Pileus Git - ~andy/linux/blob - 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 const struct ceph_connection_operations osd_con_ops;
  26
  27 static void __send_queued(struct ceph_osd_client *osdc);
  28 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  29 static void __register_request(struct ceph_osd_client *osdc,
  30                                struct ceph_osd_request *req);
  31 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  32                                         struct ceph_osd_request *req);
  33 static void __send_request(struct ceph_osd_client *osdc,
  34                            struct ceph_osd_request *req);
  35
  36 /*
  37  * Implement client access to distributed object storage cluster.
  38  *
  39  * All data objects are stored within a cluster/cloud of OSDs, or
  40  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  41  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  42  * remote daemons serving up and coordinating consistent and safe
  43  * access to storage.
  44  *
  45  * Cluster membership and the mapping of data objects onto storage devices
  46  * are described by the osd map.
  47  *
  48  * We keep track of pending OSD requests (read, write), resubmit
  49  * requests to different OSDs when the cluster topology/data layout
  50  * change, or retry the affected requests when the communications
  51  * channel with an OSD is reset.
  52  */
  53
  54 /*
  55  * calculate the mapping of a file extent onto an object, and fill out the
  56  * request accordingly.  shorten extent as necessary if it crosses an
  57  * object boundary.
  58  *
  59  * fill osd op in request message.
  60  */
  61 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
  62                         u64 *objnum, u64 *objoff, u64 *objlen)
  63 {
  64         u64 orig_len = *plen;
  65         int r;
  66
  67         /* object extent? */
  68         r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
  69                                           objoff, objlen);
  70         if (r < 0)
  71                 return r;
  72         if (*objlen < orig_len) {
  73                 *plen = *objlen;
  74                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  75                      orig_len - *plen, off, *plen);
  76         }
  77
  78         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
  79
  80         return 0;
  81 }
  82
  83 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
  84 {
  85         memset(osd_data, 0, sizeof (*osd_data));
  86         osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
  87 }
  88
  89 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
  90                         struct page **pages, u64 length, u32 alignment,
  91                         bool pages_from_pool, bool own_pages)
  92 {
  93         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
  94         osd_data->pages = pages;
  95         osd_data->length = length;
  96         osd_data->alignment = alignment;
  97         osd_data->pages_from_pool = pages_from_pool;
  98         osd_data->own_pages = own_pages;
  99 }
 100
 101 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
 102                         struct ceph_pagelist *pagelist)
 103 {
 104         osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
 105         osd_data->pagelist = pagelist;
 106 }
 107
 108 #ifdef CONFIG_BLOCK
 109 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
 110                         struct bio *bio, size_t bio_length)
 111 {
 112         osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
 113         osd_data->bio = bio;
 114         osd_data->bio_length = bio_length;
 115 }
 116 #endif /* CONFIG_BLOCK */
 117
 118 struct ceph_osd_data *
 119 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
 120                         unsigned int which, bool write_request)
 121 {
 122         BUG_ON(which >= osd_req->r_num_ops);
 123
 124         return &osd_req->r_ops[which].extent.osd_data;
 125 }
 126 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
 127
 128 struct ceph_osd_data *
 129 osd_req_op_cls_request_info(struct ceph_osd_request *osd_req,
 130                         unsigned int which)
 131 {
 132         BUG_ON(which >= osd_req->r_num_ops);
 133
 134         return &osd_req->r_ops[which].cls.request_info;
 135 }
 136 EXPORT_SYMBOL(osd_req_op_cls_request_info);     /* ??? */
 137
 138 struct ceph_osd_data *
 139 osd_req_op_cls_request_data(struct ceph_osd_request *osd_req,
 140                         unsigned int which)
 141 {
 142         BUG_ON(which >= osd_req->r_num_ops);
 143
 144         return &osd_req->r_ops[which].cls.request_data;
 145 }
 146 EXPORT_SYMBOL(osd_req_op_cls_request_data);     /* ??? */
 147
 148 struct ceph_osd_data *
 149 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
 150                         unsigned int which)
 151 {
 152         BUG_ON(which >= osd_req->r_num_ops);
 153
 154         return &osd_req->r_ops[which].cls.response_data;
 155 }
 156 EXPORT_SYMBOL(osd_req_op_cls_response_data);    /* ??? */
 157
 158 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
 159                         unsigned int which, bool write_request,
 160                         struct page **pages, u64 length, u32 alignment,
 161                         bool pages_from_pool, bool own_pages)
 162 {
 163         struct ceph_osd_data *osd_data;
 164
 165         osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request);
 166         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 167                                 pages_from_pool, own_pages);
 168 }
 169 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
 170
 171 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
 172                         unsigned int which, bool write_request,
 173                         struct ceph_pagelist *pagelist)
 174 {
 175         struct ceph_osd_data *osd_data;
 176
 177         osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request);
 178         ceph_osd_data_pagelist_init(osd_data, pagelist);
 179 }
 180 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
 181
 182 #ifdef CONFIG_BLOCK
 183 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
 184                         unsigned int which, bool write_request,
 185                         struct bio *bio, size_t bio_length)
 186 {
 187         struct ceph_osd_data *osd_data;
 188
 189         osd_data = osd_req_op_extent_osd_data(osd_req, which, write_request);
 190         ceph_osd_data_bio_init(osd_data, bio, bio_length);
 191 }
 192 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
 193 #endif /* CONFIG_BLOCK */
 194
 195 static void osd_req_op_cls_request_info_pagelist(
 196                         struct ceph_osd_request *osd_req,
 197                         unsigned int which, struct ceph_pagelist *pagelist)
 198 {
 199         struct ceph_osd_data *osd_data;
 200
 201         osd_data = osd_req_op_cls_request_info(osd_req, which);
 202         ceph_osd_data_pagelist_init(osd_data, pagelist);
 203 }
 204
 205 void osd_req_op_cls_request_data_pagelist(
 206                         struct ceph_osd_request *osd_req,
 207                         unsigned int which, struct ceph_pagelist *pagelist)
 208 {
 209         struct ceph_osd_data *osd_data;
 210
 211         osd_data = osd_req_op_cls_request_data(osd_req, which);
 212         ceph_osd_data_pagelist_init(osd_data, pagelist);
 213 }
 214 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
 215
 216 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
 217                         unsigned int which, struct page **pages, u64 length,
 218                         u32 alignment, bool pages_from_pool, bool own_pages)
 219 {
 220         struct ceph_osd_data *osd_data;
 221
 222         osd_data = osd_req_op_cls_response_data(osd_req, which);
 223         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 224                                 pages_from_pool, own_pages);
 225 }
 226 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
 227
 228 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
 229 {
 230         switch (osd_data->type) {
 231         case CEPH_OSD_DATA_TYPE_NONE:
 232                 return 0;
 233         case CEPH_OSD_DATA_TYPE_PAGES:
 234                 return osd_data->length;
 235         case CEPH_OSD_DATA_TYPE_PAGELIST:
 236                 return (u64)osd_data->pagelist->length;
 237 #ifdef CONFIG_BLOCK
 238         case CEPH_OSD_DATA_TYPE_BIO:
 239                 return (u64)osd_data->bio_length;
 240 #endif /* CONFIG_BLOCK */
 241         default:
 242                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
 243                 return 0;
 244         }
 245 }
 246
 247
 248 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
 249 {
 250         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
 251                 int num_pages;
 252
 253                 num_pages = calc_pages_for((u64)osd_data->alignment,
 254                                                 (u64)osd_data->length);
 255                 ceph_release_page_vector(osd_data->pages, num_pages);
 256         }
 257         ceph_osd_data_init(osd_data);
 258 }
 259
 260 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
 261                         unsigned int which)
 262 {
 263         struct ceph_osd_req_op *op;
 264
 265         BUG_ON(which >= osd_req->r_num_ops);
 266         op = &osd_req->r_ops[which];
 267
 268         switch (op->op) {
 269         case CEPH_OSD_OP_READ:
 270         case CEPH_OSD_OP_WRITE:
 271                 ceph_osd_data_release(&op->extent.osd_data);
 272                 break;
 273         case CEPH_OSD_OP_CALL:
 274                 ceph_osd_data_release(&op->cls.request_info);
 275                 ceph_osd_data_release(&op->cls.request_data);
 276                 ceph_osd_data_release(&op->cls.response_data);
 277                 break;
 278         default:
 279                 break;
 280         }
 281 }
 282
 283 /*
 284  * requests
 285  */
 286 void ceph_osdc_release_request(struct kref *kref)
 287 {
 288         struct ceph_osd_request *req;
 289         unsigned int which;
 290
 291         req = container_of(kref, struct ceph_osd_request, r_kref);
 292         if (req->r_request)
 293                 ceph_msg_put(req->r_request);
 294         if (req->r_reply) {
 295                 ceph_msg_revoke_incoming(req->r_reply);
 296                 ceph_msg_put(req->r_reply);
 297         }
 298
 299         for (which = 0; which < req->r_num_ops; which++)
 300                 osd_req_op_data_release(req, which);
 301
 302         ceph_put_snap_context(req->r_snapc);
 303         if (req->r_mempool)
 304                 mempool_free(req, req->r_osdc->req_mempool);
 305         else
 306                 kfree(req);
 307 }
 308 EXPORT_SYMBOL(ceph_osdc_release_request);
 309
 310 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 311                                                struct ceph_snap_context *snapc,
 312                                                unsigned int num_ops,
 313                                                bool use_mempool,
 314                                                gfp_t gfp_flags)
 315 {
 316         struct ceph_osd_request *req;
 317         struct ceph_msg *msg;
 318         size_t msg_size;
 319
 320         BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
 321         BUG_ON(num_ops > CEPH_OSD_MAX_OP);
 322
 323         msg_size = 4 + 4 + 8 + 8 + 4+8;
 324         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
 325         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
 326         msg_size += 4 + MAX_OBJ_NAME_SIZE;
 327         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
 328         msg_size += 8;  /* snapid */
 329         msg_size += 8;  /* snap_seq */
 330         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
 331         msg_size += 4;
 332
 333         if (use_mempool) {
 334                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 335                 memset(req, 0, sizeof(*req));
 336         } else {
 337                 req = kzalloc(sizeof(*req), gfp_flags);
 338         }
 339         if (req == NULL)
 340                 return NULL;
 341
 342         req->r_osdc = osdc;
 343         req->r_mempool = use_mempool;
 344         req->r_num_ops = num_ops;
 345
 346         kref_init(&req->r_kref);
 347         init_completion(&req->r_completion);
 348         init_completion(&req->r_safe_completion);
 349         RB_CLEAR_NODE(&req->r_node);
 350         INIT_LIST_HEAD(&req->r_unsafe_item);
 351         INIT_LIST_HEAD(&req->r_linger_item);
 352         INIT_LIST_HEAD(&req->r_linger_osd);
 353         INIT_LIST_HEAD(&req->r_req_lru_item);
 354         INIT_LIST_HEAD(&req->r_osd_item);
 355
 356         /* create reply message */
 357         if (use_mempool)
 358                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 359         else
 360                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 361                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 362         if (!msg) {
 363                 ceph_osdc_put_request(req);
 364                 return NULL;
 365         }
 366         req->r_reply = msg;
 367
 368         /* create request message; allow space for oid */
 369         if (use_mempool)
 370                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 371         else
 372                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 373         if (!msg) {
 374                 ceph_osdc_put_request(req);
 375                 return NULL;
 376         }
 377
 378         memset(msg->front.iov_base, 0, msg->front.iov_len);
 379
 380         req->r_request = msg;
 381
 382         return req;
 383 }
 384 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 385
 386 static bool osd_req_opcode_valid(u16 opcode)
 387 {
 388         switch (opcode) {
 389         case CEPH_OSD_OP_READ:
 390         case CEPH_OSD_OP_STAT:
 391         case CEPH_OSD_OP_MAPEXT:
 392         case CEPH_OSD_OP_MASKTRUNC:
 393         case CEPH_OSD_OP_SPARSE_READ:
 394         case CEPH_OSD_OP_NOTIFY:
 395         case CEPH_OSD_OP_NOTIFY_ACK:
 396         case CEPH_OSD_OP_ASSERT_VER:
 397         case CEPH_OSD_OP_WRITE:
 398         case CEPH_OSD_OP_WRITEFULL:
 399         case CEPH_OSD_OP_TRUNCATE:
 400         case CEPH_OSD_OP_ZERO:
 401         case CEPH_OSD_OP_DELETE:
 402         case CEPH_OSD_OP_APPEND:
 403         case CEPH_OSD_OP_STARTSYNC:
 404         case CEPH_OSD_OP_SETTRUNC:
 405         case CEPH_OSD_OP_TRIMTRUNC:
 406         case CEPH_OSD_OP_TMAPUP:
 407         case CEPH_OSD_OP_TMAPPUT:
 408         case CEPH_OSD_OP_TMAPGET:
 409         case CEPH_OSD_OP_CREATE:
 410         case CEPH_OSD_OP_ROLLBACK:
 411         case CEPH_OSD_OP_WATCH:
 412         case CEPH_OSD_OP_OMAPGETKEYS:
 413         case CEPH_OSD_OP_OMAPGETVALS:
 414         case CEPH_OSD_OP_OMAPGETHEADER:
 415         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
 416         case CEPH_OSD_OP_OMAPSETVALS:
 417         case CEPH_OSD_OP_OMAPSETHEADER:
 418         case CEPH_OSD_OP_OMAPCLEAR:
 419         case CEPH_OSD_OP_OMAPRMKEYS:
 420         case CEPH_OSD_OP_OMAP_CMP:
 421         case CEPH_OSD_OP_CLONERANGE:
 422         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
 423         case CEPH_OSD_OP_SRC_CMPXATTR:
 424         case CEPH_OSD_OP_GETXATTR:
 425         case CEPH_OSD_OP_GETXATTRS:
 426         case CEPH_OSD_OP_CMPXATTR:
 427         case CEPH_OSD_OP_SETXATTR:
 428         case CEPH_OSD_OP_SETXATTRS:
 429         case CEPH_OSD_OP_RESETXATTRS:
 430         case CEPH_OSD_OP_RMXATTR:
 431         case CEPH_OSD_OP_PULL:
 432         case CEPH_OSD_OP_PUSH:
 433         case CEPH_OSD_OP_BALANCEREADS:
 434         case CEPH_OSD_OP_UNBALANCEREADS:
 435         case CEPH_OSD_OP_SCRUB:
 436         case CEPH_OSD_OP_SCRUB_RESERVE:
 437         case CEPH_OSD_OP_SCRUB_UNRESERVE:
 438         case CEPH_OSD_OP_SCRUB_STOP:
 439         case CEPH_OSD_OP_SCRUB_MAP:
 440         case CEPH_OSD_OP_WRLOCK:
 441         case CEPH_OSD_OP_WRUNLOCK:
 442         case CEPH_OSD_OP_RDLOCK:
 443         case CEPH_OSD_OP_RDUNLOCK:
 444         case CEPH_OSD_OP_UPLOCK:
 445         case CEPH_OSD_OP_DNLOCK:
 446         case CEPH_OSD_OP_CALL:
 447         case CEPH_OSD_OP_PGLS:
 448         case CEPH_OSD_OP_PGLS_FILTER:
 449                 return true;
 450         default:
 451                 return false;
 452         }
 453 }
 454
 455 /*
 456  * This is an osd op init function for opcodes that have no data or
 457  * other information associated with them.  It also serves as a
 458  * common init routine for all the other init functions, below.
 459  */
 460 static struct ceph_osd_req_op *
 461 osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
 462                                 u16 opcode)
 463 {
 464         struct ceph_osd_req_op *op;
 465
 466         BUG_ON(which >= osd_req->r_num_ops);
 467         BUG_ON(!osd_req_opcode_valid(opcode));
 468
 469         op = &osd_req->r_ops[which];
 470         memset(op, 0, sizeof (*op));
 471         op->op = opcode;
 472
 473         return op;
 474 }
 475
 476 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 477                                 unsigned int which, u16 opcode,
 478                                 u64 offset, u64 length,
 479                                 u64 truncate_size, u32 truncate_seq)
 480 {
 481         struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode);
 482         size_t payload_len = 0;
 483
 484         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE);
 485
 486         op->extent.offset = offset;
 487         op->extent.length = length;
 488         op->extent.truncate_size = truncate_size;
 489         op->extent.truncate_seq = truncate_seq;
 490         if (opcode == CEPH_OSD_OP_WRITE)
 491                 payload_len += length;
 492
 493         op->payload_len = payload_len;
 494 }
 495 EXPORT_SYMBOL(osd_req_op_extent_init);
 496
 497 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 498                                 unsigned int which, u64 length)
 499 {
 500         struct ceph_osd_req_op *op;
 501         u64 previous;
 502
 503         BUG_ON(which >= osd_req->r_num_ops);
 504         op = &osd_req->r_ops[which];
 505         previous = op->extent.length;
 506
 507         if (length == previous)
 508                 return;         /* Nothing to do */
 509         BUG_ON(length > previous);
 510
 511         op->extent.length = length;
 512         op->payload_len -= previous - length;
 513 }
 514 EXPORT_SYMBOL(osd_req_op_extent_update);
 515
 516 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
 517                         u16 opcode, const char *class, const char *method)
 518 {
 519         struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode);
 520         struct ceph_pagelist *pagelist;
 521         size_t payload_len = 0;
 522         size_t size;
 523
 524         BUG_ON(opcode != CEPH_OSD_OP_CALL);
 525
 526         pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
 527         BUG_ON(!pagelist);
 528         ceph_pagelist_init(pagelist);
 529
 530         op->cls.class_name = class;
 531         size = strlen(class);
 532         BUG_ON(size > (size_t) U8_MAX);
 533         op->cls.class_len = size;
 534         ceph_pagelist_append(pagelist, class, size);
 535         payload_len += size;
 536
 537         op->cls.method_name = method;
 538         size = strlen(method);
 539         BUG_ON(size > (size_t) U8_MAX);
 540         op->cls.method_len = size;
 541         ceph_pagelist_append(pagelist, method, size);
 542         payload_len += size;
 543
 544         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
 545
 546         op->cls.argc = 0;       /* currently unused */
 547
 548         op->payload_len = payload_len;
 549 }
 550 EXPORT_SYMBOL(osd_req_op_cls_init);
 551
 552 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
 553                                 unsigned int which, u16 opcode,
 554                                 u64 cookie, u64 version, int flag)
 555 {
 556         struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, opcode);
 557
 558         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
 559
 560         op->watch.cookie = cookie;
 561         /* op->watch.ver = version; */  /* XXX 3847 */
 562         op->watch.ver = cpu_to_le64(version);
 563         if (opcode == CEPH_OSD_OP_WATCH && flag)
 564                 op->watch.flag = (u8)1;
 565 }
 566 EXPORT_SYMBOL(osd_req_op_watch_init);
 567
 568 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
 569                                 struct ceph_osd_data *osd_data)
 570 {
 571         u64 length = ceph_osd_data_length(osd_data);
 572
 573         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 574                 BUG_ON(length > (u64) SIZE_MAX);
 575                 if (length)
 576                         ceph_msg_data_add_pages(msg, osd_data->pages,
 577                                         length, osd_data->alignment);
 578         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 579                 BUG_ON(!length);
 580                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
 581 #ifdef CONFIG_BLOCK
 582         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
 583                 ceph_msg_data_add_bio(msg, osd_data->bio, length);
 584 #endif
 585         } else {
 586                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
 587         }
 588 }
 589
 590 static u64 osd_req_encode_op(struct ceph_osd_request *req,
 591                               struct ceph_osd_op *dst, unsigned int which)
 592 {
 593         struct ceph_osd_req_op *src;
 594         struct ceph_osd_data *osd_data;
 595         u64 request_data_len = 0;
 596         u64 data_length;
 597
 598         BUG_ON(which >= req->r_num_ops);
 599         src = &req->r_ops[which];
 600         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
 601                 pr_err("unrecognized osd opcode %d\n", src->op);
 602
 603                 return 0;
 604         }
 605
 606         switch (src->op) {
 607         case CEPH_OSD_OP_STAT:
 608                 break;
 609         case CEPH_OSD_OP_READ:
 610         case CEPH_OSD_OP_WRITE:
 611                 if (src->op == CEPH_OSD_OP_WRITE)
 612                         request_data_len = src->extent.length;
 613                 dst->extent.offset = cpu_to_le64(src->extent.offset);
 614                 dst->extent.length = cpu_to_le64(src->extent.length);
 615                 dst->extent.truncate_size =
 616                         cpu_to_le64(src->extent.truncate_size);
 617                 dst->extent.truncate_seq =
 618                         cpu_to_le32(src->extent.truncate_seq);
 619                 osd_data = &src->extent.osd_data;
 620                 if (src->op == CEPH_OSD_OP_WRITE)
 621                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 622                 else
 623                         ceph_osdc_msg_data_add(req->r_reply, osd_data);
 624                 break;
 625         case CEPH_OSD_OP_CALL:
 626                 dst->cls.class_len = src->cls.class_len;
 627                 dst->cls.method_len = src->cls.method_len;
 628                 osd_data = &src->cls.request_info;
 629                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 630                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
 631                 request_data_len = osd_data->pagelist->length;
 632
 633                 osd_data = &src->cls.request_data;
 634                 data_length = ceph_osd_data_length(osd_data);
 635                 if (data_length) {
 636                         BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
 637                         dst->cls.indata_len = cpu_to_le32(data_length);
 638                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 639                         src->payload_len += data_length;
 640                         request_data_len += data_length;
 641                 }
 642                 osd_data = &src->cls.response_data;
 643                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 644                 break;
 645         case CEPH_OSD_OP_STARTSYNC:
 646                 break;
 647         case CEPH_OSD_OP_NOTIFY_ACK:
 648         case CEPH_OSD_OP_WATCH:
 649                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 650                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 651                 dst->watch.flag = src->watch.flag;
 652                 break;
 653         default:
 654                 pr_err("unsupported osd opcode %s\n",
 655                         ceph_osd_op_name(src->op));
 656                 WARN_ON(1);
 657
 658                 return 0;
 659         }
 660         dst->op = cpu_to_le16(src->op);
 661         dst->payload_len = cpu_to_le32(src->payload_len);
 662
 663         return request_data_len;
 664 }
 665
 666 /*
 667  * build new request AND message, calculate layout, and adjust file
 668  * extent as needed.
 669  *
 670  * if the file was recently truncated, we include information about its
 671  * old and new size so that the object can be updated appropriately.  (we
 672  * avoid synchronously deleting truncated objects because it's slow.)
 673  *
 674  * if @do_sync, include a 'startsync' command so that the osd will flush
 675  * data quickly.
 676  */
 677 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 678                                                struct ceph_file_layout *layout,
 679                                                struct ceph_vino vino,
 680                                                u64 off, u64 *plen, int num_ops,
 681                                                int opcode, int flags,
 682                                                struct ceph_snap_context *snapc,
 683                                                u32 truncate_seq,
 684                                                u64 truncate_size,
 685                                                bool use_mempool)
 686 {
 687         struct ceph_osd_request *req;
 688         u64 objnum = 0;
 689         u64 objoff = 0;
 690         u64 objlen = 0;
 691         u32 object_size;
 692         u64 object_base;
 693         int r;
 694
 695         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE);
 696
 697         req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
 698                                         GFP_NOFS);
 699         if (!req)
 700                 return ERR_PTR(-ENOMEM);
 701
 702         req->r_flags = flags;
 703
 704         /* calculate max write size */
 705         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
 706         if (r < 0) {
 707                 ceph_osdc_put_request(req);
 708                 return ERR_PTR(r);
 709         }
 710
 711         object_size = le32_to_cpu(layout->fl_object_size);
 712         object_base = off - objoff;
 713         if (truncate_size <= object_base) {
 714                 truncate_size = 0;
 715         } else {
 716                 truncate_size -= object_base;
 717                 if (truncate_size > object_size)
 718                         truncate_size = object_size;
 719         }
 720
 721         osd_req_op_extent_init(req, 0, opcode, objoff, objlen,
 722                                 truncate_size, truncate_seq);
 723
 724         /*
 725          * A second op in the ops array means the caller wants to
 726          * also issue a include a 'startsync' command so that the
 727          * osd will flush data quickly.
 728          */
 729         if (num_ops > 1)
 730                 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC);
 731
 732         req->r_file_layout = *layout;  /* keep a copy */
 733
 734         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx",
 735                 vino.ino, objnum);
 736         req->r_oid_len = strlen(req->r_oid);
 737
 738         return req;
 739 }
 740 EXPORT_SYMBOL(ceph_osdc_new_request);
 741
 742 /*
 743  * We keep osd requests in an rbtree, sorted by ->r_tid.
 744  */
 745 static void __insert_request(struct ceph_osd_client *osdc,
 746                              struct ceph_osd_request *new)
 747 {
 748         struct rb_node **p = &osdc->requests.rb_node;
 749         struct rb_node *parent = NULL;
 750         struct ceph_osd_request *req = NULL;
 751
 752         while (*p) {
 753                 parent = *p;
 754                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 755                 if (new->r_tid < req->r_tid)
 756                         p = &(*p)->rb_left;
 757                 else if (new->r_tid > req->r_tid)
 758                         p = &(*p)->rb_right;
 759                 else
 760                         BUG();
 761         }
 762
 763         rb_link_node(&new->r_node, parent, p);
 764         rb_insert_color(&new->r_node, &osdc->requests);
 765 }
 766
 767 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 768                                                  u64 tid)
 769 {
 770         struct ceph_osd_request *req;
 771         struct rb_node *n = osdc->requests.rb_node;
 772
 773         while (n) {
 774                 req = rb_entry(n, struct ceph_osd_request, r_node);
 775                 if (tid < req->r_tid)
 776                         n = n->rb_left;
 777                 else if (tid > req->r_tid)
 778                         n = n->rb_right;
 779                 else
 780                         return req;
 781         }
 782         return NULL;
 783 }
 784
 785 static struct ceph_osd_request *
 786 __lookup_request_ge(struct ceph_osd_client *osdc,
 787                     u64 tid)
 788 {
 789         struct ceph_osd_request *req;
 790         struct rb_node *n = osdc->requests.rb_node;
 791
 792         while (n) {
 793                 req = rb_entry(n, struct ceph_osd_request, r_node);
 794                 if (tid < req->r_tid) {
 795                         if (!n->rb_left)
 796                                 return req;
 797                         n = n->rb_left;
 798                 } else if (tid > req->r_tid) {
 799                         n = n->rb_right;
 800                 } else {
 801                         return req;
 802                 }
 803         }
 804         return NULL;
 805 }
 806
 807 /*
 808  * Resubmit requests pending on the given osd.
 809  */
 810 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 811                                 struct ceph_osd *osd)
 812 {
 813         struct ceph_osd_request *req, *nreq;
 814         LIST_HEAD(resend);
 815         int err;
 816
 817         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 818         err = __reset_osd(osdc, osd);
 819         if (err)
 820                 return;
 821         /*
 822          * Build up a list of requests to resend by traversing the
 823          * osd's list of requests.  Requests for a given object are
 824          * sent in tid order, and that is also the order they're
 825          * kept on this list.  Therefore all requests that are in
 826          * flight will be found first, followed by all requests that
 827          * have not yet been sent.  And to resend requests while
 828          * preserving this order we will want to put any sent
 829          * requests back on the front of the osd client's unsent
 830          * list.
 831          *
 832          * So we build a separate ordered list of already-sent
 833          * requests for the affected osd and splice it onto the
 834          * front of the osd client's unsent list.  Once we've seen a
 835          * request that has not yet been sent we're done.  Those
 836          * requests are already sitting right where they belong.
 837          */
 838         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 839                 if (!req->r_sent)
 840                         break;
 841                 list_move_tail(&req->r_req_lru_item, &resend);
 842                 dout("requeueing %p tid %llu osd%d\n", req, req->r_tid,
 843                      osd->o_osd);
 844                 if (!req->r_linger)
 845                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 846         }
 847         list_splice(&resend, &osdc->req_unsent);
 848
 849         /*
 850          * Linger requests are re-registered before sending, which
 851          * sets up a new tid for each.  We add them to the unsent
 852          * list at the end to keep things in tid order.
 853          */
 854         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 855                                  r_linger_osd) {
 856                 /*
 857                  * reregister request prior to unregistering linger so
 858                  * that r_osd is preserved.
 859                  */
 860                 BUG_ON(!list_empty(&req->r_req_lru_item));
 861                 __register_request(osdc, req);
 862                 list_add_tail(&req->r_req_lru_item, &osdc->req_unsent);
 863                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
 864                 __unregister_linger_request(osdc, req);
 865                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 866                      osd->o_osd);
 867         }
 868 }
 869
 870 /*
 871  * If the osd connection drops, we need to resubmit all requests.
 872  */
 873 static void osd_reset(struct ceph_connection *con)
 874 {
 875         struct ceph_osd *osd = con->private;
 876         struct ceph_osd_client *osdc;
 877
 878         if (!osd)
 879                 return;
 880         dout("osd_reset osd%d\n", osd->o_osd);
 881         osdc = osd->o_osdc;
 882         down_read(&osdc->map_sem);
 883         mutex_lock(&osdc->request_mutex);
 884         __kick_osd_requests(osdc, osd);
 885         __send_queued(osdc);
 886         mutex_unlock(&osdc->request_mutex);
 887         up_read(&osdc->map_sem);
 888 }
 889
 890 /*
 891  * Track open sessions with osds.
 892  */
 893 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 894 {
 895         struct ceph_osd *osd;
 896
 897         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 898         if (!osd)
 899                 return NULL;
 900
 901         atomic_set(&osd->o_ref, 1);
 902         osd->o_osdc = osdc;
 903         osd->o_osd = onum;
 904         RB_CLEAR_NODE(&osd->o_node);
 905         INIT_LIST_HEAD(&osd->o_requests);
 906         INIT_LIST_HEAD(&osd->o_linger_requests);
 907         INIT_LIST_HEAD(&osd->o_osd_lru);
 908         osd->o_incarnation = 1;
 909
 910         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 911
 912         INIT_LIST_HEAD(&osd->o_keepalive_item);
 913         return osd;
 914 }
 915
 916 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 917 {
 918         if (atomic_inc_not_zero(&osd->o_ref)) {
 919                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 920                      atomic_read(&osd->o_ref));
 921                 return osd;
 922         } else {
 923                 dout("get_osd %p FAIL\n", osd);
 924                 return NULL;
 925         }
 926 }
 927
 928 static void put_osd(struct ceph_osd *osd)
 929 {
 930         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 931              atomic_read(&osd->o_ref) - 1);
 932         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 933                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 934
 935                 ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
 936                 kfree(osd);
 937         }
 938 }
 939
 940 /*
 941  * remove an osd from our map
 942  */
 943 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 944 {
 945         dout("__remove_osd %p\n", osd);
 946         BUG_ON(!list_empty(&osd->o_requests));
 947         rb_erase(&osd->o_node, &osdc->osds);
 948         list_del_init(&osd->o_osd_lru);
 949         ceph_con_close(&osd->o_con);
 950         put_osd(osd);
 951 }
 952
 953 static void remove_all_osds(struct ceph_osd_client *osdc)
 954 {
 955         dout("%s %p\n", __func__, osdc);
 956         mutex_lock(&osdc->request_mutex);
 957         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 958                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 959                                                 struct ceph_osd, o_node);
 960                 __remove_osd(osdc, osd);
 961         }
 962         mutex_unlock(&osdc->request_mutex);
 963 }
 964
 965 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 966                               struct ceph_osd *osd)
 967 {
 968         dout("__move_osd_to_lru %p\n", osd);
 969         BUG_ON(!list_empty(&osd->o_osd_lru));
 970         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 971         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 972 }
 973
 974 static void __remove_osd_from_lru(struct ceph_osd *osd)
 975 {
 976         dout("__remove_osd_from_lru %p\n", osd);
 977         if (!list_empty(&osd->o_osd_lru))
 978                 list_del_init(&osd->o_osd_lru);
 979 }
 980
 981 static void remove_old_osds(struct ceph_osd_client *osdc)
 982 {
 983         struct ceph_osd *osd, *nosd;
 984
 985         dout("__remove_old_osds %p\n", osdc);
 986         mutex_lock(&osdc->request_mutex);
 987         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 988                 if (time_before(jiffies, osd->lru_ttl))
 989                         break;
 990                 __remove_osd(osdc, osd);
 991         }
 992         mutex_unlock(&osdc->request_mutex);
 993 }
 994
 995 /*
 996  * reset osd connect
 997  */
 998 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 999 {
1000         struct ceph_entity_addr *peer_addr;
1001
1002         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1003         if (list_empty(&osd->o_requests) &&
1004             list_empty(&osd->o_linger_requests)) {
1005                 __remove_osd(osdc, osd);
1006
1007                 return -ENODEV;
1008         }
1009
1010         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1011         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1012                         !ceph_con_opened(&osd->o_con)) {
1013                 struct ceph_osd_request *req;
1014
1015                 dout(" osd addr hasn't changed and connection never opened,"
1016                      " letting msgr retry");
1017                 /* touch each r_stamp for handle_timeout()'s benfit */
1018                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1019                         req->r_stamp = jiffies;
1020
1021                 return -EAGAIN;
1022         }
1023
1024         ceph_con_close(&osd->o_con);
1025         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1026         osd->o_incarnation++;
1027
1028         return 0;
1029 }
1030
1031 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1032 {
1033         struct rb_node **p = &osdc->osds.rb_node;
1034         struct rb_node *parent = NULL;
1035         struct ceph_osd *osd = NULL;
1036
1037         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1038         while (*p) {
1039                 parent = *p;
1040                 osd = rb_entry(parent, struct ceph_osd, o_node);
1041                 if (new->o_osd < osd->o_osd)
1042                         p = &(*p)->rb_left;
1043                 else if (new->o_osd > osd->o_osd)
1044                         p = &(*p)->rb_right;
1045                 else
1046                         BUG();
1047         }
1048
1049         rb_link_node(&new->o_node, parent, p);
1050         rb_insert_color(&new->o_node, &osdc->osds);
1051 }
1052
1053 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1054 {
1055         struct ceph_osd *osd;
1056         struct rb_node *n = osdc->osds.rb_node;
1057
1058         while (n) {
1059                 osd = rb_entry(n, struct ceph_osd, o_node);
1060                 if (o < osd->o_osd)
1061                         n = n->rb_left;
1062                 else if (o > osd->o_osd)
1063                         n = n->rb_right;
1064                 else
1065                         return osd;
1066         }
1067         return NULL;
1068 }
1069
1070 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1071 {
1072         schedule_delayed_work(&osdc->timeout_work,
1073                         osdc->client->options->osd_keepalive_timeout * HZ);
1074 }
1075
1076 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1077 {
1078         cancel_delayed_work(&osdc->timeout_work);
1079 }
1080
1081 /*
1082  * Register request, assign tid.  If this is the first request, set up
1083  * the timeout event.
1084  */
1085 static void __register_request(struct ceph_osd_client *osdc,
1086                                struct ceph_osd_request *req)
1087 {
1088         req->r_tid = ++osdc->last_tid;
1089         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1090         dout("__register_request %p tid %lld\n", req, req->r_tid);
1091         __insert_request(osdc, req);
1092         ceph_osdc_get_request(req);
1093         osdc->num_requests++;
1094         if (osdc->num_requests == 1) {
1095                 dout(" first request, scheduling timeout\n");
1096                 __schedule_osd_timeout(osdc);
1097         }
1098 }
1099
1100 /*
1101  * called under osdc->request_mutex
1102  */
1103 static void __unregister_request(struct ceph_osd_client *osdc,
1104                                  struct ceph_osd_request *req)
1105 {
1106         if (RB_EMPTY_NODE(&req->r_node)) {
1107                 dout("__unregister_request %p tid %lld not registered\n",
1108                         req, req->r_tid);
1109                 return;
1110         }
1111
1112         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1113         rb_erase(&req->r_node, &osdc->requests);
1114         osdc->num_requests--;
1115
1116         if (req->r_osd) {
1117                 /* make sure the original request isn't in flight. */
1118                 ceph_msg_revoke(req->r_request);
1119
1120                 list_del_init(&req->r_osd_item);
1121                 if (list_empty(&req->r_osd->o_requests) &&
1122                     list_empty(&req->r_osd->o_linger_requests)) {
1123                         dout("moving osd to %p lru\n", req->r_osd);
1124                         __move_osd_to_lru(osdc, req->r_osd);
1125                 }
1126                 if (list_empty(&req->r_linger_item))
1127                         req->r_osd = NULL;
1128         }
1129
1130         list_del_init(&req->r_req_lru_item);
1131         ceph_osdc_put_request(req);
1132
1133         if (osdc->num_requests == 0) {
1134                 dout(" no requests, canceling timeout\n");
1135                 __cancel_osd_timeout(osdc);
1136         }
1137 }
1138
1139 /*
1140  * Cancel a previously queued request message
1141  */
1142 static void __cancel_request(struct ceph_osd_request *req)
1143 {
1144         if (req->r_sent && req->r_osd) {
1145                 ceph_msg_revoke(req->r_request);
1146                 req->r_sent = 0;
1147         }
1148 }
1149
1150 static void __register_linger_request(struct ceph_osd_client *osdc,
1151                                     struct ceph_osd_request *req)
1152 {
1153         dout("__register_linger_request %p\n", req);
1154         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1155         if (req->r_osd)
1156                 list_add_tail(&req->r_linger_osd,
1157                               &req->r_osd->o_linger_requests);
1158 }
1159
1160 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1161                                         struct ceph_osd_request *req)
1162 {
1163         dout("__unregister_linger_request %p\n", req);
1164         list_del_init(&req->r_linger_item);
1165         if (req->r_osd) {
1166                 list_del_init(&req->r_linger_osd);
1167
1168                 if (list_empty(&req->r_osd->o_requests) &&
1169                     list_empty(&req->r_osd->o_linger_requests)) {
1170                         dout("moving osd to %p lru\n", req->r_osd);
1171                         __move_osd_to_lru(osdc, req->r_osd);
1172                 }
1173                 if (list_empty(&req->r_osd_item))
1174                         req->r_osd = NULL;
1175         }
1176 }
1177
1178 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
1179                                          struct ceph_osd_request *req)
1180 {
1181         mutex_lock(&osdc->request_mutex);
1182         if (req->r_linger) {
1183                 __unregister_linger_request(osdc, req);
1184                 ceph_osdc_put_request(req);
1185         }
1186         mutex_unlock(&osdc->request_mutex);
1187 }
1188 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
1189
1190 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1191                                   struct ceph_osd_request *req)
1192 {
1193         if (!req->r_linger) {
1194                 dout("set_request_linger %p\n", req);
1195                 req->r_linger = 1;
1196                 /*
1197                  * caller is now responsible for calling
1198                  * unregister_linger_request
1199                  */
1200                 ceph_osdc_get_request(req);
1201         }
1202 }
1203 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1204
1205 /*
1206  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1207  * (as needed), and set the request r_osd appropriately.  If there is
1208  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1209  * (unsent, homeless) or leave on in-flight lru.
1210  *
1211  * Return 0 if unchanged, 1 if changed, or negative on error.
1212  *
1213  * Caller should hold map_sem for read and request_mutex.
1214  */
1215 static int __map_request(struct ceph_osd_client *osdc,
1216                          struct ceph_osd_request *req, int force_resend)
1217 {
1218         struct ceph_pg pgid;
1219         int acting[CEPH_PG_MAX_SIZE];
1220         int o = -1, num = 0;
1221         int err;
1222
1223         dout("map_request %p tid %lld\n", req, req->r_tid);
1224         err = ceph_calc_ceph_pg(&pgid, req->r_oid, osdc->osdmap,
1225                                 ceph_file_layout_pg_pool(req->r_file_layout));
1226         if (err) {
1227                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1228                 return err;
1229         }
1230         req->r_pgid = pgid;
1231
1232         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
1233         if (err > 0) {
1234                 o = acting[0];
1235                 num = err;
1236         }
1237
1238         if ((!force_resend &&
1239              req->r_osd && req->r_osd->o_osd == o &&
1240              req->r_sent >= req->r_osd->o_incarnation &&
1241              req->r_num_pg_osds == num &&
1242              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1243             (req->r_osd == NULL && o == -1))
1244                 return 0;  /* no change */
1245
1246         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1247              req->r_tid, pgid.pool, pgid.seed, o,
1248              req->r_osd ? req->r_osd->o_osd : -1);
1249
1250         /* record full pg acting set */
1251         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1252         req->r_num_pg_osds = num;
1253
1254         if (req->r_osd) {
1255                 __cancel_request(req);
1256                 list_del_init(&req->r_osd_item);
1257                 req->r_osd = NULL;
1258         }
1259
1260         req->r_osd = __lookup_osd(osdc, o);
1261         if (!req->r_osd && o >= 0) {
1262                 err = -ENOMEM;
1263                 req->r_osd = create_osd(osdc, o);
1264                 if (!req->r_osd) {
1265                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1266                         goto out;
1267                 }
1268
1269                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1270                 __insert_osd(osdc, req->r_osd);
1271
1272                 ceph_con_open(&req->r_osd->o_con,
1273                               CEPH_ENTITY_TYPE_OSD, o,
1274                               &osdc->osdmap->osd_addr[o]);
1275         }
1276
1277         if (req->r_osd) {
1278                 __remove_osd_from_lru(req->r_osd);
1279                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1280                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1281         } else {
1282                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1283         }
1284         err = 1;   /* osd or pg changed */
1285
1286 out:
1287         return err;
1288 }
1289
1290 /*
1291  * caller should hold map_sem (for read) and request_mutex
1292  */
1293 static void __send_request(struct ceph_osd_client *osdc,
1294                            struct ceph_osd_request *req)
1295 {
1296         void *p;
1297
1298         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1299              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1300              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1301
1302         /* fill in message content that changes each time we send it */
1303         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1304         put_unaligned_le32(req->r_flags, req->r_request_flags);
1305         put_unaligned_le64(req->r_pgid.pool, req->r_request_pool);
1306         p = req->r_request_pgid;
1307         ceph_encode_64(&p, req->r_pgid.pool);
1308         ceph_encode_32(&p, req->r_pgid.seed);
1309         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1310         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1311                sizeof(req->r_reassert_version));
1312
1313         req->r_stamp = jiffies;
1314         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1315
1316         ceph_msg_get(req->r_request); /* send consumes a ref */
1317
1318         /* Mark the request unsafe if this is the first timet's being sent. */
1319
1320         if (!req->r_sent && req->r_unsafe_callback)
1321                 req->r_unsafe_callback(req, true);
1322         req->r_sent = req->r_osd->o_incarnation;
1323
1324         ceph_con_send(&req->r_osd->o_con, req->r_request);
1325 }
1326
1327 /*
1328  * Send any requests in the queue (req_unsent).
1329  */
1330 static void __send_queued(struct ceph_osd_client *osdc)
1331 {
1332         struct ceph_osd_request *req, *tmp;
1333
1334         dout("__send_queued\n");
1335         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1336                 __send_request(osdc, req);
1337 }
1338
1339 /*
1340  * Timeout callback, called every N seconds when 1 or more osd
1341  * requests has been active for more than N seconds.  When this
1342  * happens, we ping all OSDs with requests who have timed out to
1343  * ensure any communications channel reset is detected.  Reset the
1344  * request timeouts another N seconds in the future as we go.
1345  * Reschedule the timeout event another N seconds in future (unless
1346  * there are no open requests).
1347  */
1348 static void handle_timeout(struct work_struct *work)
1349 {
1350         struct ceph_osd_client *osdc =
1351                 container_of(work, struct ceph_osd_client, timeout_work.work);
1352         struct ceph_osd_request *req;
1353         struct ceph_osd *osd;
1354         unsigned long keepalive =
1355                 osdc->client->options->osd_keepalive_timeout * HZ;
1356         struct list_head slow_osds;
1357         dout("timeout\n");
1358         down_read(&osdc->map_sem);
1359
1360         ceph_monc_request_next_osdmap(&osdc->client->monc);
1361
1362         mutex_lock(&osdc->request_mutex);
1363
1364         /*
1365          * ping osds that are a bit slow.  this ensures that if there
1366          * is a break in the TCP connection we will notice, and reopen
1367          * a connection with that osd (from the fault callback).
1368          */
1369         INIT_LIST_HEAD(&slow_osds);
1370         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1371                 if (time_before(jiffies, req->r_stamp + keepalive))
1372                         break;
1373
1374                 osd = req->r_osd;
1375                 BUG_ON(!osd);
1376                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1377                      req->r_tid, osd->o_osd);
1378                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1379         }
1380         while (!list_empty(&slow_osds)) {
1381                 osd = list_entry(slow_osds.next, struct ceph_osd,
1382                                  o_keepalive_item);
1383                 list_del_init(&osd->o_keepalive_item);
1384                 ceph_con_keepalive(&osd->o_con);
1385         }
1386
1387         __schedule_osd_timeout(osdc);
1388         __send_queued(osdc);
1389         mutex_unlock(&osdc->request_mutex);
1390         up_read(&osdc->map_sem);
1391 }
1392
1393 static void handle_osds_timeout(struct work_struct *work)
1394 {
1395         struct ceph_osd_client *osdc =
1396                 container_of(work, struct ceph_osd_client,
1397                              osds_timeout_work.work);
1398         unsigned long delay =
1399                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1400
1401         dout("osds timeout\n");
1402         down_read(&osdc->map_sem);
1403         remove_old_osds(osdc);
1404         up_read(&osdc->map_sem);
1405
1406         schedule_delayed_work(&osdc->osds_timeout_work,
1407                               round_jiffies_relative(delay));
1408 }
1409
1410 static void complete_request(struct ceph_osd_request *req)
1411 {
1412         if (req->r_unsafe_callback)
1413                 req->r_unsafe_callback(req, false);
1414         complete_all(&req->r_safe_completion);  /* fsync waiter */
1415 }
1416
1417 /*
1418  * handle osd op reply.  either call the callback if it is specified,
1419  * or do the completion to wake up the waiting thread.
1420  */
1421 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1422                          struct ceph_connection *con)
1423 {
1424         void *p, *end;
1425         struct ceph_osd_request *req;
1426         u64 tid;
1427         int object_len;
1428         unsigned int numops;
1429         int payload_len, flags;
1430         s32 result;
1431         s32 retry_attempt;
1432         struct ceph_pg pg;
1433         int err;
1434         u32 reassert_epoch;
1435         u64 reassert_version;
1436         u32 osdmap_epoch;
1437         int already_completed;
1438         u32 bytes;
1439         unsigned int i;
1440
1441         tid = le64_to_cpu(msg->hdr.tid);
1442         dout("handle_reply %p tid %llu\n", msg, tid);
1443
1444         p = msg->front.iov_base;
1445         end = p + msg->front.iov_len;
1446
1447         ceph_decode_need(&p, end, 4, bad);
1448         object_len = ceph_decode_32(&p);
1449         ceph_decode_need(&p, end, object_len, bad);
1450         p += object_len;
1451
1452         err = ceph_decode_pgid(&p, end, &pg);
1453         if (err)
1454                 goto bad;
1455
1456         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1457         flags = ceph_decode_64(&p);
1458         result = ceph_decode_32(&p);
1459         reassert_epoch = ceph_decode_32(&p);
1460         reassert_version = ceph_decode_64(&p);
1461         osdmap_epoch = ceph_decode_32(&p);
1462
1463         /* lookup */
1464         mutex_lock(&osdc->request_mutex);
1465         req = __lookup_request(osdc, tid);
1466         if (req == NULL) {
1467                 dout("handle_reply tid %llu dne\n", tid);
1468                 goto bad_mutex;
1469         }
1470         ceph_osdc_get_request(req);
1471
1472         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1473              req, result);
1474
1475         ceph_decode_need(&p, end, 4, bad);
1476         numops = ceph_decode_32(&p);
1477         if (numops > CEPH_OSD_MAX_OP)
1478                 goto bad_put;
1479         if (numops != req->r_num_ops)
1480                 goto bad_put;
1481         payload_len = 0;
1482         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad);
1483         for (i = 0; i < numops; i++) {
1484                 struct ceph_osd_op *op = p;
1485                 int len;
1486
1487                 len = le32_to_cpu(op->payload_len);
1488                 req->r_reply_op_len[i] = len;
1489                 dout(" op %d has %d bytes\n", i, len);
1490                 payload_len += len;
1491                 p += sizeof(*op);
1492         }
1493         bytes = le32_to_cpu(msg->hdr.data_len);
1494         if (payload_len != bytes) {
1495                 pr_warning("sum of op payload lens %d != data_len %d",
1496                            payload_len, bytes);
1497                 goto bad_put;
1498         }
1499
1500         ceph_decode_need(&p, end, 4 + numops * 4, bad);
1501         retry_attempt = ceph_decode_32(&p);
1502         for (i = 0; i < numops; i++)
1503                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1504
1505         if (!req->r_got_reply) {
1506
1507                 req->r_result = result;
1508                 dout("handle_reply result %d bytes %d\n", req->r_result,
1509                      bytes);
1510                 if (req->r_result == 0)
1511                         req->r_result = bytes;
1512
1513                 /* in case this is a write and we need to replay, */
1514                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1515                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1516
1517                 req->r_got_reply = 1;
1518         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1519                 dout("handle_reply tid %llu dup ack\n", tid);
1520                 mutex_unlock(&osdc->request_mutex);
1521                 goto done;
1522         }
1523
1524         dout("handle_reply tid %llu flags %d\n", tid, flags);
1525
1526         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1527                 __register_linger_request(osdc, req);
1528
1529         /* either this is a read, or we got the safe response */
1530         if (result < 0 ||
1531             (flags & CEPH_OSD_FLAG_ONDISK) ||
1532             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1533                 __unregister_request(osdc, req);
1534
1535         already_completed = req->r_completed;
1536         req->r_completed = 1;
1537         mutex_unlock(&osdc->request_mutex);
1538         if (already_completed)
1539                 goto done;
1540
1541         if (req->r_callback)
1542                 req->r_callback(req, msg);
1543         else
1544                 complete_all(&req->r_completion);
1545
1546         if (flags & CEPH_OSD_FLAG_ONDISK)
1547                 complete_request(req);
1548
1549 done:
1550         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1551         ceph_osdc_put_request(req);
1552         return;
1553
1554 bad_put:
1555         ceph_osdc_put_request(req);
1556 bad_mutex:
1557         mutex_unlock(&osdc->request_mutex);
1558 bad:
1559         pr_err("corrupt osd_op_reply got %d %d\n",
1560                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1561         ceph_msg_dump(msg);
1562 }
1563
1564 static void reset_changed_osds(struct ceph_osd_client *osdc)
1565 {
1566         struct rb_node *p, *n;
1567
1568         for (p = rb_first(&osdc->osds); p; p = n) {
1569                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1570
1571                 n = rb_next(p);
1572                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1573                     memcmp(&osd->o_con.peer_addr,
1574                            ceph_osd_addr(osdc->osdmap,
1575                                          osd->o_osd),
1576                            sizeof(struct ceph_entity_addr)) != 0)
1577                         __reset_osd(osdc, osd);
1578         }
1579 }
1580
1581 /*
1582  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1583  * no osd, request a new map.
1584  *
1585  * Caller should hold map_sem for read.
1586  */
1587 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1588 {
1589         struct ceph_osd_request *req, *nreq;
1590         struct rb_node *p;
1591         int needmap = 0;
1592         int err;
1593
1594         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1595         mutex_lock(&osdc->request_mutex);
1596         for (p = rb_first(&osdc->requests); p; ) {
1597                 req = rb_entry(p, struct ceph_osd_request, r_node);
1598                 p = rb_next(p);
1599
1600                 /*
1601                  * For linger requests that have not yet been
1602                  * registered, move them to the linger list; they'll
1603                  * be sent to the osd in the loop below.  Unregister
1604                  * the request before re-registering it as a linger
1605                  * request to ensure the __map_request() below
1606                  * will decide it needs to be sent.
1607                  */
1608                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1609                         dout("%p tid %llu restart on osd%d\n",
1610                              req, req->r_tid,
1611                              req->r_osd ? req->r_osd->o_osd : -1);
1612                         __unregister_request(osdc, req);
1613                         __register_linger_request(osdc, req);
1614                         continue;
1615                 }
1616
1617                 err = __map_request(osdc, req, force_resend);
1618                 if (err < 0)
1619                         continue;  /* error */
1620                 if (req->r_osd == NULL) {
1621                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1622                         needmap++;  /* request a newer map */
1623                 } else if (err > 0) {
1624                         if (!req->r_linger) {
1625                                 dout("%p tid %llu requeued on osd%d\n", req,
1626                                      req->r_tid,
1627                                      req->r_osd ? req->r_osd->o_osd : -1);
1628                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1629                         }
1630                 }
1631         }
1632
1633         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1634                                  r_linger_item) {
1635                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1636
1637                 err = __map_request(osdc, req, force_resend);
1638                 dout("__map_request returned %d\n", err);
1639                 if (err == 0)
1640                         continue;  /* no change and no osd was specified */
1641                 if (err < 0)
1642                         continue;  /* hrm! */
1643                 if (req->r_osd == NULL) {
1644                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1645                         needmap++;  /* request a newer map */
1646                         continue;
1647                 }
1648
1649                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1650                      req->r_osd ? req->r_osd->o_osd : -1);
1651                 __register_request(osdc, req);
1652                 __unregister_linger_request(osdc, req);
1653         }
1654         mutex_unlock(&osdc->request_mutex);
1655
1656         if (needmap) {
1657                 dout("%d requests for down osds, need new map\n", needmap);
1658                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1659         }
1660         reset_changed_osds(osdc);
1661 }
1662
1663
1664 /*
1665  * Process updated osd map.
1666  *
1667  * The message contains any number of incremental and full maps, normally
1668  * indicating some sort of topology change in the cluster.  Kick requests
1669  * off to different OSDs as needed.
1670  */
1671 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1672 {
1673         void *p, *end, *next;
1674         u32 nr_maps, maplen;
1675         u32 epoch;
1676         struct ceph_osdmap *newmap = NULL, *oldmap;
1677         int err;
1678         struct ceph_fsid fsid;
1679
1680         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1681         p = msg->front.iov_base;
1682         end = p + msg->front.iov_len;
1683
1684         /* verify fsid */
1685         ceph_decode_need(&p, end, sizeof(fsid), bad);
1686         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1687         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1688                 return;
1689
1690         down_write(&osdc->map_sem);
1691
1692         /* incremental maps */
1693         ceph_decode_32_safe(&p, end, nr_maps, bad);
1694         dout(" %d inc maps\n", nr_maps);
1695         while (nr_maps > 0) {
1696                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1697                 epoch = ceph_decode_32(&p);
1698                 maplen = ceph_decode_32(&p);
1699                 ceph_decode_need(&p, end, maplen, bad);
1700                 next = p + maplen;
1701                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1702                         dout("applying incremental map %u len %d\n",
1703                              epoch, maplen);
1704                         newmap = osdmap_apply_incremental(&p, next,
1705                                                           osdc->osdmap,
1706                                                           &osdc->client->msgr);
1707                         if (IS_ERR(newmap)) {
1708                                 err = PTR_ERR(newmap);
1709                                 goto bad;
1710                         }
1711                         BUG_ON(!newmap);
1712                         if (newmap != osdc->osdmap) {
1713                                 ceph_osdmap_destroy(osdc->osdmap);
1714                                 osdc->osdmap = newmap;
1715                         }
1716                         kick_requests(osdc, 0);
1717                 } else {
1718                         dout("ignoring incremental map %u len %d\n",
1719                              epoch, maplen);
1720                 }
1721                 p = next;
1722                 nr_maps--;
1723         }
1724         if (newmap)
1725                 goto done;
1726
1727         /* full maps */
1728         ceph_decode_32_safe(&p, end, nr_maps, bad);
1729         dout(" %d full maps\n", nr_maps);
1730         while (nr_maps) {
1731                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1732                 epoch = ceph_decode_32(&p);
1733                 maplen = ceph_decode_32(&p);
1734                 ceph_decode_need(&p, end, maplen, bad);
1735                 if (nr_maps > 1) {
1736                         dout("skipping non-latest full map %u len %d\n",
1737                              epoch, maplen);
1738                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1739                         dout("skipping full map %u len %d, "
1740                              "older than our %u\n", epoch, maplen,
1741                              osdc->osdmap->epoch);
1742                 } else {
1743                         int skipped_map = 0;
1744
1745                         dout("taking full map %u len %d\n", epoch, maplen);
1746                         newmap = osdmap_decode(&p, p+maplen);
1747                         if (IS_ERR(newmap)) {
1748                                 err = PTR_ERR(newmap);
1749                                 goto bad;
1750                         }
1751                         BUG_ON(!newmap);
1752                         oldmap = osdc->osdmap;
1753                         osdc->osdmap = newmap;
1754                         if (oldmap) {
1755                                 if (oldmap->epoch + 1 < newmap->epoch)
1756                                         skipped_map = 1;
1757                                 ceph_osdmap_destroy(oldmap);
1758                         }
1759                         kick_requests(osdc, skipped_map);
1760                 }
1761                 p += maplen;
1762                 nr_maps--;
1763         }
1764
1765 done:
1766         downgrade_write(&osdc->map_sem);
1767         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1768
1769         /*
1770          * subscribe to subsequent osdmap updates if full to ensure
1771          * we find out when we are no longer full and stop returning
1772          * ENOSPC.
1773          */
1774         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1775                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1776
1777         mutex_lock(&osdc->request_mutex);
1778         __send_queued(osdc);
1779         mutex_unlock(&osdc->request_mutex);
1780         up_read(&osdc->map_sem);
1781         wake_up_all(&osdc->client->auth_wq);
1782         return;
1783
1784 bad:
1785         pr_err("osdc handle_map corrupt msg\n");
1786         ceph_msg_dump(msg);
1787         up_write(&osdc->map_sem);
1788         return;
1789 }
1790
1791 /*
1792  * watch/notify callback event infrastructure
1793  *
1794  * These callbacks are used both for watch and notify operations.
1795  */
1796 static void __release_event(struct kref *kref)
1797 {
1798         struct ceph_osd_event *event =
1799                 container_of(kref, struct ceph_osd_event, kref);
1800
1801         dout("__release_event %p\n", event);
1802         kfree(event);
1803 }
1804
1805 static void get_event(struct ceph_osd_event *event)
1806 {
1807         kref_get(&event->kref);
1808 }
1809
1810 void ceph_osdc_put_event(struct ceph_osd_event *event)
1811 {
1812         kref_put(&event->kref, __release_event);
1813 }
1814 EXPORT_SYMBOL(ceph_osdc_put_event);
1815
1816 static void __insert_event(struct ceph_osd_client *osdc,
1817                              struct ceph_osd_event *new)
1818 {
1819         struct rb_node **p = &osdc->event_tree.rb_node;
1820         struct rb_node *parent = NULL;
1821         struct ceph_osd_event *event = NULL;
1822
1823         while (*p) {
1824                 parent = *p;
1825                 event = rb_entry(parent, struct ceph_osd_event, node);
1826                 if (new->cookie < event->cookie)
1827                         p = &(*p)->rb_left;
1828                 else if (new->cookie > event->cookie)
1829                         p = &(*p)->rb_right;
1830                 else
1831                         BUG();
1832         }
1833
1834         rb_link_node(&new->node, parent, p);
1835         rb_insert_color(&new->node, &osdc->event_tree);
1836 }
1837
1838 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1839                                                 u64 cookie)
1840 {
1841         struct rb_node **p = &osdc->event_tree.rb_node;
1842         struct rb_node *parent = NULL;
1843         struct ceph_osd_event *event = NULL;
1844
1845         while (*p) {
1846                 parent = *p;
1847                 event = rb_entry(parent, struct ceph_osd_event, node);
1848                 if (cookie < event->cookie)
1849                         p = &(*p)->rb_left;
1850                 else if (cookie > event->cookie)
1851                         p = &(*p)->rb_right;
1852                 else
1853                         return event;
1854         }
1855         return NULL;
1856 }
1857
1858 static void __remove_event(struct ceph_osd_event *event)
1859 {
1860         struct ceph_osd_client *osdc = event->osdc;
1861
1862         if (!RB_EMPTY_NODE(&event->node)) {
1863                 dout("__remove_event removed %p\n", event);
1864                 rb_erase(&event->node, &osdc->event_tree);
1865                 ceph_osdc_put_event(event);
1866         } else {
1867                 dout("__remove_event didn't remove %p\n", event);
1868         }
1869 }
1870
1871 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1872                            void (*event_cb)(u64, u64, u8, void *),
1873                            void *data, struct ceph_osd_event **pevent)
1874 {
1875         struct ceph_osd_event *event;
1876
1877         event = kmalloc(sizeof(*event), GFP_NOIO);
1878         if (!event)
1879                 return -ENOMEM;
1880
1881         dout("create_event %p\n", event);
1882         event->cb = event_cb;
1883         event->one_shot = 0;
1884         event->data = data;
1885         event->osdc = osdc;
1886         INIT_LIST_HEAD(&event->osd_node);
1887         RB_CLEAR_NODE(&event->node);
1888         kref_init(&event->kref);   /* one ref for us */
1889         kref_get(&event->kref);    /* one ref for the caller */
1890
1891         spin_lock(&osdc->event_lock);
1892         event->cookie = ++osdc->event_count;
1893         __insert_event(osdc, event);
1894         spin_unlock(&osdc->event_lock);
1895
1896         *pevent = event;
1897         return 0;
1898 }
1899 EXPORT_SYMBOL(ceph_osdc_create_event);
1900
1901 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1902 {
1903         struct ceph_osd_client *osdc = event->osdc;
1904
1905         dout("cancel_event %p\n", event);
1906         spin_lock(&osdc->event_lock);
1907         __remove_event(event);
1908         spin_unlock(&osdc->event_lock);
1909         ceph_osdc_put_event(event); /* caller's */
1910 }
1911 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1912
1913
1914 static void do_event_work(struct work_struct *work)
1915 {
1916         struct ceph_osd_event_work *event_work =
1917                 container_of(work, struct ceph_osd_event_work, work);
1918         struct ceph_osd_event *event = event_work->event;
1919         u64 ver = event_work->ver;
1920         u64 notify_id = event_work->notify_id;
1921         u8 opcode = event_work->opcode;
1922
1923         dout("do_event_work completing %p\n", event);
1924         event->cb(ver, notify_id, opcode, event->data);
1925         dout("do_event_work completed %p\n", event);
1926         ceph_osdc_put_event(event);
1927         kfree(event_work);
1928 }
1929
1930
1931 /*
1932  * Process osd watch notifications
1933  */
1934 static void handle_watch_notify(struct ceph_osd_client *osdc,
1935                                 struct ceph_msg *msg)
1936 {
1937         void *p, *end;
1938         u8 proto_ver;
1939         u64 cookie, ver, notify_id;
1940         u8 opcode;
1941         struct ceph_osd_event *event;
1942         struct ceph_osd_event_work *event_work;
1943
1944         p = msg->front.iov_base;
1945         end = p + msg->front.iov_len;
1946
1947         ceph_decode_8_safe(&p, end, proto_ver, bad);
1948         ceph_decode_8_safe(&p, end, opcode, bad);
1949         ceph_decode_64_safe(&p, end, cookie, bad);
1950         ceph_decode_64_safe(&p, end, ver, bad);
1951         ceph_decode_64_safe(&p, end, notify_id, bad);
1952
1953         spin_lock(&osdc->event_lock);
1954         event = __find_event(osdc, cookie);
1955         if (event) {
1956                 BUG_ON(event->one_shot);
1957                 get_event(event);
1958         }
1959         spin_unlock(&osdc->event_lock);
1960         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1961              cookie, ver, event);
1962         if (event) {
1963                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1964                 if (!event_work) {
1965                         dout("ERROR: could not allocate event_work\n");
1966                         goto done_err;
1967                 }
1968                 INIT_WORK(&event_work->work, do_event_work);
1969                 event_work->event = event;
1970                 event_work->ver = ver;
1971                 event_work->notify_id = notify_id;
1972                 event_work->opcode = opcode;
1973                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1974                         dout("WARNING: failed to queue notify event work\n");
1975                         goto done_err;
1976                 }
1977         }
1978
1979         return;
1980
1981 done_err:
1982         ceph_osdc_put_event(event);
1983         return;
1984
1985 bad:
1986         pr_err("osdc handle_watch_notify corrupt msg\n");
1987         return;
1988 }
1989
1990 /*
1991  * build new request AND message
1992  *
1993  */
1994 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
1995                                 struct ceph_snap_context *snapc, u64 snap_id,
1996                                 struct timespec *mtime)
1997 {
1998         struct ceph_msg *msg = req->r_request;
1999         void *p;
2000         size_t msg_size;
2001         int flags = req->r_flags;
2002         u64 data_len;
2003         unsigned int i;
2004
2005         req->r_snapid = snap_id;
2006         req->r_snapc = ceph_get_snap_context(snapc);
2007
2008         /* encode request */
2009         msg->hdr.version = cpu_to_le16(4);
2010
2011         p = msg->front.iov_base;
2012         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2013         req->r_request_osdmap_epoch = p;
2014         p += 4;
2015         req->r_request_flags = p;
2016         p += 4;
2017         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2018                 ceph_encode_timespec(p, mtime);
2019         p += sizeof(struct ceph_timespec);
2020         req->r_request_reassert_version = p;
2021         p += sizeof(struct ceph_eversion); /* will get filled in */
2022
2023         /* oloc */
2024         ceph_encode_8(&p, 4);
2025         ceph_encode_8(&p, 4);
2026         ceph_encode_32(&p, 8 + 4 + 4);
2027         req->r_request_pool = p;
2028         p += 8;
2029         ceph_encode_32(&p, -1);  /* preferred */
2030         ceph_encode_32(&p, 0);   /* key len */
2031
2032         ceph_encode_8(&p, 1);
2033         req->r_request_pgid = p;
2034         p += 8 + 4;
2035         ceph_encode_32(&p, -1);  /* preferred */
2036
2037         /* oid */
2038         ceph_encode_32(&p, req->r_oid_len);
2039         memcpy(p, req->r_oid, req->r_oid_len);
2040         dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len);
2041         p += req->r_oid_len;
2042
2043         /* ops--can imply data */
2044         ceph_encode_16(&p, (u16)req->r_num_ops);
2045         data_len = 0;
2046         for (i = 0; i < req->r_num_ops; i++) {
2047                 data_len += osd_req_encode_op(req, p, i);
2048                 p += sizeof(struct ceph_osd_op);
2049         }
2050
2051         /* snaps */
2052         ceph_encode_64(&p, req->r_snapid);
2053         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2054         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2055         if (req->r_snapc) {
2056                 for (i = 0; i < snapc->num_snaps; i++) {
2057                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2058                 }
2059         }
2060
2061         req->r_request_attempts = p;
2062         p += 4;
2063
2064         /* data */
2065         if (flags & CEPH_OSD_FLAG_WRITE) {
2066                 u16 data_off;
2067
2068                 /*
2069                  * The header "data_off" is a hint to the receiver
2070                  * allowing it to align received data into its
2071                  * buffers such that there's no need to re-copy
2072                  * it before writing it to disk (direct I/O).
2073                  */
2074                 data_off = (u16) (off & 0xffff);
2075                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2076         }
2077         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2078
2079         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2080         msg_size = p - msg->front.iov_base;
2081         msg->front.iov_len = msg_size;
2082         msg->hdr.front_len = cpu_to_le32(msg_size);
2083
2084         dout("build_request msg_size was %d\n", (int)msg_size);
2085 }
2086 EXPORT_SYMBOL(ceph_osdc_build_request);
2087
2088 /*
2089  * Register request, send initial attempt.
2090  */
2091 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2092                             struct ceph_osd_request *req,
2093                             bool nofail)
2094 {
2095         int rc = 0;
2096
2097         down_read(&osdc->map_sem);
2098         mutex_lock(&osdc->request_mutex);
2099         __register_request(osdc, req);
2100         WARN_ON(req->r_sent);
2101         rc = __map_request(osdc, req, 0);
2102         if (rc < 0) {
2103                 if (nofail) {
2104                         dout("osdc_start_request failed map, "
2105                                 " will retry %lld\n", req->r_tid);
2106                         rc = 0;
2107                 }
2108                 goto out_unlock;
2109         }
2110         if (req->r_osd == NULL) {
2111                 dout("send_request %p no up osds in pg\n", req);
2112                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2113         } else {
2114                 __send_queued(osdc);
2115         }
2116         rc = 0;
2117 out_unlock:
2118         mutex_unlock(&osdc->request_mutex);
2119         up_read(&osdc->map_sem);
2120         return rc;
2121 }
2122 EXPORT_SYMBOL(ceph_osdc_start_request);
2123
2124 /*
2125  * wait for a request to complete
2126  */
2127 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2128                            struct ceph_osd_request *req)
2129 {
2130         int rc;
2131
2132         rc = wait_for_completion_interruptible(&req->r_completion);
2133         if (rc < 0) {
2134                 mutex_lock(&osdc->request_mutex);
2135                 __cancel_request(req);
2136                 __unregister_request(osdc, req);
2137                 mutex_unlock(&osdc->request_mutex);
2138                 complete_request(req);
2139                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
2140                 return rc;
2141         }
2142
2143         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
2144         return req->r_result;
2145 }
2146 EXPORT_SYMBOL(ceph_osdc_wait_request);
2147
2148 /*
2149  * sync - wait for all in-flight requests to flush.  avoid starvation.
2150  */
2151 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2152 {
2153         struct ceph_osd_request *req;
2154         u64 last_tid, next_tid = 0;
2155
2156         mutex_lock(&osdc->request_mutex);
2157         last_tid = osdc->last_tid;
2158         while (1) {
2159                 req = __lookup_request_ge(osdc, next_tid);
2160                 if (!req)
2161                         break;
2162                 if (req->r_tid > last_tid)
2163                         break;
2164
2165                 next_tid = req->r_tid + 1;
2166                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2167                         continue;
2168
2169                 ceph_osdc_get_request(req);
2170                 mutex_unlock(&osdc->request_mutex);
2171                 dout("sync waiting on tid %llu (last is %llu)\n",
2172                      req->r_tid, last_tid);
2173                 wait_for_completion(&req->r_safe_completion);
2174                 mutex_lock(&osdc->request_mutex);
2175                 ceph_osdc_put_request(req);
2176         }
2177         mutex_unlock(&osdc->request_mutex);
2178         dout("sync done (thru tid %llu)\n", last_tid);
2179 }
2180 EXPORT_SYMBOL(ceph_osdc_sync);
2181
2182 /*
2183  * init, shutdown
2184  */
2185 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2186 {
2187         int err;
2188
2189         dout("init\n");
2190         osdc->client = client;
2191         osdc->osdmap = NULL;
2192         init_rwsem(&osdc->map_sem);
2193         init_completion(&osdc->map_waiters);
2194         osdc->last_requested_map = 0;
2195         mutex_init(&osdc->request_mutex);
2196         osdc->last_tid = 0;
2197         osdc->osds = RB_ROOT;
2198         INIT_LIST_HEAD(&osdc->osd_lru);
2199         osdc->requests = RB_ROOT;
2200         INIT_LIST_HEAD(&osdc->req_lru);
2201         INIT_LIST_HEAD(&osdc->req_unsent);
2202         INIT_LIST_HEAD(&osdc->req_notarget);
2203         INIT_LIST_HEAD(&osdc->req_linger);
2204         osdc->num_requests = 0;
2205         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2206         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2207         spin_lock_init(&osdc->event_lock);
2208         osdc->event_tree = RB_ROOT;
2209         osdc->event_count = 0;
2210
2211         schedule_delayed_work(&osdc->osds_timeout_work,
2212            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2213
2214         err = -ENOMEM;
2215         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2216                                         sizeof(struct ceph_osd_request));
2217         if (!osdc->req_mempool)
2218                 goto out;
2219
2220         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2221                                 OSD_OP_FRONT_LEN, 10, true,
2222                                 "osd_op");
2223         if (err < 0)
2224                 goto out_mempool;
2225         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2226                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2227                                 "osd_op_reply");
2228         if (err < 0)
2229                 goto out_msgpool;
2230
2231         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2232         if (IS_ERR(osdc->notify_wq)) {
2233                 err = PTR_ERR(osdc->notify_wq);
2234                 osdc->notify_wq = NULL;
2235                 goto out_msgpool;
2236         }
2237         return 0;
2238
2239 out_msgpool:
2240         ceph_msgpool_destroy(&osdc->msgpool_op);
2241 out_mempool:
2242         mempool_destroy(osdc->req_mempool);
2243 out:
2244         return err;
2245 }
2246
2247 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2248 {
2249         flush_workqueue(osdc->notify_wq);
2250         destroy_workqueue(osdc->notify_wq);
2251         cancel_delayed_work_sync(&osdc->timeout_work);
2252         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2253         if (osdc->osdmap) {
2254                 ceph_osdmap_destroy(osdc->osdmap);
2255                 osdc->osdmap = NULL;
2256         }
2257         remove_all_osds(osdc);
2258         mempool_destroy(osdc->req_mempool);
2259         ceph_msgpool_destroy(&osdc->msgpool_op);
2260         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2261 }
2262
2263 /*
2264  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2265  * *plen.  Return number of bytes read, or error.
