#define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
OST_TGT(lod,i)->ltd_statfs.os_bsize)
+static inline int lod_statfs_check(struct lu_tgt_descs *ltd,
+ struct lu_tgt_desc *tgt)
+{
+ struct obd_statfs *sfs = &tgt->ltd_statfs;
+
+ 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_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_STATFS_NOPRECREATE)
+ return -ENOBUFS;
+
+ return 0;
+}
+
/**
- * Check whether the target is available for new OST objects.
+ * Check whether the target is available for new objects.
*
* Request statfs data from the given target and verify it's active and not
- * read-only. If so, then it can be used to place new OST objects. This
+ * read-only. If so, then it can be used to place new objects. This
* function also maintains the number of active/inactive targets and sets
* dirty flags if those numbers change so others can run re-balance procedures.
* No external locking is required.
*
* \param[in] env execution environment for this thread
* \param[in] d LOD device
- * \param[in] index index of OST target to check
- * \param[out] sfs buffer for statfs data
+ * \param[in] ltd target table
+ * \param[in] tgt target
*
* \retval 0 if the target is good
* \retval negative negated errno on error
-
*/
static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
- int index, struct obd_statfs *sfs)
+ struct lu_tgt_descs *ltd,
+ struct lu_tgt_desc *tgt, __u64 reserve)
{
- struct lod_tgt_desc *ost;
- int rc;
+ struct obd_statfs_info info = { 0 };
+ struct lov_desc *desc = <d->ltd_lov_desc;
+ int rc;
ENTRY;
LASSERT(d);
- ost = OST_TGT(d,index);
- LASSERT(ost);
-
- rc = dt_statfs(env, ost->ltd_ost, sfs);
-
- if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
- (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
- RETURN(-ENOSPC);
+ LASSERT(tgt);
+ info.os_enable_pre = 1;
+ rc = dt_statfs_info(env, tgt->ltd_tgt, &tgt->ltd_statfs, &info);
if (rc && rc != -ENOTCONN)
CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
- /* If the OST is readonly then we can't allocate objects there */
- if (sfs->os_state & OS_STATE_READONLY)
- rc = -EROFS;
+ if (!rc) {
+ rc = lod_statfs_check(ltd, tgt);
+ if (rc == -ENOSPC)
+ return rc;
+ }
- /* object precreation is skipped on the OST with max_create_count=0 */
- if (sfs->os_state & OS_STATE_NOPRECREATE)
- rc = -ENOBUFS;
+ if (reserve &&
+ (reserve + (info.os_reserved_mb_low << 20) >
+ tgt->ltd_statfs.os_bavail * tgt->ltd_statfs.os_bsize))
+ return -ENOSPC;
/* check whether device has changed state (active, inactive) */
- if (rc != 0 && ost->ltd_active) {
+ if (rc != 0 && tgt->ltd_active) {
/* turned inactive? */
spin_lock(&d->lod_lock);
- if (ost->ltd_active) {
- ost->ltd_active = 0;
+ if (tgt->ltd_active) {
+ tgt->ltd_active = 0;
if (rc == -ENOTCONN)
- ost->ltd_connecting = 1;
+ tgt->ltd_connecting = 1;
- LASSERT(d->lod_desc.ld_active_tgt_count > 0);
- d->lod_desc.ld_active_tgt_count--;
- d->lod_qos.lq_dirty = 1;
- d->lod_qos.lq_rr.lqr_dirty = 1;
+ LASSERT(desc->ld_active_tgt_count > 0);
+ desc->ld_active_tgt_count--;
+ ltd->ltd_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_rr.lqr_dirty = 1;
CDEBUG(D_CONFIG, "%s: turns inactive\n",
- ost->ltd_exp->exp_obd->obd_name);
+ tgt->ltd_exp->exp_obd->obd_name);
}
spin_unlock(&d->lod_lock);
- } else if (rc == 0 && ost->ltd_active == 0) {
+ } else if (rc == 0 && tgt->ltd_active == 0) {
/* turned active? */
- LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
- "active tgt count %d, ost nr %d\n",
- d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
+ LASSERTF(desc->ld_active_tgt_count < desc->ld_tgt_count,
+ "active tgt count %d, tgt nr %d\n",
+ desc->ld_active_tgt_count, desc->ld_tgt_count);
spin_lock(&d->lod_lock);
- if (ost->ltd_active == 0) {
- ost->ltd_active = 1;
- ost->ltd_connecting = 0;
- d->lod_desc.ld_active_tgt_count++;
- d->lod_qos.lq_dirty = 1;
- d->lod_qos.lq_rr.lqr_dirty = 1;
+ if (tgt->ltd_active == 0) {
+ tgt->ltd_active = 1;
+ tgt->ltd_connecting = 0;
+ desc->ld_active_tgt_count++;
+ ltd->ltd_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_rr.lqr_dirty = 1;
CDEBUG(D_CONFIG, "%s: turns active\n",
- ost->ltd_exp->exp_obd->obd_name);
+ tgt->ltd_exp->exp_obd->obd_name);
}
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);
}
+static int lod_is_tgt_usable(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
+{
+ int rc;
+
+ rc = lod_statfs_check(ltd, tgt);
+ if (rc)
+ return rc;
+
+ if (!tgt->ltd_active)
+ return -ENOTCONN;
+
+ return 0;
+}
+
/**
* Maintain per-target statfs data.
*
*
* \param[in] env execution environment for this thread
* \param[in] lod LOD device
+ * \param[in] ltd tgt table
*/
-void lod_qos_statfs_update(const struct lu_env *env, struct lod_device *lod)
+void lod_qos_statfs_update(const struct lu_env *env, struct lod_device *lod,
+ struct lu_tgt_descs *ltd)
{
struct obd_device *obd = lod2obd(lod);
- struct ost_pool *osts = &(lod->lod_pool_info);
+ struct lu_tgt_desc *tgt;
time64_t max_age;
- unsigned int i;
u64 avail;
- int idx;
ENTRY;
- max_age = ktime_get_seconds() - 2 * lod->lod_desc.ld_qos_maxage;
+ max_age = ktime_get_seconds() - 2 * ltd->ltd_lov_desc.ld_qos_maxage;
if (obd->obd_osfs_age > max_age)
/* statfs data are quite recent, don't need to refresh it */
RETURN_EXIT;
- down_write(&lod->lod_qos.lq_rw_sem);
+ down_write(<d->ltd_qos.lq_rw_sem);
if (obd->obd_osfs_age > max_age)
goto out;
- for (i = 0; i < osts->op_count; i++) {
- idx = osts->op_array[i];
- avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
- if (lod_statfs_and_check(env, lod, idx,
- &OST_TGT(lod, idx)->ltd_statfs))
+ ltd_foreach_tgt(ltd, tgt) {
+ avail = tgt->ltd_statfs.os_bavail;
+ if (lod_statfs_and_check(env, lod, ltd, tgt, 0))
continue;
- if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
+
+ if (tgt->ltd_statfs.os_bavail != avail)
/* recalculate weigths */
- lod->lod_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_dirty = 1;
}
obd->obd_osfs_age = ktime_get_seconds();
out:
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_qos.lq_rw_sem);
EXIT;
}
* a new target or activation/deactivation).
*
* \param[in] lod LOD device
- * \param[in] src_pool OST pool
+ * \param[in] ltd tgt table
+ * \param[in] src_pool tgt pool
* \param[in] lqr round-robin list
*
* \retval 0 on success
* \retval -ENOMEM fails to allocate the array
*/
-static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
+static int lod_qos_calc_rr(struct lod_device *lod, struct lu_tgt_descs *ltd,
+ const struct lu_tgt_pool *src_pool,
struct lu_qos_rr *lqr)
{
- struct lu_svr_qos *oss;
- struct lod_tgt_desc *ost;
+ struct lu_svr_qos *svr;
+ struct lu_tgt_desc *tgt;
unsigned placed, real_count;
unsigned int i;
int rc;
}
/* Do actual allocation. */
- down_write(&lod->lod_qos.lq_rw_sem);
+ down_write(<d->ltd_qos.lq_rw_sem);
/*
* Check again. While we were sleeping on @lq_rw_sem something could
*/
if (!lqr->lqr_dirty) {
LASSERT(lqr->lqr_pool.op_size);
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_qos.lq_rw_sem);
RETURN(0);
}
deleting from the pool. The lq_rw_sem insures that nobody else
is reading. */
lqr->lqr_pool.op_count = real_count;
- rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
+ rc = tgt_pool_extend(&lqr->lqr_pool, real_count);
if (rc) {
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_qos.lq_rw_sem);
RETURN(rc);
}
for (i = 0; i < lqr->lqr_pool.op_count; i++)
lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
- /* Place all the OSTs from 1 OSS at the same time. */
+ /* Place all the tgts from 1 svr at the same time. */
placed = 0;
- list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) {
+ list_for_each_entry(svr, <d->ltd_qos.lq_svr_list, lsq_svr_list) {
int j = 0;
for (i = 0; i < lqr->lqr_pool.op_count; i++) {
int next;
- if (!cfs_bitmap_check(lod->lod_ost_bitmap,
- src_pool->op_array[i]))
+ if (!test_bit(src_pool->op_array[i],
+ ltd->ltd_tgt_bitmap))
continue;
- ost = OST_TGT(lod,src_pool->op_array[i]);
- LASSERT(ost && ost->ltd_ost);
- if (ost->ltd_qos.ltq_svr != oss)
+ tgt = LTD_TGT(ltd, src_pool->op_array[i]);
+ LASSERT(tgt && tgt->ltd_tgt);
+ if (tgt->ltd_qos.ltq_svr != svr)
continue;
- /* Evenly space these OSTs across arrayspace */
- next = j * lqr->lqr_pool.op_count / oss->lsq_tgt_count;
+ /* Evenly space these tgts across arrayspace */
+ next = j * lqr->lqr_pool.op_count / svr->lsq_tgt_count;
while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
next = (next + 1) % lqr->lqr_pool.op_count;
}
lqr->lqr_dirty = 0;
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_qos.lq_rw_sem);
if (placed != real_count) {
/* This should never happen */
- LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
+ LCONSOLE_ERROR_MSG(0x14e, "Failed to place all tgts in the "
"round-robin list (%d of %d).\n",
placed, real_count);
for (i = 0; i < lqr->lqr_pool.op_count; i++) {
- LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
+ LCONSOLE(D_WARNING, "rr #%d tgt idx=%d\n", i,
lqr->lqr_pool.op_array[i]);
}
lqr->lqr_dirty = 1;
ENTRY;
LASSERT(d);
- LASSERT(ost_idx < d->lod_osts_size);
+ LASSERT(ost_idx < d->lod_ost_descs.ltd_tgts_size);
ost = OST_TGT(d,ost_idx);
LASSERT(ost);
- LASSERT(ost->ltd_ost);
+ LASSERT(ost->ltd_tgt);
- nd = &ost->ltd_ost->dd_lu_dev;
+ nd = &ost->ltd_tgt->dd_lu_dev;
/*
* allocate anonymous object with zero fid, real fid
*/
o = lu_object_anon(env, nd, NULL);
if (IS_ERR(o))
- GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
+ GOTO(out, dt = ERR_CAST(o));
n = lu_object_locate(o->lo_header, nd->ld_type);
if (unlikely(n == NULL)) {
#define LOV_CREATE_RESEED_MIN 2000
/**
- * Initialize temporary OST-in-use array.
