bulk-checkin.c

   1 /*
   2  * Copyright (c) 2011, Google Inc.
   3  */
   4 #include "git-compat-util.h"
   5 #include "alloc.h"
   6 #include "bulk-checkin.h"
   7 #include "hex.h"
   8 #include "lockfile.h"
   9 #include "repository.h"
  10 #include "csum-file.h"
  11 #include "pack.h"
  12 #include "strbuf.h"
  13 #include "string-list.h"
  14 #include "tmp-objdir.h"
  15 #include "packfile.h"
  16 #include "object-store.h"
  17
  18 static int odb_transaction_nesting;
  19
  20 static struct tmp_objdir *bulk_fsync_objdir;
  21
  22 static struct bulk_checkin_packfile {
  23         char *pack_tmp_name;
  24         struct hashfile *f;
  25         off_t offset;
  26         struct pack_idx_option pack_idx_opts;
  27
  28         struct pack_idx_entry **written;
  29         uint32_t alloc_written;
  30         uint32_t nr_written;
  31 } bulk_checkin_packfile;
  32
  33 static void finish_tmp_packfile(struct strbuf *basename,
  34                                 const char *pack_tmp_name,
  35                                 struct pack_idx_entry **written_list,
  36                                 uint32_t nr_written,
  37                                 struct pack_idx_option *pack_idx_opts,
  38                                 unsigned char hash[])
  39 {
  40         char *idx_tmp_name = NULL;
  41
  42         stage_tmp_packfiles(basename, pack_tmp_name, written_list, nr_written,
  43                             NULL, pack_idx_opts, hash, &idx_tmp_name);
  44         rename_tmp_packfile_idx(basename, &idx_tmp_name);
  45
  46         free(idx_tmp_name);
  47 }
  48
  49 static void flush_bulk_checkin_packfile(struct bulk_checkin_packfile *state)
  50 {
  51         unsigned char hash[GIT_MAX_RAWSZ];
  52         struct strbuf packname = STRBUF_INIT;
  53         int i;
  54
  55         if (!state->f)
  56                 return;
  57
  58         if (state->nr_written == 0) {
  59                 close(state->f->fd);
  60                 unlink(state->pack_tmp_name);
  61                 goto clear_exit;
  62         } else if (state->nr_written == 1) {
  63                 finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK,
  64                                   CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
  65         } else {
  66                 int fd = finalize_hashfile(state->f, hash, FSYNC_COMPONENT_PACK, 0);
  67                 fixup_pack_header_footer(fd, hash, state->pack_tmp_name,
  68                                          state->nr_written, hash,
  69                                          state->offset);
  70                 close(fd);
  71         }
  72
  73         strbuf_addf(&packname, "%s/pack/pack-%s.", get_object_directory(),
  74                     hash_to_hex(hash));
  75         finish_tmp_packfile(&packname, state->pack_tmp_name,
  76                             state->written, state->nr_written,
  77                             &state->pack_idx_opts, hash);
  78         for (i = 0; i < state->nr_written; i++)
  79                 free(state->written[i]);
  80
  81 clear_exit:
  82         free(state->written);
  83         memset(state, 0, sizeof(*state));
  84
  85         strbuf_release(&packname);
  86         /* Make objects we just wrote available to ourselves */
  87         reprepare_packed_git(the_repository);
  88 }
  89
  90 /*
  91  * Cleanup after batch-mode fsync_object_files.
  92  */
  93 static void flush_batch_fsync(void)
  94 {
  95         struct strbuf temp_path = STRBUF_INIT;
  96         struct tempfile *temp;
  97
  98         if (!bulk_fsync_objdir)
  99                 return;
 100
 101         /*
 102          * Issue a full hardware flush against a temporary file to ensure
 103          * that all objects are durable before any renames occur. The code in
 104          * fsync_loose_object_bulk_checkin has already issued a writeout
 105          * request, but it has not flushed any writeback cache in the storage
 106          * hardware or any filesystem logs. This fsync call acts as a barrier
 107          * to ensure that the data in each new object file is durable before
 108          * the final name is visible.
 109          */
 110         strbuf_addf(&temp_path, "%s/bulk_fsync_XXXXXX", get_object_directory());
 111         temp = xmks_tempfile(temp_path.buf);
 112         fsync_or_die(get_tempfile_fd(temp), get_tempfile_path(temp));
 113         delete_tempfile(&temp);
 114         strbuf_release(&temp_path);
 115
 116         /*
 117          * Make the object files visible in the primary ODB after their data is
 118          * fully durable.
 119          */
 120         tmp_objdir_migrate(bulk_fsync_objdir);
 121         bulk_fsync_objdir = NULL;
 122 }
 123
 124 static int already_written(struct bulk_checkin_packfile *state, struct object_id *oid)
 125 {
 126         int i;
 127
 128         /* The object may already exist in the repository */
 129         if (has_object_file(oid))
 130                 return 1;
 131
 132         /* Might want to keep the list sorted */
 133         for (i = 0; i < state->nr_written; i++)
 134                 if (oideq(&state->written[i]->oid, oid))
 135                         return 1;
 136
 137         /* This is a new object we need to keep */
 138         return 0;
 139 }
 140
 141 /*
 142  * Read the contents from fd for size bytes, streaming it to the
 143  * packfile in state while updating the hash in ctx. Signal a failure
 144  * by returning a negative value when the resulting pack would exceed
 145  * the pack size limit and this is not the first object in the pack,
 146  * so that the caller can discard what we wrote from the current pack
 147  * by truncating it and opening a new one. The caller will then call
 148  * us again after rewinding the input fd.
 149  *
 150  * The already_hashed_to pointer is kept untouched by the caller to
 151  * make sure we do not hash the same byte when we are called
 152  * again. This way, the caller does not have to checkpoint its hash
 153  * status before calling us just in case we ask it to call us again
 154  * with a new pack.
 155  */
 156 static int stream_to_pack(struct bulk_checkin_packfile *state,
 157                           git_hash_ctx *ctx, off_t *already_hashed_to,
 158                           int fd, size_t size, enum object_type type,
 159                           const char *path, unsigned flags)
 160 {
 161         git_zstream s;
 162         unsigned char ibuf[16384];
 163         unsigned char obuf[16384];
 164         unsigned hdrlen;
 165         int status = Z_OK;
 166         int write_object = (flags & HASH_WRITE_OBJECT);
 167         off_t offset = 0;
 168
 169         git_deflate_init(&s, pack_compression_level);
 170
 171         hdrlen = encode_in_pack_object_header(obuf, sizeof(obuf), type, size);
 172         s.next_out = obuf + hdrlen;
 173         s.avail_out = sizeof(obuf) - hdrlen;
 174
 175         while (status != Z_STREAM_END) {
 176                 if (size && !s.avail_in) {
 177                         ssize_t rsize = size < sizeof(ibuf) ? size : sizeof(ibuf);
 178                         ssize_t read_result = read_in_full(fd, ibuf, rsize);
 179                         if (read_result < 0)
 180                                 die_errno("failed to read from '%s'", path);
 181                         if (read_result != rsize)
 182                                 die("failed to read %d bytes from '%s'",
 183                                     (int)rsize, path);
 184                         offset += rsize;
 185                         if (*already_hashed_to < offset) {
 186                                 size_t hsize = offset - *already_hashed_to;
 187                                 if (rsize < hsize)
 188                                         hsize = rsize;
 189                                 if (hsize)
 190                                         the_hash_algo->update_fn(ctx, ibuf, hsize);
 191                                 *already_hashed_to = offset;
 192                         }
 193                         s.next_in = ibuf;
 194                         s.avail_in = rsize;
 195                         size -= rsize;
 196                 }
 197
 198                 status = git_deflate(&s, size ? 0 : Z_FINISH);
 199
 200                 if (!s.avail_out || status == Z_STREAM_END) {
 201                         if (write_object) {
 202                                 size_t written = s.next_out - obuf;
 203
 204                                 /* would we bust the size limit? */
 205                                 if (state->nr_written &&
 206                                     pack_size_limit_cfg &&
 207                                     pack_size_limit_cfg < state->offset + written) {
 208                                         git_deflate_abort(&s);
 209                                         return -1;
 210                                 }
 211
 212                                 hashwrite(state->f, obuf, written);
 213                                 state->offset += written;
 214                         }
 215                         s.next_out = obuf;
 216                         s.avail_out = sizeof(obuf);
 217                 }
 218
 219                 switch (status) {
 220                 case Z_OK:
 221                 case Z_BUF_ERROR:
 222                 case Z_STREAM_END:
 223                         continue;
 224                 default:
 225                         die("unexpected deflate failure: %d", status);
 226                 }
 227         }
 228         git_deflate_end(&s);
 229         return 0;
 230 }
 231
 232 /* Lazily create backing packfile for the state */
 233 static void prepare_to_stream(struct bulk_checkin_packfile *state,
 234                               unsigned flags)
 235 {
 236         if (!(flags & HASH_WRITE_OBJECT) || state->f)
 237                 return;
 238
 239         state->f = create_tmp_packfile(&state->pack_tmp_name);
 240         reset_pack_idx_option(&state->pack_idx_opts);
 241
 242         /* Pretend we are going to write only one object */
 243         state->offset = write_pack_header(state->f, 1);
 244         if (!state->offset)
 245                 die_errno("unable to write pack header");
 246 }
 247
 248 static int deflate_to_pack(struct bulk_checkin_packfile *state,
 249                            struct object_id *result_oid,
 250                            int fd, size_t size,
 251                            enum object_type type, const char *path,
 252                            unsigned flags)
 253 {
 254         off_t seekback, already_hashed_to;
 255         git_hash_ctx ctx;
 256         unsigned char obuf[16384];
 257         unsigned header_len;
 258         struct hashfile_checkpoint checkpoint = {0};
 259         struct pack_idx_entry *idx = NULL;
 260
 261         seekback = lseek(fd, 0, SEEK_CUR);
 262         if (seekback == (off_t) -1)
 263                 return error("cannot find the current offset");
 264
 265         header_len = format_object_header((char *)obuf, sizeof(obuf),
 266                                           type, size);
 267         the_hash_algo->init_fn(&ctx);
 268         the_hash_algo->update_fn(&ctx, obuf, header_len);
 269
 270         /* Note: idx is non-NULL when we are writing */
 271         if ((flags & HASH_WRITE_OBJECT) != 0)
 272                 CALLOC_ARRAY(idx, 1);
 273
 274         already_hashed_to = 0;
 275
 276         while (1) {
 277                 prepare_to_stream(state, flags);
 278                 if (idx) {
 279                         hashfile_checkpoint(state->f, &checkpoint);
 280                         idx->offset = state->offset;
 281                         crc32_begin(state->f);
 282                 }
 283                 if (!stream_to_pack(state, &ctx, &already_hashed_to,
 284                                     fd, size, type, path, flags))
 285                         break;
 286                 /*
 287                  * Writing this object to the current pack will make
 288                  * it too big; we need to truncate it, start a new
 289                  * pack, and write into it.
 290                  */
 291                 if (!idx)
 292                         BUG("should not happen");
 293                 hashfile_truncate(state->f, &checkpoint);
 294                 state->offset = checkpoint.offset;
 295                 flush_bulk_checkin_packfile(state);
 296                 if (lseek(fd, seekback, SEEK_SET) == (off_t) -1)
 297                         return error("cannot seek back");
 298         }
 299         the_hash_algo->final_oid_fn(result_oid, &ctx);
 300         if (!idx)
 301                 return 0;
 302
 303         idx->crc32 = crc32_end(state->f);
 304         if (already_written(state, result_oid)) {
 305                 hashfile_truncate(state->f, &checkpoint);
 306                 state->offset = checkpoint.offset;
 307                 free(idx);
 308         } else {
 309                 oidcpy(&idx->oid, result_oid);
 310                 ALLOC_GROW(state->written,
 311                            state->nr_written + 1,
 312                            state->alloc_written);
 313                 state->written[state->nr_written++] = idx;
 314         }
 315         return 0;
 316 }
 317
 318 void prepare_loose_object_bulk_checkin(void)
 319 {
 320         /*
 321          * We lazily create the temporary object directory
 322          * the first time an object might be added, since
 323          * callers may not know whether any objects will be
 324          * added at the time they call begin_odb_transaction.
 325          */
 326         if (!odb_transaction_nesting || bulk_fsync_objdir)
 327                 return;
 328
 329         bulk_fsync_objdir = tmp_objdir_create("bulk-fsync");
 330         if (bulk_fsync_objdir)
 331                 tmp_objdir_replace_primary_odb(bulk_fsync_objdir, 0);
 332 }
 333
 334 void fsync_loose_object_bulk_checkin(int fd, const char *filename)
 335 {
 336         /*
 337          * If we have an active ODB transaction, we issue a call that
 338          * cleans the filesystem page cache but avoids a hardware flush
 339          * command. Later on we will issue a single hardware flush
 340          * before renaming the objects to their final names as part of
 341          * flush_batch_fsync.
 342          */
 343         if (!bulk_fsync_objdir ||
 344             git_fsync(fd, FSYNC_WRITEOUT_ONLY) < 0) {
 345                 if (errno == ENOSYS)
 346                         warning(_("core.fsyncMethod = batch is unsupported on this platform"));
 347                 fsync_or_die(fd, filename);
 348         }
 349 }
 350
 351 int index_bulk_checkin(struct object_id *oid,
 352                        int fd, size_t size, enum object_type type,
 353                        const char *path, unsigned flags)
 354 {
 355         int status = deflate_to_pack(&bulk_checkin_packfile, oid, fd, size, type,
 356                                      path, flags);
 357         if (!odb_transaction_nesting)
 358                 flush_bulk_checkin_packfile(&bulk_checkin_packfile);
 359         return status;
 360 }
 361
 362 void begin_odb_transaction(void)
 363 {
 364         odb_transaction_nesting += 1;
 365 }
 366
 367 void flush_odb_transaction(void)
 368 {
 369         flush_batch_fsync();
 370         flush_bulk_checkin_packfile(&bulk_checkin_packfile);
 371 }
 372
 373 void end_odb_transaction(void)
 374 {
 375         odb_transaction_nesting -= 1;
 376         if (odb_transaction_nesting < 0)
 377                 BUG("Unbalanced ODB transaction nesting");
 378
 379         if (odb_transaction_nesting)
 380                 return;
 381
 382         flush_odb_transaction();
 383 }