xref: /openbmc/linux/net/ceph/osd_client.c (revision 144cba1493fdd6e3e1980e439a31df877831ebcd)
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 	case CEPH_OSD_OP_STAT:
300 		ceph_osd_data_release(&op->raw_data_in);
301 		break;
302 	default:
303 		break;
304 	}
305 }
306 
307 /*
308  * requests
309  */
310 static void ceph_osdc_release_request(struct kref *kref)
311 {
312 	struct ceph_osd_request *req = container_of(kref,
313 					    struct ceph_osd_request, r_kref);
314 	unsigned int which;
315 
316 	dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
317 	     req->r_request, req->r_reply);
318 	WARN_ON(!RB_EMPTY_NODE(&req->r_node));
319 	WARN_ON(!list_empty(&req->r_req_lru_item));
320 	WARN_ON(!list_empty(&req->r_osd_item));
321 	WARN_ON(!list_empty(&req->r_linger_item));
322 	WARN_ON(!list_empty(&req->r_linger_osd_item));
323 	WARN_ON(req->r_osd);
324 
325 	if (req->r_request)
326 		ceph_msg_put(req->r_request);
327 	if (req->r_reply) {
328 		ceph_msg_revoke_incoming(req->r_reply);
329 		ceph_msg_put(req->r_reply);
330 	}
331 
332 	for (which = 0; which < req->r_num_ops; which++)
333 		osd_req_op_data_release(req, which);
334 
335 	ceph_put_snap_context(req->r_snapc);
336 	if (req->r_mempool)
337 		mempool_free(req, req->r_osdc->req_mempool);
338 	else
339 		kmem_cache_free(ceph_osd_request_cache, req);
340 
341 }
342 
343 void ceph_osdc_get_request(struct ceph_osd_request *req)
344 {
345 	dout("%s %p (was %d)\n", __func__, req,
346 	     atomic_read(&req->r_kref.refcount));
347 	kref_get(&req->r_kref);
348 }
349 EXPORT_SYMBOL(ceph_osdc_get_request);
350 
351 void ceph_osdc_put_request(struct ceph_osd_request *req)
352 {
353 	dout("%s %p (was %d)\n", __func__, req,
354 	     atomic_read(&req->r_kref.refcount));
355 	kref_put(&req->r_kref, ceph_osdc_release_request);
356 }
357 EXPORT_SYMBOL(ceph_osdc_put_request);
358 
359 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
360 					       struct ceph_snap_context *snapc,
361 					       unsigned int num_ops,
362 					       bool use_mempool,
363 					       gfp_t gfp_flags)
364 {
365 	struct ceph_osd_request *req;
366 	struct ceph_msg *msg;
367 	size_t msg_size;
368 
369 	BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
370 	BUG_ON(num_ops > CEPH_OSD_MAX_OP);
371 
372 	msg_size = 4 + 4 + 8 + 8 + 4+8;
373 	msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
374 	msg_size += 1 + 8 + 4 + 4;     /* pg_t */
375 	msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
376 	msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
377 	msg_size += 8;  /* snapid */
378 	msg_size += 8;  /* snap_seq */
379 	msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
380 	msg_size += 4;
381 
382 	if (use_mempool) {
383 		req = mempool_alloc(osdc->req_mempool, gfp_flags);
384 		memset(req, 0, sizeof(*req));
385 	} else {
386 		req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
387 	}
388 	if (req == NULL)
389 		return NULL;
390 
391 	req->r_osdc = osdc;
392 	req->r_mempool = use_mempool;
393 	req->r_num_ops = num_ops;
394 
395 	kref_init(&req->r_kref);
396 	init_completion(&req->r_completion);
397 	init_completion(&req->r_safe_completion);
398 	RB_CLEAR_NODE(&req->r_node);
399 	INIT_LIST_HEAD(&req->r_unsafe_item);
400 	INIT_LIST_HEAD(&req->r_linger_item);
401 	INIT_LIST_HEAD(&req->r_linger_osd_item);
402 	INIT_LIST_HEAD(&req->r_req_lru_item);
403 	INIT_LIST_HEAD(&req->r_osd_item);
404 
405 	req->r_base_oloc.pool = -1;
406 	req->r_target_oloc.pool = -1;
407 
408 	/* create reply message */
409 	if (use_mempool)
410 		msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
411 	else
412 		msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
413 				   OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
414 	if (!msg) {
415 		ceph_osdc_put_request(req);
416 		return NULL;
417 	}
418 	req->r_reply = msg;
419 
420 	/* create request message; allow space for oid */
421 	if (use_mempool)
422 		msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
423 	else
424 		msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
425 	if (!msg) {
426 		ceph_osdc_put_request(req);
427 		return NULL;
428 	}
429 
430 	memset(msg->front.iov_base, 0, msg->front.iov_len);
431 
432 	req->r_request = msg;
433 
434 	return req;
435 }
436 EXPORT_SYMBOL(ceph_osdc_alloc_request);
437 
438 static bool osd_req_opcode_valid(u16 opcode)
439 {
440 	switch (opcode) {
441 #define GENERATE_CASE(op, opcode, str)	case CEPH_OSD_OP_##op: return true;
442 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
443 #undef GENERATE_CASE
444 	default:
445 		return false;
446 	}
447 }
448 
449 /*
450  * This is an osd op init function for opcodes that have no data or
451  * other information associated with them.  It also serves as a
452  * common init routine for all the other init functions, below.
453  */
454 static struct ceph_osd_req_op *
455 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
456 		 u16 opcode, u32 flags)
457 {
458 	struct ceph_osd_req_op *op;
459 
460 	BUG_ON(which >= osd_req->r_num_ops);
461 	BUG_ON(!osd_req_opcode_valid(opcode));
462 
463 	op = &osd_req->r_ops[which];
464 	memset(op, 0, sizeof (*op));
465 	op->op = opcode;
466 	op->flags = flags;
467 
468 	return op;
469 }
470 
471 void osd_req_op_init(struct ceph_osd_request *osd_req,
472 		     unsigned int which, u16 opcode, u32 flags)
473 {
474 	(void)_osd_req_op_init(osd_req, which, opcode, flags);
475 }
476 EXPORT_SYMBOL(osd_req_op_init);
477 
478 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
479 				unsigned int which, u16 opcode,
480 				u64 offset, u64 length,
481 				u64 truncate_size, u32 truncate_seq)
482 {
483 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
484 						      opcode, 0);
485 	size_t payload_len = 0;
486 
487 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
488 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE);
489 
490 	op->extent.offset = offset;
491 	op->extent.length = length;
492 	op->extent.truncate_size = truncate_size;
493 	op->extent.truncate_seq = truncate_seq;
494 	if (opcode == CEPH_OSD_OP_WRITE)
495 		payload_len += length;
496 
497 	op->payload_len = payload_len;
498 }
499 EXPORT_SYMBOL(osd_req_op_extent_init);
500 
501 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
502 				unsigned int which, u64 length)
503 {
504 	struct ceph_osd_req_op *op;
505 	u64 previous;
506 
507 	BUG_ON(which >= osd_req->r_num_ops);
508 	op = &osd_req->r_ops[which];
509 	previous = op->extent.length;
510 
511 	if (length == previous)
512 		return;		/* Nothing to do */
513 	BUG_ON(length > previous);
514 
515 	op->extent.length = length;
516 	op->payload_len -= previous - length;
517 }
518 EXPORT_SYMBOL(osd_req_op_extent_update);
519 
520 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
521 			u16 opcode, const char *class, const char *method)
522 {
523 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
524 						      opcode, 0);
525 	struct ceph_pagelist *pagelist;
526 	size_t payload_len = 0;
527 	size_t size;
528 
529 	BUG_ON(opcode != CEPH_OSD_OP_CALL);
530 
531 	pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
532 	BUG_ON(!pagelist);
533 	ceph_pagelist_init(pagelist);
534 
535 	op->cls.class_name = class;
536 	size = strlen(class);
537 	BUG_ON(size > (size_t) U8_MAX);
538 	op->cls.class_len = size;
539 	ceph_pagelist_append(pagelist, class, size);
540 	payload_len += size;
541 
542 	op->cls.method_name = method;
543 	size = strlen(method);
544 	BUG_ON(size > (size_t) U8_MAX);
545 	op->cls.method_len = size;
546 	ceph_pagelist_append(pagelist, method, size);
547 	payload_len += size;
548 
549 	osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
550 
551 	op->cls.argc = 0;	/* currently unused */
552 
553 	op->payload_len = payload_len;
554 }
555 EXPORT_SYMBOL(osd_req_op_cls_init);
556 
557 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
558 			  u16 opcode, const char *name, const void *value,
559 			  size_t size, u8 cmp_op, u8 cmp_mode)
560 {
561 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
562 						      opcode, 0);
563 	struct ceph_pagelist *pagelist;
564 	size_t payload_len;
565 
566 	BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
567 
568 	pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
569 	if (!pagelist)
570 		return -ENOMEM;
571 
572 	ceph_pagelist_init(pagelist);
573 
574 	payload_len = strlen(name);
575 	op->xattr.name_len = payload_len;
576 	ceph_pagelist_append(pagelist, name, payload_len);
577 
578 	op->xattr.value_len = size;
579 	ceph_pagelist_append(pagelist, value, size);
580 	payload_len += size;
581 
582 	op->xattr.cmp_op = cmp_op;
583 	op->xattr.cmp_mode = cmp_mode;
584 
585 	ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
586 	op->payload_len = payload_len;
587 	return 0;
588 }
589 EXPORT_SYMBOL(osd_req_op_xattr_init);
590 
591 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
592 				unsigned int which, u16 opcode,
593 				u64 cookie, u64 version, int flag)
594 {
595 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
596 						      opcode, 0);
597 
598 	BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
599 
600 	op->watch.cookie = cookie;
601 	op->watch.ver = version;
602 	if (opcode == CEPH_OSD_OP_WATCH && flag)
603 		op->watch.flag = (u8)1;
604 }
605 EXPORT_SYMBOL(osd_req_op_watch_init);
606 
607 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
608 				unsigned int which,
609 				u64 expected_object_size,
610 				u64 expected_write_size)
611 {
612 	struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
613 						      CEPH_OSD_OP_SETALLOCHINT,
614 						      0);
615 
616 	op->alloc_hint.expected_object_size = expected_object_size;
617 	op->alloc_hint.expected_write_size = expected_write_size;
618 
619 	/*
620 	 * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
621 	 * not worth a feature bit.  Set FAILOK per-op flag to make
622 	 * sure older osds don't trip over an unsupported opcode.
