xref: /openbmc/linux/net/ceph/osd_client.c (revision 867a0e05)
1 
2 #include <linux/ceph/ceph_debug.h>
3 
4 #include <linux/module.h>
5 #include <linux/err.h>
6 #include <linux/highmem.h>
7 #include <linux/mm.h>
8 #include <linux/pagemap.h>
9 #include <linux/slab.h>
10 #include <linux/uaccess.h>
11 #ifdef CONFIG_BLOCK
12 #include <linux/bio.h>
13 #endif
14 
15 #include <linux/ceph/libceph.h>
16 #include <linux/ceph/osd_client.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/pagelist.h>
21 
22 #define OSD_OP_FRONT_LEN	4096
23 #define OSD_OPREPLY_FRONT_LEN	512
24 
25 static struct kmem_cache	*ceph_osd_request_cache;
26 
27 static const struct ceph_connection_operations osd_con_ops;
28 
29 static void __send_queued(struct ceph_osd_client *osdc);
30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
31 static void __register_request(struct ceph_osd_client *osdc,
32 			       struct ceph_osd_request *req);
33 static void __unregister_request(struct ceph_osd_client *osdc,
34 				 struct ceph_osd_request *req);
35 static void __unregister_linger_request(struct ceph_osd_client *osdc,
36 					struct ceph_osd_request *req);
37 static void __enqueue_request(struct ceph_osd_request *req);
38 static void __send_request(struct ceph_osd_client *osdc,
39 			   struct ceph_osd_request *req);
40 
41 /*
42  * Implement client access to distributed object storage cluster.
43  *
44  * All data objects are stored within a cluster/cloud of OSDs, or
45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
47  * remote daemons serving up and coordinating consistent and safe
48  * access to storage.
49  *
50  * Cluster membership and the mapping of data objects onto storage devices
51  * are described by the osd map.
52  *
53  * We keep track of pending OSD requests (read, write), resubmit
54  * requests to different OSDs when the cluster topology/data layout
55  * change, or retry the affected requests when the communications
56  * channel with an OSD is reset.
57  */
58 
59 /*
60  * calculate the mapping of a file extent onto an object, and fill out the
61  * request accordingly.  shorten extent as necessary if it crosses an
62  * object boundary.
63  *
64  * fill osd op in request message.
65  */
66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
67 			u64 *objnum, u64 *objoff, u64 *objlen)
68 {
69 	u64 orig_len = *plen;
70 	int r;
71 
72 	/* object extent? */
73 	r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
74 					  objoff, objlen);
75 	if (r < 0)
76 		return r;
77 	if (*objlen < orig_len) {
78 		*plen = *objlen;
79 		dout(" skipping last %llu, final file extent %llu~%llu\n",
80 		     orig_len - *plen, off, *plen);
81 	}
82 
83 	dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
84 
85 	return 0;
86 }
87 
88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
89 {
90 	memset(osd_data, 0, sizeof (*osd_data));
91 	osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
92 }
93 
94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
95 			struct page **pages, u64 length, u32 alignment,
96 			bool pages_from_pool, bool own_pages)
97 {
98 	osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
99 	osd_data->pages = pages;
100 	osd_data->length = length;
101 	osd_data->alignment = alignment;
102 	osd_data->pages_from_pool = pages_from_pool;
103 	osd_data->own_pages = own_pages;
104 }
105 
106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
107 			struct ceph_pagelist *pagelist)
108 {
109 	osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
110 	osd_data->pagelist = pagelist;
111 }
112 
113 #ifdef CONFIG_BLOCK
114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
115 			struct bio *bio, size_t bio_length)
116 {
117 	osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
118 	osd_data->bio = bio;
119 	osd_data->bio_length = bio_length;
120 }
121 #endif /* CONFIG_BLOCK */
122 
123 #define osd_req_op_data(oreq, whch, typ, fld)	\
124 	({						\
125 		BUG_ON(whch >= (oreq)->r_num_ops);	\
126 		&(oreq)->r_ops[whch].typ.fld;		\
127 	})
128 
129 static struct ceph_osd_data *
130 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
131 {
132 	BUG_ON(which >= osd_req->r_num_ops);
133 
134 	return &osd_req->r_ops[which].raw_data_in;
135 }
136 
137 struct ceph_osd_data *
138 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
139 			unsigned int which)
140 {
141 	return osd_req_op_data(osd_req, which, extent, osd_data);
142 }
143 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
144 
145 struct ceph_osd_data *
146 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
147 			unsigned int which)
148 {
149 	return osd_req_op_data(osd_req, which, cls, response_data);
150 }
151 EXPORT_SYMBOL(osd_req_op_cls_response_data);	/* ??? */
152 
153 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
154 			unsigned int which, struct page **pages,
155 			u64 length, u32 alignment,
156 			bool pages_from_pool, bool own_pages)
157 {
158 	struct ceph_osd_data *osd_data;
159 
160 	osd_data = osd_req_op_raw_data_in(osd_req, which);
161 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
162 				pages_from_pool, own_pages);
163 }
164 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
165 
166 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
167 			unsigned int which, struct page **pages,
168 			u64 length, u32 alignment,
169 			bool pages_from_pool, bool own_pages)
170 {
171 	struct ceph_osd_data *osd_data;
172 
173 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
174 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
175 				pages_from_pool, own_pages);
176 }
177 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
178 
179 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
180 			unsigned int which, struct ceph_pagelist *pagelist)
181 {
182 	struct ceph_osd_data *osd_data;
183 
184 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
185 	ceph_osd_data_pagelist_init(osd_data, pagelist);
186 }
187 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
188 
189 #ifdef CONFIG_BLOCK
190 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
191 			unsigned int which, struct bio *bio, size_t bio_length)
192 {
193 	struct ceph_osd_data *osd_data;
194 
195 	osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
196 	ceph_osd_data_bio_init(osd_data, bio, bio_length);
197 }
198 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
199 #endif /* CONFIG_BLOCK */
200 
201 static void osd_req_op_cls_request_info_pagelist(
202 			struct ceph_osd_request *osd_req,
203 			unsigned int which, struct ceph_pagelist *pagelist)
204 {
205 	struct ceph_osd_data *osd_data;
206 
207 	osd_data = osd_req_op_data(osd_req, which, cls, request_info);
208 	ceph_osd_data_pagelist_init(osd_data, pagelist);
209 }
210 
211 void osd_req_op_cls_request_data_pagelist(
212 			struct ceph_osd_request *osd_req,
213 			unsigned int which, struct ceph_pagelist *pagelist)
214 {
215 	struct ceph_osd_data *osd_data;
216 
217 	osd_data = osd_req_op_data(osd_req, which, cls, request_data);
218 	ceph_osd_data_pagelist_init(osd_data, pagelist);
219 }
220 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
221 
222 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
223 			unsigned int which, struct page **pages, u64 length,
224 			u32 alignment, bool pages_from_pool, bool own_pages)
225 {
226 	struct ceph_osd_data *osd_data;
227 
228 	osd_data = osd_req_op_data(osd_req, which, cls, request_data);
229 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
230 				pages_from_pool, own_pages);
231 }
232 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
233 
234 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
235 			unsigned int which, struct page **pages, u64 length,
236 			u32 alignment, bool pages_from_pool, bool own_pages)
237 {
238 	struct ceph_osd_data *osd_data;
239 
240 	osd_data = osd_req_op_data(osd_req, which, cls, response_data);
241 	ceph_osd_data_pages_init(osd_data, pages, length, alignment,
242 				pages_from_pool, own_pages);
243 }
244 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
245 
246 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
247 {
248 	switch (osd_data->type) {
249 	case CEPH_OSD_DATA_TYPE_NONE:
250 		return 0;
251 	case CEPH_OSD_DATA_TYPE_PAGES:
252 		return osd_data->length;
253 	case CEPH_OSD_DATA_TYPE_PAGELIST:
254 		return (u64)osd_data->pagelist->length;
255 #ifdef CONFIG_BLOCK
256 	case CEPH_OSD_DATA_TYPE_BIO:
257 		return (u64)osd_data->bio_length;
258 #endif /* CONFIG_BLOCK */
259 	default:
260 		WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
261 		return 0;
262 	}
263 }
264 
265 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
266 {
267 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
268 		int num_pages;
269 
270 		num_pages = calc_pages_for((u64)osd_data->alignment,
271 						(u64)osd_data->length);
272 		ceph_release_page_vector(osd_data->pages, num_pages);
273 	}
274 	ceph_osd_data_init(osd_data);
275 }
276 
277 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
278 			unsigned int which)
279 {
280 	struct ceph_osd_req_op *op;
281 
282 	BUG_ON(which >= osd_req->r_num_ops);
283 	op = &osd_req->r_ops[which];
284 
285 	switch (op->op) {
286 	case CEPH_OSD_OP_READ:
287 	case CEPH_OSD_OP_WRITE:
288 		ceph_osd_data_release(&op->extent.osd_data);
289 		break;
290 	case CEPH_OSD_OP_CALL:
291 		ceph_osd_data_release(&op->cls.request_info);
292 		ceph_osd_data_release(&op->cls.request_data);
293 		ceph_osd_data_release(&op->cls.response_data);
294 		break;
295 	case CEPH_OSD_OP_SETXATTR:
296 	case CEPH_OSD_OP_CMPXATTR:
297 		ceph_osd_data_release(&op->xattr.osd_data);
298 		break;
299 	default:
300 		break;
301 	}
302 }
303 
304 /*
305  * requests
306  */
307 static void ceph_osdc_release_request(struct kref *kref)
308 {
309 	struct ceph_osd_request *req = container_of(kref,
310 					    struct ceph_osd_request, r_kref);
311 	unsigned int which;
312 
313 	dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
314 	     req->r_request, req->r_reply);
315 	WARN_ON(!RB_EMPTY_NODE(&req->r_node));
316 	WARN_ON(!list_empty(&req->r_req_lru_item));
317 	WARN_ON(!list_empty(&req->r_osd_item));
318 	WARN_ON(!list_empty(&req->r_linger_item));
319 	WARN_ON(!list_empty(&req->r_linger_osd_item));
320 	WARN_ON(req->r_osd);
321 
322 	if (req->r_request)
323 		ceph_msg_put(req->r_request);
324 	if (req->r_reply) {
325 		ceph_msg_revoke_incoming(req->r_reply);
326 		ceph_msg_put(req->r_reply);
327 	}
328 
329 	for (which = 0; which < req->r_num_ops; which++)
330 		osd_req_op_data_release(req, which);
331 
332 	ceph_put_snap_context(req->r_snapc);
333 	if (req->r_mempool)
334 		mempool_free(req, req->r_osdc->req_mempool);
335 	else
336 		kmem_cache_free(ceph_osd_request_cache, req);
337 
338 }
339 
340 void ceph_osdc_get_request(struct ceph_osd_request *req)
341 {
342 	dout("%s %p (was %d)\n", __func__, req,
343 	     atomic_read(&req->r_kref.refcount));
344 	kref_get(&req->r_kref);
345 }
346 EXPORT_SYMBOL(ceph_osdc_get_request);
347 
348 void ceph_osdc_put_request(struct ceph_osd_request *req)
349 {
350 	dout("%s %p (was %d)\n", __func__, req,
351 	     atomic_read(&req->r_kref.refcount));
352 	kref_put(&req->r_kref, ceph_osdc_release_request);
353 }
354 EXPORT_SYMBOL(ceph_osdc_put_request);
355 
356 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
357 					       struct ceph_snap_context *snapc,
358 					       unsigned int num_ops,
359 					       bool use_mempool,
360 					       gfp_t gfp_flags)
361 {
362 	struct ceph_osd_request *req;
363 	struct ceph_msg *msg;
364 	size_t msg_size;
365 
366 	BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
367 	BUG_ON(num_ops > CEPH_OSD_MAX_OP);
368 
369 	msg_size = 4 + 4 + 8 + 8 + 4+8;
370 	msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
371 	msg_size += 1 + 8 + 4 + 4;     /* pg_t */
372 	msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
373 	msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
374 	msg_size += 8;  /* snapid */
375 	msg_size += 8;  /* snap_seq */
376 	msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
377 	msg_size += 4;
378 
379 	if (use_mempool) {
380 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
381 		memset(req, 0, sizeof(*req));
382 	} else {
383 		req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
384 	}
385 	if (req == NULL)
386 		return NULL;
387 
388 	req->r_osdc = osdc;
389 	req->r_mempool = use_mempool;
390 	req->r_num_ops = num_ops;
391 
392 	kref_init(&req->r_kref);
393 	init_completion(&req->r_completion);
394 	init_completion(&req->r_safe_completion);
395 	RB_CLEAR_NODE(&req->r_node);
396 	INIT_LIST_HEAD(&req->r_unsafe_item);
397 	INIT_LIST_HEAD(&req->r_linger_item);
398 	INIT_LIST_HEAD(&req->r_linger_osd_item);
399 	INIT_LIST_HEAD(&req->r_req_lru_item);
400 	INIT_LIST_HEAD(&req->r_osd_item);
401 
402 	req->r_base_oloc.pool = -1;
403 	req->r_target_oloc.pool = -1;
404 
405 	/* create reply message */
406 	if (use_mempool)
407 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
408 	else
409 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
410 				   OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
411 	if (!msg) {
412 		ceph_osdc_put_request(req);
413 		return NULL;
414 	}
415 	req->r_reply = msg;
416 
417 	/* create request message; allow space for oid */
418 	if (use_mempool)
419 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
420 	else
421 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
422 	if (!msg) {
423 		ceph_osdc_put_request(req);
424 		return NULL;
425 	}
426 
427 	memset(msg->front.iov_base, 0, msg->front.iov_len);
428 
429 	req->r_request = msg;
430 
431 	return req;
432 }
433 EXPORT_SYMBOL(ceph_osdc_alloc_request);
434 
435 static bool osd_req_opcode_valid(u16 opcode)
436 {
437 	switch (opcode) {
438 #define GENERATE_CASE(op, opcode, str)	case CEPH_OSD_OP_##op: return true;
439 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
440 #undef GENERATE_CASE
441 	default:
442 		return false;
443 	}
444 }
445 
446 /*
447  * This is an osd op init function for opcodes that have no data or
448  * other information associated with them.  It also serves as a
449  * common init routine for all the other init functions, below.
