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_request(struct ceph_osd_client *osdc,
  34                                  struct ceph_osd_request *req);
  35 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  36                                         struct ceph_osd_request *req);
  37 static void __enqueue_request(struct ceph_osd_request *req);
  38 static void __send_request(struct ceph_osd_client *osdc,
  39                            struct ceph_osd_request *req);
  40
  41 /*
  42  * Implement client access to distributed object storage cluster.
  43  *
  44  * All data objects are stored within a cluster/cloud of OSDs, or
  45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  47  * remote daemons serving up and coordinating consistent and safe
  48  * access to storage.
  49  *
  50  * Cluster membership and the mapping of data objects onto storage devices
  51  * are described by the osd map.
  52  *
  53  * We keep track of pending OSD requests (read, write), resubmit
  54  * requests to different OSDs when the cluster topology/data layout
  55  * change, or retry the affected requests when the communications
  56  * channel with an OSD is reset.
  57  */
  58
  59 /*
  60  * calculate the mapping of a file extent onto an object, and fill out the
  61  * request accordingly.  shorten extent as necessary if it crosses an
  62  * object boundary.
  63  *
  64  * fill osd op in request message.
  65  */
  66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
  67                         u64 *objnum, u64 *objoff, u64 *objlen)
  68 {
  69         u64 orig_len = *plen;
  70         int r;
  71
  72         /* object extent? */
  73         r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
  74                                           objoff, objlen);
  75         if (r < 0)
  76                 return r;
  77         if (*objlen < orig_len) {
  78                 *plen = *objlen;
  79                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  80                      orig_len - *plen, off, *plen);
  81         }
  82
  83         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
  84
  85         return 0;
  86 }
  87
  88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
  89 {
  90         memset(osd_data, 0, sizeof (*osd_data));
  91         osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
  92 }
  93
  94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
  95                         struct page **pages, u64 length, u32 alignment,
  96                         bool pages_from_pool, bool own_pages)
  97 {
  98         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
  99         osd_data->pages = pages;
 100         osd_data->length = length;
 101         osd_data->alignment = alignment;
 102         osd_data->pages_from_pool = pages_from_pool;
 103         osd_data->own_pages = own_pages;
 104 }
 105
 106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
 107                         struct ceph_pagelist *pagelist)
 108 {
 109         osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
 110         osd_data->pagelist = pagelist;
 111 }
 112
 113 #ifdef CONFIG_BLOCK
 114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
 115                         struct bio *bio, size_t bio_length)
 116 {
 117         osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
 118         osd_data->bio = bio;
 119         osd_data->bio_length = bio_length;
 120 }
 121 #endif /* CONFIG_BLOCK */
 122
 123 #define osd_req_op_data(oreq, whch, typ, fld)   \
 124         ({                                              \
 125                 BUG_ON(whch >= (oreq)->r_num_ops);      \
 126                 &(oreq)->r_ops[whch].typ.fld;           \
 127         })
 128
 129 static struct ceph_osd_data *
 130 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
 131 {
 132         BUG_ON(which >= osd_req->r_num_ops);
 133
 134         return &osd_req->r_ops[which].raw_data_in;
 135 }
 136
 137 struct ceph_osd_data *
 138 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
 139                         unsigned int which)
 140 {
 141         return osd_req_op_data(osd_req, which, extent, osd_data);
 142 }
 143 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
 144
 145 struct ceph_osd_data *
 146 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
 147                         unsigned int which)
 148 {
 149         return osd_req_op_data(osd_req, which, cls, response_data);
 150 }
 151 EXPORT_SYMBOL(osd_req_op_cls_response_data);    /* ??? */
 152
 153 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
 154                         unsigned int which, struct page **pages,
 155                         u64 length, u32 alignment,
 156                         bool pages_from_pool, bool own_pages)
 157 {
 158         struct ceph_osd_data *osd_data;
 159
 160         osd_data = osd_req_op_raw_data_in(osd_req, which);
 161         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 162                                 pages_from_pool, own_pages);
 163 }
 164 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
 165
 166 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
 167                         unsigned int which, struct page **pages,
 168                         u64 length, u32 alignment,
 169                         bool pages_from_pool, bool own_pages)
 170 {
 171         struct ceph_osd_data *osd_data;
 172
 173         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 174         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 175                                 pages_from_pool, own_pages);
 176 }
 177 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
 178
 179 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
 180                         unsigned int which, struct ceph_pagelist *pagelist)
 181 {
 182         struct ceph_osd_data *osd_data;
 183
 184         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 185         ceph_osd_data_pagelist_init(osd_data, pagelist);
 186 }
 187 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
 188
 189 #ifdef CONFIG_BLOCK
 190 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
 191                         unsigned int which, struct bio *bio, size_t bio_length)
 192 {
 193         struct ceph_osd_data *osd_data;
 194
 195         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 196         ceph_osd_data_bio_init(osd_data, bio, bio_length);
 197 }
 198 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
 199 #endif /* CONFIG_BLOCK */
 200
 201 static void osd_req_op_cls_request_info_pagelist(
 202                         struct ceph_osd_request *osd_req,
 203                         unsigned int which, struct ceph_pagelist *pagelist)
 204 {
 205         struct ceph_osd_data *osd_data;
 206
 207         osd_data = osd_req_op_data(osd_req, which, cls, request_info);
 208         ceph_osd_data_pagelist_init(osd_data, pagelist);
 209 }
 210
 211 void osd_req_op_cls_request_data_pagelist(
 212                         struct ceph_osd_request *osd_req,
 213                         unsigned int which, struct ceph_pagelist *pagelist)
 214 {
 215         struct ceph_osd_data *osd_data;
 216
 217         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 218         ceph_osd_data_pagelist_init(osd_data, pagelist);
 219 }
 220 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
 221
 222 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
 223                         unsigned int which, struct page **pages, u64 length,
 224                         u32 alignment, bool pages_from_pool, bool own_pages)
 225 {
 226         struct ceph_osd_data *osd_data;
 227
 228         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 229         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 230                                 pages_from_pool, own_pages);
 231 }
 232 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
 233
 234 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
 235                         unsigned int which, struct page **pages, u64 length,
 236                         u32 alignment, bool pages_from_pool, bool own_pages)
 237 {
 238         struct ceph_osd_data *osd_data;
 239
 240         osd_data = osd_req_op_data(osd_req, which, cls, response_data);
 241         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 242                                 pages_from_pool, own_pages);
 243 }
 244 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
 245
 246 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
 247 {
 248         switch (osd_data->type) {
 249         case CEPH_OSD_DATA_TYPE_NONE:
 250                 return 0;
 251         case CEPH_OSD_DATA_TYPE_PAGES:
 252                 return osd_data->length;
 253         case CEPH_OSD_DATA_TYPE_PAGELIST:
 254                 return (u64)osd_data->pagelist->length;
 255 #ifdef CONFIG_BLOCK
 256         case CEPH_OSD_DATA_TYPE_BIO:
 257                 return (u64)osd_data->bio_length;
 258 #endif /* CONFIG_BLOCK */
 259         default:
 260                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
 261                 return 0;
 262         }
 263 }
 264
 265 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
 266 {
 267         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
 268                 int num_pages;
 269
 270                 num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                                 (u64)osd_data->length);
 272                 ceph_release_page_vector(osd_data->pages, num_pages);
 273         }
 274         ceph_osd_data_init(osd_data);
 275 }
 276
 277 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
 278                         unsigned int which)
 279 {
 280         struct ceph_osd_req_op *op;
 281
 282         BUG_ON(which >= osd_req->r_num_ops);
 283         op = &osd_req->r_ops[which];
 284
 285         switch (op->op) {
 286         case CEPH_OSD_OP_READ:
 287         case CEPH_OSD_OP_WRITE:
 288                 ceph_osd_data_release(&op->extent.osd_data);
 289                 break;
 290         case CEPH_OSD_OP_CALL:
 291                 ceph_osd_data_release(&op->cls.request_info);
 292                 ceph_osd_data_release(&op->cls.request_data);
 293                 ceph_osd_data_release(&op->cls.response_data);
 294                 break;
 295         default:
 296                 break;
 297         }
 298 }
 299
 300 /*
 301  * requests
 302  */
 303 static void ceph_osdc_release_request(struct kref *kref)
 304 {
 305         struct ceph_osd_request *req = container_of(kref,
 306                                             struct ceph_osd_request, r_kref);
 307         unsigned int which;
 308
 309         dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
 310              req->r_request, req->r_reply);
 311         WARN_ON(!RB_EMPTY_NODE(&req->r_node));
 312         WARN_ON(!list_empty(&req->r_req_lru_item));
 313         WARN_ON(!list_empty(&req->r_osd_item));
 314         WARN_ON(!list_empty(&req->r_linger_item));
 315         WARN_ON(!list_empty(&req->r_linger_osd_item));
 316         WARN_ON(req->r_osd);
 317
 318         if (req->r_request)
 319                 ceph_msg_put(req->r_request);
 320         if (req->r_reply) {
 321                 ceph_msg_revoke_incoming(req->r_reply);
 322                 ceph_msg_put(req->r_reply);
 323         }
 324
 325         for (which = 0; which < req->r_num_ops; which++)
 326                 osd_req_op_data_release(req, which);
 327
 328         ceph_put_snap_context(req->r_snapc);
 329         if (req->r_mempool)
 330                 mempool_free(req, req->r_osdc->req_mempool);
 331         else
 332                 kmem_cache_free(ceph_osd_request_cache, req);
 333
 334 }
 335
 336 void ceph_osdc_get_request(struct ceph_osd_request *req)
 337 {
 338         dout("%s %p (was %d)\n", __func__, req,
 339              atomic_read(&req->r_kref.refcount));
 340         kref_get(&req->r_kref);
 341 }
 342 EXPORT_SYMBOL(ceph_osdc_get_request);
 343
 344 void ceph_osdc_put_request(struct ceph_osd_request *req)
 345 {
 346         dout("%s %p (was %d)\n", __func__, req,
 347              atomic_read(&req->r_kref.refcount));
 348         kref_put(&req->r_kref, ceph_osdc_release_request);
 349 }
 350 EXPORT_SYMBOL(ceph_osdc_put_request);
 351
 352 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 353                                                struct ceph_snap_context *snapc,
 354                                                unsigned int num_ops,
 355                                                bool use_mempool,
 356                                                gfp_t gfp_flags)
 357 {
 358         struct ceph_osd_request *req;
 359         struct ceph_msg *msg;
 360         size_t msg_size;
 361
 362         BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
 363         BUG_ON(num_ops > CEPH_OSD_MAX_OP);
 364
 365         msg_size = 4 + 4 + 8 + 8 + 4+8;
 366         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
 367         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
 368         msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
 369         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
 370         msg_size += 8;  /* snapid */
 371         msg_size += 8;  /* snap_seq */
 372         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
 373         msg_size += 4;
 374
 375         if (use_mempool) {
 376                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 377                 memset(req, 0, sizeof(*req));
 378         } else {
 379                 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
 380         }
 381         if (req == NULL)
 382                 return NULL;
 383
 384         req->r_osdc = osdc;
 385         req->r_mempool = use_mempool;
 386         req->r_num_ops = num_ops;
 387
 388         kref_init(&req->r_kref);
 389         init_completion(&req->r_completion);
 390         init_completion(&req->r_safe_completion);
 391         RB_CLEAR_NODE(&req->r_node);
 392         INIT_LIST_HEAD(&req->r_unsafe_item);
 393         INIT_LIST_HEAD(&req->r_linger_item);
 394         INIT_LIST_HEAD(&req->r_linger_osd_item);
 395         INIT_LIST_HEAD(&req->r_req_lru_item);
 396         INIT_LIST_HEAD(&req->r_osd_item);
 397
 398         req->r_base_oloc.pool = -1;
 399         req->r_target_oloc.pool = -1;
 400
 401         /* create reply message */
 402         if (use_mempool)
 403                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 404         else
 405                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 406                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 407         if (!msg) {
 408                 ceph_osdc_put_request(req);
 409                 return NULL;
 410         }
 411         req->r_reply = msg;
 412
 413         /* create request message; allow space for oid */
 414         if (use_mempool)
 415                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 416         else
 417                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 418         if (!msg) {
 419                 ceph_osdc_put_request(req);
 420                 return NULL;
 421         }
 422
 423         memset(msg->front.iov_base, 0, msg->front.iov_len);
 424
 425         req->r_request = msg;
 426
 427         return req;
 428 }
 429 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 430
 431 static bool osd_req_opcode_valid(u16 opcode)
 432 {
 433         switch (opcode) {
 434         case CEPH_OSD_OP_READ:
 435         case CEPH_OSD_OP_STAT:
 436         case CEPH_OSD_OP_MAPEXT:
 437         case CEPH_OSD_OP_MASKTRUNC:
 438         case CEPH_OSD_OP_SPARSE_READ:
 439         case CEPH_OSD_OP_NOTIFY:
 440         case CEPH_OSD_OP_NOTIFY_ACK:
 441         case CEPH_OSD_OP_ASSERT_VER:
 442         case CEPH_OSD_OP_WRITE:
 443         case CEPH_OSD_OP_WRITEFULL:
 444         case CEPH_OSD_OP_TRUNCATE:
 445         case CEPH_OSD_OP_ZERO:
 446         case CEPH_OSD_OP_DELETE:
 447         case CEPH_OSD_OP_APPEND:
 448         case CEPH_OSD_OP_STARTSYNC:
 449         case CEPH_OSD_OP_SETTRUNC:
 450         case CEPH_OSD_OP_TRIMTRUNC:
 451         case CEPH_OSD_OP_TMAPUP:
 452         case CEPH_OSD_OP_TMAPPUT:
 453         case CEPH_OSD_OP_TMAPGET:
 454         case CEPH_OSD_OP_CREATE:
 455         case CEPH_OSD_OP_ROLLBACK:
 456         case CEPH_OSD_OP_WATCH:
 457         case CEPH_OSD_OP_OMAPGETKEYS:
 458         case CEPH_OSD_OP_OMAPGETVALS:
 459         case CEPH_OSD_OP_OMAPGETHEADER:
 460         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
 461         case CEPH_OSD_OP_OMAPSETVALS:
 462         case CEPH_OSD_OP_OMAPSETHEADER:
 463         case CEPH_OSD_OP_OMAPCLEAR:
 464         case CEPH_OSD_OP_OMAPRMKEYS:
 465         case CEPH_OSD_OP_OMAP_CMP:
 466         case CEPH_OSD_OP_SETALLOCHINT:
 467         case CEPH_OSD_OP_CLONERANGE:
 468         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
 469         case CEPH_OSD_OP_SRC_CMPXATTR:
 470         case CEPH_OSD_OP_GETXATTR:
 471         case CEPH_OSD_OP_GETXATTRS:
 472         case CEPH_OSD_OP_CMPXATTR:
 473         case CEPH_OSD_OP_SETXATTR:
 474         case CEPH_OSD_OP_SETXATTRS:
 475         case CEPH_OSD_OP_RESETXATTRS:
 476         case CEPH_OSD_OP_RMXATTR:
 477         case CEPH_OSD_OP_PULL:
 478         case CEPH_OSD_OP_PUSH:
 479         case CEPH_OSD_OP_BALANCEREADS:
 480         case CEPH_OSD_OP_UNBALANCEREADS:
 481         case CEPH_OSD_OP_SCRUB:
 482         case CEPH_OSD_OP_SCRUB_RESERVE:
 483         case CEPH_OSD_OP_SCRUB_UNRESERVE:
 484         case CEPH_OSD_OP_SCRUB_STOP:
 485         case CEPH_OSD_OP_SCRUB_MAP:
 486         case CEPH_OSD_OP_WRLOCK:
 487         case CEPH_OSD_OP_WRUNLOCK:
 488         case CEPH_OSD_OP_RDLOCK:
 489         case CEPH_OSD_OP_RDUNLOCK:
 490         case CEPH_OSD_OP_UPLOCK:
 491         case CEPH_OSD_OP_DNLOCK:
 492         case CEPH_OSD_OP_CALL:
 493         case CEPH_OSD_OP_PGLS:
 494         case CEPH_OSD_OP_PGLS_FILTER:
 495                 return true;
 496         default:
 497                 return false;
 498         }
 499 }
 500
 501 /*
 502  * This is an osd op init function for opcodes that have no data or
 503  * other information associated with them.  It also serves as a
 504  * common init routine for all the other init functions, below.
 505  */
 506 static struct ceph_osd_req_op *
 507 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
 508                                 u16 opcode)
 509 {
 510         struct ceph_osd_req_op *op;
 511
 512         BUG_ON(which >= osd_req->r_num_ops);
 513         BUG_ON(!osd_req_opcode_valid(opcode));
 514
 515         op = &osd_req->r_ops[which];
 516         memset(op, 0, sizeof (*op));
 517         op->op = opcode;
 518
 519         return op;
 520 }
 521
 522 void osd_req_op_init(struct ceph_osd_request *osd_req,
 523                                 unsigned int which, u16 opcode)
 524 {
 525         (void)_osd_req_op_init(osd_req, which, opcode);
 526 }
 527 EXPORT_SYMBOL(osd_req_op_init);
 528
 529 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 530                                 unsigned int which, u16 opcode,
 531                                 u64 offset, u64 length,
 532                                 u64 truncate_size, u32 truncate_seq)
 533 {
 534         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 535         size_t payload_len = 0;
 536
 537         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 538                opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
 539                opcode != CEPH_OSD_OP_TRUNCATE);
 540
 541         op->extent.offset = offset;
 542         op->extent.length = length;
 543         op->extent.truncate_size = truncate_size;
 544         op->extent.truncate_seq = truncate_seq;
 545         if (opcode == CEPH_OSD_OP_WRITE)
 546                 payload_len += length;
 547
 548         op->payload_len = payload_len;
 549 }
 550 EXPORT_SYMBOL(osd_req_op_extent_init);
 551
 552 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 553                                 unsigned int which, u64 length)
 554 {
 555         struct ceph_osd_req_op *op;
 556         u64 previous;
 557
 558         BUG_ON(which >= osd_req->r_num_ops);
 559         op = &osd_req->r_ops[which];
 560         previous = op->extent.length;
 561
 562         if (length == previous)
 563                 return;         /* Nothing to do */
 564         BUG_ON(length > previous);
 565
 566         op->extent.length = length;
 567         op->payload_len -= previous - length;
 568 }
 569 EXPORT_SYMBOL(osd_req_op_extent_update);
 570
 571 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
 572                         u16 opcode, const char *class, const char *method)
 573 {
 574         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 575         struct ceph_pagelist *pagelist;
 576         size_t payload_len = 0;
 577         size_t size;
 578
 579         BUG_ON(opcode != CEPH_OSD_OP_CALL);
 580
 581         pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
 582         BUG_ON(!pagelist);
 583         ceph_pagelist_init(pagelist);
 584
 585         op->cls.class_name = class;
 586         size = strlen(class);
 587         BUG_ON(size > (size_t) U8_MAX);
 588         op->cls.class_len = size;
 589         ceph_pagelist_append(pagelist, class, size);
 590         payload_len += size;
 591
 592         op->cls.method_name = method;
 593         size = strlen(method);
 594         BUG_ON(size > (size_t) U8_MAX);
 595         op->cls.method_len = size;
 596         ceph_pagelist_append(pagelist, method, size);
 597         payload_len += size;
 598
 599         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
 600
 601         op->cls.argc = 0;       /* currently unused */
 602
 603         op->payload_len = payload_len;
 604 }
 605 EXPORT_SYMBOL(osd_req_op_cls_init);
 606
 607 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
 608                                 unsigned int which, u16 opcode,
 609                                 u64 cookie, u64 version, int flag)
 610 {
 611         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 612
 613         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
 614
 615         op->watch.cookie = cookie;
 616         op->watch.ver = version;
 617         if (opcode == CEPH_OSD_OP_WATCH && flag)
 618                 op->watch.flag = (u8)1;
 619 }
 620 EXPORT_SYMBOL(osd_req_op_watch_init);
 621
 622 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 623                                 unsigned int which,
 624                                 u64 expected_object_size,
 625                                 u64 expected_write_size)
 626 {
 627         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 628                                                       CEPH_OSD_OP_SETALLOCHINT);
 629
 630         op->alloc_hint.expected_object_size = expected_object_size;
 631         op->alloc_hint.expected_write_size = expected_write_size;
 632
 633         /*
 634          * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
 635          * not worth a feature bit.  Set FAILOK per-op flag to make
 636          * sure older osds don't trip over an unsupported opcode.
 637          */
 638         op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
 639 }
 640 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
 641
 642 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
 643                                 struct ceph_osd_data *osd_data)
 644 {
 645         u64 length = ceph_osd_data_length(osd_data);
 646
 647         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 648                 BUG_ON(length > (u64) SIZE_MAX);
 649                 if (length)
 650                         ceph_msg_data_add_pages(msg, osd_data->pages,
 651                                         length, osd_data->alignment);
 652         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 653                 BUG_ON(!length);
 654                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
 655 #ifdef CONFIG_BLOCK
 656         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
 657                 ceph_msg_data_add_bio(msg, osd_data->bio, length);
 658 #endif
 659         } else {
 660                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
 661         }
 662 }
 663
 664 static u64 osd_req_encode_op(struct ceph_osd_request *req,
 665                               struct ceph_osd_op *dst, unsigned int which)
 666 {
 667         struct ceph_osd_req_op *src;
 668         struct ceph_osd_data *osd_data;
 669         u64 request_data_len = 0;
 670         u64 data_length;
 671
 672         BUG_ON(which >= req->r_num_ops);
 673         src = &req->r_ops[which];
 674         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
 675                 pr_err("unrecognized osd opcode %d\n", src->op);
 676
 677                 return 0;
 678         }
 679
 680         switch (src->op) {
 681         case CEPH_OSD_OP_STAT:
 682                 osd_data = &src->raw_data_in;
 683                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 684                 break;
 685         case CEPH_OSD_OP_READ:
 686         case CEPH_OSD_OP_WRITE:
 687         case CEPH_OSD_OP_ZERO:
 688         case CEPH_OSD_OP_DELETE:
 689         case CEPH_OSD_OP_TRUNCATE:
 690                 if (src->op == CEPH_OSD_OP_WRITE)
 691                         request_data_len = src->extent.length;
 692                 dst->extent.offset = cpu_to_le64(src->extent.offset);
 693                 dst->extent.length = cpu_to_le64(src->extent.length);
 694                 dst->extent.truncate_size =
 695                         cpu_to_le64(src->extent.truncate_size);
 696                 dst->extent.truncate_seq =
 697                         cpu_to_le32(src->extent.truncate_seq);
 698                 osd_data = &src->extent.osd_data;
 699                 if (src->op == CEPH_OSD_OP_WRITE)
 700                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 701                 else
 702                         ceph_osdc_msg_data_add(req->r_reply, osd_data);
 703                 break;
 704         case CEPH_OSD_OP_CALL:
 705                 dst->cls.class_len = src->cls.class_len;
 706                 dst->cls.method_len = src->cls.method_len;
 707                 osd_data = &src->cls.request_info;
 708                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 709                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
 710                 request_data_len = osd_data->pagelist->length;
 711
 712                 osd_data = &src->cls.request_data;
 713                 data_length = ceph_osd_data_length(osd_data);
 714                 if (data_length) {
 715                         BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
 716                         dst->cls.indata_len = cpu_to_le32(data_length);
 717                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 718                         src->payload_len += data_length;
 719                         request_data_len += data_length;
 720                 }
 721                 osd_data = &src->cls.response_data;
 722                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 723                 break;
 724         case CEPH_OSD_OP_STARTSYNC:
 725                 break;
 726         case CEPH_OSD_OP_NOTIFY_ACK:
 727         case CEPH_OSD_OP_WATCH:
 728                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 729                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 730                 dst->watch.flag = src->watch.flag;
 731                 break;
 732         case CEPH_OSD_OP_SETALLOCHINT:
 733                 dst->alloc_hint.expected_object_size =
 734                     cpu_to_le64(src->alloc_hint.expected_object_size);
 735                 dst->alloc_hint.expected_write_size =
 736                     cpu_to_le64(src->alloc_hint.expected_write_size);
 737                 break;
 738         default:
 739                 pr_err("unsupported osd opcode %s\n",
 740                         ceph_osd_op_name(src->op));
 741                 WARN_ON(1);
 742
 743                 return 0;
 744         }
 745
 746         dst->op = cpu_to_le16(src->op);
 747         dst->flags = cpu_to_le32(src->flags);
 748         dst->payload_len = cpu_to_le32(src->payload_len);
 749
 750         return request_data_len;
 751 }
 752
 753 /*
 754  * build new request AND message, calculate layout, and adjust file
 755  * extent as needed.
 756  *
 757  * if the file was recently truncated, we include information about its
 758  * old and new size so that the object can be updated appropriately.  (we
 759  * avoid synchronously deleting truncated objects because it's slow.)
 760  *
 761  * if @do_sync, include a 'startsync' command so that the osd will flush
 762  * data quickly.
 763  */
 764 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 765                                                struct ceph_file_layout *layout,
 766                                                struct ceph_vino vino,
 767                                                u64 off, u64 *plen, int num_ops,
 768                                                int opcode, int flags,
 769                                                struct ceph_snap_context *snapc,
 770                                                u32 truncate_seq,
 771                                                u64 truncate_size,
 772                                                bool use_mempool)
 773 {
 774         struct ceph_osd_request *req;
 775         u64 objnum = 0;
 776         u64 objoff = 0;
 777         u64 objlen = 0;
 778         u32 object_size;
 779         u64 object_base;
 780         int r;
 781
 782         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 783                opcode != CEPH_OSD_OP_DELETE && opcode != CEPH_OSD_OP_ZERO &&
 784                opcode != CEPH_OSD_OP_TRUNCATE);
 785
 786         req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
 787                                         GFP_NOFS);
 788         if (!req)
 789                 return ERR_PTR(-ENOMEM);
 790
 791         req->r_flags = flags;
 792
 793         /* calculate max write size */
 794         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
 795         if (r < 0) {
 796                 ceph_osdc_put_request(req);
 797                 return ERR_PTR(r);
 798         }
 799
 800         object_size = le32_to_cpu(layout->fl_object_size);
 801         object_base = off - objoff;
 802         if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
 803                 if (truncate_size <= object_base) {
 804                         truncate_size = 0;
 805                 } else {
 806                         truncate_size -= object_base;
 807                         if (truncate_size > object_size)
 808                                 truncate_size = object_size;
 809                 }
 810         }
 811
 812         osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
 813                                 truncate_size, truncate_seq);
 814
 815         /*
 816          * A second op in the ops array means the caller wants to
 817          * also issue a include a 'startsync' command so that the
 818          * osd will flush data quickly.
 819          */
 820         if (num_ops > 1)
 821                 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
 822
 823         req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
 824
 825         snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
 826                  "%llx.%08llx", vino.ino, objnum);
 827         req->r_base_oid.name_len = strlen(req->r_base_oid.name);
 828
 829         return req;
 830 }
 831 EXPORT_SYMBOL(ceph_osdc_new_request);
 832
 833 /*
 834  * We keep osd requests in an rbtree, sorted by ->r_tid.
 835  */
 836 static void __insert_request(struct ceph_osd_client *osdc,
 837                              struct ceph_osd_request *new)
 838 {
 839         struct rb_node **p = &osdc->requests.rb_node;
 840         struct rb_node *parent = NULL;
 841         struct ceph_osd_request *req = NULL;
 842
 843         while (*p) {
 844                 parent = *p;
 845                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 846                 if (new->r_tid < req->r_tid)
 847                         p = &(*p)->rb_left;
 848                 else if (new->r_tid > req->r_tid)
 849                         p = &(*p)->rb_right;
 850                 else
 851                         BUG();
 852         }
 853
 854         rb_link_node(&new->r_node, parent, p);
 855         rb_insert_color(&new->r_node, &osdc->requests);
 856 }
 857
 858 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 859                                                  u64 tid)
 860 {
 861         struct ceph_osd_request *req;
 862         struct rb_node *n = osdc->requests.rb_node;
 863
 864         while (n) {
 865                 req = rb_entry(n, struct ceph_osd_request, r_node);
 866                 if (tid < req->r_tid)
 867                         n = n->rb_left;
 868                 else if (tid > req->r_tid)
 869                         n = n->rb_right;
 870                 else
 871                         return req;
 872         }
 873         return NULL;
 874 }
 875
 876 static struct ceph_osd_request *
 877 __lookup_request_ge(struct ceph_osd_client *osdc,
 878                     u64 tid)
 879 {
 880         struct ceph_osd_request *req;
 881         struct rb_node *n = osdc->requests.rb_node;
 882
 883         while (n) {
 884                 req = rb_entry(n, struct ceph_osd_request, r_node);
 885                 if (tid < req->r_tid) {
 886                         if (!n->rb_left)
 887                                 return req;
 888                         n = n->rb_left;
 889                 } else if (tid > req->r_tid) {
 890                         n = n->rb_right;
 891                 } else {
 892                         return req;
 893                 }
 894         }
 895         return NULL;
 896 }
 897
 898 static void __kick_linger_request(struct ceph_osd_request *req)
 899 {
 900         struct ceph_osd_client *osdc = req->r_osdc;
 901         struct ceph_osd *osd = req->r_osd;
 902
 903         /*
 904          * Linger requests need to be resent with a new tid to avoid
 905          * the dup op detection logic on the OSDs.  Achieve this with
 906          * a re-register dance instead of open-coding.
 907          */
 908         ceph_osdc_get_request(req);
 909         if (!list_empty(&req->r_linger_item))
 910                 __unregister_linger_request(osdc, req);
 911         else
 912                 __unregister_request(osdc, req);
 913         __register_request(osdc, req);
 914         ceph_osdc_put_request(req);
 915
 916         /*
 917          * Unless request has been registered as both normal and
 918          * lingering, __unregister{,_linger}_request clears r_osd.
 919          * However, here we need to preserve r_osd to make sure we
 920          * requeue on the same OSD.
 921          */
 922         WARN_ON(req->r_osd || !osd);
 923         req->r_osd = osd;
 924
 925         dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
 926         __enqueue_request(req);
 927 }
 928
 929 /*
 930  * Resubmit requests pending on the given osd.
 931  */
 932 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 933                                 struct ceph_osd *osd)
 934 {
 935         struct ceph_osd_request *req, *nreq;
 936         LIST_HEAD(resend);
 937         LIST_HEAD(resend_linger);
 938         int err;
 939
 940         dout("%s osd%d\n", __func__, osd->o_osd);
 941         err = __reset_osd(osdc, osd);
 942         if (err)
 943                 return;
 944
 945         /*
 946          * Build up a list of requests to resend by traversing the
 947          * osd's list of requests.  Requests for a given object are
 948          * sent in tid order, and that is also the order they're
 949          * kept on this list.  Therefore all requests that are in
 950          * flight will be found first, followed by all requests that
 951          * have not yet been sent.  And to resend requests while
 952          * preserving this order we will want to put any sent
 953          * requests back on the front of the osd client's unsent
 954          * list.
 955          *
 956          * So we build a separate ordered list of already-sent
 957          * requests for the affected osd and splice it onto the
 958          * front of the osd client's unsent list.  Once we've seen a
 959          * request that has not yet been sent we're done.  Those
 960          * requests are already sitting right where they belong.
 961          */
 962         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 963                 if (!req->r_sent)
 964                         break;
 965
 966                 if (!req->r_linger) {
 967                         dout("%s requeueing %p tid %llu\n", __func__, req,
 968                              req->r_tid);
 969                         list_move_tail(&req->r_req_lru_item, &resend);
 970                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 971                 } else {
 972                         list_move_tail(&req->r_req_lru_item, &resend_linger);
 973                 }
 974         }
 975         list_splice(&resend, &osdc->req_unsent);
 976
 977         /*
 978          * Both registered and not yet registered linger requests are
 979          * enqueued with a new tid on the same OSD.  We add/move them
 980          * to req_unsent/o_requests at the end to keep things in tid
 981          * order.
 982          */
 983         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 984                                  r_linger_osd_item) {
 985                 WARN_ON(!list_empty(&req->r_req_lru_item));
 986                 __kick_linger_request(req);
 987         }
 988
 989         list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
 990                 __kick_linger_request(req);
 991 }
 992
 993 /*
 994  * If the osd connection drops, we need to resubmit all requests.
 995  */
 996 static void osd_reset(struct ceph_connection *con)
 997 {
 998         struct ceph_osd *osd = con->private;
 999         struct ceph_osd_client *osdc;
1000
1001         if (!osd)
1002                 return;
1003         dout("osd_reset osd%d\n", osd->o_osd);
1004         osdc = osd->o_osdc;
1005         down_read(&osdc->map_sem);
1006         mutex_lock(&osdc->request_mutex);
1007         __kick_osd_requests(osdc, osd);
1008         __send_queued(osdc);
1009         mutex_unlock(&osdc->request_mutex);
1010         up_read(&osdc->map_sem);
1011 }
1012
1013 /*
1014  * Track open sessions with osds.