623 	 */
624 	op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
625 }
626 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
627 
628 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
629 				struct ceph_osd_data *osd_data)
630 {
631 	u64 length = ceph_osd_data_length(osd_data);
632 
633 	if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
634 		BUG_ON(length > (u64) SIZE_MAX);
635 		if (length)
636 			ceph_msg_data_add_pages(msg, osd_data->pages,
637 					length, osd_data->alignment);
638 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
639 		BUG_ON(!length);
640 		ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
641 #ifdef CONFIG_BLOCK
642 	} else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
643 		ceph_msg_data_add_bio(msg, osd_data->bio, length);
644 #endif
645 	} else {
646 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
647 	}
648 }
649 
650 static u64 osd_req_encode_op(struct ceph_osd_request *req,
651 			      struct ceph_osd_op *dst, unsigned int which)
652 {
653 	struct ceph_osd_req_op *src;
654 	struct ceph_osd_data *osd_data;
655 	u64 request_data_len = 0;
656 	u64 data_length;
657 
658 	BUG_ON(which >= req->r_num_ops);
659 	src = &req->r_ops[which];
660 	if (WARN_ON(!osd_req_opcode_valid(src->op))) {
661 		pr_err("unrecognized osd opcode %d\n", src->op);
662 
663 		return 0;
664 	}
665 
666 	switch (src->op) {
667 	case CEPH_OSD_OP_STAT:
668 		osd_data = &src->raw_data_in;
669 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
670 		break;
671 	case CEPH_OSD_OP_READ:
672 	case CEPH_OSD_OP_WRITE:
673 	case CEPH_OSD_OP_ZERO:
674 	case CEPH_OSD_OP_TRUNCATE:
675 		if (src->op == CEPH_OSD_OP_WRITE)
676 			request_data_len = src->extent.length;
677 		dst->extent.offset = cpu_to_le64(src->extent.offset);
678 		dst->extent.length = cpu_to_le64(src->extent.length);
679 		dst->extent.truncate_size =
680 			cpu_to_le64(src->extent.truncate_size);
681 		dst->extent.truncate_seq =
682 			cpu_to_le32(src->extent.truncate_seq);
683 		osd_data = &src->extent.osd_data;
684 		if (src->op == CEPH_OSD_OP_WRITE)
685 			ceph_osdc_msg_data_add(req->r_request, osd_data);
686 		else
687 			ceph_osdc_msg_data_add(req->r_reply, osd_data);
688 		break;
689 	case CEPH_OSD_OP_CALL:
690 		dst->cls.class_len = src->cls.class_len;
691 		dst->cls.method_len = src->cls.method_len;
692 		osd_data = &src->cls.request_info;
693 		ceph_osdc_msg_data_add(req->r_request, osd_data);
694 		BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
695 		request_data_len = osd_data->pagelist->length;
696 
697 		osd_data = &src->cls.request_data;
698 		data_length = ceph_osd_data_length(osd_data);
699 		if (data_length) {
700 			BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
701 			dst->cls.indata_len = cpu_to_le32(data_length);
702 			ceph_osdc_msg_data_add(req->r_request, osd_data);
703 			src->payload_len += data_length;
704 			request_data_len += data_length;
705 		}
706 		osd_data = &src->cls.response_data;
707 		ceph_osdc_msg_data_add(req->r_reply, osd_data);
708 		break;
709 	case CEPH_OSD_OP_STARTSYNC:
710 		break;
711 	case CEPH_OSD_OP_NOTIFY_ACK:
712 	case CEPH_OSD_OP_WATCH:
713 		dst->watch.cookie = cpu_to_le64(src->watch.cookie);
714 		dst->watch.ver = cpu_to_le64(src->watch.ver);
715 		dst->watch.flag = src->watch.flag;
716 		break;
717 	case CEPH_OSD_OP_SETALLOCHINT:
718 		dst->alloc_hint.expected_object_size =
719 		    cpu_to_le64(src->alloc_hint.expected_object_size);
720 		dst->alloc_hint.expected_write_size =
721 		    cpu_to_le64(src->alloc_hint.expected_write_size);
722 		break;
723 	case CEPH_OSD_OP_SETXATTR:
724 	case CEPH_OSD_OP_CMPXATTR:
725 		dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
726 		dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
727 		dst->xattr.cmp_op = src->xattr.cmp_op;
728 		dst->xattr.cmp_mode = src->xattr.cmp_mode;
729 		osd_data = &src->xattr.osd_data;
730 		ceph_osdc_msg_data_add(req->r_request, osd_data);
731 		request_data_len = osd_data->pagelist->length;
732 		break;
733 	case CEPH_OSD_OP_CREATE:
734 	case CEPH_OSD_OP_DELETE:
735 		break;
736 	default:
737 		pr_err("unsupported osd opcode %s\n",
738 			ceph_osd_op_name(src->op));
739 		WARN_ON(1);
740 
741 		return 0;
742 	}
743 
744 	dst->op = cpu_to_le16(src->op);
745 	dst->flags = cpu_to_le32(src->flags);
746 	dst->payload_len = cpu_to_le32(src->payload_len);
747 
748 	return request_data_len;
749 }
750 
751 /*
752  * build new request AND message, calculate layout, and adjust file
753  * extent as needed.
754  *
755  * if the file was recently truncated, we include information about its
756  * old and new size so that the object can be updated appropriately.  (we
757  * avoid synchronously deleting truncated objects because it's slow.)
758  *
759  * if @do_sync, include a 'startsync' command so that the osd will flush
760  * data quickly.
761  */
762 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
763 					       struct ceph_file_layout *layout,
764 					       struct ceph_vino vino,
765 					       u64 off, u64 *plen,
766 					       unsigned int which, int num_ops,
767 					       int opcode, int flags,
768 					       struct ceph_snap_context *snapc,
769 					       u32 truncate_seq,
770 					       u64 truncate_size,
771 					       bool use_mempool)
772 {
773 	struct ceph_osd_request *req;
774 	u64 objnum = 0;
775 	u64 objoff = 0;
776 	u64 objlen = 0;
777 	int r;
778 
779 	BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
780 	       opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
781 	       opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
782 
783 	req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
784 					GFP_NOFS);
785 	if (!req)
786 		return ERR_PTR(-ENOMEM);
787 
788 	req->r_flags = flags;
789 
790 	/* calculate max write size */
791 	r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
792 	if (r < 0) {
793 		ceph_osdc_put_request(req);
794 		return ERR_PTR(r);
795 	}
796 
797 	if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
798 		osd_req_op_init(req, which, opcode, 0);
799 	} else {
800 		u32 object_size = le32_to_cpu(layout->fl_object_size);
801 		u32 object_base = off - objoff;
802 		if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
803 			if (truncate_size <= object_base) {
804 				truncate_size = 0;
805 			} else {
806 				truncate_size -= object_base;
807 				if (truncate_size > object_size)
808 					truncate_size = object_size;
809 			}
810 		}
811 		osd_req_op_extent_init(req, which, opcode, objoff, objlen,
812 				       truncate_size, truncate_seq);
813 	}
814 
815 	req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
816 
817 	snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
818 		 "%llx.%08llx", vino.ino, objnum);
819 	req->r_base_oid.name_len = strlen(req->r_base_oid.name);
820 
821 	return req;
822 }
823 EXPORT_SYMBOL(ceph_osdc_new_request);
824 
825 /*
826  * We keep osd requests in an rbtree, sorted by ->r_tid.
827  */
828 static void __insert_request(struct ceph_osd_client *osdc,
829 			     struct ceph_osd_request *new)
830 {
831 	struct rb_node **p = &osdc->requests.rb_node;
832 	struct rb_node *parent = NULL;
833 	struct ceph_osd_request *req = NULL;
834 
835 	while (*p) {
836 		parent = *p;
837 		req = rb_entry(parent, struct ceph_osd_request, r_node);
838 		if (new->r_tid < req->r_tid)
839 			p = &(*p)->rb_left;
840 		else if (new->r_tid > req->r_tid)
841 			p = &(*p)->rb_right;
842 		else
843 			BUG();
844 	}
845 
846 	rb_link_node(&new->r_node, parent, p);
847 	rb_insert_color(&new->r_node, &osdc->requests);
848 }
849 
850 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
851 						 u64 tid)
852 {
853 	struct ceph_osd_request *req;
854 	struct rb_node *n = osdc->requests.rb_node;
855 
856 	while (n) {
857 		req = rb_entry(n, struct ceph_osd_request, r_node);
858 		if (tid < req->r_tid)
859 			n = n->rb_left;
860 		else if (tid > req->r_tid)
861 			n = n->rb_right;
862 		else
863 			return req;
864 	}
865 	return NULL;
866 }
867 
868 static struct ceph_osd_request *
869 __lookup_request_ge(struct ceph_osd_client *osdc,
870 		    u64 tid)
871 {
872 	struct ceph_osd_request *req;
873 	struct rb_node *n = osdc->requests.rb_node;
874 
875 	while (n) {
876 		req = rb_entry(n, struct ceph_osd_request, r_node);
877 		if (tid < req->r_tid) {
878 			if (!n->rb_left)
879 				return req;
880 			n = n->rb_left;
881 		} else if (tid > req->r_tid) {
882 			n = n->rb_right;
883 		} else {
884 			return req;
885 		}
886 	}
887 	return NULL;
888 }
889 
890 static void __kick_linger_request(struct ceph_osd_request *req)
891 {
892 	struct ceph_osd_client *osdc = req->r_osdc;
893 	struct ceph_osd *osd = req->r_osd;
894 
895 	/*
896 	 * Linger requests need to be resent with a new tid to avoid
897 	 * the dup op detection logic on the OSDs.  Achieve this with
898 	 * a re-register dance instead of open-coding.
899 	 */
900 	ceph_osdc_get_request(req);
901 	if (!list_empty(&req->r_linger_item))
902 		__unregister_linger_request(osdc, req);
903 	else
904 		__unregister_request(osdc, req);
905 	__register_request(osdc, req);
906 	ceph_osdc_put_request(req);
907 
908 	/*
909 	 * Unless request has been registered as both normal and
910 	 * lingering, __unregister{,_linger}_request clears r_osd.
911 	 * However, here we need to preserve r_osd to make sure we
912 	 * requeue on the same OSD.
913 	 */
914 	WARN_ON(req->r_osd || !osd);
915 	req->r_osd = osd;
916 
917 	dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
918 	__enqueue_request(req);
919 }
920 
921 /*
922  * Resubmit requests pending on the given osd.
923  */
924 static void __kick_osd_requests(struct ceph_osd_client *osdc,
925 				struct ceph_osd *osd)
926 {
927 	struct ceph_osd_request *req, *nreq;
928 	LIST_HEAD(resend);
929 	LIST_HEAD(resend_linger);
930 	int err;
931 
932 	dout("%s osd%d\n", __func__, osd->o_osd);
933 	err = __reset_osd(osdc, osd);
934 	if (err)
935 		return;
936 
937 	/*
938 	 * Build up a list of requests to resend by traversing the
939 	 * osd's list of requests.  Requests for a given object are
940 	 * sent in tid order, and that is also the order they're
941 	 * kept on this list.  Therefore all requests that are in
942 	 * flight will be found first, followed by all requests that
943 	 * have not yet been sent.  And to resend requests while
944 	 * preserving this order we will want to put any sent
945 	 * requests back on the front of the osd client's unsent
946 	 * list.
947 	 *
948 	 * So we build a separate ordered list of already-sent
949 	 * requests for the affected osd and splice it onto the
950 	 * front of the osd client's unsent list.  Once we've seen a
951 	 * request that has not yet been sent we're done.  Those
952 	 * requests are already sitting right where they belong.
953 	 */
954 	list_for_each_entry(req, &osd->o_requests, r_osd_item) {
955 		if (!req->r_sent)
956 			break;
957 
958 		if (!req->r_linger) {
959 			dout("%s requeueing %p tid %llu\n", __func__, req,
960 			     req->r_tid);
961 			list_move_tail(&req->r_req_lru_item, &resend);
962 			req->r_flags |= CEPH_OSD_FLAG_RETRY;
963 		} else {
964 			list_move_tail(&req->r_req_lru_item, &resend_linger);
965 		}
966 	}
967 	list_splice(&resend, &osdc->req_unsent);
968 
969 	/*
970 	 * Both registered and not yet registered linger requests are
971 	 * enqueued with a new tid on the same OSD.  We add/move them
972 	 * to req_unsent/o_requests at the end to keep things in tid
973 	 * order.
974 	 */
975 	list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
976 				 r_linger_osd_item) {
977 		WARN_ON(!list_empty(&req->r_req_lru_item));
978 		__kick_linger_request(req);
979 	}
980 
981 	list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
982 		__kick_linger_request(req);
983 }
984 
985 /*
986  * If the osd connection drops, we need to resubmit all requests.
