_osdc_put_event);
1519 
1520 static void __insert_event(struct ceph_osd_client *osdc,
1521 			     struct ceph_osd_event *new)
1522 {
1523 	struct rb_node **p = &osdc->event_tree.rb_node;
1524 	struct rb_node *parent = NULL;
1525 	struct ceph_osd_event *event = NULL;
1526 
1527 	while (*p) {
1528 		parent = *p;
1529 		event = rb_entry(parent, struct ceph_osd_event, node);
1530 		if (new->cookie < event->cookie)
1531 			p = &(*p)->rb_left;
1532 		else if (new->cookie > event->cookie)
1533 			p = &(*p)->rb_right;
1534 		else
1535 			BUG();
1536 	}
1537 
1538 	rb_link_node(&new->node, parent, p);
1539 	rb_insert_color(&new->node, &osdc->event_tree);
1540 }
1541 
1542 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1543 					        u64 cookie)
1544 {
1545 	struct rb_node **p = &osdc->event_tree.rb_node;
1546 	struct rb_node *parent = NULL;
1547 	struct ceph_osd_event *event = NULL;
1548 
1549 	while (*p) {
1550 		parent = *p;
1551 		event = rb_entry(parent, struct ceph_osd_event, node);
1552 		if (cookie < event->cookie)
1553 			p = &(*p)->rb_left;
1554 		else if (cookie > event->cookie)
1555 			p = &(*p)->rb_right;
1556 		else
1557 			return event;
1558 	}
1559 	return NULL;
1560 }
1561 
1562 static void __remove_event(struct ceph_osd_event *event)
1563 {
1564 	struct ceph_osd_client *osdc = event->osdc;
1565 
1566 	if (!RB_EMPTY_NODE(&event->node)) {
1567 		dout("__remove_event removed %p\n", event);
1568 		rb_erase(&event->node, &osdc->event_tree);
1569 		ceph_osdc_put_event(event);
1570 	} else {
1571 		dout("__remove_event didn't remove %p\n", event);
1572 	}
1573 }
1574 
1575 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1576 			   void (*event_cb)(u64, u64, u8, void *),
1577 			   int one_shot, void *data,
1578 			   struct ceph_osd_event **pevent)
1579 {
1580 	struct ceph_osd_event *event;
1581 
1582 	event = kmalloc(sizeof(*event), GFP_NOIO);
1583 	if (!event)
1584 		return -ENOMEM;
1585 
1586 	dout("create_event %p\n", event);
1587 	event->cb = event_cb;
1588 	event->one_shot = one_shot;
1589 	event->data = data;
1590 	event->osdc = osdc;
1591 	INIT_LIST_HEAD(&event->osd_node);
1592 	kref_init(&event->kref);   /* one ref for us */
1593 	kref_get(&event->kref);    /* one ref for the caller */
1594 	init_completion(&event->completion);
1595 
1596 	spin_lock(&osdc->event_lock);
1597 	event->cookie = ++osdc->event_count;
1598 	__insert_event(osdc, event);
1599 	spin_unlock(&osdc->event_lock);
1600 
1601 	*pevent = event;
1602 	return 0;
1603 }
1604 EXPORT_SYMBOL(ceph_osdc_create_event);
1605 
1606 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1607 {
1608 	struct ceph_osd_client *osdc = event->osdc;
1609 
1610 	dout("cancel_event %p\n", event);
1611 	spin_lock(&osdc->event_lock);
1612 	__remove_event(event);
1613 	spin_unlock(&osdc->event_lock);
1614 	ceph_osdc_put_event(event); /* caller's */
1615 }
1616 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1617 
1618 
1619 static void do_event_work(struct work_struct *work)
1620 {
1621 	struct ceph_osd_event_work *event_work =
1622 		container_of(work, struct ceph_osd_event_work, work);
1623 	struct ceph_osd_event *event = event_work->event;
1624 	u64 ver = event_work->ver;
1625 	u64 notify_id = event_work->notify_id;
1626 	u8 opcode = event_work->opcode;
1627 
1628 	dout("do_event_work completing %p\n", event);
1629 	event->cb(ver, notify_id, opcode, event->data);
1630 	complete(&event->completion);
1631 	dout("do_event_work completed %p\n", event);
1632 	ceph_osdc_put_event(event);
1633 	kfree(event_work);
1634 }
1635 
1636 
1637 /*
1638  * Process osd watch notifications
1639  */
1640 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1641 {
1642 	void *p, *end;
1643 	u8 proto_ver;
1644 	u64 cookie, ver, notify_id;
1645 	u8 opcode;
1646 	struct ceph_osd_event *event;
1647 	struct ceph_osd_event_work *event_work;
1648 
1649 	p = msg->front.iov_base;
1650 	end = p + msg->front.iov_len;
1651 
1652 	ceph_decode_8_safe(&p, end, proto_ver, bad);
1653 	ceph_decode_8_safe(&p, end, opcode, bad);
1654 	ceph_decode_64_safe(&p, end, cookie, bad);
1655 	ceph_decode_64_safe(&p, end, ver, bad);
1656 	ceph_decode_64_safe(&p, end, notify_id, bad);
1657 
1658 	spin_lock(&osdc->event_lock);
1659 	event = __find_event(osdc, cookie);
1660 	if (event) {
1661 		get_event(event);
1662 		if (event->one_shot)
1663 			__remove_event(event);
1664 	}
1665 	spin_unlock(&osdc->event_lock);
1666 	dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1667 	     cookie, ver, event);
1668 	if (event) {
1669 		event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1670 		if (!event_work) {
1671 			dout("ERROR: could not allocate event_work\n");
1672 			goto done_err;
1673 		}
1674 		INIT_WORK(&event_work->work, do_event_work);
1675 		event_work->event = event;
1676 		event_work->ver = ver;
1677 		event_work->notify_id = notify_id;
1678 		event_work->opcode = opcode;
1679 		if (!queue_work(osdc->notify_wq, &event_work->work)) {
1680 			dout("WARNING: failed to queue notify event work\n");
1681 			goto done_err;
1682 		}
1683 	}
1684 
1685 	return;
1686 
1687 done_err:
1688 	complete(&event->completion);
1689 	ceph_osdc_put_event(event);
1690 	return;
1691 
1692 bad:
1693 	pr_err("osdc handle_watch_notify corrupt msg\n");
1694 	return;
1695 }
1696 
1697 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1698 {
1699 	int err;
1700 
1701 	dout("wait_event %p\n", event);
1702 	err = wait_for_completion_interruptible_timeout(&event->completion,
1703 							timeout * HZ);
1704 	ceph_osdc_put_event(event);
1705 	if (err > 0)
1706 		err = 0;
1707 	dout("wait_event %p returns %d\n", event, err);
1708 	return err;
1709 }
1710 EXPORT_SYMBOL(ceph_osdc_wait_event);
1711 
1712 /*
1713  * Register request, send initial attempt.
1714  */
1715 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1716 			    struct ceph_osd_request *req,
1717 			    bool nofail)
1718 {
1719 	int rc = 0;
1720 
1721 	req->r_request->pages = req->r_pages;
1722 	req->r_request->nr_pages = req->r_num_pages;
1723 #ifdef CONFIG_BLOCK
1724 	req->r_request->bio = req->r_bio;
1725 #endif
1726 	req->r_request->trail = req->r_trail;
1727 
1728 	register_request(osdc, req);
1729 
1730 	down_read(&osdc->map_sem);
1731 	mutex_lock(&osdc->request_mutex);
1732 	/*
1733 	 * a racing kick_requests() may have sent the message for us
1734 	 * while we dropped request_mutex above, so only send now if
1735 	 * the request still han't been touched yet.
1736 	 */
1737 	if (req->r_sent == 0) {
1738 		rc = __map_request(osdc, req, 0);
1739 		if (rc < 0) {
1740 			if (nofail) {
1741 				dout("osdc_start_request failed map, "
1742 				     " will retry %lld\n", req->r_tid);
1743 				rc = 0;
1744 			}
1745 			goto out_unlock;
1746 		}
1747 		if (req->r_osd == NULL) {
1748 			dout("send_request %p no up osds in pg\n", req);
1749 			ceph_monc_request_next_osdmap(&osdc->client->monc);
1750 		} else {
1751 			__send_request(osdc, req);
1752 		}
1753 		rc = 0;
1754 	}
1755 
1756 out_unlock:
1757 	mutex_unlock(&osdc->request_mutex);
1758 	up_read(&osdc->map_sem);
1759 	return rc;
1760 }
1761 EXPORT_SYMBOL(ceph_osdc_start_request);
1762 
1763 /*
1764  * wait for a request to complete
1765  */
1766 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1767 			   struct ceph_osd_request *req)
1768 {
1769 	int rc;
1770 
1771 	rc = wait_for_completion_interruptible(&req->r_completion);
1772 	if (rc < 0) {
1773 		mutex_lock(&osdc->request_mutex);
1774 		__cancel_request(req);
1775 		__unregister_request(osdc, req);
1776 		mutex_unlock(&osdc->request_mutex);
1777 		complete_request(req);
1778 		dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1779 		return rc;
1780 	}
1781 
1782 	dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1783 	return req->r_result;
1784 }
1785 EXPORT_SYMBOL(ceph_osdc_wait_request);
1786 
1787 /*
1788  * sync - wait for all in-flight requests to flush.  avoid starvation.
1789  */
1790 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1791 {
1792 	struct ceph_osd_request *req;
1793 	u64 last_tid, next_tid = 0;
1794 
1795 	mutex_lock(&osdc->request_mutex);
1796 	last_tid = osdc->last_tid;
1797 	while (1) {
1798 		req = __lookup_request_ge(osdc, next_tid);
1799 		if (!req)
1800 			break;
1801 		if (req->r_tid > last_tid)
1802 			break;
1803 
1804 		next_tid = req->r_tid + 1;
1805 		if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1806 			continue;
1807 
1808 		ceph_osdc_get_request(req);
1809 		mutex_unlock(&osdc->request_mutex);
1810 		dout("sync waiting on tid %llu (last is %llu)\n",
1811 		     req->r_tid, last_tid);
1812 		wait_for_completion(&req->r_safe_completion);
1813 		mutex_lock(&osdc->request_mutex);
1814 		ceph_osdc_put_request(req);
1815 	}
1816 	mutex_unlock(&osdc->request_mutex);
1817 	dout("sync done (thru tid %llu)\n", last_tid);
1818 }
1819 EXPORT_SYMBOL(ceph_osdc_sync);
1820 
1821 /*
1822  * init, shutdown
1823  */
1824 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1825 {
1826 	int err;
1827 
1828 	dout("init\n");
1829 	osdc->client = client;
1830 	osdc->osdmap = NULL;
1831 	init_rwsem(&osdc->map_sem);
1832 	init_completion(&osdc->map_waiters);
1833 	osdc->last_requested_map = 0;
1834 	mutex_init(&osdc->request_mutex);
1835 	osdc->last_tid = 0;
1836 	osdc->osds = RB_ROOT;
1837 	INIT_LIST_HEAD(&osdc->osd_lru);
1838 	osdc->requests = RB_ROOT;
1839 	INIT_LIST_HEAD(&osdc->req_lru);
1840 	INIT_LIST_HEAD(&osdc->req_unsent);
1841 	INIT_LIST_HEAD(&osdc->req_notarget);
1842 	INIT_LIST_HEAD(&osdc->req_linger);
1843 	osdc->num_requests = 0;
1844 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1845 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1846 	spin_lock_init(&osdc->event_lock);
1847 	osdc->event_tree = RB_ROOT;
1848 	osdc->event_count = 0;
1849 
1850 	schedule_delayed_work(&osdc->osds_timeout_work,
1851 	   round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1852 
1853 	err = -ENOMEM;
1854 	osdc->req_mempool = mempool_create_kmalloc_pool(10,
1855 					sizeof(struct ceph_osd_request));
1856 	if (!osdc->req_mempool)
1857 		goto out;
1858 
1859 	err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1860 				OSD_OP_FRONT_LEN, 10, true,
1861 				"osd_op");
1862 	if (err < 0)
1863 		goto out_mempool;
1864 	err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1865 				OSD_OPREPLY_FRONT_LEN, 10, true,
1866 				"osd_op_reply");
1867 	if (err < 0)
1868 		goto out_msgpool;
1869 
1870 	osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1871 	if (IS_ERR(osdc->notify_wq)) {
1872 		err = PTR_ERR(osdc->notify_wq);
1873 		osdc->notify_wq = NULL;
1874 		goto out_msgpool;
1875 	}
1876 	return 0;
1877 
1878 out_msgpool:
1879 	ceph_msgpool_destroy(&osdc->msgpool_op);
1880 out_mempool:
1881 	mempool_destroy(osdc->req_mempool);
1882 out:
1883 	return err;
1884 }
1885 EXPORT_SYMBOL(ceph_osdc_init);
1886 
1887 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1888 {
1889 	flush_workqueue(osdc->notify_wq);
1890 	destroy_workqueue(osdc->notify_wq);
1891 	cancel_delayed_work_sync(&osdc->timeout_work);
1892 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
1893 	if (osdc->osdmap) {
1894 		ceph_osdmap_destroy(osdc->osdmap);
1895 		osdc->osdmap = NULL;
1896 	}
1897 	remove_all_osds(osdc);
1898 	mempool_destroy(osdc->req_mempool);
1899 	ceph_msgpool_destroy(&osdc->msgpool_op);
1900 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1901 }
1902 EXPORT_SYMBOL(ceph_osdc_stop);
1903 
1904 /*
1905  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1906  * *plen.  Return number of bytes read, or error.
1907  */
1908 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1909 			struct ceph_vino vino, struct ceph_file_layout *layout,
1910 			u64 off, u64 *plen,
1911 			u32 truncate_seq, u64 truncate_size,
1912 			struct page **pages, int num_pages, int page_align)
1913 {
1914 	struct ceph_osd_request *req;
1915 	int rc = 0;
1916 
1917 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1918 	     vino.snap, off, *plen);
1919 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1920 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1921 				    NULL, 0, truncate_seq, truncate_size, NULL,
1922 				    false, 1, page_align);
1923 	if (!req)
1924 		return -ENOMEM;
1925 
1926 	/* it may be a short read due to an object boundary */
1927 	req->r_pages = pages;
1928 
1929 	dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1930 	     off, *plen, req->r_num_pages, page_align);
1931 
1932 	rc = ceph_osdc_start_request(osdc, req, false);
1933 	if (!rc)
1934 		rc = ceph_osdc_wait_request(osdc, req);
1935 
1936 	ceph_osdc_put_request(req);
1937 	dout("readpages result %d\n", rc);
1938 	return rc;
1939 }
1940 EXPORT_SYMBOL(ceph_osdc_readpages);
1941 
1942 /*
1943  * do a synchronous write on N pages
1944  */
1945 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1946 			 struct ceph_file_layout *layout,
1947 			 struct ceph_snap_context *snapc,
1948 			 u64 off, u64 len,
1949 			 u32 truncate_seq, u64 truncate_size,
1950 			 struct timespec *mtime,
1951 			 struct page **pages, int num_pages,
1952 			 int flags, int do_sync, bool nofail)
1953 {
1954 	struct ceph_osd_request *req;
1955 	int rc = 0;
1956 	int page_align = off & ~PAGE_MASK;
1957 
1958 	BUG_ON(vino.snap != CEPH_NOSNAP);
1959 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1960 				    CEPH_OSD_OP_WRITE,
1961 				    flags | CEPH_OSD_FLAG_ONDISK |
1962 					    CEPH_OSD_FLAG_WRITE,
1963 				    snapc, do_sync,
1964 				    truncate_seq, truncate_size, mtime,
1965 				    nofail, 1, page_align);
1966 	if (!req)
1967 		return -ENOMEM;
1968 
1969 	/* it may be a short write due to an object boundary */
1970 	req->r_pages = pages;
1971 	dout("writepages %llu~%llu (%d pages)\n", off, len,
1972 	     req->r_num_pages);
1973 
1974 	rc = ceph_osdc_start_request(osdc, req, nofail);
1975 	if (!rc)
1976 		rc = ceph_osdc_wait_request(osdc, req);
1977 
1978 	ceph_osdc_put_request(req);
1979 	if (rc == 0)
1980 		rc = len;
1981 	dout("writepages result %d\n", rc);
1982 	return rc;
1983 }
1984 EXPORT_SYMBOL(ceph_osdc_writepages);
1985 
1986 /*
1987  * handle incoming message
1988  */
1989 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1990 {
1991 	struct ceph_osd *osd = con->private;
1992 	struct ceph_osd_client *osdc;
1993 	int type = le16_to_cpu(msg->hdr.type);
1994 
1995 	if (!osd)
1996 		goto out;
1997 	osdc = osd->o_osdc;
1998 
1999 	switch (type) {
2000 	case CEPH_MSG_OSD_MAP:
2001 		ceph_osdc_handle_map(osdc, msg);
2002 		break;
2003 	case CEPH_MSG_OSD_OPREPLY:
2004 		handle_reply(osdc, msg, con);
2005 		break;
2006 	case CEPH_MSG_WATCH_NOTIFY:
2007 		handle_watch_notify(osdc, msg);
2008 		break;
2009 
2010 	default:
2011 		pr_err("received unknown message type %d %s\n", type,
2012 		       ceph_msg_type_name(type));
2013 	}
2014 out:
2015 	ceph_msg_put(msg);
2016 }
2017 
2018 /*
2019  * lookup and return message for incoming reply.  set up reply message
2020  * pages.
2021  */
2022 static struct ceph_msg *get_reply(struct ceph_connection *con,
2023 				  struct ceph_msg_header *hdr,
2024 				  int *skip)
2025 {
2026 	struct ceph_osd *osd = con->private;
2027 	struct ceph_osd_client *osdc = osd->o_osdc;
2028 	struct ceph_msg *m;
2029 	struct ceph_osd_request *req;
2030 	int front = le32_to_cpu(hdr->front_len);
2031 	int data_len = le32_to_cpu(hdr->data_len);
2032 	u64 tid;
2033 
2034 	tid = le64_to_cpu(hdr->tid);
2035 	mutex_lock(&osdc->request_mutex);
2036 	req = __lookup_request(osdc, tid);
2037 	if (!req) {
2038 		*skip = 1;
2039 		m = NULL;
2040 		dout("get_reply unknown tid %llu from osd%d\n", tid,
2041 		     osd->o_osd);
2042 		goto out;
2043 	}
2044 
2045 	if (req->r_con_filling_msg) {
2046 		dout("%s revoking msg %p from old con %p\n", __func__,
2047 		     req->r_reply, req->r_con_filling_msg);
2048 		ceph_msg_revoke_incoming(req->r_reply);
2049 		req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2050 		req->r_con_filling_msg = NULL;
2051 	}
2052 
2053 	if (front > req->r_reply->front.iov_len) {
2054 		pr_warning("get_reply front %d > preallocated %d\n",
2055 			   front, (int)req->r_reply->front.iov_len);
2056 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2057 		if (!m)
2058 			goto out;
2059 		ceph_msg_put(req->r_reply);
2060 		req->r_reply = m;
2061 	}
2062 	m = ceph_msg_get(req->r_reply);
2063 
2064 	if (data_len > 0) {
2065 		int want = calc_pages_for(req->r_page_alignment, data_len);
2066 
2067 		if (unlikely(req->r_num_pages < want)) {
2068 			pr_warning("tid %lld reply has %d bytes %d pages, we"
2069 				   " had only %d pages ready\n", tid, data_len,
2070 				   want, req->r_num_pages);
2071 			*skip = 1;
2072 			ceph_msg_put(m);
2073 			m = NULL;
2074 			goto out;
2075 		}
2076 		m->pages = req->r_pages;
2077 		m->nr_pages = req->r_num_pages;
2078 		m->page_alignment = req->r_page_alignment;
2079 #ifdef CONFIG_BLOCK
2080 		m->bio = req->r_bio;
2081 #endif
2082 	}
2083 	*skip = 0;
2084 	req->r_con_filling_msg = con->ops->get(con);
2085 	dout("get_reply tid %lld %p\n", tid, m);
2086 
2087 out:
2088 	mutex_unlock(&osdc->request_mutex);
2089 	return m;
2090 
2091 }
2092 
2093 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2094 				  struct ceph_msg_header *hdr,
2095 				  int *skip)
2096 {
2097 	struct ceph_osd *osd = con->private;
2098 	int type = le16_to_cpu(hdr->type);
2099 	int front = le32_to_cpu(hdr->front_len);
2100 
2101 	*skip = 0;
2102 	switch (type) {
2103 	case CEPH_MSG_OSD_MAP:
2104 	case CEPH_MSG_WATCH_NOTIFY:
2105 		return ceph_msg_new(type, front, GFP_NOFS, false);
2106 	case CEPH_MSG_OSD_OPREPLY:
2107 		return get_reply(con, hdr, skip);
2108 	default:
2109 		pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2110 			osd->o_osd);
2111 		*skip = 1;
2112 		return NULL;
2113 	}
2114 }
2115 
2116 /*
2117  * Wrappers to refcount containing ceph_osd struct
2118  */
2119 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2120 {
2121 	struct ceph_osd *osd = con->private;
2122 	if (get_osd(osd))
2123 		return con;
2124 	return NULL;
2125 }
2126 
2127 static void put_osd_con(struct ceph_connection *con)
2128 {
2129 	struct ceph_osd *osd = con->private;
2130 	put_osd(osd);
2131 }
2132 
2133 /*
2134  * authentication
2135  */
2136 /*
2137  * Note: returned pointer is the address of a structure that's
2138  * managed separately.  Caller must *not* attempt to free it.
2139  */
2140 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2141 					int *proto, int force_new)
2142 {
2143 	struct ceph_osd *o = con->private;
2144 	struct ceph_osd_client *osdc = o->o_osdc;
2145 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2146 	struct ceph_auth_handshake *auth = &o->o_auth;
2147 
2148 	if (force_new && auth->authorizer) {
2149 		if (ac->ops && ac->ops->destroy_authorizer)
2150 			ac->ops->destroy_authorizer(ac, auth->authorizer);
2151 		auth->authorizer = NULL;
2152 	}
2153 	if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2154 		int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2155 							auth);
2156 		if (ret)
2157 			return ERR_PTR(ret);
2158 	}
2159 	*proto = ac->protocol;
2160 
2161 	return auth;
2162 }
2163 
2164 
2165 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2166 {
2167 	struct ceph_osd *o = con->private;
2168 	struct ceph_osd_client *osdc = o->o_osdc;
2169 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2170 
2171 	/*
2172 	 * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2173 	 * XXX which do we do:  succeed or fail?
2174 	 */
2175 	return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2176 }
2177 
2178 static int invalidate_authorizer(struct ceph_connection *con)
2179 {
2180 	struct ceph_osd *o = con->private;
2181 	struct ceph_osd_client *osdc = o->o_osdc;
2182 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2183 
2184 	if (ac->ops && ac->ops->invalidate_authorizer)
2185 		ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2186 
2187 	return ceph_monc_validate_auth(&osdc->client->monc);
2188 }
2189 
2190 static const struct ceph_connection_operations osd_con_ops = {
2191 	.get = get_osd_con,
2192 	.put = put_osd_con,
2193 	.dispatch = dispatch,
2194 	.get_authorizer = get_authorizer,
2195 	.verify_authorizer_reply = verify_authorizer_reply,
2196 	.invalidate_authorizer = invalidate_authorizer,
2197 	.alloc_msg = alloc_msg,
2198 	.fault = osd_reset,
2199 };
2200