2266  */
2267 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2268                         struct ceph_vino vino, struct ceph_file_layout *layout,
2269                         u64 off, u64 *plen,
2270                         u32 truncate_seq, u64 truncate_size,
2271                         struct page **pages, int num_pages, int page_align)
2272 {
2273         struct ceph_osd_request *req;
2274         int rc = 0;
2275
2276         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2277              vino.snap, off, *plen);
2278         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 1,
2279                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2280                                     NULL, truncate_seq, truncate_size,
2281                                     false);
2282         if (IS_ERR(req))
2283                 return PTR_ERR(req);
2284
2285         /* it may be a short read due to an object boundary */
2286
2287         osd_req_op_extent_osd_data_pages(req, 0, false,
2288                                 pages, *plen, page_align, false, false);
2289
2290         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2291              off, *plen, *plen, page_align);
2292
2293         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2294
2295         rc = ceph_osdc_start_request(osdc, req, false);
2296         if (!rc)
2297                 rc = ceph_osdc_wait_request(osdc, req);
2298
2299         ceph_osdc_put_request(req);
2300         dout("readpages result %d\n", rc);
2301         return rc;
2302 }
2303 EXPORT_SYMBOL(ceph_osdc_readpages);
2304
2305 /*
2306  * do a synchronous write on N pages
2307  */
2308 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2309                          struct ceph_file_layout *layout,
2310                          struct ceph_snap_context *snapc,
2311                          u64 off, u64 len,
2312                          u32 truncate_seq, u64 truncate_size,
2313                          struct timespec *mtime,
2314                          struct page **pages, int num_pages)
2315 {
2316         struct ceph_osd_request *req;
2317         int rc = 0;
2318         int page_align = off & ~PAGE_MASK;
2319
2320         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2321         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 1,
2322                                     CEPH_OSD_OP_WRITE,
2323                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2324                                     snapc, truncate_seq, truncate_size,
2325                                     true);
2326         if (IS_ERR(req))
2327                 return PTR_ERR(req);
2328
2329         /* it may be a short write due to an object boundary */
2330         osd_req_op_extent_osd_data_pages(req, 0, true, pages, len, page_align,
2331                                 false, false);
2332         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2333
2334         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2335
2336         rc = ceph_osdc_start_request(osdc, req, true);
2337         if (!rc)
2338                 rc = ceph_osdc_wait_request(osdc, req);
2339
2340         ceph_osdc_put_request(req);
2341         if (rc == 0)
2342                 rc = len;
2343         dout("writepages result %d\n", rc);
2344         return rc;
2345 }
2346 EXPORT_SYMBOL(ceph_osdc_writepages);
2347
2348 /*
2349  * handle incoming message
2350  */
2351 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2352 {
2353         struct ceph_osd *osd = con->private;
2354         struct ceph_osd_client *osdc;
2355         int type = le16_to_cpu(msg->hdr.type);
2356
2357         if (!osd)
2358                 goto out;
2359         osdc = osd->o_osdc;
2360
2361         switch (type) {
2362         case CEPH_MSG_OSD_MAP:
2363                 ceph_osdc_handle_map(osdc, msg);
2364                 break;
2365         case CEPH_MSG_OSD_OPREPLY:
2366                 handle_reply(osdc, msg, con);
2367                 break;
2368         case CEPH_MSG_WATCH_NOTIFY:
2369                 handle_watch_notify(osdc, msg);
2370                 break;
2371
2372         default:
2373                 pr_err("received unknown message type %d %s\n", type,
2374                        ceph_msg_type_name(type));
2375         }
2376 out:
2377         ceph_msg_put(msg);
2378 }
2379
2380 /*
2381  * lookup and return message for incoming reply.  set up reply message
2382  * pages.
2383  */
2384 static struct ceph_msg *get_reply(struct ceph_connection *con,
2385                                   struct ceph_msg_header *hdr,
2386                                   int *skip)
2387 {
2388         struct ceph_osd *osd = con->private;
2389         struct ceph_osd_client *osdc = osd->o_osdc;
2390         struct ceph_msg *m;
2391         struct ceph_osd_request *req;
2392         int front = le32_to_cpu(hdr->front_len);
2393         int data_len = le32_to_cpu(hdr->data_len);
2394         u64 tid;
2395
2396         tid = le64_to_cpu(hdr->tid);
2397         mutex_lock(&osdc->request_mutex);
2398         req = __lookup_request(osdc, tid);
2399         if (!req) {
2400                 *skip = 1;
2401                 m = NULL;
2402                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2403                      osd->o_osd);
2404                 goto out;
2405         }
2406
2407         if (req->r_reply->con)
2408                 dout("%s revoking msg %p from old con %p\n", __func__,
2409                      req->r_reply, req->r_reply->con);
2410         ceph_msg_revoke_incoming(req->r_reply);
2411
2412         if (front > req->r_reply->front.iov_len) {
2413                 pr_warning("get_reply front %d > preallocated %d\n",
2414                            front, (int)req->r_reply->front.iov_len);
2415                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2416                 if (!m)
2417                         goto out;
2418                 ceph_msg_put(req->r_reply);
2419                 req->r_reply = m;
2420         }
2421         m = ceph_msg_get(req->r_reply);
2422
2423         if (data_len > 0) {
2424                 struct ceph_osd_data *osd_data;
2425
2426                 /*
2427                  * XXX This is assuming there is only one op containing
2428                  * XXX page data.  Probably OK for reads, but this
2429                  * XXX ought to be done more generally.
2430                  */
2431                 osd_data = osd_req_op_extent_osd_data(req, 0, false);
2432                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2433                         if (osd_data->pages &&
2434                                 unlikely(osd_data->length < data_len)) {
2435
2436                                 pr_warning("tid %lld reply has %d bytes "
2437                                         "we had only %llu bytes ready\n",
2438                                         tid, data_len, osd_data->length);
2439                                 *skip = 1;
2440                                 ceph_msg_put(m);
2441                                 m = NULL;
2442                                 goto out;
2443                         }
2444                 }
2445         }
2446         *skip = 0;
2447         dout("get_reply tid %lld %p\n", tid, m);
2448
2449 out:
2450         mutex_unlock(&osdc->request_mutex);
2451         return m;
2452
2453 }
2454
2455 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2456                                   struct ceph_msg_header *hdr,
2457                                   int *skip)
2458 {
2459         struct ceph_osd *osd = con->private;
2460         int type = le16_to_cpu(hdr->type);
2461         int front = le32_to_cpu(hdr->front_len);
2462
2463         *skip = 0;
2464         switch (type) {
2465         case CEPH_MSG_OSD_MAP:
2466         case CEPH_MSG_WATCH_NOTIFY:
2467                 return ceph_msg_new(type, front, GFP_NOFS, false);
2468         case CEPH_MSG_OSD_OPREPLY:
2469                 return get_reply(con, hdr, skip);
2470         default:
2471                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2472                         osd->o_osd);
2473                 *skip = 1;
2474                 return NULL;
2475         }
2476 }
2477
2478 /*
2479  * Wrappers to refcount containing ceph_osd struct
2480  */
2481 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2482 {
2483         struct ceph_osd *osd = con->private;
2484         if (get_osd(osd))
2485                 return con;
2486         return NULL;
2487 }
2488
2489 static void put_osd_con(struct ceph_connection *con)
2490 {
2491         struct ceph_osd *osd = con->private;
2492         put_osd(osd);
2493 }
2494
2495 /*
2496  * authentication
2497  */
2498 /*
2499  * Note: returned pointer is the address of a structure that's
2500  * managed separately.  Caller must *not* attempt to free it.
2501  */
2502 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2503                                         int *proto, int force_new)
2504 {
2505         struct ceph_osd *o = con->private;
2506         struct ceph_osd_client *osdc = o->o_osdc;
2507         struct ceph_auth_client *ac = osdc->client->monc.auth;
2508         struct ceph_auth_handshake *auth = &o->o_auth;
2509
2510         if (force_new && auth->authorizer) {
2511                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2512                 auth->authorizer = NULL;
2513         }
2514         if (!auth->authorizer) {
2515                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2516                                                       auth);
2517                 if (ret)
2518                         return ERR_PTR(ret);
2519         } else {
2520                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2521                                                      auth);
2522                 if (ret)
2523                         return ERR_PTR(ret);
2524         }
2525         *proto = ac->protocol;
2526
2527         return auth;
2528 }
2529
2530
2531 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2532 {
2533         struct ceph_osd *o = con->private;
2534         struct ceph_osd_client *osdc = o->o_osdc;
2535         struct ceph_auth_client *ac = osdc->client->monc.auth;
2536
2537         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2538 }
2539
2540 static int invalidate_authorizer(struct ceph_connection *con)
2541 {
2542         struct ceph_osd *o = con->private;
2543         struct ceph_osd_client *osdc = o->o_osdc;
2544         struct ceph_auth_client *ac = osdc->client->monc.auth;
2545
2546         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2547         return ceph_monc_validate_auth(&osdc->client->monc);
2548 }
2549
2550 static const struct ceph_connection_operations osd_con_ops = {
2551         .get = get_osd_con,
2552         .put = put_osd_con,
2553         .dispatch = dispatch,
2554         .get_authorizer = get_authorizer,
2555         .verify_authorizer_reply = verify_authorizer_reply,
2556         .invalidate_authorizer = invalidate_authorizer,
2557         .alloc_msg = alloc_msg,
2558         .fault = osd_reset,
2559 };