987  */
988 static void osd_reset(struct ceph_connection *con)
989 {
990 	struct ceph_osd *osd = con->private;
991 	struct ceph_osd_client *osdc;
992 
993 	if (!osd)
994 		return;
995 	dout("osd_reset osd%d\n", osd->o_osd);
996 	osdc = osd->o_osdc;
997 	down_read(&osdc->map_sem);
998 	mutex_lock(&osdc->request_mutex);
999 	__kick_osd_requests(osdc, osd);
1000 	__send_queued(osdc);
1001 	mutex_unlock(&osdc->request_mutex);
1002 	up_read(&osdc->map_sem);
1003 }
1004 
1005 /*
1006  * Track open sessions with osds.
1007  */
1008 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1009 {
1010 	struct ceph_osd *osd;
1011 
1012 	osd = kzalloc(sizeof(*osd), GFP_NOFS);
1013 	if (!osd)
1014 		return NULL;
1015 
1016 	atomic_set(&osd->o_ref, 1);
1017 	osd->o_osdc = osdc;
1018 	osd->o_osd = onum;
1019 	RB_CLEAR_NODE(&osd->o_node);
1020 	INIT_LIST_HEAD(&osd->o_requests);
1021 	INIT_LIST_HEAD(&osd->o_linger_requests);
1022 	INIT_LIST_HEAD(&osd->o_osd_lru);
1023 	osd->o_incarnation = 1;
1024 
1025 	ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1026 
1027 	INIT_LIST_HEAD(&osd->o_keepalive_item);
1028 	return osd;
1029 }
1030 
1031 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1032 {
1033 	if (atomic_inc_not_zero(&osd->o_ref)) {
1034 		dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1035 		     atomic_read(&osd->o_ref));
1036 		return osd;
1037 	} else {
1038 		dout("get_osd %p FAIL\n", osd);
1039 		return NULL;
1040 	}
1041 }
1042 
1043 static void put_osd(struct ceph_osd *osd)
1044 {
1045 	dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1046 	     atomic_read(&osd->o_ref) - 1);
1047 	if (atomic_dec_and_test(&osd->o_ref)) {
1048 		struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1049 
1050 		if (osd->o_auth.authorizer)
1051 			ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1052 		kfree(osd);
1053 	}
1054 }
1055 
1056 /*
1057  * remove an osd from our map
1058  */
1059 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1060 {
1061 	dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1062 	WARN_ON(!list_empty(&osd->o_requests));
1063 	WARN_ON(!list_empty(&osd->o_linger_requests));
1064 
1065 	list_del_init(&osd->o_osd_lru);
1066 	rb_erase(&osd->o_node, &osdc->osds);
1067 	RB_CLEAR_NODE(&osd->o_node);
1068 }
1069 
1070 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1071 {
1072 	dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1073 
1074 	if (!RB_EMPTY_NODE(&osd->o_node)) {
1075 		ceph_con_close(&osd->o_con);
1076 		__remove_osd(osdc, osd);
1077 		put_osd(osd);
1078 	}
1079 }
1080 
1081 static void remove_all_osds(struct ceph_osd_client *osdc)
1082 {
1083 	dout("%s %p\n", __func__, osdc);
1084 	mutex_lock(&osdc->request_mutex);
1085 	while (!RB_EMPTY_ROOT(&osdc->osds)) {
1086 		struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1087 						struct ceph_osd, o_node);
1088 		remove_osd(osdc, osd);
1089 	}
1090 	mutex_unlock(&osdc->request_mutex);
1091 }
1092 
1093 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1094 			      struct ceph_osd *osd)
1095 {
1096 	dout("%s %p\n", __func__, osd);
1097 	BUG_ON(!list_empty(&osd->o_osd_lru));
1098 
1099 	list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1100 	osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1101 }
1102 
1103 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1104 				  struct ceph_osd *osd)
1105 {
1106 	dout("%s %p\n", __func__, osd);
1107 
1108 	if (list_empty(&osd->o_requests) &&
1109 	    list_empty(&osd->o_linger_requests))
1110 		__move_osd_to_lru(osdc, osd);
1111 }
1112 
1113 static void __remove_osd_from_lru(struct ceph_osd *osd)
1114 {
1115 	dout("__remove_osd_from_lru %p\n", osd);
1116 	if (!list_empty(&osd->o_osd_lru))
1117 		list_del_init(&osd->o_osd_lru);
1118 }
1119 
1120 static void remove_old_osds(struct ceph_osd_client *osdc)
1121 {
1122 	struct ceph_osd *osd, *nosd;
1123 
1124 	dout("__remove_old_osds %p\n", osdc);
1125 	mutex_lock(&osdc->request_mutex);
1126 	list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1127 		if (time_before(jiffies, osd->lru_ttl))
1128 			break;
1129 		remove_osd(osdc, osd);
1130 	}
1131 	mutex_unlock(&osdc->request_mutex);
1132 }
1133 
1134 /*
1135  * reset osd connect
1136  */
1137 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1138 {
1139 	struct ceph_entity_addr *peer_addr;
1140 
1141 	dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1142 	if (list_empty(&osd->o_requests) &&
1143 	    list_empty(&osd->o_linger_requests)) {
1144 		remove_osd(osdc, osd);
1145 		return -ENODEV;
1146 	}
1147 
1148 	peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1149 	if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1150 			!ceph_con_opened(&osd->o_con)) {
1151 		struct ceph_osd_request *req;
1152 
1153 		dout("osd addr hasn't changed and connection never opened, "
1154 		     "letting msgr retry\n");
1155 		/* touch each r_stamp for handle_timeout()'s benfit */
1156 		list_for_each_entry(req, &osd->o_requests, r_osd_item)
1157 			req->r_stamp = jiffies;
1158 
1159 		return -EAGAIN;
1160 	}
1161 
1162 	ceph_con_close(&osd->o_con);
1163 	ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1164 	osd->o_incarnation++;
1165 
1166 	return 0;
1167 }
1168 
1169 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1170 {
1171 	struct rb_node **p = &osdc->osds.rb_node;
1172 	struct rb_node *parent = NULL;
1173 	struct ceph_osd *osd = NULL;
1174 
1175 	dout("__insert_osd %p osd%d\n", new, new->o_osd);
1176 	while (*p) {
1177 		parent = *p;
1178 		osd = rb_entry(parent, struct ceph_osd, o_node);
1179 		if (new->o_osd < osd->o_osd)
1180 			p = &(*p)->rb_left;
1181 		else if (new->o_osd > osd->o_osd)
1182 			p = &(*p)->rb_right;
1183 		else
1184 			BUG();
1185 	}
1186 
1187 	rb_link_node(&new->o_node, parent, p);
1188 	rb_insert_color(&new->o_node, &osdc->osds);
1189 }
1190 
1191 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1192 {
1193 	struct ceph_osd *osd;
1194 	struct rb_node *n = osdc->osds.rb_node;
1195 
1196 	while (n) {
1197 		osd = rb_entry(n, struct ceph_osd, o_node);
1198 		if (o < osd->o_osd)
1199 			n = n->rb_left;
1200 		else if (o > osd->o_osd)
1201 			n = n->rb_right;
1202 		else
1203 			return osd;
1204 	}
1205 	return NULL;
1206 }
1207 
1208 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1209 {
1210 	schedule_delayed_work(&osdc->timeout_work,
1211 			osdc->client->options->osd_keepalive_timeout * HZ);
1212 }
1213 
1214 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1215 {
1216 	cancel_delayed_work(&osdc->timeout_work);
1217 }
1218 
1219 /*
1220  * Register request, assign tid.  If this is the first request, set up
1221  * the timeout event.
1222  */
1223 static void __register_request(struct ceph_osd_client *osdc,
1224 			       struct ceph_osd_request *req)
1225 {
1226 	req->r_tid = ++osdc->last_tid;
1227 	req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1228 	dout("__register_request %p tid %lld\n", req, req->r_tid);
1229 	__insert_request(osdc, req);
1230 	ceph_osdc_get_request(req);
1231 	osdc->num_requests++;
1232 	if (osdc->num_requests == 1) {
1233 		dout(" first request, scheduling timeout\n");
1234 		__schedule_osd_timeout(osdc);
1235 	}
1236 }
1237 
1238 /*
1239  * called under osdc->request_mutex
1240  */
1241 static void __unregister_request(struct ceph_osd_client *osdc,
1242 				 struct ceph_osd_request *req)
1243 {
1244 	if (RB_EMPTY_NODE(&req->r_node)) {
1245 		dout("__unregister_request %p tid %lld not registered\n",
1246 			req, req->r_tid);
1247 		return;
1248 	}
1249 
1250 	dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1251 	rb_erase(&req->r_node, &osdc->requests);
1252 	RB_CLEAR_NODE(&req->r_node);
1253 	osdc->num_requests--;
1254 
1255 	if (req->r_osd) {
1256 		/* make sure the original request isn't in flight. */
1257 		ceph_msg_revoke(req->r_request);
1258 
1259 		list_del_init(&req->r_osd_item);
1260 		maybe_move_osd_to_lru(osdc, req->r_osd);
1261 		if (list_empty(&req->r_linger_osd_item))
1262 			req->r_osd = NULL;
1263 	}
1264 
1265 	list_del_init(&req->r_req_lru_item);
1266 	ceph_osdc_put_request(req);
1267 
1268 	if (osdc->num_requests == 0) {
1269 		dout(" no requests, canceling timeout\n");
1270 		__cancel_osd_timeout(osdc);
1271 	}
1272 }
1273 
1274 /*
1275  * Cancel a previously queued request message
1276  */
1277 static void __cancel_request(struct ceph_osd_request *req)
1278 {
1279 	if (req->r_sent && req->r_osd) {
1280 		ceph_msg_revoke(req->r_request);
1281 		req->r_sent = 0;
1282 	}
1283 }
1284 
1285 static void __register_linger_request(struct ceph_osd_client *osdc,
1286 				    struct ceph_osd_request *req)
1287 {
1288 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1289 	WARN_ON(!req->r_linger);
1290 
1291 	ceph_osdc_get_request(req);
1292 	list_add_tail(&req->r_linger_item, &osdc->req_linger);
1293 	if (req->r_osd)
1294 		list_add_tail(&req->r_linger_osd_item,
1295 			      &req->r_osd->o_linger_requests);
1296 }
1297 
1298 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1299 					struct ceph_osd_request *req)
1300 {
1301 	WARN_ON(!req->r_linger);
1302 
1303 	if (list_empty(&req->r_linger_item)) {
1304 		dout("%s %p tid %llu not registered\n", __func__, req,
1305 		     req->r_tid);
1306 		return;
1307 	}
1308 
1309 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1310 	list_del_init(&req->r_linger_item);
1311 
1312 	if (req->r_osd) {
1313 		list_del_init(&req->r_linger_osd_item);
1314 		maybe_move_osd_to_lru(osdc, req->r_osd);
1315 		if (list_empty(&req->r_osd_item))
1316 			req->r_osd = NULL;
1317 	}
1318 	ceph_osdc_put_request(req);
1319 }
1320 
1321 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1322 				  struct ceph_osd_request *req)
1323 {
1324 	if (!req->r_linger) {
1325 		dout("set_request_linger %p\n", req);
1326 		req->r_linger = 1;
1327 	}
1328 }
1329 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1330 
1331 /*
1332  * Returns whether a request should be blocked from being sent
1333  * based on the current osdmap and osd_client settings.
1334  *
1335  * Caller should hold map_sem for read.
