| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Overhaul the third party-induced invalidation handling, making use of the
previously added volume-level event counters (cb_scrub and cb_ro_snapshot)
that are now being parsed out of the VolSync record returned by the
fileserver in many of its replies.
This allows better handling of RO (and Backup) volumes. Since these are
snapshot of a RW volume that are updated atomically simultantanously across
all servers that host them, they only require a single callback promise for
the entire volume. The currently upstream code assumes that RO volumes
operate in the same manner as RW volumes, and that each file has its own
individual callback - which means that it does a status fetch for *every*
file in a RO volume, whether or not the volume got "released" (volume
callback breaks can occur for other reasons too, such as the volumeserver
taking ownership of a volume from a fileserver).
To this end, make the following changes:
(1) Change the meaning of the volume's cb_v_break counter so that it is
now a hint that we need to issue a status fetch to work out the state
of a volume. cb_v_break is incremented by volume break callbacks and
by server initialisation callbacks.
(2) Add a second counter, cb_v_check, to the afs_volume struct such that
if this differs from cb_v_break, we need to do a check. When the
check is complete, cb_v_check is advanced to what cb_v_break was at
the start of the status fetch.
(3) Move the list of mmap'd vnodes to the volume and trigger removal of
PTEs that map to files on a volume break rather than on a server
break.
(4) When a server reinitialisation callback comes in, use the
server-to-volume reverse mapping added in a preceding patch to iterate
over all the volumes using that server and clear the volume callback
promises for that server and the general volume promise as a whole to
trigger reanalysis.
(5) Replace the AFS_VNODE_CB_PROMISED flag with an AFS_NO_CB_PROMISE
(TIME64_MIN) value in the cb_expires_at field, reducing the number of
checks we need to make.
(6) Change afs_check_validity() to quickly see if various event counters
have been incremented or if the vnode or volume callback promise is
due to expire/has expired without making any changes to the state.
That is now left to afs_validate() as this may get more complicated in
future as we may have to examine server records too.
(7) Overhaul afs_validate() so that it does a single status fetch if we
need to check the state of either the vnode or the volume - and do so
under appropriate locking. The function does the following steps:
(A) If the vnode/volume is no longer seen as valid, then we take the
vnode validation lock and, if the volume promise has expired, the
volume check lock also. The latter prevents redundant checks being
made to find out if a new version of the volume got released.
(B) If a previous RPC call found that the volsync changed unexpectedly
or that a RO volume was updated, then we unmap all PTEs pointing to
the file to stop mmap being used for access.
(C) If the vnode is still seen to be of uncertain validity, then we
perform an FS.FetchStatus RPC op to jointly update the volume status
and the vnode status. This assessment is done as part of parsing the
reply:
If the RO volume creation timestamp advances, cb_ro_snapshot is
incremented; if either the creation or update timestamps changes in
an unexpected way, the cb_scrub counter is incremented
If the Data Version returned doesn't match the copy we have
locally, then we ask for the pagecache to be zapped. This takes
care of handling RO update.
(D) If cb_scrub differs between volume and vnode, the vnode's
pagecache is zapped and the vnode's cb_scrub is updated unless the
file is marked as having been deleted.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
|
|
A number of fileserver RPC operations return a VolSync record as part of
their reply that gives some information about the state of the volume being
accessed, including:
(1) A volume Creation timestamp. For an RW volume, this is the time at
which the volume was created; if it changes, the RW volume was
presumably restored from a backup and all cached data should be
scrubbed as Data Version numbers could regress on the files in the
volume.
For an RO volume, this is the time it was last snapshotted from the RW
volume. It is expected to advance each time this happens; if it
regresses, cached data should be scrubbed.
(2) A volume Update timestamp (Auristor only). For an RW volume, this is
updated any time any change is made to a volume or its contents. If
it regresses, all cached data must be scrubbed.
For an RO volume, this is a copy of the RW volume's Update timestamp
at the point of snapshotting. It can be used as a version number when
checking to see if a callback on a RO volume was due to a snapshot.
If it regresses, all cached data must be scrubbed.
but this is currently not made use of by the in-kernel afs filesystem.
Make the afs filesystem use this by:
(1) Add an update time field to the afs_volsync struct and use a value of
TIME64_MIN in both that and the creation time to indicate that they
are unset.
(2) Add creation and update time fields to the afs_volume struct and use
this to track the two timestamps.
(3) Add a volsync_lock mutex to the afs_volume struct to control
modification access for when we detect a change in these values.
(3) Add a 'pre-op volsync' struct to the afs_operation struct to record
the state of the volume tracking before the op.
(4) Add a new counter, cb_scrub, to the afs_volume struct to count events
that require all data to be scrubbed. A copy is placed in the
afs_vnode struct (inode) and if they no longer match, a scrub takes
place.
(5) When the result of an operation is being parsed, parse the VolSync
data too, if it is provided. Note that the two timestamps are handled
separately, since they don't work in quite the same way.
- If the afs_volume tracking is unset, just set it and do nothing
else.
- If the result timestamps are the same as the ones in afs_volume, do
nothing.
- If the timestamps regress, increment cb_scrub if not already done
so.
- If the creation timestamp on a RW volume changes, increment cb_scrub
if not already done so.
- If the creation timestamp on a RO volume advances, update the server
list and see if the current server has been excluded, if so reissue
the op. Once over half of the replication sites have been updated,
increment cb_ro_snapshot to indicate updates may be required and
switch over to excluding unupdated replication sites.
- If the creation timestamp on a Backup volume advances, just
increment cb_ro_snapshot to trigger updates.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
|
|
Apply server breaks to mmap'd files that are being used from that server
from the call processor work function rather than punting it off to a
workqueue. The work item, afs_server_init_callback(), then bumps each
individual inode off to its own work item introducing a potentially lengthy
delay. This reduces that delay at the cost of extending the amount of time
we delay replying to the CB.InitCallBack3 notification RPC from the server.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
|
|
David Howells says:
(2) afs_lookup_volume_rcu().
There can be a lot of volumes known by a system. A thousand would
require a 10-step walk and this is drivable by remote operation, so I
think this should probably take a lock on the second pass too.
Make the "seq" counter odd on the 2nd pass, otherwise read_seqbegin_or_lock()
never takes the lock.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Link: https://lore.kernel.org/r/20231130115606.GA21571@redhat.com/
|
|
While randstruct was satisfied with using an open-coded "void *" offset
cast for the netfs_i_context <-> inode casting, __builtin_object_size() as
used by FORTIFY_SOURCE was not as easily fooled. This was causing the
following complaint[1] from gcc v12:
In file included from include/linux/string.h:253,
from include/linux/ceph/ceph_debug.h:7,
from fs/ceph/inode.c:2:
In function 'fortify_memset_chk',
inlined from 'netfs_i_context_init' at include/linux/netfs.h:326:2,
inlined from 'ceph_alloc_inode' at fs/ceph/inode.c:463:2:
include/linux/fortify-string.h:242:25: warning: call to '__write_overflow_field' declared with attribute warning: detected write beyond size of field (1st parameter); maybe use struct_group()? [-Wattribute-warning]
242 | __write_overflow_field(p_size_field, size);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Fix this by embedding a struct inode into struct netfs_i_context (which
should perhaps be renamed to struct netfs_inode). The struct inode
vfs_inode fields are then removed from the 9p, afs, ceph and cifs inode
structs and vfs_inode is then simply changed to "netfs.inode" in those
filesystems.
Further, rename netfs_i_context to netfs_inode, get rid of the
netfs_inode() function that converted a netfs_i_context pointer to an
inode pointer (that can now be done with &ctx->inode) and rename the
netfs_i_context() function to netfs_inode() (which is now a wrapper
around container_of()).
Most of the changes were done with:
perl -p -i -e 's/vfs_inode/netfs.inode/'g \
`git grep -l 'vfs_inode' -- fs/{9p,afs,ceph,cifs}/*.[ch]`
Kees suggested doing it with a pair structure[2] and a special
declarator to insert that into the network filesystem's inode
wrapper[3], but I think it's cleaner to embed it - and then it doesn't
matter if struct randomisation reorders things.
Dave Chinner suggested using a filesystem-specific VFS_I() function in
each filesystem to convert that filesystem's own inode wrapper struct
into the VFS inode struct[4].
Version #2:
- Fix a couple of missed name changes due to a disabled cifs option.
- Rename nfs_i_context to nfs_inode
- Use "netfs" instead of "nic" as the member name in per-fs inode wrapper
structs.
[ This also undoes commit 507160f46c55 ("netfs: gcc-12: temporarily
disable '-Wattribute-warning' for now") that is no longer needed ]
Fixes: bc899ee1c898 ("netfs: Add a netfs inode context")
Reported-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: David Howells <dhowells@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
cc: Jonathan Corbet <corbet@lwn.net>
cc: Eric Van Hensbergen <ericvh@gmail.com>
cc: Latchesar Ionkov <lucho@ionkov.net>
cc: Dominique Martinet <asmadeus@codewreck.org>
cc: Christian Schoenebeck <linux_oss@crudebyte.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: Ilya Dryomov <idryomov@gmail.com>
cc: Steve French <smfrench@gmail.com>
cc: William Kucharski <william.kucharski@oracle.com>
cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
cc: Dave Chinner <david@fromorbit.com>
cc: linux-doc@vger.kernel.org
cc: v9fs-developer@lists.sourceforge.net
cc: linux-afs@lists.infradead.org
cc: ceph-devel@vger.kernel.org
cc: linux-cifs@vger.kernel.org
cc: samba-technical@lists.samba.org
cc: linux-fsdevel@vger.kernel.org
cc: linux-hardening@vger.kernel.org
Link: https://lore.kernel.org/r/d2ad3a3d7bdd794c6efb562d2f2b655fb67756b9.camel@kernel.org/ [1]
Link: https://lore.kernel.org/r/20220517210230.864239-1-keescook@chromium.org/ [2]
Link: https://lore.kernel.org/r/20220518202212.2322058-1-keescook@chromium.org/ [3]
Link: https://lore.kernel.org/r/20220524101205.GI2306852@dread.disaster.area/ [4]
Link: https://lore.kernel.org/r/165296786831.3591209.12111293034669289733.stgit@warthog.procyon.org.uk/ # v1
Link: https://lore.kernel.org/r/165305805651.4094995.7763502506786714216.stgit@warthog.procyon.org.uk # v2
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Try to avoid taking the RCU read lock when checking the validity of a
vnode's callback state. The only thing it's needed for is to pin the
parent volume's server list whilst we search it to find the record of the
server we're currently using to see if it has been reinitialised (ie. it
sent us a CB.InitCallBackState* RPC).
Do this by the following means:
(1) Keep an additional per-cell counter (fs_s_break) that's incremented
each time any of the fileservers in the cell reinitialises.
Since the new counter can be accessed without RCU from the vnode, we
can check that first - and only if it differs, get the RCU read lock
and check the volume's server list.
(2) Replace afs_get_s_break_rcu() with afs_check_server_good() which now
indicates whether the callback promise is still expected to be present
on the server. This does the checks as described in (1).
(3) Restructure afs_check_validity() to take account of the change in (2).
We can also get rid of the valid variable and just use the need_clear
variable with the addition of the afs_cb_break_no_promise reason.
(4) afs_check_validity() probably shouldn't be altering vnode->cb_v_break
and vnode->cb_s_break when it doesn't have cb_lock exclusively locked.
Move the change to vnode->cb_v_break to __afs_break_callback().
Delegate the change to vnode->cb_s_break to afs_select_fileserver()
and set vnode->cb_fs_s_break there also.
(5) afs_validate() no longer needs to get the RCU read lock around its
call to afs_check_validity() - and can skip the call entirely if we
don't have a promise.
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Markus Suvanto <markus.suvanto@gmail.com>
cc: linux-afs@lists.infradead.org
Link: https://lore.kernel.org/r/163111669583.283156.1397603105683094563.stgit@warthog.procyon.org.uk/
|
|
Fix the coherency management of mmap'd data such that 3rd-party changes
become visible as soon as possible after the callback notification is
delivered by the fileserver. This is done by the following means:
(1) When we break a callback on a vnode specified by the CB.CallBack call
from the server, we queue a work item (vnode->cb_work) to go and
clobber all the PTEs mapping to that inode.
This causes the CPU to trip through the ->map_pages() and
->page_mkwrite() handlers if userspace attempts to access the page(s)
again.
(Ideally, this would be done in the service handler for CB.CallBack,
but the server is waiting for our reply before considering, and we
have a list of vnodes, all of which need breaking - and the process of
getting the mmap_lock and stripping the PTEs on all CPUs could be
quite slow.)
(2) Call afs_validate() from the ->map_pages() handler to check to see if
the file has changed and to get a new callback promise from the
server.
Also handle the fileserver telling us that it's dropping all callbacks,
possibly after it's been restarted by sending us a CB.InitCallBackState*
call by the following means:
(3) Maintain a per-cell list of afs files that are currently mmap'd
(cell->fs_open_mmaps).
(4) Add a work item to each server that is invoked if there are any open
mmaps when CB.InitCallBackState happens. This work item goes through
the aforementioned list and invokes the vnode->cb_work work item for
each one that is currently using this server.
This causes the PTEs to be cleared, causing ->map_pages() or
->page_mkwrite() to be called again, thereby calling afs_validate()
again.
I've chosen to simply strip the PTEs at the point of notification reception
rather than invalidate all the pages as well because (a) it's faster, (b)
we may get a notification for other reasons than the data being altered (in
which case we don't want to clobber the pagecache) and (c) we need to ask
the server to find out - and I don't want to wait for the reply before
holding up userspace.
This was tested using the attached test program:
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
int main(int argc, char *argv[])
{
size_t size = getpagesize();
unsigned char *p;
bool mod = (argc == 3);
int fd;
if (argc != 2 && argc != 3) {
fprintf(stderr, "Format: %s <file> [mod]\n", argv[0]);
exit(2);
}
fd = open(argv[1], mod ? O_RDWR : O_RDONLY);
if (fd < 0) {
perror(argv[1]);
exit(1);
}
p = mmap(NULL, size, mod ? PROT_READ|PROT_WRITE : PROT_READ,
MAP_SHARED, fd, 0);
if (p == MAP_FAILED) {
perror("mmap");
exit(1);
}
for (;;) {
if (mod) {
p[0]++;
msync(p, size, MS_ASYNC);
fsync(fd);
}
printf("%02x", p[0]);
fflush(stdout);
sleep(1);
}
}
It runs in two modes: in one mode, it mmaps a file, then sits in a loop
reading the first byte, printing it and sleeping for a second; in the
second mode it mmaps a file, then sits in a loop incrementing the first
byte and flushing, then printing and sleeping.
Two instances of this program can be run on different machines, one doing
the reading and one doing the writing. The reader should see the changes
made by the writer, but without this patch, they aren't because validity
checking is being done lazily - only on entry to the filesystem.
Testing the InitCallBackState change is more complicated. The server has
to be taken offline, the saved callback state file removed and then the
server restarted whilst the reading-mode program continues to run. The
client machine then has to poke the server to trigger the InitCallBackState
call.
Signed-off-by: David Howells <dhowells@redhat.com>
Tested-by: Markus Suvanto <markus.suvanto@gmail.com>
cc: linux-afs@lists.infradead.org
Link: https://lore.kernel.org/r/163111668833.283156.382633263709075739.stgit@warthog.procyon.org.uk/
|
|
Whilst it shouldn't happen, it is possible for multiple fileservers to
share a UUID, particularly if an entire cell has been duplicated, UUIDs and
all. In such a case, it's not necessarily possible to map the effect of
the CB.InitCallBackState3 incoming RPC to a specific server unambiguously
by UUID and thus to a specific cell.
Indeed, there's a problem whereby multiple server records may need to
occupy the same spot in the rb_tree rooted in the afs_net struct.
Fix this by allowing servers to form a list, with the head of the list in
the tree. When the front entry in the list is removed, the second in the
list just replaces it. afs_init_callback_state() then just goes down the
line, poking each server in the list.
This means that some servers will be unnecessarily poked, unfortunately.
An alternative would be to route by call parameters.
Reported-by: Jeffrey Altman <jaltman@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Fixes: d2ddc776a458 ("afs: Overhaul volume and server record caching and fileserver rotation")
|
|
Reorganise afs_volume objects such that they're in a tree keyed on volume
ID, rooted at on an afs_cell object rather than being in multiple trees,
each of which is rooted on an afs_server object.
afs_server structs become per-cell and acquire a pointer to the cell.
The process of breaking a callback then starts with finding the server by
its network address, following that to the cell and then looking up each
volume ID in the volume tree.
This is simpler than the afs_vol_interest/afs_cb_interest N:M mapping web
and allows those structs and the code for maintaining them to be simplified
or removed.
It does make a couple of things a bit more tricky, though:
(1) Operations now start with a volume, not a server, so there can be more
than one answer as to whether or not the server we'll end up using
supports the FS.InlineBulkStatus RPC.
(2) CB RPC operations that specify the server UUID. There's still a tree
of servers by UUID on the afs_net struct, but the UUIDs in it aren't
guaranteed unique.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Turn the afs_operation struct into the main way that most fileserver
operations are managed. Various things are added to the struct, including
the following:
(1) All the parameters and results of the relevant operations are moved
into it, removing corresponding fields from the afs_call struct.
afs_call gets a pointer to the op.
(2) The target volume is made the main focus of the operation, rather than
the target vnode(s), and a bunch of op->vnode->volume are made
op->volume instead.
(3) Two vnode records are defined (op->file[]) for the vnode(s) involved
in most operations. The vnode record (struct afs_vnode_param)
contains:
- The vnode pointer.
- The fid of the vnode to be included in the parameters or that was
returned in the reply (eg. FS.MakeDir).
- The status and callback information that may be returned in the
reply about the vnode.
- Callback break and data version tracking for detecting
simultaneous third-parth changes.
(4) Pointers to dentries to be updated with new inodes.
(5) An operations table pointer. The table includes pointers to functions
for issuing AFS and YFS-variant RPCs, handling the success and abort
of an operation and handling post-I/O-lock local editing of a
directory.
To make this work, the following function restructuring is made:
(A) The rotation loop that issues calls to fileservers that can be found
in each function that wants to issue an RPC (such as afs_mkdir()) is
extracted out into common code, in a new file called fs_operation.c.
(B) The rotation loops, such as the one in afs_mkdir(), are replaced with
a much smaller piece of code that allocates an operation, sets the
parameters and then calls out to the common code to do the actual
work.
(C) The code for handling the success and failure of an operation are
moved into operation functions (as (5) above) and these are called
from the core code at appropriate times.
(D) The pseudo inode getting stuff used by the dynamic root code is moved
over into dynroot.c.
(E) struct afs_iget_data is absorbed into the operation struct and
afs_iget() expects to be given an op pointer and a vnode record.
(F) Point (E) doesn't work for the root dir of a volume, but we know the
FID in advance (it's always vnode 1, unique 1), so a separate inode
getter, afs_root_iget(), is provided to special-case that.
(G) The inode status init/update functions now also take an op and a vnode
record.
(H) The RPC marshalling functions now, for the most part, just take an
afs_operation struct as their only argument. All the data they need
is held there. The result delivery functions write their answers
there as well.
(I) The call is attached to the operation and then the operation core does
the waiting.
And then the new operation code is, for the moment, made to just initialise
the operation, get the appropriate vnode I/O locks and do the same rotation
loop as before.
This lays the foundation for the following changes in the future:
(*) Overhauling the rotation (again).
(*) Support for asynchronous I/O, where the fileserver rotation must be
done asynchronously also.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
afs_vol_interest objects represent the volume IDs currently being accessed
from a fileserver. These hold lists of afs_cb_interest objects that
repesent the superblocks using that volume ID on that server.
When a callback notification from the server telling of a modification by
another client arrives, the volume ID specified in the notification is
looked up in the server's afs_vol_interest list. Through the
afs_cb_interest list, the relevant superblocks can be iterated over and the
specific inode looked up and marked in each one.
Make the following efficiency improvements:
(1) Hold rcu_read_lock() over the entire processing rather than locking it
each time.
(2) Do all the callbacks for each vid together rather than individually.
Each volume then only needs to be looked up once.
(3) afs_vol_interest objects are now stored in an rb_tree rather than a
flat list to reduce the lookup step count.
(4) afs_vol_interest lookup is now done with RCU, but because it's in an
rb_tree which may rotate under us, a seqlock is used so that if it
changes during the walk, we repeat the walk with a lock held.
With this and the preceding patch which adds RCU-based lookups in the inode
cache, target volumes/vnodes can be taken without the need to take any
locks, except on the target itself.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Make the inode hash table RCU searchable so that searches that want to
access or modify an inode without taking a ref on that inode can do so
without taking the inode hash table lock.
The main thing this requires is some RCU annotation on the list
manipulation operations. Inodes are already freed by RCU in most cases.
Users of this interface must take care as the inode may be still under
construction or may be being torn down around them.
There are at least three instances where this can be of use:
(1) Testing whether the inode number iunique() is going to return is
currently unique (the iunique_lock is still held).
(2) Ext4 date stamp updating.
(3) AFS callback breaking.
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
cc: linux-ext4@vger.kernel.org
cc: linux-afs@lists.infradead.org
|
|
Servers sending callback breaks to the YFS_CM_SERVICE service may
send up to YFSCBMAX (1024) fids in a single RPC. Anything over
AFSCBMAX (50) will cause the assert in afs_break_callbacks to trigger.
Remove the assert, as the count has already been checked against
the appropriate max values in afs_deliver_cb_callback and
afs_deliver_yfs_cb_callback.
Fixes: 35dbfba3111a ("afs: Implement the YFS cache manager service")
Signed-off-by: Marc Dionne <marc.dionne@auristor.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Add a tracepoint (afs_server) to track the afs_server object usage count.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Add a couple of tracepoints to track callback management:
(1) afs_cb_miss - Logs when we were unable to apply a callback, either due
to the inode being discarded or due to a competing thread applying a
callback first.
(2) afs_cb_break - Logs when we attempted to clear the noted callback
promise, either due to the server explicitly breaking the callback,
the callback promise lapsing or a local event obsoleting it.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Fix the cb_break_lock spinlock in afs_volume struct by initialising it when
the volume record is allocated.
Also rename the lock to cb_v_break_lock to distinguish it from the lock of
the same name in the afs_server struct.
Without this, the following trace may be observed when a volume-break
callback is received:
INFO: trying to register non-static key.
the code is fine but needs lockdep annotation.
turning off the locking correctness validator.
CPU: 2 PID: 50 Comm: kworker/2:1 Not tainted 5.2.0-rc1-fscache+ #3045
Hardware name: ASUS All Series/H97-PLUS, BIOS 2306 10/09/2014
Workqueue: afs SRXAFSCB_CallBack
Call Trace:
dump_stack+0x67/0x8e
register_lock_class+0x23b/0x421
? check_usage_forwards+0x13c/0x13c
__lock_acquire+0x89/0xf73
lock_acquire+0x13b/0x166
? afs_break_callbacks+0x1b2/0x3dd
_raw_write_lock+0x2c/0x36
? afs_break_callbacks+0x1b2/0x3dd
afs_break_callbacks+0x1b2/0x3dd
? trace_event_raw_event_afs_server+0x61/0xac
SRXAFSCB_CallBack+0x11f/0x16c
process_one_work+0x2c5/0x4ee
? worker_thread+0x234/0x2ac
worker_thread+0x1d8/0x2ac
? cancel_delayed_work_sync+0xf/0xf
kthread+0x11f/0x127
? kthread_park+0x76/0x76
ret_from_fork+0x24/0x30
Fixes: 68251f0a6818 ("afs: Fix whole-volume callback handling")
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Use RCU-based freeing for afs_cb_interest struct objects and use RCU on
vnode->cb_interest. Use that change to allow afs_check_validity() to use
read_seqbegin_or_lock() instead of read_seqlock_excl().
This also requires the caller of afs_check_validity() to hold the RCU read
lock across the call.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
__afs_break_callback() holds vnode->lock around its call of
afs_lock_may_be_available() - which also takes that lock.
Fix this by not taking the lock in __afs_break_callback().
Also, there's no point checking the granted_locks and pending_locks queues;
it's sufficient to check lock_state, so move that check out of
afs_lock_may_be_available() into __afs_break_callback() to replace the
queue checks.
Fixes: e8d6c554126b ("AFS: implement file locking")
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
The in-kernel afs filesystem client counts the number of server-level
callback invalidation events (CB.InitCallBackState* RPC operations) that it
receives from the server. This is stored in cb_s_break in various
structures, including afs_server and afs_vnode.
If an inode is examined by afs_validate(), say, the afs_server copy is
compared, along with other break counters, to those in afs_vnode, and if
one or more of the counters do not match, it is considered that the
server's callback promise is broken. At points where this happens,
AFS_VNODE_CB_PROMISED is cleared to indicate that the status must be
refetched from the server.
afs_validate() issues an FS.FetchStatus operation to get updated metadata -
and based on the updated data_version may invalidate the pagecache too.
However, the break counters are also used to determine whether to note a
new callback in the vnode (which would set the AFS_VNODE_CB_PROMISED flag)
and whether to cache the permit data included in the YFSFetchStatus record
by the server.
The problem comes when the server sends us a CB.InitCallBackState op. The
first such instance doesn't cause cb_s_break to be incremented, but rather
causes AFS_SERVER_FL_NEW to be cleared - but thereafter, say some hours
after last use and all the volumes have been automatically unmounted and
the server has forgotten about the client[*], this *will* likely cause an
increment.
[*] There are other circumstances too, such as the server restarting or
needing to make space in its callback table.
Note that the server won't send us a CB.InitCallBackState op until we talk
to it again.
So what happens is:
(1) A mount for a new volume is attempted, a inode is created for the root
vnode and vnode->cb_s_break and AFS_VNODE_CB_PROMISED aren't set
immediately, as we don't have a nominated server to talk to yet - and
we may iterate through a few to find one.
(2) Before the operation happens, afs_fetch_status(), say, notes in the
cursor (fc.cb_break) the break counter sum from the vnode, volume and
server counters, but the server->cb_s_break is currently 0.
(3) We send FS.FetchStatus to the server. The server sends us back
CB.InitCallBackState. We increment server->cb_s_break.
(4) Our FS.FetchStatus completes. The reply includes a callback record.
(5) xdr_decode_AFSCallBack()/xdr_decode_YFSCallBack() check to see whether
the callback promise was broken by checking the break counter sum from
step (2) against the current sum.
This fails because of step (3), so we don't set the callback record
and, importantly, don't set AFS_VNODE_CB_PROMISED on the vnode.
This does not preclude the syscall from progressing, and we don't loop here
rechecking the status, but rather assume it's good enough for one round
only and will need to be rechecked next time.
(6) afs_validate() it triggered on the vnode, probably called from
d_revalidate() checking the parent directory.
(7) afs_validate() notes that AFS_VNODE_CB_PROMISED isn't set, so doesn't
update vnode->cb_s_break and assumes the vnode to be invalid.
(8) afs_validate() needs to calls afs_fetch_status(). Go back to step (2)
and repeat, every time the vnode is validated.
This primarily affects volume root dir vnodes. Everything subsequent to
those inherit an already incremented cb_s_break upon mounting.
The issue is that we assume that the callback record and the cached permit
information in a reply from the server can't be trusted after getting a
server break - but this is wrong since the server makes sure things are
done in the right order, holding up our ops if necessary[*].
[*] There is an extremely unlikely scenario where a reply from before the
CB.InitCallBackState could get its delivery deferred till after - at
which point we think we have a promise when we don't. This, however,
requires unlucky mass packet loss to one call.
AFS_SERVER_FL_NEW tries to paper over the cracks for the initial mount from
a server we've never contacted before, but this should be unnecessary.
It's also further insulated from the problem on an initial mount by
querying the server first with FS.GetCapabilities, which triggers the
CB.InitCallBackState.
Fix this by
(1) Remove AFS_SERVER_FL_NEW.
(2) In afs_calc_vnode_cb_break(), don't include cb_s_break in the
calculation.
(3) In afs_cb_is_broken(), don't include cb_s_break in the check.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Implement support for talking to YFS-variant fileservers in the cache
manager and the filesystem client. These implement upgraded services on
the same port as their AFS services.
YFS fileservers provide expanded capabilities over AFS.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Remove unnecessary details of a broken callback, such as version, expiry
and type, from the afs_callback_break struct as they're not actually used
and make the list take more memory.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Increase the sizes of the volume ID to 64 bits and the vnode ID (inode
number equivalent) to 96 bits to allow the support of YFS.
This requires the iget comparator to check the vnode->fid rather than i_ino
and i_generation as i_ino is not sufficiently capacious. It also requires
this data to be placed into the vnode cache key for fscache.
For the moment, just discard the top 32 bits of the vnode ID when returning
it though stat.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
At the moment, afs_break_callbacks calls afs_break_one_callback() for each
separate FID it was given, and the latter looks up the volume individually
for each one.
However, this is inefficient if two or more FIDs have the same vid as we
could reuse the volume. This is complicated by cell aliasing whereby we
may have multiple cells sharing a volume and can therefore have multiple
callback interests for any particular volume ID.
At the moment afs_break_one_callback() scans the entire list of volumes
we're getting from a server and breaks the appropriate callback in every
matching volume, regardless of cell. This scan is done for every FID.
Optimise callback breaking by the following means:
(1) Sort the FID list by vid so that all FIDs belonging to the same volume
are clumped together.
This is done through the use of an indirection table as we cannot do
an insertion sort on the afs_callback_break array as we decode FIDs
into it as we subsequently also have to decode callback info into it
that corresponds by array index only.
We also don't really want to bubblesort afterwards if we can avoid it.
(2) Sort the server->cb_interests array by vid so that all the matching
volumes are grouped together. This permits the scan to stop after
finding a record that has a higher vid.
(3) When breaking FIDs, we try to keep server->cb_break_lock as long as
possible, caching the start point in the array for that volume group
as long as possible.
It might make sense to add another layer in that list and have a
refcounted volume ID anchor that has the matching interests attached
to it rather than being in the list. This would allow the lock to be
dropped without losing the cursor.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
It's possible for an AFS file server to issue a whole-volume notification
that callbacks on all the vnodes in the file have been broken. This is
done for R/O and backup volumes (which don't have per-file callbacks) and
for things like a volume being taken offline.
Fix callback handling to detect whole-volume notifications, to track it
across operations and to check it during inode validation.
Fixes: c435ee34551e ("afs: Overhaul the callback handling")
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
The refcounting on afs_cb_interest struct objects in
afs_register_server_cb_interest() is wrong as it uses the server list
entry's call back interest pointer without regard for the fact that it
might be replaced at any time and the object thrown away.
Fix this by:
(1) Put a lock on the afs_server_list struct that can be used to
mediate access to the callback interest pointers in the servers array.
(2) Keep a ref on the callback interest that we get from the entry.
(3) Dropping the old reference held by vnode->cb_interest if we replace
the pointer.
Fixes: c435ee34551e ("afs: Overhaul the callback handling")
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Processes like ld that do lots of small writes that aren't necessarily
contiguous result in a lot of small StoreData operations to the server, the
idea being that if someone else changes the data on the server, we only
write our changes over that and not the space between. Further, we don't
want to write back empty space if we can avoid it to make it easier for the
server to do sparse files.
However, making lots of tiny RPC ops is a lot less efficient for the server
than one big one because each op requires allocation of resources and the
taking of locks, so we want to compromise a bit.
Reduce the load by the following:
(1) If a file is just created locally or has just been truncated with
O_TRUNC locally, allow subsequent writes to the file to be merged with
intervening space if that space doesn't cross an entire intervening
page.
(2) Don't flush the file on ->flush() but rather on ->release() if the
file was open for writing.
Just linking vmlinux.o, without this patch, looking in /proc/fs/afs/stats:
file-wr : n=441 nb=513581204
and after the patch:
file-wr : n=62 nb=513668555
there were 379 fewer StoreData RPC operations at the expense of an extra
87K being written.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
When afs_lookup() is called, prospectively look up the next 50 uncached
fids also from that same directory and cache the results, rather than just
looking up the one file requested.
This allows us to use the FS.InlineBulkStatus RPC op to increase efficiency
by fetching up to 50 file statuses at a time.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Fix warnings raised by checker, including:
(*) Warnings raised by unequal comparison for the purposes of sorting,
where the endianness doesn't matter:
fs/afs/addr_list.c:246:21: warning: restricted __be16 degrades to integer
fs/afs/addr_list.c:246:30: warning: restricted __be16 degrades to integer
fs/afs/addr_list.c:248:21: warning: restricted __be32 degrades to integer
fs/afs/addr_list.c:248:49: warning: restricted __be32 degrades to integer
fs/afs/addr_list.c:283:21: warning: restricted __be16 degrades to integer
fs/afs/addr_list.c:283:30: warning: restricted __be16 degrades to integer
(*) afs_set_cb_interest() is not actually used and can be removed.
(*) afs_cell_gc_delay() should be provided with a sysctl.
(*) afs_cell_destroy() needs to use rcu_access_pointer() to read
cell->vl_addrs.
(*) afs_init_fs_cursor() should be static.
(*) struct afs_vnode::permit_cache needs to be marked __rcu.
(*) afs_server_rcu() needs to use rcu_access_pointer().
(*) afs_destroy_server() should use rcu_access_pointer() on
server->addresses as the server object is no longer accessible.
(*) afs_find_server() casts __be16/__be32 values to int in order to
directly compare them for the purpose of finding a match in a list,
but is should also annotate the cast with __force to avoid checker
warnings.
(*) afs_check_permit() accesses vnode->permit_cache outside of the RCU
readlock, though it doesn't then access the value; the extraneous
access is deleted.
False positives:
(*) Conditional locking around the code in xdr_decode_AFSFetchStatus. This
can be dealt with in a separate patch.
fs/afs/fsclient.c:148:9: warning: context imbalance in 'xdr_decode_AFSFetchStatus' - different lock contexts for basic block
(*) Incorrect handling of seq-retry lock context balance:
fs/afs/inode.c:455:38: warning: context imbalance in 'afs_getattr' - different
lock contexts for basic block
fs/afs/server.c:52:17: warning: context imbalance in 'afs_find_server' - different lock contexts for basic block
fs/afs/server.c:128:17: warning: context imbalance in 'afs_find_server_by_uuid' - different lock contexts for basic block
Errors:
(*) afs_lookup_cell_rcu() needs to break out of the seq-retry loop, not go
round again if it successfully found the workstation cell.
(*) Fix UUID decode in afs_deliver_cb_probe_uuid().
(*) afs_cache_permit() has a missing rcu_read_unlock() before one of the
jumps to the someone_else_changed_it label. Move the unlock to after
the label.
(*) afs_vl_get_addrs_u() is using ntohl() rather than htonl() when
encoding to XDR.
(*) afs_deliver_yfsvl_get_endpoints() is using htonl() rather than ntohl()
when decoding from XDR.
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
The current code assumes that volumes and servers are per-cell and are
never shared, but this is not enforced, and, indeed, public cells do exist
that are aliases of each other. Further, an organisation can, say, set up
a public cell and a private cell with overlapping, but not identical, sets
of servers. The difference is purely in the database attached to the VL
servers.
The current code will malfunction if it sees a server in two cells as it
assumes global address -> server record mappings and that each server is in
just one cell.
Further, each server may have multiple addresses - and may have addresses
of different families (IPv4 and IPv6, say).
To this end, the following structural changes are made:
(1) Server record management is overhauled:
(a) Server records are made independent of cell. The namespace keeps
track of them, volume records have lists of them and each vnode
has a server on which its callback interest currently resides.
(b) The cell record no longer keeps a list of servers known to be in
that cell.
(c) The server records are now kept in a flat list because there's no
single address to sort on.
(d) Server records are now keyed by their UUID within the namespace.
(e) The addresses for a server are obtained with the VL.GetAddrsU
rather than with VL.GetEntryByName, using the server's UUID as a
parameter.
(f) Cached server records are garbage collected after a period of
non-use and are counted out of existence before purging is allowed
to complete. This protects the work functions against rmmod.
(g) The servers list is now in /proc/fs/afs/servers.
(2) Volume record management is overhauled:
(a) An RCU-replaceable server list is introduced. This tracks both
servers and their coresponding callback interests.
(b) The superblock is now keyed on cell record and numeric volume ID.
(c) The volume record is now tied to the superblock which mounts it,
and is activated when mounted and deactivated when unmounted.
This makes it easier to handle the cache cookie without causing a
double-use in fscache.
(d) The volume record is loaded from the VLDB using VL.GetEntryByNameU
to get the server UUID list.
(e) The volume name is updated if it is seen to have changed when the
volume is updated (the update is keyed on the volume ID).
(3) The vlocation record is got rid of and VLDB records are no longer
cached. Sufficient information is stored in the volume record, though
an update to a volume record is now no longer shared between related
volumes (volumes come in bundles of three: R/W, R/O and backup).
and the following procedural changes are made:
(1) The fileserver cursor introduced previously is now fleshed out and
used to iterate over fileservers and their addresses.
(2) Volume status is checked during iteration, and the server list is
replaced if a change is detected.
(3) Server status is checked during iteration, and the address list is
replaced if a change is detected.
(4) The abort code is saved into the address list cursor and -ECONNABORTED
returned in afs_make_call() if a remote abort happened rather than
translating the abort into an error message. This allows actions to
be taken depending on the abort code more easily.
(a) If a VMOVED abort is seen then this is handled by rechecking the
volume and restarting the iteration.
(b) If a VBUSY, VRESTARTING or VSALVAGING abort is seen then this is
handled by sleeping for a short period and retrying and/or trying
other servers that might serve that volume. A message is also
displayed once until the condition has cleared.
(c) If a VOFFLINE abort is seen, then this is handled as VBUSY for the
moment.
(d) If a VNOVOL abort is seen, the volume is rechecked in the VLDB to
see if it has been deleted; if not, the fileserver is probably
indicating that the volume couldn't be attached and needs
salvaging.
(e) If statfs() sees one of these aborts, it does not sleep, but
rather returns an e |
