NFS: Fix mode bits and nlink count for v4 referral dirs
Subsystems (3)
PATCH 1/1 NFS: Fix mode bits and nlink count for v4 referral dirs
Request for mode bits and nlink count in the nfs4_get_referral call and if server returns them use them instead of hard coded values.
--- fs/nfs/nfs4proc.c | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-)
fs/nfs/nfs4proc.c +17 -3
@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED
@@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode *inode, char *list, size_t list_len)*/ static void nfs_fixup_referral_attributes (struct nfs_fattr *fattr) { +bool fix_mode = true, fix_nlink = true; +if (!(((fattr->valid &NFS_ATTR_FATTR_MOUNTED_ON_FILEID ) || (fattr->valid &NFS_ATTR_FATTR_FILEID )) && (fattr->valid &NFS_ATTR_FATTR_FSID ) && (fattr->valid &NFS_ATTR_FATTR_V4_LOCATIONS )))return ; +if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |=NFS_ATTR_FATTR_TYPE |NFS_ATTR_FATTR_MODE |NFS_ATTR_FATTR_NLINK |NFS_ATTR_FATTR_V4_REFERRAL ; - fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ; } static int _nfs4_proc_fs_locations(struct rpc_clnt *client, struct inode *dir,
Signature · 1 line
2.22.0
Quoted 55 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referralcall and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>--- fs/nfs/nfs4proc.c | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +17
-3@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode *inode,char *list, size_t list_len) */ static void nfs_fixup_referral_attributes (struct nfs_fattr *fattr) { +bool fix_mode = true, fix_nlink = true; +if (!(((fattr->valid &NFS_ATTR_FATTR_MOUNTED_ON_FILEID ) || (fattr->valid &NFS_ATTR_FATTR_FILEID )) && (fattr->valid &NFS_ATTR_FATTR_FSID ) && (fattr->valid &NFS_ATTR_FATTR_V4_LOCATIONS )))return ; +if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |=NFS_ATTR_FATTR_TYPE |NFS_ATTR_FATTR_MODE |NFS_ATTR_FATTR_NLINK |NFS_ATTR_FATTR_V4_REFERRAL ; - fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ; }
Quoted 2 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,
NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than the ones we're already asking for because it may not even have a real directory. The client is required to fake up an inode, hence the existing code.
Signature · 3 lines
Trond Myklebust Linux NFS client maintainer, Hammerspace [email protected]
Quoted 57 lines · Trond Myklebust
Quoted 55 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referralcall and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>--- fs/nfs/nfs4proc.c | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +17
-3@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode *inode,char *list, size_t list_len) */ static void nfs_fixup_referral_attributes (struct nfs_fattr *fattr) { +bool fix_mode = true, fix_nlink = true; +if (!(((fattr->valid &NFS_ATTR_FATTR_MOUNTED_ON_FILEID ) || (fattr->valid &NFS_ATTR_FATTR_FILEID )) && (fattr->valid &NFS_ATTR_FATTR_FSID ) && (fattr->valid &NFS_ATTR_FATTR_V4_LOCATIONS )))return ; +if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |=NFS_ATTR_FATTR_TYPE |NFS_ATTR_FATTR_MODE |NFS_ATTR_FATTR_NLINK |NFS_ATTR_FATTR_V4_REFERRAL ; - fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ; }
Quoted 8 lines
Quoted 2 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than the ones we're already asking for because it may not even have a real directory. The client is required to fake up an inode, hence the existing code.
Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGE
| FATTR4_WORD0_SIZE | FATTR4_WORD0_FSID | FATTR4_WORD0_FILEID | FATTR4_WORD0_FS_LOCATIONS, FATTR4_WORD1_OWNER | FATTR4_WORD1_OWNER_GROUP | FATTR4_WORD1_RAWDEV | FATTR4_WORD1_SPACE_USED | FATTR4_WORD1_TIME_ACCESS | FATTR4_WORD1_TIME_METADATA | FATTR4_WORD1_TIME_MODIFY | FATTR4_WORD1_MOUNTED_ON_FILEID,
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and numlinks?