1336  */
1337 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1338 				   struct ceph_osd_request *req)
1339 {
1340 	bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1341 	bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1342 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1343 	return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1344 		(req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1345 }
1346 
1347 /*
1348  * Calculate mapping of a request to a PG.  Takes tiering into account.
1349  */
1350 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1351 			     struct ceph_osd_request *req,
1352 			     struct ceph_pg *pg_out)
1353 {
1354 	bool need_check_tiering;
1355 
1356 	need_check_tiering = false;
1357 	if (req->r_target_oloc.pool == -1) {
1358 		req->r_target_oloc = req->r_base_oloc; /* struct */
1359 		need_check_tiering = true;
1360 	}
1361 	if (req->r_target_oid.name_len == 0) {
1362 		ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1363 		need_check_tiering = true;
1364 	}
1365 
1366 	if (need_check_tiering &&
1367 	    (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1368 		struct ceph_pg_pool_info *pi;
1369 
1370 		pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1371 		if (pi) {
1372 			if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1373 			    pi->read_tier >= 0)
1374 				req->r_target_oloc.pool = pi->read_tier;
1375 			if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1376 			    pi->write_tier >= 0)
1377 				req->r_target_oloc.pool = pi->write_tier;
1378 		}
1379 		/* !pi is caught in ceph_oloc_oid_to_pg() */
1380 	}
1381 
1382 	return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1383 				   &req->r_target_oid, pg_out);
1384 }
1385 
1386 static void __enqueue_request(struct ceph_osd_request *req)
1387 {
1388 	struct ceph_osd_client *osdc = req->r_osdc;
1389 
1390 	dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1391 	     req->r_osd ? req->r_osd->o_osd : -1);
1392 
1393 	if (req->r_osd) {
1394 		__remove_osd_from_lru(req->r_osd);
1395 		list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1396 		list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1397 	} else {
1398 		list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1399 	}
1400 }
1401 
1402 /*
1403  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1404  * (as needed), and set the request r_osd appropriately.  If there is
1405  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1406  * (unsent, homeless) or leave on in-flight lru.
1407  *
1408  * Return 0 if unchanged, 1 if changed, or negative on error.
1409  *
1410  * Caller should hold map_sem for read and request_mutex.
1411  */
1412 static int __map_request(struct ceph_osd_client *osdc,
1413 			 struct ceph_osd_request *req, int force_resend)
1414 {
1415 	struct ceph_pg pgid;
1416 	int acting[CEPH_PG_MAX_SIZE];
1417 	int num, o;
1418 	int err;
1419 	bool was_paused;
1420 
1421 	dout("map_request %p tid %lld\n", req, req->r_tid);
1422 
1423 	err = __calc_request_pg(osdc->osdmap, req, &pgid);
1424 	if (err) {
1425 		list_move(&req->r_req_lru_item, &osdc->req_notarget);
1426 		return err;
1427 	}
1428 	req->r_pgid = pgid;
1429 
1430 	num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1431 	if (num < 0)
1432 		num = 0;
1433 
1434 	was_paused = req->r_paused;
1435 	req->r_paused = __req_should_be_paused(osdc, req);
1436 	if (was_paused && !req->r_paused)
1437 		force_resend = 1;
1438 
1439 	if ((!force_resend &&
1440 	     req->r_osd && req->r_osd->o_osd == o &&
1441 	     req->r_sent >= req->r_osd->o_incarnation &&
1442 	     req->r_num_pg_osds == num &&
1443 	     memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1444 	    (req->r_osd == NULL && o == -1) ||
1445 	    req->r_paused)
1446 		return 0;  /* no change */
1447 
1448 	dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1449 	     req->r_tid, pgid.pool, pgid.seed, o,
1450 	     req->r_osd ? req->r_osd->o_osd : -1);
1451 
1452 	/* record full pg acting set */
1453 	memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1454 	req->r_num_pg_osds = num;
1455 
1456 	if (req->r_osd) {
1457 		__cancel_request(req);
1458 		list_del_init(&req->r_osd_item);
1459 		list_del_init(&req->r_linger_osd_item);
1460 		req->r_osd = NULL;
1461 	}
1462 
1463 	req->r_osd = __lookup_osd(osdc, o);
1464 	if (!req->r_osd && o >= 0) {
1465 		err = -ENOMEM;
1466 		req->r_osd = create_osd(osdc, o);
1467 		if (!req->r_osd) {
1468 			list_move(&req->r_req_lru_item, &osdc->req_notarget);
1469 			goto out;
1470 		}
1471 
1472 		dout("map_request osd %p is osd%d\n", req->r_osd, o);
1473 		__insert_osd(osdc, req->r_osd);
1474 
1475 		ceph_con_open(&req->r_osd->o_con,
1476 			      CEPH_ENTITY_TYPE_OSD, o,
1477 			      &osdc->osdmap->osd_addr[o]);
1478 	}
1479 
1480 	__enqueue_request(req);
1481 	err = 1;   /* osd or pg changed */
1482 
1483 out:
1484 	return err;
1485 }
1486 
1487 /*
1488  * caller should hold map_sem (for read) and request_mutex
1489  */
1490 static void __send_request(struct ceph_osd_client *osdc,
1491 			   struct ceph_osd_request *req)
1492 {
1493 	void *p;
1494 
1495 	dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1496 	     req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1497 	     (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1498 
1499 	/* fill in message content that changes each time we send it */
1500 	put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1501 	put_unaligned_le32(req->r_flags, req->r_request_flags);
1502 	put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1503 	p = req->r_request_pgid;
1504 	ceph_encode_64(&p, req->r_pgid.pool);
1505 	ceph_encode_32(&p, req->r_pgid.seed);
1506 	put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1507 	memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1508 	       sizeof(req->r_reassert_version));
1509 
1510 	req->r_stamp = jiffies;
1511 	list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1512 
1513 	ceph_msg_get(req->r_request); /* send consumes a ref */
1514 
1515 	req->r_sent = req->r_osd->o_incarnation;
1516 
1517 	ceph_con_send(&req->r_osd->o_con, req->r_request);
1518 }
1519 
1520 /*
1521  * Send any requests in the queue (req_unsent).
1522  */
1523 static void __send_queued(struct ceph_osd_client *osdc)
1524 {
1525 	struct ceph_osd_request *req, *tmp;
1526 
1527 	dout("__send_queued\n");
1528 	list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1529 		__send_request(osdc, req);
1530 }
1531 
1532 /*
1533  * Caller should hold map_sem for read and request_mutex.
1534  */
1535 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1536 				     struct ceph_osd_request *req,
1537 				     bool nofail)
1538 {
1539 	int rc;
1540 
1541 	__register_request(osdc, req);
1542 	req->r_sent = 0;
1543 	req->r_got_reply = 0;
1544 	rc = __map_request(osdc, req, 0);
1545 	if (rc < 0) {
1546 		if (nofail) {
1547 			dout("osdc_start_request failed map, "
1548 				" will retry %lld\n", req->r_tid);
1549 			rc = 0;
1550 		} else {
1551 			__unregister_request(osdc, req);
1552 		}
1553 		return rc;
1554 	}
1555 
1556 	if (req->r_osd == NULL) {
1557 		dout("send_request %p no up osds in pg\n", req);
1558 		ceph_monc_request_next_osdmap(&osdc->client->monc);
1559 	} else {
1560 		__send_queued(osdc);
1561 	}
1562 
1563 	return 0;
1564 }
1565 
1566 /*
1567  * Timeout callback, called every N seconds when 1 or more osd
1568  * requests has been active for more than N seconds.  When this
1569  * happens, we ping all OSDs with requests who have timed out to
1570  * ensure any communications channel reset is detected.  Reset the
1571  * request timeouts another N seconds in the future as we go.
1572  * Reschedule the timeout event another N seconds in future (unless
1573  * there are no open requests).
1574  */
1575 static void handle_timeout(struct work_struct *work)
1576 {
1577 	struct ceph_osd_client *osdc =
1578 		container_of(work, struct ceph_osd_client, timeout_work.work);
1579 	struct ceph_osd_request *req;
1580 	struct ceph_osd *osd;
1581 	unsigned long keepalive =
1582 		osdc->client->options->osd_keepalive_timeout * HZ;
1583 	struct list_head slow_osds;
1584 	dout("timeout\n");
1585 	down_read(&osdc->map_sem);
1586 
1587 	ceph_monc_request_next_osdmap(&osdc->client->monc);
1588 
1589 	mutex_lock(&osdc->request_mutex);
1590 
1591 	/*
1592 	 * ping osds that are a bit slow.  this ensures that if there
1593 	 * is a break in the TCP connection we will notice, and reopen
1594 	 * a connection with that osd (from the fault callback).
1595 	 */
1596 	INIT_LIST_HEAD(&slow_osds);
1597 	list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1598 		if (time_before(jiffies, req->r_stamp + keepalive))
1599 			break;
1600 
1601 		osd = req->r_osd;
1602 		BUG_ON(!osd);
1603 		dout(" tid %llu is slow, will send keepalive on osd%d\n",
1604 		     req->r_tid, osd->o_osd);
1605 		list_move_tail(&osd->o_keepalive_item, &slow_osds);
1606 	}
1607 	while (!list_empty(&slow_osds)) {
1608 		osd = list_entry(slow_osds.next, struct ceph_osd,
1609 				 o_keepalive_item);
1610 		list_del_init(&osd->o_keepalive_item);
1611 		ceph_con_keepalive(&osd->o_con);
1612 	}
1613 
1614 	__schedule_osd_timeout(osdc);
1615 	__send_queued(osdc);
1616 	mutex_unlock(&osdc->request_mutex);
1617 	up_read(&osdc->map_sem);
1618 }
1619 
1620 static void handle_osds_timeout(struct work_struct *work)
1621 {
1622 	struct ceph_osd_client *osdc =
1623 		container_of(work, struct ceph_osd_client,
1624 			     osds_timeout_work.work);
1625 	unsigned long delay =
1626 		osdc->client->options->osd_idle_ttl * HZ >> 2;
1627 
1628 	dout("osds timeout\n");
1629 	down_read(&osdc->map_sem);
1630 	remove_old_osds(osdc);
1631 	up_read(&osdc->map_sem);
1632 
1633 	schedule_delayed_work(&osdc->osds_timeout_work,
1634 			      round_jiffies_relative(delay));
1635 }
1636 
1637 static int ceph_oloc_decode(void **p, void *end,
1638 			    struct ceph_object_locator *oloc)
1639 {
1640 	u8 struct_v, struct_cv;
1641 	u32 len;
1642 	void *struct_end;
1643 	int ret = 0;
1644 
1645 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1646 	struct_v = ceph_decode_8(p);
1647 	struct_cv = ceph_decode_8(p);
1648 	if (struct_v < 3) {
1649 		pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1650 			struct_v, struct_cv);
1651 		goto e_inval;
1652 	}
1653 	if (struct_cv > 6) {
1654 		pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1655 			struct_v, struct_cv);
1656 		goto e_inval;
1657 	}
1658 	len = ceph_decode_32(p);
1659 	ceph_decode_need(p, end, len, e_inval);
1660 	struct_end = *p + len;
1661 
1662 	oloc->pool = ceph_decode_64(p);
1663 	*p += 4; /* skip preferred */
1664 
1665 	len = ceph_decode_32(p);
1666 	if (len > 0) {
1667 		pr_warn("ceph_object_locator::key is set\n");
1668 		goto e_inval;
1669 	}
1670 
1671 	if (struct_v >= 5) {
1672 		len = ceph_decode_32(p);
1673 		if (len > 0) {
1674 			pr_warn("ceph_object_locator::nspace is set\n");
1675 			goto e_inval;
1676 		}
1677 	}
1678 
1679 	if (struct_v >= 6) {
1680 		s64 hash = ceph_decode_64(p);
1681 		if (hash != -1) {
1682 			pr_warn("ceph_object_locator::hash is set\n");
1683 			goto e_inval;
1684 		}
1685 	}
1686 
1687 	/* skip the rest */
1688 	*p = struct_end;
1689 out:
1690 	return ret;
1691 
1692 e_inval:
1693 	ret = -EINVAL;
1694 	goto out;
1695 }
1696 
1697 static int ceph_redirect_decode(void **p, void *end,
1698 				struct ceph_request_redirect *redir)
1699 {
1700 	u8 struct_v, struct_cv;
1701 	u32 len;
1702 	void *struct_end;
1703 	int ret;
1704 
1705 	ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1706 	struct_v = ceph_decode_8(p);
1707 	struct_cv = ceph_decode_8(p);
1708 	if (struct_cv > 1) {
1709 		pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1710 			struct_v, struct_cv);
1711 		goto e_inval;
1712 	}
1713 	len = ceph_decode_32(p);
1714 	ceph_decode_need(p, end, len, e_inval);
1715 	struct_end = *p + len;
1716 
1717 	ret = ceph_oloc_decode(p, end, &redir->oloc);
1718 	if (ret)
1719 		goto out;
1720 
1721 	len = ceph_decode_32(p);
1722 	if (len > 0) {
1723 		pr_warn("ceph_request_redirect::object_name is set\n");
1724 		goto e_inval;
1725 	}
1726 
1727 	len = ceph_decode_32(p);
1728 	*p += len; /* skip osd_instructions */
1729 
1730 	/* skip the rest */
1731 	*p = struct_end;
1732 out:
1733 	return ret;
1734 
1735 e_inval:
1736 	ret = -EINVAL;
1737 	goto out;
1738 }
1739 
1740 static void complete_request(struct ceph_osd_request *req)
1741 {
1742 	complete_all(&req->r_safe_completion);  /* fsync waiter */
1743 }
1744 
1745 /*
1746  * handle osd op reply.  either call the callback if it is specified,
1747  * or do the completion to wake up the waiting thread.
