/*
* util_pool_open -- open a memory pool (set or a single file)
*
* This routine does all the work, but takes a rdonly flag so internal
* calls can map a read-only pool if required.
*/
int
util_pool_open(struct pool_set **setp, const char *path, int rdonly,
size_t minsize, size_t hdrsize, const char *sig,
uint32_t major, uint32_t compat, uint32_t incompat, uint32_t ro_compat)
{
LOG(3, "setp %p path %s rdonly %d minsize %zu "
"hdrsize %zu sig %s major %u "
"compat %#x incompat %#x ro_comapt %#x",
setp, path, rdonly, minsize, hdrsize,
sig, major, compat, incompat, ro_compat);
int flags = rdonly ? MAP_PRIVATE|MAP_NORESERVE : MAP_SHARED;
int ret = util_poolset_open(setp, path, minsize);
if (ret < 0) {
LOG(2, "cannot open pool set");
return -1;
}
struct pool_set *set = *setp;
ASSERT(set->nreplicas > 0);
set->rdonly = 0;
set->poolsize = SIZE_MAX;
for (unsigned r = 0; r < set->nreplicas; r++) {
if (util_replica_open(set, r, flags, hdrsize) != 0) {
LOG(2, "replica open failed");
goto err;
}
}
/* check headers, check UUID's, check replicas linkage */
for (unsigned r = 0; r < set->nreplicas; r++) {
struct pool_replica *rep = set->replica[r];
for (unsigned p = 0; p < rep->nparts; p++) {
if (util_header_check(set, r, p, sig, major,
compat, incompat, ro_compat) != 0) {
LOG(2, "header check failed - part #%d", p);
goto err;
}
set->rdonly |= rep->part[p].rdonly;
}
if (memcmp(HDR(REP(set, r - 1), 0)->uuid,
HDR(REP(set, r), 0)->prev_repl_uuid,
POOL_HDR_UUID_LEN) ||
memcmp(HDR(REP(set, r + 1), 0)->uuid,
HDR(REP(set, r), 0)->next_repl_uuid,
POOL_HDR_UUID_LEN)) {
ERR("wrong replica UUID");
errno = EINVAL;
goto err;
}
}
/* unmap all headers */
for (unsigned r = 0; r < set->nreplicas; r++) {
struct pool_replica *rep = set->replica[r];
for (unsigned p = 0; p < rep->nparts; p++)
util_unmap_hdr(&rep->part[p]);
}
return 0;
err:
LOG(4, "error clean up");
int oerrno = errno;
for (unsigned r = 0; r < set->nreplicas; r++)
util_replica_close(set, r);
util_poolset_close(set, 0);
errno = oerrno;
return -1;
}
/*
* util_replica_open -- (internal) open a memory pool replica
*/
static int
util_replica_open(struct pool_set *set, unsigned repidx, int flags,
size_t hdrsize, const char *sig,
uint32_t major, uint32_t compat, uint32_t incompat, uint32_t ro_compat)
{
LOG(3, "set %p repidx %u flags %d hdrsize %zu sig %s major %u "
"compat %#x incompat %#x ro_comapt %#x",
set, repidx, flags, hdrsize, sig, major,
compat, incompat, ro_compat);
struct pool_replica *rep = set->replica[repidx];
rep->repsize -= (rep->nparts - 1) * hdrsize;
/* determine a hint address for mmap() */
void *addr = util_map_hint(rep->repsize); /* XXX - randomize */
if (addr == NULL) {
ERR("cannot find a contiguous region of given size");
return -1;
}
/* map the first part and reserve space for remaining parts */
if (util_map_part(&rep->part[0], addr, rep->repsize, 0, flags) != 0) {
LOG(2, "pool mapping failed - part #0");
return -1;
}
VALGRIND_REGISTER_PMEM_MAPPING(rep->part[0].addr, rep->part[0].size);
VALGRIND_REGISTER_PMEM_FILE(rep->part[0].fd,
rep->part[0].addr, rep->part[0].size, 0);
/* map all the remaining headers - don't care about the address */
for (unsigned p = 1; p < rep->nparts; p++) {
if (util_map_part(&rep->part[p], NULL,
hdrsize, 0, flags) != 0) {
LOG(2, "header mapping failed - part #%d", p);
goto err;
}
VALGRIND_REGISTER_PMEM_FILE(rep->part[p].fd,
rep->part[p].addr, rep->part[p].size, 0);
}
/* check headers, check UUID's */
for (unsigned p = 0; p < rep->nparts; p++) {
if (util_header_check(set, repidx, p, sig, major,
compat, incompat, ro_compat) != 0) {
LOG(2, "header check failed - part #%d", p);
goto err;
}
}
set->rdonly |= rep->part[0].rdonly;
size_t mapsize = rep->part[0].filesize & ~(Pagesize - 1);
addr = rep->part[0].addr + mapsize;
/*
* unmap headers; map the remaining parts of the usable pool space
* (4K-aligned)
*/
for (unsigned p = 1; p < rep->nparts; p++) {
/* unmap header */
if (util_unmap_part(&rep->part[p]) != 0) {
LOG(2, "header unmapping failed - part #%d", p);
}
/* map data part */
if (util_map_part(&rep->part[p], addr, 0, hdrsize,
flags | MAP_FIXED) != 0) {
LOG(2, "usable space mapping failed - part #%d", p);
goto err;
}
VALGRIND_REGISTER_PMEM_FILE(rep->part[p].fd,
rep->part[p].addr, rep->part[p].size, hdrsize);
mapsize += rep->part[p].size;
set->rdonly |= rep->part[p].rdonly; /* XXX - not needed? */
addr += rep->part[p].size;
}
rep->is_pmem = pmem_is_pmem(rep->part[0].addr, rep->part[0].size);
ASSERTeq(mapsize, rep->repsize);
/* calculate pool size - choose the smallest replica size */
if (rep->repsize < set->poolsize)
set->poolsize = rep->repsize;
LOG(3, "replica addr %p", rep->part[0].addr);
return 0;
err:
LOG(4, "error clean up");
int oerrno = errno;
VALGRIND_REMOVE_PMEM_MAPPING(rep->part[0].addr, rep->part[0].size);
util_unmap(rep->part[0].addr, rep->part[0].size);
errno = oerrno;
return -1;
}