Skip to contentSkip to conversation

sgi IRIX-6.5 autofs

DISCUSSION Max Matveev 2 participants 4 messages Apr 18, 2005 – Apr 19, 2005 1 message not in the archive Kernel automounter

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"ian" == Ian Kemp ([email protected]) wrote:

Quoted 6 lines · Greg Bradner
Quoted 6 lines

Let me explain: I have one NAS1 server that has a large number of links to another NAS2. When I run a specific program, an in house tool, it stats the links; which in turn tries to mount NAS2. With the large number of links, the program will stat one, get 'permission denied', and moves to the next. On IRIX, autofs tries to mount with each stat.

That's exactly as designed - stat(2) follows the symlink, if you want the information about the directory or file which symlinks points to, it would have to mount first to do NFS' getattr call to get the info which stat would return. So you should expect a mount attempt for each stat.

ian> autofs on Linux shouldn't mount at all on stat unless the target is ian> contained within (not the mount point itself) the filesystem.

Why? IMO stat(2), as opposed to lstat(2), must prove the info about the node of the remote filesystem regardless of whatever it's the trigger point or a file inside the trigger point.

max

#2 lore raw
Quoted 18 lines · Max Matveev
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"ian" == Ian Kemp ([email protected]) wrote:

Quoted 6 lines · Greg Bradner
Quoted 6 lines

Let me explain: I have one NAS1 server that has a large number of links to another NAS2. When I run a specific program, an in house tool, it stats the links; which in turn tries to mount NAS2. With the large number of links, the program will stat one, get 'permission denied', and moves to the next. On IRIX, autofs tries to mount with each stat.

That's exactly as designed - stat(2) follows the symlink, if you want the information about the directory or file which symlinks points to, it would have to mount first to do NFS' getattr call to get the info which stat would return. So you should expect a mount attempt for each stat.

Quoted 6 lines

ian> autofs on Linux shouldn't mount at all on stat unless the target is ian> contained within (not the mount point itself) the filesystem.

Why? IMO stat(2), as opposed to lstat(2), must prove the info about the node of the remote filesystem regardless of whatever it's the trigger point or a file inside the trigger point.

Equally by design.

Colour ls will mount everything when you list an autofs directory and this (the mount storm on large autofs directories) was seen as undesirable for lazy mounting in a browsable tree.

But I'm open to suggestions and ideas.

Ian

#3 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"ian" == raven <[email protected]> writes:

ian> Colour ls will mount everything when you list an autofs ian> directory and this (the mount storm on large autofs directories) ian> was seen as undesirable for lazy mounting in a browsable tree.

It sounds like an attempt to fix appliction bug in the kernel - color ls should've used lstat(2) instead of stat(2). But if people are happy with such behaviour then there is no need to change it.

BTW, Greg's problem with 30 sec timeout could be because server does not reply - tcpdump/ethereal run on a client should confirm or refute this hypothesis.

max

maintainer
#4 lore raw
Quoted 10 lines · Max Matveev
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"ian" == raven <[email protected]> writes:

ian> Colour ls will mount everything when you list an autofs ian> directory and this (the mount storm on large autofs directories) ian> was seen as undesirable for lazy mounting in a browsable tree.

It sounds like an attempt to fix appliction bug in the kernel - color ls should've used lstat(2) instead of stat(2). But if people are happy with such behaviour then there is no need to change it.

Yea.

Oddly, Solaris automounter seems to behave this way as well.

Quoted 4 lines

BTW, Greg's problem with 30 sec timeout could be because server does not reply - tcpdump/ethereal run on a client should confirm or refute this hypothesis.

Greg?

Ian