1748  */
1749 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1750 			 struct ceph_connection *con)
1751 {
1752 	void *p, *end;
1753 	struct ceph_osd_request *req;
1754 	struct ceph_request_redirect redir;
1755 	u64 tid;
1756 	int object_len;
1757 	unsigned int numops;
1758 	int payload_len, flags;
1759 	s32 result;
1760 	s32 retry_attempt;
1761 	struct ceph_pg pg;
1762 	int err;
1763 	u32 reassert_epoch;
1764 	u64 reassert_version;
1765 	u32 osdmap_epoch;
1766 	int already_completed;
1767 	u32 bytes;
1768 	unsigned int i;
1769 
1770 	tid = le64_to_cpu(msg->hdr.tid);
1771 	dout("handle_reply %p tid %llu\n", msg, tid);
1772 
1773 	p = msg->front.iov_base;
1774 	end = p + msg->front.iov_len;
1775 
1776 	ceph_decode_need(&p, end, 4, bad);
1777 	object_len = ceph_decode_32(&p);
1778 	ceph_decode_need(&p, end, object_len, bad);
1779 	p += object_len;
1780 
1781 	err = ceph_decode_pgid(&p, end, &pg);
1782 	if (err)
1783 		goto bad;
1784 
1785 	ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1786 	flags = ceph_decode_64(&p);
1787 	result = ceph_decode_32(&p);
1788 	reassert_epoch = ceph_decode_32(&p);
1789 	reassert_version = ceph_decode_64(&p);
1790 	osdmap_epoch = ceph_decode_32(&p);
1791 
1792 	/* lookup */
1793 	down_read(&osdc->map_sem);
1794 	mutex_lock(&osdc->request_mutex);
1795 	req = __lookup_request(osdc, tid);
1796 	if (req == NULL) {
1797 		dout("handle_reply tid %llu dne\n", tid);
1798 		goto bad_mutex;
1799 	}
1800 	ceph_osdc_get_request(req);
1801 
1802 	dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1803 	     req, result);
1804 
1805 	ceph_decode_need(&p, end, 4, bad_put);
1806 	numops = ceph_decode_32(&p);
1807 	if (numops > CEPH_OSD_MAX_OP)
1808 		goto bad_put;
1809 	if (numops != req->r_num_ops)
1810 		goto bad_put;
1811 	payload_len = 0;
1812 	ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1813 	for (i = 0; i < numops; i++) {
1814 		struct ceph_osd_op *op = p;
1815 		int len;
1816 
1817 		len = le32_to_cpu(op->payload_len);
1818 		req->r_reply_op_len[i] = len;
1819 		dout(" op %d has %d bytes\n", i, len);
1820 		payload_len += len;
1821 		p += sizeof(*op);
1822 	}
1823 	bytes = le32_to_cpu(msg->hdr.data_len);
1824 	if (payload_len != bytes) {
1825 		pr_warn("sum of op payload lens %d != data_len %d\n",
1826 			payload_len, bytes);
1827 		goto bad_put;
1828 	}
1829 
1830 	ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1831 	retry_attempt = ceph_decode_32(&p);
1832 	for (i = 0; i < numops; i++)
1833 		req->r_reply_op_result[i] = ceph_decode_32(&p);
1834 
1835 	if (le16_to_cpu(msg->hdr.version) >= 6) {
1836 		p += 8 + 4; /* skip replay_version */
1837 		p += 8; /* skip user_version */
1838 
1839 		err = ceph_redirect_decode(&p, end, &redir);
1840 		if (err)
1841 			goto bad_put;
1842 	} else {
1843 		redir.oloc.pool = -1;
1844 	}
1845 
1846 	if (redir.oloc.pool != -1) {
1847 		dout("redirect pool %lld\n", redir.oloc.pool);
1848 
1849 		__unregister_request(osdc, req);
1850 
1851 		req->r_target_oloc = redir.oloc; /* struct */
1852 
1853 		/*
1854 		 * Start redirect requests with nofail=true.  If
1855 		 * mapping fails, request will end up on the notarget
1856 		 * list, waiting for the new osdmap (which can take
1857 		 * a while), even though the original request mapped
1858 		 * successfully.  In the future we might want to follow
1859 		 * original request's nofail setting here.
1860 		 */
1861 		err = __ceph_osdc_start_request(osdc, req, true);
1862 		BUG_ON(err);
1863 
1864 		goto out_unlock;
1865 	}
1866 
1867 	already_completed = req->r_got_reply;
1868 	if (!req->r_got_reply) {
1869 		req->r_result = result;
1870 		dout("handle_reply result %d bytes %d\n", req->r_result,
1871 		     bytes);
1872 		if (req->r_result == 0)
1873 			req->r_result = bytes;
1874 
1875 		/* in case this is a write and we need to replay, */
1876 		req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1877 		req->r_reassert_version.version = cpu_to_le64(reassert_version);
1878 
1879 		req->r_got_reply = 1;
1880 	} else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1881 		dout("handle_reply tid %llu dup ack\n", tid);
1882 		goto out_unlock;
1883 	}
1884 
1885 	dout("handle_reply tid %llu flags %d\n", tid, flags);
1886 
1887 	if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1888 		__register_linger_request(osdc, req);
1889 
1890 	/* either this is a read, or we got the safe response */
1891 	if (result < 0 ||
1892 	    (flags & CEPH_OSD_FLAG_ONDISK) ||
1893 	    ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1894 		__unregister_request(osdc, req);
1895 
1896 	mutex_unlock(&osdc->request_mutex);
1897 	up_read(&osdc->map_sem);
1898 
1899 	if (!already_completed) {
1900 		if (req->r_unsafe_callback &&
1901 		    result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1902 			req->r_unsafe_callback(req, true);
1903 		if (req->r_callback)
1904 			req->r_callback(req, msg);
1905 		else
1906 			complete_all(&req->r_completion);
1907 	}
1908 
1909 	if (flags & CEPH_OSD_FLAG_ONDISK) {
1910 		if (req->r_unsafe_callback && already_completed)
1911 			req->r_unsafe_callback(req, false);
1912 		complete_request(req);
1913 	}
1914 
1915 out:
1916 	dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1917 	ceph_osdc_put_request(req);
1918 	return;
1919 out_unlock:
1920 	mutex_unlock(&osdc->request_mutex);
1921 	up_read(&osdc->map_sem);
1922 	goto out;
1923 
1924 bad_put:
1925 	req->r_result = -EIO;
1926 	__unregister_request(osdc, req);
1927 	if (req->r_callback)
1928 		req->r_callback(req, msg);
1929 	else
1930 		complete_all(&req->r_completion);
1931 	complete_request(req);
1932 	ceph_osdc_put_request(req);
1933 bad_mutex:
1934 	mutex_unlock(&osdc->request_mutex);
1935 	up_read(&osdc->map_sem);
1936 bad:
1937 	pr_err("corrupt osd_op_reply got %d %d\n",
1938 	       (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1939 	ceph_msg_dump(msg);
1940 }
1941 
1942 static void reset_changed_osds(struct ceph_osd_client *osdc)
1943 {
1944 	struct rb_node *p, *n;
1945 
1946 	dout("%s %p\n", __func__, osdc);
1947 	for (p = rb_first(&osdc->osds); p; p = n) {
1948 		struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1949 
1950 		n = rb_next(p);
1951 		if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1952 		    memcmp(&osd->o_con.peer_addr,
1953 			   ceph_osd_addr(osdc->osdmap,
1954 					 osd->o_osd),
1955 			   sizeof(struct ceph_entity_addr)) != 0)
1956 			__reset_osd(osdc, osd);
1957 	}
1958 }
1959 
1960 /*
1961  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1962  * no osd, request a new map.
1963  *
1964  * Caller should hold map_sem for read.
1965  */
1966 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1967 			  bool force_resend_writes)
1968 {
1969 	struct ceph_osd_request *req, *nreq;
1970 	struct rb_node *p;
1971 	int needmap = 0;
1972 	int err;
1973 	bool force_resend_req;
1974 
1975 	dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1976 		force_resend_writes ? " (force resend writes)" : "");
1977 	mutex_lock(&osdc->request_mutex);
1978 	for (p = rb_first(&osdc->requests); p; ) {
1979 		req = rb_entry(p, struct ceph_osd_request, r_node);
1980 		p = rb_next(p);
1981 
1982 		/*
1983 		 * For linger requests that have not yet been
1984 		 * registered, move them to the linger list; they'll
1985 		 * be sent to the osd in the loop below.  Unregister
1986 		 * the request before re-registering it as a linger
1987 		 * request to ensure the __map_request() below
1988 		 * will decide it needs to be sent.
