/*
* resync raid
*/
int
meta_raid_resync(
mdsetname_t *sp,
mdname_t *raidnp,
daddr_t size,
md_error_t *ep
)
{
char *miscname;
md_resync_ioctl_t ri;
/* should have a set */
assert(sp != NULL);
assert(sp->setno == MD_MIN2SET(meta_getminor(raidnp->dev)));
/* make sure we have a raid */
if ((miscname = metagetmiscname(raidnp, ep)) == NULL)
return (-1);
if (strcmp(miscname, MD_RAID) != 0) {
return (mdmderror(ep, MDE_NOT_RAID, meta_getminor(raidnp->dev),
raidnp->cname));
}
/* start resync */
(void) memset(&ri, 0, sizeof (ri));
MD_SETDRIVERNAME(&ri, MD_RAID, sp->setno);
ri.ri_mnum = meta_getminor(raidnp->dev);
ri.ri_copysize = size;
if (metaioctl(MD_IOCSETSYNC, &ri, &ri.mde, raidnp->cname) != 0)
return (mdstealerror(ep, &ri.mde));
/* return success */
return (0);
}
/*
* Return the current root filesystem block device name. This is only valid
* when root is either a slice, a stripe or a mirror.
*/
mdname_t *
meta_get_current_root_dev(
mdsetname_t *sp,
md_error_t *ep
)
{
md_stripe_t *stripep;
md_mirror_t *mirrorp;
md_row_t *rp;
md_comp_t *cp;
mdname_t *rootnp;
void *curroot;
char *miscname;
int smi;
if ((curroot = meta_get_current_root(ep)) == NULL)
return (NULL);
if ((rootnp = metaname(&sp, curroot, UNKNOWN, ep)) == NULL)
return (NULL);
if (metaismeta(rootnp)) {
if ((miscname = metagetmiscname(rootnp, ep)) == NULL)
return (NULL);
if ((strcmp(miscname, MD_MIRROR) == 0) &&
((mirrorp = meta_get_mirror(sp, rootnp, ep)) != NULL)) {
for (smi = 0; smi < NMIRROR; smi++) {
md_submirror_t *mdsp =
&mirrorp->submirrors[smi];
rootnp = mdsp->submirnamep;
/* skip unused submirrors */
if (rootnp == NULL) {
assert(mdsp->state == SMS_UNUSED);
continue;
}
if ((miscname = metagetmiscname(rootnp, ep))
== NULL) {
(void) mdmderror(ep, MDE_UNKNOWN_TYPE,
meta_getminor(rootnp->dev),
rootnp->cname);
return (NULL);
}
break;
}
}
if ((strcmp(miscname, MD_STRIPE) == 0) &&
((stripep = meta_get_stripe(sp, rootnp, ep)) != NULL)) {
rp = &stripep->rows.rows_val[0];
cp = &rp->comps.comps_val[0];
if (metachkcomp(cp->compnamep, ep) == 0)
return (cp->compnamep);
}
/* Root is not a single stripe metadevice */
(void) mddeverror(ep, MDE_INV_ROOT, rootnp->dev, rootnp->cname);
return (NULL);
} else return (rootnp);
}
int
md_probe_ioctl(mdnamelist_t *nlp, int ndevs, char *drvname, boolean_e verbose)
{
mdnamelist_t *p;
mdname_t *np;
md_probedev_t probe_ioc,
*iocp;
int i,
retval = 0;
/*
* Allocate space for all the metadevices and fill in
* the minor numbers.
*/
memset(&probe_ioc, 0, sizeof (probe_ioc));
iocp = &probe_ioc;
if ((iocp->mnum_list = (uintptr_t)calloc(ndevs, sizeof (minor_t)))
== 0) {
monitord_print(0, "md_probe_ioctl: calloc");
return (-1);
}
(void) strcpy(iocp->test_name, MD_PROBE_OPEN_T);
MD_SETDRIVERNAME(&probe_ioc, drvname, sp->setno);
if (verbose == True) {
monitord_print(6, "\n\nmd_probe_ioctl: %s: %s\n",
(strcmp(sp->setname, MD_LOCAL_NAME) == 0) ?