Also, isn't the whole point of the server returning attribute map is to tell the client which attribute is valid? So, in the case where the server does not have the required information then it will not return those attributes and we will fall back to the old behavior. Whether the server has nlink and mode information is entirely up to the server implementation. For example, the referral's stat information could be maintained in a distributed database which can be accessed from any node in the cluster.
Thanks,
Ashish
Quoted 7 lines
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]
Quoted 91 lines · Ashish Sangwan
Quoted 59 lines · Trond Myklebust
Quoted 58 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referralcall
and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>--- fs/nfs/nfs4proc.c | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +17
-3@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode*inode, char *list, size_t list_len) */ static void nfs_fixup_referral_attributes (struct nfs_fattr *fattr) { +bool fix_mode = true, fix_nlink = true; +if (!(((fattr->valid &NFS_ATTR_FATTR_MOUNTED_ON_FILEID ) || (fattr->valid &NFS_ATTR_FATTR_FILEID )) && (fattr->valid &NFS_ATTR_FATTR_FSID ) && (fattr->valid &NFS_ATTR_FATTR_V4_LOCATIONS )))return ; +if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |=NFS_ATTR_FATTR_TYPE |NFS_ATTR_FATTR_MODE |NFS_ATTR_FATTR_NLINK |NFS_ATTR_FATTR_V4_REFERRAL ; - fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ;}
Quoted 10 lines
Quoted 3 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client,
struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than the
ones we're already asking for because it may not even have a real directory. The client is required to fake up an inode, hence the existing code.Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGEQuoted 4 lines
FATTR4_WORD0_SIZEFATTR4_WORD0_FSIDFATTR4_WORD0_FILEIDFATTR4_WORD0_FS_LOCATIONS,
FATTR4_WORD1_OWNERQuoted 7 lines
FATTR4_WORD1_OWNER_GROUPFATTR4_WORD1_RAWDEVFATTR4_WORD1_SPACE_USEDFATTR4_WORD1_TIME_ACCESSFATTR4_WORD1_TIME_METADATAFATTR4_WORD1_TIME_MODIFYFATTR4_WORD1_MOUNTED_ON_FILEID,
Quoted 3 lines
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and numlinks?
No. We shouldn't be asking for any of that information for a referral because the server isn't supposed to return any values for it.
Chuck and Anna, what's the deal with commit c05cefcc7241? That appears to have changed the original code to speculatively assume that the server will violate RFC5661 Section 11.3.1 and/or RFC7530 Section 8.3.1. Specifically, the paragraph that says:
"
Other attributes SHOULD NOT be made available for absent file systems, even when it is possible to provide them. The server should not assume that more information is always better and should avoid gratuitously providing additional information."
So why is the client asking for them?
Signature · 3 lines
Trond Myklebust Linux NFS client maintainer, Hammerspace [email protected]
Quoted 100 lines · Trond Myklebust
Quoted 91 lines · Ashish Sangwan
Quoted 59 lines · Trond Myklebust
Quoted 58 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referralcall
and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>---
fs/nfs/nfs4proc.c| 20 +++++++++++++++++---1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +3
-1@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode*inode, char *list, size_t list_len) */ static void
nfs_fixup_referral_attributes(struct nfs_fattr *fattr)
{fs/nfs/nfs4proc.c +14
-2+bool fix_mode = true, fix_nlink = true; +if (!(((fattr->valid &NFS_ATTR_FATTR_MOUNTED_ON_FILEID ) || (fattr->valid &NFS_ATTR_FATTR_FILEID )) && (fattr->valid &NFS_ATTR_FATTR_FSID ) && (fattr->valid &NFS_ATTR_FATTR_V4_LOCATIONS )))return ; +if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |=NFS_ATTR_FATTR_TYPE |NFS_ATTR_FATTR_MODE |NFS_ATTR_FATTR_NLINK |NFS_ATTR_FATTR_V4_REFERRAL ; - fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ;}
Quoted 10 lines
Quoted 3 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than the
ones we're already asking for because it may not even have a real directory. The client is required to fake up an inode, hence the existing code.Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGEQuoted 4 lines
FATTR4_WORD0_SIZEFATTR4_WORD0_FSIDFATTR4_WORD0_FILEIDFATTR4_WORD0_FS_LOCATIONS,
FATTR4_WORD1_OWNERQuoted 7 lines
FATTR4_WORD1_OWNER_GROUPFATTR4_WORD1_RAWDEVFATTR4_WORD1_SPACE_USEDFATTR4_WORD1_TIME_ACCESSFATTR4_WORD1_TIME_METADATAFATTR4_WORD1_TIME_MODIFYFATTR4_WORD1_MOUNTED_ON_FILEID,Quoted 3 lines
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and numlinks?
