nfs: don't skip CTO on v2/3 mounts, regardless of order of reference puts
Subsystems (2)
Jeff Layton
maintainerRFC PATCH 1/1 nfs: don't skip CTO on v2/3 mounts, regardless of order of reference puts
Olga reported a case of data corruption in a situation where two clients (v3 and v4) were alternately doing open/write/close the same file.
Looking at captures, the v3 client failed to perform any of the GETATTR calls for CTO during one of the events, leading it to overwrite some data that had been written by the v4 client.
We have two calls that work similarly: put_nfs_open_context and put_nfs_open_context_sync. The only difference is the is_sync parameter that gets passed to close_context. The only caller of the _sync variant is in the close codepath.
On a v2/3 mount, if the last reference is not put by put_nfs_open_context_sync, then CTO revalidation never happens. Fix this by adding a new flag to nfs_open_context and set that in put_nfs_open_context_sync. In nfs_close_context, check for that flag when we're checking is_sync and treat them as equivalent.
--- fs/nfs/inode.c | 3 ++- include/linux/nfs_fs.h | 3 ++- 2 files changed, 4 insertions(+), 2 deletions(-)
I'm not sure this is the right fix. Maybe we'd be better off just ignoring the is_sync parameter in nfs_close_context? It's probably functionally equivalent anyway.
Thoughts?
fs/nfs/inode.c +2 -1
@@ -1005,7 +1005,7 @@ void nfs_close_context(struct nfs_open_context *ctx, int is_sync)if (!(ctx->mode &FMODE_WRITE ))return ; -if (!is_sync) +if (!is_sync && !test_bit (NFS_CONTEXT_DO_CLOSE , &ctx->flags))return ; inode =d_inode (ctx->dentry);if (NFS_PROTO (inode)->have_delegation (inode,FMODE_READ ))
@@ -1091,6 +1091,7 @@ EXPORT_SYMBOL_GPL(put_nfs_open_context);static void put_nfs_open_context_sync (struct nfs_open_context *ctx) { +set_bit (NFS_CONTEXT_DO_CLOSE , &ctx->flags);__put_nfs_open_context (ctx,1 ); }
include/linux/nfs_fs.h +2 -1
@@ -85,8 +85,9 @@ struct nfs_open_context {unsigned long flags;#define NFS_CONTEXT_RESEND_WRITES (1)#define NFS_CONTEXT_BAD (2) -#define NFS_CONTEXT_UNLOCK (3) +#define NFS_CONTEXT_UNLOCK (3)#define NFS_CONTEXT_FILE_OPEN (4) +#define NFS_CONTEXT_DO_CLOSE (5)int error;struct list_head list;
Signature · 1 line
2.37.1
Quoted 3 lines · Jeff Layton
Olga reported a case of data corruption in a situation where two clients (v3 and v4) were alternately doing open/write/close the same file.
My thoughts: I did report it but this problem has been impossible to reproduce. Thus perhaps this proposal might be the right thing that needs to be done anyway and that's fine but it's still unclear to me why the problem isn't more reproducible and whether or not this proposal will fix it.
Quoted 69 lines
Looking at captures, the v3 client failed to perform any of the GETATTR calls for CTO during one of the events, leading it to overwrite some data that had been written by the v4 client.
We have two calls that work similarly:
put_nfs_open_contextandput_nfs_open_context_sync. The only difference is theis_syncparameter that gets passed toclose_context. The only caller of the_syncvariant is in the close codepath.On a v2/3 mount, if the last reference is not put by
put_nfs_open_context_sync, then CTO revalidation never happens. Fix this by adding a new flag tonfs_open_contextand set that input_nfs_open_context_sync. Innfs_close_context, check for that flag when we're checkingis_syncand treat them as equivalent.Cc Scott Mayhew <[email protected]>Cc Ben Coddington <[email protected]>Reported-by Olga Kornieskaia <[email protected]>Signed-off-by Jeff Layton <[email protected]>--- fs/nfs/inode.c | 3 ++- include/linux/nfs_fs.h | 3 ++- 2 files changed, 4 insertions(+), 2 deletions(-)I'm not sure this is the right fix. Maybe we'd be better off just ignoring the
is_syncparameter innfs_close_context? It's probably functionally equivalent anyway.Thoughts?
fs/nfs/inode.c +2
-1@@ -1005,7 +1005,7 @@ void nfs_close_context(struct nfs_open_context *ctx, int is_sync)if (!(ctx->mode &FMODE_WRITE ))return ; -if (!is_sync) +if (!is_sync && !test_bit (NFS_CONTEXT_DO_CLOSE , &ctx->flags))return ; inode =d_inode (ctx->dentry);if (NFS_PROTO (inode)->have_delegation (inode,FMODE_READ ))@@ -1091,6 +1091,7 @@ EXPORT_SYMBOL_GPL(put_nfs_open_context);static void put_nfs_open_context_sync (struct nfs_open_context *ctx) { +set_bit (NFS_CONTEXT_DO_CLOSE , &ctx->flags);__put_nfs_open_context (ctx,1 ); }include/linux/nfs_fs.h +2
-2@@ -85,8 +85,9 @@ struct nfs_open_context {unsigned long flags;#define NFS_CONTEXT_RESEND_WRITES (1)#define NFS_CONTEXT_BAD (2) -#define NFS_CONTEXT_UNLOCK (3) +#define NFS_CONTEXT_UNLOCK (3)#define NFS_CONTEXT_FILE_OPEN (4) +#define NFS_CONTEXT_DO_CLOSE (5)int error;struct list_head list; --2.37.1
Quoted 44 lines · Jeff Layton
Olga reported a case of data corruption in a situation where two clients
(v3 and v4) were alternately doing open/write/close the same file.Looking at captures, the v3 client failed to perform any of the GETATTR
calls for CTO during one of the events, leading it to overwrite some data that had been written by the v4 client.We have two calls that work similarly:
put_nfs_open_contextandput_nfs_open_context_sync. The only difference is theis_syncparameter
that gets passed toclose_context. The only caller of the_syncvariant
is in the close codepath.On a v2/3 mount, if the last reference is not put by
put_nfs_open_context_sync, then CTO revalidation never happens. Fix this
by adding a new flag tonfs_open_contextand set that input_nfs_open_context_sync. Innfs_close_context, check for that flag when we're checkingis_syncand treat them as equivalent.Cc Scott Mayhew <[email protected]>Cc Ben Coddington <[email protected]>Reported-by Olga Kornieskaia <[email protected]>Signed-off-by Jeff Layton <[email protected]>---
fs/nfs/inode.c | 3 ++-
include/linux/nfs_fs.h | 3 ++-2 files changed, 4 insertions(+), 2 deletions(-)I'm not sure this is the right fix. Maybe we'd be better off just ignoring the
is_syncparameter innfs_close_context? It's probably functionally equivalent anyway.Thoughts?
fs/nfs/inode.c +0
-0@@ -1005,7 +1005,7 @@ void nfs_close_context(struct nfs_open_context *ctx, int is_sync)
Quoted 8 lines
if (!(ctx->mode &FMODE_WRITE))
return;
- if (!is_sync)
+ if (!is_sync && !test_bit(NFS_CONTEXT_DO_CLOSE,&ctx->flags)) return;
inode =d_inode(ctx->dentry);
if (NFS_PROTO(inode)->have_delegation(inode,FMODE_READ))
NACK. If the 'is_sync' flag is not set, then it is because the function is being called from a context where it is not safe to do a synchronous RPC call.
Signature · 3 lines
Trond Myklebust Linux NFS client maintainer, Hammerspace [email protected]
Jeff Layton
maintainerQuoted 49 lines · Trond Myklebust
Quoted 48 lines · Jeff Layton
Olga reported a case of data corruption in a situation where two clients
(v3 and v4) were alternately doing open/write/close the same file.Looking at captures, the v3 client failed to perform any of the GETATTR
calls for CTO during one of the events, leading it to overwrite some data that had been written by the v4 client.We have two calls that work similarly:
put_nfs_open_contextandput_nfs_open_context_sync. The only difference is theis_syncparameter
that gets passed toclose_context. The only caller of the_syncvariant
is in the close codepath.On a v2/3 mount, if the last reference is not put by
put_nfs_open_context_sync, then CTO revalidation never happens. Fix this
by adding a new flag tonfs_open_contextand set that input_nfs_open_context_sync. Innfs_close_context, check for that flag when we're checkingis_syncand treat them as equivalent.Cc Scott Mayhew <[email protected]>Cc Ben Coddington <[email protected]>Reported-by Olga Kornieskaia <[email protected]>Signed-off-by Jeff Layton <[email protected]>---
fs/nfs/inode.c | 3 ++-
include/linux/nfs_fs.h | 3 ++-2 files changed, 4 insertions(+), 2 deletions(-)I'm not sure this is the right fix. Maybe we'd be better off just ignoring the
is_syncparameter innfs_close_context? It's probably functionally equivalent anyway.Thoughts?
fs/nfs/inode.c +0
-0@@ -1005,7 +1005,7 @@ void nfs_close_context(struct nfs_open_context *ctx, int is_sync)
if (!(ctx->mode &FMODE_WRITE))
return;
- if (!is_sync)
Quoted 9 lines
Quoted 4 lines
+ if (!is_sync && !test_bit(NFS_CONTEXT_DO_CLOSE,
&ctx->flags)) return;
inode =d_inode(ctx->dentry);
if (NFS_PROTO(inode)->have_delegation(inode,FMODE_READ))NACK. If the '
is_sync' flag is not set, then it is because the function is being called from a context where it is not safe to do a synchronous RPC call.
Ok. Any thoughts on the right way to fix this then? It seems like skipping revalidation because the last put wasn't in the close codepath is the wrong thing to do. Should we schedule it to a workqueue?
Signature · 1 line
Jeff Layton <[email protected]>
Quoted 69 lines · Jeff Layton
Quoted 61 lines · Trond Myklebust
Quoted 55 lines · Jeff Layton
Olga reported a case of data corruption in a situation where two clients
(v3 and v4) were alternately doing open/write/close the same file.Looking at captures, the v3 client failed to perform any of the GETATTR
calls for CTO during one of the events, leading it to overwrite some
data that had been written by the v4 client.We have two calls that work similarly:
put_nfs_open_contextandput_nfs_open_context_sync. The only difference is theis_syncparameter
that gets passed toclose_context. The only caller of the_syncvariant
is in the close codepath.On a v2/3 mount, if the last reference is not put by
put_nfs_open_context_sync, then CTO revalidation never happens. Fix
this
by adding a new flag tonfs_open_contextand set that input_nfs_open_context_sync. Innfs_close_context, check for that flag
when we're checkingis_syncand treat them as equivalent.Cc Scott Mayhew <[email protected]>Cc Ben Coddington <[email protected]>Reported-by Olga Kornieskaia <[email protected]>Signed-off-by Jeff Layton <[email protected]>---
fs/nfs/inode.c | 3 ++-
include/linux/nfs_fs.h | 3 ++-2 files changed, 4 insertions(+), 2 deletions(-)I'm not sure this is the right fix. Maybe we'd be better off just ignoring the
is_syncparameter innfs_close_context? It's probably
functionally equivalent anyway.Thoughts?
fs/nfs/inode.c +0
-0@@ -1005,7 +1005,7 @@ void nfs_close_context(struct
nfs_open_context
*ctx, intis_sync)
if (!(ctx->mode &FMODE_WRITE))
return;
- if (!is_sync)
+ if (!is_sync && !test_bit(NFS_CONTEXT_DO_CLOSE, &ctx-Quoted 4 lines
Quoted 1 line
flags))
return;
inode =d_inode(ctx->dentry);
if (NFS_PROTO(inode)->have_delegation(inode,FMODE_READ))Quoted 6 lines
NACK. If the '
is_sync' flag is not set, then it is because the function
is being called from a context where it is not safe to do a synchronous
RPC call.
Quoted 4 lines
Ok. Any thoughts on the right way to fix this then? It seems like skipping revalidation because the last put wasn't in the close codepath
is the wrong thing to do. Should we schedule it to a workqueue?
If the attributes are marked as up to date, because they were already revalidated after the I/O was performed, then we're fine. If they are not up to date, then that is expected to trigger a cache invalidation on the next open when we compare the client's cached values to the server-retrieved values for the file attributes.
i.e. this ought to be just a performance problem, and not a correctness problem, if attribute cache revalidation is working as designed.
Signature · 3 lines
Trond Myklebust Linux NFS client maintainer, Hammerspace [email protected]