1015  */
1016 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1017 {
1018         struct ceph_osd *osd;
1019
1020         osd = kzalloc(sizeof(*osd), GFP_NOFS);
1021         if (!osd)
1022                 return NULL;
1023
1024         atomic_set(&osd->o_ref, 1);
1025         osd->o_osdc = osdc;
1026         osd->o_osd = onum;
1027         RB_CLEAR_NODE(&osd->o_node);
1028         INIT_LIST_HEAD(&osd->o_requests);
1029         INIT_LIST_HEAD(&osd->o_linger_requests);
1030         INIT_LIST_HEAD(&osd->o_osd_lru);
1031         osd->o_incarnation = 1;
1032
1033         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1034
1035         INIT_LIST_HEAD(&osd->o_keepalive_item);
1036         return osd;
1037 }
1038
1039 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1040 {
1041         if (atomic_inc_not_zero(&osd->o_ref)) {
1042                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1043                      atomic_read(&osd->o_ref));
1044                 return osd;
1045         } else {
1046                 dout("get_osd %p FAIL\n", osd);
1047                 return NULL;
1048         }
1049 }
1050
1051 static void put_osd(struct ceph_osd *osd)
1052 {
1053         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1054              atomic_read(&osd->o_ref) - 1);
1055         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
1056                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1057
1058                 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1059                 kfree(osd);
1060         }
1061 }
1062
1063 /*
1064  * remove an osd from our map
1065  */
1066 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1067 {
1068         dout("__remove_osd %p\n", osd);
1069         BUG_ON(!list_empty(&osd->o_requests));
1070         BUG_ON(!list_empty(&osd->o_linger_requests));
1071
1072         rb_erase(&osd->o_node, &osdc->osds);
1073         list_del_init(&osd->o_osd_lru);
1074         ceph_con_close(&osd->o_con);
1075         put_osd(osd);
1076 }
1077
1078 static void remove_all_osds(struct ceph_osd_client *osdc)
1079 {
1080         dout("%s %p\n", __func__, osdc);
1081         mutex_lock(&osdc->request_mutex);
1082         while (!RB_EMPTY_ROOT(&osdc->osds)) {
1083                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1084                                                 struct ceph_osd, o_node);
1085                 __remove_osd(osdc, osd);
1086         }
1087         mutex_unlock(&osdc->request_mutex);
1088 }
1089
1090 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1091                               struct ceph_osd *osd)
1092 {
1093         dout("%s %p\n", __func__, osd);
1094         BUG_ON(!list_empty(&osd->o_osd_lru));
1095
1096         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1097         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1098 }
1099
1100 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1101                                   struct ceph_osd *osd)
1102 {
1103         dout("%s %p\n", __func__, osd);
1104
1105         if (list_empty(&osd->o_requests) &&
1106             list_empty(&osd->o_linger_requests))
1107                 __move_osd_to_lru(osdc, osd);
1108 }
1109
1110 static void __remove_osd_from_lru(struct ceph_osd *osd)
1111 {
1112         dout("__remove_osd_from_lru %p\n", osd);
1113         if (!list_empty(&osd->o_osd_lru))
1114                 list_del_init(&osd->o_osd_lru);
1115 }
1116
1117 static void remove_old_osds(struct ceph_osd_client *osdc)
1118 {
1119         struct ceph_osd *osd, *nosd;
1120
1121         dout("__remove_old_osds %p\n", osdc);
1122         mutex_lock(&osdc->request_mutex);
1123         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1124                 if (time_before(jiffies, osd->lru_ttl))
1125                         break;
1126                 __remove_osd(osdc, osd);
1127         }
1128         mutex_unlock(&osdc->request_mutex);
1129 }
1130
1131 /*
1132  * reset osd connect
1133  */
1134 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1135 {
1136         struct ceph_entity_addr *peer_addr;
1137
1138         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1139         if (list_empty(&osd->o_requests) &&
1140             list_empty(&osd->o_linger_requests)) {
1141                 __remove_osd(osdc, osd);
1142
1143                 return -ENODEV;
1144         }
1145
1146         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1147         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1148                         !ceph_con_opened(&osd->o_con)) {
1149                 struct ceph_osd_request *req;
1150
1151                 dout("osd addr hasn't changed and connection never opened, "
1152                      "letting msgr retry\n");
1153                 /* touch each r_stamp for handle_timeout()'s benfit */
1154                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1155                         req->r_stamp = jiffies;
1156
1157                 return -EAGAIN;
1158         }
1159
1160         ceph_con_close(&osd->o_con);
1161         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1162         osd->o_incarnation++;
1163
1164         return 0;
1165 }
1166
1167 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1168 {
1169         struct rb_node **p = &osdc->osds.rb_node;
1170         struct rb_node *parent = NULL;
1171         struct ceph_osd *osd = NULL;
1172
1173         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1174         while (*p) {
1175                 parent = *p;
1176                 osd = rb_entry(parent, struct ceph_osd, o_node);
1177                 if (new->o_osd < osd->o_osd)
1178                         p = &(*p)->rb_left;
1179                 else if (new->o_osd > osd->o_osd)
1180                         p = &(*p)->rb_right;
1181                 else
1182                         BUG();
1183         }
1184
1185         rb_link_node(&new->o_node, parent, p);
1186         rb_insert_color(&new->o_node, &osdc->osds);
1187 }
1188
1189 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1190 {
1191         struct ceph_osd *osd;
1192         struct rb_node *n = osdc->osds.rb_node;
1193
1194         while (n) {
1195                 osd = rb_entry(n, struct ceph_osd, o_node);
1196                 if (o < osd->o_osd)
1197                         n = n->rb_left;
1198                 else if (o > osd->o_osd)
1199                         n = n->rb_right;
1200                 else
1201                         return osd;
1202         }
1203         return NULL;
1204 }
1205
1206 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1207 {
1208         schedule_delayed_work(&osdc->timeout_work,
1209                         osdc->client->options->osd_keepalive_timeout * HZ);
1210 }
1211
1212 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1213 {
1214         cancel_delayed_work(&osdc->timeout_work);
1215 }
1216
1217 /*
1218  * Register request, assign tid.  If this is the first request, set up
1219  * the timeout event.
1220  */
1221 static void __register_request(struct ceph_osd_client *osdc,
1222                                struct ceph_osd_request *req)
1223 {
1224         req->r_tid = ++osdc->last_tid;
1225         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1226         dout("__register_request %p tid %lld\n", req, req->r_tid);
1227         __insert_request(osdc, req);
1228         ceph_osdc_get_request(req);
1229         osdc->num_requests++;
1230         if (osdc->num_requests == 1) {
1231                 dout(" first request, scheduling timeout\n");
1232                 __schedule_osd_timeout(osdc);
1233         }
1234 }
1235
1236 /*
1237  * called under osdc->request_mutex
1238  */
1239 static void __unregister_request(struct ceph_osd_client *osdc,
1240                                  struct ceph_osd_request *req)
1241 {
1242         if (RB_EMPTY_NODE(&req->r_node)) {
1243                 dout("__unregister_request %p tid %lld not registered\n",
1244                         req, req->r_tid);
1245                 return;
1246         }
1247
1248         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1249         rb_erase(&req->r_node, &osdc->requests);
1250         RB_CLEAR_NODE(&req->r_node);
1251         osdc->num_requests--;
1252
1253         if (req->r_osd) {
1254                 /* make sure the original request isn't in flight. */
1255                 ceph_msg_revoke(req->r_request);
1256
1257                 list_del_init(&req->r_osd_item);
1258                 maybe_move_osd_to_lru(osdc, req->r_osd);
1259                 if (list_empty(&req->r_linger_osd_item))
1260                         req->r_osd = NULL;
1261         }
1262
1263         list_del_init(&req->r_req_lru_item);
1264         ceph_osdc_put_request(req);
1265
1266         if (osdc->num_requests == 0) {
1267                 dout(" no requests, canceling timeout\n");
1268                 __cancel_osd_timeout(osdc);
1269         }
1270 }
1271
1272 /*
1273  * Cancel a previously queued request message
1274  */
1275 static void __cancel_request(struct ceph_osd_request *req)
1276 {
1277         if (req->r_sent && req->r_osd) {
1278                 ceph_msg_revoke(req->r_request);
1279                 req->r_sent = 0;
1280         }
1281 }
1282
1283 static void __register_linger_request(struct ceph_osd_client *osdc,
1284                                     struct ceph_osd_request *req)
1285 {
1286         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1287         WARN_ON(!req->r_linger);
1288
1289         ceph_osdc_get_request(req);
1290         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1291         if (req->r_osd)
1292                 list_add_tail(&req->r_linger_osd_item,
1293                               &req->r_osd->o_linger_requests);
1294 }
1295
1296 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1297                                         struct ceph_osd_request *req)
1298 {
1299         WARN_ON(!req->r_linger);
1300
1301         if (list_empty(&req->r_linger_item)) {
1302                 dout("%s %p tid %llu not registered\n", __func__, req,
1303                      req->r_tid);
1304                 return;
1305         }
1306
1307         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1308         list_del_init(&req->r_linger_item);
1309
1310         if (req->r_osd) {
1311                 list_del_init(&req->r_linger_osd_item);
1312                 maybe_move_osd_to_lru(osdc, req->r_osd);
1313                 if (list_empty(&req->r_osd_item))
1314                         req->r_osd = NULL;
1315         }
1316         ceph_osdc_put_request(req);
1317 }
1318
1319 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1320                                   struct ceph_osd_request *req)
1321 {
1322         if (!req->r_linger) {
1323                 dout("set_request_linger %p\n", req);
1324                 req->r_linger = 1;
1325         }
1326 }
1327 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1328
1329 /*
1330  * Returns whether a request should be blocked from being sent
1331  * based on the current osdmap and osd_client settings.
1332  *
1333  * Caller should hold map_sem for read.
1334  */
1335 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1336                                    struct ceph_osd_request *req)
1337 {
1338         bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1339         bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1340                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1341         return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1342                 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1343 }
1344
1345 /*
1346  * Calculate mapping of a request to a PG.  Takes tiering into account.
1347  */
1348 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1349                              struct ceph_osd_request *req,
1350                              struct ceph_pg *pg_out)
1351 {
1352         bool need_check_tiering;
1353
1354         need_check_tiering = false;
1355         if (req->r_target_oloc.pool == -1) {
1356                 req->r_target_oloc = req->r_base_oloc; /* struct */
1357                 need_check_tiering = true;
1358         }
1359         if (req->r_target_oid.name_len == 0) {
1360                 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1361                 need_check_tiering = true;
1362         }
1363
1364         if (need_check_tiering &&
1365             (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1366                 struct ceph_pg_pool_info *pi;
1367
1368                 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1369                 if (pi) {
1370                         if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1371                             pi->read_tier >= 0)
1372                                 req->r_target_oloc.pool = pi->read_tier;
1373                         if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1374                             pi->write_tier >= 0)
1375                                 req->r_target_oloc.pool = pi->write_tier;
1376                 }
1377                 /* !pi is caught in ceph_oloc_oid_to_pg() */
1378         }
1379
1380         return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1381                                    &req->r_target_oid, pg_out);
1382 }
1383
1384 static void __enqueue_request(struct ceph_osd_request *req)
1385 {
1386         struct ceph_osd_client *osdc = req->r_osdc;
1387
1388         dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1389              req->r_osd ? req->r_osd->o_osd : -1);
1390
1391         if (req->r_osd) {
1392                 __remove_osd_from_lru(req->r_osd);
1393                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1394                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1395         } else {
1396                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1397         }
1398 }
1399
1400 /*
1401  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1402  * (as needed), and set the request r_osd appropriately.  If there is
1403  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1404  * (unsent, homeless) or leave on in-flight lru.
1405  *
1406  * Return 0 if unchanged, 1 if changed, or negative on error.
1407  *
1408  * Caller should hold map_sem for read and request_mutex.
1409  */
1410 static int __map_request(struct ceph_osd_client *osdc,
1411                          struct ceph_osd_request *req, int force_resend)
1412 {
1413         struct ceph_pg pgid;
1414         int acting[CEPH_PG_MAX_SIZE];
1415         int num, o;
1416         int err;
1417         bool was_paused;
1418
1419         dout("map_request %p tid %lld\n", req, req->r_tid);
1420
1421         err = __calc_request_pg(osdc->osdmap, req, &pgid);
1422         if (err) {
1423                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1424                 return err;
1425         }
1426         req->r_pgid = pgid;
1427
1428         num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1429         if (num < 0)
1430                 num = 0;
1431
1432         was_paused = req->r_paused;
1433         req->r_paused = __req_should_be_paused(osdc, req);
1434         if (was_paused && !req->r_paused)
1435                 force_resend = 1;
1436
1437         if ((!force_resend &&
1438              req->r_osd && req->r_osd->o_osd == o &&
1439              req->r_sent >= req->r_osd->o_incarnation &&
1440              req->r_num_pg_osds == num &&
1441              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1442             (req->r_osd == NULL && o == -1) ||
1443             req->r_paused)
1444                 return 0;  /* no change */
1445
1446         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1447              req->r_tid, pgid.pool, pgid.seed, o,
1448              req->r_osd ? req->r_osd->o_osd : -1);
1449
1450         /* record full pg acting set */
1451         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1452         req->r_num_pg_osds = num;
1453
1454         if (req->r_osd) {
1455                 __cancel_request(req);
1456                 list_del_init(&req->r_osd_item);
1457                 req->r_osd = NULL;
1458         }
1459
1460         req->r_osd = __lookup_osd(osdc, o);
1461         if (!req->r_osd && o >= 0) {
1462                 err = -ENOMEM;
1463                 req->r_osd = create_osd(osdc, o);
1464                 if (!req->r_osd) {
1465                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1466                         goto out;
1467                 }
1468
1469                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1470                 __insert_osd(osdc, req->r_osd);
1471
1472                 ceph_con_open(&req->r_osd->o_con,
1473                               CEPH_ENTITY_TYPE_OSD, o,
1474                               &osdc->osdmap->osd_addr[o]);
1475         }
1476
1477         __enqueue_request(req);
1478         err = 1;   /* osd or pg changed */
1479
1480 out:
1481         return err;
1482 }
1483
1484 /*
1485  * caller should hold map_sem (for read) and request_mutex
1486  */
1487 static void __send_request(struct ceph_osd_client *osdc,
1488                            struct ceph_osd_request *req)
1489 {
1490         void *p;
1491
1492         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1493              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1494              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1495
1496         /* fill in message content that changes each time we send it */
1497         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1498         put_unaligned_le32(req->r_flags, req->r_request_flags);
1499         put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1500         p = req->r_request_pgid;
1501         ceph_encode_64(&p, req->r_pgid.pool);
1502         ceph_encode_32(&p, req->r_pgid.seed);
1503         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1504         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1505                sizeof(req->r_reassert_version));
1506
1507         req->r_stamp = jiffies;
1508         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1509
1510         ceph_msg_get(req->r_request); /* send consumes a ref */
1511
1512         req->r_sent = req->r_osd->o_incarnation;
1513
1514         ceph_con_send(&req->r_osd->o_con, req->r_request);
1515 }
1516
1517 /*
1518  * Send any requests in the queue (req_unsent).
