#define DEBUG_SUBSYSTEM S_FILTER
#include <dt_object.h>
+#include <linux/falloc.h>
+#include <lustre_nodemap.h>
+
#include "mdt_internal.h"
/* functions below are stubs for now, they will be implemented with
ENTRY;
res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
- &arg->lpa_resid, LDLM_EXTENT, 0);
+ &arg->lpa_resid, LDLM_IBITS, 0);
if (IS_ERR(res)) {
CDEBUG(D_DLMTRACE,
"Failed to get resource for resid %llu/%llu\n",
if (ldlm_has_dom(lock)) {
LDLM_DEBUG(lock, "DOM lock to prolong ");
ldlm_lock_prolong_one(lock, arg);
- break;
+ /* only one PW or EX lock can be granted,
+ * no need to continue search
+ */
+ if (lock->l_granted_mode & (LCK_PW | LCK_EX))
+ break;
}
}
unlock_res(res);
RETURN(0);
/* a bulk write can only hold a reference on a PW extent lock. */
- mode = LCK_PW;
+ mode = LCK_PW | LCK_GROUP;
if (opc == OST_READ)
/* whereas a bulk read can be protected by either a PR or PW
* extent lock */
LASSERT(rnb != NULL);
LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
- pa.lpa_mode = LCK_PW;
+ pa.lpa_mode = LCK_PW | LCK_GROUP;
if (opc == OST_READ)
pa.lpa_mode |= LCK_PR;
if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
RETURN(0);
- if (!(lock->l_granted_mode & LCK_PW))
+ if (!(lock->l_granted_mode & (LCK_PW | LCK_GROUP)))
RETURN(0);
RETURN(1);
LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
oa->o_flags & OBD_FL_SRVLOCK));
- pa.lpa_mode = LCK_PW;
+ pa.lpa_mode = LCK_PW | LCK_GROUP;
CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
mdt_prolong_dom_lock(tsi, &pa);
-
if (pa.lpa_blocks_cnt > 0) {
CDEBUG(D_DLMTRACE,
"%s: refreshed %u locks timeout for req %p.\n",
struct thandle *th;
int rc = 0;
int retries = 0;
- int i;
+ int i, restart = 0;
ENTRY;
retry:
if (!dt_object_exists(dob))
GOTO(out, rc = -ENOENT);
- if (lu_object_is_dying(&mo->mot_header)) {
- /* Commit to stale object can be just skipped silently. */
- CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
- PFID(mdt_object_fid(mo)));
- GOTO(out, rc = 0);
- }
if (niocount == 0) {
rc = -EPROTO;
GOTO(out, rc);
}
+ CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_COMMITRW_DELAY, cfs_fail_val);
+
th = dt_trans_create(env, dt);
if (IS_ERR(th))
GOTO(out, rc = PTR_ERR(th));
GOTO(out_stop, rc);
dt_write_lock(env, dob, 0);
+ if (lu_object_is_dying(&mo->mot_header)) {
+ /* Commit to stale object can be just skipped silently. */
+ CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
+ PFID(mdt_object_fid(mo)));
+ GOTO(unlock, rc = 0);
+ }
rc = dt_write_commit(env, dob, lnb, niocount, th, oa->o_size);
- if (rc)
+ if (rc) {
+ restart = th->th_restart_tran;
GOTO(unlock, rc);
+ }
if (la->la_valid) {
rc = dt_attr_set(env, dob, la, th);
granted = 0;
}
- th->th_result = rc;
+ th->th_result = restart ? 0 : rc;
dt_trans_stop(env, dt, th);
if (rc == -ENOSPC && retries++ < 3) {
CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
retries);
goto retry;
}
+ if (restart) {
+ retries++;
+ restart = 0;
+ if (retries % 10000 == 0)
+ CERROR("%s: restart IO write too many times: %d\n",
+ exp->exp_obd->obd_name, retries);
+ CDEBUG(D_INODE, "retry transaction, retries:%d\n",
+ retries);
+ goto retry;
+ }
out:
dt_bufs_put(env, dob, lnb, niocount);
}
void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
- bool increase_only)
+ struct obdo *oa, bool increase_only)
{
struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
struct ldlm_res_id resid;
return;
/* Update lvbo data if exists. */
- if (mdt_dom_lvb_is_valid(res))
+ if (mdt_dom_lvb_is_valid(res)) {
mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
+ if (oa) {
+ struct ost_lvb *res_lvb = res->lr_lvb_data;
+
+ lock_res(res);
+ oa->o_valid |= OBD_MD_FLBLOCKS | OBD_MD_FLSIZE |
+ OBD_MD_FLMTIME | OBD_MD_FLATIME |
+ OBD_MD_FLCTIME;
+ oa->o_blocks = res_lvb->lvb_blocks;
+ oa->o_size = res_lvb->lvb_size;
+ oa->o_atime = res_lvb->lvb_atime;
+ oa->o_mtime = res_lvb->lvb_mtime;
+ oa->o_ctime = res_lvb->lvb_ctime;
+ unlock_res(res);
+ }
+ }
ldlm_resource_putref(res);
}
struct lu_attr *la = &info->mti_attr.ma_attr;
__u64 valid;
int rc = 0;
+ int root_squash = 0;
LASSERT(mo);
if (cmd == OBD_BRW_WRITE) {
+ struct lu_nodemap *nodemap;
+ __u32 mapped_uid, mapped_gid, mapped_projid;
+
+ nodemap = nodemap_get_from_exp(exp);
+ if (IS_ERR(nodemap))
+ RETURN(PTR_ERR(nodemap));
+ mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID,
+ NODEMAP_FS_TO_CLIENT,
+ oa->o_uid);
+ mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID,
+ NODEMAP_FS_TO_CLIENT,
+ oa->o_gid);
+ mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
+ NODEMAP_FS_TO_CLIENT,
+ oa->o_projid);
+ if (!IS_ERR_OR_NULL(nodemap)) {
+ /* do not bypass quota enforcement if squashed uid */
+ if (unlikely(mapped_uid == nodemap->nm_squash_uid)) {
+ int idx;
+
+ for (idx = 0; idx < npages; idx++)
+ lnb[idx].lnb_flags &=
+ ~OBD_BRW_SYS_RESOURCE;
+ root_squash = 1;
+ }
+ nodemap_putref(nodemap);
+ }
/* Don't update timestamps if this write is older than a
* setattr which modifies the timestamps. b=10150 */
else
obdo_from_la(oa, la, LA_GID | LA_UID);
- mdt_dom_obj_lvb_update(env, mo, false);
+ mdt_dom_obj_lvb_update(env, mo, NULL, false);
/* don't report overquota flag if we failed before reaching
* commit */
if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
oa->o_flags = OBD_FL_NO_PRJQUOTA;
}
+ if (root_squash)
+ oa->o_flags |= OBD_FL_ROOT_SQUASH;
+
oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
}
+ /* Convert back to client IDs. LU-9671.
+ * nodemap_get_from_exp() may fail due to nodemap deactivated,
+ * server ID will be returned back to client in that case.
+ */
+ oa->o_uid = mapped_uid;
+ oa->o_gid = mapped_gid;
+ oa->o_projid = mapped_projid;
} else if (cmd == OBD_BRW_READ) {
/* If oa != NULL then mdt_preprw_read updated the inode
* atime and we should update the lvb so that other glimpses
* will also get the updated value. bug 5972 */
if (oa)
- mdt_dom_obj_lvb_update(env, mo, true);
+ mdt_dom_obj_lvb_update(env, mo, NULL, true);
rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
if (old_rc)
rc = old_rc;
RETURN(rc);
}
+int mdt_object_fallocate(const struct lu_env *env, struct dt_device *dt,
+ struct dt_object *dob, __u64 start, __u64 end,
+ int mode, struct lu_attr *la)
+{
+ struct thandle *th;
+ int rc;
+
+ ENTRY;
+
+ if (!dt_object_exists(dob))
+ RETURN(-ENOENT);
+
+ th = dt_trans_create(env, dt);
+ if (IS_ERR(th))
+ RETURN(PTR_ERR(th));
+
+ rc = dt_declare_attr_set(env, dob, la, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = dt_declare_fallocate(env, dob, start, end, mode, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ tgt_vbr_obj_set(env, dob);
+ rc = dt_trans_start(env, dt, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ dt_write_lock(env, dob, 0);
+ rc = dt_falloc(env, dob, start, end, mode, th);
+ if (rc)
+ GOTO(unlock, rc);
+ rc = dt_attr_set(env, dob, la, th);
+ if (rc)
+ GOTO(unlock, rc);
+unlock:
+ dt_write_unlock(env, dob);
+stop:
+ th->th_result = rc;
+ dt_trans_stop(env, dt, th);
+ RETURN(rc);
+}
+
+/**
+ * MDT request handler for OST_FALLOCATE RPC.
+ *
+ * This is part of request processing. Validate request fields,
+ * preallocate the given MDT object and pack reply.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+int mdt_fallocate_hdl(struct tgt_session_info *tsi)
+{
+ struct obdo *oa = &tsi->tsi_ost_body->oa;
+ struct ptlrpc_request *req = tgt_ses_req(tsi);
+ struct ost_body *repbody;
+ struct mdt_thread_info *info;
+ struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
+ struct obd_export *exp = tsi->tsi_exp;
+ struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
+ struct mdt_object *mo;
+ struct dt_object *dob;
+ struct lu_attr *la;
+ __u64 flags = 0;
+ struct lustre_handle lh = { 0, };
+ int rc, mode;
+ __u64 start, end;
+ bool srvlock;
+ ktime_t kstart = ktime_get();
+
+ repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
+ if (repbody == NULL)
+ RETURN(err_serious(-ENOMEM));
+
+ /*
+ * fallocate start and end are passed in o_size, o_blocks
+ * on the wire.
+ */
+ if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
+ (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
+ RETURN(err_serious(-EPROTO));
+
+ start = oa->o_size;
+ end = oa->o_blocks;
+ mode = oa->o_falloc_mode;
+
+ CDEBUG(D_INODE,
+ "fallocate: "DFID", mode = %#x, start = %lld, end = %lld\n",
+ PFID(&tsi->tsi_fid), mode, start, end);
+
+ /*
+ * mode == 0 (which is standard prealloc) and PUNCH is supported
+ * Rest of mode options are not supported yet.
+ */
+ if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
+ RETURN(-EOPNOTSUPP);
+
+ if (mode & FALLOC_FL_PUNCH_HOLE && !(mode & FALLOC_FL_KEEP_SIZE)) {
+ CWARN("%s: PUNCH mode misses KEEP_SIZE flag, setting it\n",
+ tsi->tsi_tgt->lut_obd->obd_name);
+ mode |= FALLOC_FL_KEEP_SIZE;
+ }
+
+ info = tsi2mdt_info(tsi);
+ la = &info->mti_attr.ma_attr;
+
+ repbody->oa.o_oi = oa->o_oi;
+ repbody->oa.o_valid = OBD_MD_FLID;
+
+ srvlock = oa->o_valid & OBD_MD_FLFLAGS &&
+ oa->o_flags & OBD_FL_SRVLOCK;
+
+ if (srvlock) {
+ rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
+ &flags);
+ if (rc != 0)
+ GOTO(out, rc);
+ }
+
+ mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
+ if (IS_ERR(mo))
+ GOTO(out_unlock, rc = PTR_ERR(mo));
+
+ if (!mdt_object_exists(mo))
+ GOTO(out_put, rc = -ENOENT);
+
+ /* Shouldn't happen on dirs */
+ if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
+ rc = -EPERM;
+ CERROR("%s: fallocate on dir "DFID": rc = %d\n",
+ exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
+ GOTO(out_put, rc);
+ }
+
+ la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
+
+ mdt_dom_write_lock(mo);
+ dob = mdt_obj2dt(mo);
+
+ if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
+ tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
+ tgt_ses_req(tsi)->rq_xid);
+
+ rc = mdt_object_fallocate(tsi->tsi_env, mdt->mdt_bottom, dob, start,
+ end, mode, la);
+ mdt_dom_write_unlock(mo);
+ if (rc)
+ GOTO(out_put, rc);
+
+ mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false);
+
+ mdt_counter_incr(req, LPROC_MDT_FALLOCATE,
+ ktime_us_delta(ktime_get(), kstart));
+
+ EXIT;
+out_put:
+ lu_object_put(tsi->tsi_env, &mo->mot_obj);
+out_unlock:
+ if (srvlock)
+ tgt_data_unlock(&lh, LCK_PW);
+out:
+ mdt_thread_info_fini(info);
+ return rc;
+}
+
int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
struct dt_object *dob, __u64 start, __u64 end,
struct lu_attr *la)
ENTRY;
- /* check that we do support OBD_CONNECT_TRUNCLOCK. */
- BUILD_BUG_ON(!(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK));
-
if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
RETURN(err_serious(-EPROTO));
if (rc)
GOTO(out_put, rc);
- mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
+ mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false);
+
mdt_counter_incr(req, LPROC_MDT_IO_PUNCH,
ktime_us_delta(ktime_get(), kstart));
EXIT;
lu_object_put(tsi->tsi_env, &mo->mot_obj);
out_unlock:
if (srvlock)
- tgt_extent_unlock(&lh, LCK_PW);
+ tgt_data_unlock(&lh, LCK_PW);
out:
mdt_thread_info_fini(info);
return rc;
return rc;
}
-static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
+static void mdt_lvb2reply(struct ldlm_resource *res, struct mdt_body *mb,
+ struct ost_lvb *lvb)
{
struct ost_lvb *res_lvb;