1989 		 */
1990 		if (req->r_linger && list_empty(&req->r_linger_item)) {
1991 			dout("%p tid %llu restart on osd%d\n",
1992 			     req, req->r_tid,
1993 			     req->r_osd ? req->r_osd->o_osd : -1);
1994 			ceph_osdc_get_request(req);
1995 			__unregister_request(osdc, req);
1996 			__register_linger_request(osdc, req);
1997 			ceph_osdc_put_request(req);
1998 			continue;
1999 		}
2000 
2001 		force_resend_req = force_resend ||
2002 			(force_resend_writes &&
2003 				req->r_flags & CEPH_OSD_FLAG_WRITE);
2004 		err = __map_request(osdc, req, force_resend_req);
2005 		if (err < 0)
2006 			continue;  /* error */
2007 		if (req->r_osd == NULL) {
2008 			dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2009 			needmap++;  /* request a newer map */
2010 		} else if (err > 0) {
2011 			if (!req->r_linger) {
2012 				dout("%p tid %llu requeued on osd%d\n", req,
2013 				     req->r_tid,
2014 				     req->r_osd ? req->r_osd->o_osd : -1);
2015 				req->r_flags |= CEPH_OSD_FLAG_RETRY;
2016 			}
2017 		}
2018 	}
2019 
2020 	list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2021 				 r_linger_item) {
2022 		dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2023 
2024 		err = __map_request(osdc, req,
2025 				    force_resend || force_resend_writes);
2026 		dout("__map_request returned %d\n", err);
2027 		if (err < 0)
2028 			continue;  /* hrm! */
2029 		if (req->r_osd == NULL || err > 0) {
2030 			if (req->r_osd == NULL) {
2031 				dout("lingering %p tid %llu maps to no osd\n",
2032 				     req, req->r_tid);
2033 				/*
2034 				 * A homeless lingering request makes
2035 				 * no sense, as it's job is to keep
2036 				 * a particular OSD connection open.
2037 				 * Request a newer map and kick the
2038 				 * request, knowing that it won't be
2039 				 * resent until we actually get a map
2040 				 * that can tell us where to send it.
2041 				 */
2042 				needmap++;
2043 			}
2044 
2045 			dout("kicking lingering %p tid %llu osd%d\n", req,
2046 			     req->r_tid, req->r_osd ? req->r_osd->o_osd : -1);
2047 			__register_request(osdc, req);
2048 			__unregister_linger_request(osdc, req);
2049 		}
2050 	}
2051 	reset_changed_osds(osdc);
2052 	mutex_unlock(&osdc->request_mutex);
2053 
2054 	if (needmap) {
2055 		dout("%d requests for down osds, need new map\n", needmap);
2056 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2057 	}
2058 }
2059 
2060 
2061 /*
2062  * Process updated osd map.
2063  *
2064  * The message contains any number of incremental and full maps, normally
2065  * indicating some sort of topology change in the cluster.  Kick requests
2066  * off to different OSDs as needed.
2067  */
2068 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2069 {
2070 	void *p, *end, *next;
2071 	u32 nr_maps, maplen;
2072 	u32 epoch;
2073 	struct ceph_osdmap *newmap = NULL, *oldmap;
2074 	int err;
2075 	struct ceph_fsid fsid;
2076 	bool was_full;
2077 
2078 	dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2079 	p = msg->front.iov_base;
2080 	end = p + msg->front.iov_len;
2081 
2082 	/* verify fsid */
2083 	ceph_decode_need(&p, end, sizeof(fsid), bad);
2084 	ceph_decode_copy(&p, &fsid, sizeof(fsid));
2085 	if (ceph_check_fsid(osdc->client, &fsid) < 0)
2086 		return;
2087 
2088 	down_write(&osdc->map_sem);
2089 
2090 	was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2091 
2092 	/* incremental maps */
2093 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2094 	dout(" %d inc maps\n", nr_maps);
2095 	while (nr_maps > 0) {
2096 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2097 		epoch = ceph_decode_32(&p);
2098 		maplen = ceph_decode_32(&p);
2099 		ceph_decode_need(&p, end, maplen, bad);
2100 		next = p + maplen;
2101 		if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2102 			dout("applying incremental map %u len %d\n",
2103 			     epoch, maplen);
2104 			newmap = osdmap_apply_incremental(&p, next,
2105 							  osdc->osdmap,
2106 							  &osdc->client->msgr);
2107 			if (IS_ERR(newmap)) {
2108 				err = PTR_ERR(newmap);
2109 				goto bad;
2110 			}
2111 			BUG_ON(!newmap);
2112 			if (newmap != osdc->osdmap) {
2113 				ceph_osdmap_destroy(osdc->osdmap);
2114 				osdc->osdmap = newmap;
2115 			}
2116 			was_full = was_full ||
2117 				ceph_osdmap_flag(osdc->osdmap,
2118 						 CEPH_OSDMAP_FULL);
2119 			kick_requests(osdc, 0, was_full);
2120 		} else {
2121 			dout("ignoring incremental map %u len %d\n",
2122 			     epoch, maplen);
2123 		}
2124 		p = next;
2125 		nr_maps--;
2126 	}
2127 	if (newmap)
2128 		goto done;
2129 
2130 	/* full maps */
2131 	ceph_decode_32_safe(&p, end, nr_maps, bad);
2132 	dout(" %d full maps\n", nr_maps);
2133 	while (nr_maps) {
2134 		ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2135 		epoch = ceph_decode_32(&p);
2136 		maplen = ceph_decode_32(&p);
2137 		ceph_decode_need(&p, end, maplen, bad);
2138 		if (nr_maps > 1) {
2139 			dout("skipping non-latest full map %u len %d\n",
2140 			     epoch, maplen);
2141 		} else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2142 			dout("skipping full map %u len %d, "
2143 			     "older than our %u\n", epoch, maplen,
2144 			     osdc->osdmap->epoch);
2145 		} else {
2146 			int skipped_map = 0;
2147 
2148 			dout("taking full map %u len %d\n", epoch, maplen);
2149 			newmap = ceph_osdmap_decode(&p, p+maplen);
2150 			if (IS_ERR(newmap)) {
2151 				err = PTR_ERR(newmap);
2152 				goto bad;
2153 			}
2154 			BUG_ON(!newmap);
2155 			oldmap = osdc->osdmap;
2156 			osdc->osdmap = newmap;
2157 			if (oldmap) {
2158 				if (oldmap->epoch + 1 < newmap->epoch)
2159 					skipped_map = 1;
2160 				ceph_osdmap_destroy(oldmap);
2161 			}
2162 			was_full = was_full ||
2163 				ceph_osdmap_flag(osdc->osdmap,
2164 						 CEPH_OSDMAP_FULL);
2165 			kick_requests(osdc, skipped_map, was_full);
2166 		}
2167 		p += maplen;
2168 		nr_maps--;
2169 	}
2170 
2171 	if (!osdc->osdmap)
2172 		goto bad;
2173 done:
2174 	downgrade_write(&osdc->map_sem);
2175 	ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2176 
2177 	/*
2178 	 * subscribe to subsequent osdmap updates if full to ensure
2179 	 * we find out when we are no longer full and stop returning
2180 	 * ENOSPC.
2181 	 */
2182 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2183 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2184 		ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2185 		ceph_monc_request_next_osdmap(&osdc->client->monc);
2186 
2187 	mutex_lock(&osdc->request_mutex);
2188 	__send_queued(osdc);
2189 	mutex_unlock(&osdc->request_mutex);
2190 	up_read(&osdc->map_sem);
2191 	wake_up_all(&osdc->client->auth_wq);
2192 	return;
2193 
2194 bad:
2195 	pr_err("osdc handle_map corrupt msg\n");
2196 	ceph_msg_dump(msg);
2197 	up_write(&osdc->map_sem);
2198 }
2199 
2200 /*
2201  * watch/notify callback event infrastructure
2202  *
2203  * These callbacks are used both for watch and notify operations.
2204  */
2205 static void __release_event(struct kref *kref)
2206 {
2207 	struct ceph_osd_event *event =
2208 		container_of(kref, struct ceph_osd_event, kref);
2209 
2210 	dout("__release_event %p\n", event);
2211 	kfree(event);
2212 }
2213 
2214 static void get_event(struct ceph_osd_event *event)
2215 {
2216 	kref_get(&event->kref);
2217 }
2218 
2219 void ceph_osdc_put_event(struct ceph_osd_event *event)
2220 {
2221 	kref_put(&event->kref, __release_event);
2222 }
2223 EXPORT_SYMBOL(ceph_osdc_put_event);
2224 
2225 static void __insert_event(struct ceph_osd_client *osdc,
2226 			     struct ceph_osd_event *new)
2227 {
2228 	struct rb_node **p = &osdc->event_tree.rb_node;
2229 	struct rb_node *parent = NULL;
2230 	struct ceph_osd_event *event = NULL;
2231 
2232 	while (*p) {
2233 		parent = *p;
2234 		event = rb_entry(parent, struct ceph_osd_event, node);
2235 		if (new->cookie < event->cookie)
2236 			p = &(*p)->rb_left;
2237 		else if (new->cookie > event->cookie)
2238 			p = &(*p)->rb_right;
2239 		else
2240 			BUG();
2241 	}
2242 
2243 	rb_link_node(&new->node, parent, p);
2244 	rb_insert_color(&new->node, &osdc->event_tree);
2245 }
2246 
2247 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2248 					        u64 cookie)
2249 {
2250 	struct rb_node **p = &osdc->event_tree.rb_node;
2251 	struct rb_node *parent = NULL;
2252 	struct ceph_osd_event *event = NULL;
2253 
2254 	while (*p) {
2255 		parent = *p;
2256 		event = rb_entry(parent, struct ceph_osd_event, node);
2257 		if (cookie < event->cookie)
2258 			p = &(*p)->rb_left;
2259 		else if (cookie > event->cookie)
2260 			p = &(*p)->rb_right;
2261 		else
2262 			return event;
2263 	}
2264 	return NULL;
2265 }
2266 
2267 static void __remove_event(struct ceph_osd_event *event)
2268 {
2269 	struct ceph_osd_client *osdc = event->osdc;
2270 
2271 	if (!RB_EMPTY_NODE(&event->node)) {
2272 		dout("__remove_event removed %p\n", event);
2273 		rb_erase(&event->node, &osdc->event_tree);
2274 		ceph_osdc_put_event(event);
2275 	} else {
2276 		dout("__remove_event didn't remove %p\n", event);
2277 	}
2278 }
2279 
2280 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2281 			   void (*event_cb)(u64, u64, u8, void *),
2282 			   void *data, struct ceph_osd_event **pevent)
2283 {
2284 	struct ceph_osd_event *event;
2285 
2286 	event = kmalloc(sizeof(*event), GFP_NOIO);
2287 	if (!event)
2288 		return -ENOMEM;
2289 
2290 	dout("create_event %p\n", event);
2291 	event->cb = event_cb;
2292 	event->one_shot = 0;
2293 	event->data = data;
2294 	event->osdc = osdc;
2295 	INIT_LIST_HEAD(&event->osd_node);
2296 	RB_CLEAR_NODE(&event->node);
2297 	kref_init(&event->kref);   /* one ref for us */
2298 	kref_get(&event->kref);    /* one ref for the caller */
2299 
2300 	spin_lock(&osdc->event_lock);
2301 	event->cookie = ++osdc->event_count;
2302 	__insert_event(osdc, event);
2303 	spin_unlock(&osdc->event_lock);
2304 
2305 	*pevent = event;
2306 	return 0;
2307 }
2308 EXPORT_SYMBOL(ceph_osdc_create_event);
2309 
2310 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2311 {
2312 	struct ceph_osd_client *osdc = event->osdc;
2313 
2314 	dout("cancel_event %p\n", event);
2315 	spin_lock(&osdc->event_lock);
2316 	__remove_event(event);
2317 	spin_unlock(&osdc->event_lock);
2318 	ceph_osdc_put_event(event); /* caller's */
2319 }
2320 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2321 
2322 
2323 static void do_event_work(struct work_struct *work)
2324 {
2325 	struct ceph_osd_event_work *event_work =
2326 		container_of(work, struct ceph_osd_event_work, work);
2327 	struct ceph_osd_event *event = event_work->event;
2328 	u64 ver = event_work->ver;
2329 	u64 notify_id = event_work->notify_id;
2330 	u8 opcode = event_work->opcode;
2331 
2332 	dout("do_event_work completing %p\n", event);