1519  */
1520 static void __send_queued(struct ceph_osd_client *osdc)
1521 {
1522         struct ceph_osd_request *req, *tmp;
1523
1524         dout("__send_queued\n");
1525         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1526                 __send_request(osdc, req);
1527 }
1528
1529 /*
1530  * Caller should hold map_sem for read and request_mutex.
1531  */
1532 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1533                                      struct ceph_osd_request *req,
1534                                      bool nofail)
1535 {
1536         int rc;
1537
1538         __register_request(osdc, req);
1539         req->r_sent = 0;
1540         req->r_got_reply = 0;
1541         rc = __map_request(osdc, req, 0);
1542         if (rc < 0) {
1543                 if (nofail) {
1544                         dout("osdc_start_request failed map, "
1545                                 " will retry %lld\n", req->r_tid);
1546                         rc = 0;
1547                 } else {
1548                         __unregister_request(osdc, req);
1549                 }
1550                 return rc;
1551         }
1552
1553         if (req->r_osd == NULL) {
1554                 dout("send_request %p no up osds in pg\n", req);
1555                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1556         } else {
1557                 __send_queued(osdc);
1558         }
1559
1560         return 0;
1561 }
1562
1563 /*
1564  * Timeout callback, called every N seconds when 1 or more osd
1565  * requests has been active for more than N seconds.  When this
1566  * happens, we ping all OSDs with requests who have timed out to
1567  * ensure any communications channel reset is detected.  Reset the
1568  * request timeouts another N seconds in the future as we go.
1569  * Reschedule the timeout event another N seconds in future (unless
1570  * there are no open requests).
1571  */
1572 static void handle_timeout(struct work_struct *work)
1573 {
1574         struct ceph_osd_client *osdc =
1575                 container_of(work, struct ceph_osd_client, timeout_work.work);
1576         struct ceph_osd_request *req;
1577         struct ceph_osd *osd;
1578         unsigned long keepalive =
1579                 osdc->client->options->osd_keepalive_timeout * HZ;
1580         struct list_head slow_osds;
1581         dout("timeout\n");
1582         down_read(&osdc->map_sem);
1583
1584         ceph_monc_request_next_osdmap(&osdc->client->monc);
1585
1586         mutex_lock(&osdc->request_mutex);
1587
1588         /*
1589          * ping osds that are a bit slow.  this ensures that if there
1590          * is a break in the TCP connection we will notice, and reopen
1591          * a connection with that osd (from the fault callback).
1592          */
1593         INIT_LIST_HEAD(&slow_osds);
1594         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1595                 if (time_before(jiffies, req->r_stamp + keepalive))
1596                         break;
1597
1598                 osd = req->r_osd;
1599                 BUG_ON(!osd);
1600                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1601                      req->r_tid, osd->o_osd);
1602                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1603         }
1604         while (!list_empty(&slow_osds)) {
1605                 osd = list_entry(slow_osds.next, struct ceph_osd,
1606                                  o_keepalive_item);
1607                 list_del_init(&osd->o_keepalive_item);
1608                 ceph_con_keepalive(&osd->o_con);
1609         }
1610
1611         __schedule_osd_timeout(osdc);
1612         __send_queued(osdc);
1613         mutex_unlock(&osdc->request_mutex);
1614         up_read(&osdc->map_sem);
1615 }
1616
1617 static void handle_osds_timeout(struct work_struct *work)
1618 {
1619         struct ceph_osd_client *osdc =
1620                 container_of(work, struct ceph_osd_client,
1621                              osds_timeout_work.work);
1622         unsigned long delay =
1623                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1624
1625         dout("osds timeout\n");
1626         down_read(&osdc->map_sem);
1627         remove_old_osds(osdc);
1628         up_read(&osdc->map_sem);
1629
1630         schedule_delayed_work(&osdc->osds_timeout_work,
1631                               round_jiffies_relative(delay));
1632 }
1633
1634 static int ceph_oloc_decode(void **p, void *end,
1635                             struct ceph_object_locator *oloc)
1636 {
1637         u8 struct_v, struct_cv;
1638         u32 len;
1639         void *struct_end;
1640         int ret = 0;
1641
1642         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1643         struct_v = ceph_decode_8(p);
1644         struct_cv = ceph_decode_8(p);
1645         if (struct_v < 3) {
1646                 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1647                         struct_v, struct_cv);
1648                 goto e_inval;
1649         }
1650         if (struct_cv > 6) {
1651                 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1652                         struct_v, struct_cv);
1653                 goto e_inval;
1654         }
1655         len = ceph_decode_32(p);
1656         ceph_decode_need(p, end, len, e_inval);
1657         struct_end = *p + len;
1658
1659         oloc->pool = ceph_decode_64(p);
1660         *p += 4; /* skip preferred */
1661
1662         len = ceph_decode_32(p);
1663         if (len > 0) {
1664                 pr_warn("ceph_object_locator::key is set\n");
1665                 goto e_inval;
1666         }
1667
1668         if (struct_v >= 5) {
1669                 len = ceph_decode_32(p);
1670                 if (len > 0) {
1671                         pr_warn("ceph_object_locator::nspace is set\n");
1672                         goto e_inval;
1673                 }
1674         }
1675
1676         if (struct_v >= 6) {
1677                 s64 hash = ceph_decode_64(p);
1678                 if (hash != -1) {
1679                         pr_warn("ceph_object_locator::hash is set\n");
1680                         goto e_inval;
1681                 }
1682         }
1683
1684         /* skip the rest */
1685         *p = struct_end;
1686 out:
1687         return ret;
1688
1689 e_inval:
1690         ret = -EINVAL;
1691         goto out;
1692 }
1693
1694 static int ceph_redirect_decode(void **p, void *end,
1695                                 struct ceph_request_redirect *redir)
1696 {
1697         u8 struct_v, struct_cv;
1698         u32 len;
1699         void *struct_end;
1700         int ret;
1701
1702         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1703         struct_v = ceph_decode_8(p);
1704         struct_cv = ceph_decode_8(p);
1705         if (struct_cv > 1) {
1706                 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1707                         struct_v, struct_cv);
1708                 goto e_inval;
1709         }
1710         len = ceph_decode_32(p);
1711         ceph_decode_need(p, end, len, e_inval);
1712         struct_end = *p + len;
1713
1714         ret = ceph_oloc_decode(p, end, &redir->oloc);
1715         if (ret)
1716                 goto out;
1717
1718         len = ceph_decode_32(p);
1719         if (len > 0) {
1720                 pr_warn("ceph_request_redirect::object_name is set\n");
1721                 goto e_inval;
1722         }
1723
1724         len = ceph_decode_32(p);
1725         *p += len; /* skip osd_instructions */
1726
1727         /* skip the rest */
1728         *p = struct_end;
1729 out:
1730         return ret;
1731
1732 e_inval:
1733         ret = -EINVAL;
1734         goto out;
1735 }
1736
1737 static void complete_request(struct ceph_osd_request *req)
1738 {
1739         complete_all(&req->r_safe_completion);  /* fsync waiter */
1740 }
1741
1742 /*
1743  * handle osd op reply.  either call the callback if it is specified,
1744  * or do the completion to wake up the waiting thread.
1745  */
1746 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1747                          struct ceph_connection *con)
1748 {
1749         void *p, *end;
1750         struct ceph_osd_request *req;
1751         struct ceph_request_redirect redir;
1752         u64 tid;
1753         int object_len;
1754         unsigned int numops;
1755         int payload_len, flags;
1756         s32 result;
1757         s32 retry_attempt;
1758         struct ceph_pg pg;
1759         int err;
1760         u32 reassert_epoch;
1761         u64 reassert_version;
1762         u32 osdmap_epoch;
1763         int already_completed;
1764         u32 bytes;
1765         unsigned int i;
1766
1767         tid = le64_to_cpu(msg->hdr.tid);
1768         dout("handle_reply %p tid %llu\n", msg, tid);
1769
1770         p = msg->front.iov_base;
1771         end = p + msg->front.iov_len;
1772
1773         ceph_decode_need(&p, end, 4, bad);
1774         object_len = ceph_decode_32(&p);
1775         ceph_decode_need(&p, end, object_len, bad);
1776         p += object_len;
1777
1778         err = ceph_decode_pgid(&p, end, &pg);
1779         if (err)
1780                 goto bad;
1781
1782         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1783         flags = ceph_decode_64(&p);
1784         result = ceph_decode_32(&p);
1785         reassert_epoch = ceph_decode_32(&p);
1786         reassert_version = ceph_decode_64(&p);
1787         osdmap_epoch = ceph_decode_32(&p);
1788
1789         /* lookup */
1790         down_read(&osdc->map_sem);
1791         mutex_lock(&osdc->request_mutex);
1792         req = __lookup_request(osdc, tid);
1793         if (req == NULL) {
1794                 dout("handle_reply tid %llu dne\n", tid);
1795                 goto bad_mutex;
1796         }
1797         ceph_osdc_get_request(req);
1798
1799         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1800              req, result);
1801
1802         ceph_decode_need(&p, end, 4, bad_put);
1803         numops = ceph_decode_32(&p);
1804         if (numops > CEPH_OSD_MAX_OP)
1805                 goto bad_put;
1806         if (numops != req->r_num_ops)
1807                 goto bad_put;
1808         payload_len = 0;
1809         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1810         for (i = 0; i < numops; i++) {
1811                 struct ceph_osd_op *op = p;
1812                 int len;
1813
1814                 len = le32_to_cpu(op->payload_len);
1815                 req->r_reply_op_len[i] = len;
1816                 dout(" op %d has %d bytes\n", i, len);
1817                 payload_len += len;
1818                 p += sizeof(*op);
1819         }
1820         bytes = le32_to_cpu(msg->hdr.data_len);
1821         if (payload_len != bytes) {
1822                 pr_warning("sum of op payload lens %d != data_len %d",
1823                            payload_len, bytes);
1824                 goto bad_put;
1825         }
1826
1827         ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1828         retry_attempt = ceph_decode_32(&p);
1829         for (i = 0; i < numops; i++)
1830                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1831
1832         if (le16_to_cpu(msg->hdr.version) >= 6) {
1833                 p += 8 + 4; /* skip replay_version */
1834                 p += 8; /* skip user_version */
1835
1836                 err = ceph_redirect_decode(&p, end, &redir);
1837                 if (err)
1838                         goto bad_put;
1839         } else {
1840                 redir.oloc.pool = -1;
1841         }
1842
1843         if (redir.oloc.pool != -1) {
1844                 dout("redirect pool %lld\n", redir.oloc.pool);
1845
1846                 __unregister_request(osdc, req);
1847
1848                 req->r_target_oloc = redir.oloc; /* struct */
1849
1850                 /*
1851                  * Start redirect requests with nofail=true.  If
1852                  * mapping fails, request will end up on the notarget
1853                  * list, waiting for the new osdmap (which can take
1854                  * a while), even though the original request mapped
1855                  * successfully.  In the future we might want to follow
1856                  * original request's nofail setting here.
1857                  */
1858                 err = __ceph_osdc_start_request(osdc, req, true);
1859                 BUG_ON(err);
1860
1861                 goto out_unlock;
1862         }
1863
1864         already_completed = req->r_got_reply;
1865         if (!req->r_got_reply) {
1866                 req->r_result = result;
1867                 dout("handle_reply result %d bytes %d\n", req->r_result,
1868                      bytes);
1869                 if (req->r_result == 0)
1870                         req->r_result = bytes;
1871
1872                 /* in case this is a write and we need to replay, */
1873                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1874                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1875
1876                 req->r_got_reply = 1;
1877         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1878                 dout("handle_reply tid %llu dup ack\n", tid);
1879                 goto out_unlock;
1880         }
1881
1882         dout("handle_reply tid %llu flags %d\n", tid, flags);
1883
1884         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1885                 __register_linger_request(osdc, req);
1886
1887         /* either this is a read, or we got the safe response */
1888         if (result < 0 ||
1889             (flags & CEPH_OSD_FLAG_ONDISK) ||
1890             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1891                 __unregister_request(osdc, req);
1892
1893         mutex_unlock(&osdc->request_mutex);
1894         up_read(&osdc->map_sem);
1895
1896         if (!already_completed) {
1897                 if (req->r_unsafe_callback &&
1898                     result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1899                         req->r_unsafe_callback(req, true);
1900                 if (req->r_callback)
1901                         req->r_callback(req, msg);
1902                 else
1903                         complete_all(&req->r_completion);
1904         }
1905
1906         if (flags & CEPH_OSD_FLAG_ONDISK) {
1907                 if (req->r_unsafe_callback && already_completed)
1908                         req->r_unsafe_callback(req, false);
1909                 complete_request(req);
1910         }
1911
1912 out:
1913         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1914         ceph_osdc_put_request(req);
1915         return;
1916 out_unlock:
1917         mutex_unlock(&osdc->request_mutex);
1918         up_read(&osdc->map_sem);
1919         goto out;
1920
1921 bad_put:
1922         req->r_result = -EIO;
1923         __unregister_request(osdc, req);
1924         if (req->r_callback)
1925                 req->r_callback(req, msg);
1926         else
1927                 complete_all(&req->r_completion);
1928         complete_request(req);
1929         ceph_osdc_put_request(req);
1930 bad_mutex:
1931         mutex_unlock(&osdc->request_mutex);
1932         up_read(&osdc->map_sem);
1933 bad:
1934         pr_err("corrupt osd_op_reply got %d %d\n",
1935                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1936         ceph_msg_dump(msg);
1937 }
1938
1939 static void reset_changed_osds(struct ceph_osd_client *osdc)
1940 {
1941         struct rb_node *p, *n;
1942
1943         for (p = rb_first(&osdc->osds); p; p = n) {
1944                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1945
1946                 n = rb_next(p);
1947                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1948                     memcmp(&osd->o_con.peer_addr,
1949                            ceph_osd_addr(osdc->osdmap,
1950                                          osd->o_osd),
1951                            sizeof(struct ceph_entity_addr)) != 0)
1952                         __reset_osd(osdc, osd);
1953         }
1954 }
1955
1956 /*
1957  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1958  * no osd, request a new map.
1959  *
1960  * Caller should hold map_sem for read.
1961  */
1962 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1963                           bool force_resend_writes)
1964 {
1965         struct ceph_osd_request *req, *nreq;
1966         struct rb_node *p;
1967         int needmap = 0;
1968         int err;
1969         bool force_resend_req;
1970
1971         dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1972                 force_resend_writes ? " (force resend writes)" : "");
1973         mutex_lock(&osdc->request_mutex);
1974         for (p = rb_first(&osdc->requests); p; ) {
1975                 req = rb_entry(p, struct ceph_osd_request, r_node);
1976                 p = rb_next(p);
1977
1978                 /*
1979                  * For linger requests that have not yet been
1980                  * registered, move them to the linger list; they'll
1981                  * be sent to the osd in the loop below.  Unregister
1982                  * the request before re-registering it as a linger
1983                  * request to ensure the __map_request() below
1984                  * will decide it needs to be sent.
1985                  */
1986                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1987                         dout("%p tid %llu restart on osd%d\n",
1988                              req, req->r_tid,
1989                              req->r_osd ? req->r_osd->o_osd : -1);
1990                         ceph_osdc_get_request(req);
1991                         __unregister_request(osdc, req);
1992                         __register_linger_request(osdc, req);
1993                         ceph_osdc_put_request(req);
1994                         continue;
1995                 }
1996
1997                 force_resend_req = force_resend ||
1998                         (force_resend_writes &&
1999                                 req->r_flags & CEPH_OSD_FLAG_WRITE);
2000                 err = __map_request(osdc, req, force_resend_req);
2001                 if (err < 0)
2002                         continue;  /* error */
2003                 if (req->r_osd == NULL) {
2004                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2005                         needmap++;  /* request a newer map */
2006                 } else if (err > 0) {
2007                         if (!req->r_linger) {
2008                                 dout("%p tid %llu requeued on osd%d\n", req,
2009                                      req->r_tid,
2010                                      req->r_osd ? req->r_osd->o_osd : -1);
2011                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2012                         }
2013                 }
2014         }
2015
2016         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2017                                  r_linger_item) {
2018                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2019
2020                 err = __map_request(osdc, req,
2021                                     force_resend || force_resend_writes);
2022                 dout("__map_request returned %d\n", err);
2023                 if (err == 0)
2024                         continue;  /* no change and no osd was specified */
2025                 if (err < 0)
2026                         continue;  /* hrm! */
2027                 if (req->r_osd == NULL) {
2028                         dout("tid %llu maps to no valid osd\n", req->r_tid);
2029                         needmap++;  /* request a newer map */
2030                         continue;
2031                 }
2032
2033                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
2034                      req->r_osd ? req->r_osd->o_osd : -1);
2035                 __register_request(osdc, req);
2036                 __unregister_linger_request(osdc, req);
2037         }
2038         reset_changed_osds(osdc);
2039         mutex_unlock(&osdc->request_mutex);
2040
2041         if (needmap) {
2042                 dout("%d requests for down osds, need new map\n", needmap);
2043                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2044         }
2045 }
2046
2047
2048 /*
2049  * Process updated osd map.
2050  *
2051  * The message contains any number of incremental and full maps, normally
2052  * indicating some sort of topology change in the cluster.  Kick requests
2053  * off to different OSDs as needed.
2054  */
2055 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2056 {
2057         void *p, *end, *next;
2058         u32 nr_maps, maplen;
2059         u32 epoch;
2060         struct ceph_osdmap *newmap = NULL, *oldmap;
2061         int err;
2062         struct ceph_fsid fsid;
2063         bool was_full;
2064
2065         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2066         p = msg->front.iov_base;
2067         end = p + msg->front.iov_len;
2068
2069         /* verify fsid */
2070         ceph_decode_need(&p, end, sizeof(fsid), bad);
2071         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2072         if (ceph_check_fsid(osdc->client, &fsid) < 0)
2073                 return;
2074
2075         down_write(&osdc->map_sem);
2076
2077         was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2078
2079         /* incremental maps */
2080         ceph_decode_32_safe(&p, end, nr_maps, bad);
2081         dout(" %d inc maps\n", nr_maps);
2082         while (nr_maps > 0) {
2083                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2084                 epoch = ceph_decode_32(&p);
2085                 maplen = ceph_decode_32(&p);
2086                 ceph_decode_need(&p, end, maplen, bad);
2087                 next = p + maplen;
2088                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2089                         dout("applying incremental map %u len %d\n",
2090                              epoch, maplen);
2091                         newmap = osdmap_apply_incremental(&p, next,
2092                                                           osdc->osdmap,
2093                                                           &osdc->client->msgr);
2094                         if (IS_ERR(newmap)) {
2095                                 err = PTR_ERR(newmap);
2096                                 goto bad;
2097                         }
2098                         BUG_ON(!newmap);
2099                         if (newmap != osdc->osdmap) {
2100                                 ceph_osdmap_destroy(osdc->osdmap);
2101                                 osdc->osdmap = newmap;
2102                         }
2103                         was_full = was_full ||
2104                                 ceph_osdmap_flag(osdc->osdmap,
2105                                                  CEPH_OSDMAP_FULL);
2106                         kick_requests(osdc, 0, was_full);
2107                 } else {
2108                         dout("ignoring incremental map %u len %d\n",
2109                              epoch, maplen);
2110                 }
2111                 p = next;
2112                 nr_maps--;
2113         }
2114         if (newmap)
2115                 goto done;
2116
2117         /* full maps */
2118         ceph_decode_32_safe(&p, end, nr_maps, bad);
2119         dout(" %d full maps\n", nr_maps);
2120         while (nr_maps) {
2121                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2122                 epoch = ceph_decode_32(&p);
2123                 maplen = ceph_decode_32(&p);
2124                 ceph_decode_need(&p, end, maplen, bad);
2125                 if (nr_maps > 1) {
2126                         dout("skipping non-latest full map %u len %d\n",
2127                              epoch, maplen);
2128                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2129                         dout("skipping full map %u len %d, "
2130                              "older than our %u\n", epoch, maplen,
2131                              osdc->osdmap->epoch);
2132                 } else {
2133                         int skipped_map = 0;
2134
2135                         dout("taking full map %u len %d\n", epoch, maplen);
2136                         newmap = ceph_osdmap_decode(&p, p+maplen);
2137                         if (IS_ERR(newmap)) {
2138                                 err = PTR_ERR(newmap);
2139                                 goto bad;
2140                         }
2141                         BUG_ON(!newmap);
2142                         oldmap = osdc->osdmap;
2143                         osdc->osdmap = newmap;
2144                         if (oldmap) {
2145                                 if (oldmap->epoch + 1 < newmap->epoch)
2146                                         skipped_map = 1;
2147                                 ceph_osdmap_destroy(oldmap);
2148                         }
2149                         was_full = was_full ||
2150                                 ceph_osdmap_flag(osdc->osdmap,
2151                                                  CEPH_OSDMAP_FULL);
2152                         kick_requests(osdc, skipped_map, was_full);
2153                 }
2154                 p += maplen;
2155                 nr_maps--;
2156         }
2157
2158         if (!osdc->osdmap)
2159                 goto bad;
2160 done:
2161         downgrade_write(&osdc->map_sem);
2162         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2163
2164         /*
2165          * subscribe to subsequent osdmap updates if full to ensure
2166          * we find out when we are no longer full and stop returning
2167          * ENOSPC.
2168          */
2169         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2170                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2171                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2172                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2173
2174         mutex_lock(&osdc->request_mutex);
2175         __send_queued(osdc);
2176         mutex_unlock(&osdc->request_mutex);
2177         up_read(&osdc->map_sem);
2178         wake_up_all(&osdc->client->auth_wq);
2179         return;
2180
2181 bad:
2182         pr_err("osdc handle_map corrupt msg\n");
2183         ceph_msg_dump(msg);
2184         up_write(&osdc->map_sem);
2185 }
2186
2187 /*
2188  * watch/notify callback event infrastructure
2189  *
2190  * These callbacks are used both for watch and notify operations.
2191  */
2192 static void __release_event(struct kref *kref)
2193 {
2194         struct ceph_osd_event *event =
2195                 container_of(kref, struct ceph_osd_event, kref);
2196
2197         dout("__release_event %p\n", event);
2198         kfree(event);
2199 }
2200
2201 static void get_event(struct ceph_osd_event *event)
2202 {
2203         kref_get(&event->kref);
2204 }
2205
2206 void ceph_osdc_put_event(struct ceph_osd_event *event)
2207 {
2208         kref_put(&event->kref, __release_event);
2209 }
2210 EXPORT_SYMBOL(ceph_osdc_put_event);
2211
2212 static void __insert_event(struct ceph_osd_client *osdc,
2213                              struct ceph_osd_event *new)
2214 {
2215         struct rb_node **p = &osdc->event_tree.rb_node;
2216         struct rb_node *parent = NULL;
2217         struct ceph_osd_event *event = NULL;
2218
2219         while (*p) {
2220                 parent = *p;
2221                 event = rb_entry(parent, struct ceph_osd_event, node);
2222                 if (new->cookie < event->cookie)
2223                         p = &(*p)->rb_left;
2224                 else if (new->cookie > event->cookie)
2225                         p = &(*p)->rb_right;
2226                 else
2227                         BUG();
2228         }
2229
2230         rb_link_node(&new->node, parent, p);
2231         rb_insert_color(&new->node, &osdc->event_tree);
2232 }
2233
2234 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2235                                                 u64 cookie)
2236 {
2237         struct rb_node **p = &osdc->event_tree.rb_node;
2238         struct rb_node *parent = NULL;
2239         struct ceph_osd_event *event = NULL;
2240
2241         while (*p) {
2242                 parent = *p;
2243                 event = rb_entry(parent, struct ceph_osd_event, node);
2244                 if (cookie < event->cookie)
2245                         p = &(*p)->rb_left;
2246                 else if (cookie > event->cookie)
2247                         p = &(*p)->rb_right;
2248                 else
2249                         return event;
2250         }
2251         return NULL;
2252 }
2253
2254 static void __remove_event(struct ceph_osd_event *event)
2255 {
2256         struct ceph_osd_client *osdc = event->osdc;
2257
2258         if (!RB_EMPTY_NODE(&event->node)) {
2259                 dout("__remove_event removed %p\n", event);
2260                 rb_erase(&event->node, &osdc->event_tree);
2261                 ceph_osdc_put_event(event);
2262         } else {
2263                 dout("__remove_event didn't remove %p\n", event);
2264         }
2265 }
2266
2267 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2268                            void (*event_cb)(u64, u64, u8, void *),
2269                            void *data, struct ceph_osd_event **pevent)
2270 {
2271         struct ceph_osd_event *event;
2272
2273         event = kmalloc(sizeof(*event), GFP_NOIO);
2274         if (!event)
2275                 return -ENOMEM;
2276
2277         dout("create_event %p\n", event);
2278         event->cb = event_cb;
2279         event->one_shot = 0;
2280         event->data = data;
2281         event->osdc = osdc;
2282         INIT_LIST_HEAD(&event->osd_node);
2283         RB_CLEAR_NODE(&event->node);
2284         kref_init(&event->kref);   /* one ref for us */
2285         kref_get(&event->kref);    /* one ref for the caller */
2286
2287         spin_lock(&osdc->event_lock);
2288         event->cookie = ++osdc->event_count;
2289         __insert_event(osdc, event);
2290         spin_unlock(&osdc->event_lock);
2291
2292         *pevent = event;
2293         return 0;
2294 }
2295 EXPORT_SYMBOL(ceph_osdc_create_event);
2296
2297 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2298 {
2299         struct ceph_osd_client *osdc = event->osdc;
2300
2301         dout("cancel_event %p\n", event);
2302         spin_lock(&osdc->event_lock);
2303         __remove_event(event);
2304         spin_unlock(&osdc->event_lock);
2305         ceph_osdc_put_event(event); /* caller's */
2306 }
2307 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2308
2309
2310 static void do_event_work(struct work_struct *work)
2311 {
2312         struct ceph_osd_event_work *event_work =
2313                 container_of(work, struct ceph_osd_event_work, work);
2314         struct ceph_osd_event *event = event_work->event;
2315         u64 ver = event_work->ver;
2316         u64 notify_id = event_work->notify_id;
2317         u8 opcode = event_work->opcode;
2318
2319         dout("do_event_work completing %p\n", event);
2320         event->cb(ver, notify_id, opcode, event->data);
2321         dout("do_event_work completed %p\n", event);
2322         ceph_osdc_put_event(event);
2323         kfree(event_work);
2324 }
2325
2326
2327 /*
2328  * Process osd watch notifications
2329  */
2330 static void handle_watch_notify(struct ceph_osd_client *osdc,
2331                                 struct ceph_msg *msg)
2332 {
2333         void *p, *end;
2334         u8 proto_ver;
2335         u64 cookie, ver, notify_id;
2336         u8 opcode;
2337         struct ceph_osd_event *event;
2338         struct ceph_osd_event_work *event_work;
2339
2340         p = msg->front.iov_base;
2341         end = p + msg->front.iov_len;
2342
2343         ceph_decode_8_safe(&p, end, proto_ver, bad);
2344         ceph_decode_8_safe(&p, end, opcode, bad);
2345         ceph_decode_64_safe(&p, end, cookie, bad);
2346         ceph_decode_64_safe(&p, end, ver, bad);
2347         ceph_decode_64_safe(&p, end, notify_id, bad);
2348
2349         spin_lock(&osdc->event_lock);
2350         event = __find_event(osdc, cookie);
2351         if (event) {
2352                 BUG_ON(event->one_shot);
2353                 get_event(event);
2354         }
2355         spin_unlock(&osdc->event_lock);
2356         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2357              cookie, ver, event);
2358         if (event) {
2359                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2360                 if (!event_work) {
2361                         dout("ERROR: could not allocate event_work\n");
2362                         goto done_err;
2363                 }
2364                 INIT_WORK(&event_work->work, do_event_work);
2365                 event_work->event = event;
2366                 event_work->ver = ver;
2367                 event_work->notify_id = notify_id;
2368                 event_work->opcode = opcode;
2369                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
2370                         dout("WARNING: failed to queue notify event work\n");
2371                         goto done_err;
2372                 }
2373         }
2374
2375         return;
2376
2377 done_err:
2378         ceph_osdc_put_event(event);
2379         return;
2380
2381 bad:
2382         pr_err("osdc handle_watch_notify corrupt msg\n");
2383 }
2384
2385 /*
2386  * build new request AND message
2387  *
2388  */
2389 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2390                                 struct ceph_snap_context *snapc, u64 snap_id,
2391                                 struct timespec *mtime)
2392 {
2393         struct ceph_msg *msg = req->r_request;
2394         void *p;
2395         size_t msg_size;
2396         int flags = req->r_flags;
2397         u64 data_len;
2398         unsigned int i;
2399
2400         req->r_snapid = snap_id;
2401         req->r_snapc = ceph_get_snap_context(snapc);
2402
2403         /* encode request */
2404         msg->hdr.version = cpu_to_le16(4);
2405
2406         p = msg->front.iov_base;
2407         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2408         req->r_request_osdmap_epoch = p;
2409         p += 4;
2410         req->r_request_flags = p;
2411         p += 4;
2412         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2413                 ceph_encode_timespec(p, mtime);
2414         p += sizeof(struct ceph_timespec);
2415         req->r_request_reassert_version = p;
2416         p += sizeof(struct ceph_eversion); /* will get filled in */
2417
2418         /* oloc */
2419         ceph_encode_8(&p, 4);
2420         ceph_encode_8(&p, 4);
2421         ceph_encode_32(&p, 8 + 4 + 4);
2422         req->r_request_pool = p;
2423         p += 8;
2424         ceph_encode_32(&p, -1);  /* preferred */
2425         ceph_encode_32(&p, 0);   /* key len */
2426
2427         ceph_encode_8(&p, 1);
2428         req->r_request_pgid = p;
2429         p += 8 + 4;
2430         ceph_encode_32(&p, -1);  /* preferred */
2431
2432         /* oid */
2433         ceph_encode_32(&p, req->r_base_oid.name_len);
2434         memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2435         dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2436              req->r_base_oid.name, req->r_base_oid.name_len);
2437         p += req->r_base_oid.name_len;
2438
2439         /* ops--can imply data */
2440         ceph_encode_16(&p, (u16)req->r_num_ops);
2441         data_len = 0;
2442         for (i = 0; i < req->r_num_ops; i++) {
2443                 data_len += osd_req_encode_op(req, p, i);
2444                 p += sizeof(struct ceph_osd_op);
2445         }
2446
2447         /* snaps */
2448         ceph_encode_64(&p, req->r_snapid);
2449         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2450         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2451         if (req->r_snapc) {
2452                 for (i = 0; i < snapc->num_snaps; i++) {
2453                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2454                 }
2455         }
2456
2457         req->r_request_attempts = p;
2458         p += 4;
2459
2460         /* data */
2461         if (flags & CEPH_OSD_FLAG_WRITE) {
2462                 u16 data_off;
2463
2464                 /*
2465                  * The header "data_off" is a hint to the receiver
2466                  * allowing it to align received data into its
2467                  * buffers such that there's no need to re-copy
2468                  * it before writing it to disk (direct I/O).
2469                  */
2470                 data_off = (u16) (off & 0xffff);
2471                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2472         }
2473         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2474
2475         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2476         msg_size = p - msg->front.iov_base;
2477         msg->front.iov_len = msg_size;
2478         msg->hdr.front_len = cpu_to_le32(msg_size);
2479
2480         dout("build_request msg_size was %d\n", (int)msg_size);
2481 }
2482 EXPORT_SYMBOL(ceph_osdc_build_request);
2483
2484 /*
2485  * Register request, send initial attempt.
2486  */
2487 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2488                             struct ceph_osd_request *req,
2489                             bool nofail)
2490 {
2491         int rc;
2492
2493         down_read(&osdc->map_sem);
2494         mutex_lock(&osdc->request_mutex);
2495
2496         rc = __ceph_osdc_start_request(osdc, req, nofail);
2497
2498         mutex_unlock(&osdc->request_mutex);
2499         up_read(&osdc->map_sem);
2500
2501         return rc;
2502 }
2503 EXPORT_SYMBOL(ceph_osdc_start_request);
2504
2505 /*
2506  * Unregister a registered request.  The request is not completed (i.e.
2507  * no callbacks or wakeups) - higher layers are supposed to know what
2508  * they are canceling.
2509  */
2510 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2511 {
2512         struct ceph_osd_client *osdc = req->r_osdc;
2513
2514         mutex_lock(&osdc->request_mutex);
2515         if (req->r_linger)
2516                 __unregister_linger_request(osdc, req);
2517         __unregister_request(osdc, req);
2518         mutex_unlock(&osdc->request_mutex);
2519
2520         dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2521 }
2522 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2523
2524 /*
2525  * wait for a request to complete
2526  */
2527 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2528                            struct ceph_osd_request *req)
2529 {
2530         int rc;
2531
2532         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2533
2534         rc = wait_for_completion_interruptible(&req->r_completion);
2535         if (rc < 0) {
2536                 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2537                 ceph_osdc_cancel_request(req);
2538                 complete_request(req);
2539                 return rc;
2540         }
2541
2542         dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2543              req->r_result);
2544         return req->r_result;
2545 }
2546 EXPORT_SYMBOL(ceph_osdc_wait_request);
2547
2548 /*
2549  * sync - wait for all in-flight requests to flush.  avoid starvation.
2550  */
2551 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2552 {
2553         struct ceph_osd_request *req;
2554         u64 last_tid, next_tid = 0;
2555
2556         mutex_lock(&osdc->request_mutex);
2557         last_tid = osdc->last_tid;
2558         while (1) {
2559                 req = __lookup_request_ge(osdc, next_tid);
2560                 if (!req)
2561                         break;
2562                 if (req->r_tid > last_tid)
2563                         break;
2564
2565                 next_tid = req->r_tid + 1;
2566                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2567                         continue;
2568
2569                 ceph_osdc_get_request(req);
2570                 mutex_unlock(&osdc->request_mutex);
2571                 dout("sync waiting on tid %llu (last is %llu)\n",
2572                      req->r_tid, last_tid);
2573                 wait_for_completion(&req->r_safe_completion);
2574                 mutex_lock(&osdc->request_mutex);
2575                 ceph_osdc_put_request(req);
2576         }
2577         mutex_unlock(&osdc->request_mutex);
2578         dout("sync done (thru tid %llu)\n", last_tid);
2579 }
2580 EXPORT_SYMBOL(ceph_osdc_sync);
2581
2582 /*
2583  * Call all pending notify callbacks - for use after a watch is
2584  * unregistered, to make sure no more callbacks for it will be invoked
2585  */
2586 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2587 {
2588         flush_workqueue(osdc->notify_wq);
2589 }
2590 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2591
2592
2593 /*
2594  * init, shutdown
2595  */
2596 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2597 {
2598         int err;
2599
2600         dout("init\n");
2601         osdc->client = client;
2602         osdc->osdmap = NULL;
2603         init_rwsem(&osdc->map_sem);
2604         init_completion(&osdc->map_waiters);
2605         osdc->last_requested_map = 0;
2606         mutex_init(&osdc->request_mutex);
2607         osdc->last_tid = 0;
2608         osdc->osds = RB_ROOT;
2609         INIT_LIST_HEAD(&osdc->osd_lru);
2610         osdc->requests = RB_ROOT;
2611         INIT_LIST_HEAD(&osdc->req_lru);
2612         INIT_LIST_HEAD(&osdc->req_unsent);
2613         INIT_LIST_HEAD(&osdc->req_notarget);
2614         INIT_LIST_HEAD(&osdc->req_linger);
2615         osdc->num_requests = 0;
2616         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2617         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2618         spin_lock_init(&osdc->event_lock);
2619         osdc->event_tree = RB_ROOT;
2620         osdc->event_count = 0;
2621
2622         schedule_delayed_work(&osdc->osds_timeout_work,
2623            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2624
2625         err = -ENOMEM;
2626         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2627                                         sizeof(struct ceph_osd_request));
2628         if (!osdc->req_mempool)
2629                 goto out;
2630
2631         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2632                                 OSD_OP_FRONT_LEN, 10, true,
2633                                 "osd_op");
2634         if (err < 0)
2635                 goto out_mempool;
2636         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2637                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2638                                 "osd_op_reply");
2639         if (err < 0)
2640                 goto out_msgpool;
2641
2642         err = -ENOMEM;
2643         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2644         if (!osdc->notify_wq)
2645                 goto out_msgpool_reply;
2646
2647         return 0;
2648
2649 out_msgpool_reply:
2650         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2651 out_msgpool:
2652         ceph_msgpool_destroy(&osdc->msgpool_op);
2653 out_mempool:
2654         mempool_destroy(osdc->req_mempool);
2655 out:
2656         return err;
2657 }
2658
2659 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2660 {
2661         flush_workqueue(osdc->notify_wq);
2662         destroy_workqueue(osdc->notify_wq);
2663         cancel_delayed_work_sync(&osdc->timeout_work);
2664         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2665         if (osdc->osdmap) {
2666                 ceph_osdmap_destroy(osdc->osdmap);
2667                 osdc->osdmap = NULL;
2668         }
2669         remove_all_osds(osdc);
2670         mempool_destroy(osdc->req_mempool);
2671         ceph_msgpool_destroy(&osdc->msgpool_op);
2672         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2673 }
2674
2675 /*
2676  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2677  * *plen.  Return number of bytes read, or error.
2678  */
2679 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2680                         struct ceph_vino vino, struct ceph_file_layout *layout,
2681                         u64 off, u64 *plen,
2682                         u32 truncate_seq, u64 truncate_size,
2683                         struct page **pages, int num_pages, int page_align)
2684 {
2685         struct ceph_osd_request *req;
2686         int rc = 0;
2687
2688         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2689              vino.snap, off, *plen);
2690         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2691                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2692                                     NULL, truncate_seq, truncate_size,
2693                                     false);
2694         if (IS_ERR(req))
2695                 return PTR_ERR(req);
2696
2697         /* it may be a short read due to an object boundary */
2698
2699         osd_req_op_extent_osd_data_pages(req, 0,
2700                                 pages, *plen, page_align, false, false);
2701
2702         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2703              off, *plen, *plen, page_align);
2704
2705         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2706
2707         rc = ceph_osdc_start_request(osdc, req, false);
2708         if (!rc)
2709                 rc = ceph_osdc_wait_request(osdc, req);
2710
2711         ceph_osdc_put_request(req);
2712         dout("readpages result %d\n", rc);
2713         return rc;
2714 }
2715 EXPORT_SYMBOL(ceph_osdc_readpages);
2716
2717 /*
2718  * do a synchronous write on N pages
2719  */
2720 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2721                          struct ceph_file_layout *layout,
2722                          struct ceph_snap_context *snapc,
2723                          u64 off, u64 len,
2724                          u32 truncate_seq, u64 truncate_size,
2725                          struct timespec *mtime,
2726                          struct page **pages, int num_pages)
2727 {
2728         struct ceph_osd_request *req;
2729         int rc = 0;
2730         int page_align = off & ~PAGE_MASK;
2731
2732         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2733         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2734                                     CEPH_OSD_OP_WRITE,
2735                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2736                                     snapc, truncate_seq, truncate_size,
2737                                     true);
2738         if (IS_ERR(req))
2739                 return PTR_ERR(req);
2740
2741         /* it may be a short write due to an object boundary */
2742         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2743                                 false, false);
2744         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2745
2746         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2747
2748         rc = ceph_osdc_start_request(osdc, req, true);
2749         if (!rc)
2750                 rc = ceph_osdc_wait_request(osdc, req);
2751
2752         ceph_osdc_put_request(req);
2753         if (rc == 0)
2754                 rc = len;
2755         dout("writepages result %d\n", rc);
2756         return rc;
2757 }
2758 EXPORT_SYMBOL(ceph_osdc_writepages);
2759
2760 int ceph_osdc_setup(void)
2761 {
2762         BUG_ON(ceph_osd_request_cache);
2763         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2764                                         sizeof (struct ceph_osd_request),
2765                                         __alignof__(struct ceph_osd_request),
2766                                         0, NULL);
2767
2768         return ceph_osd_request_cache ? 0 : -ENOMEM;
2769 }
2770 EXPORT_SYMBOL(ceph_osdc_setup);
2771
2772 void ceph_osdc_cleanup(void)
2773 {
2774         BUG_ON(!ceph_osd_request_cache);
2775         kmem_cache_destroy(ceph_osd_request_cache);
2776         ceph_osd_request_cache = NULL;
2777 }
2778 EXPORT_SYMBOL(ceph_osdc_cleanup);
2779
2780 /*
2781  * handle incoming message
2782  */
2783 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2784 {
2785         struct ceph_osd *osd = con->private;
2786         struct ceph_osd_client *osdc;
2787         int type = le16_to_cpu(msg->hdr.type);
2788
2789         if (!osd)
2790                 goto out;
2791         osdc = osd->o_osdc;
2792
2793         switch (type) {
2794         case CEPH_MSG_OSD_MAP:
2795                 ceph_osdc_handle_map(osdc, msg);
2796                 break;
2797         case CEPH_MSG_OSD_OPREPLY:
2798                 handle_reply(osdc, msg, con);
2799                 break;
2800         case CEPH_MSG_WATCH_NOTIFY:
2801                 handle_watch_notify(osdc, msg);
2802                 break;
2803
2804         default:
2805                 pr_err("received unknown message type %d %s\n", type,
2806                        ceph_msg_type_name(type));
2807         }
2808 out:
2809         ceph_msg_put(msg);
2810 }
2811
2812 /*
2813  * lookup and return message for incoming reply.  set up reply message
2814  * pages.
2815  */
2816 static struct ceph_msg *get_reply(struct ceph_connection *con,
2817                                   struct ceph_msg_header *hdr,
2818                                   int *skip)
2819 {
2820         struct ceph_osd *osd = con->private;
2821         struct ceph_osd_client *osdc = osd->o_osdc;
2822         struct ceph_msg *m;
2823         struct ceph_osd_request *req;
2824         int front_len = le32_to_cpu(hdr->front_len);
2825         int data_len = le32_to_cpu(hdr->data_len);
2826         u64 tid;
2827
2828         tid = le64_to_cpu(hdr->tid);
2829         mutex_lock(&osdc->request_mutex);
2830         req = __lookup_request(osdc, tid);
2831         if (!req) {
2832                 *skip = 1;
2833                 m = NULL;
2834                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2835                      osd->o_osd);
2836                 goto out;
2837         }
2838
2839         if (req->r_reply->con)
2840                 dout("%s revoking msg %p from old con %p\n", __func__,
2841                      req->r_reply, req->r_reply->con);
2842         ceph_msg_revoke_incoming(req->r_reply);
2843
2844         if (front_len > req->r_reply->front_alloc_len) {
2845                 pr_warning("get_reply front %d > preallocated %d (%u#%llu)\n",
2846                            front_len, req->r_reply->front_alloc_len,
2847                            (unsigned int)con->peer_name.type,
2848                            le64_to_cpu(con->peer_name.num));
2849                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2850                                  false);
2851                 if (!m)
2852                         goto out;
2853                 ceph_msg_put(req->r_reply);
2854                 req->r_reply = m;
2855         }
2856         m = ceph_msg_get(req->r_reply);
2857
2858         if (data_len > 0) {
2859                 struct ceph_osd_data *osd_data;
2860
2861                 /*
2862                  * XXX This is assuming there is only one op containing
2863                  * XXX page data.  Probably OK for reads, but this
2864                  * XXX ought to be done more generally.
2865                  */
2866                 osd_data = osd_req_op_extent_osd_data(req, 0);
2867                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2868                         if (osd_data->pages &&
2869                                 unlikely(osd_data->length < data_len)) {
2870
2871                                 pr_warning("tid %lld reply has %d bytes "
2872                                         "we had only %llu bytes ready\n",
2873                                         tid, data_len, osd_data->length);
2874                                 *skip = 1;
2875                                 ceph_msg_put(m);
2876                                 m = NULL;
2877                                 goto out;
2878                         }
2879                 }
2880         }
2881         *skip = 0;
2882         dout("get_reply tid %lld %p\n", tid, m);
2883
2884 out:
2885         mutex_unlock(&osdc->request_mutex);
2886         return m;
2887
2888 }
2889
2890 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2891                                   struct ceph_msg_header *hdr,
2892                                   int *skip)
2893 {
2894         struct ceph_osd *osd = con->private;
2895         int type = le16_to_cpu(hdr->type);
2896         int front = le32_to_cpu(hdr->front_len);
2897
2898         *skip = 0;
2899         switch (type) {
2900         case CEPH_MSG_OSD_MAP:
2901         case CEPH_MSG_WATCH_NOTIFY:
2902                 return ceph_msg_new(type, front, GFP_NOFS, false);
2903         case CEPH_MSG_OSD_OPREPLY:
2904                 return get_reply(con, hdr, skip);
2905         default:
2906                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2907                         osd->o_osd);
2908                 *skip = 1;
2909                 return NULL;
2910         }
2911 }
2912
2913 /*
2914  * Wrappers to refcount containing ceph_osd struct
2915  */
2916 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2917 {
2918         struct ceph_osd *osd = con->private;
2919         if (get_osd(osd))
2920                 return con;
2921         return NULL;
2922 }
2923
2924 static void put_osd_con(struct ceph_connection *con)
2925 {
2926         struct ceph_osd *osd = con->private;
2927         put_osd(osd);
2928 }
2929
2930 /*
2931  * authentication
2932  */
2933 /*
2934  * Note: returned pointer is the address of a structure that's
2935  * managed separately.  Caller must *not* attempt to free it.
2936  */
2937 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2938                                         int *proto, int force_new)
2939 {
2940         struct ceph_osd *o = con->private;
2941         struct ceph_osd_client *osdc = o->o_osdc;
2942         struct ceph_auth_client *ac = osdc->client->monc.auth;
2943         struct ceph_auth_handshake *auth = &o->o_auth;
2944
2945         if (force_new && auth->authorizer) {
2946                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2947                 auth->authorizer = NULL;
2948         }
2949         if (!auth->authorizer) {
2950                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2951                                                       auth);
2952                 if (ret)
2953                         return ERR_PTR(ret);
2954         } else {
2955                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2956                                                      auth);
2957                 if (ret)
2958                         return ERR_PTR(ret);
2959         }
2960         *proto = ac->protocol;
2961
2962         return auth;
2963 }
2964
2965
2966 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2967 {
2968         struct ceph_osd *o = con->private;
2969         struct ceph_osd_client *osdc = o->o_osdc;
2970         struct ceph_auth_client *ac = osdc->client->monc.auth;
2971
2972         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2973 }
2974
2975 static int invalidate_authorizer(struct ceph_connection *con)
2976 {
2977         struct ceph_osd *o = con->private;
2978         struct ceph_osd_client *osdc = o->o_osdc;
2979         struct ceph_auth_client *ac = osdc->client->monc.auth;
2980
2981         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2982         return ceph_monc_validate_auth(&osdc->client->monc);
2983 }
2984
2985 static const struct ceph_connection_operations osd_con_ops = {
2986         .get = get_osd_con,
2987         .put = put_osd_con,
2988         .dispatch = dispatch,
2989         .get_authorizer = get_authorizer,
2990         .verify_authorizer_reply = verify_authorizer_reply,
2991         .invalidate_authorizer = invalidate_authorizer,
2992         .alloc_msg = alloc_msg,
2993         .fault = osd_reset,
2994 };