pNFS: Enable layoutreturn operation for return-on-close
authorTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 16 Nov 2016 06:11:25 +0000 (01:11 -0500)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Thu, 1 Dec 2016 22:21:47 +0000 (17:21 -0500)
Amend the pnfs return on close helper functions to enable sending the
layoutreturn op in CLOSE/DELEGRETURN. This closes a potential race between
CLOSE/DELEGRETURN and parallel OPEN calls to the same file, and allows the
client and the server to agree on whether or not there is an outstanding
layout.

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/nfs4proc.c
fs/nfs/pnfs.c
fs/nfs/pnfs.h

index 765af663257ae8b20b038ae3ef4392b7b6ae7624..221d97de0e2c2f2a88d0af5d569f9275b539de02 100644 (file)
@@ -3052,7 +3052,8 @@ static void nfs4_free_closedata(void *data)
        struct super_block *sb = calldata->state->inode->i_sb;
 
        if (calldata->lr.roc)
-               pnfs_roc_release(calldata->state->inode);
+               pnfs_roc_release(&calldata->lr.arg, &calldata->lr.res,
+                               calldata->res.lr_ret);
        nfs4_put_open_state(calldata->state);
        nfs_free_seqid(calldata->arg.seqid);
        nfs4_put_state_owner(sp);
@@ -3103,9 +3104,6 @@ static void nfs4_close_done(struct rpc_task *task, void *data)
        switch (task->tk_status) {
                case 0:
                        res_stateid = &calldata->res.stateid;
-                       if (calldata->lr.roc)
-                               pnfs_roc_set_barrier(state->inode,
-                                                    calldata->lr.roc_barrier);
                        renew_lease(server, calldata->timestamp);
                        break;
                case -NFS4ERR_ADMIN_REVOKED:
@@ -3181,7 +3179,7 @@ static void nfs4_close_prepare(struct rpc_task *task, void *data)
                goto out_no_action;
        }
 
-       if (!calldata->arg.lr_args && nfs4_wait_on_layoutreturn(inode, task)) {
+       if (!calldata->lr.roc && nfs4_wait_on_layoutreturn(inode, task)) {
                nfs_release_seqid(calldata->arg.seqid);
                goto out_wait;
        }
@@ -3195,8 +3193,6 @@ static void nfs4_close_prepare(struct rpc_task *task, void *data)
                else
                        calldata->arg.bitmask = NULL;
        }
-       if (calldata->lr.roc)
-               pnfs_roc_get_barrier(inode, &calldata->lr.roc_barrier);
 
        calldata->arg.share_access =
                nfs4_map_atomic_open_share(NFS_SERVER(inode),
@@ -3223,13 +3219,6 @@ static const struct rpc_call_ops nfs4_close_ops = {
        .rpc_release = nfs4_free_closedata,
 };
 
-static bool nfs4_roc(struct inode *inode)
-{
-       if (!nfs_have_layout(inode))
-               return false;
-       return pnfs_roc(inode);
-}
-
 /* 
  * It is possible for data to be read/written from a mem-mapped file 
  * after the sys_close call (which hits the vfs layer as a flush).
@@ -3281,7 +3270,12 @@ int nfs4_do_close(struct nfs4_state *state, gfp_t gfp_mask, int wait)
        calldata->res.seqid = calldata->arg.seqid;
        calldata->res.server = server;
        calldata->res.lr_ret = -NFS4ERR_NOMATCHING_LAYOUT;
-       calldata->lr.roc = nfs4_roc(state->inode);
+       calldata->lr.roc = pnfs_roc(state->inode,
+                       &calldata->lr.arg, &calldata->lr.res, msg.rpc_cred);
+       if (calldata->lr.roc) {
+               calldata->arg.lr_args = &calldata->lr.arg;
+               calldata->res.lr_res = &calldata->lr.res;
+       }
        nfs_sb_active(calldata->inode->i_sb);
 
        msg.rpc_argp = &calldata->arg;
@@ -5676,8 +5670,6 @@ static void nfs4_delegreturn_done(struct rpc_task *task, void *calldata)
                }
        }
        data->rpc_status = task->tk_status;
-       if (data->lr.roc && data->rpc_status == 0)
-               pnfs_roc_set_barrier(data->inode, data->lr.roc_barrier);
 }
 
 static void nfs4_delegreturn_release(void *calldata)
@@ -5687,7 +5679,8 @@ static void nfs4_delegreturn_release(void *calldata)
 
        if (inode) {
                if (data->lr.roc)
-                       pnfs_roc_release(inode);
+                       pnfs_roc_release(&data->lr.arg, &data->lr.res,
+                                       data->res.lr_ret);
                nfs_iput_and_deactive(inode);
        }
        kfree(calldata);
@@ -5699,13 +5692,9 @@ static void nfs4_delegreturn_prepare(struct rpc_task *task, void *data)
 
        d_data = (struct nfs4_delegreturndata *)data;
 