2333 	event->cb(ver, notify_id, opcode, event->data);
2334 	dout("do_event_work completed %p\n", event);
2335 	ceph_osdc_put_event(event);
2336 	kfree(event_work);
2337 }
2338 
2339 
2340 /*
2341  * Process osd watch notifications
2342  */
2343 static void handle_watch_notify(struct ceph_osd_client *osdc,
2344 				struct ceph_msg *msg)
2345 {
2346 	void *p, *end;
2347 	u8 proto_ver;
2348 	u64 cookie, ver, notify_id;
2349 	u8 opcode;
2350 	struct ceph_osd_event *event;
2351 	struct ceph_osd_event_work *event_work;
2352 
2353 	p = msg->front.iov_base;
2354 	end = p + msg->front.iov_len;
2355 
2356 	ceph_decode_8_safe(&p, end, proto_ver, bad);
2357 	ceph_decode_8_safe(&p, end, opcode, bad);
2358 	ceph_decode_64_safe(&p, end, cookie, bad);
2359 	ceph_decode_64_safe(&p, end, ver, bad);
2360 	ceph_decode_64_safe(&p, end, notify_id, bad);
2361 
2362 	spin_lock(&osdc->event_lock);
2363 	event = __find_event(osdc, cookie);
2364 	if (event) {
2365 		BUG_ON(event->one_shot);
2366 		get_event(event);
2367 	}
2368 	spin_unlock(&osdc->event_lock);
2369 	dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2370 	     cookie, ver, event);
2371 	if (event) {
2372 		event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2373 		if (!event_work) {
2374 			pr_err("couldn't allocate event_work\n");
2375 			ceph_osdc_put_event(event);
2376 			return;
2377 		}
2378 		INIT_WORK(&event_work->work, do_event_work);
2379 		event_work->event = event;
2380 		event_work->ver = ver;
2381 		event_work->notify_id = notify_id;
2382 		event_work->opcode = opcode;
2383 
2384 		queue_work(osdc->notify_wq, &event_work->work);
2385 	}
2386 
2387 	return;
2388 
2389 bad:
2390 	pr_err("osdc handle_watch_notify corrupt msg\n");
2391 }
2392 
2393 /*
2394  * build new request AND message
2395  *
2396  */
2397 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2398 				struct ceph_snap_context *snapc, u64 snap_id,
2399 				struct timespec *mtime)
2400 {
2401 	struct ceph_msg *msg = req->r_request;
2402 	void *p;
2403 	size_t msg_size;
2404 	int flags = req->r_flags;
2405 	u64 data_len;
2406 	unsigned int i;
2407 
2408 	req->r_snapid = snap_id;
2409 	req->r_snapc = ceph_get_snap_context(snapc);
2410 
2411 	/* encode request */
2412 	msg->hdr.version = cpu_to_le16(4);
2413 
2414 	p = msg->front.iov_base;
2415 	ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2416 	req->r_request_osdmap_epoch = p;
2417 	p += 4;
2418 	req->r_request_flags = p;
2419 	p += 4;
2420 	if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2421 		ceph_encode_timespec(p, mtime);
2422 	p += sizeof(struct ceph_timespec);
2423 	req->r_request_reassert_version = p;
2424 	p += sizeof(struct ceph_eversion); /* will get filled in */
2425 
2426 	/* oloc */
2427 	ceph_encode_8(&p, 4);
2428 	ceph_encode_8(&p, 4);
2429 	ceph_encode_32(&p, 8 + 4 + 4);
2430 	req->r_request_pool = p;
2431 	p += 8;
2432 	ceph_encode_32(&p, -1);  /* preferred */
2433 	ceph_encode_32(&p, 0);   /* key len */
2434 
2435 	ceph_encode_8(&p, 1);
2436 	req->r_request_pgid = p;
2437 	p += 8 + 4;
2438 	ceph_encode_32(&p, -1);  /* preferred */
2439 
2440 	/* oid */
2441 	ceph_encode_32(&p, req->r_base_oid.name_len);
2442 	memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2443 	dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2444 	     req->r_base_oid.name, req->r_base_oid.name_len);
2445 	p += req->r_base_oid.name_len;
2446 
2447 	/* ops--can imply data */
2448 	ceph_encode_16(&p, (u16)req->r_num_ops);
2449 	data_len = 0;
2450 	for (i = 0; i < req->r_num_ops; i++) {
2451 		data_len += osd_req_encode_op(req, p, i);
2452 		p += sizeof(struct ceph_osd_op);
2453 	}
2454 
2455 	/* snaps */
2456 	ceph_encode_64(&p, req->r_snapid);
2457 	ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2458 	ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2459 	if (req->r_snapc) {
2460 		for (i = 0; i < snapc->num_snaps; i++) {
2461 			ceph_encode_64(&p, req->r_snapc->snaps[i]);
2462 		}
2463 	}
2464 
2465 	req->r_request_attempts = p;
2466 	p += 4;
2467 
2468 	/* data */
2469 	if (flags & CEPH_OSD_FLAG_WRITE) {
2470 		u16 data_off;
2471 
2472 		/*
2473 		 * The header "data_off" is a hint to the receiver
2474 		 * allowing it to align received data into its
2475 		 * buffers such that there's no need to re-copy
2476 		 * it before writing it to disk (direct I/O).
2477 		 */
2478 		data_off = (u16) (off & 0xffff);
2479 		req->r_request->hdr.data_off = cpu_to_le16(data_off);
2480 	}
2481 	req->r_request->hdr.data_len = cpu_to_le32(data_len);
2482 
2483 	BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2484 	msg_size = p - msg->front.iov_base;
2485 	msg->front.iov_len = msg_size;
2486 	msg->hdr.front_len = cpu_to_le32(msg_size);
2487 
2488 	dout("build_request msg_size was %d\n", (int)msg_size);
2489 }
2490 EXPORT_SYMBOL(ceph_osdc_build_request);
2491 
2492 /*
2493  * Register request, send initial attempt.
2494  */
2495 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2496 			    struct ceph_osd_request *req,
2497 			    bool nofail)
2498 {
2499 	int rc;
2500 
2501 	down_read(&osdc->map_sem);
2502 	mutex_lock(&osdc->request_mutex);
2503 
2504 	rc = __ceph_osdc_start_request(osdc, req, nofail);
2505 
2506 	mutex_unlock(&osdc->request_mutex);
2507 	up_read(&osdc->map_sem);
2508 
2509 	return rc;
2510 }
2511 EXPORT_SYMBOL(ceph_osdc_start_request);
2512 
2513 /*
2514  * Unregister a registered request.  The request is not completed (i.e.
2515  * no callbacks or wakeups) - higher layers are supposed to know what
2516  * they are canceling.
2517  */
2518 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2519 {
2520 	struct ceph_osd_client *osdc = req->r_osdc;
2521 
2522 	mutex_lock(&osdc->request_mutex);
2523 	if (req->r_linger)
2524 		__unregister_linger_request(osdc, req);
2525 	__unregister_request(osdc, req);
2526 	mutex_unlock(&osdc->request_mutex);
2527 
2528 	dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2529 }
2530 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2531 
2532 /*
2533  * wait for a request to complete
2534  */
2535 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2536 			   struct ceph_osd_request *req)
2537 {
2538 	int rc;
2539 
2540 	dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2541 
2542 	rc = wait_for_completion_interruptible(&req->r_completion);
2543 	if (rc < 0) {
2544 		dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2545 		ceph_osdc_cancel_request(req);
2546 		complete_request(req);
2547 		return rc;
2548 	}
2549 
2550 	dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2551 	     req->r_result);
2552 	return req->r_result;
2553 }
2554 EXPORT_SYMBOL(ceph_osdc_wait_request);
2555 
2556 /*
2557  * sync - wait for all in-flight requests to flush.  avoid starvation.
2558  */
2559 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2560 {
2561 	struct ceph_osd_request *req;
2562 	u64 last_tid, next_tid = 0;
2563 
2564 	mutex_lock(&osdc->request_mutex);
2565 	last_tid = osdc->last_tid;
2566 	while (1) {
2567 		req = __lookup_request_ge(osdc, next_tid);
2568 		if (!req)
2569 			break;
2570 		if (req->r_tid > last_tid)
2571 			break;
2572 
2573 		next_tid = req->r_tid + 1;
2574 		if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2575 			continue;
2576 
2577 		ceph_osdc_get_request(req);
2578 		mutex_unlock(&osdc->request_mutex);
2579 		dout("sync waiting on tid %llu (last is %llu)\n",
2580 		     req->r_tid, last_tid);
2581 		wait_for_completion(&req->r_safe_completion);
2582 		mutex_lock(&osdc->request_mutex);
2583 		ceph_osdc_put_request(req);
2584 	}
2585 	mutex_unlock(&osdc->request_mutex);
2586 	dout("sync done (thru tid %llu)\n", last_tid);
2587 }
2588 EXPORT_SYMBOL(ceph_osdc_sync);
2589 
2590 /*
2591  * Call all pending notify callbacks - for use after a watch is
2592  * unregistered, to make sure no more callbacks for it will be invoked
2593  */
2594 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2595 {
2596 	flush_workqueue(osdc->notify_wq);
2597 }
2598 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2599 
2600 
2601 /*
2602  * init, shutdown
2603  */
2604 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2605 {
2606 	int err;
2607 
2608 	dout("init\n");
2609 	osdc->client = client;
2610 	osdc->osdmap = NULL;
2611 	init_rwsem(&osdc->map_sem);
2612 	init_completion(&osdc->map_waiters);
2613 	osdc->last_requested_map = 0;
2614 	mutex_init(&osdc->request_mutex);
2615 	osdc->last_tid = 0;
2616 	osdc->osds = RB_ROOT;
2617 	INIT_LIST_HEAD(&osdc->osd_lru);
2618 	osdc->requests = RB_ROOT;
2619 	INIT_LIST_HEAD(&osdc->req_lru);
2620 	INIT_LIST_HEAD(&osdc->req_unsent);
2621 	INIT_LIST_HEAD(&osdc->req_notarget);
2622 	INIT_LIST_HEAD(&osdc->req_linger);
2623 	osdc->num_requests = 0;
2624 	INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2625 	INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2626 	spin_lock_init(&osdc->event_lock);
2627 	osdc->event_tree = RB_ROOT;
2628 	osdc->event_count = 0;
2629 
2630 	schedule_delayed_work(&osdc->osds_timeout_work,
2631 	   round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2632 
2633 	err = -ENOMEM;
2634 	osdc->req_mempool = mempool_create_kmalloc_pool(10,
2635 					sizeof(struct ceph_osd_request));
2636 	if (!osdc->req_mempool)
2637 		goto out;
2638 
2639 	err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2640 				OSD_OP_FRONT_LEN, 10, true,
2641 				"osd_op");
2642 	if (err < 0)
2643 		goto out_mempool;
2644 	err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2645 				OSD_OPREPLY_FRONT_LEN, 10, true,
2646 				"osd_op_reply");
2647 	if (err < 0)
2648 		goto out_msgpool;
2649 
2650 	err = -ENOMEM;
2651 	osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2652 	if (!osdc->notify_wq)
2653 		goto out_msgpool_reply;
2654 
2655 	return 0;
2656 
2657 out_msgpool_reply:
2658 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2659 out_msgpool:
2660 	ceph_msgpool_destroy(&osdc->msgpool_op);
2661 out_mempool:
2662 	mempool_destroy(osdc->req_mempool);
2663 out:
2664 	return err;
2665 }
2666 
2667 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2668 {
2669 	flush_workqueue(osdc->notify_wq);
2670 	destroy_workqueue(osdc->notify_wq);
2671 	cancel_delayed_work_sync(&osdc->timeout_work);
2672 	cancel_delayed_work_sync(&osdc->osds_timeout_work);
2673 	if (osdc->osdmap) {
2674 		ceph_osdmap_destroy(osdc->osdmap);
2675 		osdc->osdmap = NULL;
2676 	}
2677 	remove_all_osds(osdc);
2678 	mempool_destroy(osdc->req_mempool);
2679 	ceph_msgpool_destroy(&osdc->msgpool_op);
2680 	ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2681 }
2682 
2683 /*
2684  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2685  * *plen.  Return number of bytes read, or error.
2686  */
2687 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2688 			struct ceph_vino vino, struct ceph_file_layout *layout,
2689 			u64 off, u64 *plen,
2690 			u32 truncate_seq, u64 truncate_size,
2691 			struct page **pages, int num_pages, int page_align)
2692 {
2693 	struct ceph_osd_request *req;
2694 	int rc = 0;
2695 
2696 	dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2697 	     vino.snap, off, *plen);
2698 	req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2699 				    CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2700 				    NULL, truncate_seq, truncate_size,
2701 				    false);
2702 	if (IS_ERR(req))
2703 		return PTR_ERR(req);
2704 
2705 	/* it may be a short read due to an object boundary */
2706 
2707 	osd_req_op_extent_osd_data_pages(req, 0,
2708 				pages, *plen, page_align, false, false);
2709 
2710 	dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2711 	     off, *plen, *plen, page_align);
2712 
2713 	ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2714 
2715 	rc = ceph_osdc_start_request(osdc, req, false);
2716 	if (!rc)
2717 		rc = ceph_osdc_wait_request(osdc, req);
2718 
2719 	ceph_osdc_put_request(req);
2720 	dout("readpages result %d\n", rc);
2721 	return rc;
2722 }
2723 EXPORT_SYMBOL(ceph_osdc_readpages);
2724 
2725 /*
2726  * do a synchronous write on N pages
2727  */
2728 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2729 			 struct ceph_file_layout *layout,
2730 			 struct ceph_snap_context *snapc,
2731 			 u64 off, u64 len,
2732 			 u32 truncate_seq, u64 truncate_size,
2733 			 struct timespec *mtime,
2734 			 struct page **pages, int num_pages)
2735 {
2736 	struct ceph_osd_request *req;
2737 	int rc = 0;
2738 	int page_align = off & ~PAGE_MASK;
2739 
2740 	BUG_ON(vino.snap != CEPH_NOSNAP);	/* snapshots aren't writeable */
2741 	req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2742 				    CEPH_OSD_OP_WRITE,
2743 				    CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2744 				    snapc, truncate_seq, truncate_size,
2745 				    true);
2746 	if (IS_ERR(req))
2747 		return PTR_ERR(req);
2748 
2749 	/* it may be a short write due to an object boundary */
2750 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2751 				false, false);
2752 	dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2753 
2754 	ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2755 
2756 	rc = ceph_osdc_start_request(osdc, req, true);
2757 	if (!rc)
2758 		rc = ceph_osdc_wait_request(osdc, req);
2759 
2760 	ceph_osdc_put_request(req);
2761 	if (rc == 0)
2762 		rc = len;
2763 	dout("writepages result %d\n", rc);
2764 	return rc;
2765 }
2766 EXPORT_SYMBOL(ceph_osdc_writepages);
2767 
2768 int ceph_osdc_setup(void)
2769 {
2770 	BUG_ON(ceph_osd_request_cache);
2771 	ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2772 					sizeof (struct ceph_osd_request),
2773 					__alignof__(struct ceph_osd_request),
2774 					0, NULL);
2775 
2776 	return ceph_osd_request_cache ? 0 : -ENOMEM;
2777 }
2778 EXPORT_SYMBOL(ceph_osdc_setup);
2779 
2780 void ceph_osdc_cleanup(void)
2781 {
2782 	BUG_ON(!ceph_osd_request_cache);
2783 	kmem_cache_destroy(ceph_osd_request_cache);
2784 	ceph_osd_request_cache = NULL;
2785 }
2786 EXPORT_SYMBOL(ceph_osdc_cleanup);
2787 
2788 /*
2789  * handle incoming message
2790  */
2791 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2792 {
2793 	struct ceph_osd *osd = con->private;
2794 	struct ceph_osd_client *osdc;
2795 	int type = le16_to_cpu(msg->hdr.type);
2796 
2797 	if (!osd)
2798 		goto out;
2799 	osdc = osd->o_osdc;
2800 
2801 	switch (type) {
2802 	case CEPH_MSG_OSD_MAP:
2803 		ceph_osdc_handle_map(osdc, msg);
2804 		break;
2805 	case CEPH_MSG_OSD_OPREPLY:
2806 		handle_reply(osdc, msg, con);
2807 		break;
2808 	case CEPH_MSG_WATCH_NOTIFY:
2809 		handle_watch_notify(osdc, msg);
2810 		break;
2811 
2812 	default:
2813 		pr_err("received unknown message type %d %s\n", type,
2814 		       ceph_msg_type_name(type));
2815 	}
2816 out:
2817 	ceph_msg_put(msg);
2818 }
2819 
2820 /*
2821  * lookup and return message for incoming reply.  set up reply message
2822  * pages.
2823  */
2824 static struct ceph_msg *get_reply(struct ceph_connection *con,
2825 				  struct ceph_msg_header *hdr,
2826 				  int *skip)
2827 {
2828 	struct ceph_osd *osd = con->private;
2829 	struct ceph_osd_client *osdc = osd->o_osdc;
2830 	struct ceph_msg *m;
2831 	struct ceph_osd_request *req;
2832 	int front_len = le32_to_cpu(hdr->front_len);
2833 	int data_len = le32_to_cpu(hdr->data_len);
2834 	u64 tid;
2835 
2836 	tid = le64_to_cpu(hdr->tid);
2837 	mutex_lock(&osdc->request_mutex);
2838 	req = __lookup_request(osdc, tid);
2839 	if (!req) {
2840 		*skip = 1;
2841 		m = NULL;
2842 		dout("get_reply unknown tid %llu from osd%d\n", tid,
2843 		     osd->o_osd);
2844 		goto out;
2845 	}
2846 
2847 	if (req->r_reply->con)
2848 		dout("%s revoking msg %p from old con %p\n", __func__,
2849 		     req->r_reply, req->r_reply->con);
2850 	ceph_msg_revoke_incoming(req->r_reply);
2851 
2852 	if (front_len > req->r_reply->front_alloc_len) {
2853 		pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2854 			front_len, req->r_reply->front_alloc_len,
2855 			(unsigned int)con->peer_name.type,
2856 			le64_to_cpu(con->peer_name.num));
2857 		m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2858 				 false);
2859 		if (!m)
2860 			goto out;
2861 		ceph_msg_put(req->r_reply);
2862 		req->r_reply = m;
2863 	}
2864 	m = ceph_msg_get(req->r_reply);
2865 
2866 	if (data_len > 0) {
2867 		struct ceph_osd_data *osd_data;
2868 
2869 		/*
2870 		 * XXX This is assuming there is only one op containing
2871 		 * XXX page data.  Probably OK for reads, but this
2872 		 * XXX ought to be done more generally.
2873 		 */
2874 		osd_data = osd_req_op_extent_osd_data(req, 0);
2875 		if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2876 			if (osd_data->pages &&
2877 				unlikely(osd_data->length < data_len)) {
2878 
2879 				pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2880 					tid, data_len, osd_data->length);
2881 				*skip = 1;
2882 				ceph_msg_put(m);
2883 				m = NULL;
2884 				goto out;
2885 			}
2886 		}
2887 	}
2888 	*skip = 0;
2889 	dout("get_reply tid %lld %p\n", tid, m);
2890 
2891 out:
2892 	mutex_unlock(&osdc->request_mutex);
2893 	return m;
2894 
2895 }
2896 
2897 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2898 				  struct ceph_msg_header *hdr,
2899 				  int *skip)
2900 {
2901 	struct ceph_osd *osd = con->private;
2902 	int type = le16_to_cpu(hdr->type);
2903 	int front = le32_to_cpu(hdr->front_len);
2904 
2905 	*skip = 0;
2906 	switch (type) {
2907 	case CEPH_MSG_OSD_MAP:
2908 	case CEPH_MSG_WATCH_NOTIFY:
2909 		return ceph_msg_new(type, front, GFP_NOFS, false);
2910 	case CEPH_MSG_OSD_OPREPLY:
2911 		return get_reply(con, hdr, skip);
2912 	default:
2913 		pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2914 			osd->o_osd);
2915 		*skip = 1;
2916 		return NULL;
2917 	}
2918 }
2919 
2920 /*
2921  * Wrappers to refcount containing ceph_osd struct
2922  */
2923 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2924 {
2925 	struct ceph_osd *osd = con->private;
2926 	if (get_osd(osd))
2927 		return con;
2928 	return NULL;
2929 }
2930 
2931 static void put_osd_con(struct ceph_connection *con)
2932 {
2933 	struct ceph_osd *osd = con->private;
2934 	put_osd(osd);
2935 }
2936 
2937 /*
2938  * authentication
2939  */
2940 /*
2941  * Note: returned pointer is the address of a structure that's
2942  * managed separately.  Caller must *not* attempt to free it.
2943  */
2944 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2945 					int *proto, int force_new)
2946 {
2947 	struct ceph_osd *o = con->private;
2948 	struct ceph_osd_client *osdc = o->o_osdc;
2949 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2950 	struct ceph_auth_handshake *auth = &o->o_auth;
2951 
2952 	if (force_new && auth->authorizer) {
2953 		ceph_auth_destroy_authorizer(ac, auth->authorizer);
2954 		auth->authorizer = NULL;
2955 	}
2956 	if (!auth->authorizer) {
2957 		int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2958 						      auth);
2959 		if (ret)
2960 			return ERR_PTR(ret);
2961 	} else {
2962 		int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2963 						     auth);
2964 		if (ret)
2965 			return ERR_PTR(ret);
2966 	}
2967 	*proto = ac->protocol;
2968 
2969 	return auth;
2970 }
2971 
2972 
2973 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2974 {
2975 	struct ceph_osd *o = con->private;
2976 	struct ceph_osd_client *osdc = o->o_osdc;
2977 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2978 
2979 	return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2980 }
2981 
2982 static int invalidate_authorizer(struct ceph_connection *con)
2983 {
2984 	struct ceph_osd *o = con->private;
2985 	struct ceph_osd_client *osdc = o->o_osdc;
2986 	struct ceph_auth_client *ac = osdc->client->monc.auth;
2987 
2988 	ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2989 	return ceph_monc_validate_auth(&osdc->client->monc);
2990 }
2991 
2992 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
2993 {
2994 	struct ceph_osd *o = con->private;
2995 	struct ceph_auth_handshake *auth = &o->o_auth;
2996 	return ceph_auth_sign_message(auth, msg);
2997 }
2998 
2999 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
3000 {
3001 	struct ceph_osd *o = con->private;
3002 	struct ceph_auth_handshake *auth = &o->o_auth;
3003 	return ceph_auth_check_message_signature(auth, msg);
3004 }
3005 
3006 static const struct ceph_connection_operations osd_con_ops = {
3007 	.get = get_osd_con,
3008 	.put = put_osd_con,
3009 	.dispatch = dispatch,
3010 	.get_authorizer = get_authorizer,
3011 	.verify_authorizer_reply = verify_authorizer_reply,
3012 	.invalidate_authorizer = invalidate_authorizer,
3013 	.alloc_msg = alloc_msg,
3014 	.sign_message = sign_message,
3015 	.check_message_signature = check_message_signature,
3016 	.fault = osd_reset,
3017 };
3018