make knfsd interact cleanly with HSMs
Subsystems (3)
PATCH RESEND 1/1 make knfsd interact cleanly with HSMs
G'day,
The NFSv3 protocol specifies an error, NFS3ERR_JUKEBOX, which a server should return when an I/O operation will take a very long time. This causes a different pattern of retries in clients, and avoids a number of serious problems associated with I/Os which take longer than an RPC timeout. The Linux knfsd server has code to generate the jukebox error and many NFS clients are known to have working code to handle it.
One scenario in which a server should emit the JUKEBOX error is when a file data which the client is attempting to access is managed by an HSM (Hierarchical Storage Manager) and is not present on the disk and needs to be brought in from tape. Due to the nature of tapes this operation can take minutes rather than the milliseconds normally seen for local file data.
Currently the Linux knfsd handles this situation poorly. A READ NFS call will cause the nfsd thread handling it to block until the file is available, without sending a reply to the NFS client. After a few seconds the client retries, and this second READ call causes another nfsd to block behind the first one. A few seconds later and the client's retries have blocked *all* the nfsd threads, and all NFS service from the server stops until the original file arrives on disk.
WRITEs and SETATTRs which truncate the file are marginally better, in that the knfsd dupcache will catch the retries and drop them without blocking an nfsd (the dupcache *will* catch the retries because the cache entry remains in RC_INPROG state and is not reused until the first call finishes). However the first call still blocks, so given WRITEs to enough offline files the server can still be locked up.
There are also client-side implications, depending on the client implementation. For example, on a Linux client an RPC retry loop uses an RPC request slot, so reads from enough separate offline files can lock up a mountpoint.
This patch seeks to remedy the interaction between knfsd and HSMs by providing mechanisms to allow knfsd to tell an underlying filesystem (which supports HSMs) not to block for reads, writes and truncates of offline files. It's a port of a Linux 2.4 patch used in SGI's ProPack distro since 2004 and in SLES9 since SP2. The patch:
* provides a new ATTR_NO_BLOCK flag which the kernel can use to tell a filesystem's inode_ops->setattr() operation not to block when truncating an offline file. XFS already obeys this flag (inside a #ifdef)
* changes knfsd to provide ATTR_NO_BLOCK when it does the VFS calls to implement the SETATTR NFS call.
* changes knfsd to supply the O_NONBLOCK flag in the temporary struct file it uses for VFS reads and writes, in order to ask the filesystem not to block when reading or writing an offline file. XFS already obeys this new semantic for O_NONBLOCK (and in SLES9 so does JFS).
* adds code to translate the -EAGAIN the filesystem returns when it would have blocked, to the -ETIMEDOUT that knfsd expects.
(SLES9 patch Acked-by: [email protected])
---
This is a resend of
http://marc.theaimsgroup.com/?l=linux-nfs&m=111087383132762&w=2
fs/nfsd/vfs.c | 33 +++++++++++++++++++++++++++++++-- include/linux/fs.h | 1 + 2 files changed, 32 insertions(+), 2 deletions(-)
fs/nfsd/vfs.c +31 -2
@@ -327,6 +327,16 @@ nfsd_setattr(struct svc_rqst *rqstp, str goto out_nfserr; }DQUOT_INIT (inode); + + +/* +* Tell a Hierarchical Storage Manager (e.g. via DMAPI) to +* return EAGAIN when an action would take minutes instead of +* milliseconds so that NFS can reply to the client with +* NFSERR_JUKEBOX instead of blocking an nfsd thread. +*/ +if (rqstp->rq_vers >=3 ) + iap->ia_valid |=ATTR_NO_BLOCK ; } imode = inode->i_mode;
@@ -349,6 +359,9 @@ nfsd_setattr(struct svc_rqst *rqstp, strif (!check_guard || guardtime == inode->i_ctime.tv_sec) {fh_lock (fhp); err =notify_change (dentry, iap); +/* to get NFSERR_JUKEBOX on the wire, need -ETIMEDOUT */ +if (err == -EAGAIN ) + err = -ETIMEDOUT ; err =nfserrno (err);fh_unlock (fhp); }
@@ -834,6 +847,10 @@ nfsd_vfs_read(struct svc_rqst *rqstp, stif (ra && ra->p_set) file->f_ra = ra->p_ra; +/* Support HSMs -- see comment in nfsd_setattr() */ +if (rqstp->rq_vers >=3 ) + file->f_flags |=O_NONBLOCK ; +if (file->f_op->sendfile) {svc_pushback_unused_pages (rqstp); err = file->f_op->sendfile(file, &offset, *count,
@@ -859,8 +876,12 @@ nfsd_vfs_read(struct svc_rqst *rqstp, st *count = err; err =0 ;fsnotify_access (file->f_dentry); - }else + }else { +/* to get NFSERR_JUKEBOX on the wire, need -ETIMEDOUT */ +if (err == -EAGAIN ) + err = -ETIMEDOUT ; err =nfserrno (err); + } out:return err; }
@@ -918,6 +939,10 @@ nfsd_vfs_write(struct svc_rqst *rqstp, sif (stable && !EX_WGATHER (exp)) file->f_flags |=O_SYNC ; +/* Support HSMs -- see comment in nfsd_setattr() */ +if (rqstp->rq_vers >=3 ) + file->f_flags |=O_NONBLOCK ; +/* Write the data. */ oldfs =get_fs ();set_fs (KERNEL_DS ); err =vfs_writev (file, (struct iovec __user *)vec, vlen, &offset);
@@ -970,8 +995,12 @@ nfsd_vfs_write(struct svc_rqst *rqstp, sdprintk ("nfsd: write complete err=%d\n" , err);if (err >=0 ) err =0 ; -else +else { +/* to get NFSERR_JUKEBOX on the wire, need -ETIMEDOUT */ +if (err == -EAGAIN ) + err = -ETIMEDOUT ; err =nfserrno (err); + } out:return err; }
include/linux/fs.h +1 -0
@@ -273,6 +273,7 @@ typedef void (dio_iodone_t)(struct kiocb#define ATTR_KILL_SUID 2048#define ATTR_KILL_SGID 4096#define ATTR_FILE 8192 +#define ATTR_NO_BLOCK 32768/* Return EAGAIN and don't block on long truncates */ /* * This is the Inode Attributes structure, used for notify_change(). It
Signature · 2 lines
Greg Banks, R&D Software Engineer, SGI Australian Software Group. I don't speak for SGI.
Quoted 15 lines · On Friday May 5, [email protected] wrote:
This patch seeks to remedy the interaction between knfsd and HSMs by providing mechanisms to allow knfsd to tell an underlying filesystem (which supports HSMs) not to block for reads, writes and truncates of offline files. It's a port of a Linux 2.4 patch used in SGI's ProPack distro since 2004 and in SLES9 since SP2. The patch:
* provides a new
ATTR_NO_BLOCKflag which the kernel can use to tell a filesystem'sinode_ops->setattr()operation not to block when truncating an offline file. XFS already obeys this flag (inside a #ifdef)* changes knfsd to provide
ATTR_NO_BLOCKwhen it does the VFS calls to implement the SETATTR NFS call.
Quoted 9 lines
* changes knfsd to supply the
O_NONBLOCKflag in the temporary struct file it uses for VFS reads and writes, in order to ask the filesystem not to block when reading or writing an offline file. XFS already obeys this new semantic forO_NONBLOCK(and in SLES9 so does JFS).* adds code to translate the
-EAGAINthe filesystem returns when it would have blocked, to the-ETIMEDOUTthat knfsd expects.
Yes, I'm happy with this. All the changes make sense and look right.
Thanks for persisting with it.
I'll send it on up the chain...
NeilBrown
NACK. As long as we have no HSM support in the tree there's no reason to add this. From the kernel's point it's just untested and unused code that can break.
Neil Brown
Acked-byQuoted 3 lines · On Monday May 8, [email protected] wrote:
NACK. As long as we have no HSM support in the tree there's no reason to add this. From the kernel's point it's just untested and unused code that can break.
Greg: you seemed to suggest that there was already code in XFS that could make use of this. If that is so: could you point us to it please.
Alternately, if this will only be used by future patches, please feel free to add this patch to that set, and add
to it.
Thanks,
NeilBrown
Quoted 8 lines · Neil Brown
Quoted 3 lines · On Monday May 8, [email protected] wrote:
NACK. As long as we have no HSM support in the tree there's no reason to add this. From the kernel's point it's just untested and unused code that can break.
Greg: you seemed to suggest that there was already code in XFS that could make use of this. If that is so: could you point us to it please.
The code in question is in fs/xfs/linux-2.6/xfs_iops.c::xfs_vn_setattr, fs/xfs/xfs_dmapi.h::AT_DELAY_FLAG, and its users in xfs_setattr and xfs_free_file_space and one or two other spots in XFS. The DMAPI code (driver, outside XFS) can be found in the XFS CVS tree on oss.sgi.com.
Quoted 2 lines
Alternately, if this will only be used by future patches, please feel free to add this patch to that set, and add
The DMAPI code was released at the same time as XFS, but it seems to be a spec that is not widely used (unsurprising since not may people have tape robots doing their bidding) and, perhaps more importantly, its not viewed as a particularly clean spec by, erm, certain kernel developers. ;)
cheers.
Signature · 1 line
Nathan
Quoted 9 lines · Neil Brown
Quoted 3 lines · On Monday May 8, [email protected] wrote:
NACK. As long as we have no HSM support in the tree there's no reason to add this. From the kernel's point it's just untested and unused code that can break.
Greg: you seemed to suggest that there was already code in XFS that could make use of this. If that is so: could you point us to it please.
It's used by SGI's out of tree dmapi implementation. Because dmapi is such an utterly braindead standard I don't expect anyone to submit a kernel-based implementation for inclusion, although support for a big enough subset of that standard could be archived by proper kernel <-> userspace cooperation.
Quoted 5 lines · Christoph Hellwig
Quoted 4 lines · Neil Brown
Quoted 3 lines · On Monday May 8, [email protected] wrote:
NACK. As long as we have no HSM support in the tree there's no reason to add this. From the kernel's point it's just untested and unused code that can break.
So your only objection is the absence of DMAPI or some equivalent? If so, perhaps the best way forward in the short term is for SUSE to add it as an out-of-tree patch.
Quoted 6 lines
Quoted 3 lines
Greg: you seemed to suggest that there was already code in XFS that could make use of this. If that is so: could you point us to it please.
It's used by SGI's out of tree dmapi implementation. Because dmapi is such an utterly braindead standard
Sure it's ugly, but to be fair it's no worse than SysV semaphores and shmem, and it does fill a real (albeit niche) need which nothing else does AFAIK. Compare to TLI, which is pretty much entirely useless.
Quoted 3 lines
I don't expect anyone to submit a kernel-based
implementation for inclusion, although support for a big enough subset of that standard could be archived by proper kernel <-> userspace cooperation.
I don't understand what you mean here: surely almost all of what DMAPI specifies is for achieving proper kernel/userspace co-operation? Do you have some ideas for a better way of doing that?
Greg.
Signature · 2 lines
Greg Banks, R&D Software Engineer, SGI Australian Software Group. I don't speak for SGI.
Quoted 3 lines · Greg Banks
So your only objection is the absence of DMAPI or some equivalent? If so, perhaps the best way forward in the short term is for SUSE to add it as an out-of-tree patch.
The objection doesn't have anything to do with DMAPI per see. It has to do with adding kernel code that's never used in tree and thus bloats the
kernel. Also it has a fair chance to bitrot as there's no way it can be
tested.
p.s. -fsdevel couldn't care less about what patches you send to SuSE, please stop posting such things here and talk to your SuSE technical contacts.
Quoted 3 lines
Sure it's ugly, but to be fair it's no worse than SysV semaphores and shmem, and it does fill a real (albeit niche) need which nothing else does AFAIK. Compare to TLI, which is pretty much entirely useless.
It's fat worse because it makes assumptions that simply aren't true in a unix or linux enviroment. SysV IPC is ugly but doesn't do this.
Quoted 6 lines
Quoted 3 lines
I don't expect anyone to submit a kernel-based
implementation for inclusion, although support for a big enough subset of that standard could be archived by proper kernel <-> userspace cooperation.I don't understand what you mean here: surely almost all of what DMAPI specifies is for achieving proper kernel/userspace co-operation?
no. the DMAPI spec doesn't even know about the user <-> kernel boundary nor should it.
Quoted 1 line
Do you have some ideas for a better way of doing that?
Yes, and I've written them up quite a few times, please look at the fsdevel linux-kernel archives and possible some SGI-internal lists from years ago.