gettext("local_set") :
sp->setname, iocp->md_driver.md_drivername);
}
iocp->nmdevs = ndevs;
if (verbose == True)
monitord_print(6, "...ndevs 0x%x\n", ndevs);
for (p = nlp, i = 0; p; p = p->next, i++) {
np = p->namep;
((minor_t *)(uintptr_t)iocp->mnum_list)[i] =
meta_getminor(np->dev);
if (verbose == True)
monitord_print(6, "...%s 0x%lx\n", np->cname,
((minor_t *)(uintptr_t)iocp->mnum_list)[i]);
}
if (issue_ioctl == True) {
if (metaioctl(MD_IOCPROBE_DEV, iocp, &(iocp->mde), NULL) != 0)
retval = -1;
}
return (retval);
}
/*
* grow generic device
*/
int
meta_concat_generic(
mdsetname_t *sp,
mdname_t *namep,
u_longlong_t big_or_little,
md_error_t *ep
)
{
md_grow_params_t mgp;
char *miscname;
/* should have a set */
assert(sp != NULL);
assert(sp->setno == MD_MIN2SET(meta_getminor(namep->dev)));
/* get type */
if ((miscname = metagetmiscname(namep, ep)) == NULL)
return (-1);
/* grow device */
(void) memset(&mgp, 0, sizeof (mgp));
if (big_or_little & MD_64BIT_META_DEV)
mgp.options = MD_CRO_64BIT;
else
mgp.options = MD_CRO_32BIT;
mgp.mnum = meta_getminor(namep->dev);
MD_SETDRIVERNAME(&mgp, miscname, sp->setno);
if (metaioctl(MD_IOCGROW, &mgp, &mgp.mde, namep->cname) != 0)
return (mdstealerror(ep, &mgp.mde));
/* clear cache */
meta_invalidate_name(namep);
/* return success */
return (0);
}
/*
* grow the parent of a device
*/
int
meta_concat_parent(
mdsetname_t *sp,
mdname_t *childnp,
md_error_t *ep
)
{
md_common_t *mdp;
mdname_t *parentnp;
md_unit_t *mup;
/* should have a set */
assert(sp != NULL);
assert(sp->setno == MD_MIN2SET(meta_getminor(childnp->dev)));
/* get parent */
if ((mdp = meta_get_unit(sp, childnp, ep)) == NULL)
return (-1);
if (! MD_HAS_PARENT(mdp->parent))
return (0);
if (mdp->parent == MD_MULTI_PARENT)
return (0);
/* single parent */
if ((parentnp = metamnumname(&sp, mdp->parent, 0, ep)) == NULL)
return (-1);
/* don't grow non-metadevices or soft partitions */
if (! metaismeta(parentnp) || meta_sp_issp(sp, parentnp, ep) == 0)
return (0);
if ((mup = meta_get_mdunit(sp, childnp, ep)) == NULL)
return (-1);
/* grow parent */
if (meta_concat_generic(sp, parentnp, mup->c.un_revision, ep) != 0)
return (-1);
/* recursively check for parents of parents */
return (meta_concat_parent(sp, parentnp, ep));
}
char *
obj2devname(uint32_t tag, set_t setno, md_dev64_t dev)
{
char *setname;
char *uname;
char name[MD_MAX_CTDLEN];
mdsetname_t *sp;
md_error_t status = mdnullerror;
md_set_record *md_sr;
minor_t mnum = meta_getminor(dev);
int rtn = 0;
setname = NULL;
if ((setno != MD_SET_BAD) &&
((sp = metasetnosetname(setno, &status)) != NULL)) {
setname = sp->setname;
}
name[0] = '\0';
switch (tag) {
case SVM_TAG_HS:
case SVM_TAG_METADEVICE:
case SVM_TAG_MIRROR:
case SVM_TAG_RAID5:
case SVM_TAG_STRIPE:
case SVM_TAG_TRANS:
uname = get_mdname(sp, mnum);
if (uname == NULL)
return (NULL);
(void) strcpy(name, uname);
break;
case SVM_TAG_HSP:
uname = get_hspname(sp, mnum);
if (uname == NULL)
return (NULL);
(void) strcpy(name, uname);
break;
case SVM_TAG_DRIVE:
(void) sprintf(name, "drive");
break;
case SVM_TAG_HOST:
md_sr = NULL;
if (setname != NULL) {
md_sr = getsetbyname(setname, &status);
}
if ((md_sr != NULL) && (md_sr->sr_nodes[mnum] != NULL)) {
/*
* Get the host data from the node array.
*/
rtn = snprintf(name, sizeof (name), "%s",
md_sr->sr_nodes[mnum]);
}
if ((name[0] == '\0') || (rtn >= sizeof (name))) {
(void) sprintf(name, "host");
rtn = 0;
}
break;
case SVM_TAG_SET:
if (setname == NULL) {
(void) sprintf(name, "diskset");
} else {
rtn = snprintf(name, sizeof (name), "%s", setname);
}
break;
default:
if ((setname = get_devname(setno, dev)) != NULL) {
rtn = snprintf(name, sizeof (name), "%s", setname);
}
break;
}
mdclrerror(&status);
/* Check if we got any rubbish for any of the snprintf's */
if ((name[0] == '\0') || (rtn >= sizeof (name))) {
return (NULL);
}
return (strdup(name));
}
/*
* FUNCTION: meta_repartition_drive()
* INPUT: sp - the set name for the device to check
* dnp - the name of the drive to partition
* options - options (see NOTES)
* OUTPUT: vtocp - pointer to an mdvtoc_t structure in which
* to return the new VTOC to the caller
* ep - pointer to an md_error_t structure in which
* to return errors to the caller
* RETURNS: int - 0 - drive was or can be repartitioned
* -1 - drive could not or should not be
* repartitioned
* PURPOSE: Repartition a disk for use in a disk set or in order
* to create soft partitions on it. Alternatively,
* return the VTOC that the disk would have if it were
* repartitioned without actually repartitioning it.
*
* NOTES:
*
* This routine will repartition a drive to make it suitable for
* inclusion in a diskset. Specifically, it will create a
* proposed VTOC that specifies a replica slice that begins at the
* first valid lba, is large enough to hold a label and a metadb
* replica, does not overlap any other slices, and is unmountable.
* If the current replica slice already satisfies those criteria,
* the routine will neither create a proposed VTOC nor repartition
* the drive unless the MD_REPART_FORCE flag is passed into the
* routine in the options argument. If the routine does create a
* proposed VTOC, it will return the proposed VTOC in *vtocp if
* vtocp isn't NULL.
*
* The slice to be used as the replica slice is determined by the
* function meta_replicaslice().
*
* If the replica slice does not satisfy the above criteria or the
* MD_REPART_FORCE flag is set, the proposed VTOC will specify a
* replica slice that satisfies the above criteria, a slice zero
* that contains the remaining space on the disk, and no other
* slices. If that repartitioning would cause the replica slice
* to move or shrink, and the MD_REPART_LEAVE_REP option is set,
* the routine will return -1 without creating or returning a
* proposed vtoc, and without repartitioning the disk. Otherwise
* the routine will repartition the disk unless the
* MD_REPART_DONT_LABEL flag is set in the options argument.
*
* If the MD_REPART_DONT_LABEL flag is set in the options argument,
* but the routine would otherwise repartition the drive, the
* routine won't repartition the drive, but will create a proposed
* VTOC that satisfies the criteria defined above and return it
* it in *vtocp if vtocp isn't NULL, The MD_REPART_DONT_LABEL
* option allows calling routines to determine what the contents of
* the drive's VTOC would be if the drive were repartitioned without
* actually repartitioning the drive.
*/
int
meta_repartition_drive(
mdsetname_t *sp,
mddrivename_t *dnp,
int options,
mdvtoc_t *vtocp,
md_error_t *ep
)
{
uint_t replicaslice;
diskaddr_t first_lba, last_lba;
int round_sizes = 1;
unsigned long long cylsize;
unsigned long long drvsize;
int i;
mdgeom_t *mdgp;
mdvtoc_t *mdvp;
mdvtoc_t proposed_vtoc;
uint_t reservedcyl;
ushort_t resflag;
mdname_t *resnp;
unsigned long long ressize;
md_set_desc *sd;
daddr_t dbsize;
diskaddr_t replica_start;
diskaddr_t replica_size;
diskaddr_t replica_end;
diskaddr_t data_start;
diskaddr_t data_size;
if (meta_replicaslice(dnp, &replicaslice, ep) != 0) {
return (-1);
}
/* Don't round for EFI disks */
if (replicaslice == MD_SLICE6)
round_sizes = 0;
/*
* We took as argument a drive name pointer, but we need a
* slice name pointer to retrieve vtoc information. So get
* the name pointer for slice zero first, then use it to get
* the vtoc info for the disk.
*/
if ((resnp = metaslicename(dnp, MD_SLICE0, ep)) == NULL)
return (-1);
if ((mdvp = metagetvtoc(resnp, FALSE, NULL, ep)) == NULL)
return (-1);
/*
* Determine the metadb size.
*/
dbsize = MD_DBSIZE;
if (!metaislocalset(sp)) {
if ((sd = metaget_setdesc(sp, ep)) == NULL)
return (-1);
if (MD_MNSET_DESC(sd))
dbsize = MD_MN_DBSIZE;
}
/* If we've got an efi disk, we better have lba info */
first_lba = mdvp->first_lba;
last_lba = mdvp->last_lba;
ASSERT((round_sizes != 0) || (last_lba > 0));
/*
* At this point, ressize is used as a minimum value. Later
* it will be rounded up to a cylinder boundary if
* appropriate. ressize is in units of disk sectors.
*/
ressize = dbsize + VTOC_SIZE;
resflag = V_UNMNT;
/*
* If we're forcing the repartition, we can skip the replica
* slice and overlap tests.
*/
if (options & MD_REPART_FORCE) {
goto do_repartition;
}
/*
* Replica slice tests: it must begin at first_lba, be long
* enough, have the right flags, and not overlap any other
* slices. If any of these conditions is violated, we need to
* repartition the disk.
*/
if (mdvp->parts[replicaslice].start != first_lba) {
goto do_repartition;
}
if (mdvp->parts[replicaslice].size < ressize) {
goto do_repartition;
}
if (mdvp->parts[replicaslice].flag != resflag) {
goto do_repartition;
}
/*
* Check for overlap: this test should use the actual size of
* the replica slice, as contained in the vtoc, and NOT the
* minimum size calculated above.
*/
replica_end = first_lba + mdvp->parts[replicaslice].size;
for (i = 0; i < mdvp->nparts; i++) {
if (i != replicaslice) {
if ((mdvp->parts[i].size > 0) &&
(mdvp->parts[i].start < replica_end)) {
goto do_repartition;
}
}
}
/*
* If we passed the above tests, then the disk is already
* partitioned appropriately, and we're not being told to
* force a change.
*/
return (0);
do_repartition:
/* Retrieve disk geometry info and round to cylinder sizes */
if (round_sizes != 0) {
if ((mdgp = metagetgeom(resnp, ep)) == NULL)
return (-1);
/*
* Both cylsize and drvsize are in units of disk
* sectors.
*
* The intended results are of type unsigned long
* long. Since each operand of the first
* multiplication is of type unsigned int, we risk
* overflow by multiplying and then converting the
* result. Therefore we explicitly cast (at least)
* one of the operands, forcing conversion BEFORE
* multiplication, and avoiding overflow. The second
* assignment is OK, since one of the operands is
* already of the desired type.
*/
cylsize =
((unsigned long long)mdgp->nhead) * mdgp->nsect;
drvsize = cylsize * mdgp->ncyl;
/*
* How many cylinders must we reserve for the replica
* slice to ensure that it meets the previously
* calculated minimum size?
*/
reservedcyl = (ressize + cylsize - 1) / cylsize;
ressize = reservedcyl * cylsize;
} else {
drvsize = last_lba - first_lba;
}
/* Would this require a forbidden change? */
if (options & MD_REPART_LEAVE_REP) {
if ((mdvp->parts[replicaslice].start != first_lba) ||
(mdvp->parts[replicaslice].size < ressize)) {
return (mddeverror(ep, MDE_REPART_REPLICA,
resnp->dev, NULL));
}
}
/*
* It seems unlikely that someone would pass us too small a
* disk, but it's still worth checking for...
*/
if (((round_sizes != 0) && (reservedcyl >= (int)mdgp->ncyl)) ||
((round_sizes == 0) && (ressize + first_lba >= last_lba))) {
return (mdmddberror(ep, MDE_DB_TOOSMALL,
meta_getminor(resnp->dev), sp->setno, 0, NULL));
}
replica_start = first_lba;
replica_size = ressize;
data_start = first_lba + ressize;
data_size = drvsize - ressize;
/*
* Create the proposed VTOC. First copy the current VTOC
* into the proposed VTOC to duplicate the values that don't
* need to change. Then change the partition table and set
* the flag value for the replica slice to resflag to reserve it
* for metadata.
*/
proposed_vtoc = *mdvp;
/* We need at least replicaslice partitions in the proposed vtoc */
if (replicaslice >= proposed_vtoc.nparts) {
proposed_vtoc.nparts = replicaslice + 1;
}
for (i = 0; i < proposed_vtoc.nparts; i++) {
/* don't change the reserved partition of an EFI device */
if (proposed_vtoc.parts[i].tag == V_RESERVED)
data_size = proposed_vtoc.parts[i].start - data_start;
else
(void) memset(&proposed_vtoc.parts[i], '\0',
sizeof (proposed_vtoc.parts[i]));
}
proposed_vtoc.parts[MD_SLICE0].start = data_start;
proposed_vtoc.parts[MD_SLICE0].size = data_size;
proposed_vtoc.parts[MD_SLICE0].tag = V_USR;
proposed_vtoc.parts[replicaslice].start = replica_start;
proposed_vtoc.parts[replicaslice].size = replica_size;
proposed_vtoc.parts[replicaslice].flag = resflag;
proposed_vtoc.parts[replicaslice].tag = V_USR;
if (!(options & MD_REPART_DONT_LABEL)) {
/*
* Label the disk with the proposed VTOC.
*/
*mdvp = proposed_vtoc;
if (metasetvtoc(resnp, ep) != 0) {
return (-1);
}
}
if (vtocp != NULL) {
/*
* Return the proposed VTOC.
*/
*vtocp = proposed_vtoc;
}
return (0);
}