lustre/mdt/mdt_batch.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2020, DDN Storage Corporation.
  24  */
  25 /*
  26  * This file is part of Lustre, http://www.lustre.org/
  27  */
  28 /*
  29  * lustre/mdt/mdt_batch.c
  30  *
  31  * Batch Metadata Updating on the server (MDT)
  32  *
  33  * Author: Qian Yingjin <qian@ddn.com>
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_MDS
  37
  38 #include <linux/module.h>
  39
  40 #include <lustre_mds.h>
  41 #include "mdt_internal.h"
  42
  43 static int mdt_batch_unpack(struct mdt_thread_info *info, __u32 opc)
  44 {
  45         int rc = 0;
  46
  47         switch (opc) {
  48         default:
  49                 rc = -EOPNOTSUPP;
  50                 CERROR("%s: Unexpected opcode %d: rc = %d\n",
  51                        mdt_obd_name(info->mti_mdt), opc, rc);
  52                 break;
  53         }
  54
  55         RETURN(rc);
  56 }
  57
  58 static int mdt_batch_pack_repmsg(struct mdt_thread_info *info)
  59 {
  60         return 0;
  61 }
  62
  63 typedef int (*mdt_batch_reconstructor)(struct tgt_session_info *tsi);
  64
  65 static mdt_batch_reconstructor reconstructors[BUT_LAST_OPC];
  66
  67 static int mdt_batch_reconstruct(struct tgt_session_info *tsi, long opc)
  68 {
  69         mdt_batch_reconstructor reconst;
  70         int rc;
  71
  72         ENTRY;
  73
  74         if (opc >= BUT_LAST_OPC)
  75                 RETURN(-EOPNOTSUPP);
  76
  77         reconst = reconstructors[opc];
  78         LASSERT(reconst != NULL);
  79         rc = reconst(tsi);
  80         RETURN(rc);
  81 }
  82
  83 /* Batch UpdaTe Request with a format known in advance */
  84 #define TGT_BUT_HDL(flags, opc, fn)                     \
  85 [opc - BUT_FIRST_OPC] = {                               \
  86         .th_name        = #opc,                         \
  87         .th_fail_id     = 0,                            \
  88         .th_opc         = opc,                          \
  89         .th_flags       = flags,                        \
  90         .th_act         = fn,                           \
  91         .th_fmt         = &RQF_ ## opc,                 \
  92         .th_version     = LUSTRE_MDS_VERSION,           \
  93         .th_hp          = NULL,                         \
  94 }
  95
  96 static struct tgt_handler mdt_batch_handlers[BUT_LAST_OPC];
  97
  98 static struct tgt_handler *mdt_batch_handler_find(__u32 opc)
  99 {
 100         struct tgt_handler *h;
 101
 102         h = NULL;
 103         if (opc >= BUT_FIRST_OPC && opc < BUT_LAST_OPC) {
 104                 h = &mdt_batch_handlers[opc - BUT_FIRST_OPC];
 105                 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
 106                          h->th_opc, opc);
 107         } else {
 108                 h = NULL; /* unsupported opc */
 109         }
 110         return h;
 111 }
 112
 113 int mdt_batch(struct tgt_session_info *tsi)
 114 {
 115         struct mdt_thread_info *info = tsi2mdt_info(tsi);
 116         struct req_capsule *pill = &info->mti_sub_pill;
 117         struct ptlrpc_request *req = tgt_ses_req(tsi);
 118         struct but_update_header *buh;
 119         struct but_update_buffer *bub = NULL;
 120         struct batch_update_reply *reply = NULL;
 121         struct ptlrpc_bulk_desc *desc = NULL;
 122         struct tg_reply_data *trd = NULL;
 123         struct lustre_msg *repmsg = NULL;
 124         bool need_reconstruct;
 125         __u32 handled_update_count = 0;
 126         __u32 update_buf_count;
 127         __u32 packed_replen;
 128         void **update_bufs;
 129         int buh_size;
 130         int rc;
 131         int i;
 132
 133         ENTRY;
 134
 135         buh_size = req_capsule_get_size(&req->rq_pill, &RMF_BUT_HEADER,
 136                                         RCL_CLIENT);
 137         if (buh_size <= 0)
 138                 RETURN(err_serious(-EPROTO));
 139
 140         buh = req_capsule_client_get(&req->rq_pill, &RMF_BUT_HEADER);
 141         if (buh == NULL)
 142                 RETURN(err_serious(-EPROTO));
 143
 144         if (buh->buh_magic != BUT_HEADER_MAGIC) {
 145                 CERROR("%s: invalid update header magic %x expect %x: "
 146                        "rc = %d\n", tgt_name(tsi->tsi_tgt), buh->buh_magic,
 147                        BUT_HEADER_MAGIC, -EPROTO);
 148                 RETURN(err_serious(-EPROTO));
 149         }
 150
 151         update_buf_count = buh->buh_count;
 152         if (update_buf_count == 0)
 153                 RETURN(err_serious(-EPROTO));
 154
 155         OBD_ALLOC_PTR_ARRAY(update_bufs, update_buf_count);
 156         if (update_bufs == NULL)
 157                 RETURN(err_serious(-ENOMEM));
 158
 159         if (buh->buh_inline_length > 0) {
 160                 update_bufs[0] = buh->buh_inline_data;
 161         } else {
 162                 struct but_update_buffer *tmp;
 163                 int page_count = 0;
 164
 165                 bub = req_capsule_client_get(&req->rq_pill, &RMF_BUT_BUF);
 166                 if (bub == NULL)
 167                         GOTO(out, rc = err_serious(-EPROTO));
 168
 169                 for (i = 0; i < update_buf_count; i++)
 170                         /* First *and* last might be partial pages, hence +1 */
 171                         page_count += DIV_ROUND_UP(bub[i].bub_size,
 172                                                    PAGE_SIZE) + 1;
 173
 174                 desc = ptlrpc_prep_bulk_exp(req, page_count,
 175                                             PTLRPC_BULK_OPS_COUNT,
 176                                             PTLRPC_BULK_GET_SINK,
 177                                             MDS_BULK_PORTAL,
 178                                             &ptlrpc_bulk_kiov_nopin_ops);
 179                 if (desc == NULL)
 180                         GOTO(out, rc = err_serious(-ENOMEM));
 181
 182                 tmp = bub;
 183                 for (i = 0; i < update_buf_count; i++, tmp++) {
 184                         if (tmp->bub_size >= OUT_MAXREQSIZE)
 185                                 GOTO(out, rc = err_serious(-EPROTO));
 186
 187                         OBD_ALLOC_LARGE(update_bufs[i], tmp->bub_size);
 188                         if (update_bufs[i] == NULL)
 189                                 GOTO(out, rc = err_serious(-ENOMEM));
 190
 191                         desc->bd_frag_ops->add_iov_frag(desc, update_bufs[i],
 192                                                         tmp->bub_size);
 193                 }
 194
 195                 req->rq_bulk_write = 1;
 196                 rc = sptlrpc_svc_prep_bulk(req, desc);
 197                 if (rc != 0)
 198                         GOTO(out, rc = err_serious(rc));
 199
 200                 rc = target_bulk_io(req->rq_export, desc);
 201                 if (rc < 0)
 202                         GOTO(out, rc = err_serious(rc));
 203         }
 204
 205         req_capsule_set_size(&req->rq_pill, &RMF_BUT_REPLY, RCL_SERVER,
 206                              buh->buh_reply_size);
 207         rc = req_capsule_server_pack(&req->rq_pill);
 208         if (rc != 0) {
 209                 DEBUG_REQ(D_ERROR, req, "%s: Can't pack response: rc = %d\n",
 210                        tgt_name(tsi->tsi_tgt), rc);
 211                 GOTO(out, rc);
 212         }
 213
 214         /* Prepare the update reply buffer */
 215         reply = req_capsule_server_get(&req->rq_pill, &RMF_BUT_REPLY);
 216         if (reply == NULL)
 217                 GOTO(out, rc = -EPROTO);
 218
 219         reply->burp_magic = BUT_REPLY_MAGIC;
 220         packed_replen = sizeof(*reply);
 221         info->mti_max_repsize = buh->buh_reply_size;
 222         info->mti_batch_env = 1;
 223         info->mti_pill = pill;
 224         tsi->tsi_batch_env = true;
 225
 226         OBD_ALLOC_PTR(trd);
 227         if (trd == NULL)
 228                 GOTO(out, rc = -ENOMEM);
 229
 230         need_reconstruct = tgt_check_resent(req, trd);
 231         /* Walk through sub requests in the batch request to execute them. */
 232         for (i = 0; i < update_buf_count; i++) {
 233                 struct batch_update_request *bur;
 234                 struct lustre_msg *reqmsg = NULL;
 235                 struct tgt_handler *h;
 236                 int update_count;
 237                 int j;
 238
 239                 bur = update_bufs[i];
 240                 update_count = bur->burq_count;
 241                 for (j = 0; j < update_count; j++) {
 242                         __u32 replen;
 243
 244                         reqmsg = batch_update_reqmsg_next(bur, reqmsg);
 245                         repmsg = batch_update_repmsg_next(reply, repmsg);
 246
 247                         if (handled_update_count > buh->buh_update_count)
 248                                 GOTO(out, rc = -EOVERFLOW);
 249
 250                         LASSERT(reqmsg != NULL && repmsg != NULL);
 251                         LASSERTF(reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V2,
 252                                  "Invalid reqmsg magic %x expected %x\n",
 253                                  reqmsg->lm_magic, LUSTRE_MSG_MAGIC_V2);
 254
 255                         h = mdt_batch_handler_find(reqmsg->lm_opc);
 256                         if (unlikely(h == NULL)) {
 257                                 CERROR("%s: unsupported opc: 0x%x\n",
 258                                        tgt_name(tsi->tsi_tgt), reqmsg->lm_opc);
 259                                 GOTO(out, rc = -ENOTSUPP);
 260                         }
 261
 262                         LASSERT(h->th_fmt != NULL);
 263                         req_capsule_subreq_init(pill, h->th_fmt, req,
 264                                                 reqmsg, repmsg, RCL_SERVER);
 265
 266                         rc = mdt_batch_unpack(info, reqmsg->lm_opc);
 267                         if (rc) {
 268                                 CERROR("%s: Can't unpack subreq, rc = %d\n",
 269                                        mdt_obd_name(info->mti_mdt), rc);
 270                                 GOTO(out, rc);
 271                         }
 272
 273                         rc = mdt_batch_pack_repmsg(info);
 274                         if (rc)
 275                                 GOTO(out, rc);
 276
 277                         /* Need to reconstruct the reply for committed sub
 278                          * requests in a batched RPC.
 279                          * It only calls reconstruct for modification sub
 280                          * requests.
 281                          * For uncommitted or read-only sub requests, the server
 282                          * should re-execute them via the ->th_act() below.
 283                          */
 284                         if ((h->th_flags & IS_MUTABLE) && need_reconstruct &&
 285                             handled_update_count <=
 286                             trd->trd_reply.lrd_batch_idx) {
 287                                 rc = mdt_batch_reconstruct(tsi, reqmsg->lm_opc);
 288                                 if (rc)
 289                                         GOTO(out, rc);
 290                                 GOTO(next, rc);
 291                         }
 292
 293                         tsi->tsi_batch_idx = handled_update_count;
 294                         rc = h->th_act(tsi);
 295                         if (rc)
 296                                 GOTO(out, rc);
 297 next:
 298                         repmsg->lm_result = rc;
 299                         mdt_thread_info_reset(info);
 300                         /*
 301                          * TODO: Check whether overflow reply buffer.
 302                          * Fix reply, shrink and/or grow reply buffers.
 303                          */
 304                         replen = lustre_packed_msg_size(repmsg);
 305                         info->mti_max_repsize -= replen;
 306                         packed_replen += replen;
 307                         handled_update_count++;
 308                 }
 309         }
 310
 311         /*
 312          * TODO: Grow/shrink the reply buffer.
 313          */
 314         CDEBUG(D_INFO, "reply size %u packed replen %u\n",
 315                buh->buh_reply_size, packed_replen);
 316         if (buh->buh_reply_size > packed_replen)
 317                 req_capsule_shrink(&req->rq_pill, &RMF_BUT_REPLY,
 318                                    packed_replen, RCL_SERVER);
 319 out:
 320         if (reply != NULL)
 321                 reply->burp_count = handled_update_count;
 322
 323         if (update_bufs != NULL) {
 324                 if (bub != NULL) {
 325                         for (i = 0; i < update_buf_count; i++, bub++) {
 326                                 if (update_bufs[i] != NULL)
 327                                         OBD_FREE_LARGE(update_bufs[i],
 328                                                        bub->bub_size);
 329                         }
 330                 }
 331
 332                 OBD_FREE_PTR_ARRAY(update_bufs, update_buf_count);
 333         }
 334
 335         if (trd)
 336                 OBD_FREE_PTR(trd);
 337
 338         if (desc != NULL)
 339                 ptlrpc_free_bulk(desc);
 340
 341         mdt_thread_info_fini(info);
 342         tsi->tsi_reply_fail_id = OBD_FAIL_BUT_UPDATE_NET_REP;
 343         RETURN(rc);
 344 }
 345