-       if (!d_data->args.lr_args &&
-           nfs4_wait_on_layoutreturn(d_data->inode, task))
+       if (!d_data->lr.roc && nfs4_wait_on_layoutreturn(d_data->inode, task))
                return;
 
-       if (d_data->lr.roc)
-               pnfs_roc_get_barrier(d_data->inode, &d_data->lr.roc_barrier);
-
        nfs4_setup_sequence(d_data->res.server,
                        &d_data->args.seq_args,
                        &d_data->res.seq_res,
@@ -5756,8 +5745,14 @@ static int _nfs4_proc_delegreturn(struct inode *inode, struct rpc_cred *cred, co
        data->timestamp = jiffies;
        data->rpc_status = 0;
        data->inode = nfs_igrab_and_active(inode);
-       if (data->inode)
-               data->lr.roc = nfs4_roc(inode);
+       if (data->inode) {
+               data->lr.roc = pnfs_roc(inode, &data->lr.arg, &data->lr.res,
+                               cred);
+               if (data->lr.roc) {
+                       data->args.lr_args = &data->lr.arg;
+                       data->res.lr_res = &data->lr.res;
+               }
+       }
 
        task_setup_data.callback_data = data;
        msg.rpc_argp = &data->args;
index a93afdd37203bb2a9539e28f7a5a74f20a42a92c..f61cb81eb5ab6531f4af07a21d8e1cc6b5e57970 100644 (file)
@@ -984,6 +984,20 @@ out_unlock:
 
 }
 
+static void
+pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
+                        u32 seq)
+{
+       if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
+               iomode = IOMODE_ANY;
+       lo->plh_return_iomode = iomode;
+       set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
+       if (seq != 0) {
+               WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
+               lo->plh_return_seq = seq;
+       }
+}
+
 static bool
 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
                nfs4_stateid *stateid,
@@ -1188,17 +1202,22 @@ pnfs_commit_and_return_layout(struct inode *inode)
        return ret;
 }
 
-bool pnfs_roc(struct inode *ino)
+bool pnfs_roc(struct inode *ino,
+               struct nfs4_layoutreturn_args *args,
+               struct nfs4_layoutreturn_res *res,
+               const struct rpc_cred *cred)
 {
        struct nfs_inode *nfsi = NFS_I(ino);
        struct nfs_open_context *ctx;
        struct nfs4_state *state;
        struct pnfs_layout_hdr *lo;
-       struct pnfs_layout_segment *lseg, *tmp;
+       struct pnfs_layout_segment *lseg, *next;
        nfs4_stateid stateid;
-       LIST_HEAD(tmp_list);
-       bool found = false, layoutreturn = false, roc = false;
+       enum pnfs_iomode iomode = 0;
+       bool layoutreturn = false, roc = false;
 
+       if (!nfs_have_layout(ino))
+               return false;
        spin_lock(&ino->i_lock);
        lo = nfsi->layout;
        if (!lo || !pnfs_layout_is_valid(lo) ||
@@ -1217,83 +1236,63 @@ bool pnfs_roc(struct inode *ino)
        }
 
 
-       list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list) {
+       list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
                /* If we are sending layoutreturn, invalidate all valid lsegs */
-               if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
-                       mark_lseg_invalid(lseg, &tmp_list);
-                       found = true;
-               }
+               if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
+                       continue;
+               /*
+                * Note: mark lseg for return so pnfs_layout_remove_lseg
+                * doesn't invalidate the layout for us.
+                */
+               set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
+               if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
+                       continue;
+               pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
        }
 
-       /* always send layoutreturn if being marked so */
-       if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
-               layoutreturn = pnfs_prepare_layoutreturn(lo,
-                               &stateid, NULL);
-               if (layoutreturn)
-                       goto out_noroc;
-       }
+       if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
+               goto out_noroc;
 
        /* ROC in two conditions:
         * 1. there are ROC lsegs
         * 2. we don't send layoutreturn
         */
-       if (found) {
-               /* lo ref dropped in pnfs_roc_release() */
-               pnfs_get_layout_hdr(lo);
-               roc = true;
-       }
+       /* lo ref dropped in pnfs_roc_release() */
+       layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
+       /* If the creds don't match, we can't compound the layoutreturn */
+       if (!layoutreturn || cred != lo->plh_lc_cred)
+               goto out_noroc;
+
+       roc = layoutreturn;
+       pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
+       res->lrs_present = 0;
+       layoutreturn = false;
 
 out_noroc:
        spin_unlock(&ino->i_lock);
-       pnfs_free_lseg_list(&tmp_list);
        pnfs_layoutcommit_inode(ino, true);
        if (layoutreturn)
-               pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
+               pnfs_send_layoutreturn(lo, &stateid, iomode, true);
        return roc;
 }
 
-void pnfs_roc_release(struct inode *ino)
+void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
+               struct nfs4_layoutreturn_res *res,
+               int ret)
 {
-       struct pnfs_layout_hdr *lo;
+       struct pnfs_layout_hdr *lo = args->layout;
+       const nfs4_stateid *arg_stateid = NULL;
+       const nfs4_stateid *res_stateid = NULL;
 
-       spin_lock(&ino->i_lock);
-       lo = NFS_I(ino)->layout;
-       pnfs_clear_layoutreturn_waitbit(lo);
-       if (atomic_dec_and_test(&lo->plh_refcount)) {
-               pnfs_detach_layout_hdr(lo);
-               spin_unlock(&ino->i_lock);
-               pnfs_free_layout_hdr(lo);
-       } else
-               spin_unlock(&ino->i_lock);
-}
-
-void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
-{
-       struct pnfs_layout_hdr *lo;
-
-       spin_lock(&ino->i_lock);
-       lo = NFS_I(ino)->layout;
-       if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
-               lo->plh_barrier = barrier;
-       spin_unlock(&ino->i_lock);
-       trace_nfs4_layoutreturn_on_close(ino, 0);
-}
-
-void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
-{
-       struct nfs_inode *nfsi = NFS_I(ino);
-       struct pnfs_layout_hdr *lo;
-       u32 current_seqid;
-
-       spin_lock(&ino->i_lock);
-       lo = nfsi->layout;
-       current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
-
-       /* Since close does not return a layout stateid for use as
-        * a barrier, we choose the worst-case barrier.
-        */
-       *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
-       spin_unlock(&ino->i_lock);
+       if (ret == 0) {
+               arg_stateid = &args->stateid;
+               if (res->lrs_present)
+                       res_stateid = &res->stateid;
+       }
+       pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
+                       res_stateid);
+       pnfs_put_layout_hdr(lo);
+       trace_nfs4_layoutreturn_on_close(args->inode, 0);
 }
 
 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
@@ -1931,20 +1930,6 @@ out_forget:
        return ERR_PTR(-EAGAIN);
 }
 
-static void
-pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
-                        u32 seq)
-{
-       if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
-               iomode = IOMODE_ANY;
-       lo->plh_return_iomode = iomode;
-       set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
-       if (seq != 0) {
-               WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
-               lo->plh_return_seq = seq;
-       }
-}
-
 /**
  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
  * @lo: pointer to layout header
index 75ff9392127f5e0c09b604a82d3b8606f5ed5f59..f55c065664e124dff22caecc949f94c8670c12bb 100644 (file)
@@ -271,10 +271,13 @@ int pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
                                u32 seq);
 int pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
                struct list_head *lseg_list);
-bool pnfs_roc(struct inode *ino);
-void pnfs_roc_release(struct inode *ino);
-void pnfs_roc_set_barrier(struct inode *ino, u32 barrier);
-void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier);
+bool pnfs_roc(struct inode *ino,
+               struct nfs4_layoutreturn_args *args,
+               struct nfs4_layoutreturn_res *res,
+               const struct rpc_cred *cred);
+void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
+               struct nfs4_layoutreturn_res *res,
+               int ret);
 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task);
 void pnfs_set_layoutcommit(struct inode *, struct pnfs_layout_segment *, loff_t);
 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data);
@@ -666,23 +669,18 @@ pnfs_layoutcommit_outstanding(struct inode *inode)
 
 
 static inline bool
-pnfs_roc(struct inode *ino)
+pnfs_roc(struct inode *ino,
+               struct nfs4_layoutreturn_args *args,
+               struct nfs4_layoutreturn_res *res,
+               const struct rpc_cred *cred)
 {
        return false;
 }
 
 static inline void
-pnfs_roc_release(struct inode *ino)
-{
-}
-
-static inline void
-pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
-{
-}
-
-static inline void
-pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
+pnfs_roc_release(struct nfs4_layoutreturn_args *args,
+               struct nfs4_layoutreturn_res *res,
+               int ret)
 {
 }