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