{
struct obd_statfs *sfs = &tgt->ltd_statfs;
- if (((sfs->os_state & OS_STATE_ENOSPC) ||
- (!ltd->ltd_is_mdt && sfs->os_state & OS_STATE_ENOINO &&
+ if (((sfs->os_state & OS_STATFS_ENOSPC) ||
+ (!ltd->ltd_is_mdt && sfs->os_state & OS_STATFS_ENOINO &&
sfs->os_fprecreated == 0)))
return -ENOSPC;
/* If the OST is readonly then we can't allocate objects there */
- if (sfs->os_state & OS_STATE_READONLY)
+ if (sfs->os_state & OS_STATFS_READONLY)
return -EROFS;
/* object precreation is skipped on the OST with max_create_count=0 */
- if (!ltd->ltd_is_mdt && sfs->os_state & OS_STATE_NOPRECREATE)
+ if (!ltd->ltd_is_mdt && sfs->os_state & OS_STATFS_NOPRECREATE)
return -ENOBUFS;
return 0;
{
struct lov_desc *desc = <d->ltd_lov_desc;
int rc;
+ ENTRY;
LASSERT(d);
LASSERT(tgt);
}
spin_unlock(&d->lod_lock);
}
+ if (rc == -ENOTCONN) {
+ /* In case that the ENOTCONN for inactive OST state is
+ * mistreated as MDT disconnection state by the client,
+ * this error should be changed to someone else.
+ */
+ rc = -EREMOTEIO;
+ }
- return rc;
+ RETURN(rc);
}
static int lod_is_tgt_usable(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
deleting from the pool. The lq_rw_sem insures that nobody else
is reading. */
lqr->lqr_pool.op_count = real_count;
- rc = lod_tgt_pool_extend(&lqr->lqr_pool, real_count);
+ rc = tgt_pool_extend(&lqr->lqr_pool, real_count);
if (rc) {
up_write(<d->ltd_qos.lq_rw_sem);
RETURN(rc);
for (i = 0; i < lqr->lqr_pool.op_count; i++) {
int next;
- if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap,
- src_pool->op_array[i]))
+ if (!test_bit(src_pool->op_array[i],
+ ltd->ltd_tgt_bitmap))
continue;
tgt = LTD_TGT(ltd, src_pool->op_array[i]);
if (info->lti_ea_store_size < sizeof(int) * stripes)
lod_ea_store_resize(info, stripes * sizeof(int));
if (info->lti_ea_store_size < sizeof(int) * stripes) {
- CERROR("can't allocate memory for ost-in-use array\n");
+ CERROR("can't allocate memory for tgt-in-use array\n");
return -ENOMEM;
}
memset(info->lti_ea_store, -1, sizeof(int) * stripes);
bool used = false;
int i;
- if (!cfs_bitmap_check(lod->lod_ost_bitmap, index)) {
+ if (!test_bit(index, lod->lod_ost_bitmap)) {
QOS_DEBUG("OST%d: been used in conflicting mirror component\n",
index);
return true;
/*
* try to use another OSP if this one is degraded
*/
- if (ost->ltd_statfs.os_state & OS_STATE_DEGRADED && speed < 2) {
+ if (ost->ltd_statfs.os_state & OS_STATFS_DEGRADED && speed < 2) {
QOS_DEBUG("#%d: degraded\n", ost_idx);
RETURN(rc);
}
stripe_idx, array_idx, ost_idx);
if ((ost_idx == LOV_QOS_EMPTY) ||
- !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
+ !test_bit(ost_idx, m->lod_ost_bitmap))
continue;
/* Fail Check before osc_precreate() is called
RETURN(rc);
}
+static int
+lod_qos_mdt_in_use_init(const struct lu_env *env,
+ const struct lu_tgt_descs *ltd,
+ u32 stripe_idx, u32 stripe_count,
+ const struct lu_tgt_pool *pool,
+ struct dt_object **stripes)
+{
+ u32 mdt_idx;
+ struct lu_tgt_desc *mdt;
+ int i, j;
+ int rc;
+
+ rc = lod_qos_tgt_in_use_clear(env, stripe_count);
+ if (rc)
+ return rc;
+
+ /* if stripe_idx > 1, we are splitting directory, mark existing stripes
+ * in_use. Because for either split or creation, stripe 0 is local,
+ * don't mark it in use.
+ */
+ for (i = 1; i < stripe_idx; i++) {
+ LASSERT(stripes[i]);
+ for (j = 0; j < pool->op_count; j++) {
+ mdt_idx = pool->op_array[j];
+
+ if (!test_bit(mdt_idx, ltd->ltd_tgt_bitmap))
+ continue;
+
+ mdt = LTD_TGT(ltd, mdt_idx);
+ if (&mdt->ltd_tgt->dd_lu_dev ==
+ stripes[i]->do_lu.lo_dev)
+ lod_qos_tgt_in_use(env, i, mdt_idx);
+ }
+ }
+
+ return 0;
+}
+
/**
* Allocate a striping using round-robin algorithm.