+ * Initialize temporary tgt-in-use array.
*
* Allocate or extend the array used to mark targets already assigned to a new
* striping so they are not used more than once.
* \retval 0 on success
* \retval -ENOMEM on error
*/
-static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
+static inline int lod_qos_tgt_in_use_clear(const struct lu_env *env,
__u32 stripes)
{
struct lod_thread_info *info = lod_env_info(env);
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);
* Remember a target in the array of used targets.
*
* Mark the given target as used for a new striping being created. The status
- * of an OST in a striping can be checked with lod_qos_is_ost_used().
+ * of an tgt in a striping can be checked with lod_qos_is_tgt_used().
*
* \param[in] env execution environment for this thread
* \param[in] idx index in the array
- * \param[in] ost OST target index to mark as used
+ * \param[in] tgt_idx target index to mark as used
*/
-static inline void lod_qos_ost_in_use(const struct lu_env *env,
- int idx, int ost)
+static inline void lod_qos_tgt_in_use(const struct lu_env *env,
+ int idx, int tgt_idx)
{
struct lod_thread_info *info = lod_env_info(env);
- int *osts = info->lti_ea_store;
+ int *tgts = info->lti_ea_store;
LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
- osts[idx] = ost;
+ tgts[idx] = tgt_idx;
}
/**
- * Check is OST used in a striping.
+ * Check is tgt used in a striping.
*
- * Checks whether OST with the given index is marked as used in the temporary
- * array (see lod_qos_ost_in_use()).
+ * Checks whether tgt with the given index is marked as used in the temporary
+ * array (see lod_qos_tgt_in_use()).
*
* \param[in] env execution environment for this thread
- * \param[in] ost OST target index to check
+ * \param[in] tgt_idx target index to check
* \param[in] stripes the number of items used in the array already
*
* \retval 0 not used
* \retval 1 used
*/
-static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
+static int lod_qos_is_tgt_used(const struct lu_env *env, int tgt_idx,
+ __u32 stripes)
{
struct lod_thread_info *info = lod_env_info(env);
- int *osts = info->lti_ea_store;
+ int *tgts = info->lti_ea_store;
__u32 j;
for (j = 0; j < stripes; j++) {
- if (osts[j] == ost)
+ if (tgts[j] == tgt_idx)
return 1;
}
return 0;
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;
static int lod_check_and_reserve_ost(const struct lu_env *env,
struct lod_object *lo,
struct lod_layout_component *lod_comp,
- struct obd_statfs *sfs, __u32 ost_idx,
- __u32 speed, __u32 *s_idx,
+ __u32 ost_idx, __u32 speed, __u32 *s_idx,
struct dt_object **stripe,
__u32 *ost_indices,
struct thandle *th,
- bool *overstriped)
+ bool *overstriped,
+ __u64 reserve)
{
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
+ struct lu_tgt_desc *ost = OST_TGT(lod, ost_idx);
struct dt_object *o;
__u32 stripe_idx = *s_idx;
int rc;
+
ENTRY;
- rc = lod_statfs_and_check(env, lod, ost_idx, sfs);
+ rc = lod_statfs_and_check(env, lod, &lod->lod_ost_descs, ost, reserve);
if (rc)
RETURN(rc);
* We expect number of precreated objects in f_ffree at
* the first iteration, skip OSPs with no objects ready
*/
- if (sfs->os_fprecreated == 0 && speed == 0) {
+ if (ost->ltd_statfs.os_fprecreated == 0 && speed == 0) {
QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
RETURN(rc);
}
/*
* try to use another OSP if this one is degraded
*/
- if (sfs->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);
}
* for the first and second time.
*/
if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) {
- QOS_DEBUG("iter %d: OST%d used by conflicting mirror "
- "component\n", speed, ost_idx);
+ QOS_DEBUG("iter %d: OST%d used by conflicting mirror component\n",
+ speed, ost_idx);
RETURN(rc);
}
/* do not put >1 objects on a single OST, except for overstriping */
- if (lod_qos_is_ost_used(env, ost_idx, stripe_idx)) {
+ if (lod_qos_is_tgt_used(env, ost_idx, stripe_idx)) {
if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
*overstriped = true;
else
* We've successfully declared (reserved) an object
*/
lod_avoid_update(lo, lag);
- lod_qos_ost_in_use(env, stripe_idx, ost_idx);
+ lod_qos_tgt_in_use(env, stripe_idx, ost_idx);
stripe[stripe_idx] = o;
ost_indices[stripe_idx] = ost_idx;
OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2);
* \retval -ENOSPC if not enough OSTs are found
* \retval negative negated errno for other failures
*/
-static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripe, __u32 *ost_indices,
- int flags, struct thandle *th, int comp_idx)
+static int lod_ost_alloc_rr(const struct lu_env *env, struct lod_object *lo,
+ struct dt_object **stripe, __u32 *ost_indices,
+ int flags, struct thandle *th, int comp_idx,
+ __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
struct pool_desc *pool = NULL;
- struct ost_pool *osts;
+ struct lu_tgt_pool *osts;
struct lu_qos_rr *lqr;
- unsigned int i, array_idx;
+ unsigned int i, array_idx;
__u32 ost_start_idx_temp;
__u32 stripe_idx = 0;
__u32 stripe_count, stripe_count_min, ost_idx;
osts = &(pool->pool_obds);
lqr = &(pool->pool_rr);
} else {
- osts = &(m->lod_pool_info);
- lqr = &(m->lod_qos.lq_rr);
+ osts = &m->lod_ost_descs.ltd_tgt_pool;
+ lqr = &(m->lod_ost_descs.ltd_qos.lq_rr);
}
- rc = lod_qos_calc_rr(m, osts, lqr);
+ rc = lod_qos_calc_rr(m, &m->lod_ost_descs, osts, lqr);
if (rc)
GOTO(out, rc);
- rc = lod_qos_ost_in_use_clear(env, stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, stripe_count);
if (rc)
GOTO(out, rc);
- down_read(&m->lod_qos.lq_rw_sem);
+ down_read(&m->lod_ost_descs.ltd_qos.lq_rw_sem);
spin_lock(&lqr->lqr_alloc);
if (--lqr->lqr_start_count <= 0) {
lqr->lqr_start_idx = prandom_u32_max(osts->op_count);
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
continue;
spin_unlock(&lqr->lqr_alloc);
- rc = lod_check_and_reserve_ost(env, lo, lod_comp, sfs, ost_idx,
+ rc = lod_check_and_reserve_ost(env, lo, lod_comp, ost_idx,
speed, &stripe_idx, stripe,
- ost_indices, th, &overstriped);
+ ost_indices, th, &overstriped,
+ reserve);
spin_lock(&lqr->lqr_alloc);
if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
}
spin_unlock(&lqr->lqr_alloc);
- up_read(&m->lod_qos.lq_rw_sem);
+ up_read(&m->lod_ost_descs.ltd_qos.lq_rw_sem);
/* If there are enough OSTs, a component with overstriping requested
* will not actually end up overstriped. The comp should reflect this.
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.
+ *
+ * Allocates a new striping using round-robin algorithm. The function refreshes
+ * all the internal structures (statfs cache, array of available remote MDTs
+ * sorted with regard to MDS, etc). The number of stripes required is taken from
+ * the object (must be prepared by the caller). The caller should ensure nobody
+ * else is trying to create a striping on the object in parallel. All the
+ * internal structures (like pools, etc) are protected and no additional locking
+ * is required. The function succeeds even if a single stripe is allocated.
+ *
+ * \param[in] env execution environment for this thread
+ * \param[in] lo LOD object
+ * \param[out] stripes striping created
+ *
+ * \retval positive stripe objects allocated, including the first stripe
+ * allocated outside
+ * \retval -ENOSPC if not enough MDTs are found
+ * \retval negative negated errno for other failures
+ */
+int lod_mdt_alloc_rr(const struct lu_env *env, struct lod_object *lo,
+ 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_pool *pool;
+ struct lu_qos_rr *lqr;
+ struct lu_tgt_desc *mdt;
+ struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
+ struct lu_fid fid = { 0 };
+ struct dt_object *dto;
+ unsigned int pool_idx;
+ unsigned int i;
+ u32 saved_idx = stripe_idx;
+ u32 start_mdt;
+ u32 mdt_idx;
+ bool use_degraded = false;
+ int tgt_connecting = 0;
+ int rc;
+
+ ENTRY;
+
+ pool = <d->ltd_tgt_pool;
+ lqr = <d->ltd_qos.lq_rr;
+ rc = lod_qos_calc_rr(lod, ltd, pool, lqr);
+ if (rc)
+ RETURN(rc);
+
+ rc = lod_qos_mdt_in_use_init(env, ltd, stripe_idx, stripe_count, pool,
+ stripes);
+ if (rc)
+ RETURN(rc);
+
+ down_read(<d->ltd_qos.lq_rw_sem);
+ spin_lock(&lqr->lqr_alloc);
+ if (--lqr->lqr_start_count <= 0) {
+ lqr->lqr_start_idx = prandom_u32_max(pool->op_count);
+ lqr->lqr_start_count =
+ (LOV_CREATE_RESEED_MIN / max(pool->op_count, 1U) +
+ LOV_CREATE_RESEED_MULT) * max(pool->op_count, 1U);
+ } else if (stripe_count - 1 >= pool->op_count ||
+ lqr->lqr_start_idx > pool->op_count) {
+ /* If we have allocated from all of the tgts, slowly
+ * precess the next start if the tgt/stripe count isn't
+ * already doing this for us. */
+ lqr->lqr_start_idx %= pool->op_count;
+ if (stripe_count - 1 > 1 &&
+ (pool->op_count % (stripe_count - 1)) != 1)
+ ++lqr->lqr_offset_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",
+ stripe_count - 1, lqr->lqr_start_idx, lqr->lqr_start_count,
+ lqr->lqr_offset_idx, pool->op_count, pool->op_count);
+
+ for (i = 0; i < pool->op_count && stripe_idx < stripe_count; i++) {
+ pool_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
+ pool->op_count;
+ ++lqr->lqr_start_idx;
+ mdt_idx = lqr->lqr_pool.op_array[pool_idx];
+ mdt = LTD_TGT(ltd, mdt_idx);
+
+ QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
+ i, lqr->lqr_start_idx, /* XXX: active*/ 0,
+ stripe_idx, pool_idx, mdt_idx);
+
+ if (mdt_idx == LOV_QOS_EMPTY ||
+ !test_bit(mdt_idx, ltd->ltd_tgt_bitmap))
+ continue;
+
+ /* do not put >1 objects on one MDT */
+ if (lod_qos_is_tgt_used(env, mdt_idx, stripe_idx))
+ continue;
+
+ rc = lod_is_tgt_usable(ltd, mdt);
+ if (rc) {
+ if (mdt->ltd_connecting)
+ tgt_connecting = 1;
+ continue;
+ }
+
+ /* try to use another OSP if this one is 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 = 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;
+ }
+
+ dto = dt_locate_at(env, mdt->ltd_tgt, &fid,
+ lo->ldo_obj.do_lu.lo_dev->ld_site->ls_top_dev,
+ &conf);
+
+ spin_lock(&lqr->lqr_alloc);
+ if (IS_ERR(dto)) {
+ QOS_DEBUG("can't alloc stripe on #%u: %d\n",
+ mdt->ltd_index, (int) PTR_ERR(dto));
+
+ if (mdt->ltd_connecting)
+ tgt_connecting = 1;
+ continue;
+ }
+
+ lod_qos_tgt_in_use(env, stripe_idx, mdt_idx);
+ stripes[stripe_idx++] = dto;
+ }
+
+ if (!use_degraded && stripe_idx < stripe_count) {
+ /* Try again, allowing slower MDTs */
+ use_degraded = true;
+ 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 > saved_idx)
+ /* at least one stripe is allocated */
+ RETURN(stripe_idx);
+
+ /* nobody provided us with a single object */
+ if (tgt_connecting)
+ RETURN(-EINPROGRESS);
+
+ RETURN(-ENOSPC);
+}
+
/**
* Allocate a specific striping layout on a user defined set of OSTs.
*
*/
static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
struct dt_object **stripe, __u32 *ost_indices,
- struct thandle *th, int comp_idx)
+ struct thandle *th, int comp_idx, __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
struct dt_object *o;
unsigned int array_idx = 0;
int stripe_count = 0;
LASSERT(lod_comp->llc_ostlist.op_array);
LASSERT(lod_comp->llc_ostlist.op_count);
- rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, lod_comp->llc_stripe_count);
if (rc < 0)
RETURN(rc);
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;
}
/* do not put >1 objects on a single OST, except for
* overstriping
*/
- if (lod_qos_is_ost_used(env, ost_idx, stripe_count) &&
+ if (lod_qos_is_tgt_used(env, ost_idx, stripe_count) &&
!(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)) {
rc = -EINVAL;
break;
}
- rc = lod_statfs_and_check(env, m, ost_idx, sfs);
+ rc = lod_statfs_and_check(env, m, &m->lod_ost_descs,
+ LTD_TGT(&m->lod_ost_descs, ost_idx),
+ reserve);
if (rc < 0) /* this OSP doesn't feel well */
break;
/*
* We've successfully declared (reserved) an object
*/
- lod_qos_ost_in_use(env, stripe_count, ost_idx);
+ lod_qos_tgt_in_use(env, stripe_count, ost_idx);
stripe[stripe_count] = o;
ost_indices[stripe_count] = ost_idx;
stripe_count++;
* \retval -EINVAL requested offset is invalid
* \retval negative errno on failure
*/
-static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripe, __u32 *ost_indices,
- int flags, struct thandle *th, int comp_idx)
+static int lod_ost_alloc_specific(const struct lu_env *env,
+ struct lod_object *lo,
+ struct dt_object **stripe, __u32 *ost_indices,
+ int flags, struct thandle *th, int comp_idx,
+ __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
struct dt_object *o;
+ struct lu_tgt_desc *tgt;
__u32 ost_idx;
unsigned int i, array_idx, ost_count;
int rc, stripe_num = 0;
int speed = 0;
- struct pool_desc *pool = NULL;
- struct ost_pool *osts;
+ struct pool_desc *pool = NULL;
+ struct lu_tgt_pool *osts;
int stripes_per_ost = 1;
bool overstriped = false;
ENTRY;
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
lod_comp = &lo->ldo_comp_entries[comp_idx];
- rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, lod_comp->llc_stripe_count);
if (rc)
GOTO(out, rc);
down_read(&pool_tgt_rw_sem(pool));
osts = &(pool->pool_obds);
} else {
- osts = &(m->lod_pool_info);
+ osts = &m->lod_ost_descs.ltd_tgt_pool;
}
ost_count = osts->op_count;
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
* do not put >1 objects on a single OST, except for
* overstriping, where it is intended
*/
- if (lod_qos_is_ost_used(env, ost_idx, stripe_num)) {
+ if (lod_qos_is_tgt_used(env, ost_idx, stripe_num)) {
if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
overstriped = true;
else
lod_comp_is_ost_used(env, lo, ost_idx))
continue;
+ tgt = LTD_TGT(&m->lod_ost_descs, ost_idx);
+
/* Drop slow OSCs if we can, but not for requested start idx.
*
* This means "if OSC is slow and it is not the requested
* start OST, then it can be skipped, otherwise skip it only
* if it is inactive/recovering/out-of-space." */
- rc = lod_statfs_and_check(env, m, ost_idx, sfs);
+ rc = lod_statfs_and_check(env, m, &m->lod_ost_descs,
+ tgt, reserve);
if (rc) {
/* this OSP doesn't feel well */
continue;
* iteration. Skip OSPs with no objects ready. Don't apply
* this logic to OST specified with stripe_offset.
*/
- if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
+ if (i && !tgt->ltd_statfs.os_fprecreated && !speed)
continue;
o = lod_qos_declare_object_on(env, m, ost_idx, th);
/*
* We've successfully declared (reserved) an object
*/
- lod_qos_ost_in_use(env, stripe_num, ost_idx);
+ lod_qos_tgt_in_use(env, stripe_num, ost_idx);
stripe[stripe_num] = o;
ost_indices[stripe_num] = ost_idx;
stripe_num++;
* \retval -EINVAL requested OST index is invalid
* \retval negative errno on failure
*/
-static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripe, __u32 *ost_indices,
- int flags, struct thandle *th, int comp_idx)
+static int lod_ost_alloc_qos(const struct lu_env *env, struct lod_object *lo,
+ struct dt_object **stripe, __u32 *ost_indices,
+ int flags, struct thandle *th, int comp_idx,
+ __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
struct lod_tgt_desc *ost;
struct dt_object *o;
__u64 total_weight = 0;
struct pool_desc *pool = NULL;
- struct ost_pool *osts;
+ struct lu_tgt_pool *osts;
unsigned int i;
__u32 nfound, good_osts, stripe_count, stripe_count_min;
bool overstriped = false;
down_read(&pool_tgt_rw_sem(pool));
osts = &(pool->pool_obds);
} else {
- osts = &(lod->lod_pool_info);
+ osts = &lod->lod_ost_descs.ltd_tgt_pool;
}
/* Detect -EAGAIN early, before expensive lock is taken. */
- if (!lqos_is_usable(&lod->lod_qos, lod->lod_desc.ld_active_tgt_count))
+ if (!ltd_qos_is_usable(&lod->lod_ost_descs))
GOTO(out_nolock, rc = -EAGAIN);
if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
(lod_comp->llc_stripe_count - 1)/osts->op_count + 1;
/* Do actual allocation, use write lock here. */
- down_write(&lod->lod_qos.lq_rw_sem);
+ down_write(&lod->lod_ost_descs.ltd_qos.lq_rw_sem);
/*
* Check again, while we were sleeping on @lq_rw_sem things could
* change.
*/
- if (!lqos_is_usable(&lod->lod_qos, lod->lod_desc.ld_active_tgt_count))
+ if (!ltd_qos_is_usable(&lod->lod_ost_descs))
GOTO(out, rc = -EAGAIN);
- rc = lqos_calc_penalties(&lod->lod_qos, &lod->lod_ost_descs,
- lod->lod_desc.ld_active_tgt_count,
- lod->lod_desc.ld_qos_maxage, false);
+ rc = ltd_qos_penalties_calc(&lod->lod_ost_descs);
if (rc)
GOTO(out, rc);
- rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, lod_comp->llc_stripe_count);
if (rc)
GOTO(out, rc);
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]);
ost->ltd_qos.ltq_usable = 0;
- rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs);
+ rc = lod_statfs_and_check(env, lod, &lod->lod_ost_descs,
+ ost, reserve);
if (rc) {
/* this OSP doesn't feel well */
continue;
}
- if (sfs->os_state & OS_STATE_DEGRADED)
+ if (ost->ltd_statfs.os_state & OS_STATFS_DEGRADED)
continue;
/* Fail Check before osc_precreate() is called
- so we can only 'fail' single OSC. */
+ * so we can only 'fail' single OSC.
+ */
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
osts->op_array[i] == 0)
continue;
ost->ltd_qos.ltq_usable = 1;
- lqos_calc_weight(ost);
+ lu_tgt_qos_weight_calc(ost);
total_weight += ost->ltd_qos.ltq_weight;
good_osts++;
rand = lu_prandom_u64_max(total_weight);
/* On average, this will hit larger-weighted OSTs more often.
- * 0-weight OSTs will always get used last (only when rand=0) */
+ * 0-weight OSTs will always get used last (only when rand=0)
+ */
for (i = 0; i < osts->op_count; i++) {
__u32 idx = osts->op_array[i];
!(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING))
continue;
- if (lod_qos_is_ost_used(env, idx, nfound)) {
+ if (lod_qos_is_tgt_used(env, idx, nfound)) {
if (lod_comp->llc_pattern &
LOV_PATTERN_OVERSTRIPING)
overstriped = true;
}
lod_avoid_update(lo, lag);
- lod_qos_ost_in_use(env, nfound, idx);
+ lod_qos_tgt_in_use(env, nfound, idx);
stripe[nfound] = o;
ost_indices[nfound] = idx;
- lqos_recalc_weight(&lod->lod_qos, &lod->lod_ost_descs,
- ost,
- lod->lod_desc.ld_active_tgt_count,
- &total_weight);
+ ltd_qos_update(&lod->lod_ost_descs, ost, &total_weight);
nfound++;
rc = 0;
break;
}
/* makes sense to rebalance next time */
- lod->lod_qos.lq_dirty = 1;
- lod->lod_qos.lq_same_space = 0;
+ lod->lod_ost_descs.ltd_qos.lq_dirty = 1;
+ lod->lod_ost_descs.ltd_qos.lq_same_space = 0;
rc = -EAGAIN;
}
lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
out:
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(&lod->lod_ost_descs.ltd_qos.lq_rw_sem);
out_nolock:
if (pool != NULL) {
}
/**
+ * Allocate a striping using an algorithm with weights.
+ *
+ * The function allocates remote MDT objects to create a striping, the first
+ * object was already allocated on current MDT to ensure master object and
+ * the first object are on the same MDT. The algorithm used is based on weights
+ * (both free space and inodes), and it's trying to ensure the space/inodes are
+ * used evenly by MDTs and MDSs. The striping configuration (# of stripes,
+ * offset, pool) is taken from the object and is prepared by the caller.
+ *
+ * If prepared configuration can't be met due to too few MDTs, then allocation
+ * fails.
+ *
+ * No concurrent allocation is allowed on the object and this must be ensured
+ * by the caller. All the internal structures are protected by the function.
+ *
+ * The algorithm has two steps: find available MDTs and calculate their
+ * weights, then select the MDTs with their weights used as the probability.
+ * An MDT with a higher weight is proportionately more likely to be selected
+ * than one with a lower weight.
+ *
+ * \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
+ * lo->ldo_dir_stripe_count
+ * \retval -EAGAIN not enough tgts are found for specified stripe count
+ * \retval -EINVAL requested MDT index is invalid
+ * \retval negative errno on failure
+ */
+int lod_mdt_alloc_qos(const struct lu_env *env, struct lod_object *lo,
+ 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_object_conf conf = { .loc_flags = LOC_F_NEW };
+ struct lu_fid fid = { 0 };
+ const struct lu_tgt_pool *pool;
+ struct lu_tgt_desc *mdt;
+ struct dt_object *dto;
+ u64 total_weight = 0;
+ u32 saved_idx = stripe_idx;
+ u32 mdt_idx;
+ unsigned int good_mdts;
+ unsigned int i;
+ int rc = 0;
+
+ ENTRY;
+
+ 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);
+
+ /*
+ * Check again, while we were sleeping on @lq_rw_sem things could
+ * change.
+ */
+ if (!ltd_qos_is_usable(ltd))
+ GOTO(unlock, rc = -EAGAIN);
+
+ rc = ltd_qos_penalties_calc(ltd);
+ if (rc)
+ GOTO(unlock, rc);
+
+ good_mdts = 0;
+ /* Find all the MDTs that are valid stripe candidates */
+ for (i = 0; i < pool->op_count; i++) {
+ if (!test_bit(pool->op_array[i], ltd->ltd_tgt_bitmap))
+ continue;
+
+ mdt = LTD_TGT(ltd, pool->op_array[i]);
+ mdt->ltd_qos.ltq_usable = 0;
+
+ rc = lod_is_tgt_usable(ltd, mdt);
+ if (rc)
+ continue;
+
+ if (mdt->ltd_statfs.os_state & OS_STATFS_DEGRADED)
+ continue;
+
+ mdt->ltd_qos.ltq_usable = 1;
+ lu_tgt_qos_weight_calc(mdt);
+ total_weight += mdt->ltd_qos.ltq_weight;
+
+ good_mdts++;
+ }
+
+ QOS_DEBUG("found %d good MDTs\n", good_mdts);
+
+ if (good_mdts < stripe_count - stripe_idx)
+ GOTO(unlock, rc = -EAGAIN);
+
+ /* Find enough MDTs with weighted random allocation. */
+ while (stripe_idx < stripe_count) {
+ u64 rand, cur_weight;
+
+ cur_weight = 0;
+ rc = -ENOSPC;
+
+ rand = lu_prandom_u64_max(total_weight);
+
+ /* 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++) {
+ int rc2;
+
+ 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("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",
+ stripe_idx, mdt_idx);
+
+ if (lod_qos_is_tgt_used(env, mdt_idx, stripe_idx))
+ continue;
+
+ rc2 = dt_fid_alloc(env, mdt->ltd_tgt, &fid, NULL, NULL);
+ if (rc2 < 0) {
+ QOS_DEBUG("can't alloc FID on #%u: %d\n",
+ mdt_idx, rc2);
+ continue;
+ }
+
+ conf.loc_flags = LOC_F_NEW;
+ dto = dt_locate_at(env, mdt->ltd_tgt, &fid,
+ lo->ldo_obj.do_lu.lo_dev->ld_site->ls_top_dev,
+ &conf);
+ if (IS_ERR(dto)) {
+ QOS_DEBUG("can't alloc stripe on #%u: %d\n",
+ mdt_idx, (int) PTR_ERR(dto));
+ continue;
+ }
+
+ lod_qos_tgt_in_use(env, stripe_idx, mdt_idx);
+ stripes[stripe_idx] = dto;
+ ltd_qos_update(ltd, mdt, &total_weight);
+ stripe_idx++;
+ rc = 0;
+ break;
+ }
+
+ /* no MDT found on this iteration, give up */
+ if (rc)
+ break;
+ }
+
+ if (unlikely(stripe_idx != stripe_count)) {
+ /*
+ * when the decision to use weighted algorithm was made
+ * we had enough appropriate OSPs, but this state can
+ * change anytime (no space on MDT, broken connection, etc)
+ * so it's possible OSP won't be able to provide us with
+ * an object due to just changed state
+ */
+ QOS_DEBUG("%s: wanted %d objects, found only %d\n",
+ 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;
+ }
+
+ /* makes sense to rebalance next time */
+ ltd->ltd_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_same_space = 0;
+
+ rc = -EAGAIN;
+ } else {
+ rc = stripe_idx;
+ }
+
+unlock:
+ up_write(<d->ltd_qos.lq_rw_sem);
+
+ RETURN(rc);
+}
+
+/**
* Check stripe count the caller can use.
*
* For new layouts (no initialized components), check the total size of the
if (!stripe_count)
- stripe_count = lod->lod_desc.ld_default_stripe_count;
+ stripe_count =
+ lod->lod_ost_descs.ltd_lov_desc.ld_default_stripe_count;
if (!stripe_count)
stripe_count = 1;
/* Overstriping allows more stripes than targets */
- if (stripe_count > lod->lod_desc.ld_active_tgt_count && !overstriping)
- stripe_count = lod->lod_desc.ld_active_tgt_count;
+ if (stripe_count >
+ lod->lod_ost_descs.ltd_lov_desc.ld_active_tgt_count &&
+ !overstriping)
+ stripe_count =
+ lod->lod_ost_descs.ltd_lov_desc.ld_active_tgt_count;
if (lo->ldo_is_composite) {
struct lod_layout_component *lod_comp;
}
max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
+ max_stripes = (max_stripes == 0) ? 0 : max_stripes - 1;
return (stripe_count < max_stripes) ? stripe_count : max_stripes;
}
{
struct lod_layout_component *lod_comp;
struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
- struct lov_desc *desc = &d->lod_desc;
+ struct lov_desc *desc = &d->lod_ost_descs.ltd_lov_desc;
struct lov_user_md_v1 *v1 = NULL;
struct lov_user_md_v3 *v3 = NULL;
struct lov_comp_md_v1 *comp_v1 = NULL;
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 ||
lov_pattern(v1->lmm_pattern) == LOV_PATTERN_MDT)
lod_comp->llc_stripe_count = v1->lmm_stripe_count;
+ if (lov_pattern(lod_comp->llc_pattern) == LOV_PATTERN_MDT &&
+ lod_comp->llc_stripe_count != 0) {
+ CDEBUG(D_LAYOUT, "%s: invalid stripe count: %u\n",
+ lod2obd(d)->obd_name,
+ lod_comp->llc_stripe_count);
+ GOTO(free_comp, rc = -EINVAL);
+ }
+
lod_comp->llc_stripe_offset = v1->lmm_stripe_offset;
lod_obj_set_pool(lo, i, pool_name);
- LASSERT(ergo(lov_pattern(lod_comp->llc_pattern) ==
- LOV_PATTERN_MDT, lod_comp->llc_stripe_count == 0));
-
if (pool_name == NULL)
continue;
int lod_prepare_avoidance(const struct lu_env *env, struct lod_object *lo)
{
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct lod_tgt_descs *ltds = &lod->lod_ost_descs;
struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
struct cfs_bitmap *bitmap = NULL;
__u32 *new_oss = NULL;
- lag->lag_ost_avail = ltds->ltd_tgtnr;
+ lag->lag_ost_avail = lod->lod_ost_count;
/* reset OSS avoid guide array */
lag->lag_oaa_count = 0;
- if (lag->lag_oss_avoid_array && lag->lag_oaa_size < ltds->ltd_tgtnr) {
- OBD_FREE(lag->lag_oss_avoid_array,
- sizeof(__u32) * lag->lag_oaa_size);
+ if (lag->lag_oss_avoid_array &&
+ lag->lag_oaa_size < lod->lod_ost_count) {
+ OBD_FREE_PTR_ARRAY(lag->lag_oss_avoid_array, lag->lag_oaa_size);
lag->lag_oss_avoid_array = NULL;
lag->lag_oaa_size = 0;
}
/* init OST avoid guide bitmap */
if (lag->lag_ost_avoid_bitmap) {
- if (ltds->ltd_tgtnr <= lag->lag_ost_avoid_bitmap->size) {
+ if (lod->lod_ost_count <= lag->lag_ost_avoid_bitmap->size) {
CFS_RESET_BITMAP(lag->lag_ost_avoid_bitmap);
} else {
CFS_FREE_BITMAP(lag->lag_ost_avoid_bitmap);
}
if (!lag->lag_ost_avoid_bitmap) {
- bitmap = CFS_ALLOCATE_BITMAP(ltds->ltd_tgtnr);
+ bitmap = CFS_ALLOCATE_BITMAP(lod->lod_ost_count);
if (!bitmap)
return -ENOMEM;
}
* using OST count to allocate the array to store the OSS
* id.
*/
- OBD_ALLOC(new_oss, sizeof(*new_oss) * ltds->ltd_tgtnr);
+ OBD_ALLOC_PTR_ARRAY(new_oss, lod->lod_ost_count);
if (!new_oss) {
CFS_FREE_BITMAP(bitmap);
return -ENOMEM;
if (new_oss) {
lag->lag_oss_avoid_array = new_oss;
- lag->lag_oaa_size = ltds->ltd_tgtnr;
+ lag->lag_oaa_size = lod->lod_ost_count;
}
if (bitmap)
lag->lag_ost_avoid_bitmap = bitmap;
*/
int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
struct lu_attr *attr, struct thandle *th,
- int comp_idx)
+ int comp_idx, __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
* statfs and check OST targets now, since ld_active_tgt_count
* could be changed if some OSTs are [de]activated manually.
*/
- lod_qos_statfs_update(env, d);
+ lod_qos_statfs_update(env, d, &d->lod_ost_descs);
stripe_len = lod_get_stripe_count(d, lo,
lod_comp->llc_stripe_count,
lod_comp->llc_pattern &
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);
+repeat:
lod_getref(&d->lod_ost_descs);
/* XXX: support for non-0 files w/o objects */
CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n",
- d->lod_desc.ld_tgt_count, stripe_len);
+ d->lod_ost_count, stripe_len);
if (lod_comp->llc_ostlist.op_array &&
lod_comp->llc_ostlist.op_count) {
rc = lod_alloc_ost_list(env, lo, stripe, ost_indices,
- th, comp_idx);
+ th, comp_idx, reserve);
} else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) {
/**
* collect OSTs and OSSs used in other mirrors whose
comp_idx);
lod_collect_avoidance(lo, lag, comp_idx);
- rc = lod_alloc_qos(env, lo, stripe, ost_indices, flag,
- th, comp_idx);
+ rc = lod_ost_alloc_qos(env, lo, stripe, ost_indices,
+ flag, th, comp_idx, reserve);
if (rc == -EAGAIN)
- rc = lod_alloc_rr(env, lo, stripe, ost_indices,
- flag, th, comp_idx);
+ rc = lod_ost_alloc_rr(env, lo, stripe,
+ ost_indices, flag, th,
+ comp_idx, reserve);
} else {
- rc = lod_alloc_specific(env, lo, stripe, ost_indices,
- flag, th, comp_idx);
+ rc = lod_ost_alloc_specific(env, lo, stripe,
+ ost_indices, flag, th,
+ comp_idx, reserve);
}
put_ldts:
lod_putref(d, &d->lod_ost_descs);
for (i = 0; i < stripe_len; i++)
if (stripe[i] != NULL)
dt_object_put(env, stripe[i]);
+
+ /* In case there is no space on any OST, let's ignore
+ * the @reserve space to avoid an error at the init
+ * time, probably the actual IO will be less than the
+ * given @reserve space (aka extension_size). */
+ if (reserve) {
+ reserve = 0;
+ goto repeat;
+ }
lod_comp->llc_stripe_count = 0;
} else {
lod_comp->llc_stripe = stripe;
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);
}
/* no OST available */
/* XXX: should we be waiting a bit to prevent failures during
* cluster initialization? */
- if (d->lod_ostnr == 0)
+ if (!d->lod_ost_count)
RETURN(-EIO);
/*
extent = &lod_comp->llc_extent;
QOS_DEBUG("comp[%d] %lld "DEXT"\n", i, size, PEXT(extent));
if (!lo->ldo_is_composite || size >= extent->e_start) {
- rc = lod_qos_prep_create(env, lo, attr, th, i);
+ rc = lod_qos_prep_create(env, lo, attr, th, i, 0);
if (rc)
break;
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff