We have multiple code paths that hold the following locks at the same
time:
- avc->lock for a vcache
- The page lock for a page in 'avc'
In order to avoid deadlocks, we need a consistent ordering for obtaining
these two locks. The code in afs_putpage() currently obtains avc->lock
before the page lock (Obtain*Lock is called before pvn_vplist_dirty).
The code in afs_getpages() also obtains avc->lock before the page lock,
but it does so in a loop for all requested pages (via pvn_getpages()).
On the second iteration of that loop, it obtains avc->lock, and the page
from the first iteration of the loop is still locked. Thus, it obtains a
page lock before locking avc->lock in some cases.
Since we have two code paths that obtain those two locks in a different
order, a deadlock can occur. Fixing this properly requires changing at
least one of those code paths, so the locks are taken in a consistent
order. However, doing so is complex and will be done in a separate
future commit.
For this commit, we can avoid the deadlock for RO volumes by simply
avoiding taking avc->lock in afs_putpages() at all while the pages are
locked. Normally, we lock avc->lock because pvn_vplist_dirty() will call
afs_putapage() for each dirty page (and afs_putapage() requires
avc->lock held). But for RO volumes, we will have no dirty pages
(because RO volumes cannot be written to from a client), and so
afs_putapage() will never be called.
So to avoid this deadlock issue for RO volumes, avoid taking avc->lock
across the pvn_vplist_dirty() call in afs_putpage(). We now pass a dummy
pageout callback function to pvn_vplist_dirty() instead, which should
never be called, and which panics if it ever is.
We still need to hold avc->lock a few other times during afs_putpage()
for other minor reasons, but none of these hold page locks at the same
time, so the deadlock issue is still avoided.
[mmeffie: comments, and fix missing write lock, fix lock releases]
[adeason: revised commit message]
Change-Id: Iec11101147220828f319dae4027e7ab1f08483a6
Reviewed-on: https://gerrit.openafs.org/12247
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Andrew Deason <adeason@dson.org>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Export this header to the kernel sources in the libafs_tree, since it is
needed for the kernel module build.
FIXES 134476
Change-Id: Id359c6d065c259601d14ee5c02b93647f86a0288
Reviewed-on: https://gerrit.openafs.org/12882
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Update all remaining copies of CellServDB in the tree, and make the
Red Hat packaging use it by default too.
Change-Id: I5a70a7c658ad0056cd10945bb730e84f0edfb730
Reviewed-on: https://gerrit.openafs.org/12880
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Add files for FreeBSD 10.4, 11.1, and 12.0 (12-CURRENT), for i386 and amd64.
Change-Id: I904f576914bb965a659750e6302f011acf66ba81
Reviewed-on: https://gerrit.openafs.org/12863
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Stephan Wiesand <stephan.wiesand@desy.de>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Add sysnames for i386 and amd64 10.4, 11.1, and 12.0 (12-CURRENT, at present).
Change-Id: If38ecca7b2b3e40c186b7e9321ce017b4711139c
Reviewed-on: https://gerrit.openafs.org/12862
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Stephan Wiesand <stephan.wiesand@desy.de>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
The sync-site relabels its database at the end of the first write
transaction. The new label will be equal to the time at which the
sync-site in question first received its coordinator mandate. This time
is stored by a global called ubik_epochTime. In order to make sure that
the new database label is sane, only relabel the database if
ubik_epochTime is within a specific range.
Change-Id: I2408569e5de46d387f63cbc2fab05ea1264a505c
Reviewed-on: https://gerrit.openafs.org/12640
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Andrew Deason <adeason@dson.org>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
The ubik_dbVersion global represents the sync site's database version
and it is mostly used by the remote sites for sanity checks. Currently,
this global is updated when database changes are made on the sync site
(SDISK_Commit or SDISK_SetVersion), as well as every time we vote "yes"
for the sync-site in a beacon reply. Unfortunately, ubik_dbVersion is
not updated when a copy of the sync site's database is received via
DISK_SendFile, and it won't get updated until our next "yes" vote.
During this window, the current database version will not match
ubik_dbVersion. As a result, any write transaction during this time
frame will fail on the remote site in question.
To fix this problem, do not wait for the next beacon packet to update
ubik_dbVersion when the sync site's database is received; just update
it when we get the new database. Since no write transactions are
allowed while the db is transferring, ubik_dbVersion can be safely
updated.
Change-Id: Ide7a695a69cb3229ad585d9e56c5ddc2efb76dd7
Reviewed-on: https://gerrit.openafs.org/12716
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Andrew Deason <adeason@dson.org>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Currently, check_dentry_race locks the parent inode in order to ensure
it is not running in parallel with d_splice_alias for the same inode.
(For old Linux kernel versions; see commit b0461f2d: "LINUX:
Workaround d_splice_alias/d_lookup race".)
However, it is possible to hit this area of code when the parent inode
is already locked. When someone tries to create a file, directory, or
symlink, Linux tries to lookup the dentry for the target path, to see
if it already exists. While looking up the last component of the path,
Linux locks the directory, and if it finds a dentry for the target
name, it calls d_invalidate on it while the parent directory is
locked.
For a dentry with a NULL inode, we'll then try to lock the parent
inode in check_dentry_race. But since the inode is already locked, we
will deadlock.
From a user's point of view, the hang can be reproduced by doing
something similar to:
$ mkdir dir # succeeds
$ rmdir dir
$ ls -l dir
ls: cannot access dir: No such file or directory
$ mkdir dir # hangs
To avoid this, we can just change which lock we're using to avoid
check_dentry_race/d_splice_alias from running in parallel. Instead of
locking the parent inode, introduce a new global lock (called
dentry_race_sem), and lock that in check_dentry_race and around our
d_splice_alias call. We know that those are the only two users of this
new lock, so this should avoid any such deadlocks.
This does potentially reduce performance, since all tasks that hit
check_dentry_race or d_splice_alias will take the same global lock.
However, this at least still allows us to make use of negative
dentries, and this entire code path only applies to older Linux
kernels. It could be possible to add a new lock into struct vcache
instead, but using a global lock like this commit does is much
simpler.
Change-Id: Ide0f21145c83d6fbb34c637d8a36c8cd21549940
Reviewed-on: https://gerrit.openafs.org/12868
Tested-by: Benjamin Kaduk <kaduk@mit.edu>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Commit 443dd5367e added support for a
separate debuginfo package for the kernel module. Unfortunately, the
%files directive for the kernel module debuginfo package was incorrectly
placed in the %if stanza of the build_userspace condition, so the
rpmbuild fails when attempting to build just the kernel module.
That is, when running rpmbuild with the options:
rpmbuild --define "build_userspace 0" --define "build_modules 1" ...
rpmbuild fails with:
RPM build errors:
Installed (but unpackaged) file(s) found:
/usr/lib/debug/lib/modules/.../extra/openafs/openafs.ko.debug
Fix this by moving the new %files directive out of the build_userspace
conditional.
Change-Id: I46e74b660048022a4cc4327835c6055402a34ccf
Reviewed-on: https://gerrit.openafs.org/12874
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: Benjamin Kaduk <kaduk@mit.edu>
Further refactoring of the autoconf macros. Divy up the linux kernel
checks into smaller files.
This is a non-functional change. Care has been taken preserve the
ordering of the autoconf tests. Except for whitespace, the generated
configure file has not been changed by this refactoring. This has been
verified with a 'diff -u -w -B' comparison of the generated configure
file before and after applying this commit.
Change-Id: I5ea4c9e3a0aeff1767ef561bdb8361781694ee28
Reviewed-on: https://gerrit.openafs.org/12844
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Further refactoring of the autoconf macros. Move more linux and solaris
specific checks into their own files.
This is a non-functional change. Care has been taken preserve the
ordering of the autoconf tests. Except for whitespace, the generated
configure file has not been changed by this refactoring. This has been
verified with a 'diff -u -w -B' comparison of the generated configure
file before and after applying this commit.
Change-Id: Ib3e7b1270826970c541a695230f4e3cd13cf9e3d
Reviewed-on: https://gerrit.openafs.org/12843
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
The acinclude.m4 is very large and often requires to be changed for
unrelated commits. Divy up the large acinclude.m4 into a number of
smaller files to avoid so many contentions and to make the autoconf
system easier to maintain.
This is a non-functional change. Care has been taken preserve the
ordering of the autoconf tests. Except for whitespace, the generated
configure file has not been changed by this refactoring. This has been
verified with a 'diff -u -w -B' comparison of the generated configure
file before and after applying this commit.
Change-Id: I70e7f846dea0055d00a60a47422aa73bff25c4c6
Reviewed-on: https://gerrit.openafs.org/12842
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Since the first introduction of FreeBSD support, the macros
(MUTEX_ENTER, etc.) for kernel mutex operations have included
trailing semicolons, unique among all the platforms.
This did not cause problems until the recent work on rx event
handlers, which put a MUTEX_ENTER() in the body of an 'if' clause
with no brackets, and attempted to follow it with an 'else' clause.
This results in the following (rather obtuse) compiler error:
/root/openafs/src/rx/rx.c:3666:5: error: expected expression
else
^
Which is more visible in the preprocessed source, as
if (condition)
expression;;
else
other_expression;
is clearly invalid C.
To fix the FreeBSD kernel module build, remove the unneeded semicolons.
Change-Id: I191009ad412852dcc03cd71a0982fe41a953301d
Reviewed-on: https://gerrit.openafs.org/12853
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
afs_nfsdisp.lo is not used, so we do not need a build rule for it.
Change-Id: I4ca53a4823b0ccd5bfd769867f6766bd05ea4ceb
Reviewed-on: https://gerrit.openafs.org/12802
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Our in-tree xdr.h appears to have started life as a concatenation of
rpc/types.h and rpc/xdr.h, and should include all the needed functionality.
Indeed, commit 7293ddf325 even indicates
that we expect to be using our in-tree XDR everywhere anyway, so the
system XDR is superfluous.
Note that afs/sysincludes.h (not afsincludes.h!) already includes
rx/xdr.h ifndef AFS_LINUX22_ENV.
This change should help systems running glibc 2.26 or newer, which has
stopped providing the Sun RPC headers by default.
While here remove some duplicate includes of rpc/types.h in the
AIX-specific sources.
The Solaris NFS translator bits cannot really be changed, since the system
headers are used and have tight interdependencies.
Update rxgen to not emit rpc/types.h inclusion.
[mmeffie: squash 12801 to not emit rpc/types.h from rxgen]
Change-Id: I0b195216affa06ab9e259cb0bab0c8286a1636d9
Reviewed-on: https://gerrit.openafs.org/12800
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
With recent changes to d_invalidate's semantics (it returns void in Linux 3.11,
and always returns success in RHEL 7.4), it has become increasingly clear that
d_invalidate() is not the best function for use in our best-effort
(nondisruptive) attempt to free up vcaches that is afs_ShakeLooseVCaches().
The new d_invalidate() semantics always force the invalidation of a directory
dentry, which contradicts our desire to be nondisruptive, especially when
that directory is being used as the current working directory for a process.
Our call to d_invalidate(), intended to merely probe for whether a dentry
can be discarded without affecting other consumers, instead would cause
processes using that dentry as a CWD to receive ENOENT errors from getcwd().
A previous commit (c3bbf0b444) tried to address
this issue by calling d_prune_aliases() instead of d_invalidate(), but
d_prune_aliases() does not recursively descend into children of the given
dentry while pruning, leaving it an incomplete solution for our use-case.
To address these issues, modify the shakeloose routine TryEvictDentries() to
call shrink_dcache_parent() and maybe __d_drop() for directories, and
d_prune_aliases() for non-directories, instead of d_invalidate(). (Calls to
d_prune_aliases() for directories have already been removed by reverting commit
c3bbf0b4444db88192eea4580ac9e9ca3de0d286.)
Just like d_invalidate(), shrink_dcache_parent() has been around "forever"
(since pre-git v2.6.12). Also like d_invalidate(), it "walks" the parent
dentry's subdirectories and "shrinks" (unhashes) unused dentries. But unlike
d_invalidate(), shrink_dcache_parent() will not unhash an in-use dentry, and
has never changed its signature or semantics.
d_prune_aliases() has also been available "forever", and has also never changed
its signature or semantics. The lack of recursive descent is not an issue for
non-directories, which cannot have such children.
[kaduk@mit.edu: apply review feedback to fix locking and avoid extraneous
changes, and reword commit message]
Change-Id: Icb6138ee5785e0ef82a9b85b1d2651dfd0830043
Reviewed-on: https://gerrit.openafs.org/12830
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
The two stanzas for HAVE_DCACHE_LOCK are now functionally identical;
remove the preprocessor conditionals and duplicate code.
Minor functional change is incurrred for very old (before 2.6.38) Linux
versions that have dcache_lock; we are now obtaining the d_lock as well.
This is safe because d_lock is also quite old (pre-git, 2.6.12), and it
is a spinlock that's only held for checking d_unhashed. Therefore, it
should have negligible performance impact. It cannot cause deadlocks or
violate locking order, because spinlocks can't be held across sleeps.
Change-Id: I08faf204e6bd82c4401cdf6048d12cd551dd18fc
Reviewed-on: https://gerrit.openafs.org/12792
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Reviewed-by: Andrew Deason <adeason@dson.org>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
The two stanzas for HAVE_DCACHE_LOCK are now identical;
remove the preprocessor conditionals and duplicate code.
No functional change should be incurred by this commit.
Change-Id: I15cd4631d1932dcfb920313acb82fcbe570087e8
Reviewed-on: https://gerrit.openafs.org/12791
Reviewed-by: Andrew Deason <adeason@dson.org>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Simplify some #ifdefs for HAVE_DCACHE_LOCK by pushing them down into
new helpers in osi_compat.h.
No functional change should be incurred by this commit.
Change-Id: Ia0dc560bc84c8db4b84ddcc77a17bab5fbf93af9
Reviewed-on: https://gerrit.openafs.org/12790
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
For dentry operations that cover multiple dentry aliases of
a single inode, create a compatibility wrapper to hide differences
between the older dget_locked() and the current dget().
No functional change should be incurred by this commit.
Change-Id: I2bb0d453417f37707018f6ba5859903c3d34c8ff
Reviewed-on: https://gerrit.openafs.org/12789
Reviewed-by: Andrew Deason <adeason@dson.org>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Linux recently changed the semantics of d_invalidate() to:
- return void
- invalidate even a current working directory
OpenAFS commit c3bbf0b444 switched libafs
to use d_prune_aliases() instead.
However, since that commit, several things have happened:
- RHEL 7.4 changed the semantics of d_invalidate() such that it
invalidates the cwd, but did NOT change the return type to void.
This broke our autoconf test for detecting the new semantics.
- Further research reveals that d_prune_aliases() was not the best
choice for replacing d_invalidate(). This is because for directories,
d_prune_aliases() doesn't invalidate dentries when they are referenced
by its children, and it doesn't walk the tree trying to invalidate
child dentries. So it can leave dentries dangling, if the only
references to thos dentries are via children.
In preparation for future commits, revert
c3bbf0b444 .
Change-Id: Iafbef23a6070180c0e21eb01a2d59385ef52f55c
Reviewed-on: https://gerrit.openafs.org/12788
Reviewed-by: Andrew Deason <adeason@dson.org>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Linux 4.15 removes the distinction between "hot" and "cold" cache
pages, and pagevec_init() no longer takes a "cold" flag as the
second argument. Add a configure test and use it in osi_vnodeops.c .
Change-Id: Ia5287b409b2a811d2250c274579e6f15fd18fdbb
Reviewed-on: https://gerrit.openafs.org/12824
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Marcio Brito Barbosa <mbarbosa@sinenomine.net>
Tested-by: Marcio Brito Barbosa <mbarbosa@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Linux 4.15 removes the distinction between "hot" and "cold" cache
pages, and no longer provides page_cache_alloc_cold(). Simply use
page_cache_alloc() instead, rather than adding yet another test.
Change-Id: I34e734223927030f7ff252acb61120366a808ad6
Reviewed-on: https://gerrit.openafs.org/12823
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Marcio Brito Barbosa <mbarbosa@sinenomine.net>
Tested-by: Marcio Brito Barbosa <mbarbosa@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Place the debuginfo for the kmod into its own rpm so that
it doesn't have to track against the userspace packages.
FIXES 132034
Change-Id: I60a753275d896a89c1f6896c653d78a4e1fe7e2c
Reviewed-on: https://gerrit.openafs.org/11867
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
When using the configure option --enable-checking with gcc 7.2.1,
the compilation fails with
vutil.c:860:20: error: ‘%s’ directive writing up to 255 bytes into \
a region of size 63 [-Werror=format-overflow=]
This can be seen in the logs of the openSUSE Tumbleweed builder
for e.g. build 2368.
Avoid this warning by using snprintf which is provided by libroken
for all platforms.
Change-Id: I6acd3a1c06760abc8144c0892812c3bb50477227
Reviewed-on: https://gerrit.openafs.org/12813
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Apple has introduced a new file system called APFS. Starting from High
Sierra, APFS replaces Mac OS Extended (HFS+) as the default file system
for solid-state drives and other flash storage devices.
The current OpenAFS client is not aware of APFS. As a result, the
installation of the current client into an APFS volume will panic the
machine.
To fix this problem, make the OpenAFS client aware of APFS.
Change-Id: Ib5ac88b87f348744864f4e33f1f222efbc852d41
Reviewed-on: https://gerrit.openafs.org/12743
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
This commit introduces the new set of changes / files required to
successfully create the dmg installer on OS X 10.13 "High Sierra".
Change-Id: Id9da3cf959627a13d8cfd1d1d7412820e46ad63e
Reviewed-on: https://gerrit.openafs.org/12742
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Marcio Brito Barbosa <mbarbosa@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
This commit introduces the new set of changes / files required to
successfully build the OpenAFS source code on OS X 10.13 "High Sierra".
Change-Id: I51928279d97c9d86c67db7de5eb7fc9d317fd381
Reviewed-on: https://gerrit.openafs.org/12741
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
The m4 macro implementing the configure check is called
LINUX_KERNEL_READ_OFFSET_IS_LAST, but it defines a preprocessor symbol
that is just KERNEL_READ_OFFSET_IS_LAST. Our code needs to check
for the latter being defined, not the former.
Reported by Aaron Ucko.
Change-Id: Id7cd3245b6a8eb05f83c03faee9c15bab8d0f6e8
Reviewed-on: https://gerrit.openafs.org/12808
Reviewed-by: Anders Kaseorg <andersk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Rather than blindly trusting the values received in the
(unauthenticated) ack packet trailer, apply some minmial sanity checks
to received values. natMTU and regular MTU values are subject to
Rx minmium/maximum packet sizes, and the transmit window cannot drop
below one without risk of deadlock.
The maxDgramPackets value that can also be present in the trailer
already has sufficient sanity checking.
Extremely low MTU values (less than 28 == RX_HEADER_SIZE) can cause us
to set a negative "maximum usable data" size that gets used as an
(unsigned) packet length for subsequent allocation and computation,
triggering an assertion when the connection is used to transmit data.
FIXES 134450
Change-Id: I37698ff166da47a57aa0d1962ae8effc74e30851
Reported by the opensuse buildbot:
CC [M] /home/buildbot/opensuse-tumbleweed-i386-builder/build/src/libafs/MODLOAD-4.13.12-1-default-MP/rx_packet.o
/home/buildbot/opensuse-tumbleweed-i386-builder/build/src/afs/afs_pioctl.c: In function ‘PNewCell’:
/home/buildbot/opensuse-tumbleweed-i386-builder/build/src/afs/afs_pioctl.c:3075:55: error: ‘*’ in boolean context, suggest ‘&&’ instead [-Werror=int-in-bool-context]
if ((afs_pd_remaining(ain) < AFS_MAXCELLHOSTS +3) * sizeof(afs_int32))
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~
The bug was introduced in commit 718f85a8b6.
Change-Id: Iae55a99e35266aa763fb431f2acc4eba09fa5357
Reviewed-on: https://gerrit.openafs.org/12782
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
The recent event handling normalization in commit
304d758983 had event handlers switch
to dropping their reference on the associated connection/call just
before return. An early return case was missed in the conversion,
leading to a refcount leak in an error case.
Change-Id: Ie3d0bc9474fdbc09be9c753f4d0192c8cca68351
Reviewed-on: https://gerrit.openafs.org/12781
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
The order / content of the arguments passed to kernel_write and
kernel_read are not right. As a result, the kernel will panic if one of
the functions in question is called.
[kaduk@mit.edu: include configure check for multiple kernel_read()
variants, per linux commits bdd1d2d3d251c65b74ac4493e08db18971c09240
and e13ec939e96b13e664bb6cee361cc976a0ee621a]
FIXES 134440
Change-Id: I4753dee61f1b986bbe6a12b5568d1a8db30c65f8
Reviewed-on: https://gerrit.openafs.org/12769
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Tested-by: Marcio Brito Barbosa <mbarbosa@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Use the NUMEVENTS symbol which defines the array size instead of an
incorrect hard coded number when checking if a second event can be added
to be fired at the same time. This fixes a potential out of bounds
access of the event test array.
Also update the comment which incorrectly mentions the incorrect number
of events in the test.
Change-Id: I4f993b42e53e7e6a42fa31302fd1baa70e9f5041
Reviewed-on: https://gerrit.openafs.org/12762
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Instead of inlining the body (taking the lock, incrementing the
refcount, and dropping the lock), use the convenience function
designed for this purpose.
Change-Id: I674d389e61e42710ef340e202992748e66c5e763
Reviewed-on: https://gerrit.openafs.org/12772
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Reported by Mark Vitale
Change-Id: I3269fbb0f87285bcb9af64f4ad81791177582e6d
Reviewed-on: https://gerrit.openafs.org/12771
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
In utility functions that access fields of type struct rxevent *,
assert that the appropriate lock is held for the access in question.
These assertions are only compiled in when built with -DOPR_DEBUG_LOCKS,
which can be enbled by --debug-locks at configure time.
Change-Id: I16885a4d37a0f094f0d365c54e8157ed92070c69
Reviewed-on: https://gerrit.openafs.org/12757
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Go over all consumers of the rx event framework and normalize its usage
according to the following principles:
rxevent_Post() is used to create an event, and it returns an event
handle (with a reference on the event structure) that can be used
to cancel the event before its timeout fires. (There is also an
additional reference on the event held by the global event tree.)
In all(*) usage within the tree, that event handle is stored within
either an rx_connection or an rx_call. Reads/writes to the member variable
that holds the event handle require either the conn_data_lock or call
lock, respectively -- that means that in most cases, callers of
rxevent_Post() and rxevent_Cancel() will be holding one of those
aforementioned locks. The event handlers themselves will need to
modify the call/connection object according to the nature of the
event, which requires holding those same locks, and also a guarantee
that the call/connection is still a live object and has not been
deallocated! Whether or not rxevent_Cancel() succeeds in cancelling
the event before it fires, whenever passed a non-NULL event structure
it will NULL out the supplied pointer and drop a reference on the
event structure. This is the correct behavior, since the caller
has asked to cancel the event and has no further use for the event
handle or its reference on the event structure. The caller of
rxevent_Cancel() must check its return value to know whether or
not the event was cancelled before its handler was able to run.
The interaction window between the call/connection lock and the lock
protecting the red/black tree of pending events opens up a somewhat
problematic race window. Because the application thread is expected
to hold the call/connection lock around rxevent_Cancel() (to protect
the write to the field in the call/connection structure that holds
an event handle), and rxevent_Cancel() must take the lock protecting
the red/black tree of events, this establishes a lock order with the
call/connection lock taken before the eventTree lock. This is in
conflict with the event handler thread, which must take the eventTree
lock first, in order to select an event to run (and thus know what
additional lock would need to be taken, by virtue of what handler
function is to be run). The conflict is easy to resolve in the
standard way, by having a local pointer to the event that is obtained
while the event is removed from the red/black tree under the eventTree
lock, and then the eventTree lock can be dropped and the event run
based on the local variable referring to it. The race window occurs
when the caller of rxevent_Cancel() holds the call/connection lock,
and rxevent_Cancel() obtains the eventTree lock just after the event
handler thread drops it in order to run the event. The event handler
function begins to execute, and immediately blocks trying to obtain
the call/connection lock. Now that rxevent_Cancel() has the eventTree
lock it can proceed to search the tree, fail to find the indicated event
in the tree, clear out the event pointer from the call/connection
data structure, drop its caller's reference to the event structure,
and return failure (the event was not cancelled). Only then does the
caller of rxevent_Cancel() drop the call/connection lock and allow
the event handler to make progress.
This race is not necessarily problematic if appropriate care is taken,
but in the previous code such was not the case. In particular, it
is a common idiom for the firing event to call rxevent_Put() on itself,
to release the handle stored in the call/connection that could have
been used to cancel the event before it fired. Failing to do so would
result in a memory leak of event structures; however, rxevent_Put() does
not check for a NULL argument, so a segfault (NULL dereference) was
observed in the test suite when the race occurred and the event handler
tried to rxevent_Put() the reference that had already been released by
the unsuccessful rxevent_Cancel() call. Upon inspection, many (but not
all) of the uses in rx.c were susceptible to a similar race condition
and crash.
The test suite also papers over a related issue in that the event handler
in the test suite always knows that the data structure containing the
event handle will remain live, since it is a global array that is allocated
for the entire scope of the test. In rx.c, events are associated with
calls and connections that have a finite lifetime, so we need to take care
to ensure that the call/connection pointer stored in the event remains
valid for the duration of the event's lifecycle. In particular, even an
attempt to take the call/connection lock to check whether the corresponding
event field is NULL is fraught with risk, as it could crash if the lock
(and containing call/connection) has already been destroyed! There are
several potential ways to ensure the liveness of the associated
call/connection while the event handler runs, most notably to take care
in the call/connection destruction path to ensure that all associated
events are either successfully cancelled or run to completion before
tearing down the call/connection structure, and to give the pending event
its own reference on the associated call/connection. Here, we opt for
the latter, acknowledging that this may result in the event handler thread
doing the full call/connection teardown and delay the firing of subsequent
events. This is deemed acceptable, as pending events are for intentionally
delayed tasks, and some extra delay is probably acceptable. (The various
keepalive events and the challenge event could delay the user experience
and/or security properties if significantly delayed, but I do not believe
that this change admits completely unbounded delay in the event handler
thread, so the practical risk seems minimal.)
Accordingly, this commit attempts to ensure that:
* Each event holds a formal reference on its associated call/connection.
* The appropriate lock is held for all accesses to event pointers in
call/connection structures.
* Each event handler (after taking the appropriate lock) checks whether
it raced with rxevent_Cancel() and only drops the call/connection's
reference to the event if the race did not occur.
* Each event handler drops its reference to the associated call/connection
*after* doing any actions that might access/modify the call/connection.
* The per-event reference on the associated call/connection is dropped by
the thread that removes the event from the red/black tree. That is,
the event handler function if the event runs, or by the caller of
rxevent_Cancel() when the cancellation succeed.
* No non-NULL event handles remain in a call/connection being destroyed,
which would indicate a refcounting error.
(*) There is an additional event used in practice, to reap old connections,
but it is effectively a background task that reschedules itself
periodically, with no handle to the event retained so as to be able
to cancel it. As such, it is unaffected by the concerns raised here.
While here, standardize on the rx_GetConnection() function for incrementing
the reference count on a connection object, instead of inlining the
corresponding mutex lock/unlock and variable access.
Also enable refcount checking unconditionally on unix, as this is a
rather invasive change late in the 1.8.0 release process and we want
to get as much sanity checking coverage as possible.
Change-Id: I27bcb932ec200ff20364fb1b83ea811221f9871c
Reviewed-on: https://gerrit.openafs.org/12756
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
We currently do not properly handle the case where a thread runs
rxevent_Cancel() in parallel with the event-handler thread attempting
to fire that event, but the test suite only picked up on this issue
in a handful of the Debian automated builds (somewhat less-resourced
ones, perhaps).
Modify the event scheduling algorithm in the test so as to create a
larger chunk of events scheduled to fire "right away" and thereby
exercise the race condition more often when we proceed to cancel
a quarter of events "right away".
Change-Id: I50f55fd532901147cfda1a5f40ef949bf3270401
Reviewed-on: https://gerrit.openafs.org/12755
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
If ncurses is built with "./configure --with-termlib=tinfo", gtx fails
to link because of an undefined reference to the LINES symbol which is
then provided by libtinfo.so and not libncurses.so.
If ncurses is present, additionally check whether LINES is provided by
ncurses or tinfo and set $LIB_curses accordingly.
This change is based on a patch provided by Bastian Beischer.
FIXES 134420
Change-Id: I3e29c61405d90d0b850bafe4c51125bef433452b
Reviewed-on: https://gerrit.openafs.org/12760
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
AS_IF does not invoke the test(1) shell builtin for us, so we must
take care to consistently use it ourself.
While here, sprinkle some missing double-quotes around variable
expansions in AS_IF statements in this file.
Submitted by Bastian Beischer.
FIXES 134414
Change-Id: Iccfe311011f17de6317cf64abdc58b0812b81b8c
Reviewed-on: https://gerrit.openafs.org/12738
Reviewed-by: Michael Meffie <mmeffie@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Tested-by: Benjamin Kaduk <kaduk@mit.edu>
We hide the uses of set_fs/get_fs behind a macro, as those functions
are likely to soon become unavailable:
> Christoph Hellwig suggested removing all calls outside of the core
> filesystem and architecture code; Andy Lutomirski went one step
> further and said they should all go.
https://lwn.net/Articles/722267/
Change-Id: Ib668f8fdb62ca01fe14321c07bd14d218744d909
Reviewed-on: https://gerrit.openafs.org/12729
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Older versions of rpmbuild do not support the files exclude directive,
so fall back to the old way in which we remove the files to be excluded
and list the files to be included.
Change-Id: If64df382ef372aa1078f1703a34942a1930bdc88
Reviewed-on: https://gerrit.openafs.org/12733
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Move the deprecated klog.krb, pagsh.krb, and tokens.krb programs and man
pages to the optional openafs-kauth-client subpackage.
Change-Id: I09a2e36b60f9d47726a6a314a26db88e44575567
Reviewed-on: https://gerrit.openafs.org/12732
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Currently, some of the man pages are specified with the full name and
some are specified with a wildcard for the filename extension. Instead,
specify all the man pages without a wildcards to be more consistent and
to avoid putting incorrect man pages in packages.
This change removes a stray copy the klog.krb5.1 man page from
openafs-kauth-client subpackage and moves the AuthLog/AuthLog.dir man
pages to the optional openafs-kauth-server subpackage.
Change-Id: Id30a6174c532a9a00f850d6ca2722158293d5118
Reviewed-on: https://gerrit.openafs.org/12731
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
The afsd.fuse binary is not currently packaged; do not package the man
page.
Change-Id: Ia0dd4fa72dc8a87e2c835798b6fbe1213d71da5f
Reviewed-on: https://gerrit.openafs.org/12730
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
As already described on 7c708506, SDISK_Begin fails on remotes if
lastYesState is not set. To fix this problem, 7c708506 does not allow
write transactions until we know that lastYesState is set on at least
quorum (ubik_syncSiteAdvertised == 1). In other words, if enough sites
received a beacon packet informing that a sync-site was elected, write
transactions will be allowed. This means that ubik_syncSiteAdvertised
can be true while lastYesState is not set in a few sites.
Consider the following scenario in a cell with frequent write
transactions:
Site A => Sync-site (up)
Site B => Remote 1 (up)
Site C => Remote 2 (down - unreachable)
Since A and B are up, we have quorum. After the second wave of beacons,
ubik_syncSiteAdvertised will be true and write transactions will be
allowed. At some point, C is not unreachable anymore. Site A sends a
copy of its database to C, but C did not vote for A yet (lastYesState ==
0). A new write transaction is initialized and, since lastYesState is
not set on C, DISK_Begin fails on this remote site and C is marked as
down. Since C is reachable, A will mark this remote site as up. The
sync-site will send its database to C, but C did not vote for A yet. A
new write transaction is initialized and, since lastYesState is not set
on C, DISK_Begin fails on this remote site and C is marked as down. In a
cell with frequent write transactions, this cycle will repeat forever.
As a result, the sync-site will be constantly sending its database to C
and quorum will be operating with less sites, increasing the chances
of re-elections.
To fix this problem, do not call DISK_Begin on remotes that did not
vote for the sync-site yet.
Change-Id: I27f5122a089064e7b83beba3533261d8a4e31c64
Reviewed-on: https://gerrit.openafs.org/12715
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Mark Vitale <mvitale@sinenomine.net>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
The following commit:
commit eb031849d52e61d24ba54e9d27553189ff328174
Author: Christoph Hellwig <hch@lst.de>
Date: Fri Sep 1 17:39:23 2017 +0200
fs: unexport __vfs_read/__vfs_write
unexports both __vfs_read and __vfs_write, but keeps the former in
fs.h--as it is is still being used by another part of the tree.
This situation results in a false positive in our Autoconf check,
which does not see the export statements, and ends up marking the
corresponding API as available.
That, in turn, causes some code which assumes symmetry with
__vfs_write to fail to compile.
Switch to testing for __vfs_write, which correctly marks the API as
unavailable.
Change-Id: I392f2b17b4de7bd81d549c84e6f7b5ef05e1b999
Reviewed-on: https://gerrit.openafs.org/12728
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