*
*
* \param[in] env execution environment for this thread
* \param[in] lo LOD object
- * \param[out] stripe striping created
+ * \param[out] stripes striping created
*
* \retval positive stripe objects allocated, including the first stripe
* allocated outside
* \retval negative negated errno for other failures
*/
int lod_mdt_alloc_rr(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripe)
+ struct dt_object **stripes, u32 stripe_idx,
+ u32 stripe_count)
{
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
struct lu_tgt_descs *ltd = &lod->lod_mdt_descs;
struct dt_object *dto;
unsigned int pool_idx;
unsigned int i;
- u32 start_idx_temp;
- u32 stripe_count = lo->ldo_dir_stripe_count;
- u32 stripe_idx = 1;
+ u32 saved_idx = stripe_idx;
+ u32 start_mdt;
u32 mdt_idx;
bool use_degraded = false;
int tgt_connecting = 0;
if (rc)
RETURN(rc);
- rc = lod_qos_tgt_in_use_clear(env, stripe_count);
+ rc = lod_qos_mdt_in_use_init(env, ltd, stripe_idx, stripe_count, pool,
+ stripes);
if (rc)
RETURN(rc);
(pool->op_count % (stripe_count - 1)) != 1)
++lqr->lqr_offset_idx;
}
- start_idx_temp = lqr->lqr_start_idx;
+ start_mdt = lqr->lqr_start_idx;
repeat_find:
- QOS_DEBUG("want %d start_idx %d start_count %d offset %d active %d count %d\n",
+ QOS_DEBUG("want=%d start_idx=%d start_count=%d offset=%d active=%d count=%d\n",
stripe_count - 1, lqr->lqr_start_idx, lqr->lqr_start_count,
lqr->lqr_offset_idx, pool->op_count, pool->op_count);
stripe_idx, pool_idx, mdt_idx);
if (mdt_idx == LOV_QOS_EMPTY ||
- !cfs_bitmap_check(ltd->ltd_tgt_bitmap, mdt_idx))
+ !test_bit(mdt_idx, ltd->ltd_tgt_bitmap))
continue;
/* do not put >1 objects on one MDT */
}
/* try to use another OSP if this one is degraded */
- if (mdt->ltd_statfs.os_state & OS_STATE_DEGRADED &&
+ if (mdt->ltd_statfs.os_state & OS_STATFS_DEGRADED &&
!use_degraded) {
QOS_DEBUG("#%d: degraded\n", mdt_idx);
continue;
}
spin_unlock(&lqr->lqr_alloc);
- rc = obd_fid_alloc(env, mdt->ltd_exp, &fid, NULL);
- if (rc) {
+ rc = dt_fid_alloc(env, mdt->ltd_tgt, &fid, NULL, NULL);
+ if (rc < 0) {
QOS_DEBUG("#%d: alloc FID failed: %dl\n", mdt_idx, rc);
spin_lock(&lqr->lqr_alloc);
continue;
}
lod_qos_tgt_in_use(env, stripe_idx, mdt_idx);
- stripe[stripe_idx] = dto;
- stripe_idx++;
+ stripes[stripe_idx++] = dto;
}
if (!use_degraded && stripe_idx < stripe_count) {
- /* Try again, allowing slower OSCs */
+ /* Try again, allowing slower MDTs */
use_degraded = true;
- lqr->lqr_start_idx = start_idx_temp;
+ lqr->lqr_start_idx = start_mdt;
tgt_connecting = 0;
goto repeat_find;
spin_unlock(&lqr->lqr_alloc);
up_read(<d->ltd_qos.lq_rw_sem);
- if (stripe_idx > 1)
+ if (stripe_idx > saved_idx)
/* at least one stripe is allocated */
RETURN(stripe_idx);
i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) {
__u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx];
- if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
+ if (!test_bit(ost_idx, m->lod_ost_bitmap)) {
rc = -ENODEV;
break;
}
i++, array_idx = (array_idx + 1) % ost_count) {
ost_idx = osts->op_array[array_idx];
- if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
+ if (!test_bit(ost_idx, m->lod_ost_bitmap))
continue;
/* Fail Check before osc_precreate() is called
good_osts = 0;
/* Find all the OSTs that are valid stripe candidates */
for (i = 0; i < osts->op_count; i++) {
- if (!cfs_bitmap_check(lod->lod_ost_bitmap, osts->op_array[i]))
+ if (!test_bit(osts->op_array[i], lod->lod_ost_bitmap))
continue;
ost = OST_TGT(lod, osts->op_array[i]);
continue;
}
- if (ost->ltd_statfs.os_state & OS_STATE_DEGRADED)
+ if (ost->ltd_statfs.os_state & OS_STATFS_DEGRADED)
continue;
/* Fail Check before osc_precreate() is called
*
* \param[in] env execution environment for this thread
* \param[in] lo LOD object
+ * \param[in] stripe_idx starting stripe index to allocate, if it's not
+ * 0, we are restriping directory
+ * \param[in] stripe_count total stripe count
* \param[out] stripes striping created
*
* \retval positive stripes allocated, and it should be equal to
* \retval negative errno on failure
*/
int lod_mdt_alloc_qos(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripes)
+ struct dt_object **stripes, u32 stripe_idx,
+ u32 stripe_count)
{
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
struct lu_tgt_descs *ltd = &lod->lod_mdt_descs;
struct lu_tgt_desc *mdt;
struct dt_object *dto;
u64 total_weight = 0;
- u32 stripe_count = lo->ldo_dir_stripe_count;
- unsigned int nfound;
+ u32 saved_idx = stripe_idx;
+ u32 mdt_idx;
unsigned int good_mdts;
unsigned int i;
int rc = 0;
ENTRY;
- if (stripe_count == 1)
- RETURN(1);
+ LASSERT(stripe_idx <= stripe_count);
+ if (stripe_idx == stripe_count)
+ RETURN(stripe_count);
+ /* use MDT pool in @ltd, once MDT pool is supported in the future, it
+ * can be passed in as argument like OST object allocation.
+ */
pool = <d->ltd_tgt_pool;
/* Detect -EAGAIN early, before expensive lock is taken. */
if (!ltd_qos_is_usable(ltd))
RETURN(-EAGAIN);
+ rc = lod_qos_mdt_in_use_init(env, ltd, stripe_idx, stripe_count, pool,
+ stripes);
+ if (rc)
+ RETURN(rc);
+
/* Do actual allocation, use write lock here. */
down_write(<d->ltd_qos.lq_rw_sem);
if (rc)
GOTO(unlock, rc);
- rc = lod_qos_tgt_in_use_clear(env, stripe_count);
- if (rc)
- GOTO(unlock, rc);
-
good_mdts = 0;
- /* Find all the tgts that are valid stripe candidates */
+ /* Find all the MDTs that are valid stripe candidates */
for (i = 0; i < pool->op_count; i++) {
- if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, pool->op_array[i]))
+ if (!test_bit(pool->op_array[i], ltd->ltd_tgt_bitmap))
continue;
mdt = LTD_TGT(ltd, pool->op_array[i]);
if (rc)
continue;
- if (mdt->ltd_statfs.os_state & OS_STATE_DEGRADED)
+ if (mdt->ltd_statfs.os_state & OS_STATFS_DEGRADED)
continue;
mdt->ltd_qos.ltq_usable = 1;
good_mdts++;
}
- QOS_DEBUG("found %d good tgts\n", good_mdts);
+ QOS_DEBUG("found %d good MDTs\n", good_mdts);
- if (good_mdts < stripe_count - 1)
+ if (good_mdts < stripe_count - stripe_idx)
GOTO(unlock, rc = -EAGAIN);
- /* Find enough tgts with weighted random allocation. */
- nfound = 1;
- while (nfound < stripe_count) {
+ /* Find enough MDTs with weighted random allocation. */
+ while (stripe_idx < stripe_count) {
u64 rand, cur_weight;
cur_weight = 0;
rand = lu_prandom_u64_max(total_weight);
- /* On average, this will hit larger-weighted tgts more often.
- * 0-weight tgts will always get used last (only when rand=0) */
+ /* On average, this will hit larger-weighted MDTs more often.
+ * 0-weight MDT will always get used last (only when rand=0) */
for (i = 0; i < pool->op_count; i++) {
- __u32 idx = pool->op_array[i];
int rc2;
- mdt = LTD_TGT(ltd, idx);
+ mdt_idx = pool->op_array[i];
+ mdt = LTD_TGT(ltd, mdt_idx);
if (!mdt->ltd_qos.ltq_usable)
continue;
cur_weight += mdt->ltd_qos.ltq_weight;
- QOS_DEBUG("idx=%d nfound=%d cur_weight=%llu rand=%llu total_weight=%llu\n",
- idx, nfound, cur_weight, rand,
+ QOS_DEBUG("stripe_count=%d stripe_index=%d cur_weight=%llu rand=%llu total_weight=%llu\n",
+ stripe_count, stripe_idx, cur_weight, rand,
total_weight);
if (cur_weight < rand)
continue;
- QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
+ QOS_DEBUG("stripe=%d to idx=%d\n",
+ stripe_idx, mdt_idx);
- if (lod_qos_is_tgt_used(env, idx, nfound))
+ if (lod_qos_is_tgt_used(env, mdt_idx, stripe_idx))
continue;
- rc2 = obd_fid_alloc(env, mdt->ltd_exp, &fid, NULL);
- if (rc2) {
+ rc2 = dt_fid_alloc(env, mdt->ltd_tgt, &fid, NULL, NULL);
+ if (rc2 < 0) {
QOS_DEBUG("can't alloc FID on #%u: %d\n",
- idx, rc2);
+ mdt_idx, rc2);
continue;
}
&conf);
if (IS_ERR(dto)) {
QOS_DEBUG("can't alloc stripe on #%u: %d\n",
- idx, (int) PTR_ERR(dto));
+ mdt_idx, (int) PTR_ERR(dto));
continue;
}
- lod_qos_tgt_in_use(env, nfound, idx);
- stripes[nfound] = dto;
+ lod_qos_tgt_in_use(env, stripe_idx, mdt_idx);
+ stripes[stripe_idx] = dto;
ltd_qos_update(ltd, mdt, &total_weight);
- nfound++;
+ stripe_idx++;
rc = 0;
break;
}
break;
}
- if (unlikely(nfound != stripe_count)) {
+ if (unlikely(stripe_idx != stripe_count)) {
/*
* when the decision to use weighted algorithm was made
* we had enough appropriate OSPs, but this state can
* an object due to just changed state
*/
QOS_DEBUG("%s: wanted %d objects, found only %d\n",
- lod2obd(lod)->obd_name, stripe_count, nfound);
- for (i = 1; i < nfound; i++) {
+ lod2obd(lod)->obd_name, stripe_count, stripe_idx);
+ for (i = saved_idx; i < stripe_idx; i++) {
LASSERT(stripes[i] != NULL);
dt_object_put(env, stripes[i]);
stripes[i] = NULL;
rc = -EAGAIN;
} else {
- rc = nfound;
+ rc = stripe_idx;
}
unlock:
else
lod_free_comp_entries(lo);
- rc = lod_verify_striping(d, lo, buf, false);
+ rc = lod_verify_striping(env, d, lo, buf, false);
if (rc)
RETURN(-EINVAL);
}
lod_comp->llc_pattern = v1->lmm_pattern;
- lod_comp->llc_stripe_size = desc->ld_default_stripe_size;
- if (v1->lmm_stripe_size)
- lod_comp->llc_stripe_size = v1->lmm_stripe_size;
+ lod_comp->llc_stripe_size = v1->lmm_stripe_size;
+ lod_adjust_stripe_size(lod_comp, desc->ld_default_stripe_size);
lod_comp->llc_stripe_count = desc->ld_default_stripe_count;
if (v1->lmm_stripe_count ||
lag->lag_oaa_count = 0;
if (lag->lag_oss_avoid_array &&
lag->lag_oaa_size < lod->lod_ost_count) {
- OBD_FREE(lag->lag_oss_avoid_array,
- sizeof(__u32) * lag->lag_oaa_size);
+ OBD_FREE_PTR_ARRAY(lag->lag_oss_avoid_array, lag->lag_oaa_size);
lag->lag_oss_avoid_array = NULL;
lag->lag_oaa_size = 0;
}
* using OST count to allocate the array to store the OSS
* id.
*/
- OBD_ALLOC(new_oss, sizeof(*new_oss) * lod->lod_ost_count);
+ OBD_ALLOC_PTR_ARRAY(new_oss, lod->lod_ost_count);
if (!new_oss) {
CFS_FREE_BITMAP(bitmap);
return -ENOMEM;
if (stripe_len == 0)
GOTO(out, rc = -ERANGE);
lod_comp->llc_stripe_count = stripe_len;
- OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
+ OBD_ALLOC_PTR_ARRAY(stripe, stripe_len);
if (stripe == NULL)
GOTO(out, rc = -ENOMEM);
- OBD_ALLOC(ost_indices, sizeof(*ost_indices) * stripe_len);
+ OBD_ALLOC_PTR_ARRAY(ost_indices, stripe_len);
if (!ost_indices)
GOTO(out, rc = -ENOMEM);
out:
if (rc < 0) {
if (stripe)
- OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
+ OBD_FREE_PTR_ARRAY(stripe, stripe_len);
if (ost_indices)
- OBD_FREE(ost_indices,
- sizeof(*ost_indices) * stripe_len);
+ OBD_FREE_PTR_ARRAY(ost_indices, stripe_len);
}
RETURN(rc);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff