#define DEBUG_SUBSYSTEM S_LOV
#include <asm/div64.h>
+#include <linux/random.h>
+
#include <libcfs/libcfs.h>
#include <uapi/linux/lustre/lustre_idl.h>
#include <lustre_swab.h>
OST_TGT(lod,i)->ltd_statfs.os_bsize)
/**
- * Add a new target to Quality of Service (QoS) target table.
- *
- * Add a new OST target to the structure representing an OSS. Resort the list
- * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
- * protected internally and no external locking is required.
- *
- * \param[in] lod LOD device
- * \param[in] ost_desc OST description
- *
- * \retval 0 on success
- * \retval -ENOMEM on error
- */
-int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
-{
- struct lod_qos_oss *oss = NULL, *temposs;
- struct obd_export *exp = ost_desc->ltd_exp;
- int rc = 0, found = 0;
- struct list_head *list;
- __u32 id = 0;
- ENTRY;
-
- down_write(&lod->lod_qos.lq_rw_sem);
- /*
- * a bit hacky approach to learn NID of corresponding connection
- * but there is no official API to access information like this
- * with OSD API.
- */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
- if (obd_uuid_equals(&oss->lqo_uuid,
- &exp->exp_connection->c_remote_uuid)) {
- found++;
- break;
- }
- if (oss->lqo_id > id)
- id = oss->lqo_id;
- }
-
- if (!found) {
- OBD_ALLOC_PTR(oss);
- if (!oss)
- GOTO(out, rc = -ENOMEM);
- memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
- sizeof(oss->lqo_uuid));
- ++id;
- oss->lqo_id = id;
- } else {
- /* Assume we have to move this one */
- list_del(&oss->lqo_oss_list);
- }
-
- oss->lqo_ost_count++;
- ost_desc->ltd_qos.ltq_oss = oss;
-
- CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
- obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
- oss->lqo_ost_count);
-
- /* Add sorted by # of OSTs. Find the first entry that we're
- bigger than... */
- list = &lod->lod_qos.lq_oss_list;
- list_for_each_entry(temposs, list, lqo_oss_list) {
- if (oss->lqo_ost_count > temposs->lqo_ost_count)
- break;
- }
- /* ...and add before it. If we're the first or smallest, temposs
- points to the list head, and we add to the end. */
- list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
-
- lod->lod_qos.lq_dirty = 1;
- lod->lod_qos.lq_rr.lqr_dirty = 1;
-
-out:
- up_write(&lod->lod_qos.lq_rw_sem);
- RETURN(rc);
-}
-
-/**
- * Remove OST target from QoS table.
- *
- * Removes given OST target from QoS table and releases related OSS structure
- * if no OSTs remain on the OSS.
- *
- * \param[in] lod LOD device
- * \param[in] ost_desc OST description
- *
- * \retval 0 on success
- * \retval -ENOENT if no OSS was found
- */
-int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
-{
- struct lod_qos_oss *oss;
- int rc = 0;
- ENTRY;
-
- down_write(&lod->lod_qos.lq_rw_sem);
- oss = ost_desc->ltd_qos.ltq_oss;
- if (!oss)
- GOTO(out, rc = -ENOENT);
-
- oss->lqo_ost_count--;
- if (oss->lqo_ost_count == 0) {
- CDEBUG(D_QOS, "removing OSS %s\n",
- obd_uuid2str(&oss->lqo_uuid));
- list_del(&oss->lqo_oss_list);
- ost_desc->ltd_qos.ltq_oss = NULL;
- OBD_FREE_PTR(oss);
- }
-
- lod->lod_qos.lq_dirty = 1;
- lod->lod_qos.lq_rr.lqr_dirty = 1;
-out:
- up_write(&lod->lod_qos.lq_rw_sem);
- RETURN(rc);
-}
-
-/**
* Check whether the target is available for new OST objects.
