/* searchs for name and returns a newly allocated hash of its value if set, * returns NULL if not found */ scr_hash* scr_param_get_hash(char* name) { scr_hash* hash = NULL; scr_hash* value_hash = NULL; /* see if this parameter is one which is restricted from user */ scr_hash* no_user = scr_hash_get(scr_no_user_hash, name); /* if parameter is set in environment, return that value */ if (no_user == NULL && getenv(name) != NULL) { /* TODO: need to strdup here to be safe? */ hash = scr_hash_new(); scr_hash_set(hash, getenv(name), scr_hash_new()); return hash; } /* otherwise, if parameter is set in user configuration file, return that value */ value_hash = scr_hash_get(scr_user_hash, name); if (no_user == NULL && value_hash != NULL) { hash = scr_hash_new(); scr_hash_merge(hash, value_hash); return hash; } /* otherwise, if parameter is set in system configuration file, return that value */ value_hash = scr_hash_get(scr_system_hash, name); if (value_hash != NULL) { hash = scr_hash_new(); scr_hash_merge(hash, value_hash); return hash; } /* parameter not found, return NULL */ return NULL; }
/* read config files and store contents */ int scr_param_init() { /* allocate storage and read in config files if we haven't already */ if (scr_param_ref_count == 0) { /* allocate hash object to hold names we cannot read from the * environment */ scr_no_user_hash = scr_hash_new(); scr_hash_set(scr_no_user_hash, "SCR_CNTL_BASE", scr_hash_new()); /* allocate hash object to store values from user config file, * if specified */ char* user_file = user_config_path(); if (user_file != NULL) { scr_user_hash = scr_hash_new(); scr_config_read(user_file, scr_user_hash); } scr_free(&user_file); /* allocate hash object to store values from system config file */ scr_system_hash = scr_hash_new(); scr_config_read(scr_config_file, scr_system_hash); /* initialize our hash to cache lookups to getenv */ scr_env_hash = scr_hash_new(); /* warn user if he set any parameters in his environment or user * config file which aren't permitted */ scr_hash_elem* elem; for (elem = scr_hash_elem_first(scr_no_user_hash); elem != NULL; elem = scr_hash_elem_next(elem)) { /* get the parameter name */ char* key = scr_hash_elem_key(elem); char* env_val = getenv(key); scr_hash* env_hash = scr_hash_get(scr_user_hash, key); /* check whether this is set in the environment */ if (env_val != NULL || env_hash != NULL) { scr_err("%s cannot be set in the environment or user configuration file, ignoring setting", key ); } } } /* increment our reference count */ scr_param_ref_count++; return SCR_SUCCESS; }
/* flush files specified in list, and record corresponding entries for summary file */ static int scr_flush_files_list(scr_hash* file_list, scr_hash* summary) { /* assume we will succeed in this flush */ int rc = SCR_SUCCESS; /* flush each of my files and fill in summary data structure */ scr_hash_elem* elem = NULL; scr_hash* files = scr_hash_get(file_list, SCR_KEY_FILE); for (elem = scr_hash_elem_first(files); elem != NULL; elem = scr_hash_elem_next(elem)) { /* get the filename */ char* file = scr_hash_elem_key(elem); /* convert file to path and extract name of file */ scr_path* path_name = scr_path_from_str(file); scr_path_basename(path_name); /* get the hash for this element */ scr_hash* hash = scr_hash_elem_hash(elem); /* get meta data for this file */ scr_meta* meta = scr_hash_get(hash, SCR_KEY_META); /* if segments are defined, we flush the file to its containers, * otherwise we copy the file out as is */ scr_hash* segments = scr_hash_get(hash, SCR_SUMMARY_6_KEY_SEGMENT); if (segments != NULL) { /* TODO: PRESERVE get original filename here */ /* add this file to the summary file */ char* name = scr_path_strdup(path_name); scr_hash* file_hash = scr_hash_set_kv(summary, SCR_SUMMARY_6_KEY_FILE, name); scr_free(&name); // USERDEF fixme! /* flush the file to the containers listed in its segmenets */ if (scr_flush_file_to_containers(file, meta, segments, scr_prefix) == SCR_SUCCESS) { /* successfully flushed this file, record the filesize */ unsigned long filesize = 0; if (scr_meta_get_filesize(meta, &filesize) == SCR_SUCCESS) { scr_hash_util_set_bytecount(file_hash, SCR_SUMMARY_6_KEY_SIZE, filesize); } /* record the crc32 if one was computed */ uLong crc = 0; if (scr_meta_get_crc32(meta, &crc) == SCR_SUCCESS) { scr_hash_util_set_crc32(file_hash, SCR_SUMMARY_6_KEY_CRC, crc); } /* record segment information in summary file */ scr_hash* segments_copy = scr_hash_new(); scr_hash_merge(segments_copy, segments); scr_hash_set(file_hash, SCR_SUMMARY_6_KEY_SEGMENT, segments_copy); } else { /* the flush failed */ rc = SCR_FAILURE; /* explicitly mark file as incomplete */ scr_hash_set_kv_int(file_hash, SCR_SUMMARY_6_KEY_COMPLETE, 0); } } else { /* get directory to flush file to */ char* dir; if (scr_hash_util_get_str(hash, SCR_KEY_PATH, &dir) == SCR_SUCCESS) { /* create full path of destination file */ scr_path* path_full = scr_path_from_str(dir); scr_path_append(path_full, path_name); /* get relative path to flushed file from SCR_PREFIX directory */ scr_path* path_relative = scr_path_relative(scr_prefix_path, path_full); if (! scr_path_is_null(path_relative)) { /* record the name of the file in the summary hash, and get reference to a hash for this file */ char* name = scr_path_strdup(path_relative); scr_hash* file_hash = scr_hash_set_kv(summary, SCR_SUMMARY_6_KEY_FILE, name); scr_free(&name); /* flush the file and fill in the meta data for this file */ if (scr_flush_a_file(file, dir, meta) == SCR_SUCCESS) { /* successfully flushed this file, record the filesize */ unsigned long filesize = 0; if (scr_meta_get_filesize(meta, &filesize) == SCR_SUCCESS) { scr_hash_util_set_bytecount(file_hash, SCR_SUMMARY_6_KEY_SIZE, filesize); } /* record the crc32 if one was computed */ uLong crc = 0; if (scr_meta_get_crc32(meta, &crc) == SCR_SUCCESS) { scr_hash_util_set_crc32(file_hash, SCR_SUMMARY_6_KEY_CRC, crc); } } else { /* the flush failed */ rc = SCR_FAILURE; /* explicitly mark incomplete files */ scr_hash_set_kv_int(file_hash, SCR_SUMMARY_6_KEY_COMPLETE, 0); } } else { scr_abort(-1, "Failed to get relative path to directory %s from %s @ %s:%d", dir, scr_prefix, __FILE__, __LINE__ ); } /* free relative and full paths */ scr_path_delete(&path_relative); scr_path_delete(&path_full); } else { scr_abort(-1, "Failed to read directory to flush file to @ %s:%d", __FILE__, __LINE__ ); } } /* free the file name path */ scr_path_delete(&path_name); } return rc; }
int main(int argc, char* argv[]) { int i, j; int index = 1; /* print usage if not enough arguments were given */ if (argc < 2) { printf("Usage: scr_rebuild_xor <size> <root> <missing_xor_filename> <ordered_remaining_xor_filenames>\n"); return 1; } /* TODO: want to pass this on command line? */ /* get current working directory */ char dsetdir[SCR_MAX_FILENAME]; scr_getcwd(dsetdir, sizeof(dsetdir)); /* create and reduce path for dataset */ scr_path* path_dset = scr_path_from_str(dsetdir); scr_path_reduce(path_dset); /* allocate buffers */ char* buffer_A = malloc(buffer_size * sizeof(char)); char* buffer_B = malloc(buffer_size * sizeof(char)); if (buffer_A == NULL || buffer_B == NULL) { scr_err("Failed to allocate buffer memory @ %s:%d", __FILE__, __LINE__ ); return 1; } /* read in the size of the XOR set */ int xor_set_size = (int) strtol(argv[index++], (char **)NULL, 10); if (xor_set_size <= 0) { scr_err("Invalid XOR set size argument %s @ %s:%d", argv[index-1], __FILE__, __LINE__ ); return 1; } /* allocate memory for data structures based on the XOR set size */ int* num_files = malloc(xor_set_size * sizeof(int)); int* offsets = malloc(xor_set_size * sizeof(int)); char** xor_files = malloc(xor_set_size * sizeof(char*)); int* xor_fds = malloc(xor_set_size * sizeof(int)); scr_hash** xor_headers = malloc(xor_set_size * sizeof(scr_hash*)); if (num_files == NULL || offsets == NULL || xor_files == NULL || xor_fds == NULL || xor_headers == NULL) { scr_err("Failed to allocate buffer memory @ %s:%d", __FILE__, __LINE__ ); return 1; } /* read in the rank of the missing process (the root) */ int root = (int) strtol(argv[index++], (char **)NULL, 10); if (root < 0 || root >= xor_set_size) { scr_err("Invalid root argument %s @ %s:%d", argv[index-1], __FILE__, __LINE__ ); return 1; } /* read in the missing xor filename */ xor_files[0] = strdup(argv[index++]); if (xor_files[0] == NULL) { scr_err("Failed to dup XOR filename @ %s:%d", __FILE__, __LINE__ ); return 1; } /* read in the xor filenames (expected to be in order of XOR segment number) */ /* we order ranks so that root is index 0, the rank to the right of root is index 1, and so on */ for (i=0; i < xor_set_size; i++) { xor_headers[i] = scr_hash_new(); /* we'll get the XOR file name for root from the header stored in the XOR file of the partner */ if (i == root) { continue; } /* adjust the index relative to root */ j = i - root; if (j < 0) { j += xor_set_size; } /* copy the XOR file name */ xor_files[j] = strdup(argv[index++]); if (xor_files[j] == NULL) { scr_err("Failed to dup XOR filename @ %s:%d", __FILE__, __LINE__ ); return 1; } } /* open each of the xor files and read in the headers */ for (i=1; i < xor_set_size; i++) { /* open each xor file for reading */ xor_fds[i] = scr_open(xor_files[i], O_RDONLY); if (xor_fds[i] < 0) { scr_err("Opening xor segment file: scr_open(%s) errno=%d %s @ %s:%d", xor_files[i], errno, strerror(errno), __FILE__, __LINE__ ); return 1; } /* read the header from this xor file */ if (scr_hash_read_fd(xor_files[i], xor_fds[i], xor_headers[i]) < 0) { scr_err("Failed to read XOR header from %s @ %s:%d", xor_files[i], __FILE__, __LINE__ ); return 1; } } /* build header for missing XOR file */ int partner_rank = -1; if (xor_set_size >= 2) { scr_hash_merge(xor_headers[0], xor_headers[1]); /* fetch our own file list from rank to our right */ scr_hash* rhs_hash = scr_hash_get(xor_headers[1], SCR_KEY_COPY_XOR_PARTNER); scr_hash* current_hash = scr_hash_new(); scr_hash_merge(current_hash, rhs_hash); scr_hash_set(xor_headers[0], SCR_KEY_COPY_XOR_CURRENT, current_hash); /* we are the partner to the rank to our left */ scr_hash* lhs_hash = scr_hash_get(xor_headers[xor_set_size-1], SCR_KEY_COPY_XOR_CURRENT); scr_hash* partner_hash = scr_hash_new(); scr_hash_merge(partner_hash, lhs_hash); scr_hash_set(xor_headers[0], SCR_KEY_COPY_XOR_PARTNER, partner_hash); /* get global rank of partner */ if (scr_hash_util_get_int(lhs_hash, SCR_KEY_COPY_XOR_RANK, &partner_rank) != SCR_SUCCESS) { scr_err("Failed to read partner rank from XOR file header in %s @ %s:%d", xor_files[xor_set_size-1], __FILE__, __LINE__ ); return 1; } } /* get a pointer to the current hash for the missing rank */ scr_hash* missing_current_hash = scr_hash_get(xor_headers[0], SCR_KEY_COPY_XOR_CURRENT); /* read the rank */ int my_rank = -1; if (scr_hash_util_get_int(missing_current_hash, SCR_KEY_COPY_XOR_RANK, &my_rank) != SCR_SUCCESS) { scr_err("Failed to read rank from XOR file header in %s @ %s:%d", xor_files[0], __FILE__, __LINE__ ); return 1; } /* get the dataset */ scr_dataset* dataset = scr_hash_get(xor_headers[0], SCR_KEY_COPY_XOR_DATASET); /* read the dataset id */ int dset_id = -1; if (scr_dataset_get_id(dataset, &dset_id) != SCR_SUCCESS) { scr_err("Failed to read dataset id from XOR file header in %s @ %s:%d", xor_files[0], __FILE__, __LINE__ ); return 1; } /* read the ranks */ int num_ranks = -1; if (scr_hash_util_get_int(xor_headers[0], SCR_KEY_COPY_XOR_RANKS, &num_ranks) != SCR_SUCCESS) { scr_err("Failed to read ranks from XOR file header in %s @ %s:%d", xor_files[0], __FILE__, __LINE__ ); return 1; } /* get name of partner's fmap */ scr_path* path_partner_map = scr_path_from_str(".scr"); scr_path_append_strf(path_partner_map, "fmap.%d.scr", partner_rank); /* extract partner's flush descriptor */ scr_hash* flushdesc = scr_hash_new(); scr_filemap* partner_map = scr_filemap_new(); scr_filemap_read(path_partner_map, partner_map); scr_filemap_get_flushdesc(partner_map, dset_id, partner_rank, flushdesc); scr_filemap_delete(&partner_map); /* delete partner map path */ scr_path_delete(&path_partner_map); /* determine whether we should preserve user directories */ int preserve_dirs = 0; scr_hash_util_get_int(flushdesc, SCR_SCAVENGE_KEY_PRESERVE, &preserve_dirs); /* read the chunk size */ unsigned long chunk_size = 0; if (scr_hash_util_get_unsigned_long(xor_headers[0], SCR_KEY_COPY_XOR_CHUNK, &chunk_size) != SCR_SUCCESS) { scr_err("Failed to read chunk size from XOR file header in %s @ %s:%d", xor_files[0], __FILE__, __LINE__ ); return 1; } /* determine number of files each member wrote in XOR set */ for (i=0; i < xor_set_size; i++) { /* record the number of files for this rank */ scr_hash* current_hash = scr_hash_get(xor_headers[i], SCR_KEY_COPY_XOR_CURRENT); if (scr_hash_util_get_int(current_hash, SCR_KEY_COPY_XOR_FILES, &num_files[i]) != SCR_SUCCESS) { scr_err("Failed to read number of files from %s @ %s:%d", xor_files[i], __FILE__, __LINE__ ); return 1; } } /* count the total number of files and set the offsets array */ int total_num_files = 0; for (i=0; i < xor_set_size; i++) { offsets[i] = total_num_files; total_num_files += num_files[i]; } /* allocate space for a file descriptor, file name pointer, and filesize for each user file */ int* user_fds = (int*) malloc(total_num_files * sizeof(int)); char** user_files = (char**) malloc(total_num_files * sizeof(char*)); char** user_rel_files = (char**) malloc(total_num_files * sizeof(char*)); unsigned long* user_filesizes = (unsigned long*) malloc(total_num_files * sizeof(unsigned long)); if (user_fds == NULL || user_files == NULL || user_rel_files == NULL || user_filesizes == NULL) { scr_err("Failed to allocate buffer memory @ %s:%d", __FILE__, __LINE__ ); return 1; } /* get file name, file size, and open each of the user files that we have */ for (i=0; i < xor_set_size; i++) { scr_hash* current_hash = scr_hash_get(xor_headers[i], SCR_KEY_COPY_XOR_CURRENT); /* for each file belonging to this rank, get filename, filesize, and open file */ for (j=0; j < num_files[i]; j++) { int offset = offsets[i] + j; /* get the meta data for this file */ scr_meta* meta = scr_hash_get_kv_int(current_hash, SCR_KEY_COPY_XOR_FILE, j); if (meta == NULL) { scr_err("Failed to read meta data for file %d in %s @ %s:%d", j, xor_files[i], __FILE__, __LINE__ ); return 1; } /* record the filesize of this file */ if (scr_meta_get_filesize(meta, &user_filesizes[offset]) != SCR_SUCCESS) { scr_err("Failed to read filesize field for file %d in %s @ %s:%d", j, xor_files[i], __FILE__, __LINE__ ); return 1; } /* get filename */ char* origname; if (scr_meta_get_origname(meta, &origname) != SCR_SUCCESS) { scr_err("Failed to read original name for file %d in %s @ %s:%d", j, xor_files[i], __FILE__, __LINE__ ); return 1; } /* construct full path to user file */ scr_path* path_user_full = scr_path_from_str(origname); if (preserve_dirs) { /* get original path of file */ char* origpath; if (scr_meta_get_origpath(meta, &origpath) != SCR_SUCCESS) { scr_err("Failed to read original path for file %d in %s @ %s:%d", j, xor_files[i], __FILE__, __LINE__ ); return 1; } /* construct full path to file */ scr_path_prepend_str(path_user_full, origpath); } else { /* construct full path to file */ scr_path_prepend(path_user_full, path_dset); } /* reduce path to user file */ scr_path_reduce(path_user_full); /* make a copy of the full path */ user_files[offset] = scr_path_strdup(path_user_full); /* make a copy of relative path */ scr_path* path_user_rel = scr_path_relative(path_dset, path_user_full); user_rel_files[offset] = scr_path_strdup(path_user_rel); scr_path_delete(&path_user_rel); /* free the full path */ scr_path_delete(&path_user_full); /* open the file */ if (i == 0) { /* create directory for file */ scr_path* user_dir_path = scr_path_from_str(user_files[offset]); scr_path_reduce(user_dir_path); scr_path_dirname(user_dir_path); if (! scr_path_is_null(user_dir_path)) { char* user_dir = scr_path_strdup(user_dir_path); mode_t mode_dir = scr_getmode(1, 1, 1); if (scr_mkdir(user_dir, mode_dir) != SCR_SUCCESS) { scr_err("Failed to create directory for user file %s @ %s:%d", user_dir, __FILE__, __LINE__ ); return 1; } scr_free(&user_dir); } scr_path_delete(&user_dir_path); /* open missing file for writing */ mode_t mode_file = scr_getmode(1, 1, 0); user_fds[offset] = scr_open(user_files[offset], O_WRONLY | O_CREAT | O_TRUNC, mode_file); if (user_fds[offset] < 0) { scr_err("Opening user file for writing: scr_open(%s) errno=%d %s @ %s:%d", user_files[offset], errno, strerror(errno), __FILE__, __LINE__ ); return 1; } } else { /* open existing file for reading */ user_fds[offset] = scr_open(user_files[offset], O_RDONLY); if (user_fds[offset] < 0) { scr_err("Opening user file for reading: scr_open(%s) errno=%d %s @ %s:%d", user_files[offset], errno, strerror(errno), __FILE__, __LINE__ ); return 1; } } } } /* finally, open the xor file for the missing rank */ mode_t mode_file = scr_getmode(1, 1, 0); xor_fds[0] = scr_open(xor_files[0], O_WRONLY | O_CREAT | O_TRUNC, mode_file); if (xor_fds[0] < 0) { scr_err("Opening xor file to be reconstructed: scr_open(%s) errno=%d %s @ %s:%d", xor_files[0], errno, strerror(errno), __FILE__, __LINE__ ); return 1; } int rc = 0; /* write the header to the XOR file of the missing rank */ if (scr_hash_write_fd(xor_files[0], xor_fds[0], xor_headers[0]) < 0) { rc = 1; } /* this offset array records the current position we are in the logical file for each rank */ unsigned long* offset = malloc(xor_set_size * sizeof(unsigned long)); if (offset == NULL) { scr_err("Failed to allocate buffer memory @ %s:%d", __FILE__, __LINE__ ); return 1; } for (i=0; i < xor_set_size; i++) { offset[i] = 0; } unsigned long write_pos = 0; int chunk_id; for (chunk_id = 0; chunk_id < xor_set_size && rc == 0; chunk_id++) { size_t nread = 0; while (nread < chunk_size && rc == 0) { /* read upto buffer_size bytes at a time */ size_t count = chunk_size - nread; if (count > buffer_size) { count = buffer_size; } /* clear our buffer */ memset(buffer_A, 0, count); /* read a segment from each rank and XOR it into our buffer */ for (i=1; i < xor_set_size; i++) { /* read the next set of bytes for this chunk from my file into send_buf */ if (chunk_id != ((i + root) % xor_set_size)) { /* read chunk from the logical file for this rank */ if (scr_read_pad_n(num_files[i], &user_files[offsets[i]], &user_fds[offsets[i]], buffer_B, count, offset[i], &user_filesizes[offsets[i]]) != SCR_SUCCESS) { /* our read failed, set the return code to an error */ rc = 1; count = 0; } offset[i] += count; } else { /* read chunk from the XOR file for this rank */ if (scr_read_attempt(xor_files[i], xor_fds[i], buffer_B, count) != count) { /* our read failed, set the return code to an error */ rc = 1; count = 0; } } /* TODO: XORing with unsigned long would be faster here (if chunk size is multiple of this size) */ /* merge the blocks via xor operation */ for (j = 0; j < count; j++) { buffer_A[j] ^= buffer_B[j]; } } /* at this point, we have the data from the missing rank, write it out */ if (chunk_id != root) { /* write chunk to logical file for the missing rank */ if (scr_write_pad_n(num_files[0], &user_files[0], &user_fds[0], buffer_A, count, write_pos, &user_filesizes[0]) != SCR_SUCCESS) { /* our write failed, set the return code to an error */ rc = 1; } write_pos += count; } else { /* write chunk to xor file for the missing rank */ if (scr_write_attempt(xor_files[0], xor_fds[0], buffer_A, count) != count) { /* our write failed, set the return code to an error */ rc = 1; } } nread += count; } } /* close each of the user files */ for (i=0; i < total_num_files; i++) { if (scr_close(user_files[i], user_fds[i]) != SCR_SUCCESS) { rc = 1; } } /* close each of the XOR files */ for (i=0; i < xor_set_size; i++) { if (scr_close(xor_files[i], xor_fds[i]) != SCR_SUCCESS) { rc = 1; } } /* if the write failed, delete the files we just wrote, and return an error */ if (rc != 0) { for (j=0; j < num_files[0]; j++) { scr_file_unlink(user_files[j]); } scr_file_unlink(xor_files[0]); return 1; } /* check that filesizes are correct */ unsigned long filesize; for (j=0; j < num_files[0]; j++) { filesize = scr_file_size(user_files[j]); if (filesize != user_filesizes[j]) { /* the filesize check failed, so delete the file */ scr_file_unlink(user_files[j]); /* mark the file as incomplete */ scr_meta* meta = scr_hash_get_kv_int(missing_current_hash, SCR_KEY_COPY_XOR_FILE, j); scr_meta_set_complete(meta, 0); rc = 1; } } /* TODO: we didn't record the filesize of the XOR file for the missing rank anywhere */ /* create a filemap for this rank */ scr_filemap* map = scr_filemap_new(); if (map == NULL) { scr_err("Failed to allocate filemap @ %s:%d", __FILE__, __LINE__ ); return 1; } /* record the dataset information in the filemap */ scr_filemap_set_dataset(map, dset_id, my_rank, dataset); /* write meta data for each of the user files and add each one to the filemap */ for (j=0; j < num_files[0]; j++) { /* add user file to filemap and record meta data */ char* user_file_relative = user_rel_files[j]; scr_filemap_add_file(map, dset_id, my_rank, user_file_relative); scr_meta* meta = scr_hash_get_kv_int(missing_current_hash, SCR_KEY_COPY_XOR_FILE, j); scr_filemap_set_meta(map, dset_id, my_rank, user_file_relative, meta); } /* write meta data for xor file and add it to the filemap */ scr_filemap_add_file(map, dset_id, my_rank, xor_files[0]); unsigned long full_chunk_filesize = scr_file_size(xor_files[0]); int missing_complete = 1; scr_meta* meta_chunk = scr_meta_new(); scr_meta_set_filename(meta_chunk, xor_files[0]); scr_meta_set_filetype(meta_chunk, SCR_META_FILE_XOR); scr_meta_set_filesize(meta_chunk, full_chunk_filesize); /* TODO: remove this from meta file, for now it's needed in scr_index.c */ scr_meta_set_ranks(meta_chunk, num_ranks); scr_meta_set_complete(meta_chunk, missing_complete); scr_filemap_set_meta(map, dset_id, my_rank, xor_files[0], meta_chunk); /* set expected number of files for the missing rank */ int expected_num_files = scr_filemap_num_files(map, dset_id, my_rank); scr_filemap_set_expected_files(map, dset_id, my_rank, expected_num_files); /* compute, check, and store crc values with files */ for (j=0; j < num_files[0]; j++) { /* compute crc on user file */ char* user_file_relative = user_rel_files[j]; if (scr_compute_crc(map, dset_id, my_rank, user_file_relative) != SCR_SUCCESS) { /* the crc check failed, so delete the file */ scr_file_unlink(user_files[j]); rc = 1; } } if (scr_compute_crc(map, dset_id, my_rank, xor_files[0]) != SCR_SUCCESS) { /* the crc check failed, so delete the file */ scr_file_unlink(xor_files[0]); rc = 1; } /* store flush descriptor */ scr_filemap_set_flushdesc(map, dset_id, my_rank, flushdesc); /* write filemap for this rank */ scr_path* path_map = scr_path_from_str(".scr"); scr_path_append_strf(path_map, "fmap.%d.scr", my_rank); if (scr_filemap_write(path_map, map) != SCR_SUCCESS) { rc = 1; } scr_path_delete(&path_map); /* delete the map */ scr_filemap_delete(&map); scr_meta_delete(&meta_chunk); /* delete the flush/scavenge descriptor */ scr_hash_delete(&flushdesc); scr_free(&offset); for (i=0; i < total_num_files; i++) { scr_free(&user_rel_files[i]); scr_free(&user_files[i]); } scr_free(&user_filesizes); scr_free(&user_rel_files); scr_free(&user_files); scr_free(&user_fds); for (i=0; i < xor_set_size; i++) { scr_hash_delete(&xor_headers[i]); } for (i=0; i < xor_set_size; i++) { scr_free(&xor_files[i]); } scr_free(&xor_headers); scr_free(&xor_fds); scr_free(&xor_files); scr_free(&offsets); scr_free(&num_files); scr_free(&buffer_B); scr_free(&buffer_A); scr_path_delete(&path_dset); return rc; }
/* read contents of summary file */ static int scr_fetch_summary( const char* summary_dir, scr_hash* file_list) { /* assume that we won't succeed in our fetch attempt */ int rc = SCR_SUCCESS; /* check whether summary file exists and is readable */ if (scr_my_rank_world == 0) { /* check that we can access the directory */ if (scr_file_is_readable(summary_dir) != SCR_SUCCESS) { scr_err("Failed to access summary directory %s @ %s:%d", summary_dir, __FILE__, __LINE__ ); rc = SCR_FAILURE; } } /* broadcast success code from rank 0 */ MPI_Bcast(&rc, 1, MPI_INT, 0, scr_comm_world); if (rc != SCR_SUCCESS) { return rc; } /* add path to file list */ scr_hash_util_set_str(file_list, SCR_KEY_PATH, summary_dir); /* build path to summary file */ scr_path* dataset_path = scr_path_from_str(summary_dir); scr_path* meta_path = scr_path_dup(dataset_path); scr_path_append_str(meta_path, ".scr"); scr_path_reduce(meta_path); /* rank 0 reads the summary file */ scr_hash* header = scr_hash_new(); if (scr_my_rank_world == 0) { /* build path to summary file */ scr_path* summary_path = scr_path_dup(meta_path); scr_path_append_str(summary_path, "summary.scr"); const char* summary_file = scr_path_strdup(summary_path); /* open file for reading */ int fd = scr_open(summary_file, O_RDONLY); if (fd >= 0) { /* read summary hash */ ssize_t header_size = scr_hash_read_fd(summary_file, fd, header); if (header_size < 0) { rc = SCR_FAILURE; } /* TODO: check that the version is correct */ /* close the file */ scr_close(summary_file, fd); } else { scr_err("Failed to open summary file %s @ %s:%d", summary_file, __FILE__, __LINE__ ); rc = SCR_FAILURE; } /* free summary path and string */ scr_free(&summary_file); scr_path_delete(&summary_path); } /* broadcast success code from rank 0 */ MPI_Bcast(&rc, 1, MPI_INT, 0, scr_comm_world); if (rc != SCR_SUCCESS) { goto cleanup; } /* broadcast the summary hash */ scr_hash_bcast(header, 0, scr_comm_world); /* extract and record the datast in file list */ scr_hash* dataset_hash = scr_hash_new(); scr_dataset* dataset = scr_hash_get(header, SCR_SUMMARY_6_KEY_DATASET); scr_hash_merge(dataset_hash, dataset); scr_hash_set(file_list, SCR_SUMMARY_6_KEY_DATASET, dataset_hash); /* build path to rank2file map */ scr_path* rank2file_path = scr_path_dup(meta_path); scr_path_append_str(rank2file_path, "rank2file.scr"); /* fetch file names and offsets containing file hash data */ int valid = 0; char* file = NULL; unsigned long offset = 0; if (scr_my_rank_world == 0) { /* rank 0 is only valid reader to start with */ valid = 1; file = scr_path_strdup(rank2file_path); offset = 0; } if (scr_fetch_rank2file_map(dataset_path, 1, &valid, &file, &offset) != SCR_SUCCESS) { rc = SCR_FAILURE; } /* create hashes to exchange data */ scr_hash* send = scr_hash_new(); scr_hash* recv = scr_hash_new(); /* read data from file */ if (valid) { /* open file if necessary */ int fd = scr_open(file, O_RDONLY); if (fd >= 0) { /* create hash to hold file contents */ scr_hash* save = scr_hash_new(); /* read hash from file */ scr_lseek(file, fd, offset, SEEK_SET); ssize_t readsize = scr_hash_read_fd(file, fd, save); if (readsize < 0) { scr_err("Failed to read rank2file map file %s @ %s:%d", file, __FILE__, __LINE__ ); rc = SCR_FAILURE; } /* check that the number of ranks match */ int ranks = 0; scr_hash_util_get_int(save, SCR_SUMMARY_6_KEY_RANKS, &ranks); if (ranks != scr_ranks_world) { scr_err("Invalid number of ranks in %s, got %d expected %d @ %s:%d", file, ranks, scr_ranks_world, __FILE__, __LINE__ ); rc = SCR_FAILURE; } /* delete current send hash, set it to values from file, * delete file hash */ scr_hash_delete(&send); send = scr_hash_extract(save, SCR_SUMMARY_6_KEY_RANK); scr_hash_delete(&save); /* close the file */ scr_close(file, fd); } else { scr_err("Failed to open rank2file map %s @ %s:%d", file, __FILE__, __LINE__ ); rc = SCR_FAILURE; } /* delete file name string */ scr_free(&file); } /* check that everyone read the data ok */ if (! scr_alltrue(rc == SCR_SUCCESS)) { rc = SCR_FAILURE; goto cleanup_hashes; } /* scatter to groups */ scr_hash_exchange_direction(send, recv, scr_comm_world, SCR_HASH_EXCHANGE_RIGHT); /* iterate over the ranks that sent data to us, and set up our * list of files */ scr_hash_elem* elem; for (elem = scr_hash_elem_first(recv); elem != NULL; elem = scr_hash_elem_next(elem)) { /* the key is the source rank, which we don't care about, * the info we need is in the element hash */ scr_hash* elem_hash = scr_hash_elem_hash(elem); /* get pointer to file hash */ scr_hash* file_hash = scr_hash_get(elem_hash, SCR_SUMMARY_6_KEY_FILE); if (file_hash != NULL) { /* TODO: parse summary file format */ scr_hash_merge(file_list, elem_hash); } else { rc = SCR_FAILURE; } } /* fill in file list parameters */ if (rc == SCR_SUCCESS) { /* if we're not using containers, add PATH entry for each of our * files */ scr_hash* files = scr_hash_get(file_list, SCR_KEY_FILE); for (elem = scr_hash_elem_first(files); elem != NULL; elem = scr_hash_elem_next(elem)) { /* get the file name */ char* file = scr_hash_elem_key(elem); /* combine the file name with the summary directory to build a * full path to the file */ scr_path* path_full = scr_path_dup(dataset_path); scr_path_append_str(path_full, file); /* subtract off last component to get just the path */ scr_path_dirname(path_full); char* path = scr_path_strdup(path_full); /* record path in file list */ scr_hash* hash = scr_hash_elem_hash(elem); scr_hash_util_set_str(hash, SCR_KEY_PATH, path); /* free the path and string */ scr_free(&path); scr_path_delete(&path_full); } } /* check that everyone read the data ok */ if (! scr_alltrue(rc == SCR_SUCCESS)) { rc = SCR_FAILURE; goto cleanup_hashes; } cleanup_hashes: /* delete send and receive hashes */ scr_hash_delete(&recv); scr_hash_delete(&send); /* free string and path for rank2file map */ scr_path_delete(&rank2file_path); cleanup: /* free the header hash */ scr_hash_delete(&header); /* free path for dataset directory */ scr_path_delete(&meta_path); scr_path_delete(&dataset_path); return rc; }
/* apply XOR redundancy scheme to dataset files */ static int scr_reddesc_apply_xor(scr_filemap* map, const scr_reddesc* c, int id) { int rc = SCR_SUCCESS; int i; /* get pointer to XOR state structure */ scr_reddesc_xor* state = (scr_reddesc_xor*) c->copy_state; /* allocate buffer to read a piece of my file */ char* send_buf = (char*) scr_align_malloc(scr_mpi_buf_size, scr_page_size); if (send_buf == NULL) { scr_abort(-1, "Allocating memory for send buffer: malloc(%d) errno=%d %s @ %s:%d", scr_mpi_buf_size, errno, strerror(errno), __FILE__, __LINE__ ); } /* allocate buffer to read a piece of the recevied chunk file */ char* recv_buf = (char*) scr_align_malloc(scr_mpi_buf_size, scr_page_size); if (recv_buf == NULL) { scr_abort(-1, "Allocating memory for recv buffer: malloc(%d) errno=%d %s @ %s:%d", scr_mpi_buf_size, errno, strerror(errno), __FILE__, __LINE__ ); } /* count the number of files I have and allocate space in structures for each of them */ int num_files = scr_filemap_num_files(map, id, scr_my_rank_world); int* fds = (int*) SCR_MALLOC(num_files * sizeof(int)); char** filenames = (char**) SCR_MALLOC(num_files * sizeof(char*)); unsigned long* filesizes = (unsigned long*) SCR_MALLOC(num_files * sizeof(unsigned long)); /* record partner's redundancy descriptor hash in our filemap */ scr_hash* lhs_desc_hash = scr_hash_new(); scr_hash* my_desc_hash = scr_hash_new(); scr_reddesc_store_to_hash(c, my_desc_hash); scr_hash_sendrecv(my_desc_hash, state->rhs_rank, lhs_desc_hash, state->lhs_rank, c->comm); scr_filemap_set_desc(map, id, state->lhs_rank_world, lhs_desc_hash); scr_hash_delete(&my_desc_hash); scr_hash_delete(&lhs_desc_hash); /* allocate a new xor file header hash */ scr_hash* header = scr_hash_new(); /* record the global ranks of the processes in our xor group */ scr_hash_merge(header, state->group_map); /* record dataset in header */ scr_hash* dataset = scr_hash_new(); scr_filemap_get_dataset(map, id, scr_my_rank_world, dataset); scr_hash_set(header, SCR_KEY_COPY_XOR_DATASET, dataset); /* open each file, get the filesize of each, and read the meta data of each */ scr_hash* current_files = scr_hash_new(); int file_count = 0; unsigned long my_bytes = 0; scr_hash_elem* file_elem; for (file_elem = scr_filemap_first_file(map, id, scr_my_rank_world); file_elem != NULL; file_elem = scr_hash_elem_next(file_elem)) { /* get the filename */ filenames[file_count] = scr_hash_elem_key(file_elem); /* get the filesize of this file and add the byte count to the total */ filesizes[file_count] = scr_file_size(filenames[file_count]); my_bytes += filesizes[file_count]; /* read the meta data for this file and insert it into the current_files hash */ scr_meta* file_hash = scr_meta_new(); scr_filemap_get_meta(map, id, scr_my_rank_world, filenames[file_count], file_hash); scr_hash_setf(current_files, file_hash, "%d", file_count); /* open the file */ fds[file_count] = scr_open(filenames[file_count], O_RDONLY); if (fds[file_count] < 0) { /* TODO: try again? */ scr_abort(-1, "Opening checkpoint file for copying: scr_open(%s, O_RDONLY) errno=%d %s @ %s:%d", filenames[file_count], errno, strerror(errno), __FILE__, __LINE__ ); } file_count++; } /* set total number of files we have, plus our rank */ scr_hash* current_hash = scr_hash_new(); scr_hash_set_kv_int(current_hash, SCR_KEY_COPY_XOR_RANK, scr_my_rank_world); scr_hash_set_kv_int(current_hash, SCR_KEY_COPY_XOR_FILES, file_count); scr_hash_set(current_hash, SCR_KEY_COPY_XOR_FILE, current_files); /* exchange file info with partners and add data to our header */ scr_hash* partner_hash = scr_hash_new(); scr_hash_sendrecv(current_hash, state->rhs_rank, partner_hash, state->lhs_rank, c->comm); scr_hash_set(header, SCR_KEY_COPY_XOR_CURRENT, current_hash); scr_hash_set(header, SCR_KEY_COPY_XOR_PARTNER, partner_hash); /* allreduce to get maximum filesize */ unsigned long max_bytes; MPI_Allreduce(&my_bytes, &max_bytes, 1, MPI_UNSIGNED_LONG, MPI_MAX, c->comm); /* TODO: use unsigned long integer arithmetic (with proper byte padding) instead of char to speed things up */ /* compute chunk size according to maximum file length and number of ranks in xor set */ /* if filesize doesn't divide evenly, then add one byte to chunk_size */ /* TODO: check that ranks > 1 for this divide to be safe (or at partner selection time) */ size_t chunk_size = max_bytes / (unsigned long) (c->ranks - 1); if ((c->ranks - 1) * chunk_size < max_bytes) { chunk_size++; } /* TODO: need something like this to handle 0-byte files? */ if (chunk_size == 0) { chunk_size++; } /* record the dataset id and the chunk size in the xor chunk header */ scr_hash_util_set_bytecount(header, SCR_KEY_COPY_XOR_CHUNK, chunk_size); /* set chunk filenames of form: xor.<group_id>_<xor_rank+1>_of_<xor_ranks>.scr */ char my_chunk_file[SCR_MAX_FILENAME]; char* dir = scr_cache_dir_hidden_get(c, id); sprintf(my_chunk_file, "%s/xor.%d_%d_of_%d.scr", dir, c->group_id, c->rank+1, c->ranks); scr_free(&dir); /* record chunk file in filemap before creating it */ scr_filemap_add_file(map, id, scr_my_rank_world, my_chunk_file); scr_filemap_write(scr_map_file, map); /* open my chunk file */ mode_t mode_file = scr_getmode(1, 1, 0); int fd_chunk = scr_open(my_chunk_file, O_WRONLY | O_CREAT | O_TRUNC, mode_file); if (fd_chunk < 0) { /* TODO: try again? */ scr_abort(-1, "Opening XOR chunk file for writing: scr_open(%s) errno=%d %s @ %s:%d", my_chunk_file, errno, strerror(errno), __FILE__, __LINE__ ); } /* write out the xor chunk header */ scr_hash_write_fd(my_chunk_file, fd_chunk, header); scr_hash_delete(&header); MPI_Request request[2]; MPI_Status status[2]; /* XOR Reduce_scatter */ size_t nread = 0; while (nread < chunk_size) { size_t count = chunk_size - nread; if (count > scr_mpi_buf_size) { count = scr_mpi_buf_size; } int chunk_id; for(chunk_id = c->ranks-1; chunk_id >= 0; chunk_id--) { /* read the next set of bytes for this chunk from my file into send_buf */ if (chunk_id > 0) { int chunk_id_rel = (c->rank + c->ranks + chunk_id) % c->ranks; if (chunk_id_rel > c->rank) { chunk_id_rel--; } unsigned long offset = chunk_size * (unsigned long) chunk_id_rel + nread; if (scr_read_pad_n(num_files, filenames, fds, send_buf, count, offset, filesizes) != SCR_SUCCESS) { rc = SCR_FAILURE; } } else { memset(send_buf, 0, count); } /* TODO: XORing with unsigned long would be faster here (if chunk size is multiple of this size) */ /* merge the blocks via xor operation */ if (chunk_id < c->ranks-1) { for (i = 0; i < count; i++) { send_buf[i] ^= recv_buf[i]; } } if (chunk_id > 0) { /* not our chunk to write, forward it on and get the next */ MPI_Irecv(recv_buf, count, MPI_BYTE, state->lhs_rank, 0, c->comm, &request[0]); MPI_Isend(send_buf, count, MPI_BYTE, state->rhs_rank, 0, c->comm, &request[1]); MPI_Waitall(2, request, status); } else { /* write send block to send chunk file */ if (scr_write_attempt(my_chunk_file, fd_chunk, send_buf, count) != count) { rc = SCR_FAILURE; } } } nread += count; } /* close my chunkfile, with fsync */ if (scr_close(my_chunk_file, fd_chunk) != SCR_SUCCESS) { rc = SCR_FAILURE; } /* close my dataset files */ for (i=0; i < num_files; i++) { scr_close(filenames[i], fds[i]); } /* free the buffers */ scr_free(&filesizes); /* in this case, we don't free each name, since we copied the pointer to the string in the filemap */ scr_free(&filenames); scr_free(&fds); scr_align_free(&send_buf); scr_align_free(&recv_buf); /* TODO: need to check for errors */ /* write meta file for xor chunk */ unsigned long my_chunk_file_size = scr_file_size(my_chunk_file); scr_meta* meta = scr_meta_new(); scr_meta_set_filename(meta, my_chunk_file); scr_meta_set_filetype(meta, SCR_META_FILE_XOR); scr_meta_set_filesize(meta, my_chunk_file_size); scr_meta_set_complete(meta, 1); /* TODODSET: move the ranks field elsewhere, for now it's needed by scr_index.c */ scr_meta_set_ranks(meta, scr_ranks_world); scr_filemap_set_meta(map, id, scr_my_rank_world, my_chunk_file, meta); scr_filemap_write(scr_map_file, map); scr_meta_delete(&meta); /* if crc_on_copy is set, compute and store CRC32 value for chunk file */ if (scr_crc_on_copy) { scr_compute_crc(map, id, scr_my_rank_world, my_chunk_file); /* TODO: would be nice to save this CRC in our partner's XOR file so we can check correctness on a rebuild */ } return rc; }