int
pfl_odt_create(const char *fn, int64_t nitems, size_t itemsz,
int overwrite, size_t startoff, size_t pad, int tflg)
{
int rc;
int64_t item;
struct pfl_odt_slotftr f;
struct pfl_odt_hdr *h;
struct pfl_odt *t;
t = PSCALLOC(sizeof(*t));
t->odt_ops = pfl_odtops;
INIT_SPINLOCK(&t->odt_lock);
snprintf(t->odt_name, sizeof(t->odt_name), "%s", pfl_basename(fn));
t->odt_iostats.rd = pfl_opstat_init("odt-%s-rd", t->odt_name);
t->odt_iostats.wr = pfl_opstat_init("odt-%s-wr", t->odt_name);
h = PSCALLOC(sizeof(*h));
memset(h, 0, sizeof(*h));
h->odth_nitems = nitems;
h->odth_itemsz = itemsz;
h->odth_slotsz = itemsz + pad + sizeof(f);
h->odth_options = tflg;
h->odth_start = startoff;
t->odt_hdr = h;
psc_crc64_calc(&h->odth_crc, h, sizeof(*h) - sizeof(h->odth_crc));
/* pfl_odt_new() and slm_odt_new() */
rc = t->odt_ops.odtop_new(t, fn, overwrite);
if (rc)
return (rc);
for (item = 0; item < nitems; item++)
_pfl_odt_doput(t, item, NULL, &f, 0);
PFLOG_ODT(PLL_DIAG, t, "created");
pfl_odt_release(t);
return (0);
}
/*
* Traverse a file hierarchy and perform an operation on each file
* system entry.
* @fn: file root.
* @flags: behavorial flags.
* @cmpf: optional dirent comparator for ordering.
* @cbf: callback to invoke on each file.
* @arg: optional argument to supply to callback.
* Notes: the callback will be invoked with a fully resolved absolute
* path name unless the file in question is a symbolic link.
*/
int
pfl_filewalk(const char *fn, int flags, void *cmpf, int (*cbf)(FTSENT *,
void *), void *arg)
{
char * const pathv[] = { (char *)fn, NULL };
int rc = 0, f_flags = 0;
struct stat stb;
FTSENT *f;
FTS *fp;
if (flags & PFL_FILEWALKF_RECURSIVE) {
if (flags & PFL_FILEWALKF_NOSTAT)
f_flags |= FTS_NOSTAT;
if (flags & PFL_FILEWALKF_NOCHDIR)
f_flags |= FTS_NOCHDIR;
fp = pfl_fts_open(pathv, f_flags | FTS_COMFOLLOW |
FTS_PHYSICAL, cmpf);
if (fp == NULL)
psc_fatal("fts_open %s", fn);
while ((f = pfl_fts_read(fp)) != NULL) {
switch (f->fts_info) {
case FTS_NS:
psclog_warnx("%s: %s", f->fts_path,
strerror(f->fts_errno));
break;
case FTS_F:
case FTS_D:
case FTS_SL:
if (flags & PFL_FILEWALKF_VERBOSE)
warnx("processing %s%s",
fn, f->fts_info == FTS_D ?
"/" : "");
case FTS_DP:
rc = cbf(f, arg);
if (rc) {
pfl_fts_close(fp);
return (rc);
}
break;
default:
if (f->fts_errno == 0)
f->fts_errno = EOPNOTSUPP;
psclog_warnx("%s: %s", f->fts_path,
strerror(f->fts_errno));
break;
}
}
pfl_fts_close(fp);
} else {
const char *basefn;
size_t baselen;
if (lstat(fn, &stb) == -1)
err(1, "%s", fn);
basefn = pfl_basename(fn);
baselen = strlen(basefn);
f = PSCALLOC(sizeof(*f) + baselen);
f->fts_accpath = (char *)fn;
f->fts_path = (char *)fn;
f->fts_pathlen = strlen(fn);
strlcpy(f->fts_name, basefn, baselen + 1);
f->fts_namelen = baselen;
f->fts_ino = stb.st_ino;
f->fts_statp = &stb;
switch (stb.st_mode & S_IFMT) {
case S_IFDIR: f->fts_info = FTS_D; break;
case S_IFREG: f->fts_info = FTS_F; break;
case S_IFLNK: f->fts_info = FTS_SL; break;
case S_IFBLK: f->fts_info = FTS_DEFAULT; break;
default:
psclog_warnx("%s: %s", fn,
strerror(EOPNOTSUPP));
break;
}
rc = cbf(f, arg);
PSCFREE(f);
}
return (rc);
}
/**
* pjournal_dump - Dump the contents of a journal file.
* @fn: journal filename to query.
* @verbose: whether to report stats summary or full dump.
*
* Each time mds restarts, it writes log entries starting from the very
* first slot of the log. Anyway, the function dumps all log entries,
* some of them may be from previous incarnations of the MDS.
*/
void
pjournal_dump(const char *fn)
{
int i, ntotal, nmagic, nchksum, nformat, ndump, first = 1;
uint32_t slot, highest_slot = -1, lowest_slot = -1;
uint64_t chksum, highest_xid = 0, lowest_xid = 0;
struct psc_journal_enthdr *pje;
struct psc_journal_hdr *pjh;
struct psc_journal *pj;
struct stat statbuf;
unsigned char *jbuf;
ssize_t nb, pjhlen;
time_t ts;
ntotal = nmagic = nchksum = nformat = ndump = 0;
pj = PSCALLOC(sizeof(*pj));
strlcpy(pj->pj_name, pfl_basename(fn), sizeof(pj->pj_name));
pj->pj_fd = open(fn, O_RDWR | O_DIRECT);
if (pj->pj_fd == -1)
psc_fatal("failed to open journal %s", fn);
if (fstat(pj->pj_fd, &statbuf) == -1)
psc_fatal("failed to stat journal %s", fn);
/*
* O_DIRECT may impose alignment restrictions so align the
* buffer and perform I/O in multiples of file system block
* size.
*/
pjhlen = PSC_ALIGN(sizeof(*pjh), statbuf.st_blksize);
pjh = psc_alloc(pjhlen, PAF_PAGEALIGN);
nb = pread(pj->pj_fd, pjh, pjhlen, 0);
if (nb != pjhlen)
psc_fatal("failed to read journal header");
pj->pj_hdr = pjh;
if (pjh->pjh_magic != PJH_MAGIC)
psc_fatalx("journal header has a bad magic number "
"%#"PRIx64, pjh->pjh_magic);
if (pjh->pjh_version != PJH_VERSION)
psc_fatalx("journal header has an invalid version "
"number %d", pjh->pjh_version);
psc_crc64_init(&chksum);
psc_crc64_add(&chksum, pjh, offsetof(struct psc_journal_hdr,
pjh_chksum));
psc_crc64_fini(&chksum);
if (pjh->pjh_chksum != chksum)
psc_fatalx("journal header has an invalid checksum "
"value %"PSCPRIxCRC64" vs %"PSCPRIxCRC64,
pjh->pjh_chksum, chksum);
if (S_ISREG(statbuf.st_mode) && statbuf.st_size !=
(off_t)(pjhlen + pjh->pjh_nents * PJ_PJESZ(pj)))
psc_fatalx("size of the journal log %"PSCPRIdOFFT"d does "
"not match specs in its header", statbuf.st_size);
if (pjh->pjh_nents % pjh->pjh_readsize)
psc_fatalx("number of entries %d is not a multiple of the "
"readsize %d", pjh->pjh_nents, pjh->pjh_readsize);
ts = pjh->pjh_timestamp;
printf("%s:\n"
" version: %u\n"
" entry size: %u\n"
" number of entries: %u\n"
" batch read size: %u\n"
" entry start offset: %"PRId64"\n"
" format time: %s"
" uuid: %"PRIx64"\n"
" %4s %3s %4s %4s %s\n",
fn, pjh->pjh_version, PJ_PJESZ(pj), pjh->pjh_nents,
pjh->pjh_readsize, pjh->pjh_start_off,
ctime(&ts), pjh->pjh_fsuuid,
"idx", "typ", "xid", "txg", "details");
jbuf = psc_alloc(PJ_PJESZ(pj) * pj->pj_hdr->pjh_readsize,
PAF_PAGEALIGN);
for (slot = 0; slot < pjh->pjh_nents;
slot += pjh->pjh_readsize) {
nb = pread(pj->pj_fd, jbuf, PJ_PJESZ(pj) *
pjh->pjh_readsize, PJ_GETENTOFF(pj, slot));
if (nb != PJ_PJESZ(pj) * pjh->pjh_readsize)
warn("failed to read %d log entries at slot %d",
pjh->pjh_readsize, slot);
for (i = 0; i < pjh->pjh_readsize; i++) {
ntotal++;
pje = (void *)&jbuf[PJ_PJESZ(pj) * i];
if (pje->pje_magic != PJE_MAGIC) {
nmagic++;
warnx("journal slot %d has a bad magic"
"number", slot + i);
continue;
}
/*
* If we hit a new entry that is never used, we
* assume that the rest of the journal is never
* used.
*/
if (pje->pje_type == PJE_FORMAT) {
nformat = nformat + pjh->pjh_nents -
(slot + i);
goto done;
}
psc_crc64_init(&chksum);
psc_crc64_add(&chksum, pje, offsetof(
struct psc_journal_enthdr, pje_chksum));
psc_crc64_add(&chksum, pje->pje_data,
pje->pje_len);
psc_crc64_fini(&chksum);
if (pje->pje_chksum != chksum) {
nchksum++;
warnx("journal slot %d has a corrupt "
"checksum", slot + i);
goto done;
}
ndump++;
if (verbose)
pjournal_dump_entry(slot + i, pje);
if (first) {
first = 0;
highest_xid = lowest_xid = pje->pje_xid;
highest_slot = lowest_slot = slot + i;
continue;
}
if (highest_xid < pje->pje_xid) {
highest_xid = pje->pje_xid;
highest_slot = slot + i;
}
if (lowest_xid > pje->pje_xid) {
lowest_xid = pje->pje_xid;
lowest_slot = slot + i;
}
}
}
done:
if (close(pj->pj_fd) == -1)
printf("failed closing journal %s", fn);
psc_free(jbuf, PAF_PAGEALIGN, PJ_PJESZ(pj));
PSCFREE(pj);
printf("----------------------------------------------\n"
"%8d slot(s) scanned\n"
"%8d in use\n"
"%8d formatted\n"
"%8d bad magic\n"
"%8d bad checksum(s)\n"
"lowest transaction ID=%#"PRIx64" (slot=%d)\n"
"highest transaction ID=%#"PRIx64" (slot=%d)\n",
ntotal, ndump, nformat, nmagic, nchksum,
lowest_xid, lowest_slot,
highest_xid, highest_slot);
}