*
* Request statfs data from the given target and verify it's active and not
* \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)
+int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
+ int index, struct obd_statfs *sfs,
+ struct obd_statfs_info *info)
{
struct lod_tgt_desc *ost;
int rc;
ost = OST_TGT(d,index);
LASSERT(ost);
- rc = dt_statfs(env, ost->ltd_ost, sfs);
+ rc = dt_statfs(env, ost->ltd_ost, sfs, info);
if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
(sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
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))
+ &OST_TGT(lod, idx)->ltd_statfs, NULL))
continue;
if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
/* recalculate weigths */
*/
static int lod_qos_calc_ppo(struct lod_device *lod)
{
- struct lod_qos_oss *oss;
- __u64 ba_max, ba_min, temp;
- __u32 num_active;
- unsigned int i;
- int rc, prio_wide;
- time64_t now, age;
+ struct lu_svr_qos *oss;
+ __u64 ba_max, ba_min, temp;
+ __u32 num_active;
+ unsigned int i;
+ int rc, prio_wide;
+ time64_t now, age;
+
ENTRY;
if (!lod->lod_qos.lq_dirty)
GOTO(out, rc = -EAGAIN);
/* find bavail on each OSS */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
- oss->lqo_bavail = 0;
- lod->lod_qos.lq_active_oss_count = 0;
+ list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list)
+ oss->lsq_bavail = 0;
+ lod->lod_qos.lq_active_svr_count = 0;
/*
* How badly user wants to select OSTs "widely" (not recently chosen
ba_max = max(temp, ba_max);
/* Count the number of usable OSS's */
- if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
- lod->lod_qos.lq_active_oss_count++;
- OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
+ if (OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_bavail == 0)
+ lod->lod_qos.lq_active_svr_count++;
+ OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_bavail += temp;
/* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
temp >>= 1;
(age / lod->lod_desc.ld_qos_maxage);
}
- num_active = lod->lod_qos.lq_active_oss_count - 1;
+ num_active = lod->lod_qos.lq_active_svr_count - 1;
if (num_active < 1) {
/* If there's only 1 OSS, we can't penalize it, so instead
we have to double the OST penalty */
}
/* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
- temp = oss->lqo_bavail >> 1;
- do_div(temp, oss->lqo_ost_count * num_active);
- oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
+ list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) {
+ temp = oss->lsq_bavail >> 1;
+ do_div(temp, oss->lsq_tgt_count * num_active);
+ oss->lsq_penalty_per_obj = (temp * prio_wide) >> 8;
- age = (now - oss->lqo_used) >> 3;
+ age = (now - oss->lsq_used) >> 3;
if (lod->lod_qos.lq_reset ||
age > 32 * lod->lod_desc.ld_qos_maxage)
- oss->lqo_penalty = 0;
+ oss->lsq_penalty = 0;
else if (age > lod->lod_desc.ld_qos_maxage)
/* Decay OSS penalty. */
- oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
+ oss->lsq_penalty >>= age / lod->lod_desc.ld_qos_maxage;
}
lod->lod_qos.lq_dirty = 0;
__u64 temp, temp2;
temp = TGT_BAVAIL(i);
- temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
- OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
+ temp2 = OST_TGT(lod, i)->ltd_qos.ltq_penalty +
+ OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_penalty;
if (temp < temp2)
- OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
+ OST_TGT(lod, i)->ltd_qos.ltq_weight = 0;
else
- OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
+ OST_TGT(lod, i)->ltd_qos.ltq_weight = temp - temp2;
return 0;
}
__u32 index, __u64 *total_wt)
{
struct lod_tgt_desc *ost;
- struct lod_qos_oss *oss;
+ struct lu_svr_qos *oss;
unsigned int j;
ENTRY;
/* Don't allocate on this devuce anymore, until the next alloc_qos */
ost->ltd_qos.ltq_usable = 0;
- oss = ost->ltd_qos.ltq_oss;
+ oss = ost->ltd_qos.ltq_svr;
/* Decay old penalty by half (we're adding max penalty, and don't
want it to run away.) */
ost->ltd_qos.ltq_penalty >>= 1;
- oss->lqo_penalty >>= 1;
+ oss->lsq_penalty >>= 1;
/* mark the OSS and OST as recently used */
- ost->ltd_qos.ltq_used = oss->lqo_used = ktime_get_real_seconds();
+ ost->ltd_qos.ltq_used = oss->lsq_used = ktime_get_real_seconds();
/* Set max penalties for this OST and OSS */
ost->ltd_qos.ltq_penalty +=
ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
- oss->lqo_penalty += oss->lqo_penalty_per_obj *
- lod->lod_qos.lq_active_oss_count;
+ oss->lsq_penalty += oss->lsq_penalty_per_obj *
+ lod->lod_qos.lq_active_svr_count;
/* Decrease all OSS penalties */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
- if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
- oss->lqo_penalty = 0;
+ list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) {
+ if (oss->lsq_penalty < oss->lsq_penalty_per_obj)
+ oss->lsq_penalty = 0;
else
- oss->lqo_penalty -= oss->lqo_penalty_per_obj;
+ oss->lsq_penalty -= oss->lsq_penalty_per_obj;
}
*total_wt = 0;
i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
ost->ltd_qos.ltq_penalty_per_obj >> 10,
ost->ltd_qos.ltq_penalty >> 10,
- ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
- ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
+ ost->ltd_qos.ltq_svr->lsq_penalty_per_obj >> 10,
+ ost->ltd_qos.ltq_svr->lsq_penalty >> 10,
ost->ltd_qos.ltq_weight >> 10);
}
RETURN(0);
}
-void lod_qos_rr_init(struct lod_qos_rr *lqr)
-{
- spin_lock_init(&lqr->lqr_alloc);
- lqr->lqr_dirty = 1;
-}
-
-
#define LOV_QOS_EMPTY ((__u32)-1)
/**
* \retval -ENOMEM fails to allocate the array
*/
static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
- struct lod_qos_rr *lqr)
+ struct lu_qos_rr *lqr)
{
- struct lod_qos_oss *oss;
+ struct lu_svr_qos *oss;
struct lod_tgt_desc *ost;
unsigned placed, real_count;
unsigned int i;
/* Place all the OSTs from 1 OSS at the same time. */
placed = 0;
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
+ list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) {
int j = 0;
for (i = 0; i < lqr->lqr_pool.op_count; i++) {
ost = OST_TGT(lod,src_pool->op_array[i]);
LASSERT(ost && ost->ltd_ost);
- if (ost->ltd_qos.ltq_oss != oss)
+ if (ost->ltd_qos.ltq_svr != oss)
continue;
/* Evenly space these OSTs across arrayspace */
- next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
+ next = j * lqr->lqr_pool.op_count / oss->lsq_tgt_count;
while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
next = (next + 1) % lqr->lqr_pool.op_count;
{
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
struct lod_tgt_desc *ost = OST_TGT(lod, index);
- struct lod_qos_oss *lqo = ost->ltd_qos.ltq_oss;
+ struct lu_svr_qos *lsq = ost->ltd_qos.ltq_svr;
bool used = false;
int i;
/* check OSS use */
for (i = 0; i < lag->lag_oaa_count; i++) {
- if (lag->lag_oss_avoid_array[i] == lqo->lqo_id) {
+ if (lag->lag_oss_avoid_array[i] == lsq->lsq_id) {
used = true;
break;
}
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,
struct dt_object **stripe,
__u32 *ost_indices,
- struct thandle *th)
+ struct thandle *th,
+ bool *overstriped)
{
struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
int rc;
ENTRY;
- rc = lod_statfs_and_check(env, lod, ost_idx, sfs);
+ rc = lod_statfs_and_check(env, lod, ost_idx, sfs, NULL);
if (rc)
RETURN(rc);
"component\n", speed, ost_idx);
RETURN(rc);
}
- /*
- * do not put >1 objects on a single OST
- */
- if (lod_qos_is_ost_used(env, ost_idx, stripe_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_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ *overstriped = true;
+ else
+ RETURN(rc);
+ }
o = lod_qos_declare_object_on(env, lod, ost_idx, th);
if (IS_ERR(o)) {
struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
struct pool_desc *pool = NULL;
struct ost_pool *osts;
- struct lod_qos_rr *lqr;
+ struct lu_qos_rr *lqr;
unsigned int i, array_idx;
__u32 ost_start_idx_temp;
__u32 stripe_idx = 0;
__u32 stripe_count, stripe_count_min, ost_idx;
int rc, speed = 0, ost_connecting = 0;
+ int stripes_per_ost = 1;
+ bool overstriped = false;
ENTRY;
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
down_read(&m->lod_qos.lq_rw_sem);
spin_lock(&lqr->lqr_alloc);
if (--lqr->lqr_start_count <= 0) {
- lqr->lqr_start_idx = cfs_rand() % osts->op_count;
+ lqr->lqr_start_idx = prandom_u32_max(osts->op_count);
lqr->lqr_start_count =
(LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count,
lqr->lqr_offset_idx, osts->op_count, osts->op_count);
- for (i = 0; i < osts->op_count && stripe_idx < stripe_count; i++) {
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ stripes_per_ost =
+ (lod_comp->llc_stripe_count - 1)/osts->op_count + 1;
+
+ for (i = 0; i < osts->op_count * stripes_per_ost
+ && stripe_idx < stripe_count; i++) {
array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
osts->op_count;
++lqr->lqr_start_idx;
continue;
spin_unlock(&lqr->lqr_alloc);
- rc = lod_check_and_reserve_ost(env, lo, sfs, ost_idx, speed,
- &stripe_idx, stripe, ost_indices,
- th);
+ rc = lod_check_and_reserve_ost(env, lo, lod_comp, sfs, ost_idx,
+ speed, &stripe_idx, stripe,
+ ost_indices, th, &overstriped);
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);
+ /* If there are enough OSTs, a component with overstriping requested
+ * will not actually end up overstriped. The comp should reflect this.
+ */
+ if (!overstriped)
+ lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
+
if (stripe_idx) {
lod_comp->llc_stripe_count = stripe_idx;
/* at least one stripe is allocated */
break;
}
- /*
- * do not put >1 objects on a single OST
+ /* 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_ost_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, ost_idx, sfs, NULL);
if (rc < 0) /* this OSP doesn't feel well */
break;
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;
- __u32 ost_idx;
- unsigned int i, array_idx, ost_count;
- int rc, stripe_num = 0;
- int speed = 0;
+ struct dt_object *o;
+ __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;
+ int stripes_per_ost = 1;
+ bool overstriped = false;
ENTRY;
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
GOTO(out, rc = -EINVAL);
}
- for (i = 0; i < ost_count;
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ stripes_per_ost =
+ (lod_comp->llc_stripe_count - 1)/ost_count + 1;
+
+ for (i = 0; i < ost_count * stripes_per_ost;
i++, array_idx = (array_idx + 1) % ost_count) {
ost_idx = osts->op_array[array_idx];
continue;
/*
- * do not put >1 objects on a single OST
+ * 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))
- continue;
+ if (lod_qos_is_ost_used(env, ost_idx, stripe_num)) {
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ overstriped = true;
+ else
+ continue;
+ }
/*
* try not allocate on the OST used by other component
* 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, ost_idx, sfs, NULL);
if (rc) {
/* this OSP doesn't feel well */
continue;
PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
lod_comp->llc_stripe_count);
rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
+
+ /* If there are enough OSTs, a component with overstriping requessted
+ * will not actually end up overstriped. The comp should reflect this.
+ */
+ if (rc == 0 && !overstriped)
+ lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
+
out:
if (pool != NULL) {
up_read(&pool_tgt_rw_sem(pool));
struct ost_pool *osts;
unsigned int i;
__u32 nfound, good_osts, stripe_count, stripe_count_min;
+ bool overstriped = false;
+ int stripes_per_ost = 1;
int rc = 0;
ENTRY;
if (!lod_qos_is_usable(lod))
GOTO(out_nolock, rc = -EAGAIN);
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ stripes_per_ost =
+ (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);
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, osts->op_array[i],
+ sfs, NULL);
if (rc) {
/* this OSP doesn't feel well */
continue;
if (good_osts < stripe_count_min)
GOTO(out, rc = -EAGAIN);
- /* We have enough osts */
- if (good_osts < stripe_count)
- stripe_count = good_osts;
+ /* If we do not have enough OSTs for the requested stripe count, do not
+ * put more stripes per OST than requested.
+ */
+ if (stripe_count / stripes_per_ost > good_osts)
+ stripe_count = good_osts * stripes_per_ost;
/* Find enough OSTs with weighted random allocation. */
nfound = 0;
while (nfound < stripe_count) {
- __u64 rand, cur_weight;
+ u64 rand, cur_weight;
cur_weight = 0;
rc = -ENOSPC;
- if (total_weight) {
-#if BITS_PER_LONG == 32
- rand = cfs_rand() % (unsigned)total_weight;
- /* If total_weight > 32-bit, first generate the high
- * 32 bits of the random number, then add in the low
- * 32 bits (truncated to the upper limit, if needed) */
- if (total_weight > 0xffffffffULL)
- rand = (__u64)(cfs_rand() %
- (unsigned)(total_weight >> 32)) << 32;
- else
- rand = 0;
-
- if (rand == (total_weight & 0xffffffff00000000ULL))
- rand |= cfs_rand() % (unsigned)total_weight;
- else
- rand |= cfs_rand();
-
-#else
- rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
- total_weight;
-#endif
- } else {
- rand = 0;
- }
+ 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) */
QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
/*
- * do not put >1 objects on a single OST
+ * do not put >1 objects on a single OST, except for
+ * overstriping
*/
- if (lod_qos_is_ost_used(env, idx, nfound) ||
- lod_comp_is_ost_used(env, lo, idx))
+ if ((lod_comp_is_ost_used(env, lo, idx)) &&
+ !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING))
continue;
+ if (lod_qos_is_ost_used(env, idx, nfound)) {
+ if (lod_comp->llc_pattern &
+ LOV_PATTERN_OVERSTRIPING)
+ overstriped = true;
+ else
+ continue;
+ }
+
o = lod_qos_declare_object_on(env, lod, idx, th);
if (IS_ERR(o)) {
QOS_DEBUG("can't declare object on #%u: %d\n",
rc = -EAGAIN;
}
+ /* If there are enough OSTs, a component with overstriping requessted
+ * will not actually end up overstriped. The comp should reflect this.
+ */
+ if (rc == 0 && !overstriped)
+ lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
+
out:
up_write(&lod->lod_qos.lq_rw_sem);
}
/**
- * Find largest stripe count the caller can use.
+ * Check stripe count the caller can use.
+ *
+ * For new layouts (no initialized components), check the total size of the
+ * layout against the maximum EA size from the backing file system. This
+ * stops us from creating a layout which will be too large once initialized.
*
+ * For existing layouts (with initialized components):
* Find the maximal possible stripe count not greater than \a stripe_count.
- * Sometimes suggested stripecount can't be reached for a number of reasons:
- * lack of enough active OSTs or the backend does not support EAs that large.
- * If the passed one is 0, then the filesystem's default one is used.
+ * If the provided stripe count is 0, then the filesystem's default is used.
*
* \param[in] lod LOD device
* \param[in] lo The lod_object
* \retval the maximum usable stripe count
*/
__u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo,
- __u16 stripe_count)
+ __u16 stripe_count, bool overstriping)
{
__u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
+ /* max stripe count is based on OSD ea size */
+ unsigned int easize = lod->lod_osd_max_easize;
+ int i;
+
if (!stripe_count)
stripe_count = lod->lod_desc.ld_default_stripe_count;
- if (stripe_count > lod->lod_desc.ld_active_tgt_count)
- stripe_count = lod->lod_desc.ld_active_tgt_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;
- /* stripe count is based on whether OSD can handle larger EA sizes */
- if (lod->lod_osd_max_easize > 0) {
- unsigned int easize = lod->lod_osd_max_easize;
- int i;
-
- if (lo->ldo_is_composite) {
- struct lod_layout_component *lod_comp;
- unsigned int header_sz = sizeof(struct lov_comp_md_v1);
-
- header_sz += sizeof(struct lov_comp_md_entry_v1) *
- lo->ldo_comp_cnt;
- for (i = 0; i < lo->ldo_comp_cnt; i++) {
- lod_comp = &lo->ldo_comp_entries[i];
- if (lod_comp->llc_flags & LCME_FL_INIT)
- header_sz += lov_mds_md_size(
- lod_comp->llc_stripe_count,
- LOV_MAGIC_V3);
- }
- if (easize > header_sz)
- easize -= header_sz;
- else
- easize = 0;
+ if (lo->ldo_is_composite) {
+ struct lod_layout_component *lod_comp;
+ unsigned int header_sz = sizeof(struct lov_comp_md_v1);
+ unsigned int init_comp_sz = 0;
+ unsigned int total_comp_sz = 0;
+ unsigned int comp_sz;
+
+ header_sz += sizeof(struct lov_comp_md_entry_v1) *
+ lo->ldo_comp_cnt;
+
+ for (i = 0; i < lo->ldo_comp_cnt; i++) {
+ lod_comp = &lo->ldo_comp_entries[i];
+ comp_sz = lov_mds_md_size(lod_comp->llc_stripe_count,
+ LOV_MAGIC_V3);
+ total_comp_sz += comp_sz;
+ if (lod_comp->llc_flags & LCME_FL_INIT)
+ init_comp_sz += comp_sz;
}
- max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
+ if (init_comp_sz > 0)
+ total_comp_sz = init_comp_sz;
+
+ header_sz += total_comp_sz;
+
+ if (easize > header_sz)
+ easize -= header_sz;
+ else
+ easize = 0;
}
+ max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
+
return (stripe_count < max_stripes) ? stripe_count : max_stripes;
}
magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEFINED;
if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
- magic != LOV_MAGIC_COMP_V1)
+ magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN)
GOTO(unlock, rc = -EINVAL);
if (magic == LOV_MAGIC_COMP_V1) {
mo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
LCM_FL_FLR_MASK;
mo->ldo_is_composite = 1;
+ } else if (magic == LOV_MAGIC_FOREIGN) {
+ struct lov_foreign_md *foreign;
+ size_t length;
+
+ if (buf->lb_len < offsetof(typeof(*foreign), lfm_value)) {
+ CDEBUG(D_LAYOUT,
+ "buf len %zu < min lov_foreign_md size (%zu)\n",
+ buf->lb_len,
+ offsetof(typeof(*foreign), lfm_value));
+ GOTO(out, rc = -EINVAL);
+ }
+ foreign = (struct lov_foreign_md *)buf->lb_buf;
+ length = foreign_size_le(foreign);
+ if (buf->lb_len < length) {
+ CDEBUG(D_LAYOUT,
+ "buf len %zu < this lov_foreign_md size (%zu)\n",
+ buf->lb_len, length);
+ GOTO(out, rc = -EINVAL);
+ }
+
+ /* just cache foreign LOV EA raw */
+ rc = lod_alloc_foreign_lov(mo, length);
+ if (rc)
+ GOTO(out, rc);
+ memcpy(mo->ldo_foreign_lov, buf->lb_buf, length);
+ GOTO(out, rc);
} else {
mo->ldo_is_composite = 0;
comp_cnt = 1;
if (mo->ldo_is_composite) {
offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
+ v3 = (struct lov_mds_md_v3 *)v1;
magic = le32_to_cpu(v1->lmm_magic);
ext = &comp_v1->lcm_entries[i].lcme_extent;
lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
lod_comp->llc_flags =
le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
+ if (lod_comp->llc_flags & LCME_FL_NOSYNC)
+ lod_comp->llc_timestamp = le64_to_cpu(
+ comp_v1->lcm_entries[i].lcme_timestamp);
lod_comp->llc_id =
le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
if (lod_comp->llc_id == LCME_ID_INVAL)
const struct lu_buf *buf)
{
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_user_md_v1 *v1 = NULL;
- struct lov_user_md_v3 *v3 = NULL;
- struct lov_comp_md_v1 *comp_v1 = NULL;
- char def_pool[LOV_MAXPOOLNAME + 1];
- __u32 magic;
- __u16 comp_cnt;
- __u16 mirror_cnt;
- int i, rc;
+ struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
+ struct lov_desc *desc = &d->lod_desc;
+ struct lov_user_md_v1 *v1 = NULL;
+ struct lov_user_md_v3 *v3 = NULL;
+ struct lov_comp_md_v1 *comp_v1 = NULL;
+ struct lov_foreign_md *lfm = NULL;
+ char def_pool[LOV_MAXPOOLNAME + 1];
+ __u32 magic;
+ __u16 comp_cnt;
+ __u16 mirror_cnt;
+ int i, rc;
ENTRY;
if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
def_pool, sizeof(def_pool));
/* free default striping info */
- lod_free_comp_entries(lo);
+ if (lo->ldo_is_foreign)
+ lod_free_foreign_lov(lo);
+ else
+ lod_free_comp_entries(lo);
rc = lod_verify_striping(d, lo, buf, false);
if (rc)
v3 = buf->lb_buf;
v1 = buf->lb_buf;
comp_v1 = buf->lb_buf;
+ /* {lmm,lfm}_magic position/length work for all LOV formats */
magic = v1->lmm_magic;
if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) {
/* fall trhough */
case LOV_USER_MAGIC_COMP_V1:
break;
+ case __swab32(LOV_USER_MAGIC_FOREIGN):
+ lfm = buf->lb_buf;
+ __swab32s(&lfm->lfm_magic);
+ __swab32s(&lfm->lfm_length);
+ __swab32s(&lfm->lfm_type);
+ __swab32s(&lfm->lfm_flags);
+ magic = lfm->lfm_magic;
+ /* fall through */
+ case LOV_USER_MAGIC_FOREIGN:
+ if (!lfm)
+ lfm = buf->lb_buf;
+ rc = lod_alloc_foreign_lov(lo, foreign_size(lfm));
+ if (rc)
+ RETURN(rc);
+ memcpy(lo->ldo_foreign_lov, buf->lb_buf, foreign_size(lfm));
+ RETURN(0);
default:
CERROR("%s: unrecognized magic %X\n",
lod2obd(d)->obd_name, magic);
lod_comp->llc_extent = *ext;
lod_comp->llc_flags =
comp_v1->lcm_entries[i].lcme_flags &
- LCME_USER_FLAGS;
+ LCME_CL_COMP_FLAGS;
}
pool_name = NULL;
if (v1->lmm_pattern == 0)
v1->lmm_pattern = LOV_PATTERN_RAID0;
if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0 &&
- lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT) {
+ lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT &&
+ lov_pattern(v1->lmm_pattern) !=
+ (LOV_PATTERN_RAID0 | LOV_PATTERN_OVERSTRIPING)) {
CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n",
lod2obd(d)->obd_name, v1->lmm_pattern);
GOTO(free_comp, rc = -EINVAL);
}
}
- if (lod_comp->llc_stripe_count > pool_tgt_count(pool))
+ if (lod_comp->llc_stripe_count > pool_tgt_count(pool) &&
+ !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING))
lod_comp->llc_stripe_count = pool_tgt_count(pool);
lod_pool_putref(pool);
*/
for (j = 0; j < comp->llc_stripe_count; j++) {
struct lod_tgt_desc *ost;
- struct lod_qos_oss *lqo;
+ struct lu_svr_qos *lsq;
int k;
ost = OST_TGT(lod, comp->llc_ost_indices[j]);
- lqo = ost->ltd_qos.ltq_oss;
+ lsq = ost->ltd_qos.ltq_svr;
if (cfs_bitmap_check(bitmap, ost->ltd_index))
continue;
for (k = 0; k < lag->lag_oaa_count; k++) {
if (lag->lag_oss_avoid_array[k] ==
- lqo->lqo_id)
+ lsq->lsq_id)
break;
}
if (k == lag->lag_oaa_count) {
lag->lag_oss_avoid_array[k] =
- lqo->lqo_id;
+ lsq->lsq_id;
lag->lag_oaa_count++;
}
}
LASSERT(lo);
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
lod_comp = &lo->ldo_comp_entries[comp_idx];
+ LASSERT(!(lod_comp->llc_flags & LCME_FL_EXTENSION));
/* A released component is being created */
if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
*/
lod_qos_statfs_update(env, d);
stripe_len = lod_get_stripe_count(d, lo,
- lod_comp->llc_stripe_count);
+ lod_comp->llc_stripe_count,
+ lod_comp->llc_pattern &
+ LOV_PATTERN_OVERSTRIPING);
+
if (stripe_len == 0)
GOTO(out, rc = -ERANGE);
lod_comp->llc_stripe_count = stripe_len;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff