net/ceph/osd_client.c

   1
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/module.h>
   5 #include <linux/err.h>
   6 #include <linux/highmem.h>
   7 #include <linux/mm.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/slab.h>
  10 #include <linux/uaccess.h>
  11 #ifdef CONFIG_BLOCK
  12 #include <linux/bio.h>
  13 #endif
  14
  15 #include <linux/ceph/libceph.h>
  16 #include <linux/ceph/osd_client.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/auth.h>
  20 #include <linux/ceph/pagelist.h>
  21
  22 #define OSD_OP_FRONT_LEN        4096
  23 #define OSD_OPREPLY_FRONT_LEN   512
  24
  25 static struct kmem_cache        *ceph_osd_request_cache;
  26
  27 static const struct ceph_connection_operations osd_con_ops;
  28
  29 static void __send_queued(struct ceph_osd_client *osdc);
  30 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  31 static void __register_request(struct ceph_osd_client *osdc,
  32                                struct ceph_osd_request *req);
  33 static void __unregister_request(struct ceph_osd_client *osdc,
  34                                  struct ceph_osd_request *req);
  35 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  36                                         struct ceph_osd_request *req);
  37 static void __enqueue_request(struct ceph_osd_request *req);
  38 static void __send_request(struct ceph_osd_client *osdc,
  39                            struct ceph_osd_request *req);
  40
  41 /*
  42  * Implement client access to distributed object storage cluster.
  43  *
  44  * All data objects are stored within a cluster/cloud of OSDs, or
  45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  47  * remote daemons serving up and coordinating consistent and safe
  48  * access to storage.
  49  *
  50  * Cluster membership and the mapping of data objects onto storage devices
  51  * are described by the osd map.
  52  *
  53  * We keep track of pending OSD requests (read, write), resubmit
  54  * requests to different OSDs when the cluster topology/data layout
  55  * change, or retry the affected requests when the communications
  56  * channel with an OSD is reset.
  57  */
  58
  59 /*
  60  * calculate the mapping of a file extent onto an object, and fill out the
  61  * request accordingly.  shorten extent as necessary if it crosses an
  62  * object boundary.
  63  *
  64  * fill osd op in request message.
  65  */
  66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
  67                         u64 *objnum, u64 *objoff, u64 *objlen)
  68 {
  69         u64 orig_len = *plen;
  70         int r;
  71
  72         /* object extent? */
  73         r = ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
  74                                           objoff, objlen);
  75         if (r < 0)
  76                 return r;
  77         if (*objlen < orig_len) {
  78                 *plen = *objlen;
  79                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  80                      orig_len - *plen, off, *plen);
  81         }
  82
  83         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
  84
  85         return 0;
  86 }
  87
  88 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
  89 {
  90         memset(osd_data, 0, sizeof (*osd_data));
  91         osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
  92 }
  93
  94 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
  95                         struct page **pages, u64 length, u32 alignment,
  96                         bool pages_from_pool, bool own_pages)
  97 {
  98         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
  99         osd_data->pages = pages;
 100         osd_data->length = length;
 101         osd_data->alignment = alignment;
 102         osd_data->pages_from_pool = pages_from_pool;
 103         osd_data->own_pages = own_pages;
 104 }
 105
 106 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
 107                         struct ceph_pagelist *pagelist)
 108 {
 109         osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
 110         osd_data->pagelist = pagelist;
 111 }
 112
 113 #ifdef CONFIG_BLOCK
 114 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
 115                         struct bio *bio, size_t bio_length)
 116 {
 117         osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
 118         osd_data->bio = bio;
 119         osd_data->bio_length = bio_length;
 120 }
 121 #endif /* CONFIG_BLOCK */
 122
 123 #define osd_req_op_data(oreq, whch, typ, fld)   \
 124         ({                                              \
 125                 BUG_ON(whch >= (oreq)->r_num_ops);      \
 126                 &(oreq)->r_ops[whch].typ.fld;           \
 127         })
 128
 129 static struct ceph_osd_data *
 130 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
 131 {
 132         BUG_ON(which >= osd_req->r_num_ops);
 133
 134         return &osd_req->r_ops[which].raw_data_in;
 135 }
 136
 137 struct ceph_osd_data *
 138 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
 139                         unsigned int which)
 140 {
 141         return osd_req_op_data(osd_req, which, extent, osd_data);
 142 }
 143 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
 144
 145 struct ceph_osd_data *
 146 osd_req_op_cls_response_data(struct ceph_osd_request *osd_req,
 147                         unsigned int which)
 148 {
 149         return osd_req_op_data(osd_req, which, cls, response_data);
 150 }
 151 EXPORT_SYMBOL(osd_req_op_cls_response_data);    /* ??? */
 152
 153 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
 154                         unsigned int which, struct page **pages,
 155                         u64 length, u32 alignment,
 156                         bool pages_from_pool, bool own_pages)
 157 {
 158         struct ceph_osd_data *osd_data;
 159
 160         osd_data = osd_req_op_raw_data_in(osd_req, which);
 161         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 162                                 pages_from_pool, own_pages);
 163 }
 164 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
 165
 166 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
 167                         unsigned int which, struct page **pages,
 168                         u64 length, u32 alignment,
 169                         bool pages_from_pool, bool own_pages)
 170 {
 171         struct ceph_osd_data *osd_data;
 172
 173         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 174         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 175                                 pages_from_pool, own_pages);
 176 }
 177 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
 178
 179 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
 180                         unsigned int which, struct ceph_pagelist *pagelist)
 181 {
 182         struct ceph_osd_data *osd_data;
 183
 184         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 185         ceph_osd_data_pagelist_init(osd_data, pagelist);
 186 }
 187 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
 188
 189 #ifdef CONFIG_BLOCK
 190 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
 191                         unsigned int which, struct bio *bio, size_t bio_length)
 192 {
 193         struct ceph_osd_data *osd_data;
 194
 195         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 196         ceph_osd_data_bio_init(osd_data, bio, bio_length);
 197 }
 198 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
 199 #endif /* CONFIG_BLOCK */
 200
 201 static void osd_req_op_cls_request_info_pagelist(
 202                         struct ceph_osd_request *osd_req,
 203                         unsigned int which, struct ceph_pagelist *pagelist)
 204 {
 205         struct ceph_osd_data *osd_data;
 206
 207         osd_data = osd_req_op_data(osd_req, which, cls, request_info);
 208         ceph_osd_data_pagelist_init(osd_data, pagelist);
 209 }
 210
 211 void osd_req_op_cls_request_data_pagelist(
 212                         struct ceph_osd_request *osd_req,
 213                         unsigned int which, struct ceph_pagelist *pagelist)
 214 {
 215         struct ceph_osd_data *osd_data;
 216
 217         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 218         ceph_osd_data_pagelist_init(osd_data, pagelist);
 219 }
 220 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
 221
 222 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
 223                         unsigned int which, struct page **pages, u64 length,
 224                         u32 alignment, bool pages_from_pool, bool own_pages)
 225 {
 226         struct ceph_osd_data *osd_data;
 227
 228         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 229         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 230                                 pages_from_pool, own_pages);
 231 }
 232 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
 233
 234 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
 235                         unsigned int which, struct page **pages, u64 length,
 236                         u32 alignment, bool pages_from_pool, bool own_pages)
 237 {
 238         struct ceph_osd_data *osd_data;
 239
 240         osd_data = osd_req_op_data(osd_req, which, cls, response_data);
 241         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 242                                 pages_from_pool, own_pages);
 243 }
 244 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
 245
 246 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
 247 {
 248         switch (osd_data->type) {
 249         case CEPH_OSD_DATA_TYPE_NONE:
 250                 return 0;
 251         case CEPH_OSD_DATA_TYPE_PAGES:
 252                 return osd_data->length;
 253         case CEPH_OSD_DATA_TYPE_PAGELIST:
 254                 return (u64)osd_data->pagelist->length;
 255 #ifdef CONFIG_BLOCK
 256         case CEPH_OSD_DATA_TYPE_BIO:
 257                 return (u64)osd_data->bio_length;
 258 #endif /* CONFIG_BLOCK */
 259         default:
 260                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
 261                 return 0;
 262         }
 263 }
 264
 265 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
 266 {
 267         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
 268                 int num_pages;
 269
 270                 num_pages = calc_pages_for((u64)osd_data->alignment,
 271                                                 (u64)osd_data->length);
 272                 ceph_release_page_vector(osd_data->pages, num_pages);
 273         }
 274         ceph_osd_data_init(osd_data);
 275 }
 276
 277 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
 278                         unsigned int which)
 279 {
 280         struct ceph_osd_req_op *op;
 281
 282         BUG_ON(which >= osd_req->r_num_ops);
 283         op = &osd_req->r_ops[which];
 284
 285         switch (op->op) {
 286         case CEPH_OSD_OP_READ:
 287         case CEPH_OSD_OP_WRITE:
 288                 ceph_osd_data_release(&op->extent.osd_data);
 289                 break;
 290         case CEPH_OSD_OP_CALL:
 291                 ceph_osd_data_release(&op->cls.request_info);
 292                 ceph_osd_data_release(&op->cls.request_data);
 293                 ceph_osd_data_release(&op->cls.response_data);
 294                 break;
 295         case CEPH_OSD_OP_SETXATTR:
 296         case CEPH_OSD_OP_CMPXATTR:
 297                 ceph_osd_data_release(&op->xattr.osd_data);
 298                 break;
 299         default:
 300                 break;
 301         }
 302 }
 303
 304 /*
 305  * requests
 306  */
 307 static void ceph_osdc_release_request(struct kref *kref)
 308 {
 309         struct ceph_osd_request *req = container_of(kref,
 310                                             struct ceph_osd_request, r_kref);
 311         unsigned int which;
 312
 313         dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
 314              req->r_request, req->r_reply);
 315         WARN_ON(!RB_EMPTY_NODE(&req->r_node));
 316         WARN_ON(!list_empty(&req->r_req_lru_item));
 317         WARN_ON(!list_empty(&req->r_osd_item));
 318         WARN_ON(!list_empty(&req->r_linger_item));
 319         WARN_ON(!list_empty(&req->r_linger_osd_item));
 320         WARN_ON(req->r_osd);
 321
 322         if (req->r_request)
 323                 ceph_msg_put(req->r_request);
 324         if (req->r_reply) {
 325                 ceph_msg_revoke_incoming(req->r_reply);
 326                 ceph_msg_put(req->r_reply);
 327         }
 328
 329         for (which = 0; which < req->r_num_ops; which++)
 330                 osd_req_op_data_release(req, which);
 331
 332         ceph_put_snap_context(req->r_snapc);
 333         if (req->r_mempool)
 334                 mempool_free(req, req->r_osdc->req_mempool);
 335         else
 336                 kmem_cache_free(ceph_osd_request_cache, req);
 337
 338 }
 339
 340 void ceph_osdc_get_request(struct ceph_osd_request *req)
 341 {
 342         dout("%s %p (was %d)\n", __func__, req,
 343              atomic_read(&req->r_kref.refcount));
 344         kref_get(&req->r_kref);
 345 }
 346 EXPORT_SYMBOL(ceph_osdc_get_request);
 347
 348 void ceph_osdc_put_request(struct ceph_osd_request *req)
 349 {
 350         dout("%s %p (was %d)\n", __func__, req,
 351              atomic_read(&req->r_kref.refcount));
 352         kref_put(&req->r_kref, ceph_osdc_release_request);
 353 }
 354 EXPORT_SYMBOL(ceph_osdc_put_request);
 355
 356 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 357                                                struct ceph_snap_context *snapc,
 358                                                unsigned int num_ops,
 359                                                bool use_mempool,
 360                                                gfp_t gfp_flags)
 361 {
 362         struct ceph_osd_request *req;
 363         struct ceph_msg *msg;
 364         size_t msg_size;
 365
 366         BUILD_BUG_ON(CEPH_OSD_MAX_OP > U16_MAX);
 367         BUG_ON(num_ops > CEPH_OSD_MAX_OP);
 368
 369         msg_size = 4 + 4 + 8 + 8 + 4+8;
 370         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
 371         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
 372         msg_size += 4 + CEPH_MAX_OID_NAME_LEN; /* oid */
 373         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
 374         msg_size += 8;  /* snapid */
 375         msg_size += 8;  /* snap_seq */
 376         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
 377         msg_size += 4;
 378
 379         if (use_mempool) {
 380                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 381                 memset(req, 0, sizeof(*req));
 382         } else {
 383                 req = kmem_cache_zalloc(ceph_osd_request_cache, gfp_flags);
 384         }
 385         if (req == NULL)
 386                 return NULL;
 387
 388         req->r_osdc = osdc;
 389         req->r_mempool = use_mempool;
 390         req->r_num_ops = num_ops;
 391
 392         kref_init(&req->r_kref);
 393         init_completion(&req->r_completion);
 394         init_completion(&req->r_safe_completion);
 395         RB_CLEAR_NODE(&req->r_node);
 396         INIT_LIST_HEAD(&req->r_unsafe_item);
 397         INIT_LIST_HEAD(&req->r_linger_item);
 398         INIT_LIST_HEAD(&req->r_linger_osd_item);
 399         INIT_LIST_HEAD(&req->r_req_lru_item);
 400         INIT_LIST_HEAD(&req->r_osd_item);
 401
 402         req->r_base_oloc.pool = -1;
 403         req->r_target_oloc.pool = -1;
 404
 405         /* create reply message */
 406         if (use_mempool)
 407                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 408         else
 409                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 410                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 411         if (!msg) {
 412                 ceph_osdc_put_request(req);
 413                 return NULL;
 414         }
 415         req->r_reply = msg;
 416
 417         /* create request message; allow space for oid */
 418         if (use_mempool)
 419                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 420         else
 421                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 422         if (!msg) {
 423                 ceph_osdc_put_request(req);
 424                 return NULL;
 425         }
 426
 427         memset(msg->front.iov_base, 0, msg->front.iov_len);
 428
 429         req->r_request = msg;
 430
 431         return req;
 432 }
 433 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 434
 435 static bool osd_req_opcode_valid(u16 opcode)
 436 {
 437         switch (opcode) {
 438 #define GENERATE_CASE(op, opcode, str)  case CEPH_OSD_OP_##op: return true;
 439 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
 440 #undef GENERATE_CASE
 441         default:
 442                 return false;
 443         }
 444 }
 445
 446 /*
 447  * This is an osd op init function for opcodes that have no data or
 448  * other information associated with them.  It also serves as a
 449  * common init routine for all the other init functions, below.