Quoted 7 lines
No. We shouldn't be asking for any of that information for a referral because the server isn't supposed to return any values for it.
Chuck and Anna, what's the deal with commit
c05cefcc7241? That appears to have changed the original code to speculatively assume that the server will violate RFC5661 Section 11.3.1 and/or RFC7530 Section 8.3.1.
The commit is an attempt to address the many complaints we've had about the ugly appearance of referral anchors. The strange "special" default values made the client appear to be broken, and was confusing to some. I consider this to be a UX issue: the information displayed in this case is not meant to be factual, but rather to prevent the user concluding that something is wrong.
I'm not attached to this particular solution, though. Does it make sense to perform the referral mount before returning "ls" results so that the target server has a chance to supply reasonable attribute values for the mounted-on directory object? Just spit balling here.
Quoted 9 lines
Specifically, the paragraph that says:
"
Other attributes SHOULD NOT be made available for absent file systems, even when it is possible to provide them. The server should not assume that more information is always better and should avoid gratuitously providing additional information."So why is the client asking for them?
This paragraph (and it's most modern incarnation in RFC 8881 Section 11.4.1) describes server behavior. The current client behavior is spec-compliant because there is no explicit prohibition in the spec language against a client requesting additional attributes in this case.
Either the server can clear those bitmap flags on the GETATTR reply and not supply those attributes, and clients must be prepared for that.
Or, it's also possible to read this paragraph to mean that the server can provide those attributes and the values should not reflect attributes for the absent file system, but rather something else (eg, server-manufactured defaults, or the attributes from the object on the source server).
And since this is a SHOULD NOT rather than a MUST NOT, servers are still free to return information about the absent file system. Clients are not guaranteed this will be the case, however.
I don't think c05cefcc7241 makes any assumption about whether the server is lying about the extra attributes. Perhaps the server has no better values for these attributes than the client's defaults were.
--
Chuck Lever
Quoted 161 lines · Chuck Lever
Quoted 113 lines
On Oct 15, 2020, at 8:06 AM, Trond Myklebust <
[email protected]> wrote:Quoted 100 lines · Ashish Sangwan
Quoted 64 lines · Trond Myklebust
Quoted 63 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referral
call
and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>---
fs/nfs/nfs4proc.c| 20 +++++++++++++++++---1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +3
-1@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode*inode, char *list, size_t list_len) */ static void
nfs_fixup_referral_attributes(struct nfs_fattr *fattr)
{fs/nfs/nfs4proc.c +2
-0+bool fix_mode = true, fix_nlink = true; + if (!(((fattr->valid & NFS_ATTR_FATTR_MOUNTED_ON_FILEID)||
(fattr->valid & NFS_ATTR_FATTR_FILEID)) && (fattr->valid & NFS_ATTR_FATTR_FSID) && (fattr->valid & NFS_ATTR_FATTR_V4_LOCATIONS))) return;fs/nfs/nfs4proc.c +4
-0+if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |= NFS_ATTR_FATTR_TYPE |
NFS_ATTR_FATTR_MODE|
NFS_ATTR_FATTR_NLINK|NFS_ATTR_FATTR_V4_REFERRAL;fs/nfs/nfs4proc.c +8
-2- fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ;}
Quoted 13 lines
Quoted 3 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than
the
ones we're already asking for because it may not even have a real
directory. The client is required to fake up an inode, hence the
existing code.Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please
elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGEQuoted 4 lines
FATTR4_WORD0_SIZEFATTR4_WORD0_FSIDFATTR4_WORD0_FILEIDFATTR4_WORD0_FS_LOCATIONS,
FATTR4_WORD1_OWNERQuoted 7 lines
FATTR4_WORD1_OWNER_GROUPFATTR4_WORD1_RAWDEVFATTR4_WORD1_SPACE_USEDFATTR4_WORD1_TIME_ACCESSFATTR4_WORD1_TIME_METADATAFATTR4_WORD1_TIME_MODIFYFATTR4_WORD1_MOUNTED_ON_FILEID,Quoted 4 lines
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and
numlinks?No. We shouldn't be asking for any of that information for a referral
because the server isn't supposed to return any values for it.Quoted 5 lines
Chuck and Anna, what's the deal with commit
c05cefcc7241? That appears
to have changed the original code to speculatively assume that the server will violate RFC5661 Section 11.3.1 and/or RFC7530 Section 8.3.1.The commit is an attempt to address the many complaints we've had about the ugly appearance of referral anchors. The strange "special" default values made the client appear to be broken, and was confusing to some. I consider this to be a UX issue: the information displayed in this case is not meant to be factual, but rather to prevent the user concluding that something is wrong.