lock_res(res);
res_lvb = res->lr_lvb_data;
- mb->mbo_dom_size = res_lvb->lvb_size;
- mb->mbo_dom_blocks = res_lvb->lvb_blocks;
- mb->mbo_mtime = res_lvb->lvb_mtime;
- mb->mbo_ctime = res_lvb->lvb_ctime;
- mb->mbo_atime = res_lvb->lvb_atime;
-
- CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
-
- mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
- OBD_MD_DOM_SIZE;
+ if (res_lvb) {
+ if (lvb)
+ *lvb = *res_lvb;
+
+ if (mb) {
+ mb->mbo_dom_size = res_lvb->lvb_size;
+ mb->mbo_dom_blocks = res_lvb->lvb_blocks;
+ mb->mbo_mtime = res_lvb->lvb_mtime;
+ mb->mbo_ctime = res_lvb->lvb_ctime;
+ mb->mbo_atime = res_lvb->lvb_atime;
+ mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME |
+ OBD_MD_FLMTIME | OBD_MD_DOM_SIZE;
+ }
+ CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
+ }
unlock_res(res);
}
if (dom_lock || !mdt_dom_lvb_is_valid(res))
mdt_dom_lvbo_update(res, NULL, NULL, false);
- mdt_lvb2body(res, mb);
+ mdt_lvb2reply(res, mb, NULL);
ldlm_resource_putref(res);
RETURN(rc);
}
ldlm_processing_policy policy;
struct ldlm_reply *rep;
struct mdt_body *mbo;
+ struct ost_lvb *lvb;
+ bool old_client = !exp_connect_dom_lvb(mti->mti_exp);
int rc;
ENTRY;
policy = ldlm_get_processing_policy(res);
LASSERT(policy != NULL);
- req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
- req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
+ if (unlikely(old_client)) {
+ req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
+ req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
+ } else {
+ req_capsule_set_size(mti->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
+ sizeof(*lvb));
+ }
rc = req_capsule_server_pack(mti->mti_pill);
if (rc)
RETURN(err_serious(rc));
rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
- if (rep == NULL)
- RETURN(-EPROTO);
- mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
- if (mbo == NULL)
- RETURN(-EPROTO);
+ if (unlikely(old_client)) {
+ mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
+ LASSERT(mbo);
+ lvb = NULL;
+ } else {
+ lvb = req_capsule_server_get(mti->mti_pill, &RMF_DLM_LVB);
+ LASSERT(lvb);
+ mbo = NULL;
+ }
lock_res(res);
/* Check if this is a resend case (MSG_RESENT is set on RPC) and a
if (rc == -ENOENT) {
/* We are racing with unlink(); just return -ENOENT */
rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
- rc = 0;
} else if (rc == -EINVAL) {
/* this is possible is client lock has been cancelled but
* still exists on server. If that lock was found on server
* as only conflicting lock then the client has already
* size authority and glimpse is not needed. */
CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
- rc = 0;
} else if (rc < 0) {
RETURN(rc);
}
/* LVB can be without valid data in case of DOM */
if (!mdt_dom_lvb_is_valid(res))
mdt_dom_lvbo_update(res, lock, NULL, false);
- mdt_lvb2body(res, mbo);
+ mdt_lvb2reply(res, mbo, lvb);
+
RETURN(rc);
}
ENTRY;
- /* Get lock from request for possible resent case. */
- mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
rc = req_capsule_server_pack(mti->mti_pill);
if (mdt_object_remote(mo))
GOTO(out, rc = -EPROTO);
+ /* Get lock from request for possible resent case. */
+ mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
/* resent case */
if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
mdt_lock_handle_init(lhc);
- mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
+ mdt_lh_reg_init(lhc, *lockp);
+
/* This will block MDT thread but it should be fine until
* client caches small amount of data for DoM, which should be
* smaller than one BRW RPC and should be able to be
GOTO(out_fail, rc);
mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
}
- mdt_lvb2body(res, mbo);
+ mdt_lvb2reply(res, mbo, NULL);
out_fail:
rep->lock_policy_res2 = clear_serious(rc);
if (rep->lock_policy_res2) {
rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
out:
+ if (rc < 0)
+ lhc->mlh_reg_lh.cookie = 0ull;
mdt_object_put(mti->mti_env, mo);
RETURN(rc);
}
rc = 0;
out:
if (srvlock)
- tgt_mdt_data_unlock(&lh, lock_mode);
+ tgt_data_unlock(&lh, lock_mode);
repbody->mbo_valid |= OBD_MD_FLFLAGS;
repbody->mbo_flags = OBD_FL_FLUSH;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff