/* this transfers redundancy descriptors for the given dataset id */
static int scr_distribute_reddescs(scr_filemap* map, int id, scr_reddesc* red)
{
int i;
/* create a new hash to record redundancy descriptors that we have */
scr_hash* send_hash = scr_hash_new();
/* for this dataset, get list of ranks we have data for */
int nranks = 0;
int* ranks = NULL;
scr_filemap_list_ranks_by_dataset(map, id, &nranks, &ranks);
/* for each rank we have files for, check whether we also have
* its redundancy descriptor */
int invalid_rank_found = 0;
for (i=0; i < nranks; i++) {
/* get the rank id */
int rank = ranks[i];
/* check that the rank is within range */
if (rank < 0 || rank >= scr_ranks_world) {
scr_err("Invalid rank id %d in world of %d @ %s:%d",
rank, scr_ranks_world, __FILE__, __LINE__
);
invalid_rank_found = 1;
}
/* lookup the redundancy descriptor hash for this rank */
scr_hash* desc = scr_hash_new();
scr_filemap_get_desc(map, id, rank, desc);
/* if this descriptor has entries, add it to our send hash,
* delete the hash otherwise */
if (scr_hash_size(desc) > 0) {
scr_hash_setf(send_hash, desc, "%d", rank);
} else {
scr_hash_delete(&desc);
}
}
/* free off our list of ranks */
scr_free(&ranks);
/* check that we didn't find an invalid rank on any process */
if (! scr_alltrue(invalid_rank_found == 0)) {
scr_hash_delete(&send_hash);
return SCR_FAILURE;
}
/* create an empty hash to receive any incoming descriptors */
/* exchange descriptors with other ranks */
scr_hash* recv_hash = scr_hash_new();
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* check that everyone can get their descriptor */
int num_desc = scr_hash_size(recv_hash);
if (! scr_alltrue(num_desc > 0)) {
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
scr_dbg(2, "Cannot find process that has my redundancy descriptor @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* just go with the first redundancy descriptor in our list,
* they should all be the same */
scr_hash_elem* desc_elem = scr_hash_elem_first(recv_hash);
scr_hash* desc_hash = scr_hash_elem_hash(desc_elem);
/* record the descriptor in our filemap */
scr_filemap_set_desc(map, id, scr_my_rank_world, desc_hash);
scr_filemap_write(scr_map_file, map);
/* TODO: at this point, we could delete descriptors for other
* ranks for this checkpoint */
/* create our redundancy descriptor struct from the map */
scr_reddesc_create_from_filemap(map, id, scr_my_rank_world, red);
/* free off our send and receive hashes */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
return SCR_SUCCESS;
}
/* this moves all files of the specified dataset in the cache to
* make them accessible to new rank mapping */
static int scr_distribute_files(scr_filemap* map, const scr_reddesc* red, int id)
{
int i, round;
int rc = SCR_SUCCESS;
/* TODO: mark dataset as being distributed in filemap,
* because if we fail in the middle of a distribute,
* we can't trust the contents of the files anymore,
* at which point it should be deleted */
/* clean out any incomplete files before we start */
scr_cache_clean(map);
/* for this dataset, get list of ranks we have data for */
int nranks = 0;
int* ranks = NULL;
scr_filemap_list_ranks_by_dataset(map, id, &nranks, &ranks);
/* walk backwards through the list of ranks, and set our start index
* to the rank which is the first rank that is equal to or higher
* than our own rank -- when we assign round ids below, this offsetting
* helps distribute the load */
int start_index = 0;
int invalid_rank_found = 0;
for (i = nranks-1; i >= 0; i--) {
int rank = ranks[i];
/* pick the first rank whose rank id is equal to or higher than our own */
if (rank >= scr_my_rank_world) {
start_index = i;
}
/* while we're at it, check that the rank is within range */
if (rank < 0 || rank >= scr_ranks_world) {
scr_err("Invalid rank id %d in world of %d @ %s:%d",
rank, scr_ranks_world, __FILE__, __LINE__
);
invalid_rank_found = 1;
}
}
/* check that we didn't find an invalid rank on any process */
if (! scr_alltrue(invalid_rank_found == 0)) {
scr_free(&ranks);
return SCR_FAILURE;
}
/* allocate array to record the rank we can send to in each round */
int* have_rank_by_round = (int*) SCR_MALLOC(sizeof(int) * nranks);
int* send_flag_by_round = (int*) SCR_MALLOC(sizeof(int) * nranks);
/* check that we have all of the files for each rank,
* and determine the round we can send them */
scr_hash* send_hash = scr_hash_new();
scr_hash* recv_hash = scr_hash_new();
for (round = 0; round < nranks; round++) {
/* get the rank id */
int index = (start_index + round) % nranks;
int rank = ranks[index];
/* record the rank indexed by the round number */
have_rank_by_round[round] = rank;
/* assume we won't be sending to this rank in this round */
send_flag_by_round[round] = 0;
/* if we have files for this rank, specify the round we can
* send those files in */
if (scr_bool_have_files(map, id, rank)) {
scr_hash_setf(send_hash, NULL, "%d %d", rank, round);
}
}
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* search for the minimum round we can get our files */
int retrieve_rank = -1;
int retrieve_round = -1;
scr_hash_elem* elem = NULL;
for (elem = scr_hash_elem_first(recv_hash);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the rank id */
int rank = scr_hash_elem_key_int(elem);
/* get the round id */
scr_hash* round_hash = scr_hash_elem_hash(elem);
scr_hash_elem* round_elem = scr_hash_elem_first(round_hash);
char* round_str = scr_hash_elem_key(round_elem);
int round = atoi(round_str);
/* record this round and rank number if it's less than the current round */
if (round < retrieve_round || retrieve_round == -1) {
retrieve_round = round;
retrieve_rank = rank;
}
}
/* done with the round hashes, free them off */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
/* free off our list of ranks */
scr_free(&ranks);
/* for some redundancy schemes, we know at this point whether we
* can recover all files */
int can_get_files = (retrieve_rank != -1);
if (red->copy_type != SCR_COPY_XOR && !scr_alltrue(can_get_files)) {
/* print a debug message indicating which rank is missing files */
if (! can_get_files) {
scr_dbg(2, "Cannot find process that has my checkpoint files @ %s:%d",
__FILE__, __LINE__
);
}
return SCR_FAILURE;
}
/* get the maximum retrieve round */
int max_rounds = 0;
MPI_Allreduce(
&retrieve_round, &max_rounds, 1, MPI_INT, MPI_MAX, scr_comm_world
);
/* tell destination which round we'll take our files in */
send_hash = scr_hash_new();
recv_hash = scr_hash_new();
if (retrieve_rank != -1) {
scr_hash_setf(send_hash, NULL, "%d %d", retrieve_rank, retrieve_round);
}
scr_hash_exchange(send_hash, recv_hash, scr_comm_world);
/* determine which ranks want to fetch their files from us */
for(elem = scr_hash_elem_first(recv_hash);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the round id */
scr_hash* round_hash = scr_hash_elem_hash(elem);
scr_hash_elem* round_elem = scr_hash_elem_first(round_hash);
char* round_str = scr_hash_elem_key(round_elem);
int round = atoi(round_str);
/* record whether this rank wants its files from us */
if (round >= 0 && round < nranks) {
send_flag_by_round[round] = 1;
}
}
/* done with the round hashes, free them off */
scr_hash_delete(&recv_hash);
scr_hash_delete(&send_hash);
int tmp_rc = 0;
/* run through rounds and exchange files */
for (round = 0; round <= max_rounds; round++) {
/* assume we don't need to send or receive any files this round */
int send_rank = MPI_PROC_NULL;
int recv_rank = MPI_PROC_NULL;
int send_num = 0;
int recv_num = 0;
/* check whether I can potentially send to anyone in this round */
if (round < nranks) {
/* have someone's files, check whether they are asking
* for them this round */
if (send_flag_by_round[round]) {
/* need to send files this round, remember to whom and how many */
int dst_rank = have_rank_by_round[round];
send_rank = dst_rank;
send_num = scr_filemap_num_files(map, id, dst_rank);
}
}
/* if I'm supposed to get my files this round, set the recv_rank */
if (retrieve_round == round) {
recv_rank = retrieve_rank;
}
/* TODO: another special case is to just move files if the
* processes are on the same node */
/* if i'm sending to myself, just move (rename) each file */
if (send_rank == scr_my_rank_world) {
/* get our file list */
int numfiles = 0;
char** files = NULL;
scr_filemap_list_files(map, id, send_rank, &numfiles, &files);
/* TODO: sort files in reverse order by size */
/* iterate over and rename each file */
for (i=0; i < numfiles; i++) {
/* get the current file name */
char* file = files[i];
/* lookup meta data for this file */
scr_meta* meta = scr_meta_new();
scr_filemap_get_meta(map, id, send_rank, file, meta);
/* get the path for this file based on its type
* and dataset id */
char* dir = NULL;
if (scr_meta_check_filetype(meta, SCR_META_FILE_USER) == SCR_SUCCESS) {
dir = scr_cache_dir_get(red, id);
} else {
dir = scr_cache_dir_hidden_get(red, id);
}
/* build the new file name */
scr_path* path_newfile = scr_path_from_str(file);
scr_path_basename(path_newfile);
scr_path_prepend_str(path_newfile, dir);
char* newfile = scr_path_strdup(path_newfile);
/* if the new file name is different from the old name, rename it */
if (strcmp(file, newfile) != 0) {
/* record the new filename to our map and write it to disk */
scr_filemap_add_file(map, id, send_rank, newfile);
scr_filemap_set_meta(map, id, send_rank, newfile, meta);
scr_filemap_write(scr_map_file, map);
/* rename the file */
scr_dbg(2, "Round %d: rename(%s, %s)", round, file, newfile);
tmp_rc = rename(file, newfile);
if (tmp_rc != 0) {
/* TODO: to cross mount points, if tmp_rc == EXDEV,
* open new file, copy, and delete orig */
scr_err("Moving checkpoint file: rename(%s, %s) %s errno=%d @ %s:%d",
file, newfile, strerror(errno), errno, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* remove the old name from the filemap and write it to disk */
scr_filemap_remove_file(map, id, send_rank, file);
scr_filemap_write(scr_map_file, map);
}
/* free the path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
/* free directory string */
scr_free(&dir);
/* free meta data */
scr_meta_delete(&meta);
}
/* free the list of filename pointers */
scr_free(&files);
} else {
/* if we have files for this round, but the correspdonding
* rank doesn't need them, delete the files */
if (round < nranks && send_rank == MPI_PROC_NULL) {
int dst_rank = have_rank_by_round[round];
scr_unlink_rank(map, id, dst_rank);
}
/* sending to and/or recieving from another node */
if (send_rank != MPI_PROC_NULL || recv_rank != MPI_PROC_NULL) {
/* have someone to send to or receive from */
int have_outgoing = 0;
int have_incoming = 0;
if (send_rank != MPI_PROC_NULL) {
have_outgoing = 1;
}
if (recv_rank != MPI_PROC_NULL) {
have_incoming = 1;
}
/* first, determine how many files I will be receiving and
* tell how many I will be sending */
MPI_Request request[2];
MPI_Status status[2];
int num_req = 0;
if (have_incoming) {
MPI_Irecv(
&recv_num, 1, MPI_INT, recv_rank, 0,
scr_comm_world, &request[num_req]
);
num_req++;
}
if (have_outgoing) {
MPI_Isend(
&send_num, 1, MPI_INT, send_rank, 0,
scr_comm_world, &request[num_req]
);
num_req++;
}
if (num_req > 0) {
MPI_Waitall(num_req, request, status);
}
/* record how many files I will receive (need to distinguish
* between 0 files and not knowing) */
if (have_incoming) {
scr_filemap_set_expected_files(map, id, scr_my_rank_world, recv_num);
}
/* turn off send or receive flags if the file count is 0,
* nothing else to do */
if (send_num == 0) {
have_outgoing = 0;
send_rank = MPI_PROC_NULL;
}
if (recv_num == 0) {
have_incoming = 0;
recv_rank = MPI_PROC_NULL;
}
/* TODO: since we overwrite files in place in order to avoid
* running out of storage space, we should sort files in order
* of descending size for the next step */
/* get our file list for the destination */
int numfiles = 0;
char** files = NULL;
if (have_outgoing) {
scr_filemap_list_files(map, id, send_rank, &numfiles, &files);
}
/* while we have a file to send or receive ... */
while (have_incoming || have_outgoing) {
/* get the filename */
char* file = NULL;
scr_meta* send_meta = NULL;
if (have_outgoing) {
file = files[numfiles - send_num];
send_meta = scr_meta_new();
scr_filemap_get_meta(map, id, send_rank, file, send_meta);
}
/* exchange meta data so we can determine type of incoming file */
scr_meta* recv_meta = scr_meta_new();
scr_hash_sendrecv(send_meta, send_rank, recv_meta, recv_rank, scr_comm_world);
/* get the path for this file based on its type and dataset id */
char* dir = NULL;
if (have_incoming) {
if (scr_meta_check_filetype(recv_meta, SCR_META_FILE_USER) == SCR_SUCCESS) {
dir = scr_cache_dir_get(red, id);
} else {
dir = scr_cache_dir_hidden_get(red, id);
}
}
/* exhange file names with partners,
* building full path of incoming file */
char file_partner[SCR_MAX_FILENAME];
scr_swap_file_names(
file, send_rank, file_partner, sizeof(file_partner), recv_rank,
dir, scr_comm_world
);
/* free directory string */
scr_free(&dir);
/* free incoming meta data (we'll get this again later) */
scr_meta_delete(&recv_meta);
/* if we'll receive a file, record the name of our file
* in the filemap and write it to disk */
recv_meta = NULL;
if (recv_rank != MPI_PROC_NULL) {
recv_meta = scr_meta_new();
scr_filemap_add_file(map, id, scr_my_rank_world, file_partner);
scr_filemap_write(scr_map_file, map);
}
/* either sending or receiving a file this round, since we move files,
* it will be deleted or overwritten */
if (scr_swap_files(MOVE_FILES, file, send_meta, send_rank,
file_partner, recv_meta, recv_rank, scr_comm_world) != SCR_SUCCESS)
{
scr_err("Swapping files: %s to %d, %s from %d @ %s:%d",
file, send_rank, file_partner, recv_rank, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* if we received a file, record its meta data and decrement
* our receive count */
if (have_incoming) {
/* record meta data for the file we received */
scr_filemap_set_meta(map, id, scr_my_rank_world, file_partner, recv_meta);
scr_meta_delete(&recv_meta);
/* decrement receive count */
recv_num--;
if (recv_num == 0) {
have_incoming = 0;
recv_rank = MPI_PROC_NULL;
}
}
/* if we sent a file, remove it from the filemap and decrement
* our send count */
if (have_outgoing) {
/* remove file from the filemap */
scr_filemap_remove_file(map, id, send_rank, file);
scr_meta_delete(&send_meta);
/* decrement our send count */
send_num--;
if (send_num == 0) {
have_outgoing = 0;
send_rank = MPI_PROC_NULL;
}
}
/* update filemap on disk */
scr_filemap_write(scr_map_file, map);
}
/* free our file list */
scr_free(&files);
}
}
}
/* if we have more rounds than max rounds, delete the remainder of our files */
for (round = max_rounds+1; round < nranks; round++) {
/* have someone's files for this round, so delete them */
int dst_rank = have_rank_by_round[round];
scr_unlink_rank(map, id, dst_rank);
}
scr_free(&send_flag_by_round);
scr_free(&have_rank_by_round);
/* write out new filemap and free the memory resources */
scr_filemap_write(scr_map_file, map);
/* clean out any incomplete files */
scr_cache_clean(map);
/* TODO: if the exchange or redundancy rebuild failed,
* we should also delete any *good* files we received */
/* return whether distribute succeeded, it does not ensure we have
* all of our files, only that the transfer completed without failure */
return rc;
}
/* given a filename, its meta data, its list of segments, and list of destination containers,
* copy file to container files */
static int scr_flush_file_to_containers(
const char* file,
scr_meta* meta,
scr_hash* segments,
const char* dst_dir)
{
/* check that we got something for a source file */
if (file == NULL || strcmp(file, "") == 0) {
scr_err("Invalid source file @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* check that our other arguments are valid */
if (meta == NULL || segments == NULL) {
scr_err("Invalid metadata or segments @ %s:%d",
__FILE__, __LINE__
);
return SCR_FAILURE;
}
/* open the file for reading */
int fd_src = scr_open(file, O_RDONLY);
if (fd_src < 0) {
scr_err("Opening file to copy: scr_open(%s) errno=%d %s @ %s:%d",
file, errno, strerror(errno), __FILE__, __LINE__
);
return SCR_FAILURE;
}
#if !defined(__APPLE__)
/* TODO:
posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL)
that tells the kernel that you don't ever need the pages
from the file again, and it won't bother keeping them in the page cache.
*/
posix_fadvise(fd_src, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL);
#endif
/* get the buffer size we'll use to write to the file */
unsigned long buf_size = scr_file_buf_size;
/* allocate buffer to read in file chunks */
char* buf = (char*) SCR_MALLOC(buf_size);
/* initialize crc value */
uLong crc;
if (scr_crc_on_flush) {
crc = crc32(0L, Z_NULL, 0);
}
int rc = SCR_SUCCESS;
/* write out each segment */
scr_hash_sort_int(segments, SCR_HASH_SORT_ASCENDING);
scr_hash_elem* elem;
for (elem = scr_hash_elem_first(segments);
elem != NULL;
elem = scr_hash_elem_next(elem))
{
/* get the container info for this segment */
scr_hash* hash = scr_hash_elem_hash(elem);
/* get the offset into the container and the length of the segment (both in bytes) */
char* container_name;
unsigned long container_offset, segment_length;
if (scr_container_get_name_offset_length(hash,
&container_name, &container_offset, &segment_length) != SCR_SUCCESS)
{
scr_err("Failed to get segment offset and length @ %s:%d",
__FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
/* build full name to destination file */
scr_path* dst_path = scr_path_from_str(dst_dir);
scr_path_append_str(dst_path, container_name);
scr_path_reduce(dst_path);
char* dst_file = scr_path_strdup(dst_path);
/* open container file for writing -- we don't truncate here because more than one
* process may be writing to the same file */
int fd_container = scr_open(dst_file, O_WRONLY);
if (fd_container < 0) {
scr_err("Opening file for writing: scr_open(%s) errno=%d %s @ %s:%d",
dst_file, errno, strerror(errno), __FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
#if !defined(__APPLE__)
/* TODO:
posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL)
that tells the kernel that you don't ever need the pages
from the file again, and it won't bother keeping them in the page cache.
*/
posix_fadvise(fd_container, 0, 0, POSIX_FADV_DONTNEED | POSIX_FADV_SEQUENTIAL);
#endif
/* seek to offset within container */
off_t pos = (off_t) container_offset;
if (lseek(fd_container, pos, SEEK_SET) == (off_t)-1) {
/* our seek failed, return an error */
scr_err("Failed to seek to byte %lu in %s @ %s:%d",
pos, dst_file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
break;
}
/* copy data from file into container in chunks */
unsigned long remaining = segment_length;
while (remaining > 0) {
/* read / write up to buf_size bytes at a time from file */
unsigned long count = remaining;
if (count > buf_size) {
count = buf_size;
}
/* attempt to read buf_size bytes from file */
int nread = scr_read_attempt(file, fd_src, buf, count);
/* if we read some bytes, write them out */
if (nread > 0) {
/* optionally compute crc value as we go */
if (scr_crc_on_flush) {
crc = crc32(crc, (const Bytef*) buf, (uInt) nread);
}
/* write our nread bytes out */
int nwrite = scr_write_attempt(dst_file, fd_container, buf, nread);
/* check for a write error or a short write */
if (nwrite != nread) {
/* write had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
/* subtract the bytes we've processed from the number remaining */
remaining -= (unsigned long) nread;
}
/* assume a short read is an error */
if (nread < count) {
/* read had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
/* check for a read error, stop copying and return an error */
if (nread < 0) {
/* read had a problem, stop copying and return an error */
rc = SCR_FAILURE;
break;
}
}
/* close container */
if (scr_close(dst_file, fd_container) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* free the container file name and path */
scr_free(&dst_file);
scr_path_delete(&dst_path);
}
/* close the source file */
if (scr_close(file, fd_src) != SCR_SUCCESS) {
rc = SCR_FAILURE;
}
/* free buffer */
scr_free(&buf);
/* verify / set crc value */
if (rc == SCR_SUCCESS) {
uLong crc2;
if (scr_crc_on_flush) {
if (scr_meta_get_crc32(meta, &crc2) == SCR_SUCCESS) {
/* if a crc is already set in the meta data, check that we computed the same value */
if (crc != crc2) {
scr_err("CRC32 mismatch detected when flushing file %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
} else {
/* if there is no crc set, let's set it now */
scr_meta_set_crc32(meta, crc);
}
}
}
return rc;
}
/* 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;
}
/* 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;
}
static int scr_fetch_rank2file_map(
const scr_path* dataset_path,
int depth,
int* ptr_valid,
char** ptr_file,
unsigned long* ptr_offset)
{
int rc = SCR_SUCCESS;
/* get local variables so we don't have to deference everything */
int valid = *ptr_valid;
char* file = *ptr_file;
unsigned long offset = *ptr_offset;
/* create a hash to hold section of file */
scr_hash* hash = scr_hash_new();
/* if we can read from file do it */
if (valid) {
/* open file if we haven't already */
int fd = scr_open(file, O_RDONLY);
if (fd >= 0) {
/* read our segment from the file */
scr_lseek(file, fd, offset, SEEK_SET);
ssize_t read_rc = scr_hash_read_fd(file, fd, hash);
if (read_rc < 0) {
scr_err("Failed to read from %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
/* close the file */
scr_close(file, fd);
} else {
scr_err("Failed to open rank2file map %s @ %s:%d",
file, __FILE__, __LINE__
);
rc = SCR_FAILURE;
}
}
/* check for read errors */
if (! scr_alltrue(rc == SCR_SUCCESS)) {
rc = SCR_FAILURE;
goto cleanup;
}
/* create hashes to exchange data */
scr_hash* send = scr_hash_new();
scr_hash* recv = scr_hash_new();
/* copy rank data into send hash */
if (valid) {
scr_hash* rank_hash = scr_hash_get(hash, SCR_SUMMARY_6_KEY_RANK);
scr_hash_merge(send, rank_hash);
}
/* exchange hashes */
scr_hash_exchange_direction(send, recv, scr_comm_world, SCR_HASH_EXCHANGE_RIGHT);
/* see if anyone sent us anything */
int newvalid = 0;
char* newfile = NULL;
unsigned long newoffset = 0;
scr_hash_elem* elem = scr_hash_elem_first(recv);
if (elem != NULL) {
/* got something, so now we'll read in the next step */
newvalid = 1;
/* get file name we should read */
scr_hash* elem_hash = scr_hash_elem_hash(elem);
char* value;
if (scr_hash_util_get_str(elem_hash, SCR_SUMMARY_6_KEY_FILE, &value)
== SCR_SUCCESS)
{
/* return string of full path to file to caller */
scr_path* newpath = scr_path_dup(dataset_path);
scr_path_append_str(newpath, value);
newfile = scr_path_strdup(newpath);
scr_path_delete(&newpath);
} else {
rc = SCR_FAILURE;
}
/* get offset we should start reading from */
if (scr_hash_util_get_bytecount(elem_hash, SCR_SUMMARY_6_KEY_OFFSET, &newoffset)
!= SCR_SUCCESS)
{
rc = SCR_FAILURE;
}
}
/* free the send and receive hashes */
scr_hash_delete(&recv);
scr_hash_delete(&send);
/* get level id, and broadcast it from rank 0,
* which we assume to be a reader in all steps */
int level_id = -1;
if (valid) {
if (scr_hash_util_get_int(hash, SCR_SUMMARY_6_KEY_LEVEL, &level_id)
!= SCR_SUCCESS)
{
rc = SCR_FAILURE;
}
}
MPI_Bcast(&level_id, 1, MPI_INT, 0, scr_comm_world);
/* check for read errors */
if (! scr_alltrue(rc == SCR_SUCCESS)) {
rc = SCR_FAILURE;
goto cleanup;
}
/* set parameters for output or next iteration,
* we already took care of updating ptr_fd earlier */
if (valid) {
scr_free(ptr_file);
}
*ptr_valid = newvalid;
*ptr_file = newfile;
*ptr_offset = newoffset;
/* recurse if we still have levels to read */
if (level_id > 1) {
rc = scr_fetch_rank2file_map(dataset_path, depth+1, ptr_valid, ptr_file, ptr_offset);
}
cleanup:
/* free the hash */
scr_hash_delete(&hash);
return rc;
}
/* fetch files listed in hash into specified cache directory,
* update filemap and fill in total number of bytes fetched,
* returns SCR_SUCCESS if successful */
static int scr_fetch_files_list(
const scr_hash* file_list,
const char* dir,
scr_filemap* map)
{
/* assume we'll succeed in fetching our files */
int rc = SCR_SUCCESS;
/* assume we don't have any files to fetch */
int my_num_files = 0;
/* get dataset id */
int id;
scr_dataset* dataset = scr_hash_get(file_list, SCR_KEY_DATASET);
scr_dataset_get_id(dataset, &id);
/* now iterate through the file list and fetch each file */
scr_hash_elem* file_elem = NULL;
scr_hash* files = scr_hash_get(file_list, SCR_KEY_FILE);
for (file_elem = scr_hash_elem_first(files);
file_elem != NULL;
file_elem = scr_hash_elem_next(file_elem))
{
/* get the filename */
char* file = scr_hash_elem_key(file_elem);
/* get a pointer to the hash for this file */
scr_hash* hash = scr_hash_elem_hash(file_elem);
/* check whether we are supposed to fetch this file */
/* TODO: this is a hacky way to avoid reading a redundancy file
* back in under the assumption that it's an original file, which
* breaks our redundancy computation due to a name conflict on
* the file names */
scr_hash_elem* no_fetch_hash = scr_hash_elem_get(hash, SCR_SUMMARY_6_KEY_NOFETCH);
if (no_fetch_hash != NULL) {
continue;
}
/* increment our file count */
my_num_files++;
/* build the destination file name */
scr_path* path_newfile = scr_path_from_str(file);
scr_path_basename(path_newfile);
scr_path_prepend_str(path_newfile, dir);
char* newfile = scr_path_strdup(path_newfile);
/* add the file to our filemap and write it to disk before creating
* the file, this way we have a record that it may exist before we
* actually start to fetch it */
scr_filemap_add_file(map, id, scr_my_rank_world, newfile);
scr_filemap_write(scr_map_file, map);
/* get the file size */
unsigned long filesize = 0;
if (scr_hash_util_get_unsigned_long(hash, SCR_KEY_SIZE, &filesize) != SCR_SUCCESS) {
scr_err("Failed to read file size from summary data @ %s:%d",
__FILE__, __LINE__
);
rc = SCR_FAILURE;
/* free path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
break;
}
/* check for a complete flag */
int complete = 1;
if (scr_hash_util_get_int(hash, SCR_KEY_COMPLETE, &complete) != SCR_SUCCESS) {
/* in summary file, the absence of a complete flag on a file
* implies the file is complete */
complete = 1;
}
/* create a new meta data object for this file */
scr_meta* meta = scr_meta_new();
/* set the meta data */
scr_meta_set_filename(meta, newfile);
scr_meta_set_filetype(meta, SCR_META_FILE_USER);
scr_meta_set_filesize(meta, filesize);
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);
/* get the crc, if set, and add it to the meta data */
uLong crc;
if (scr_hash_util_get_crc32(hash, SCR_KEY_CRC, &crc) == SCR_SUCCESS) {
scr_meta_set_crc32(meta, crc);
}
/* fetch file from containers if they are defined, otherwise fetch
* the native file */
scr_hash* segments = scr_hash_get(hash, SCR_SUMMARY_6_KEY_SEGMENT);
if (segments != NULL) {
/* get source path */
char* from_dir;
if (scr_hash_util_get_str(file_list, SCR_KEY_PATH, &from_dir) == SCR_SUCCESS) {
/* fetch file from containers */
if (scr_fetch_file_from_containers(newfile, meta, segments, from_dir) != SCR_SUCCESS) {
/* failed to fetch file, mark it as incomplete */
scr_meta_set_complete(meta, 0);
rc = SCR_FAILURE;
}
} else {
/* failed to find base dataset directory in file list */
rc = SCR_FAILURE;
}
} else {
/* fetch native file, lookup directory for this file */
char* from_dir;
if (scr_hash_util_get_str(hash, SCR_KEY_PATH, &from_dir) == SCR_SUCCESS) {
if (scr_fetch_file(newfile, from_dir, meta) != SCR_SUCCESS) {
/* failed to fetch file, mark it as incomplete */
scr_meta_set_complete(meta, 0);
rc = SCR_FAILURE;
}
} else {
/* failed to read source directory, mark file as incomplete */
scr_meta_set_complete(meta, 0);
rc = SCR_FAILURE;
}
}
/* TODODSET: want to write out filemap before we start to fetch
* each file? */
/* mark the file as complete */
scr_filemap_set_meta(map, id, scr_my_rank_world, newfile, meta);
/* free the meta data object */
scr_meta_delete(&meta);
/* free path and string */
scr_free(&newfile);
scr_path_delete(&path_newfile);
}
/* set the expected number of files for this dataset */
scr_filemap_set_expected_files(map, id, scr_my_rank_world, my_num_files);
scr_filemap_write(scr_map_file, map);
return rc;
}