 450  */
 451 static struct ceph_osd_req_op *
 452 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
 453                                 u16 opcode)
 454 {
 455         struct ceph_osd_req_op *op;
 456
 457         BUG_ON(which >= osd_req->r_num_ops);
 458         BUG_ON(!osd_req_opcode_valid(opcode));
 459
 460         op = &osd_req->r_ops[which];
 461         memset(op, 0, sizeof (*op));
 462         op->op = opcode;
 463
 464         return op;
 465 }
 466
 467 void osd_req_op_init(struct ceph_osd_request *osd_req,
 468                                 unsigned int which, u16 opcode)
 469 {
 470         (void)_osd_req_op_init(osd_req, which, opcode);
 471 }
 472 EXPORT_SYMBOL(osd_req_op_init);
 473
 474 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 475                                 unsigned int which, u16 opcode,
 476                                 u64 offset, u64 length,
 477                                 u64 truncate_size, u32 truncate_seq)
 478 {
 479         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 480         size_t payload_len = 0;
 481
 482         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 483                opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE);
 484
 485         op->extent.offset = offset;
 486         op->extent.length = length;
 487         op->extent.truncate_size = truncate_size;
 488         op->extent.truncate_seq = truncate_seq;
 489         if (opcode == CEPH_OSD_OP_WRITE)
 490                 payload_len += length;
 491
 492         op->payload_len = payload_len;
 493 }
 494 EXPORT_SYMBOL(osd_req_op_extent_init);
 495
 496 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 497                                 unsigned int which, u64 length)
 498 {
 499         struct ceph_osd_req_op *op;
 500         u64 previous;
 501
 502         BUG_ON(which >= osd_req->r_num_ops);
 503         op = &osd_req->r_ops[which];
 504         previous = op->extent.length;
 505
 506         if (length == previous)
 507                 return;         /* Nothing to do */
 508         BUG_ON(length > previous);
 509
 510         op->extent.length = length;
 511         op->payload_len -= previous - length;
 512 }
 513 EXPORT_SYMBOL(osd_req_op_extent_update);
 514
 515 void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
 516                         u16 opcode, const char *class, const char *method)
 517 {
 518         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 519         struct ceph_pagelist *pagelist;
 520         size_t payload_len = 0;
 521         size_t size;
 522
 523         BUG_ON(opcode != CEPH_OSD_OP_CALL);
 524
 525         pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS);
 526         BUG_ON(!pagelist);
 527         ceph_pagelist_init(pagelist);
 528
 529         op->cls.class_name = class;
 530         size = strlen(class);
 531         BUG_ON(size > (size_t) U8_MAX);
 532         op->cls.class_len = size;
 533         ceph_pagelist_append(pagelist, class, size);
 534         payload_len += size;
 535
 536         op->cls.method_name = method;
 537         size = strlen(method);
 538         BUG_ON(size > (size_t) U8_MAX);
 539         op->cls.method_len = size;
 540         ceph_pagelist_append(pagelist, method, size);
 541         payload_len += size;
 542
 543         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
 544
 545         op->cls.argc = 0;       /* currently unused */
 546
 547         op->payload_len = payload_len;
 548 }
 549 EXPORT_SYMBOL(osd_req_op_cls_init);
 550
 551 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 552                           u16 opcode, const char *name, const void *value,
 553                           size_t size, u8 cmp_op, u8 cmp_mode)
 554 {
 555         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 556         struct ceph_pagelist *pagelist;
 557         size_t payload_len;
 558
 559         BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
 560
 561         pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
 562         if (!pagelist)
 563                 return -ENOMEM;
 564
 565         ceph_pagelist_init(pagelist);
 566
 567         payload_len = strlen(name);
 568         op->xattr.name_len = payload_len;
 569         ceph_pagelist_append(pagelist, name, payload_len);
 570
 571         op->xattr.value_len = size;
 572         ceph_pagelist_append(pagelist, value, size);
 573         payload_len += size;
 574
 575         op->xattr.cmp_op = cmp_op;
 576         op->xattr.cmp_mode = cmp_mode;
 577
 578         ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
 579         op->payload_len = payload_len;
 580         return 0;
 581 }
 582 EXPORT_SYMBOL(osd_req_op_xattr_init);
 583
 584 void osd_req_op_watch_init(struct ceph_osd_request *osd_req,
 585                                 unsigned int which, u16 opcode,
 586                                 u64 cookie, u64 version, int flag)
 587 {
 588         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, opcode);
 589
 590         BUG_ON(opcode != CEPH_OSD_OP_NOTIFY_ACK && opcode != CEPH_OSD_OP_WATCH);
 591
 592         op->watch.cookie = cookie;
 593         op->watch.ver = version;
 594         if (opcode == CEPH_OSD_OP_WATCH && flag)
 595                 op->watch.flag = (u8)1;
 596 }
 597 EXPORT_SYMBOL(osd_req_op_watch_init);
 598
 599 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 600                                 unsigned int which,
 601                                 u64 expected_object_size,
 602                                 u64 expected_write_size)
 603 {
 604         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 605                                                       CEPH_OSD_OP_SETALLOCHINT);
 606
 607         op->alloc_hint.expected_object_size = expected_object_size;
 608         op->alloc_hint.expected_write_size = expected_write_size;
 609
 610         /*
 611          * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
 612          * not worth a feature bit.  Set FAILOK per-op flag to make
 613          * sure older osds don't trip over an unsupported opcode.
 614          */
 615         op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
 616 }
 617 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
 618
 619 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
 620                                 struct ceph_osd_data *osd_data)
 621 {
 622         u64 length = ceph_osd_data_length(osd_data);
 623
 624         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 625                 BUG_ON(length > (u64) SIZE_MAX);
 626                 if (length)
 627                         ceph_msg_data_add_pages(msg, osd_data->pages,
 628                                         length, osd_data->alignment);
 629         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 630                 BUG_ON(!length);
 631                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
 632 #ifdef CONFIG_BLOCK
 633         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
 634                 ceph_msg_data_add_bio(msg, osd_data->bio, length);
 635 #endif
 636         } else {
 637                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
 638         }
 639 }
 640
 641 static u64 osd_req_encode_op(struct ceph_osd_request *req,
 642                               struct ceph_osd_op *dst, unsigned int which)
 643 {
 644         struct ceph_osd_req_op *src;
 645         struct ceph_osd_data *osd_data;
 646         u64 request_data_len = 0;
 647         u64 data_length;
 648
 649         BUG_ON(which >= req->r_num_ops);
 650         src = &req->r_ops[which];
 651         if (WARN_ON(!osd_req_opcode_valid(src->op))) {
 652                 pr_err("unrecognized osd opcode %d\n", src->op);
 653
 654                 return 0;
 655         }
 656
 657         switch (src->op) {
 658         case CEPH_OSD_OP_STAT:
 659                 osd_data = &src->raw_data_in;
 660                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 661                 break;
 662         case CEPH_OSD_OP_READ:
 663         case CEPH_OSD_OP_WRITE:
 664         case CEPH_OSD_OP_ZERO:
 665         case CEPH_OSD_OP_TRUNCATE:
 666                 if (src->op == CEPH_OSD_OP_WRITE)
 667                         request_data_len = src->extent.length;
 668                 dst->extent.offset = cpu_to_le64(src->extent.offset);
 669                 dst->extent.length = cpu_to_le64(src->extent.length);
 670                 dst->extent.truncate_size =
 671                         cpu_to_le64(src->extent.truncate_size);
 672                 dst->extent.truncate_seq =
 673                         cpu_to_le32(src->extent.truncate_seq);
 674                 osd_data = &src->extent.osd_data;
 675                 if (src->op == CEPH_OSD_OP_WRITE)
 676                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 677                 else
 678                         ceph_osdc_msg_data_add(req->r_reply, osd_data);
 679                 break;
 680         case CEPH_OSD_OP_CALL:
 681                 dst->cls.class_len = src->cls.class_len;
 682                 dst->cls.method_len = src->cls.method_len;
 683                 osd_data = &src->cls.request_info;
 684                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 685                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGELIST);
 686                 request_data_len = osd_data->pagelist->length;
 687
 688                 osd_data = &src->cls.request_data;
 689                 data_length = ceph_osd_data_length(osd_data);
 690                 if (data_length) {
 691                         BUG_ON(osd_data->type == CEPH_OSD_DATA_TYPE_NONE);
 692                         dst->cls.indata_len = cpu_to_le32(data_length);
 693                         ceph_osdc_msg_data_add(req->r_request, osd_data);
 694                         src->payload_len += data_length;
 695                         request_data_len += data_length;
 696                 }
 697                 osd_data = &src->cls.response_data;
 698                 ceph_osdc_msg_data_add(req->r_reply, osd_data);
 699                 break;
 700         case CEPH_OSD_OP_STARTSYNC:
 701                 break;
 702         case CEPH_OSD_OP_NOTIFY_ACK:
 703         case CEPH_OSD_OP_WATCH:
 704                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 705                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 706                 dst->watch.flag = src->watch.flag;
 707                 break;
 708         case CEPH_OSD_OP_SETALLOCHINT:
 709                 dst->alloc_hint.expected_object_size =
 710                     cpu_to_le64(src->alloc_hint.expected_object_size);
 711                 dst->alloc_hint.expected_write_size =
 712                     cpu_to_le64(src->alloc_hint.expected_write_size);
 713                 break;
 714         case CEPH_OSD_OP_SETXATTR:
 715         case CEPH_OSD_OP_CMPXATTR:
 716                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 717                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 718                 dst->xattr.cmp_op = src->xattr.cmp_op;
 719                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 720                 osd_data = &src->xattr.osd_data;
 721                 ceph_osdc_msg_data_add(req->r_request, osd_data);
 722                 request_data_len = osd_data->pagelist->length;
 723                 break;
 724         case CEPH_OSD_OP_CREATE:
 725         case CEPH_OSD_OP_DELETE:
 726                 break;
 727         default:
 728                 pr_err("unsupported osd opcode %s\n",
 729                         ceph_osd_op_name(src->op));
 730                 WARN_ON(1);
 731
 732                 return 0;
 733         }
 734
 735         dst->op = cpu_to_le16(src->op);
 736         dst->flags = cpu_to_le32(src->flags);
 737         dst->payload_len = cpu_to_le32(src->payload_len);
 738
 739         return request_data_len;
 740 }
 741
 742 /*
 743  * build new request AND message, calculate layout, and adjust file
 744  * extent as needed.
 745  *
 746  * if the file was recently truncated, we include information about its
 747  * old and new size so that the object can be updated appropriately.  (we
 748  * avoid synchronously deleting truncated objects because it's slow.)
 749  *
 750  * if @do_sync, include a 'startsync' command so that the osd will flush
 751  * data quickly.
 752  */
 753 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 754                                                struct ceph_file_layout *layout,
 755                                                struct ceph_vino vino,
 756                                                u64 off, u64 *plen,
 757                                                unsigned int which, int num_ops,
 758                                                int opcode, int flags,
 759                                                struct ceph_snap_context *snapc,
 760                                                u32 truncate_seq,
 761                                                u64 truncate_size,
 762                                                bool use_mempool)
 763 {
 764         struct ceph_osd_request *req;
 765         u64 objnum = 0;
 766         u64 objoff = 0;
 767         u64 objlen = 0;
 768         int r;
 769
 770         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 771                opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
 772                opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
 773
 774         req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
 775                                         GFP_NOFS);
 776         if (!req)
 777                 return ERR_PTR(-ENOMEM);
 778
 779         req->r_flags = flags;
 780
 781         /* calculate max write size */
 782         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
 783         if (r < 0) {
 784                 ceph_osdc_put_request(req);
 785                 return ERR_PTR(r);
 786         }
 787
 788         if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
 789                 osd_req_op_init(req, which, opcode);
 790         } else {
 791                 u32 object_size = le32_to_cpu(layout->fl_object_size);
 792                 u32 object_base = off - objoff;
 793                 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
 794                         if (truncate_size <= object_base) {
 795                                 truncate_size = 0;
 796                         } else {
 797                                 truncate_size -= object_base;
 798                                 if (truncate_size > object_size)
 799                                         truncate_size = object_size;
 800                         }
 801                 }
 802                 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
 803                                        truncate_size, truncate_seq);
 804         }
 805
 806         req->r_base_oloc.pool = ceph_file_layout_pg_pool(*layout);
 807
 808         snprintf(req->r_base_oid.name, sizeof(req->r_base_oid.name),
 809                  "%llx.%08llx", vino.ino, objnum);
 810         req->r_base_oid.name_len = strlen(req->r_base_oid.name);
 811
 812         return req;
 813 }
 814 EXPORT_SYMBOL(ceph_osdc_new_request);
 815
 816 /*
 817  * We keep osd requests in an rbtree, sorted by ->r_tid.
 818  */
 819 static void __insert_request(struct ceph_osd_client *osdc,
 820                              struct ceph_osd_request *new)
 821 {
 822         struct rb_node **p = &osdc->requests.rb_node;
 823         struct rb_node *parent = NULL;
 824         struct ceph_osd_request *req = NULL;
 825
 826         while (*p) {
 827                 parent = *p;
 828                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 829                 if (new->r_tid < req->r_tid)
 830                         p = &(*p)->rb_left;
 831                 else if (new->r_tid > req->r_tid)
 832                         p = &(*p)->rb_right;
 833                 else
 834                         BUG();
 835         }
 836
 837         rb_link_node(&new->r_node, parent, p);
 838         rb_insert_color(&new->r_node, &osdc->requests);
 839 }
 840
 841 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 842                                                  u64 tid)
 843 {
 844         struct ceph_osd_request *req;
 845         struct rb_node *n = osdc->requests.rb_node;
 846
 847         while (n) {
 848                 req = rb_entry(n, struct ceph_osd_request, r_node);
 849                 if (tid < req->r_tid)
 850                         n = n->rb_left;
 851                 else if (tid > req->r_tid)
 852                         n = n->rb_right;
 853                 else
 854                         return req;
 855         }
 856         return NULL;
 857 }
 858
 859 static struct ceph_osd_request *
 860 __lookup_request_ge(struct ceph_osd_client *osdc,
 861                     u64 tid)
 862 {
 863         struct ceph_osd_request *req;
 864         struct rb_node *n = osdc->requests.rb_node;
 865
 866         while (n) {
 867                 req = rb_entry(n, struct ceph_osd_request, r_node);
 868                 if (tid < req->r_tid) {
 869                         if (!n->rb_left)
 870                                 return req;
 871                         n = n->rb_left;
 872                 } else if (tid > req->r_tid) {
 873                         n = n->rb_right;
 874                 } else {
 875                         return req;
 876                 }
 877         }
 878         return NULL;
 879 }
 880
 881 static void __kick_linger_request(struct ceph_osd_request *req)
 882 {
 883         struct ceph_osd_client *osdc = req->r_osdc;
 884         struct ceph_osd *osd = req->r_osd;
 885
 886         /*
 887          * Linger requests need to be resent with a new tid to avoid
 888          * the dup op detection logic on the OSDs.  Achieve this with
 889          * a re-register dance instead of open-coding.
 890          */
 891         ceph_osdc_get_request(req);
 892         if (!list_empty(&req->r_linger_item))
 893                 __unregister_linger_request(osdc, req);
 894         else
 895                 __unregister_request(osdc, req);
 896         __register_request(osdc, req);
 897         ceph_osdc_put_request(req);
 898
 899         /*
 900          * Unless request has been registered as both normal and
 901          * lingering, __unregister{,_linger}_request clears r_osd.
 902          * However, here we need to preserve r_osd to make sure we
 903          * requeue on the same OSD.
 904          */
 905         WARN_ON(req->r_osd || !osd);
 906         req->r_osd = osd;
 907
 908         dout("%s requeueing %p tid %llu\n", __func__, req, req->r_tid);
 909         __enqueue_request(req);
 910 }
 911
 912 /*
 913  * Resubmit requests pending on the given osd.
 914  */
 915 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 916                                 struct ceph_osd *osd)
 917 {
 918         struct ceph_osd_request *req, *nreq;
 919         LIST_HEAD(resend);
 920         LIST_HEAD(resend_linger);
 921         int err;
 922
 923         dout("%s osd%d\n", __func__, osd->o_osd);
 924         err = __reset_osd(osdc, osd);
 925         if (err)
 926                 return;
 927
 928         /*
 929          * Build up a list of requests to resend by traversing the
 930          * osd's list of requests.  Requests for a given object are
 931          * sent in tid order, and that is also the order they're
 932          * kept on this list.  Therefore all requests that are in
 933          * flight will be found first, followed by all requests that
 934          * have not yet been sent.  And to resend requests while
 935          * preserving this order we will want to put any sent
 936          * requests back on the front of the osd client's unsent
 937          * list.
 938          *
 939          * So we build a separate ordered list of already-sent
 940          * requests for the affected osd and splice it onto the
 941          * front of the osd client's unsent list.  Once we've seen a
 942          * request that has not yet been sent we're done.  Those
 943          * requests are already sitting right where they belong.
 944          */
 945         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 946                 if (!req->r_sent)
 947                         break;
 948
 949                 if (!req->r_linger) {
 950                         dout("%s requeueing %p tid %llu\n", __func__, req,
 951                              req->r_tid);
 952                         list_move_tail(&req->r_req_lru_item, &resend);
 953                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 954                 } else {
 955                         list_move_tail(&req->r_req_lru_item, &resend_linger);
 956                 }
 957         }
 958         list_splice(&resend, &osdc->req_unsent);
 959
 960         /*
 961          * Both registered and not yet registered linger requests are
 962          * enqueued with a new tid on the same OSD.  We add/move them
 963          * to req_unsent/o_requests at the end to keep things in tid
 964          * order.
 965          */
 966         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 967                                  r_linger_osd_item) {
 968                 WARN_ON(!list_empty(&req->r_req_lru_item));
 969                 __kick_linger_request(req);
 970         }
 971
 972         list_for_each_entry_safe(req, nreq, &resend_linger, r_req_lru_item)
 973                 __kick_linger_request(req);
 974 }
 975
 976 /*
 977  * If the osd connection drops, we need to resubmit all requests.
 978  */
 979 static void osd_reset(struct ceph_connection *con)
 980 {
 981         struct ceph_osd *osd = con->private;
 982         struct ceph_osd_client *osdc;
 983
 984         if (!osd)
 985                 return;
 986         dout("osd_reset osd%d\n", osd->o_osd);
 987         osdc = osd->o_osdc;
 988         down_read(&osdc->map_sem);
 989         mutex_lock(&osdc->request_mutex);
 990         __kick_osd_requests(osdc, osd);
 991         __send_queued(osdc);
 992         mutex_unlock(&osdc->request_mutex);
 993         up_read(&osdc->map_sem);
 994 }
 995
 996 /*
 997  * Track open sessions with osds.
 998  */
 999 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1000 {
1001         struct ceph_osd *osd;
1002
1003         osd = kzalloc(sizeof(*osd), GFP_NOFS);
1004         if (!osd)
1005                 return NULL;
1006
1007         atomic_set(&osd->o_ref, 1);
1008         osd->o_osdc = osdc;
1009         osd->o_osd = onum;
1010         RB_CLEAR_NODE(&osd->o_node);
1011         INIT_LIST_HEAD(&osd->o_requests);
1012         INIT_LIST_HEAD(&osd->o_linger_requests);
1013         INIT_LIST_HEAD(&osd->o_osd_lru);
1014         osd->o_incarnation = 1;
1015
1016         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1017
1018         INIT_LIST_HEAD(&osd->o_keepalive_item);
1019         return osd;
1020 }
1021
1022 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1023 {
1024         if (atomic_inc_not_zero(&osd->o_ref)) {
1025                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
1026                      atomic_read(&osd->o_ref));
1027                 return osd;
1028         } else {
1029                 dout("get_osd %p FAIL\n", osd);
1030                 return NULL;
1031         }
1032 }
1033
1034 static void put_osd(struct ceph_osd *osd)
1035 {
1036         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
1037              atomic_read(&osd->o_ref) - 1);
1038         if (atomic_dec_and_test(&osd->o_ref)) {
1039                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
1040
1041                 if (osd->o_auth.authorizer)
1042                         ceph_auth_destroy_authorizer(ac, osd->o_auth.authorizer);
1043                 kfree(osd);
1044         }
1045 }
1046
1047 /*
1048  * remove an osd from our map
1049  */
1050 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1051 {
1052         dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1053         WARN_ON(!list_empty(&osd->o_requests));
1054         WARN_ON(!list_empty(&osd->o_linger_requests));
1055
1056         list_del_init(&osd->o_osd_lru);
1057         rb_erase(&osd->o_node, &osdc->osds);
1058         RB_CLEAR_NODE(&osd->o_node);
1059 }
1060
1061 static void remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1062 {
1063         dout("%s %p osd%d\n", __func__, osd, osd->o_osd);
1064
1065         if (!RB_EMPTY_NODE(&osd->o_node)) {
1066                 ceph_con_close(&osd->o_con);
1067                 __remove_osd(osdc, osd);
1068                 put_osd(osd);
1069         }
1070 }
1071
1072 static void remove_all_osds(struct ceph_osd_client *osdc)
1073 {
1074         dout("%s %p\n", __func__, osdc);
1075         mutex_lock(&osdc->request_mutex);
1076         while (!RB_EMPTY_ROOT(&osdc->osds)) {
1077                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
1078                                                 struct ceph_osd, o_node);
1079                 remove_osd(osdc, osd);
1080         }
1081         mutex_unlock(&osdc->request_mutex);
1082 }
1083
1084 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
1085                               struct ceph_osd *osd)
1086 {
1087         dout("%s %p\n", __func__, osd);
1088         BUG_ON(!list_empty(&osd->o_osd_lru));
1089
1090         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1091         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
1092 }
1093
1094 static void maybe_move_osd_to_lru(struct ceph_osd_client *osdc,
1095                                   struct ceph_osd *osd)
1096 {
1097         dout("%s %p\n", __func__, osd);
1098
1099         if (list_empty(&osd->o_requests) &&
1100             list_empty(&osd->o_linger_requests))
1101                 __move_osd_to_lru(osdc, osd);
1102 }
1103
1104 static void __remove_osd_from_lru(struct ceph_osd *osd)
1105 {
1106         dout("__remove_osd_from_lru %p\n", osd);
1107         if (!list_empty(&osd->o_osd_lru))
1108                 list_del_init(&osd->o_osd_lru);
1109 }
1110
1111 static void remove_old_osds(struct ceph_osd_client *osdc)
1112 {
1113         struct ceph_osd *osd, *nosd;
1114
1115         dout("__remove_old_osds %p\n", osdc);
1116         mutex_lock(&osdc->request_mutex);
1117         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
1118                 if (time_before(jiffies, osd->lru_ttl))
1119                         break;
1120                 remove_osd(osdc, osd);
1121         }
1122         mutex_unlock(&osdc->request_mutex);
1123 }
1124
1125 /*
1126  * reset osd connect
1127  */
1128 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
1129 {
1130         struct ceph_entity_addr *peer_addr;
1131
1132         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
1133         if (list_empty(&osd->o_requests) &&
1134             list_empty(&osd->o_linger_requests)) {
1135                 remove_osd(osdc, osd);
1136                 return -ENODEV;
1137         }
1138
1139         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
1140         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1141                         !ceph_con_opened(&osd->o_con)) {
1142                 struct ceph_osd_request *req;
1143
1144                 dout("osd addr hasn't changed and connection never opened, "
1145                      "letting msgr retry\n");
1146                 /* touch each r_stamp for handle_timeout()'s benfit */
1147                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
1148                         req->r_stamp = jiffies;
1149
1150                 return -EAGAIN;
1151         }
1152
1153         ceph_con_close(&osd->o_con);
1154         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1155         osd->o_incarnation++;
1156
1157         return 0;
1158 }
1159
1160 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
1161 {
1162         struct rb_node **p = &osdc->osds.rb_node;
1163         struct rb_node *parent = NULL;
1164         struct ceph_osd *osd = NULL;
1165
1166         dout("__insert_osd %p osd%d\n", new, new->o_osd);
1167         while (*p) {
1168                 parent = *p;
1169                 osd = rb_entry(parent, struct ceph_osd, o_node);
1170                 if (new->o_osd < osd->o_osd)
1171                         p = &(*p)->rb_left;
1172                 else if (new->o_osd > osd->o_osd)
1173                         p = &(*p)->rb_right;
1174                 else
1175                         BUG();
1176         }
1177
1178         rb_link_node(&new->o_node, parent, p);
1179         rb_insert_color(&new->o_node, &osdc->osds);
1180 }
1181
1182 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
1183 {
1184         struct ceph_osd *osd;
1185         struct rb_node *n = osdc->osds.rb_node;
1186
1187         while (n) {
1188                 osd = rb_entry(n, struct ceph_osd, o_node);
1189                 if (o < osd->o_osd)
1190                         n = n->rb_left;
1191                 else if (o > osd->o_osd)
1192                         n = n->rb_right;
1193                 else
1194                         return osd;
1195         }
1196         return NULL;
1197 }
1198
1199 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
1200 {
1201         schedule_delayed_work(&osdc->timeout_work,
1202                         osdc->client->options->osd_keepalive_timeout * HZ);
1203 }
1204
1205 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
1206 {
1207         cancel_delayed_work(&osdc->timeout_work);
1208 }
1209
1210 /*
1211  * Register request, assign tid.  If this is the first request, set up
1212  * the timeout event.
1213  */
1214 static void __register_request(struct ceph_osd_client *osdc,
1215                                struct ceph_osd_request *req)
1216 {
1217         req->r_tid = ++osdc->last_tid;
1218         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
1219         dout("__register_request %p tid %lld\n", req, req->r_tid);
1220         __insert_request(osdc, req);
1221         ceph_osdc_get_request(req);
1222         osdc->num_requests++;
1223         if (osdc->num_requests == 1) {
1224                 dout(" first request, scheduling timeout\n");
1225                 __schedule_osd_timeout(osdc);
1226         }
1227 }
1228
1229 /*
1230  * called under osdc->request_mutex
1231  */
1232 static void __unregister_request(struct ceph_osd_client *osdc,
1233                                  struct ceph_osd_request *req)
1234 {
1235         if (RB_EMPTY_NODE(&req->r_node)) {
1236                 dout("__unregister_request %p tid %lld not registered\n",
1237                         req, req->r_tid);
1238                 return;
1239         }
1240
1241         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1242         rb_erase(&req->r_node, &osdc->requests);
1243         RB_CLEAR_NODE(&req->r_node);
1244         osdc->num_requests--;
1245
1246         if (req->r_osd) {
1247                 /* make sure the original request isn't in flight. */
1248                 ceph_msg_revoke(req->r_request);
1249
1250                 list_del_init(&req->r_osd_item);
1251                 maybe_move_osd_to_lru(osdc, req->r_osd);
1252                 if (list_empty(&req->r_linger_osd_item))
1253                         req->r_osd = NULL;
1254         }
1255
1256         list_del_init(&req->r_req_lru_item);
1257         ceph_osdc_put_request(req);
1258
1259         if (osdc->num_requests == 0) {
1260                 dout(" no requests, canceling timeout\n");
1261                 __cancel_osd_timeout(osdc);
1262         }
1263 }
1264
1265 /*
1266  * Cancel a previously queued request message
1267  */
1268 static void __cancel_request(struct ceph_osd_request *req)
1269 {
1270         if (req->r_sent && req->r_osd) {
1271                 ceph_msg_revoke(req->r_request);
1272                 req->r_sent = 0;
1273         }
1274 }
1275
1276 static void __register_linger_request(struct ceph_osd_client *osdc,
1277                                     struct ceph_osd_request *req)
1278 {
1279         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1280         WARN_ON(!req->r_linger);
1281
1282         ceph_osdc_get_request(req);
1283         list_add_tail(&req->r_linger_item, &osdc->req_linger);
1284         if (req->r_osd)
1285                 list_add_tail(&req->r_linger_osd_item,
1286                               &req->r_osd->o_linger_requests);
1287 }
1288
1289 static void __unregister_linger_request(struct ceph_osd_client *osdc,
1290                                         struct ceph_osd_request *req)
1291 {
1292         WARN_ON(!req->r_linger);
1293
1294         if (list_empty(&req->r_linger_item)) {
1295                 dout("%s %p tid %llu not registered\n", __func__, req,
1296                      req->r_tid);
1297                 return;
1298         }
1299
1300         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
1301         list_del_init(&req->r_linger_item);
1302
1303         if (req->r_osd) {
1304                 list_del_init(&req->r_linger_osd_item);
1305                 maybe_move_osd_to_lru(osdc, req->r_osd);
1306                 if (list_empty(&req->r_osd_item))
1307                         req->r_osd = NULL;
1308         }
1309
1310         list_del_init(&req->r_req_lru_item); /* can be on notarget */
1311         ceph_osdc_put_request(req);
1312 }
1313
1314 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
1315                                   struct ceph_osd_request *req)
1316 {
1317         if (!req->r_linger) {
1318                 dout("set_request_linger %p\n", req);
1319                 req->r_linger = 1;
1320         }
1321 }
1322 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
1323
1324 /*
1325  * Returns whether a request should be blocked from being sent
1326  * based on the current osdmap and osd_client settings.
1327  *
1328  * Caller should hold map_sem for read.
1329  */
1330 static bool __req_should_be_paused(struct ceph_osd_client *osdc,
1331                                    struct ceph_osd_request *req)
1332 {
1333         bool pauserd = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD);
1334         bool pausewr = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR) ||
1335                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
1336         return (req->r_flags & CEPH_OSD_FLAG_READ && pauserd) ||
1337                 (req->r_flags & CEPH_OSD_FLAG_WRITE && pausewr);
1338 }
1339
1340 /*
1341  * Calculate mapping of a request to a PG.  Takes tiering into account.
1342  */
1343 static int __calc_request_pg(struct ceph_osdmap *osdmap,
1344                              struct ceph_osd_request *req,
1345                              struct ceph_pg *pg_out)
1346 {
1347         bool need_check_tiering;
1348
1349         need_check_tiering = false;
1350         if (req->r_target_oloc.pool == -1) {
1351                 req->r_target_oloc = req->r_base_oloc; /* struct */
1352                 need_check_tiering = true;
1353         }
1354         if (req->r_target_oid.name_len == 0) {
1355                 ceph_oid_copy(&req->r_target_oid, &req->r_base_oid);
1356                 need_check_tiering = true;
1357         }
1358
1359         if (need_check_tiering &&
1360             (req->r_flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1361                 struct ceph_pg_pool_info *pi;
1362
1363                 pi = ceph_pg_pool_by_id(osdmap, req->r_target_oloc.pool);
1364                 if (pi) {
1365                         if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
1366                             pi->read_tier >= 0)
1367                                 req->r_target_oloc.pool = pi->read_tier;
1368                         if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
1369                             pi->write_tier >= 0)
1370                                 req->r_target_oloc.pool = pi->write_tier;
1371                 }
1372                 /* !pi is caught in ceph_oloc_oid_to_pg() */
1373         }
1374
1375         return ceph_oloc_oid_to_pg(osdmap, &req->r_target_oloc,
1376                                    &req->r_target_oid, pg_out);
1377 }
1378
1379 static void __enqueue_request(struct ceph_osd_request *req)
1380 {
1381         struct ceph_osd_client *osdc = req->r_osdc;
1382
1383         dout("%s %p tid %llu to osd%d\n", __func__, req, req->r_tid,
1384              req->r_osd ? req->r_osd->o_osd : -1);
1385
1386         if (req->r_osd) {
1387                 __remove_osd_from_lru(req->r_osd);
1388                 list_add_tail(&req->r_osd_item, &req->r_osd->o_requests);
1389                 list_move_tail(&req->r_req_lru_item, &osdc->req_unsent);
1390         } else {
1391                 list_move_tail(&req->r_req_lru_item, &osdc->req_notarget);
1392         }
1393 }
1394
1395 /*
1396  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
1397  * (as needed), and set the request r_osd appropriately.  If there is
1398  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
1399  * (unsent, homeless) or leave on in-flight lru.
1400  *
1401  * Return 0 if unchanged, 1 if changed, or negative on error.
1402  *
1403  * Caller should hold map_sem for read and request_mutex.
1404  */
1405 static int __map_request(struct ceph_osd_client *osdc,
1406                          struct ceph_osd_request *req, int force_resend)
1407 {
1408         struct ceph_pg pgid;
1409         int acting[CEPH_PG_MAX_SIZE];
1410         int num, o;
1411         int err;
1412         bool was_paused;
1413
1414         dout("map_request %p tid %lld\n", req, req->r_tid);
1415
1416         err = __calc_request_pg(osdc->osdmap, req, &pgid);
1417         if (err) {
1418                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1419                 return err;
1420         }
1421         req->r_pgid = pgid;
1422
1423         num = ceph_calc_pg_acting(osdc->osdmap, pgid, acting, &o);
1424         if (num < 0)
1425                 num = 0;
1426
1427         was_paused = req->r_paused;
1428         req->r_paused = __req_should_be_paused(osdc, req);
1429         if (was_paused && !req->r_paused)
1430                 force_resend = 1;
1431
1432         if ((!force_resend &&
1433              req->r_osd && req->r_osd->o_osd == o &&
1434              req->r_sent >= req->r_osd->o_incarnation &&
1435              req->r_num_pg_osds == num &&
1436              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1437             (req->r_osd == NULL && o == -1) ||
1438             req->r_paused)
1439                 return 0;  /* no change */
1440
1441         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
1442              req->r_tid, pgid.pool, pgid.seed, o,
1443              req->r_osd ? req->r_osd->o_osd : -1);
1444
1445         /* record full pg acting set */
1446         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1447         req->r_num_pg_osds = num;
1448
1449         if (req->r_osd) {
1450                 __cancel_request(req);
1451                 list_del_init(&req->r_osd_item);
1452                 list_del_init(&req->r_linger_osd_item);
1453                 req->r_osd = NULL;
1454         }
1455
1456         req->r_osd = __lookup_osd(osdc, o);
1457         if (!req->r_osd && o >= 0) {
1458                 err = -ENOMEM;
1459                 req->r_osd = create_osd(osdc, o);
1460                 if (!req->r_osd) {
1461                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1462                         goto out;
1463                 }
1464
1465                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1466                 __insert_osd(osdc, req->r_osd);
1467
1468                 ceph_con_open(&req->r_osd->o_con,
1469                               CEPH_ENTITY_TYPE_OSD, o,
1470                               &osdc->osdmap->osd_addr[o]);
1471         }
1472
1473         __enqueue_request(req);
1474         err = 1;   /* osd or pg changed */
1475
1476 out:
1477         return err;
1478 }
1479
1480 /*
1481  * caller should hold map_sem (for read) and request_mutex
1482  */
1483 static void __send_request(struct ceph_osd_client *osdc,
1484                            struct ceph_osd_request *req)
1485 {
1486         void *p;
1487
1488         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1489              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1490              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1491
1492         /* fill in message content that changes each time we send it */
1493         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1494         put_unaligned_le32(req->r_flags, req->r_request_flags);
1495         put_unaligned_le64(req->r_target_oloc.pool, req->r_request_pool);
1496         p = req->r_request_pgid;
1497         ceph_encode_64(&p, req->r_pgid.pool);
1498         ceph_encode_32(&p, req->r_pgid.seed);
1499         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1500         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1501                sizeof(req->r_reassert_version));
1502
1503         req->r_stamp = jiffies;
1504         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1505
1506         ceph_msg_get(req->r_request); /* send consumes a ref */
1507
1508         req->r_sent = req->r_osd->o_incarnation;
1509
1510         ceph_con_send(&req->r_osd->o_con, req->r_request);
1511 }
1512
1513 /*
1514  * Send any requests in the queue (req_unsent).
1515  */
1516 static void __send_queued(struct ceph_osd_client *osdc)
1517 {
1518         struct ceph_osd_request *req, *tmp;
1519
1520         dout("__send_queued\n");
1521         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1522                 __send_request(osdc, req);
1523 }
1524
1525 /*
1526  * Caller should hold map_sem for read and request_mutex.
1527  */
1528 static int __ceph_osdc_start_request(struct ceph_osd_client *osdc,
1529                                      struct ceph_osd_request *req,
1530                                      bool nofail)
1531 {
1532         int rc;
1533
1534         __register_request(osdc, req);
1535         req->r_sent = 0;
1536         req->r_got_reply = 0;
1537         rc = __map_request(osdc, req, 0);
1538         if (rc < 0) {
1539                 if (nofail) {
1540                         dout("osdc_start_request failed map, "
1541                                 " will retry %lld\n", req->r_tid);
1542                         rc = 0;
1543                 } else {
1544                         __unregister_request(osdc, req);
1545                 }
1546                 return rc;
1547         }
1548
1549         if (req->r_osd == NULL) {
1550                 dout("send_request %p no up osds in pg\n", req);
1551                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1552         } else {
1553                 __send_queued(osdc);
1554         }
1555
1556         return 0;
1557 }
1558
1559 /*
1560  * Timeout callback, called every N seconds when 1 or more osd
1561  * requests has been active for more than N seconds.  When this
1562  * happens, we ping all OSDs with requests who have timed out to
1563  * ensure any communications channel reset is detected.  Reset the
1564  * request timeouts another N seconds in the future as we go.
1565  * Reschedule the timeout event another N seconds in future (unless
1566  * there are no open requests).
1567  */
1568 static void handle_timeout(struct work_struct *work)
1569 {
1570         struct ceph_osd_client *osdc =
1571                 container_of(work, struct ceph_osd_client, timeout_work.work);
1572         struct ceph_osd_request *req;
1573         struct ceph_osd *osd;
1574         unsigned long keepalive =
1575                 osdc->client->options->osd_keepalive_timeout * HZ;
1576         struct list_head slow_osds;
1577         dout("timeout\n");
1578         down_read(&osdc->map_sem);
1579
1580         ceph_monc_request_next_osdmap(&osdc->client->monc);
1581
1582         mutex_lock(&osdc->request_mutex);
1583
1584         /*
1585          * ping osds that are a bit slow.  this ensures that if there
1586          * is a break in the TCP connection we will notice, and reopen
1587          * a connection with that osd (from the fault callback).
1588          */
1589         INIT_LIST_HEAD(&slow_osds);
1590         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1591                 if (time_before(jiffies, req->r_stamp + keepalive))
1592                         break;
1593
1594                 osd = req->r_osd;
1595                 BUG_ON(!osd);
1596                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1597                      req->r_tid, osd->o_osd);
1598                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1599         }
1600         while (!list_empty(&slow_osds)) {
1601                 osd = list_entry(slow_osds.next, struct ceph_osd,
1602                                  o_keepalive_item);
1603                 list_del_init(&osd->o_keepalive_item);
1604                 ceph_con_keepalive(&osd->o_con);
1605         }
1606
1607         __schedule_osd_timeout(osdc);
1608         __send_queued(osdc);
1609         mutex_unlock(&osdc->request_mutex);
1610         up_read(&osdc->map_sem);
1611 }
1612
1613 static void handle_osds_timeout(struct work_struct *work)
1614 {
1615         struct ceph_osd_client *osdc =
1616                 container_of(work, struct ceph_osd_client,
1617                              osds_timeout_work.work);
1618         unsigned long delay =
1619                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1620
1621         dout("osds timeout\n");
1622         down_read(&osdc->map_sem);
1623         remove_old_osds(osdc);
1624         up_read(&osdc->map_sem);
1625
1626         schedule_delayed_work(&osdc->osds_timeout_work,
1627                               round_jiffies_relative(delay));
1628 }
1629
1630 static int ceph_oloc_decode(void **p, void *end,
1631                             struct ceph_object_locator *oloc)
1632 {
1633         u8 struct_v, struct_cv;
1634         u32 len;
1635         void *struct_end;
1636         int ret = 0;
1637
1638         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1639         struct_v = ceph_decode_8(p);
1640         struct_cv = ceph_decode_8(p);
1641         if (struct_v < 3) {
1642                 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
1643                         struct_v, struct_cv);
1644                 goto e_inval;
1645         }
1646         if (struct_cv > 6) {
1647                 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
1648                         struct_v, struct_cv);
1649                 goto e_inval;
1650         }
1651         len = ceph_decode_32(p);
1652         ceph_decode_need(p, end, len, e_inval);
1653         struct_end = *p + len;
1654
1655         oloc->pool = ceph_decode_64(p);
1656         *p += 4; /* skip preferred */
1657
1658         len = ceph_decode_32(p);
1659         if (len > 0) {
1660                 pr_warn("ceph_object_locator::key is set\n");
1661                 goto e_inval;
1662         }
1663
1664         if (struct_v >= 5) {
1665                 len = ceph_decode_32(p);
1666                 if (len > 0) {
1667                         pr_warn("ceph_object_locator::nspace is set\n");
1668                         goto e_inval;
1669                 }
1670         }
1671
1672         if (struct_v >= 6) {
1673                 s64 hash = ceph_decode_64(p);
1674                 if (hash != -1) {
1675                         pr_warn("ceph_object_locator::hash is set\n");
1676                         goto e_inval;
1677                 }
1678         }
1679
1680         /* skip the rest */
1681         *p = struct_end;
1682 out:
1683         return ret;
1684
1685 e_inval:
1686         ret = -EINVAL;
1687         goto out;
1688 }
1689
1690 static int ceph_redirect_decode(void **p, void *end,
1691                                 struct ceph_request_redirect *redir)
1692 {
1693         u8 struct_v, struct_cv;
1694         u32 len;
1695         void *struct_end;
1696         int ret;
1697
1698         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
1699         struct_v = ceph_decode_8(p);
1700         struct_cv = ceph_decode_8(p);
1701         if (struct_cv > 1) {
1702                 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
1703                         struct_v, struct_cv);
1704                 goto e_inval;
1705         }
1706         len = ceph_decode_32(p);
1707         ceph_decode_need(p, end, len, e_inval);
1708         struct_end = *p + len;
1709
1710         ret = ceph_oloc_decode(p, end, &redir->oloc);
1711         if (ret)
1712                 goto out;
1713
1714         len = ceph_decode_32(p);
1715         if (len > 0) {
1716                 pr_warn("ceph_request_redirect::object_name is set\n");
1717                 goto e_inval;
1718         }
1719
1720         len = ceph_decode_32(p);
1721         *p += len; /* skip osd_instructions */
1722
1723         /* skip the rest */
1724         *p = struct_end;
1725 out:
1726         return ret;
1727
1728 e_inval:
1729         ret = -EINVAL;
1730         goto out;
1731 }
1732
1733 static void complete_request(struct ceph_osd_request *req)
1734 {
1735         complete_all(&req->r_safe_completion);  /* fsync waiter */
1736 }
1737
1738 /*
1739  * handle osd op reply.  either call the callback if it is specified,
1740  * or do the completion to wake up the waiting thread.
1741  */
1742 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1743                          struct ceph_connection *con)
1744 {
1745         void *p, *end;
1746         struct ceph_osd_request *req;
1747         struct ceph_request_redirect redir;
1748         u64 tid;
1749         int object_len;
1750         unsigned int numops;
1751         int payload_len, flags;
1752         s32 result;
1753         s32 retry_attempt;
1754         struct ceph_pg pg;
1755         int err;
1756         u32 reassert_epoch;
1757         u64 reassert_version;
1758         u32 osdmap_epoch;
1759         int already_completed;
1760         u32 bytes;
1761         unsigned int i;
1762
1763         tid = le64_to_cpu(msg->hdr.tid);
1764         dout("handle_reply %p tid %llu\n", msg, tid);
1765
1766         p = msg->front.iov_base;
1767         end = p + msg->front.iov_len;
1768
1769         ceph_decode_need(&p, end, 4, bad);
1770         object_len = ceph_decode_32(&p);
1771         ceph_decode_need(&p, end, object_len, bad);
1772         p += object_len;
1773
1774         err = ceph_decode_pgid(&p, end, &pg);
1775         if (err)
1776                 goto bad;
1777
1778         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1779         flags = ceph_decode_64(&p);
1780         result = ceph_decode_32(&p);
1781         reassert_epoch = ceph_decode_32(&p);
1782         reassert_version = ceph_decode_64(&p);
1783         osdmap_epoch = ceph_decode_32(&p);
1784
1785         /* lookup */
1786         down_read(&osdc->map_sem);
1787         mutex_lock(&osdc->request_mutex);
1788         req = __lookup_request(osdc, tid);
1789         if (req == NULL) {
1790                 dout("handle_reply tid %llu dne\n", tid);
1791                 goto bad_mutex;
1792         }
1793         ceph_osdc_get_request(req);
1794
1795         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1796              req, result);
1797
1798         ceph_decode_need(&p, end, 4, bad_put);
1799         numops = ceph_decode_32(&p);
1800         if (numops > CEPH_OSD_MAX_OP)
1801                 goto bad_put;
1802         if (numops != req->r_num_ops)
1803                 goto bad_put;
1804         payload_len = 0;
1805         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad_put);
1806         for (i = 0; i < numops; i++) {
1807                 struct ceph_osd_op *op = p;
1808                 int len;
1809
1810                 len = le32_to_cpu(op->payload_len);
1811                 req->r_reply_op_len[i] = len;
1812                 dout(" op %d has %d bytes\n", i, len);
1813                 payload_len += len;
1814                 p += sizeof(*op);
1815         }
1816         bytes = le32_to_cpu(msg->hdr.data_len);
1817         if (payload_len != bytes) {
1818                 pr_warn("sum of op payload lens %d != data_len %d\n",
1819                         payload_len, bytes);
1820                 goto bad_put;
1821         }
1822
1823         ceph_decode_need(&p, end, 4 + numops * 4, bad_put);
1824         retry_attempt = ceph_decode_32(&p);
1825         for (i = 0; i < numops; i++)
1826                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1827
1828         if (le16_to_cpu(msg->hdr.version) >= 6) {
1829                 p += 8 + 4; /* skip replay_version */
1830                 p += 8; /* skip user_version */
1831
1832                 err = ceph_redirect_decode(&p, end, &redir);
1833                 if (err)
1834                         goto bad_put;
1835         } else {
1836                 redir.oloc.pool = -1;
1837         }
1838
1839         if (redir.oloc.pool != -1) {
1840                 dout("redirect pool %lld\n", redir.oloc.pool);
1841
1842                 __unregister_request(osdc, req);
1843
1844                 req->r_target_oloc = redir.oloc; /* struct */
1845
1846                 /*
1847                  * Start redirect requests with nofail=true.  If
1848                  * mapping fails, request will end up on the notarget
1849                  * list, waiting for the new osdmap (which can take
1850                  * a while), even though the original request mapped
1851                  * successfully.  In the future we might want to follow
1852                  * original request's nofail setting here.
1853                  */
1854                 err = __ceph_osdc_start_request(osdc, req, true);
1855                 BUG_ON(err);
1856
1857                 goto out_unlock;
1858         }
1859
1860         already_completed = req->r_got_reply;
1861         if (!req->r_got_reply) {
1862                 req->r_result = result;
1863                 dout("handle_reply result %d bytes %d\n", req->r_result,
1864                      bytes);
1865                 if (req->r_result == 0)
1866                         req->r_result = bytes;
1867
1868                 /* in case this is a write and we need to replay, */
1869                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1870                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1871
1872                 req->r_got_reply = 1;
1873         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1874                 dout("handle_reply tid %llu dup ack\n", tid);
1875                 goto out_unlock;
1876         }
1877
1878         dout("handle_reply tid %llu flags %d\n", tid, flags);
1879
1880         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1881                 __register_linger_request(osdc, req);
1882
1883         /* either this is a read, or we got the safe response */
1884         if (result < 0 ||
1885             (flags & CEPH_OSD_FLAG_ONDISK) ||
1886             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1887                 __unregister_request(osdc, req);
1888
1889         mutex_unlock(&osdc->request_mutex);
1890         up_read(&osdc->map_sem);
1891
1892         if (!already_completed) {
1893                 if (req->r_unsafe_callback &&
1894                     result >= 0 && !(flags & CEPH_OSD_FLAG_ONDISK))
1895                         req->r_unsafe_callback(req, true);
1896                 if (req->r_callback)
1897                         req->r_callback(req, msg);
1898                 else
1899                         complete_all(&req->r_completion);
1900         }
1901
1902         if (flags & CEPH_OSD_FLAG_ONDISK) {
1903                 if (req->r_unsafe_callback && already_completed)
1904                         req->r_unsafe_callback(req, false);
1905                 complete_request(req);
1906         }
1907
1908 out:
1909         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1910         ceph_osdc_put_request(req);
1911         return;
1912 out_unlock:
1913         mutex_unlock(&osdc->request_mutex);
1914         up_read(&osdc->map_sem);
1915         goto out;
1916
1917 bad_put:
1918         req->r_result = -EIO;
1919         __unregister_request(osdc, req);
1920         if (req->r_callback)
1921                 req->r_callback(req, msg);
1922         else
1923                 complete_all(&req->r_completion);
1924         complete_request(req);
1925         ceph_osdc_put_request(req);
1926 bad_mutex:
1927         mutex_unlock(&osdc->request_mutex);
1928         up_read(&osdc->map_sem);
1929 bad:
1930         pr_err("corrupt osd_op_reply got %d %d\n",
1931                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1932         ceph_msg_dump(msg);
1933 }
1934
1935 static void reset_changed_osds(struct ceph_osd_client *osdc)
1936 {
1937         struct rb_node *p, *n;
1938
1939         dout("%s %p\n", __func__, osdc);
1940         for (p = rb_first(&osdc->osds); p; p = n) {
1941                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1942
1943                 n = rb_next(p);
1944                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1945                     memcmp(&osd->o_con.peer_addr,
1946                            ceph_osd_addr(osdc->osdmap,
1947                                          osd->o_osd),
1948                            sizeof(struct ceph_entity_addr)) != 0)
1949                         __reset_osd(osdc, osd);
1950         }
1951 }
1952
1953 /*
1954  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1955  * no osd, request a new map.
1956  *
1957  * Caller should hold map_sem for read.
1958  */
1959 static void kick_requests(struct ceph_osd_client *osdc, bool force_resend,
1960                           bool force_resend_writes)
1961 {
1962         struct ceph_osd_request *req, *nreq;
1963         struct rb_node *p;
1964         int needmap = 0;
1965         int err;
1966         bool force_resend_req;
1967
1968         dout("kick_requests %s %s\n", force_resend ? " (force resend)" : "",
1969                 force_resend_writes ? " (force resend writes)" : "");
1970         mutex_lock(&osdc->request_mutex);
1971         for (p = rb_first(&osdc->requests); p; ) {
1972                 req = rb_entry(p, struct ceph_osd_request, r_node);
1973                 p = rb_next(p);
1974
1975                 /*
1976                  * For linger requests that have not yet been
1977                  * registered, move them to the linger list; they'll
1978                  * be sent to the osd in the loop below.  Unregister
1979                  * the request before re-registering it as a linger
1980                  * request to ensure the __map_request() below
1981                  * will decide it needs to be sent.
1982                  */
1983                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1984                         dout("%p tid %llu restart on osd%d\n",
1985                              req, req->r_tid,
1986                              req->r_osd ? req->r_osd->o_osd : -1);
1987                         ceph_osdc_get_request(req);
1988                         __unregister_request(osdc, req);
1989                         __register_linger_request(osdc, req);
1990                         ceph_osdc_put_request(req);
1991                         continue;
1992                 }
1993
1994                 force_resend_req = force_resend ||
1995                         (force_resend_writes &&
1996                                 req->r_flags & CEPH_OSD_FLAG_WRITE);
1997                 err = __map_request(osdc, req, force_resend_req);
1998                 if (err < 0)
1999                         continue;  /* error */
2000                 if (req->r_osd == NULL) {
2001                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
2002                         needmap++;  /* request a newer map */
2003                 } else if (err > 0) {
2004                         if (!req->r_linger) {
2005                                 dout("%p tid %llu requeued on osd%d\n", req,
2006                                      req->r_tid,
2007                                      req->r_osd ? req->r_osd->o_osd : -1);
2008                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2009                         }
2010                 }
2011         }
2012
2013         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
2014                                  r_linger_item) {
2015                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
2016
2017                 err = __map_request(osdc, req,
2018                                     force_resend || force_resend_writes);
2019                 dout("__map_request returned %d\n", err);
2020                 if (err == 0)
2021                         continue;  /* no change and no osd was specified */
2022                 if (err < 0)
2023                         continue;  /* hrm! */
2024                 if (req->r_osd == NULL) {
2025                         dout("tid %llu maps to no valid osd\n", req->r_tid);
2026                         needmap++;  /* request a newer map */
2027                         continue;
2028                 }
2029
2030                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
2031                      req->r_osd ? req->r_osd->o_osd : -1);
2032                 __register_request(osdc, req);
2033                 __unregister_linger_request(osdc, req);
2034         }
2035         reset_changed_osds(osdc);
2036         mutex_unlock(&osdc->request_mutex);
2037
2038         if (needmap) {
2039                 dout("%d requests for down osds, need new map\n", needmap);
2040                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2041         }
2042 }
2043
2044
2045 /*
2046  * Process updated osd map.
2047  *
2048  * The message contains any number of incremental and full maps, normally
2049  * indicating some sort of topology change in the cluster.  Kick requests
2050  * off to different OSDs as needed.
2051  */
2052 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
2053 {
2054         void *p, *end, *next;
2055         u32 nr_maps, maplen;
2056         u32 epoch;
2057         struct ceph_osdmap *newmap = NULL, *oldmap;
2058         int err;
2059         struct ceph_fsid fsid;
2060         bool was_full;
2061
2062         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
2063         p = msg->front.iov_base;
2064         end = p + msg->front.iov_len;
2065
2066         /* verify fsid */
2067         ceph_decode_need(&p, end, sizeof(fsid), bad);
2068         ceph_decode_copy(&p, &fsid, sizeof(fsid));
2069         if (ceph_check_fsid(osdc->client, &fsid) < 0)
2070                 return;
2071
2072         down_write(&osdc->map_sem);
2073
2074         was_full = ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL);
2075
2076         /* incremental maps */
2077         ceph_decode_32_safe(&p, end, nr_maps, bad);
2078         dout(" %d inc maps\n", nr_maps);
2079         while (nr_maps > 0) {
2080                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2081                 epoch = ceph_decode_32(&p);
2082                 maplen = ceph_decode_32(&p);
2083                 ceph_decode_need(&p, end, maplen, bad);
2084                 next = p + maplen;
2085                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
2086                         dout("applying incremental map %u len %d\n",
2087                              epoch, maplen);
2088                         newmap = osdmap_apply_incremental(&p, next,
2089                                                           osdc->osdmap,
2090                                                           &osdc->client->msgr);
2091                         if (IS_ERR(newmap)) {
2092                                 err = PTR_ERR(newmap);
2093                                 goto bad;
2094                         }
2095                         BUG_ON(!newmap);
2096                         if (newmap != osdc->osdmap) {
2097                                 ceph_osdmap_destroy(osdc->osdmap);
2098                                 osdc->osdmap = newmap;
2099                         }
2100                         was_full = was_full ||
2101                                 ceph_osdmap_flag(osdc->osdmap,
2102                                                  CEPH_OSDMAP_FULL);
2103                         kick_requests(osdc, 0, was_full);
2104                 } else {
2105                         dout("ignoring incremental map %u len %d\n",
2106                              epoch, maplen);
2107                 }
2108                 p = next;
2109                 nr_maps--;
2110         }
2111         if (newmap)
2112                 goto done;
2113
2114         /* full maps */
2115         ceph_decode_32_safe(&p, end, nr_maps, bad);
2116         dout(" %d full maps\n", nr_maps);
2117         while (nr_maps) {
2118                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2119                 epoch = ceph_decode_32(&p);
2120                 maplen = ceph_decode_32(&p);
2121                 ceph_decode_need(&p, end, maplen, bad);
2122                 if (nr_maps > 1) {
2123                         dout("skipping non-latest full map %u len %d\n",
2124                              epoch, maplen);
2125                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
2126                         dout("skipping full map %u len %d, "
2127                              "older than our %u\n", epoch, maplen,
2128                              osdc->osdmap->epoch);
2129                 } else {
2130                         int skipped_map = 0;
2131
2132                         dout("taking full map %u len %d\n", epoch, maplen);
2133                         newmap = ceph_osdmap_decode(&p, p+maplen);
2134                         if (IS_ERR(newmap)) {
2135                                 err = PTR_ERR(newmap);
2136                                 goto bad;
2137                         }
2138                         BUG_ON(!newmap);
2139                         oldmap = osdc->osdmap;
2140                         osdc->osdmap = newmap;
2141                         if (oldmap) {
2142                                 if (oldmap->epoch + 1 < newmap->epoch)
2143                                         skipped_map = 1;
2144                                 ceph_osdmap_destroy(oldmap);
2145                         }
2146                         was_full = was_full ||
2147                                 ceph_osdmap_flag(osdc->osdmap,
2148                                                  CEPH_OSDMAP_FULL);
2149                         kick_requests(osdc, skipped_map, was_full);
2150                 }
2151                 p += maplen;
2152                 nr_maps--;
2153         }
2154
2155         if (!osdc->osdmap)
2156                 goto bad;
2157 done:
2158         downgrade_write(&osdc->map_sem);
2159         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
2160
2161         /*
2162          * subscribe to subsequent osdmap updates if full to ensure
2163          * we find out when we are no longer full and stop returning
2164          * ENOSPC.
2165          */
2166         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) ||
2167                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSERD) ||
2168                 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_PAUSEWR))
2169                 ceph_monc_request_next_osdmap(&osdc->client->monc);
2170
2171         mutex_lock(&osdc->request_mutex);
2172         __send_queued(osdc);
2173         mutex_unlock(&osdc->request_mutex);
2174         up_read(&osdc->map_sem);
2175         wake_up_all(&osdc->client->auth_wq);
2176         return;
2177
2178 bad:
2179         pr_err("osdc handle_map corrupt msg\n");
2180         ceph_msg_dump(msg);
2181         up_write(&osdc->map_sem);
2182 }
2183
2184 /*
2185  * watch/notify callback event infrastructure
2186  *
2187  * These callbacks are used both for watch and notify operations.
2188  */
2189 static void __release_event(struct kref *kref)
2190 {
2191         struct ceph_osd_event *event =
2192                 container_of(kref, struct ceph_osd_event, kref);
2193
2194         dout("__release_event %p\n", event);
2195         kfree(event);
2196 }
2197
2198 static void get_event(struct ceph_osd_event *event)
2199 {
2200         kref_get(&event->kref);
2201 }
2202
2203 void ceph_osdc_put_event(struct ceph_osd_event *event)
2204 {
2205         kref_put(&event->kref, __release_event);
2206 }
2207 EXPORT_SYMBOL(ceph_osdc_put_event);
2208
2209 static void __insert_event(struct ceph_osd_client *osdc,
2210                              struct ceph_osd_event *new)
2211 {
2212         struct rb_node **p = &osdc->event_tree.rb_node;
2213         struct rb_node *parent = NULL;
2214         struct ceph_osd_event *event = NULL;
2215
2216         while (*p) {
2217                 parent = *p;
2218                 event = rb_entry(parent, struct ceph_osd_event, node);
2219                 if (new->cookie < event->cookie)
2220                         p = &(*p)->rb_left;
2221                 else if (new->cookie > event->cookie)
2222                         p = &(*p)->rb_right;
2223                 else
2224                         BUG();
2225         }
2226
2227         rb_link_node(&new->node, parent, p);
2228         rb_insert_color(&new->node, &osdc->event_tree);
2229 }
2230
2231 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
2232                                                 u64 cookie)
2233 {
2234         struct rb_node **p = &osdc->event_tree.rb_node;
2235         struct rb_node *parent = NULL;
2236         struct ceph_osd_event *event = NULL;
2237
2238         while (*p) {
2239                 parent = *p;
2240                 event = rb_entry(parent, struct ceph_osd_event, node);
2241                 if (cookie < event->cookie)
2242                         p = &(*p)->rb_left;
2243                 else if (cookie > event->cookie)
2244                         p = &(*p)->rb_right;
2245                 else
2246                         return event;
2247         }
2248         return NULL;
2249 }
2250
2251 static void __remove_event(struct ceph_osd_event *event)
2252 {
2253         struct ceph_osd_client *osdc = event->osdc;
2254
2255         if (!RB_EMPTY_NODE(&event->node)) {
2256                 dout("__remove_event removed %p\n", event);
2257                 rb_erase(&event->node, &osdc->event_tree);
2258                 ceph_osdc_put_event(event);
2259         } else {
2260                 dout("__remove_event didn't remove %p\n", event);
2261         }
2262 }
2263
2264 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
2265                            void (*event_cb)(u64, u64, u8, void *),
2266                            void *data, struct ceph_osd_event **pevent)
2267 {
2268         struct ceph_osd_event *event;
2269
2270         event = kmalloc(sizeof(*event), GFP_NOIO);
2271         if (!event)
2272                 return -ENOMEM;
2273
2274         dout("create_event %p\n", event);
2275         event->cb = event_cb;
2276         event->one_shot = 0;
2277         event->data = data;
2278         event->osdc = osdc;
2279         INIT_LIST_HEAD(&event->osd_node);
2280         RB_CLEAR_NODE(&event->node);
2281         kref_init(&event->kref);   /* one ref for us */
2282         kref_get(&event->kref);    /* one ref for the caller */
2283
2284         spin_lock(&osdc->event_lock);
2285         event->cookie = ++osdc->event_count;
2286         __insert_event(osdc, event);
2287         spin_unlock(&osdc->event_lock);
2288
2289         *pevent = event;
2290         return 0;
2291 }
2292 EXPORT_SYMBOL(ceph_osdc_create_event);
2293
2294 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
2295 {
2296         struct ceph_osd_client *osdc = event->osdc;
2297
2298         dout("cancel_event %p\n", event);
2299         spin_lock(&osdc->event_lock);
2300         __remove_event(event);
2301         spin_unlock(&osdc->event_lock);
2302         ceph_osdc_put_event(event); /* caller's */
2303 }
2304 EXPORT_SYMBOL(ceph_osdc_cancel_event);
2305
2306
2307 static void do_event_work(struct work_struct *work)
2308 {
2309         struct ceph_osd_event_work *event_work =
2310                 container_of(work, struct ceph_osd_event_work, work);
2311         struct ceph_osd_event *event = event_work->event;
2312         u64 ver = event_work->ver;
2313         u64 notify_id = event_work->notify_id;
2314         u8 opcode = event_work->opcode;
2315
2316         dout("do_event_work completing %p\n", event);
2317         event->cb(ver, notify_id, opcode, event->data);
2318         dout("do_event_work completed %p\n", event);
2319         ceph_osdc_put_event(event);
2320         kfree(event_work);
2321 }
2322
2323
2324 /*
2325  * Process osd watch notifications
2326  */
2327 static void handle_watch_notify(struct ceph_osd_client *osdc,
2328                                 struct ceph_msg *msg)
2329 {
2330         void *p, *end;
2331         u8 proto_ver;
2332         u64 cookie, ver, notify_id;
2333         u8 opcode;
2334         struct ceph_osd_event *event;
2335         struct ceph_osd_event_work *event_work;
2336
2337         p = msg->front.iov_base;
2338         end = p + msg->front.iov_len;
2339
2340         ceph_decode_8_safe(&p, end, proto_ver, bad);
2341         ceph_decode_8_safe(&p, end, opcode, bad);
2342         ceph_decode_64_safe(&p, end, cookie, bad);
2343         ceph_decode_64_safe(&p, end, ver, bad);
2344         ceph_decode_64_safe(&p, end, notify_id, bad);
2345
2346         spin_lock(&osdc->event_lock);
2347         event = __find_event(osdc, cookie);
2348         if (event) {
2349                 BUG_ON(event->one_shot);
2350                 get_event(event);
2351         }
2352         spin_unlock(&osdc->event_lock);
2353         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
2354              cookie, ver, event);
2355         if (event) {
2356                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
2357                 if (!event_work) {
2358                         pr_err("couldn't allocate event_work\n");
2359                         ceph_osdc_put_event(event);
2360                         return;
2361                 }
2362                 INIT_WORK(&event_work->work, do_event_work);
2363                 event_work->event = event;
2364                 event_work->ver = ver;
2365                 event_work->notify_id = notify_id;
2366                 event_work->opcode = opcode;
2367
2368                 queue_work(osdc->notify_wq, &event_work->work);
2369         }
2370
2371         return;
2372
2373 bad:
2374         pr_err("osdc handle_watch_notify corrupt msg\n");
2375 }
2376
2377 /*
2378  * build new request AND message
2379  *
2380  */
2381 void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off,
2382                                 struct ceph_snap_context *snapc, u64 snap_id,
2383                                 struct timespec *mtime)
2384 {
2385         struct ceph_msg *msg = req->r_request;
2386         void *p;
2387         size_t msg_size;
2388         int flags = req->r_flags;
2389         u64 data_len;
2390         unsigned int i;
2391
2392         req->r_snapid = snap_id;
2393         req->r_snapc = ceph_get_snap_context(snapc);
2394
2395         /* encode request */
2396         msg->hdr.version = cpu_to_le16(4);
2397
2398         p = msg->front.iov_base;
2399         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
2400         req->r_request_osdmap_epoch = p;
2401         p += 4;
2402         req->r_request_flags = p;
2403         p += 4;
2404         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
2405                 ceph_encode_timespec(p, mtime);
2406         p += sizeof(struct ceph_timespec);
2407         req->r_request_reassert_version = p;
2408         p += sizeof(struct ceph_eversion); /* will get filled in */
2409
2410         /* oloc */
2411         ceph_encode_8(&p, 4);
2412         ceph_encode_8(&p, 4);
2413         ceph_encode_32(&p, 8 + 4 + 4);
2414         req->r_request_pool = p;
2415         p += 8;
2416         ceph_encode_32(&p, -1);  /* preferred */
2417         ceph_encode_32(&p, 0);   /* key len */
2418
2419         ceph_encode_8(&p, 1);
2420         req->r_request_pgid = p;
2421         p += 8 + 4;
2422         ceph_encode_32(&p, -1);  /* preferred */
2423
2424         /* oid */
2425         ceph_encode_32(&p, req->r_base_oid.name_len);
2426         memcpy(p, req->r_base_oid.name, req->r_base_oid.name_len);
2427         dout("oid '%.*s' len %d\n", req->r_base_oid.name_len,
2428              req->r_base_oid.name, req->r_base_oid.name_len);
2429         p += req->r_base_oid.name_len;
2430
2431         /* ops--can imply data */
2432         ceph_encode_16(&p, (u16)req->r_num_ops);
2433         data_len = 0;
2434         for (i = 0; i < req->r_num_ops; i++) {
2435                 data_len += osd_req_encode_op(req, p, i);
2436                 p += sizeof(struct ceph_osd_op);
2437         }
2438
2439         /* snaps */
2440         ceph_encode_64(&p, req->r_snapid);
2441         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
2442         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
2443         if (req->r_snapc) {
2444                 for (i = 0; i < snapc->num_snaps; i++) {
2445                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2446                 }
2447         }
2448
2449         req->r_request_attempts = p;
2450         p += 4;
2451
2452         /* data */
2453         if (flags & CEPH_OSD_FLAG_WRITE) {
2454                 u16 data_off;
2455
2456                 /*
2457                  * The header "data_off" is a hint to the receiver
2458                  * allowing it to align received data into its
2459                  * buffers such that there's no need to re-copy
2460                  * it before writing it to disk (direct I/O).
2461                  */
2462                 data_off = (u16) (off & 0xffff);
2463                 req->r_request->hdr.data_off = cpu_to_le16(data_off);
2464         }
2465         req->r_request->hdr.data_len = cpu_to_le32(data_len);
2466
2467         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
2468         msg_size = p - msg->front.iov_base;
2469         msg->front.iov_len = msg_size;
2470         msg->hdr.front_len = cpu_to_le32(msg_size);
2471
2472         dout("build_request msg_size was %d\n", (int)msg_size);
2473 }
2474 EXPORT_SYMBOL(ceph_osdc_build_request);
2475
2476 /*
2477  * Register request, send initial attempt.
2478  */
2479 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
2480                             struct ceph_osd_request *req,
2481                             bool nofail)
2482 {
2483         int rc;
2484
2485         down_read(&osdc->map_sem);
2486         mutex_lock(&osdc->request_mutex);
2487
2488         rc = __ceph_osdc_start_request(osdc, req, nofail);
2489
2490         mutex_unlock(&osdc->request_mutex);
2491         up_read(&osdc->map_sem);
2492
2493         return rc;
2494 }
2495 EXPORT_SYMBOL(ceph_osdc_start_request);
2496
2497 /*
2498  * Unregister a registered request.  The request is not completed (i.e.
2499  * no callbacks or wakeups) - higher layers are supposed to know what
2500  * they are canceling.
2501  */
2502 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
2503 {
2504         struct ceph_osd_client *osdc = req->r_osdc;
2505
2506         mutex_lock(&osdc->request_mutex);
2507         if (req->r_linger)
2508                 __unregister_linger_request(osdc, req);
2509         __unregister_request(osdc, req);
2510         mutex_unlock(&osdc->request_mutex);
2511
2512         dout("%s %p tid %llu canceled\n", __func__, req, req->r_tid);
2513 }
2514 EXPORT_SYMBOL(ceph_osdc_cancel_request);
2515
2516 /*
2517  * wait for a request to complete
2518  */
2519 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
2520                            struct ceph_osd_request *req)
2521 {
2522         int rc;
2523
2524         dout("%s %p tid %llu\n", __func__, req, req->r_tid);
2525
2526         rc = wait_for_completion_interruptible(&req->r_completion);
2527         if (rc < 0) {
2528                 dout("%s %p tid %llu interrupted\n", __func__, req, req->r_tid);
2529                 ceph_osdc_cancel_request(req);
2530                 complete_request(req);
2531                 return rc;
2532         }
2533
2534         dout("%s %p tid %llu result %d\n", __func__, req, req->r_tid,
2535              req->r_result);
2536         return req->r_result;
2537 }
2538 EXPORT_SYMBOL(ceph_osdc_wait_request);
2539
2540 /*
2541  * sync - wait for all in-flight requests to flush.  avoid starvation.
2542  */
2543 void ceph_osdc_sync(struct ceph_osd_client *osdc)
2544 {
2545         struct ceph_osd_request *req;
2546         u64 last_tid, next_tid = 0;
2547
2548         mutex_lock(&osdc->request_mutex);
2549         last_tid = osdc->last_tid;
2550         while (1) {
2551                 req = __lookup_request_ge(osdc, next_tid);
2552                 if (!req)
2553                         break;
2554                 if (req->r_tid > last_tid)
2555                         break;
2556
2557                 next_tid = req->r_tid + 1;
2558                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
2559                         continue;
2560
2561                 ceph_osdc_get_request(req);
2562                 mutex_unlock(&osdc->request_mutex);
2563                 dout("sync waiting on tid %llu (last is %llu)\n",
2564                      req->r_tid, last_tid);
2565                 wait_for_completion(&req->r_safe_completion);
2566                 mutex_lock(&osdc->request_mutex);
2567                 ceph_osdc_put_request(req);
2568         }
2569         mutex_unlock(&osdc->request_mutex);
2570         dout("sync done (thru tid %llu)\n", last_tid);
2571 }
2572 EXPORT_SYMBOL(ceph_osdc_sync);
2573
2574 /*
2575  * Call all pending notify callbacks - for use after a watch is
2576  * unregistered, to make sure no more callbacks for it will be invoked
2577  */
2578 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
2579 {
2580         flush_workqueue(osdc->notify_wq);
2581 }
2582 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
2583
2584
2585 /*
2586  * init, shutdown
2587  */
2588 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
2589 {
2590         int err;
2591
2592         dout("init\n");
2593         osdc->client = client;
2594         osdc->osdmap = NULL;
2595         init_rwsem(&osdc->map_sem);
2596         init_completion(&osdc->map_waiters);
2597         osdc->last_requested_map = 0;
2598         mutex_init(&osdc->request_mutex);
2599         osdc->last_tid = 0;
2600         osdc->osds = RB_ROOT;
2601         INIT_LIST_HEAD(&osdc->osd_lru);
2602         osdc->requests = RB_ROOT;
2603         INIT_LIST_HEAD(&osdc->req_lru);
2604         INIT_LIST_HEAD(&osdc->req_unsent);
2605         INIT_LIST_HEAD(&osdc->req_notarget);
2606         INIT_LIST_HEAD(&osdc->req_linger);
2607         osdc->num_requests = 0;
2608         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
2609         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
2610         spin_lock_init(&osdc->event_lock);
2611         osdc->event_tree = RB_ROOT;
2612         osdc->event_count = 0;
2613
2614         schedule_delayed_work(&osdc->osds_timeout_work,
2615            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
2616
2617         err = -ENOMEM;
2618         osdc->req_mempool = mempool_create_kmalloc_pool(10,
2619                                         sizeof(struct ceph_osd_request));
2620         if (!osdc->req_mempool)
2621                 goto out;
2622
2623         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
2624                                 OSD_OP_FRONT_LEN, 10, true,
2625                                 "osd_op");
2626         if (err < 0)
2627                 goto out_mempool;
2628         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
2629                                 OSD_OPREPLY_FRONT_LEN, 10, true,
2630                                 "osd_op_reply");
2631         if (err < 0)
2632                 goto out_msgpool;
2633
2634         err = -ENOMEM;
2635         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
2636         if (!osdc->notify_wq)
2637                 goto out_msgpool_reply;
2638
2639         return 0;
2640
2641 out_msgpool_reply:
2642         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2643 out_msgpool:
2644         ceph_msgpool_destroy(&osdc->msgpool_op);
2645 out_mempool:
2646         mempool_destroy(osdc->req_mempool);
2647 out:
2648         return err;
2649 }
2650
2651 void ceph_osdc_stop(struct ceph_osd_client *osdc)
2652 {
2653         flush_workqueue(osdc->notify_wq);
2654         destroy_workqueue(osdc->notify_wq);
2655         cancel_delayed_work_sync(&osdc->timeout_work);
2656         cancel_delayed_work_sync(&osdc->osds_timeout_work);
2657         if (osdc->osdmap) {
2658                 ceph_osdmap_destroy(osdc->osdmap);
2659                 osdc->osdmap = NULL;
2660         }
2661         remove_all_osds(osdc);
2662         mempool_destroy(osdc->req_mempool);
2663         ceph_msgpool_destroy(&osdc->msgpool_op);
2664         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
2665 }
2666
2667 /*
2668  * Read some contiguous pages.  If we cross a stripe boundary, shorten
2669  * *plen.  Return number of bytes read, or error.
2670  */
2671 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
2672                         struct ceph_vino vino, struct ceph_file_layout *layout,
2673                         u64 off, u64 *plen,
2674                         u32 truncate_seq, u64 truncate_size,
2675                         struct page **pages, int num_pages, int page_align)
2676 {
2677         struct ceph_osd_request *req;
2678         int rc = 0;
2679
2680         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
2681              vino.snap, off, *plen);
2682         req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1,
2683                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
2684                                     NULL, truncate_seq, truncate_size,
2685                                     false);
2686         if (IS_ERR(req))
2687                 return PTR_ERR(req);
2688
2689         /* it may be a short read due to an object boundary */
2690
2691         osd_req_op_extent_osd_data_pages(req, 0,
2692                                 pages, *plen, page_align, false, false);
2693
2694         dout("readpages  final extent is %llu~%llu (%llu bytes align %d)\n",
2695              off, *plen, *plen, page_align);
2696
2697         ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
2698
2699         rc = ceph_osdc_start_request(osdc, req, false);
2700         if (!rc)
2701                 rc = ceph_osdc_wait_request(osdc, req);
2702
2703         ceph_osdc_put_request(req);
2704         dout("readpages result %d\n", rc);
2705         return rc;
2706 }
2707 EXPORT_SYMBOL(ceph_osdc_readpages);
2708
2709 /*
2710  * do a synchronous write on N pages
2711  */
2712 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
2713                          struct ceph_file_layout *layout,
2714                          struct ceph_snap_context *snapc,
2715                          u64 off, u64 len,
2716                          u32 truncate_seq, u64 truncate_size,
2717                          struct timespec *mtime,
2718                          struct page **pages, int num_pages)
2719 {
2720         struct ceph_osd_request *req;
2721         int rc = 0;
2722         int page_align = off & ~PAGE_MASK;
2723
2724         BUG_ON(vino.snap != CEPH_NOSNAP);       /* snapshots aren't writeable */
2725         req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1,
2726                                     CEPH_OSD_OP_WRITE,
2727                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
2728                                     snapc, truncate_seq, truncate_size,
2729                                     true);
2730         if (IS_ERR(req))
2731                 return PTR_ERR(req);
2732
2733         /* it may be a short write due to an object boundary */
2734         osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
2735                                 false, false);
2736         dout("writepages %llu~%llu (%llu bytes)\n", off, len, len);
2737
2738         ceph_osdc_build_request(req, off, snapc, CEPH_NOSNAP, mtime);
2739
2740         rc = ceph_osdc_start_request(osdc, req, true);
2741         if (!rc)
2742                 rc = ceph_osdc_wait_request(osdc, req);
2743
2744         ceph_osdc_put_request(req);
2745         if (rc == 0)
2746                 rc = len;
2747         dout("writepages result %d\n", rc);
2748         return rc;
2749 }
2750 EXPORT_SYMBOL(ceph_osdc_writepages);
2751
2752 int ceph_osdc_setup(void)
2753 {
2754         BUG_ON(ceph_osd_request_cache);
2755         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request",
2756                                         sizeof (struct ceph_osd_request),
2757                                         __alignof__(struct ceph_osd_request),
2758                                         0, NULL);
2759
2760         return ceph_osd_request_cache ? 0 : -ENOMEM;
2761 }
2762 EXPORT_SYMBOL(ceph_osdc_setup);
2763
2764 void ceph_osdc_cleanup(void)
2765 {
2766         BUG_ON(!ceph_osd_request_cache);
2767         kmem_cache_destroy(ceph_osd_request_cache);
2768         ceph_osd_request_cache = NULL;
2769 }
2770 EXPORT_SYMBOL(ceph_osdc_cleanup);
2771
2772 /*
2773  * handle incoming message
2774  */
2775 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2776 {
2777         struct ceph_osd *osd = con->private;
2778         struct ceph_osd_client *osdc;
2779         int type = le16_to_cpu(msg->hdr.type);
2780
2781         if (!osd)
2782                 goto out;
2783         osdc = osd->o_osdc;
2784
2785         switch (type) {
2786         case CEPH_MSG_OSD_MAP:
2787                 ceph_osdc_handle_map(osdc, msg);
2788                 break;
2789         case CEPH_MSG_OSD_OPREPLY:
2790                 handle_reply(osdc, msg, con);
2791                 break;
2792         case CEPH_MSG_WATCH_NOTIFY:
2793                 handle_watch_notify(osdc, msg);
2794                 break;
2795
2796         default:
2797                 pr_err("received unknown message type %d %s\n", type,
2798                        ceph_msg_type_name(type));
2799         }
2800 out:
2801         ceph_msg_put(msg);
2802 }
2803
2804 /*
2805  * lookup and return message for incoming reply.  set up reply message
2806  * pages.
2807  */
2808 static struct ceph_msg *get_reply(struct ceph_connection *con,
2809                                   struct ceph_msg_header *hdr,
2810                                   int *skip)
2811 {
2812         struct ceph_osd *osd = con->private;
2813         struct ceph_osd_client *osdc = osd->o_osdc;
2814         struct ceph_msg *m;
2815         struct ceph_osd_request *req;
2816         int front_len = le32_to_cpu(hdr->front_len);
2817         int data_len = le32_to_cpu(hdr->data_len);
2818         u64 tid;
2819
2820         tid = le64_to_cpu(hdr->tid);
2821         mutex_lock(&osdc->request_mutex);
2822         req = __lookup_request(osdc, tid);
2823         if (!req) {
2824                 *skip = 1;
2825                 m = NULL;
2826                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2827                      osd->o_osd);
2828                 goto out;
2829         }
2830
2831         if (req->r_reply->con)
2832                 dout("%s revoking msg %p from old con %p\n", __func__,
2833                      req->r_reply, req->r_reply->con);
2834         ceph_msg_revoke_incoming(req->r_reply);
2835
2836         if (front_len > req->r_reply->front_alloc_len) {
2837                 pr_warn("get_reply front %d > preallocated %d (%u#%llu)\n",
2838                         front_len, req->r_reply->front_alloc_len,
2839                         (unsigned int)con->peer_name.type,
2840                         le64_to_cpu(con->peer_name.num));
2841                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
2842                                  false);
2843                 if (!m)
2844                         goto out;
2845                 ceph_msg_put(req->r_reply);
2846                 req->r_reply = m;
2847         }
2848         m = ceph_msg_get(req->r_reply);
2849
2850         if (data_len > 0) {
2851                 struct ceph_osd_data *osd_data;
2852
2853                 /*
2854                  * XXX This is assuming there is only one op containing
2855                  * XXX page data.  Probably OK for reads, but this
2856                  * XXX ought to be done more generally.
2857                  */
2858                 osd_data = osd_req_op_extent_osd_data(req, 0);
2859                 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
2860                         if (osd_data->pages &&
2861                                 unlikely(osd_data->length < data_len)) {
2862
2863                                 pr_warn("tid %lld reply has %d bytes we had only %llu bytes ready\n",
2864                                         tid, data_len, osd_data->length);
2865                                 *skip = 1;
2866                                 ceph_msg_put(m);
2867                                 m = NULL;
2868                                 goto out;
2869                         }
2870                 }
2871         }
2872         *skip = 0;
2873         dout("get_reply tid %lld %p\n", tid, m);
2874
2875 out:
2876         mutex_unlock(&osdc->request_mutex);
2877         return m;
2878
2879 }
2880
2881 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2882                                   struct ceph_msg_header *hdr,
2883                                   int *skip)
2884 {
2885         struct ceph_osd *osd = con->private;
2886         int type = le16_to_cpu(hdr->type);
2887         int front = le32_to_cpu(hdr->front_len);
2888
2889         *skip = 0;
2890         switch (type) {
2891         case CEPH_MSG_OSD_MAP:
2892         case CEPH_MSG_WATCH_NOTIFY:
2893                 return ceph_msg_new(type, front, GFP_NOFS, false);
2894         case CEPH_MSG_OSD_OPREPLY:
2895                 return get_reply(con, hdr, skip);
2896         default:
2897                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2898                         osd->o_osd);
2899                 *skip = 1;
2900                 return NULL;
2901         }
2902 }
2903
2904 /*
2905  * Wrappers to refcount containing ceph_osd struct
2906  */
2907 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2908 {
2909         struct ceph_osd *osd = con->private;
2910         if (get_osd(osd))
2911                 return con;
2912         return NULL;
2913 }
2914
2915 static void put_osd_con(struct ceph_connection *con)
2916 {
2917         struct ceph_osd *osd = con->private;
2918         put_osd(osd);
2919 }
2920
2921 /*
2922  * authentication
2923  */
2924 /*
2925  * Note: returned pointer is the address of a structure that's
2926  * managed separately.  Caller must *not* attempt to free it.
2927  */
2928 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2929                                         int *proto, int force_new)
2930 {
2931         struct ceph_osd *o = con->private;
2932         struct ceph_osd_client *osdc = o->o_osdc;
2933         struct ceph_auth_client *ac = osdc->client->monc.auth;
2934         struct ceph_auth_handshake *auth = &o->o_auth;
2935
2936         if (force_new && auth->authorizer) {
2937                 ceph_auth_destroy_authorizer(ac, auth->authorizer);
2938                 auth->authorizer = NULL;
2939         }
2940         if (!auth->authorizer) {
2941                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2942                                                       auth);
2943                 if (ret)
2944                         return ERR_PTR(ret);
2945         } else {
2946                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2947                                                      auth);
2948                 if (ret)
2949                         return ERR_PTR(ret);
2950         }
2951         *proto = ac->protocol;
2952
2953         return auth;
2954 }
2955
2956
2957 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2958 {
2959         struct ceph_osd *o = con->private;
2960         struct ceph_osd_client *osdc = o->o_osdc;
2961         struct ceph_auth_client *ac = osdc->client->monc.auth;
2962
2963         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2964 }
2965
2966 static int invalidate_authorizer(struct ceph_connection *con)
2967 {
2968         struct ceph_osd *o = con->private;
2969         struct ceph_osd_client *osdc = o->o_osdc;
2970         struct ceph_auth_client *ac = osdc->client->monc.auth;
2971
2972         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2973         return ceph_monc_validate_auth(&osdc->client->monc);
2974 }
2975
2976 static int sign_message(struct ceph_connection *con, struct ceph_msg *msg)
2977 {
2978         struct ceph_osd *o = con->private;
2979         struct ceph_auth_handshake *auth = &o->o_auth;
2980         return ceph_auth_sign_message(auth, msg);
2981 }
2982
2983 static int check_message_signature(struct ceph_connection *con, struct ceph_msg *msg)
2984 {
2985         struct ceph_osd *o = con->private;
2986         struct ceph_auth_handshake *auth = &o->o_auth;
2987         return ceph_auth_check_message_signature(auth, msg);
2988 }
2989
2990 static const struct ceph_connection_operations osd_con_ops = {
2991         .get = get_osd_con,
2992         .put = put_osd_con,
2993         .dispatch = dispatch,
2994         .get_authorizer = get_authorizer,
2995         .verify_authorizer_reply = verify_authorizer_reply,
2996         .invalidate_authorizer = invalidate_authorizer,
2997         .alloc_msg = alloc_msg,
2998         .sign_message = sign_message,
2999         .check_message_signature = check_message_signature,
3000         .fault = osd_reset,
3001 };