I'm not attached to this particular solution, though. Does it make sense to perform the referral mount before returning "ls" results so that the target server has a chance to supply reasonable attribute values for the mounted-on directory object? Just spit balling here.
Quoted 11 lines
Specifically, the paragraph that says:
"
Other attributes SHOULD NOT be made available for absent file systems, even when it is possible to provide them. The server
should
not assume that more information is always better and should
avoid
gratuitously providing additional information."So why is the client asking for them?
This paragraph (and it's most modern incarnation in RFC 8881 Section 11.4.1) describes server behavior. The current client behavior is spec-compliant because there is no explicit prohibition in the spec language against a client requesting additional attributes in this case.
Either the server can clear those bitmap flags on the GETATTR reply and not supply those attributes, and clients must be prepared for that.
Or, it's also possible to read this paragraph to mean that the server can provide those attributes and the values should not reflect attributes for the absent file system, but rather something else (eg, server-manufactured defaults, or the attributes from the object on the source server).
Quoted 9 lines
And since this is a SHOULD NOT rather than a MUST NOT, servers are still free to return information about the absent file system. Clients are not guaranteed this will be the case, however.
I don't think
c05cefcc7241makes any assumption about whether the server is lying about the extra attributes. Perhaps the server has no better values for these attributes than the client's defaults were.
SHOULD / SHOULD NOT indicates actions that the server is required to take in the absence of a very good reason to do otherwise. In other words, the client should expect the majority of servers to behave in a certain manner.
It doesn't matter that the client's behaviour is spec compliant. We're asking for information that is not supposed to be divulged by the majority of servers, Furthermore, that information is, quite frankly, utterly irrelevant to the client and application running on it. Any attempt to access that fake object will result in a submount of something completely different on top of that object.
IOW: the only difference here is you're asking that the server provide us with a faked up object (which it is not supposed to do), whereas previously, we were faking that object up ourselves. What's the big deal here?
Signature · 3 lines
Trond Myklebust Linux NFS client maintainer, Hammerspace [email protected]
Quoted 188 lines · Trond Myklebust
Quoted 161 lines · Chuck Lever
Quoted 113 lines
On Oct 15, 2020, at 8:06 AM, Trond Myklebust <
[email protected]> wrote:Quoted 100 lines · Ashish Sangwan
Quoted 64 lines · Trond Myklebust
Quoted 63 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referral
call
and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>---
fs/nfs/nfs4proc.c| 20 +++++++++++++++++---1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +3
-1@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode*inode, char *list, size_t list_len) */ static void
nfs_fixup_referral_attributes(struct nfs_fattr *fattr)
{fs/nfs/nfs4proc.c +2
-0+bool fix_mode = true, fix_nlink = true; + if (!(((fattr->valid & NFS_ATTR_FATTR_MOUNTED_ON_FILEID)||
(fattr->valid & NFS_ATTR_FATTR_FILEID)) && (fattr->valid & NFS_ATTR_FATTR_FSID) && (fattr->valid & NFS_ATTR_FATTR_V4_LOCATIONS))) return;fs/nfs/nfs4proc.c +4
-0+if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |= NFS_ATTR_FATTR_TYPE |
NFS_ATTR_FATTR_MODE|
NFS_ATTR_FATTR_NLINK|NFS_ATTR_FATTR_V4_REFERRAL;fs/nfs/nfs4proc.c +8
-2- fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ;}
Quoted 13 lines
Quoted 3 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than
the
ones we're already asking for because it may not even have a real
directory. The client is required to fake up an inode, hence the
existing code.Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please
elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGEQuoted 4 lines
FATTR4_WORD0_SIZEFATTR4_WORD0_FSIDFATTR4_WORD0_FILEIDFATTR4_WORD0_FS_LOCATIONS,
FATTR4_WORD1_OWNERQuoted 7 lines
FATTR4_WORD1_OWNER_GROUPFATTR4_WORD1_RAWDEVFATTR4_WORD1_SPACE_USEDFATTR4_WORD1_TIME_ACCESSFATTR4_WORD1_TIME_METADATAFATTR4_WORD1_TIME_MODIFYFATTR4_WORD1_MOUNTED_ON_FILEID,Quoted 4 lines
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and
numlinks?No. We shouldn't be asking for any of that information for a referral
because the server isn't supposed to return any values for it.Quoted 5 lines
Chuck and Anna, what's the deal with commit
c05cefcc7241? That appears
to have changed the original code to speculatively assume that the server will violate RFC5661 Section 11.3.1 and/or RFC7530 Section 8.3.1.The commit is an attempt to address the many complaints we've had about the ugly appearance of referral anchors. The strange "special" default values made the client appear to be broken, and was confusing to some. I consider this to be a UX issue: the information displayed in this case is not meant to be factual, but rather to prevent the user concluding that something is wrong.
I'm not attached to this particular solution, though. Does it make sense to perform the referral mount before returning "ls" results so that the target server has a chance to supply reasonable attribute values for the mounted-on directory object? Just spit balling here.
Quoted 11 lines
Specifically, the paragraph that says:
"
Other attributes SHOULD NOT be made available for absent file systems, even when it is possible to provide them. The server
should
not assume that more information is always better and should
avoid
gratuitously providing additional information."So why is the client asking for them?
This paragraph (and it's most modern incarnation in RFC 8881 Section 11.4.1) describes server behavior. The current client behavior is spec-compliant because there is no explicit prohibition in the spec language against a client requesting additional attributes in this case.
Either the server can clear those bitmap flags on the GETATTR reply and not supply those attributes, and clients must be prepared for that.
Or, it's also possible to read this paragraph to mean that the server can provide those attributes and the values should not reflect attributes for the absent file system, but rather something else (eg, server-manufactured defaults, or the attributes from the object on the source server).
Quoted 9 lines
And since this is a SHOULD NOT rather than a MUST NOT, servers are still free to return information about the absent file system. Clients are not guaranteed this will be the case, however.
I don't think
c05cefcc7241makes any assumption about whether the server is lying about the extra attributes. Perhaps the server has no better values for these attributes than the client's defaults were.SHOULD / SHOULD NOT indicates actions that the server is required to take in the absence of a very good reason to do otherwise. In other words, the client should expect the majority of servers to behave in a certain manner.
It doesn't matter that the client's behaviour is spec compliant. We're asking for information that is not supposed to be divulged by the majority of servers, Furthermore, that information is, quite frankly, utterly irrelevant to the client and application running on it. Any attempt to access that fake object will result in a submount of something completely different on top of that object.
Quoted 4 lines
IOW: the only difference here is you're asking that the server provide us with a faked up object (which it is not supposed to do), whereas previously, we were faking that object up ourselves. What's the big deal here?
Right, that boils it down nicely.
The difference has been that by and large the server-provided values don't look broken to users. Perhaps all we need to do is select better defaults for these attributes on Linux clients. I haven't followed Ashish's requirements, so I can't speak to them.
Here is some history.
--
Chuck Lever
Quoted 194 lines · Chuck Lever
Quoted 189 lines · Trond Myklebust
Quoted 161 lines · Chuck Lever
Quoted 113 lines
On Oct 15, 2020, at 8:06 AM, Trond Myklebust <
[email protected]> wrote:Quoted 100 lines · Ashish Sangwan
Quoted 64 lines · Trond Myklebust
Quoted 63 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referral
call
and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>---
fs/nfs/nfs4proc.c| 20 +++++++++++++++++---1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +3
-1@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode*inode, char *list, size_t list_len) */ static void
nfs_fixup_referral_attributes(struct nfs_fattr *fattr)
{fs/nfs/nfs4proc.c +2
-0+bool fix_mode = true, fix_nlink = true; + if (!(((fattr->valid & NFS_ATTR_FATTR_MOUNTED_ON_FILEID)||
(fattr->valid & NFS_ATTR_FATTR_FILEID)) && (fattr->valid & NFS_ATTR_FATTR_FSID) && (fattr->valid & NFS_ATTR_FATTR_V4_LOCATIONS))) return;fs/nfs/nfs4proc.c +4
-0+if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |= NFS_ATTR_FATTR_TYPE |
NFS_ATTR_FATTR_MODE|
NFS_ATTR_FATTR_NLINK|NFS_ATTR_FATTR_V4_REFERRAL;fs/nfs/nfs4proc.c +8
-2- fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ;}
Quoted 13 lines
Quoted 3 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than
the
ones we're already asking for because it may not even have a real
directory. The client is required to fake up an inode, hence the
existing code.Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please
elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGEQuoted 4 lines
FATTR4_WORD0_SIZEFATTR4_WORD0_FSIDFATTR4_WORD0_FILEIDFATTR4_WORD0_FS_LOCATIONS,
FATTR4_WORD1_OWNERQuoted 7 lines
FATTR4_WORD1_OWNER_GROUPFATTR4_WORD1_RAWDEVFATTR4_WORD1_SPACE_USEDFATTR4_WORD1_TIME_ACCESSFATTR4_WORD1_TIME_METADATAFATTR4_WORD1_TIME_MODIFYFATTR4_WORD1_MOUNTED_ON_FILEID,Quoted 4 lines
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and
numlinks?No. We shouldn't be asking for any of that information for a referral
because the server isn't supposed to return any values for it.Quoted 5 lines
Chuck and Anna, what's the deal with commit
c05cefcc7241? That appears
to have changed the original code to speculatively assume that the server will violate RFC5661 Section 11.3.1 and/or RFC7530 Section 8.3.1.The commit is an attempt to address the many complaints we've had about the ugly appearance of referral anchors. The strange "special" default values made the client appear to be broken, and was confusing to some. I consider this to be a UX issue: the information displayed in this case is not meant to be factual, but rather to prevent the user concluding that something is wrong.
I'm not attached to this particular solution, though. Does it make sense to perform the referral mount before returning "ls" results so that the target server has a chance to supply reasonable attribute values for the mounted-on directory object? Just spit balling here.
Quoted 11 lines
Specifically, the paragraph that says:
"
Other attributes SHOULD NOT be made available for absent file systems, even when it is possible to provide them. The server
should
not assume that more information is always better and should
avoid
gratuitously providing additional information."So why is the client asking for them?
This paragraph (and it's most modern incarnation in RFC 8881 Section 11.4.1) describes server behavior. The current client behavior is spec-compliant because there is no explicit prohibition in the spec language against a client requesting additional attributes in this case.
Either the server can clear those bitmap flags on the GETATTR reply and not supply those attributes, and clients must be prepared for that.
Or, it's also possible to read this paragraph to mean that the server can provide those attributes and the values should not reflect attributes for the absent file system, but rather something else (eg, server-manufactured defaults, or the attributes from the object on the source server).
Quoted 9 lines
And since this is a SHOULD NOT rather than a MUST NOT, servers are still free to return information about the absent file system. Clients are not guaranteed this will be the case, however.
I don't think
c05cefcc7241makes any assumption about whether the server is lying about the extra attributes. Perhaps the server has no better values for these attributes than the client's defaults were.SHOULD / SHOULD NOT indicates actions that the server is required to take in the absence of a very good reason to do otherwise. In other words, the client should expect the majority of servers to behave in a certain manner.
It doesn't matter that the client's behaviour is spec compliant. We're asking for information that is not supposed to be divulged by the majority of servers, Furthermore, that information is, quite frankly, utterly irrelevant to the client and application running on it. Any attempt to access that fake object will result in a submount of something completely different on top of that object.
Quoted 4 lines
IOW: the only difference here is you're asking that the server provide us with a faked up object (which it is not supposed to do), whereas previously, we were faking that object up ourselves. What's the big deal here?
Quoted 7 lines
Right, that boils it down nicely.
The difference has been that by and large the server-provided values don't look broken to users. Perhaps all we need to do is select better defaults for these attributes on Linux clients. I haven't followed Ashish's requirements, so I can't speak to them.
The current patch only intended to fix the UX issue, it has no practical purpose.
If it's breaking the RFC then I agree that the patch should not be included. Thanks a lot to both of you for explaining the issue in detail!
Quoted 10 lines
Here is some history.
--
Chuck Lever
Quoted 176 lines · Trond Myklebust
Quoted 161 lines · Chuck Lever
Quoted 113 lines
On Oct 15, 2020, at 8:06 AM, Trond Myklebust <
[email protected]> wrote:Quoted 100 lines · Ashish Sangwan
Quoted 64 lines · Trond Myklebust
Quoted 63 lines · Ashish Sangwan
Request for mode bits and nlink count in the
nfs4_get_referral
call
and if server returns them use them instead of hard coded values.Signed-off-by Ashish Sangwan <[email protected]>---
fs/nfs/nfs4proc.c| 20 +++++++++++++++++---1 file changed, 17 insertions(+), 3 deletions(-)fs/nfs/nfs4proc.c +3
-1@@ -266,7 +266,9 @@ const u32 nfs4_fs_locations_bitmap[3] = { | FATTR4_WORD0_FSID |FATTR4_WORD0_FILEID |FATTR4_WORD0_FS_LOCATIONS , -FATTR4_WORD1_OWNER +FATTR4_WORD1_MODE + |FATTR4_WORD1_NUMLINKS + |FATTR4_WORD1_OWNER |FATTR4_WORD1_OWNER_GROUP |FATTR4_WORD1_RAWDEV |FATTR4_WORD1_SPACE_USED @@ -7594,16 +7596,28 @@ nfs4_listxattr_nfs4_user(struct inode*inode, char *list, size_t list_len) */ static void
nfs_fixup_referral_attributes(struct nfs_fattr *fattr)
{fs/nfs/nfs4proc.c +2
-0+bool fix_mode = true, fix_nlink = true; + if (!(((fattr->valid & NFS_ATTR_FATTR_MOUNTED_ON_FILEID)||
(fattr->valid & NFS_ATTR_FATTR_FILEID)) && (fattr->valid & NFS_ATTR_FATTR_FSID) && (fattr->valid & NFS_ATTR_FATTR_V4_LOCATIONS))) return;fs/nfs/nfs4proc.c +4
-0+if (fattr->valid &NFS_ATTR_FATTR_MODE ) + fix_mode = false; +if (fattr->valid &NFS_ATTR_FATTR_NLINK ) + fix_nlink = false; fattr->valid |= NFS_ATTR_FATTR_TYPE |
NFS_ATTR_FATTR_MODE|
NFS_ATTR_FATTR_NLINK|NFS_ATTR_FATTR_V4_REFERRAL;fs/nfs/nfs4proc.c +8
-2- fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; - fattr->nlink =2 ; + +if (fix_mode) + fattr->mode =S_IFDIR |S_IRUGO |S_IXUGO ; +else + fattr->mode |=S_IFDIR ; + +if (fix_nlink) + fattr->nlink =2 ;}
Quoted 13 lines
Quoted 3 lines
static int _nfs4_proc_fs_locations(struct
rpc_clnt*client, struct
inode *dir,NACK to this patch. The whole point is that if the server has a referral, then it is not going to give us any attributes other than
the
ones we're already asking for because it may not even have a real
directory. The client is required to fake up an inode, hence the
existing code.Hi Trond, thanks for reviewing the patch!
Sorry but I didn't understand the reason to NACK it. Could you please
elaborate your concern?
These are the current attributes we request from the server on a referral:FATTR4_WORD0_CHANGEQuoted 4 lines
FATTR4_WORD0_SIZEFATTR4_WORD0_FSIDFATTR4_WORD0_FILEIDFATTR4_WORD0_FS_LOCATIONS,
FATTR4_WORD1_OWNERQuoted 7 lines
FATTR4_WORD1_OWNER_GROUPFATTR4_WORD1_RAWDEVFATTR4_WORD1_SPACE_USEDFATTR4_WORD1_TIME_ACCESSFATTR4_WORD1_TIME_METADATAFATTR4_WORD1_TIME_MODIFYFATTR4_WORD1_MOUNTED_ON_FILEID,Quoted 4 lines
So you are suggesting that it's ok to ask for SIZE, OWNER, OWNER GROUP, SPACE USED, TIMESTAMPs etc but not ok to ask for mode bits and
numlinks?No. We shouldn't be asking for any of that information for a referral
because the server isn't supposed to return any values for it.Quoted 5 lines
Chuck and Anna, what's the deal with commit
c05cefcc7241? That appears
to have changed the original code to speculatively assume that the server will violate RFC5661 Section 11.3.1 and/or RFC7530 Section 8.3.1.The commit is an attempt to address the many complaints we've had about the ugly appearance of referral anchors. The strange "special" default values made the client appear to be broken, and was confusing to some. I consider this to be a UX issue: the information displayed in this case is not meant to be factual, but rather to prevent the user concluding that something is wrong.
I'm not attached to this particular solution, though. Does it make sense to perform the referral mount before returning "ls" results so that the target server has a chance to supply reasonable attribute values for the mounted-on directory object? Just spit balling here.
Quoted 11 lines
Specifically, the paragraph that says:
"
Other attributes SHOULD NOT be made available for absent file systems, even when it is possible to provide them. The server
should
not assume that more information is always better and should
avoid
gratuitously providing additional information."So why is the client asking for them?
This paragraph (and it's most modern incarnation in RFC 8881 Section 11.4.1) describes server behavior. The current client behavior is spec-compliant because there is no explicit prohibition in the spec language against a client requesting additional attributes in this case.
Either the server can clear those bitmap flags on the GETATTR reply and not supply those attributes, and clients must be prepared for that.
Or, it's also possible to read this paragraph to mean that the server can provide those attributes and the values should not reflect attributes for the absent file system, but rather something else (eg, server-manufactured defaults, or the attributes from the object on the source server).
Quoted 9 lines
And since this is a SHOULD NOT rather than a MUST NOT, servers are still free to return information about the absent file system. Clients are not guaranteed this will be the case, however.
I don't think
c05cefcc7241makes any assumption about whether the server is lying about the extra attributes. Perhaps the server has no better values for these attributes than the client's defaults were.
Quoted 8 lines
SHOULD / SHOULD NOT indicates actions that the server is required to take in the absence of a very good reason to do otherwise. In other words, the client should expect the majority of servers to behave in a certain manner.
It doesn't matter that the client's behaviour is spec compliant. We're asking for information that is not supposed to be divulged by the majority of servers,
We might be reading the spec differently.
I read that SHOULD NOT as saying the server should not hand out attributes for the absent file system, not that it shouldn't hand out attributes at all. My experience at that time was that servers handed out attributes for the referral object that was present on that server. That seems to be completely allowed by the spec language.
But you're correct: it's not relevant to application behavior. As I said, it's merely to prevent users from deciding the Linux client is somehow not working right.
We're working around the client's behavior: it doesn't follow the referral until after a user cd's into that directory. Thus the typical pattern of
$ ls $ cd
Gives surprising results.
Quoted 9 lines
Furthermore, that information is, quite frankly,
utterly irrelevant to the client and application running on it. Any attempt to access that fake object will result in a submount of something completely different on top of that object.IOW: the only difference here is you're asking that the server provide us with a faked up object (which it is not supposed to do), whereas previously, we were faking that object up ourselves. What's the big deal here?
--
Chuck Lever