450  */
451 static struct ceph_osd_req_op *
452 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
453 				u16 opcode)
454 {
455 	struct ceph_osd_req_op *op;
456 
457 	BUG_ON(which >= osd_req->r_num_ops);
458 	BUG_ON(!osd_req_opcode_valid(opcode));
459 
460 	op = &osd_req->r_ops[which];
461 	memset(op, 0, sizeof (*op));
462 	op->op = opcode;
463 
464 	return op;
465 }
466 
467 void osd_req_op_init(struct ceph_osd_request *osd_req,
468 				unsigned int which, u16 opcode)
469 {
470 	(void)_osd_req_op_init(osd_req, which, opcode);
471 }
472 EXPORT_SYMBOL(osd_req_op_init);
473 
474 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
475 				unsigned int which, u16 opcode,
476 				u64 offset, u64 length,
477 				u64 truncate_size, u32 truncate_seq)
478 {
479 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
480 	size_t payload_len = 0;
481 
482 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
483 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE);
484 
485 	op->extent.offset = offset;
486 	op->extent.length = length;
487 	op->extent.truncate_size = truncate_size;
488 	op->extent.truncate_seq = truncate_seq;
489 	if (opcode == CEPH_OSD_OP_WRITE)
490 		payload_len += length;
491 
492 	op->payload_len = payload_len;
493 }
494 EXPORT_SYMBOL(osd_req_op_extent_init);
495 
496 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
497 				unsigned int which, u64 length)
498 {
499 	struct ceph_osd_req_op *op;
500 	u64 previous;
501 
502 	BUG_ON(which >= osd_req->r_num_ops);
503 	op = &osd_req->r_ops[which];
504 	previous = op->extent.length;
505 
506 	if (length == previous)
507 		return;		/* Nothing to do */
508 	BUG_ON(length > previous);
509 
510 	op->extent.length = length;
511 	op->payload_len -= previous - length;
512 }
513 EXPORT_SYMBOL(osd_req_op_extent_update);
514 
515 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
516 			u16 opcode, const char *class, const char *method)
517 {
518 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
519 	struct ceph_pagelist *pagelist;
520 	size_t payload_len = 0;
521 	size_t size;
522 
523 	BUG_ON(opcode != CEPH_OSD_OP_CALL);
524 
525 	pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
526 	BUG_ON(!pagelist);
527 	ceph_pagelist_init(pagelist);
528 
529 	op->cls.class_name = class;
530 	size = strlen(class);
531 	BUG_ON(size > (size_t) U8_MAX);
532 	op->cls.class_len = size;
533 	ceph_pagelist_append(pagelist, class, size);
534 	payload_len += size;
535 
536 	op->cls.method_name = method;
537 	size = strlen(method);
538 	BUG_ON(size > (size_t) U8_MAX);
539 	op->cls.method_len = size;
540 	ceph_pagelist_append(pagelist, method, size);
541 	payload_len += size;
542 
543 	osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
544 
545 	op->cls.argc = 0;	/* currently unused */
546 
547 	op->payload_len = payload_len;
548 }
549 EXPORT_SYMBOL(osd_req_op_cls_init);
550 
551 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
552 			  u16 opcode, const char *name, const void *value,
553 			  size_t size, u8 cmp_op, u8 cmp_mode)
554 {
555 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
556 	struct ceph_pagelist *pagelist;
557 	size_t payload_len;
558 
559 	BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
560 
561 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
562 	if (!pagelist)
563 		return -ENOMEM;
564 
565 	ceph_pagelist_init(pagelist);
566 
567 	payload_len = strlen(name);
568 	op->xattr.name_len = payload_len;
569 	ceph_pagelist_append(pagelist, name, payload_len);
570 
571 	op->xattr.value_len = size;
572 	ceph_pagelist_append(pagelist, value, size);
573 	payload_len += size;
574 
575 	op->xattr.cmp_op = cmp_op;
576 	op->xattr.cmp_mode = cmp_mode;
577 
578 	ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
579 	op->payload_len = payload_len;
580 	return 0;
581 }
582 EXPORT_SYMBOL(osd_req_op_xattr_init);
583 
584 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
585 				unsigned int which, u16 opcode,
586 				u64 cookie, u64 version, int flag)
587 {
588 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
589 
590 	BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
591 
592 	op->watch.cookie = cookie;
593 	op->watch.ver = version;
594 	if (opcode == CEPH_OSD_OP_WATCH && flag)
595 		op->watch.flag = (u8)1;
596 }
597 EXPORT_SYMBOL(osd_req_op_watch_init);
598 
599 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
600 				unsigned int which,
601 				u64 expected_object_size,
602 				u64 expected_write_size)
603 {
604 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
605 						      CEPH_OSD_OP_SETALLOCHINT);
606 
607 	op->alloc_hint.expected_object_size = expected_object_size;
608 	op->alloc_hint.expected_write_size = expected_write_size;
609 
610 	/*
611 	 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
612 	 * not worth a feature bit.  Set FAILOK per-op flag to make
613 	 * sure older osds don't trip over an unsupported opcode.
614 	 */
615 	op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
616 }
617 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
618 
619 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
620 				struct ceph_osd_data *osd_data)
621 {
622 	u64 length = ceph_osd_data_length(osd_data);
623 
624 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
625 		BUG_ON(length > (u64) SIZE_MAX);
626 		if (length)
627 			ceph_msg_data_add_pages(msg, osd_data->pages,
628 					length, osd_data->alignment);
629 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
630 		BUG_ON(!length);
631 		ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
632 #ifdef CONFIG_BLOCK
633 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
634 		ceph_msg_data_add_bio(msg, osd_data->bio, length);
635 #endif
636 	} else {
637 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
638 	}
639 }
640 
641 static u64 osd_req_encode_op(struct ceph_osd_request *req,
642 			      struct ceph_osd_op *dst, unsigned int which)
643 {
644 	struct ceph_osd_req_op *src;
645 	struct ceph_osd_data *osd_data;
646 	u64 request_data_len = 0;
647 	u64 data_length;
648 
649 	BUG_ON(which >= req->r_num_ops);
650 	src = &req->r_ops[which];
651 	if (WARN_ON(!osd_req_opcode_valid(src->op))) {
652 		pr_err("unrecognized osd opcode %d\n", src->op);
653 
654 		return 0;
655 	}
656 
657 	switch (src->op) {
658 	case CEPH_OSD_OP_STAT:
659 		osd_data = &src->raw_data_in;
660 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
661 		break;
662 	case CEPH_OSD_OP_READ:
663 	case CEPH_OSD_OP_WRITE:
664 	case CEPH_OSD_OP_ZERO:
665 	case CEPH_OSD_OP_TRUNCATE:
666 		if (src->op == CEPH_OSD_OP_WRITE)
667 			request_data_len = src->extent.length;
668 		dst->extent.offset = cpu_to_le64(src->extent.offset);
669 		dst->extent.length = cpu_to_le64(src->extent.length);
670 		dst->extent.truncate_size =
671 			cpu_to_le64(src->extent.truncate_size);
672 		dst->extent.truncate_seq =
673 			cpu_to_le32(src->extent.truncate_seq);
674 		osd_data = &src->extent.osd_data;
675 		if (src->op == CEPH_OSD_OP_WRITE)
676 			ceph_osdc_msg_data_add(req->r_request, osd_data);
677 		else
678 			ceph_osdc_msg_data_add(req->r_reply, osd_data);
679 		break;
680 	case CEPH_OSD_OP_CALL:
681 		dst->cls.class_len = src->cls.class_len;
682 		dst->cls.method_len = src->cls.method_len;
683 		osd_data = &src->cls.request_info;
684 		ceph_osdc_msg_data_add(req->r_request, osd_data);
685 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
686 		request_data_len = osd_data->pagelist->length;
687 
688 		osd_data = &src->cls.request_data;
689 		data_length = ceph_osd_data_length(osd_data);
690 		if (data_length) {
691 			BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
692 			dst->cls.indata_len = cpu_to_le32(data_length);
693 			ceph_osdc_msg_data_add(req->r_request, osd_data);
694 			src->payload_len += data_length;
695 			request_data_len += data_length;
696 		}
697 		osd_data = &src->cls.response_data;
698 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
699 		break;
700 	case CEPH_OSD_OP_STARTSYNC:
701 		break;
702 	case CEPH_OSD_OP_NOTIFY_ACK:
703 	case CEPH_OSD_OP_WATCH:
704 		dst->watch.cookie = cpu_to_le64(src->watch.cookie);
705 		dst->watch.ver = cpu_to_le64(src->watch.ver);
706 		dst->watch.flag = src->watch.flag;
707 		break;
708 	case CEPH_OSD_OP_SETALLOCHINT:
709 		dst->alloc_hint.expected_object_size =
710 		    cpu_to_le64(src->alloc_hint.expected_object_size);
711 		dst->alloc_hint.expected_write_size =
712 		    cpu_to_le64(src->alloc_hint.expected_write_size);
713 		break;
714 	case CEPH_OSD_OP_SETXATTR:
715 	case CEPH_OSD_OP_CMPXATTR:
716 		dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
717 		dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
718 		dst->xattr.cmp_op = src->xattr.cmp_op;
719 		dst->xattr.cmp_mode = src->xattr.cmp_mode;
720 		osd_data = &src->xattr.osd_data;
721 		ceph_osdc_msg_data_add(req->r_request, osd_data);
722 		request_data_len = osd_data->pagelist->length;
723 		break;
724 	case CEPH_OSD_OP_CREATE:
725 	case CEPH_OSD_OP_DELETE:
726 		break;
727 	default:
728 		pr_err("unsupported osd opcode %s\n",
729 			ceph_osd_op_name(src->op));
730 		WARN_ON(1);
731 
732 		return 0;
733 	}
734 
735 	dst->op = cpu_to_le16(src->op);
736 	dst->flags = cpu_to_le32(src->flags);
737 	dst->payload_len = cpu_to_le32(src->payload_len);
738 
739 	return request_data_len;
740 }
741 
742 /*
743  * build new request AND message, calculate layout, and adjust file
744  * extent as needed.
745  *
746  * if the file was recently truncated, we include information about its
747  * old and new size so that the object can be updated appropriately.  (we
748  * avoid synchronously deleting truncated objects because it's slow.)
749  *
750  * if @do_sync, include a 'startsync' command so that the osd will flush
751  * data quickly.
752  */
753 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
754 					       struct ceph_file_layout *layout,
755 					       struct ceph_vino vino,
756 					       u64 off, u64 *plen,
757 					       unsigned int which, int num_ops,
758 					       int opcode, int flags,
759 					       struct ceph_snap_context *snapc,
760 					       u32 truncate_seq,
761 					       u64 truncate_size,
762 					       bool use_mempool)
763 {
764 	struct ceph_osd_request *req;
765 	u64 objnum = 0;
766 	u64 objoff = 0;
767 	u64 objlen = 0;
768 	int r;
769 
770 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
771 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
772 	       opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
773 
774 	req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
775 					GFP_NOFS);
776 	if (!req)
777 		return ERR_PTR(-ENOMEM);
778 
779 	req->r_flags = flags;
780 
781 	/* calculate max write size */
782 	r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
783 	if (r < 0) {
784 		ceph_osdc_put_request(req);
785 		return ERR_PTR(r);
786 	}
787 
788 	if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
789 		osd_req_op_init(req, which, opcode);
790 	} else {
791 		u32 object_size = le32_to_cpu(layout->fl_object_size);
792 		u32 object_base = off - objoff;
793 		if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
794 			if (truncate_size <= object_base) {
795 				truncate_size = 0;
796 			} else {
797 				truncate_size -= object_base;
798 				if (truncate_size > object_size)
799 					truncate_size = object_size;
800 			}
801 		}
802 		osd_req_op_extent_init(req, which, opcode, objoff, objlen,
803 				       truncate_size, truncate_seq);
804 	}
805 
806 	req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
807 
808 	snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
809 		 "%llx.%08llx", vino.ino, objnum);
810 	req->r_base_oid.name_len = strlen(req->r_base_oid.name);
811 
812 	return req;
813 }
814 EXPORT_SYMBOL(ceph_osdc_new_request);
815 
816 /*
817  * We keep osd requests in an rbtree, sorted by ->r_tid.
818  */
819 static void __insert_request(struct ceph_osd_client *osdc,
820 			     struct ceph_osd_request *new)
821 {
822 	struct rb_node **p = &osdc->requests.rb_node;
823 	struct rb_node *parent = NULL;
824 	struct ceph_osd_request *req = NULL;
825 
826 	while (*p) {
827 		parent = *p;
828 		req = rb_entry(parent, struct ceph_osd_request, r_node);
829 		if (new->r_tid < req->r_tid)
830 			p = &(*p)->rb_left;
831 		else if (new->r_tid > req->r_tid)
832 			p = &(*p)->rb_right;
833 		else
834 			BUG();
835 	}
836 
837 	rb_link_node(&new->r_node, parent, p);
838 	rb_insert_color(&new->r_node, &osdc->requests);
839 }
840 
841 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
842 						 u64 tid)
843 {
844 	struct ceph_osd_request *req;
845 	struct rb_node *n = osdc->requests.rb_node;
846 
847 	while (n) {
848 		req = rb_entry(n, struct ceph_osd_request, r_node);
849 		if (tid < req->r_tid)
850 			n = n->rb_left;
851 		else if (tid > req->r_tid)
852 			n = n->rb_right;
853 		else
854 			return req;
855 	}
856 	return NULL;
857 }
858 
859 static struct ceph_osd_request *
860 __lookup_request_ge(struct ceph_osd_client *osdc,
861 		    u64 tid)
862 {
863 	struct ceph_osd_request *req;
864 	struct rb_node *n = osdc->requests.rb_node;
865 
866 	while (n) {
867 		req = rb_entry(n, struct ceph_osd_request, r_node);
868 		if (tid < req->r_tid) {
869 			if (!n->rb_left)
870 				return req;
871 			n = n->rb_left;
872 		} else if (tid > req->r_tid) {
873 			n = n->rb_right;
874 		} else {
875 			return req;
876 		}
877 	}
878 	return NULL;
879 }
880 
881 static void __kick_linger_request(struct ceph_osd_request *req)
882 {
883 	struct ceph_osd_client *osdc = req->r_osdc;
884 	struct ceph_osd *osd = req->r_osd;
885 
886 	/*
887 	 * Linger requests need to be resent with a new tid to avoid
888 	 * the dup op detection logic on the OSDs.  Achieve this with
889 	 * a re-register dance instead of open-coding.
890 	 */
891 	ceph_osdc_get_request(req);
892 	if (!list_empty(&req->r_linger_item))
893 		__unregister_linger_request(osdc, req);
894 	else
895 		__unregister_request(osdc, req);
896 	__register_request(osdc, req);
897 	ceph_osdc_put_request(req);
898 
899 	/*
900 	 * Unless request has been registered as both normal and
901 	 * lingering, __unregister{,_linger}_request clears r_osd.
902 	 * However, here we need to preserve r_osd to make sure we
903 	 * requeue on the same OSD.
904 	 */
905 	WARN_ON(req->r_osd || !osd);
906 	req->r_osd = osd;
907 
908 	dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
909 	__enqueue_request(req);
910 }
911 
912 /*
913  * Resubmit requests pending on the given osd.
914  */
915 static void __kick_osd_requests(struct ceph_osd_client *osdc,
916 				struct ceph_osd *osd)
917 {
918 	struct ceph_osd_request *req, *nreq;
919 	LIST_HEAD(resend);
920 	LIST_HEAD(resend_linger);
921 	int err;
922 
923 	dout("%s osd%d\n", __func__, osd->o_osd);
924 	err = __reset_osd(osdc, osd);
925 	if (err)
926 		return;
927 
928 	/*
929 	 * Build up a list of requests to resend by traversing the
930 	 * osd's list of requests.  Requests for a given object are
931 	 * sent in tid order, and that is also the order they're
932 	 * kept on this list.  Therefore all requests that are in
933 	 * flight will be found first, followed by all requests that
934 	 * have not yet been sent.  And to resend requests while
935 	 * preserving this order we will want to put any sent
936 	 * requests back on the front of the osd client's unsent
937 	 * list.
938 	 *
939 	 * So we build a separate ordered list of already-sent
940 	 * requests for the affected osd and splice it onto the
941 	 * front of the osd client's unsent list.  Once we've seen a
942 	 * request that has not yet been sent we're done.  Those
943 	 * requests are already sitting right where they belong.
944 	 */
945 	list_for_each_entry(req, &osd->o_requests, r_osd_item) {
946 		if (!req->r_sent)
947 			break;
948 
949 		if (!req->r_linger) {
950 			dout("%s requeueing %p tid %llu\n", __func__, req,
951 			     req->r_tid);
952 			list_move_tail(&req->r_req_lru_item, &resend);
953 			req->r_flags |= CEPH_OSD_FLAG_RETRY;
954 		} else {
955 			list_move_tail(&req->r_req_lru_item, &resend_linger);
956 		}
957 	}
958 	list_splice(&resend, &osdc->req_unsent);
959 
960 	/*
961 	 * Both registered and not yet registered linger requests are
962 	 * enqueued with a new tid on the same OSD.  We add/move them
963 	 * to req_unsent/o_requests at the end to keep things in tid
964 	 * order.
965 	 */
966 	list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
967 				 r_linger_osd_item) {
968 		WARN_ON(!list_empty(&req->r_req_lru_item));
969 		__kick_linger_request(req);
970 	}
971 
972 	list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
973 		__kick_linger_request(req);
974 }
975 
976 /*
977  * If the osd connection drops, we need to resubmit all requests.
978  */
979 static void osd_reset(struct ceph_connection *con)
980 {
981 	struct ceph_osd *osd = con->private;
982 	struct ceph_osd_client *osdc;
983 
984 	if (!osd)
985 		return;
986 	dout("osd_reset osd%d\n", osd->o_osd);
987 	osdc = osd->o_osdc;
988 	down_read(&osdc->map_sem);
989 	mutex_lock(&osdc->request_mutex);
990 	__kick_osd_requests(osdc, osd);
991 	__send_queued(osdc);
992 	mutex_unlock(&osdc->request_mutex);
993 	up_read(&osdc->map_sem);
994 }
995 
996 /*
997  * Track open sessions with osds.
998  */
999 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1000 {
1001 	struct ceph_osd *osd;
1002 
1003 	osd = kzalloc(sizeof(*osd), GFP_NOFS);
1004 	if (!osd)
1005 		return NULL;
1006 
1007 	atomic_set(&osd->o_ref, 1);
1008 	osd->o_osdc = osdc;
1009 	osd->o_osd = onum;
1010 	RB_CLEAR_NODE(&osd->o_node);
1011 	INIT_LIST_HEAD(&osd->o_requests);
1012 	INIT_LIST_HEAD(&osd->o_linger_requests);
1013 	INIT_LIST_HEAD(&osd->o_osd_lru);
1014 	osd->o_incarnation = 1;
1015 
1016 	ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1017 
1018 	INIT_LIST_HEAD(&osd->o_keepalive_item);
1019 	return osd;
1020 }
1021 
1022 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1023 {
1024 	if (atomic_inc_not_zero(&osd->o_ref)) {
1025 		dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1026 		     atomic_read(&osd->o_ref));
1027 		return osd;
1028 	} else {
1029 		dout("get_osd %p FAIL\n", osd);
1030 		return NULL;
1031 	}
1032 }
1033 
1034 static void put_osd(struct ceph_osd *osd)
1035 {
1036 	dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1037 	     atomic_read(&osd->o_ref) - 1);
1038 	if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
1039 		struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1040 
1041 		ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1042 		kfree(osd);
1043 	}
1044 }
1045 
1046 /*
1047  * remove an osd from our map
1048  */
1049 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1050 {
1051 	dout("__remove_osd %p\n", osd);
1052 	WARN_ON(!list_empty(&osd->o_requests));
1053 	WARN_ON(!list_empty(&osd->o_linger_requests));
1054 
1055 	rb_erase(&osd->o_node, &osdc->osds);
1056 	list_del_init(&osd->o_osd_lru);
1057 	ceph_con_close(&osd->o_con);
1058 	put_osd(osd);
1059 }
1060 
1061 static void remove_all_osds(struct ceph_osd_client *osdc)
1062 {
1063 	dout("%s %p\n", __func__, osdc);
1064 	mutex_lock(&osdc->request_mutex);
1065 	while (!RB_EMPTY_ROOT(&osdc->osds)) {
1066 		struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1067 						struct ceph_osd, o_node);
1068 		__remove_osd(osdc, osd);
1069 	}
1070 	mutex_unlock(&osdc->request_mutex);
1071 }
1072 
1073 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1074 			      struct ceph_osd *osd)
1075 {
1076 	dout("%s %p\n", __func__, osd);
1077 	BUG_ON(!list_empty(&osd->o_osd_lru));
1078 
1079 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1080 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1081 }
1082 
1083 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1084 				  struct ceph_osd *osd)
1085 {
1086 	dout("%s %p\n", __func__, osd);
1087 
1088 	if (list_empty(&osd->o_requests) &&
1089 	    list_empty(&osd->o_linger_requests))
1090 		__move_osd_to_lru(osdc, osd);
1091 }
1092 
1093 static void __remove_osd_from_lru(struct ceph_osd *osd)
1094 {
1095 	dout("__remove_osd_from_lru %p\n", osd);
1096 	if (!list_empty(&osd->o_osd_lru))
1097 		list_del_init(&osd->o_osd_lru);
1098 }
1099 
1100 static void remove_old_osds(struct ceph_osd_client *osdc)
1101 {
1102 	struct ceph_osd *osd, *nosd;
1103 
1104 	dout("__remove_old_osds %p\n", osdc);
1105 	mutex_lock(&osdc->request_mutex);
1106 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1107 		if (time_before(jiffies, osd->lru_ttl))
1108 			break;
1109 		__remove_osd(osdc, osd);
1110 	}
1111 	mutex_unlock(&osdc->request_mutex);
1112 }
1113 
1114 /*
1115  * reset osd connect
1116  */
1117 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1118 {
1119 	struct ceph_entity_addr *peer_addr;
1120 
1121 	dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1122 	if (list_empty(&osd->o_requests) &&
1123 	    list_empty(&osd->o_linger_requests)) {
1124 		__remove_osd(osdc, osd);
1125 
1126 		return -ENODEV;
1127 	}
1128 
1129 	peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1130 	if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1131 			!ceph_con_opened(&osd->o_con)) {
1132 		struct ceph_osd_request *req;
1133 
1134 		dout("osd addr hasn't changed and connection never opened, "
1135 		     "letting msgr retry\n");
1136 		/* touch each r_stamp for handle_timeout()'s benfit */
1137 		list_for_each_entry(req, &osd->o_requests, r_osd_item)
1138 			req->r_stamp = jiffies;
1139 
1140 		return -EAGAIN;
1141 	}
1142 
1143 	ceph_con_close(&osd->o_con);
1144 	ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1145 	osd->o_incarnation++;
1146 
1147 	return 0;
1148 }
1149 
1150 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1151 {
1152 	struct rb_node **p = &osdc->osds.rb_node;
1153 	struct rb_node *parent = NULL;
1154 	struct ceph_osd *osd = NULL;
1155 
1156 	dout("__insert_osd %p osd%d\n", new, new->o_osd);
1157 	while (*p) {
1158 		parent = *p;
1159 		osd = rb_entry(parent, struct ceph_osd, o_node);
1160 		if (new->o_osd < osd->o_osd)
1161 			p = &(*p)->rb_left;
1162 		else if (new->o_osd > osd->o_osd)
1163 			p = &(*p)->rb_right;
1164 		else
1165 			BUG();
1166 	}
1167 
1168 	rb_link_node(&new->o_node, parent, p);
1169 	rb_insert_color(&new->o_node, &osdc->osds);
1170 }
1171 
1172 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1173 {
1174 	struct ceph_osd *osd;
1175 	struct rb_node *n = osdc->osds.rb_node;
1176 
1177 	while (n) {
1178 		osd = rb_entry(n, struct ceph_osd, o_node);
1179 		if (o < osd->o_osd)
1180 			n = n->rb_left;
1181 		else if (o > osd->o_osd)
1182 			n = n->rb_right;
1183 		else
1184 			return osd;
1185 	}
1186 	return NULL;
1187 }
1188 
1189 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1190 {
1191 	schedule_delayed_work(&osdc->timeout_work,
1192 			osdc->client->options->osd_keepalive_timeout * HZ);
1193 }
1194 
1195 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1196 {
1197 	cancel_delayed_work(&osdc->timeout_work);
1198 }
1199 
1200 /*
1201  * Register request, assign tid.  If this is the first request, set up
1202  * the timeout event.
1203  */
1204 static void __register_request(struct ceph_osd_client *osdc,
1205 			       struct ceph_osd_request *req)
1206 {
1207 	req->r_tid = ++osdc->last_tid;
1208 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1209 	dout("__register_request %p tid %lld\n", req, req->r_tid);
1210 	__insert_request(osdc, req);
1211 	ceph_osdc_get_request(req);
1212 	osdc->num_requests++;
1213 	if (osdc->num_requests == 1) {
1214 		dout(" first request, scheduling timeout\n");
1215 		__schedule_osd_timeout(osdc);
1216 	}
1217 }
1218 
1219 /*
1220  * called under osdc->request_mutex
1221  */
1222 static void __unregister_request(struct ceph_osd_client *osdc,
1223 				 struct ceph_osd_request *req)
1224 {
1225 	if (RB_EMPTY_NODE(&req->r_node)) {
1226 		dout("__unregister_request %p tid %lld not registered\n",
1227 			req, req->r_tid);
1228 		return;
1229 	}
1230 
1231 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1232 	rb_erase(&req->r_node, &osdc->requests);
1233 	RB_CLEAR_NODE(&req->r_node);
1234 	osdc->num_requests--;
1235 
1236 	if (req->r_osd) {
1237 		/* make sure the original request isn't in flight. */
1238 		ceph_msg_revoke(req->r_request);
1239 
1240 		list_del_init(&req->r_osd_item);
1241 		maybe_move_osd_to_lru(osdc, req->r_osd);
1242 		if (list_empty(&req->r_linger_osd_item))
1243 			req->r_osd = NULL;
1244 	}
1245 
1246 	list_del_init(&req->r_req_lru_item);
1247 	ceph_osdc_put_request(req);
1248 
1249 	if (osdc->num_requests == 0) {
1250 		dout(" no requests, canceling timeout\n");
1251 		__cancel_osd_timeout(osdc);
1252 	}
1253 }
1254 
1255 /*
1256  * Cancel a previously queued request message
1257  */
1258 static void __cancel_request(struct ceph_osd_request *req)
1259 {
1260 	if (req->r_sent && req->r_osd) {
1261 		ceph_msg_revoke(req->r_request);
1262 		req->r_sent = 0;
1263 	}
1264 }
1265 
1266 static void __register_linger_request(struct ceph_osd_client *osdc,
1267 				    struct ceph_osd_request *req)
1268 {
1269 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1270 	WARN_ON(!req->r_linger);
1271 
1272 	ceph_osdc_get_request(req);
1273 	list_add_tail(&req->r_linger_item, &osdc->req_linger);
1274 	if (req->r_osd)
1275 		list_add_tail(&req->r_linger_osd_item,
1276 			      &req->r_osd->o_linger_requests);
1277 }
1278 
1279 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1280 					struct ceph_osd_request *req)
1281 {
1282 	WARN_ON(!req->r_linger);
1283 
1284 	if (list_empty(&req->r_linger_item)) {
1285 		dout("%s %p tid %llu not registered\n", __func__, req,
1286 		     req->r_tid);
1287 		return;
1288 	}
1289 
1290 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1291 	list_del_init(&req->r_linger_item);
1292 
1293 	if (req->r_osd) {
1294 		list_del_init(&req->r_linger_osd_item);
1295 		maybe_move_osd_to_lru(osdc, req->r_osd);
1296 		if (list_empty(&req->r_osd_item))
1297 			req->r_osd = NULL;
1298 	}
1299 
1300 	list_del_init(&req->r_req_lru_item); /* can be on notarget */
1301 	ceph_osdc_put_request(req);
1302 }
1303 
1304 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1305 				  struct ceph_osd_request *req)
1306 {
1307 	if (!req->r_linger) {
1308 		dout("set_request_linger %p\n", req);
1309 		req->r_linger = 1;
1310 	}
1311 }
1312 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1313 
1314 /*
1315  * Returns whether a request should be blocked from being sent
1316  * based on the current osdmap and osd_client settings.
1317  *
1318  * Caller should hold map_sem for read.
1319  */
1320 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1321 				   struct ceph_osd_request *req)
1322 {
1323 	bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1324 	bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1325 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1326 	return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1327 		(req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1328 }
1329 
1330 /*
1331  * Calculate mapping of a request to a PG.  Takes tiering into account.
1332  */
1333 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1334 			     struct ceph_osd_request *req,
1335 			     struct ceph_pg *pg_out)
1336 {
1337 	bool need_check_tiering;
1338 
1339 	need_check_tiering = false;
1340 	if (req->r_target_oloc.pool == -1) {
1341 		req->r_target_oloc = req->r_base_oloc; /* struct */
1342 		need_check_tiering = true;
1343 	}
1344 	if (req->r_target_oid.name_len == 0) {
1345 		ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1346 		need_check_tiering = true;
1347 	}
1348 
1349 	if (need_check_tiering &&
1350 	    (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1351 		struct ceph_pg_pool_info *pi;
1352 
1353 		pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1354 		if (pi) {
1355 			if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1356 			    pi->read_tier >= 0)
1357 				req->r_target_oloc.pool = pi->read_tier;
1358 			if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1359 			    pi->write_tier >= 0)
1360 				req->r_target_oloc.pool = pi->write_tier;
1361 		}
1362 		/* !pi is caught in ceph_oloc_oid_to_pg() */
1363 	}
1364 
1365 	return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1366 				   &req->r_target_oid, pg_out);
1367 }
1368 
1369 static void __enqueue_request(struct ceph_osd_request *req)
1370 {
1371 	struct ceph_osd_client *osdc = req->r_osdc;
1372 
1373 	dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1374 	     req->r_osd ? req->r_osd->o_osd : -1);
1375 
1376 	if (req->r_osd) {
1377 		__remove_osd_from_lru(req->r_osd);
1378 		list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1379 		list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1380 	} else {
1381 		list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1382 	}
1383 }
1384 
1385 /*
1386  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1387  * (as needed), and set the request r_osd appropriately.  If there is
1388  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1389  * (unsent, homeless) or leave on in-flight lru.
1390  *
1391  * Return 0 if unchanged, 1 if changed, or negative on error.
1392  *
1393  * Caller should hold map_sem for read and request_mutex.
1394  */
1395 static int __map_request(struct ceph_osd_client *osdc,
1396 			 struct ceph_osd_request *req, int force_resend)
1397 {
1398 	struct ceph_pg pgid;
1399 	int acting[CEPH_PG_MAX_SIZE];
1400 	int num, o;
1401 	int err;
1402 	bool was_paused;
1403 
1404 	dout("map_request %p tid %lld\n", req, req->r_tid);
1405 
1406 	err = __calc_request_pg(osdc->osdmap, req, &pgid);
1407 	if (err) {
1408 		list_move(&req->r_req_lru_item, &osdc->req_notarget);
1409 		return err;
1410 	}
1411 	req->r_pgid = pgid;
1412 
1413 	num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1414 	if (num < 0)
1415 		num = 0;
1416 
1417 	was_paused = req->r_paused;
1418 	req->r_paused = __req_should_be_paused(osdc, req);
1419 	if (was_paused && !req->r_paused)
1420 		force_resend = 1;
1421 
1422 	if ((!force_resend &&
1423 	     req->r_osd && req->r_osd->o_osd == o &&
1424 	     req->r_sent >= req->r_osd->o_incarnation &&
1425 	     req->r_num_pg_osds == num &&
1426 	     memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1427 	    (req->r_osd == NULL && o == -1) ||
1428 	    req->r_paused)
1429 		return 0;  /* no change */
1430 
1431 	dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1432 	     req->r_tid, pgid.pool, pgid.seed, o,
1433 	     req->r_osd ? req->r_osd->o_osd : -1);
1434 
1435 	/* record full pg acting set */
1436 	memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1437 	req->r_num_pg_osds = num;
1438 
1439 	if (req->r_osd) {
1440 		__cancel_request(req);
1441 		list_del_init(&req->r_osd_item);
1442 		list_del_init(&req->r_linger_osd_item);
1443 		req->r_osd = NULL;
1444 	}
1445 
1446 	req->r_osd = __lookup_osd(osdc, o);
1447 	if (!req->r_osd && o >= 0) {
1448 		err = -ENOMEM;
1449 		req->r_osd = create_osd(osdc, o);
1450 		if (!req->r_osd) {
1451 			list_move(&req->r_req_lru_item, &osdc->req_notarget);
1452 			goto out;
1453 		}
1454 
1455 		dout("map_request osd %p is osd%d\n", req->r_osd, o);
1456 		__insert_osd(osdc, req->r_osd);
1457 
1458 		ceph_con_open(&req->r_osd->o_con,
1459 			      CEPH_ENTITY_TYPE_OSD, o,
1460 			      &osdc->osdmap->osd_addr[o]);
1461 	}
1462 
1463 	__enqueue_request(req);
1464 	err = 1;   /* osd or pg changed */
1465 
1466 out:
1467 	return err;
1468 }
1469 
1470 /*
1471  * caller should hold map_sem (for read) and request_mutex
1472  */
1473 static void __send_request(struct ceph_osd_client *osdc,
1474 			   struct ceph_osd_request *req)
1475 {
1476 	void *p;
1477 
1478 	dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1479 	     req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1480 	     (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1481 
1482 	/* fill in message content that changes each time we send it */
1483 	put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1484 	put_unaligned_le32(req->r_flags, req->r_request_flags);
1485 	put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1486 	p = req->r_request_pgid;
1487 	ceph_encode_64(&p, req->r_pgid.pool);
1488 	ceph_encode_32(&p, req->r_pgid.seed);
1489 	put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1490 	memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1491 	       sizeof(req->r_reassert_version));
1492 
1493 	req->r_stamp = jiffies;
1494 	list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1495 
1496 	ceph_msg_get(req->r_request); /* send consumes a ref */
1497 
1498 	req->r_sent = req->r_osd->o_incarnation;
1499 
1500 	ceph_con_send(&req->r_osd->o_con, req->r_request);
1501 }
1502 
1503 /*
1504  * Send any requests in the queue (req_unsent).
1505  */
1506 static void __send_queued(struct ceph_osd_client *osdc)
1507 {
1508 	struct ceph_osd_request *req, *tmp;
1509 
1510 	dout("__send_queued\n");
1511 	list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1512 		__send_request(osdc, req);
1513 }
1514 
1515 /*
1516  * Caller should hold map_sem for read and request_mutex.
1517  */
1518 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1519 				     struct ceph_osd_request *req,
1520 				     bool nofail)
1521 {
1522 	int rc;
1523 
1524 	__register_request(osdc, req);
1525 	req->r_sent = 0;
1526 	req->r_got_reply = 0;
1527 	rc = __map_request(osdc, req, 0);
1528 	if (rc < 0) {
1529 		if (nofail) {
1530 			dout("osdc_start_request failed map, "
1531 				" will retry %lld\n", req->r_tid);
1532 			rc = 0;
1533 		} else {
1534 			__unregister_request(osdc, req);
1535 		}
1536 		return rc;
1537 	}
1538 
1539 	if (req->r_osd == NULL) {
1540 		dout("send_request %p no up osds in pg\n", req);
1541 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1542 	} else {
1543 		__send_queued(osdc);
1544 	}
1545 
1546 	return 0;
1547 }
1548 
1549 /*
1550  * Timeout callback, called every N seconds when 1 or more osd
1551  * requests has been active for more than N seconds.  When this
1552  * happens, we ping all OSDs with requests who have timed out to
1553  * ensure any communications channel reset is detected.  Reset the
1554  * request timeouts another N seconds in the future as we go.
1555  * Reschedule the timeout event another N seconds in future (unless
1556  * there are no open requests).
1557  */
1558 static void handle_timeout(struct work_struct *work)
1559 {
1560 	struct ceph_osd_client *osdc =
1561 		container_of(work, struct ceph_osd_client, timeout_work.work);
1562 	struct ceph_osd_request *req;
1563 	struct ceph_osd *osd;
1564 	unsigned long keepalive =
1565 		osdc->client->options->osd_keepalive_timeout * HZ;
1566 	struct list_head slow_osds;
1567 	dout("timeout\n");
1568 	down_read(&osdc->map_sem);
1569 
1570 	ceph_monc_request_next_osdmap(&osdc->client->monc);
1571 
1572 	mutex_lock(&osdc->request_mutex);
1573 
1574 	/*
1575 	 * ping osds that are a bit slow.  this ensures that if there
1576 	 * is a break in the TCP connection we will notice, and reopen
1577 	 * a connection with that osd (from the fault callback).
1578 	 */
1579 	INIT_LIST_HEAD(&slow_osds);
1580 	list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1581 		if (time_before(jiffies, req->r_stamp + keepalive))
1582 			break;
1583 
1584 		osd = req->r_osd;
1585 		BUG_ON(!osd);
1586 		dout(" tid %llu is slow, will send keepalive on osd%d\n",
1587 		     req->r_tid, osd->o_osd);
1588 		list_move_tail(&osd->o_keepalive_item, &slow_osds);
1589 	}
1590 	while (!list_empty(&slow_osds)) {
1591 		osd = list_entry(slow_osds.next, struct ceph_osd,
1592 				 o_keepalive_item);
1593 		list_del_init(&osd->o_keepalive_item);
1594 		ceph_con_keepalive(&osd->o_con);
1595 	}
1596 
1597 	__schedule_osd_timeout(osdc);
1598 	__send_queued(osdc);
1599 	mutex_unlock(&osdc->request_mutex);
1600 	up_read(&osdc->map_sem);
1601 }
1602 
1603 static void handle_osds_timeout(struct work_struct *work)
1604 {
1605 	struct ceph_osd_client *osdc =
1606 		container_of(work, struct ceph_osd_client,
1607 			     osds_timeout_work.work);
1608 	unsigned long delay =
1609 		osdc->client->options->osd_idle_ttl * HZ >> 2;
1610 
1611 	dout("osds timeout\n");
1612 	down_read(&osdc->map_sem);
1613 	remove_old_osds(osdc);
1614 	up_read(&osdc->map_sem);
1615 
1616 	schedule_delayed_work(&osdc->osds_timeout_work,
1617 			      round_jiffies_relative(delay));
1618 }
1619 
1620 static int ceph_oloc_decode(void **p, void *end,
1621 			    struct ceph_object_locator *oloc)
1622 {
1623 	u8 struct_v, struct_cv;
1624 	u32 len;
1625 	void *struct_end;
1626 	int ret = 0;
1627 
1628 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1629 	struct_v = ceph_decode_8(p);
1630 	struct_cv = ceph_decode_8(p);
1631 	if (struct_v < 3) {
1632 		pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1633 			struct_v, struct_cv);
1634 		goto e_inval;
1635 	}
1636 	if (struct_cv > 6) {
1637 		pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1638 			struct_v, struct_cv);
1639 		goto e_inval;
1640 	}
1641 	len = ceph_decode_32(p);
1642 	ceph_decode_need(p, end, len, e_inval);
1643 	struct_end = *p + len;
1644 
1645 	oloc->pool = ceph_decode_64(p);
1646 	*p += 4; /* skip preferred */
1647 
1648 	len = ceph_decode_32(p);
1649 	if (len > 0) {
1650 		pr_warn("ceph_object_locator::key is set\n");
1651 		goto e_inval;
1652 	}
1653 
1654 	if (struct_v >= 5) {
1655 		len = ceph_decode_32(p);
1656 		if (len > 0) {
1657 			pr_warn("ceph_object_locator::nspace is set\n");
1658 			goto e_inval;
1659 		}
1660 	}
1661 
1662 	if (struct_v >= 6) {
1663 		s64 hash = ceph_decode_64(p);
1664 		if (hash != -1) {
1665 			pr_warn("ceph_object_locator::hash is set\n");
1666 			goto e_inval;
1667 		}
1668 	}
1669 
1670 	/* skip the rest */
1671 	*p = struct_end;
1672 out:
1673 	return ret;
1674 
1675 e_inval:
1676 	ret = -EINVAL;
1677 	goto out;
1678 }
1679 
1680 static int ceph_redirect_decode(void **p, void *end,
1681 				struct ceph_request_redirect *redir)
1682 {
1683 	u8 struct_v, struct_cv;
1684 	u32 len;
1685 	void *struct_end;
1686 	int ret;
1687 
1688 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1689 	struct_v = ceph_decode_8(p);
1690 	struct_cv = ceph_decode_8(p);
1691 	if (struct_cv > 1) {
1692 		pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1693 			struct_v, struct_cv);
1694 		goto e_inval;
1695 	}
1696 	len = ceph_decode_32(p);
1697 	ceph_decode_need(p, end, len, e_inval);
1698 	struct_end = *p + len;
1699 
1700 	ret = ceph_oloc_decode(p, end, &redir->oloc);
1701 	if (ret)
1702 		goto out;
1703 
1704 	len = ceph_decode_32(p);
1705 	if (len > 0) {
1706 		pr_warn("ceph_request_redirect::object_name is set\n");
1707 		goto e_inval;
1708 	}
1709 
1710 	len = ceph_decode_32(p);
1711 	*p += len; /* skip osd_instructions */
1712 
1713 	/* skip the rest */
1714 	*p = struct_end;
1715 out:
1716 	return ret;
1717 
1718 e_inval:
1719 	ret = -EINVAL;
1720 	goto out;
1721 }
1722 
1723 static void complete_request(struct ceph_osd_request *req)
1724 {
1725 	complete_all(&req->r_safe_completion);  /* fsync waiter */
1726 }
1727 
1728 /*
1729  * handle osd op reply.  either call the callback if it is specified,
1730  * or do the completion to wake up the waiting thread.
1731  */
1732 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1733 			 struct ceph_connection *con)
1734 {
1735 	void *p, *end;
1736 	struct ceph_osd_request *req;
1737 	struct ceph_request_redirect redir;
1738 	u64 tid;
1739 	int object_len;
1740 	unsigned int numops;
1741 	int payload_len, flags;
1742 	s32 result;
1743 	s32 retry_attempt;
1744 	struct ceph_pg pg;
1745 	int err;
1746 	u32 reassert_epoch;
1747 	u64 reassert_version;
1748 	u32 osdmap_epoch;
1749 	int already_completed;
1750 	u32 bytes;
1751 	unsigned int i;
1752 
1753 	tid = le64_to_cpu(msg->hdr.tid);
1754 	dout("handle_reply %p tid %llu\n", msg, tid);
1755 
1756 	p = msg->front.iov_base;
1757 	end = p + msg->front.iov_len;
1758 
1759 	ceph_decode_need(&p, end, 4, bad);
1760 	object_len = ceph_decode_32(&p);
1761 	ceph_decode_need(&p, end, object_len, bad);
1762 	p += object_len;
1763 
1764 	err = ceph_decode_pgid(&p, end, &pg);
1765 	if (err)
1766 		goto bad;
1767 
1768 	ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1769 	flags = ceph_decode_64(&p);
1770 	result = ceph_decode_32(&p);
1771 	reassert_epoch = ceph_decode_32(&p);
1772 	reassert_version = ceph_decode_64(&p);
1773 	osdmap_epoch = ceph_decode_32(&p);
1774 
1775 	/* lookup */
1776 	down_read(&osdc->map_sem);
1777 	mutex_lock(&osdc->request_mutex);
1778 	req = __lookup_request(osdc, tid);
1779 	if (req == NULL) {
1780 		dout("handle_reply tid %llu dne\n", tid);
1781 		goto bad_mutex;
1782 	}
1783 	ceph_osdc_get_request(req);
1784 
1785 	dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1786 	     req, result);
1787 
1788 	ceph_decode_need(&p, end, 4, bad_put);
1789 	numops = ceph_decode_32(&p);
1790 	if (numops > CEPH_OSD_MAX_OP)
1791 		goto bad_put;
1792 	if (numops != req->r_num_ops)
1793 		goto bad_put;
1794 	payload_len = 0;
1795 	ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1796 	for (i = 0; i < numops; i++) {
1797 		struct ceph_osd_op *op = p;
1798 		int len;
1799 
1800 		len = le32_to_cpu(op->payload_len);
1801 		req->r_reply_op_len[i] = len;
1802 		dout(" op %d has %d bytes\n", i, len);
1803 		payload_len += len;
1804 		p += sizeof(*op);
1805 	}
1806 	bytes = le32_to_cpu(msg->hdr.data_len);
1807 	if (payload_len != bytes) {
1808 		pr_warn("sum of op payload lens %d != data_len %d\n",
1809 			payload_len, bytes);
1810 		goto bad_put;
1811 	}
1812 
1813 	ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1814 	retry_attempt = ceph_decode_32(&p);
1815 	for (i = 0; i < numops; i++)
1816 		req->r_reply_op_result[i] = ceph_decode_32(&p);
1817 
1818 	if (le16_to_cpu(msg->hdr.version) >= 6) {
1819 		p += 8 + 4; /* skip replay_version */
1820 		p += 8; /* skip user_version */
1821 
1822 		err = ceph_redirect_decode(&p, end, &redir);
1823 		if (err)
1824 			goto bad_put;
1825 	} else {
1826 		redir.oloc.pool = -1;
1827 	}
1828 
1829 	if (redir.oloc.pool != -1) {
1830 		dout("redirect pool %lld\n", redir.oloc.pool);
1831 
1832 		__unregister_request(osdc, req);
1833 
1834 		req->r_target_oloc = redir.oloc; /* struct */
1835 
1836 		/*
1837 		 * Start redirect requests with nofail=true.  If
1838 		 * mapping fails, request will end up on the notarget
1839 		 * list, waiting for the new osdmap (which can take
1840 		 * a while), even though the original request mapped
1841 		 * successfully.  In the future we might want to follow
1842 		 * original request's nofail setting here.
1843 		 */
1844 		err = __ceph_osdc_start_request(osdc, req, true);
1845 		BUG_ON(err);
1846 
1847 		goto out_unlock;
1848 	}
1849 
1850 	already_completed = req->r_got_reply;
1851 	if (!req->r_got_reply) {
1852 		req->r_result = result;
1853 		dout("handle_reply result %d bytes %d\n", req->r_result,
1854 		     bytes);
1855 		if (req->r_result == 0)
1856 			req->r_result = bytes;
1857 
1858 		/* in case this is a write and we need to replay, */
1859 		req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1860 		req->r_reassert_version.version = cpu_to_le64(reassert_version);
1861 
1862 		req->r_got_reply = 1;
1863 	} else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1864 		dout("handle_reply tid %llu dup ack\n", tid);
1865 		goto out_unlock;
1866 	}
1867 
1868 	dout("handle_reply tid %llu flags %d\n", tid, flags);
1869 
1870 	if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1871 		__register_linger_request(osdc, req);
1872 
1873 	/* either this is a read, or we got the safe response */
1874 	if (result < 0 ||
1875 	    (flags & CEPH_OSD_FLAG_ONDISK) ||
1876 	    ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1877 		__unregister_request(osdc, req);
1878 
1879 	mutex_unlock(&osdc->request_mutex);
1880 	up_read(&osdc->map_sem);
1881 
1882 	if (!already_completed) {
1883 		if (req->r_unsafe_callback &&
1884 		    result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1885 			req->r_unsafe_callback(req, true);
1886 		if (req->r_callback)
1887 			req->r_callback(req, msg);
1888 		else
1889 			complete_all(&req->r_completion);
1890 	}
1891 
1892 	if (flags & CEPH_OSD_FLAG_ONDISK) {
1893 		if (req->r_unsafe_callback && already_completed)
1894 			req->r_unsafe_callback(req, false);
1895 		complete_request(req);
1896 	}
1897 
1898 out:
1899 	dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1900 	ceph_osdc_put_request(req);
1901 	return;
1902 out_unlock:
1903 	mutex_unlock(&osdc->request_mutex);
1904 	up_read(&osdc->map_sem);
1905 	goto out;
1906 
1907 bad_put:
1908 	req->r_result = -EIO;
1909 	__unregister_request(osdc, req);
1910 	if (req->r_callback)
1911 		req->r_callback(req, msg);
1912 	else
1913 		complete_all(&req->r_completion);
1914 	complete_request(req);
1915 	ceph_osdc_put_request(req);
1916 bad_mutex:
1917 	mutex_unlock(&osdc->request_mutex);
1918 	up_read(&osdc->map_sem);
1919 bad:
1920 	pr_err("corrupt osd_op_reply got %d %d\n",
1921 	       (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1922 	ceph_msg_dump(msg);
1923 }
1924 
1925 static void reset_changed_osds(struct ceph_osd_client *osdc)
1926 {
1927 	struct rb_node *p, *n;
1928 
1929 	for (p = rb_first(&osdc->osds); p; p = n) {
1930 		struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1931 
1932 		n = rb_next(p);
1933 		if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1934 		    memcmp(&osd->o_con.peer_addr,
1935 			   ceph_osd_addr(osdc->osdmap,
1936 					 osd->o_osd),
1937 			   sizeof(struct ceph_entity_addr)) != 0)
1938 			__reset_osd(osdc, osd);
1939 	}
1940 }
1941 
1942 /*
1943  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1944  * no osd, request a new map.
1945  *
1946  * Caller should hold map_sem for read.
1947  */
1948 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1949 			  bool force_resend_writes)
1950 {
1951 	struct ceph_osd_request *req, *nreq;
1952 	struct rb_node *p;
1953 	int needmap = 0;
1954 	int err;
1955 	bool force_resend_req;
1956 
1957 	dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1958 		force_resend_writes ? " (force resend writes)" : "");
1959 	mutex_lock(&osdc->request_mutex);
1960 	for (p = rb_first(&osdc->requests); p; ) {
1961 		req = rb_entry(p, struct ceph_osd_request, r_node);
1962 		p = rb_next(p);
1963 
1964 		/*
1965 		 * For linger requests that have not yet been
1966 		 * registered, move them to the linger list; they'll
1967 		 * be sent to the osd in the loop below.  Unregister
1968 		 * the request before re-registering it as a linger
1969 		 * request to ensure the __map_request() below
1970 		 * will decide it needs to be sent.
1971 		 */
1972 		if (req->r_linger && list_empty(&req->r_linger_item)) {
1973 			dout("%p tid %llu restart on osd%d\n",
1974 			     req, req->r_tid,
1975 			     req->r_osd ? req->r_osd->o_osd : -1);
1976 			ceph_osdc_get_request(req);
1977 			__unregister_request(osdc, req);
1978 			__register_linger_request(osdc, req);
1979 			ceph_osdc_put_request(req);
1980 			continue;
1981 		}
1982 
1983 		force_resend_req = force_resend ||
1984 			(force_resend_writes &&
1985 				req->r_flags & CEPH_OSD_FLAG_WRITE);
1986 		err = __map_request(osdc, req, force_resend_req);
1987 		if (err < 0)
1988 			continue;  /* error */
1989 		if (req->r_osd == NULL) {
1990 			dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1991 			needmap++;  /* request a newer map */
1992 		} else if (err > 0) {
1993 			if (!req->r_linger) {
1994 				dout("%p tid %llu requeued on osd%d\n", req,
1995 				     req->r_tid,
1996 				     req->r_osd ? req->r_osd->o_osd : -1);
1997 				req->r_flags |= CEPH_OSD_FLAG_RETRY;
1998 			}
1999 		}
2000 	}
2001 
2002 	list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2003 				 r_linger_item) {
2004 		dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2005 
2006 		err = __map_request(osdc, req,
2007 				    force_resend || force_resend_writes);
2008 		dout("__map_request returned %d\n", err);
2009 		if (err == 0)
2010 			continue;  /* no change and no osd was specified */
2011 		if (err < 0)
2012 			continue;  /* hrm! */
2013 		if (req->r_osd == NULL) {
2014 			dout("tid %llu maps to no valid osd\n", req->r_tid);
2015 			needmap++;  /* request a newer map */
2016 			continue;
2017 		}
2018 
2019 		dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
2020 		     req->r_osd ? req->r_osd->o_osd : -1);
2021 		__register_request(osdc, req);
2022 		__unregister_linger_request(osdc, req);
2023 	}
2024 	reset_changed_osds(osdc);
2025 	mutex_unlock(&osdc->request_mutex);
2026 
2027 	if (needmap) {
2028 		dout("%d requests for down osds, need new map\n", needmap);
2029 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2030 	}
2031 }
2032 
2033 
2034 /*
2035  * Process updated osd map.
2036  *
2037  * The message contains any number of incremental and full maps, normally
2038  * indicating some sort of topology change in the cluster.  Kick requests
2039  * off to different OSDs as needed.
2040  */
2041 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2042 {
2043 	void *p, *end, *next;
2044 	u32 nr_maps, maplen;
2045 	u32 epoch;
2046 	struct ceph_osdmap *newmap = NULL, *oldmap;
2047 	int err;
2048 	struct ceph_fsid fsid;
2049 	bool was_full;
2050 
2051 	dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2052 	p = msg->front.iov_base;
2053 	end = p + msg->front.iov_len;
2054 
2055 	/* verify fsid */
2056 	ceph_decode_need(&p, end, sizeof(fsid), bad);
2057 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
2058 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
2059 		return;
2060 
2061 	down_write(&osdc->map_sem);
2062 
2063 	was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2064 
2065 	/* incremental maps */
2066 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2067 	dout(" %d inc maps\n", nr_maps);
2068 	while (nr_maps > 0) {
2069 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2070 		epoch = ceph_decode_32(&p);
2071 		maplen = ceph_decode_32(&p);
2072 		ceph_decode_need(&p, end, maplen, bad);
2073 		next = p + maplen;
2074 		if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2075 			dout("applying incremental map %u len %d\n",
2076 			     epoch, maplen);
2077 			newmap = osdmap_apply_incremental(&p, next,
2078 							  osdc->osdmap,
2079 							  &osdc->client->msgr);
2080 			if (IS_ERR(newmap)) {
2081 				err = PTR_ERR(newmap);
2082 				goto bad;
2083 			}
2084 			BUG_ON(!newmap);
2085 			if (newmap != osdc->osdmap) {
2086 				ceph_osdmap_destroy(osdc->osdmap);
2087 				osdc->osdmap = newmap;
2088 			}
2089 			was_full = was_full ||
2090 				ceph_osdmap_flag(osdc->osdmap,
2091 						 CEPH_OSDMAP_FULL);
2092 			kick_requests(osdc, 0, was_full);
2093 		} else {
2094 			dout("ignoring incremental map %u len %d\n",
2095 			     epoch, maplen);
2096 		}
2097 		p = next;
2098 		nr_maps--;
2099 	}
2100 	if (newmap)
2101 		goto done;
2102 
2103 	/* full maps */
2104 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2105 	dout(" %d full maps\n", nr_maps);
2106 	while (nr_maps) {
2107 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2108 		epoch = ceph_decode_32(&p);
2109 		maplen = ceph_decode_32(&p);
2110 		ceph_decode_need(&p, end, maplen, bad);
2111 		if (nr_maps > 1) {
2112 			dout("skipping non-latest full map %u len %d\n",
2113 			     epoch, maplen);
2114 		} else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2115 			dout("skipping full map %u len %d, "
2116 			     "older than our %u\n", epoch, maplen,
2117 			     osdc->osdmap->epoch);
2118 		} else {
2119 			int skipped_map = 0;
2120 
2121 			dout("taking full map %u len %d\n", epoch, maplen);
2122 			newmap = ceph_osdmap_decode(&p, p+maplen);
2123 			if (IS_ERR(newmap)) {
2124 				err = PTR_ERR(newmap);
2125 				goto bad;
2126 			}
2127 			BUG_ON(!newmap);
2128 			oldmap = osdc->osdmap;
2129 			osdc->osdmap = newmap;
2130 			if (oldmap) {
2131 				if (oldmap->epoch + 1 < newmap->epoch)
2132 					skipped_map = 1;
2133 				ceph_osdmap_destroy(oldmap);
2134 			}
2135 			was_full = was_full ||
2136 				ceph_osdmap_flag(osdc->osdmap,
2137 						 CEPH_OSDMAP_FULL);
2138 			kick_requests(osdc, skipped_map, was_full);
2139 		}
2140 		p += maplen;
2141 		nr_maps--;
2142 	}
2143 
2144 	if (!osdc->osdmap)
2145 		goto bad;
2146 done:
2147 	downgrade_write(&osdc->map_sem);
2148 	ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2149 
2150 	/*
2151 	 * subscribe to subsequent osdmap updates if full to ensure
2152 	 * we find out when we are no longer full and stop returning
2153 	 * ENOSPC.
2154 	 */
2155 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2156 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2157 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2158 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2159 
2160 	mutex_lock(&osdc->request_mutex);
2161 	__send_queued(osdc);
2162 	mutex_unlock(&osdc->request_mutex);
2163 	up_read(&osdc->map_sem);
2164 	wake_up_all(&osdc->client->auth_wq);
2165 	return;
2166 
2167 bad:
2168 	pr_err("osdc handle_map corrupt msg\n");
2169 	ceph_msg_dump(msg);
2170 	up_write(&osdc->map_sem);
2171 }
2172 
2173 /*
2174  * watch/notify callback event infrastructure
2175  *
2176  * These callbacks are used both for watch and notify operations.
2177  */
2178 static void __release_event(struct kref *kref)
2179 {
2180 	struct ceph_osd_event *event =
2181 		container_of(kref, struct ceph_osd_event, kref);
2182 
2183 	dout("__release_event %p\n", event);
2184 	kfree(event);
2185 }
2186 
2187 static void get_event(struct ceph_osd_event *event)
2188 {
2189 	kref_get(&event->kref);
2190 }
2191 
2192 void ceph_osdc_put_event(struct ceph_osd_event *event)
2193 {
2194 	kref_put(&event->kref, __release_event);
2195 }
2196 EXPORT_SYMBOL(ceph_osdc_put_event);
2197 
2198 static void __insert_event(struct ceph_osd_client *osdc,
2199 			     struct ceph_osd_event *new)
2200 {
2201 	struct rb_node **p = &osdc->event_tree.rb_node;
2202 	struct rb_node *parent = NULL;
2203 	struct ceph_osd_event *event = NULL;
2204 
2205 	while (*p) {
2206 		parent = *p;
2207 		event = rb_entry(parent, struct ceph_osd_event, node);
2208 		if (new->cookie < event->cookie)
2209 			p = &(*p)->rb_left;
2210 		else if (new->cookie > event->cookie)
2211 			p = &(*p)->rb_right;
2212 		else
2213 			BUG();
2214 	}
2215 
2216 	rb_link_node(&new->node, parent, p);
2217 	rb_insert_color(&new->node, &osdc->event_tree);
2218 }
2219 
2220 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2221 					        u64 cookie)
2222 {
2223 	struct rb_node **p = &osdc->event_tree.rb_node;
2224 	struct rb_node *parent = NULL;
2225 	struct ceph_osd_event *event = NULL;
2226 
2227 	while (*p) {
2228 		parent = *p;
2229 		event = rb_entry(parent, struct ceph_osd_event, node);
2230 		if (cookie < event->cookie)
2231 			p = &(*p)->rb_left;
2232 		else if (cookie > event->cookie)
2233 			p = &(*p)->rb_right;
2234 		else
2235 			return event;
2236 	}
2237 	return NULL;
2238 }
2239 
2240 static void __remove_event(struct ceph_osd_event *event)
2241 {
2242 	struct ceph_osd_client *osdc = event->osdc;
2243 
2244 	if (!RB_EMPTY_NODE(&event->node)) {
2245 		dout("__remove_event removed %p\n", event);
2246 		rb_erase(&event->node, &osdc->event_tree);
2247 		ceph_osdc_put_event(event);
2248 	} else {
2249 		dout("__remove_event didn't remove %p\n", event);
2250 	}
2251 }
2252 
2253 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2254 			   void (*event_cb)(u64, u64, u8, void *),
2255 			   void *data, struct ceph_osd_event **pevent)
2256 {
2257 	struct ceph_osd_event *event;
2258 
2259 	event = kmalloc(sizeof(*event), GFP_NOIO);
2260 	if (!event)
2261 		return -ENOMEM;
2262 
2263 	dout("create_event %p\n", event);
2264 	event->cb = event_cb;
2265 	event->one_shot = 0;
2266 	event->data = data;
2267 	event->osdc = osdc;
2268 	INIT_LIST_HEAD(&event->osd_node);
2269 	RB_CLEAR_NODE(&event->node);
2270 	kref_init(&event->kref);   /* one ref for us */
2271 	kref_get(&event->kref);    /* one ref for the caller */
2272 
2273 	spin_lock(&osdc->event_lock);
2274 	event->cookie = ++osdc->event_count;
2275 	__insert_event(osdc, event);
2276 	spin_unlock(&osdc->event_lock);
2277 
2278 	*pevent = event;
2279 	return 0;
2280 }
2281 EXPORT_SYMBOL(ceph_osdc_create_event);
2282 
2283 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2284 {
2285 	struct ceph_osd_client *osdc = event->osdc;
2286 
2287 	dout("cancel_event %p\n", event);
2288 	spin_lock(&osdc->event_lock);
2289 	__remove_event(event);
2290 	spin_unlock(&osdc->event_lock);
2291 	ceph_osdc_put_event(event); /* caller's */
2292 }
2293 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2294 
2295 
2296 static void do_event_work(struct work_struct *work)
2297 {
2298 	struct ceph_osd_event_work *event_work =
2299 		container_of(work, struct ceph_osd_event_work, work);
2300 	struct ceph_osd_event *event = event_work->event;
2301 	u64 ver = event_work->ver;
2302 	u64 notify_id = event_work->notify_id;
2303 	u8 opcode = event_work->opcode;
2304 
2305 	dout("do_event_work completing %p\n", event);
2306 	event->cb(ver, notify_id, opcode, event->data);
2307 	dout("do_event_work completed %p\n", event);
2308 	ceph_osdc_put_event(event);
2309 	kfree(event_work);
2310 }
2311 
2312 
2313 /*
2314  * Process osd watch notifications
2315  */
2316 static void handle_watch_notify(struct ceph_osd_client *osdc,
2317 				struct ceph_msg *msg)
2318 {
2319 	void *p, *end;
2320 	u8 proto_ver;
2321 	u64 cookie, ver, notify_id;
2322 	u8 opcode;
2323 	struct ceph_osd_event *event;
2324 	struct ceph_osd_event_work *event_work;
2325 
2326 	p = msg->front.iov_base;
2327 	end = p + msg->front.iov_len;
2328 
2329 	ceph_decode_8_safe(&p, end, proto_ver, bad);
2330 	ceph_decode_8_safe(&p, end, opcode, bad);
2331 	ceph_decode_64_safe(&p, end, cookie, bad);
2332 	ceph_decode_64_safe(&p, end, ver, bad);
2333 	ceph_decode_64_safe(&p, end, notify_id, bad);
2334 
2335 	spin_lock(&osdc->event_lock);
2336 	event = __find_event(osdc, cookie);
2337 	if (event) {
2338 		BUG_ON(event->one_shot);
2339 		get_event(event);
2340 	}
2341 	spin_unlock(&osdc->event_lock);
2342 	dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2343 	     cookie, ver, event);
2344 	if (event) {
2345 		event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2346 		if (!event_work) {
2347 			pr_err("couldn't allocate event_work\n");
2348 			ceph_osdc_put_event(event);
2349 			return;
2350 		}
2351 		INIT_WORK(&event_work->work, do_event_work);
2352 		event_work->event = event;
2353 		event_work->ver = ver;
2354 		event_work->notify_id = notify_id;
2355 		event_work->opcode = opcode;
2356 
2357 		queue_work(osdc->notify_wq, &event_work->work);
2358 	}
2359 
2360 	return;
2361 
2362 bad:
2363 	pr_err("osdc handle_watch_notify corrupt msg\n");
2364 }
2365 
2366 /*
2367  * build new request AND message
2368  *
2369  */
2370 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2371 				struct ceph_snap_context *snapc, u64 snap_id,
2372 				struct timespec *mtime)
2373 {
2374 	struct ceph_msg *msg = req->r_request;
2375 	void *p;
2376 	size_t msg_size;
2377 	int flags = req->r_flags;
2378 	u64 data_len;
2379 	unsigned int i;
2380 
2381 	req->r_snapid = snap_id;
2382 	req->r_snapc = ceph_get_snap_context(snapc);
2383 
2384 	/* encode request */
2385 	msg->hdr.version = cpu_to_le16(4);
2386 
2387 	p = msg->front.iov_base;
2388 	ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2389 	req->r_request_osdmap_epoch = p;
2390 	p += 4;
2391 	req->r_request_flags = p;
2392 	p += 4;
2393 	if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2394 		ceph_encode_timespec(p, mtime);
2395 	p += sizeof(struct ceph_timespec);
2396 	req->r_request_reassert_version = p;
2397 	p += sizeof(struct ceph_eversion); /* will get filled in */
2398 
2399 	/* oloc */
2400 	ceph_encode_8(&p, 4);
2401 	ceph_encode_8(&p, 4);
2402 	ceph_encode_32(&p, 8 + 4 + 4);
2403 	req->r_request_pool = p;
2404 	p += 8;
2405 	ceph_encode_32(&p, -1);  /* preferred */
2406 	ceph_encode_32(&p, 0);   /* key len */
2407 
2408 	ceph_encode_8(&p, 1);
2409 	req->r_request_pgid = p;
2410 	p += 8 + 4;
2411 	ceph_encode_32(&p, -1);  /* preferred */
2412 
2413 	/* oid */
2414 	ceph_encode_32(&p, req->r_base_oid.name_len);
2415 	memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2416 	dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2417 	     req->r_base_oid.name, req->r_base_oid.name_len);
2418 	p += req->r_base_oid.name_len;
2419 
2420 	/* ops--can imply data */
2421 	ceph_encode_16(&p, (u16)req->r_num_ops);
2422 	data_len = 0;
2423 	for (i = 0; i < req->r_num_ops; i++) {
2424 		data_len += osd_req_encode_op(req, p, i);
2425 		p += sizeof(struct ceph_osd_op);
2426 	}
2427 
2428 	/* snaps */
2429 	ceph_encode_64(&p, req->r_snapid);
2430 	ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2431 	ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2432 	if (req->r_snapc) {
2433 		for (i = 0; i < snapc->num_snaps; i++) {
2434 			ceph_encode_64(&p, req->r_snapc->snaps[i]);
2435 		}
2436 	}
2437 
2438 	req->r_request_attempts = p;
2439 	p += 4;
2440 
2441 	/* data */
2442 	if (flags & CEPH_OSD_FLAG_WRITE) {
2443 		u16 data_off;
2444 
2445 		/*
2446 		 * The header "data_off" is a hint to the receiver
2447 		 * allowing it to align received data into its
2448 		 * buffers such that there's no need to re-copy
2449 		 * it before writing it to disk (direct I/O).
2450 		 */
2451 		data_off = (u16) (off & 0xffff);
2452 		req->r_request->hdr.data_off = cpu_to_le16(data_off);
2453 	}
2454 	req->r_request->hdr.data_len = cpu_to_le32(data_len);
2455 
2456 	BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2457 	msg_size = p - msg->front.iov_base;
2458 	msg->front.iov_len = msg_size;
2459 	msg->hdr.front_len = cpu_to_le32(msg_size);
2460 
2461 	dout("build_request msg_size was %d\n", (int)msg_size);
2462 }
2463 EXPORT_SYMBOL(ceph_osdc_build_request);
2464 
2465 /*
2466  * Register request, send initial attempt.
2467  */
2468 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2469 			    struct ceph_osd_request *req,
2470 			    bool nofail)
2471 {
2472 	int rc;
2473 
2474 	down_read(&osdc->map_sem);
2475 	mutex_lock(&osdc->request_mutex);
2476 
2477 	rc = __ceph_osdc_start_request(osdc, req, nofail);
2478 
2479 	mutex_unlock(&osdc->request_mutex);
2480 	up_read(&osdc->map_sem);
2481 
2482 	return rc;
2483 }
2484 EXPORT_SYMBOL(ceph_osdc_start_request);
2485 
2486 /*
2487  * Unregister a registered request.  The request is not completed (i.e.
2488  * no callbacks or wakeups) - higher layers are supposed to know what
2489  * they are canceling.
2490  */
2491 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2492 {
2493 	struct ceph_osd_client *osdc = req->r_osdc;
2494 
2495 	mutex_lock(&osdc->request_mutex);
2496 	if (req->r_linger)
2497 		__unregister_linger_request(osdc, req);
2498 	__unregister_request(osdc, req);
2499 	mutex_unlock(&osdc->request_mutex);
2500 
2501 	dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2502 }
2503 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2504 
2505 /*
2506  * wait for a request to complete
2507  */
2508 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2509 			   struct ceph_osd_request *req)
2510 {
2511 	int rc;
2512 
2513 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2514 
2515 	rc = wait_for_completion_interruptible(&req->r_completion);
2516 	if (rc < 0) {
2517 		dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2518 		ceph_osdc_cancel_request(req);
2519 		complete_request(req);
2520 		return rc;
2521 	}
2522 
2523 	dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2524 	     req->r_result);
2525 	return req->r_result;
2526 }
2527 EXPORT_SYMBOL(ceph_osdc_wait_request);
2528 
2529 /*
2530  * sync - wait for all in-flight requests to flush.  avoid starvation.
2531  */
2532 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2533 {
2534 	struct ceph_osd_request *req;
2535 	u64 last_tid, next_tid = 0;
2536 
2537 	mutex_lock(&osdc->request_mutex);
2538 	last_tid = osdc->last_tid;
2539 	while (1) {
2540 		req = __lookup_request_ge(osdc, next_tid);
2541 		if (!req)
2542 			break;
2543 		if (req->r_tid > last_tid)
2544 			break;
2545 
2546 		next_tid = req->r_tid + 1;
2547 		if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2548 			continue;
2549 
2550 		ceph_osdc_get_request(req);
2551 		mutex_unlock(&osdc->request_mutex);
2552 		dout("sync waiting on tid %llu (last is %llu)\n",
2553 		     req->r_tid, last_tid);
2554 		wait_for_completion(&req->r_safe_completion);
2555 		mutex_lock(&osdc->request_mutex);
2556 		ceph_osdc_put_request(req);
2557 	}
2558 	mutex_unlock(&osdc->request_mutex);
2559 	dout("sync done (thru tid %llu)\n", last_tid);
2560 }
2561 EXPORT_SYMBOL(ceph_osdc_sync);
2562 
2563 /*
2564  * Call all pending notify callbacks - for use after a watch is
2565  * unregistered, to make sure no more callbacks for it will be invoked
2566  */
2567 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2568 {
2569 	flush_workqueue(osdc->notify_wq);
2570 }
2571 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2572 
2573 
2574 /*
2575  * init, shutdown
2576  */
2577 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2578 {
2579 	int err;
2580 
2581 	dout("init\n");
2582 	osdc->client = client;
2583 	osdc->osdmap = NULL;
2584 	init_rwsem(&osdc->map_sem);
2585 	init_completion(&osdc->map_waiters);
2586 	osdc->last_requested_map = 0;
2587 	mutex_init(&osdc->request_mutex);
2588 	osdc->last_tid = 0;
2589 	osdc->osds = RB_ROOT;
2590 	INIT_LIST_HEAD(&osdc->osd_lru);
2591 	osdc->requests = RB_ROOT;
2592 	INIT_LIST_HEAD(&osdc->req_lru);
2593 	INIT_LIST_HEAD(&osdc->req_unsent);
2594 	INIT_LIST_HEAD(&osdc->req_notarget);
2595 	INIT_LIST_HEAD(&osdc->req_linger);
2596 	osdc->num_requests = 0;
2597 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2598 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2599 	spin_lock_init(&osdc->event_lock);
2600 	osdc->event_tree = RB_ROOT;
2601 	osdc->event_count = 0;
2602 
2603 	schedule_delayed_work(&osdc->osds_timeout_work,
2604 	   round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2605 
2606 	err = -ENOMEM;
2607 	osdc->req_mempool = mempool_create_kmalloc_pool(10,
2608 					sizeof(struct ceph_osd_request));
2609 	if (!osdc->req_mempool)
2610 		goto out;
2611 
2612 	err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2613 				OSD_OP_FRONT_LEN, 10, true,
2614 				"osd_op");
2615 	if (err < 0)
2616 		goto out_mempool;
2617 	err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2618 				OSD_OPREPLY_FRONT_LEN, 10, true,
2619 				"osd_op_reply");
2620 	if (err < 0)
2621 		goto out_msgpool;
2622 
2623 	err = -ENOMEM;
2624 	osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2625 	if (!osdc->notify_wq)
2626 		goto out_msgpool_reply;
2627 
2628 	return 0;
2629 
2630 out_msgpool_reply:
2631 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2632 out_msgpool:
2633 	ceph_msgpool_destroy(&osdc->msgpool_op);
2634 out_mempool:
2635 	mempool_destroy(osdc->req_mempool);
2636 out:
2637 	return err;
2638 }
2639 
2640 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2641 {
2642 	flush_workqueue(osdc->notify_wq);
2643 	destroy_workqueue(osdc->notify_wq);
2644 	cancel_delayed_work_sync(&osdc->timeout_work);
2645 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
2646 	if (osdc->osdmap) {
2647 		ceph_osdmap_destroy(osdc->osdmap);
2648 		osdc->osdmap = NULL;
2649 	}
2650 	remove_all_osds(osdc);
2651 	mempool_destroy(osdc->req_mempool);
2652 	ceph_msgpool_destroy(&osdc->msgpool_op);
2653 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2654 }
2655 
2656 /*
2657  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2658  * *plen.  Return number of bytes read, or error.
2659  */
2660 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2661 			struct ceph_vino vino, struct ceph_file_layout *layout,
2662 			u64 off, u64 *plen,
2663 			u32 truncate_seq, u64 truncate_size,
2664 			struct page **pages, int num_pages, int page_align)
2665 {
2666 	struct ceph_osd_request *req;
2667 	int rc = 0;
2668 
2669 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2670 	     vino.snap, off, *plen);
2671 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2672 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2673 				    NULL, truncate_seq, truncate_size,
2674 				    false);
2675 	if (IS_ERR(req))
2676 		return PTR_ERR(req);
2677 
2678 	/* it may be a short read due to an object boundary */
2679 
2680 	osd_req_op_extent_osd_data_pages(req, 0,
2681 				pages, *plen, page_align, false, false);
2682 
2683 	dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2684 	     off, *plen, *plen, page_align);
2685 
2686 	ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2687 
2688 	rc = ceph_osdc_start_request(osdc, req, false);
2689 	if (!rc)
2690 		rc = ceph_osdc_wait_request(osdc, req);
2691 
2692 	ceph_osdc_put_request(req);
2693 	dout("readpages result %d\n", rc);
2694 	return rc;
2695 }
2696 EXPORT_SYMBOL(ceph_osdc_readpages);
2697 
2698 /*
2699  * do a synchronous write on N pages
2700  */
2701 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2702 			 struct ceph_file_layout *layout,
2703 			 struct ceph_snap_context *snapc,
2704 			 u64 off, u64 len,
2705 			 u32 truncate_seq, u64 truncate_size,
2706 			 struct timespec *mtime,
2707 			 struct page **pages, int num_pages)
2708 {
2709 	struct ceph_osd_request *req;
2710 	int rc = 0;
2711 	int page_align = off & ~PAGE_MASK;
2712 
2713 	BUG_ON(vino.snap != CEPH_NOSNAP);	/* snapshots aren't writeable */
2714 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2715 				    CEPH_OSD_OP_WRITE,
2716 				    CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2717 				    snapc, truncate_seq, truncate_size,
2718 				    true);
2719 	if (IS_ERR(req))
2720 		return PTR_ERR(req);
2721 
2722 	/* it may be a short write due to an object boundary */
2723 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2724 				false, false);
2725 	dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2726 
2727 	ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2728 
2729 	rc = ceph_osdc_start_request(osdc, req, true);
2730 	if (!rc)
2731 		rc = ceph_osdc_wait_request(osdc, req);
2732 
2733 	ceph_osdc_put_request(req);
2734 	if (rc == 0)
2735 		rc = len;
2736 	dout("writepages result %d\n", rc);
2737 	return rc;
2738 }
2739 EXPORT_SYMBOL(ceph_osdc_writepages);
2740 
2741 int ceph_osdc_setup(void)
2742 {
2743 	BUG_ON(ceph_osd_request_cache);
2744 	ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2745 					sizeof (struct ceph_osd_request),
2746 					__alignof__(struct ceph_osd_request),
2747 					0, NULL);
2748 
2749 	return ceph_osd_request_cache ? 0 : -ENOMEM;
2750 }
2751 EXPORT_SYMBOL(ceph_osdc_setup);
2752 
2753 void ceph_osdc_cleanup(void)
2754 {
2755 	BUG_ON(!ceph_osd_request_cache);
2756 	kmem_cache_destroy(ceph_osd_request_cache);
2757 	ceph_osd_request_cache = NULL;
2758 }
2759 EXPORT_SYMBOL(ceph_osdc_cleanup);
2760 
2761 /*
2762  * handle incoming message
2763  */
2764 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2765 {
2766 	struct ceph_osd *osd = con->private;
2767 	struct ceph_osd_client *osdc;
2768 	int type = le16_to_cpu(msg->hdr.type);
2769 
2770 	if (!osd)
2771 		goto out;
2772 	osdc = osd->o_osdc;
2773 
2774 	switch (type) {
2775 	case CEPH_MSG_OSD_MAP:
2776 		ceph_osdc_handle_map(osdc, msg);
2777 		break;
2778 	case CEPH_MSG_OSD_OPREPLY:
2779 		handle_reply(osdc, msg, con);
2780 		break;
2781 	case CEPH_MSG_WATCH_NOTIFY:
2782 		handle_watch_notify(osdc, msg);
2783 		break;
2784 
2785 	default:
2786 		pr_err("received unknown message type %d %s\n", type,
2787 		       ceph_msg_type_name(type));
2788 	}
2789 out:
2790 	ceph_msg_put(msg);
2791 }
2792 
2793 /*
2794  * lookup and return message for incoming reply.  set up reply message
2795  * pages.
2796  */
2797 static struct ceph_msg *get_reply(struct ceph_connection *con,
2798 				  struct ceph_msg_header *hdr,
2799 				  int *skip)
2800 {
2801 	struct ceph_osd *osd = con->private;
2802 	struct ceph_osd_client *osdc = osd->o_osdc;
2803 	struct ceph_msg *m;
2804 	struct ceph_osd_request *req;
2805 	int front_len = le32_to_cpu(hdr->front_len);
2806 	int data_len = le32_to_cpu(hdr->data_len);
2807 	u64 tid;
2808 
2809 	tid = le64_to_cpu(hdr->tid);
2810 	mutex_lock(&osdc->request_mutex);
2811 	req = __lookup_request(osdc, tid);
2812 	if (!req) {
2813 		*skip = 1;
2814 		m = NULL;
2815 		dout("get_reply unknown tid %llu from osd%d\n", tid,
2816 		     osd->o_osd);
2817 		goto out;
2818 	}
2819 
2820 	if (req->r_reply->con)
2821 		dout("%s revoking msg %p from old con %p\n", __func__,
2822 		     req->r_reply, req->r_reply->con);
2823 	ceph_msg_revoke_incoming(req->r_reply);
2824 
2825 	if (front_len > req->r_reply->front_alloc_len) {
2826 		pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2827 			front_len, req->r_reply->front_alloc_len,
2828 			(unsigned int)con->peer_name.type,
2829 			le64_to_cpu(con->peer_name.num));
2830 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2831 				 false);
2832 		if (!m)
2833 			goto out;
2834 		ceph_msg_put(req->r_reply);
2835 		req->r_reply = m;
2836 	}
2837 	m = ceph_msg_get(req->r_reply);
2838 
2839 	if (data_len > 0) {
2840 		struct ceph_osd_data *osd_data;
2841 
2842 		/*
2843 		 * XXX This is assuming there is only one op containing
2844 		 * XXX page data.  Probably OK for reads, but this
2845 		 * XXX ought to be done more generally.
2846 		 */
2847 		osd_data = osd_req_op_extent_osd_data(req, 0);
2848 		if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2849 			if (osd_data->pages &&
2850 				unlikely(osd_data->length < data_len)) {
2851 
2852 				pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2853 					tid, data_len, osd_data->length);
2854 				*skip = 1;
2855 				ceph_msg_put(m);
2856 				m = NULL;
2857 				goto out;
2858 			}
2859 		}
2860 	}
2861 	*skip = 0;
2862 	dout("get_reply tid %lld %p\n", tid, m);
2863 
2864 out:
2865 	mutex_unlock(&osdc->request_mutex);
2866 	return m;
2867 
2868 }
2869 
2870 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2871 				  struct ceph_msg_header *hdr,
2872 				  int *skip)
2873 {
2874 	struct ceph_osd *osd = con->private;
2875 	int type = le16_to_cpu(hdr->type);
2876 	int front = le32_to_cpu(hdr->front_len);
2877 
2878 	*skip = 0;
2879 	switch (type) {
2880 	case CEPH_MSG_OSD_MAP:
2881 	case CEPH_MSG_WATCH_NOTIFY:
2882 		return ceph_msg_new(type, front, GFP_NOFS, false);
2883 	case CEPH_MSG_OSD_OPREPLY:
2884 		return get_reply(con, hdr, skip);
2885 	default:
2886 		pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2887 			osd->o_osd);
2888 		*skip = 1;
2889 		return NULL;
2890 	}
2891 }
2892 
2893 /*
2894  * Wrappers to refcount containing ceph_osd struct
2895  */
2896 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2897 {
2898 	struct ceph_osd *osd = con->private;
2899 	if (get_osd(osd))
2900 		return con;
2901 	return NULL;
2902 }
2903 
2904 static void put_osd_con(struct ceph_connection *con)
2905 {
2906 	struct ceph_osd *osd = con->private;
2907 	put_osd(osd);
2908 }
2909 
2910 /*
2911  * authentication
2912  */
2913 /*
2914  * Note: returned pointer is the address of a structure that's
2915  * managed separately.  Caller must *not* attempt to free it.
2916  */
2917 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2918 					int *proto, int force_new)
2919 {
2920 	struct ceph_osd *o = con->private;
2921 	struct ceph_osd_client *osdc = o->o_osdc;
2922 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2923 	struct ceph_auth_handshake *auth = &o->o_auth;
2924 
2925 	if (force_new && auth->authorizer) {
2926 		ceph_auth_destroy_authorizer(ac, auth->authorizer);
2927 		auth->authorizer = NULL;
2928 	}
2929 	if (!auth->authorizer) {
2930 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2931 						      auth);
2932 		if (ret)
2933 			return ERR_PTR(ret);
2934 	} else {
2935 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2936 						     auth);
2937 		if (ret)
2938 			return ERR_PTR(ret);
2939 	}
2940 	*proto = ac->protocol;
2941 
2942 	return auth;
2943 }
2944 
2945 
2946 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2947 {
2948 	struct ceph_osd *o = con->private;
2949 	struct ceph_osd_client *osdc = o->o_osdc;
2950 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2951 
2952 	return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2953 }
2954 
2955 static int invalidate_authorizer(struct ceph_connection *con)
2956 {
2957 	struct ceph_osd *o = con->private;
2958 	struct ceph_osd_client *osdc = o->o_osdc;
2959 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2960 
2961 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2962 	return ceph_monc_validate_auth(&osdc->client->monc);
2963 }
2964 
2965 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
2966 {
2967 	struct ceph_osd *o = con->private;
2968 	struct ceph_auth_handshake *auth = &o->o_auth;
2969 	return ceph_auth_sign_message(auth, msg);
2970 }
2971 
2972 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
2973 {
2974 	struct ceph_osd *o = con->private;
2975 	struct ceph_auth_handshake *auth = &o->o_auth;
2976 	return ceph_auth_check_message_signature(auth, msg);
2977 }
2978 
2979 static const struct ceph_connection_operations osd_con_ops = {
2980 	.get = get_osd_con,
2981 	.put = put_osd_con,
2982 	.dispatch = dispatch,
2983 	.get_authorizer = get_authorizer,
2984 	.verify_authorizer_reply = verify_authorizer_reply,
2985 	.invalidate_authorizer = invalidate_authorizer,
2986 	.alloc_msg = alloc_msg,
2987 	.sign_message = sign_message,
2988 	.check_message_signature = check_message_signature,
2989 	.fault = osd_reset,
2990 };
2991