net/ceph/osd_client.c

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