Ben Kaduk [Wed, 19 May 2010 05:44:44 +0000 (01:44 -0400)]
Do not sleep in FBSD's getevent()
Calling osi_AllocSmallSpace drops the glock, which allows
a race condition that caused a rather fast deadlock on
multiprocessor systems.
Use afs_osi_Alloc_NoSleep instead.
Define the PageCheck family of functions using PageFsMisc
equivalents if those are found. I encountered a need for this
on Centos 5.5 (2.6.18-8.el5xen), but it may be more general.
Simon Wilkinson [Mon, 17 May 2010 19:15:46 +0000 (20:15 +0100)]
Fileserver: Don't sync every 10 seconds
The patch which moved our calls to fsync into the background, added
a sync prior to each batch of fsync calls. This is an incredibly bad
idea.
POSIX says that sync "shall cause all information in memory that
updates file systems to be scheduled for writing out to all file
systems". On Linux this means that we in effect perform an fsync()
on every open filehandle on the entire system, and in addition flush
all superblocks and journals out to disk. This makes the following
fsync() calls superfluous - sync() will have already written out
all of the data.
Add to this the fact that the fileserver is doing this every 10
seconds, and this becomes a major performance bottleneck,
particularly if the machine uses a journalled fs - as any disk
operations will end up blocking whilst the journal is written to
disk.
Jeffrey Altman [Mon, 17 May 2010 20:01:03 +0000 (16:01 -0400)]
rx: work harder to notice and handle MorePackets request
in particular, we did badly at handling kernel requests for
more packets, but other cases did not properly keep packet
stats either. attempt to globally better handle demand for
more packets.
Simon Wilkinson [Mon, 17 May 2010 07:40:49 +0000 (08:40 +0100)]
Autoconf: Use AC_CACHE_CHECK
AC_CACHE_CHECK has been available since Autoconf 2.5 (released in
1995), and provides a simpler mecahnism for performing cached checks
than AC_CACHE_VAL does. Modify all of our simple AC_CACHE_VAL calls
to use AC_CACHE_CHECK.
Rod Widdowson [Wed, 12 May 2010 14:48:25 +0000 (15:48 +0100)]
Clean up logging from vldb_check.
vldb_check used to say "%<index>: Volume '<name>' <id>not found in <type> hash <hash>" for two sorts of failures
- If the volume really wasn't in a hash chain of that type at all
- If volume's next pointer (for the hash type) pointed to a volume with a different hash.
In both cases it would then print out some confusing (and spurious)
information about the next element in the chain, prior to applying the
same fix (of zeroing the next pointer).
Rod Widdowson [Sat, 15 May 2010 15:25:26 +0000 (16:25 +0100)]
Stop overanxious htonl in vldb_check -fix
vldb_check -fix is corrupting the server entries (Number, Partition,
Flags). This is because it is doing a htonl on what is a char field.
The read code already has a comment "make sure not to ntohl these,
as they're chars, not ints".
This change strips the corresponding htonl from writeheader (and puts
in a similar comment).
Simon Wilkinson [Sun, 25 Apr 2010 19:23:27 +0000 (20:23 +0100)]
Demand Attach: Simplify __VVGC_entry_cl_add
The __VVGC_entry_cl_add was giving an uninitialised variable warning
for empty_idx, because it was using a separate variable, empty_found
as a sentinel for when empty_idx was holding a real value. Simplify all
of this by removing the sentinel, and making empty_idx being -1 mean
'not found', and a real value imply found.
Change-Id: I6c485993df1ff8dfd27824234c811af998ff94fa
Reviewed-on: http://gerrit.openafs.org/1833 Reviewed-by: Tom Keiser <tkeiser@sinenomine.net> Reviewed-by: Alistair Ferguson <alistair.ferguson@mac.com> Reviewed-by: Andrew Deason <adeason@sinenomine.net> Reviewed-by: Derrick Brashear <shadow@dementia.org> Tested-by: Derrick Brashear <shadow@dementia.org>
#ifdef A
if (foo) {
...
} else
#endif
#ifdef B
if (bar) {
...
}
#endif
something_else()
In a situation where we have A && !B, then something_else() ends up
only being run when foo is false. Given that something_else() is
VOL_LOCK, this will not end well.
In the real world, we hit this problen when we build the volume
package with SALVSYNC_BUILD_CLIENT and !FSYNC_BUILD_CLIENT - in
other words, whilst building the fileserver.
Ben Kaduk [Mon, 17 May 2010 04:22:32 +0000 (00:22 -0400)]
Enable PutVCache for FBSD80 and higher
Matt's patches in a123bd8ccd brought in support for FBSD 7.0, but
also disabled several calls to afs_PutVCache() for FBSD80 and
higher. Matt says that that patch was focused on 7.0, so the
8.0-related bits may not be relevant.
I have run with the PutVCache calls enabled, and if anything,
my system seems more stable with them.
Ben Kaduk [Sun, 16 May 2010 04:43:35 +0000 (00:43 -0400)]
Kill FBSD4X with fire
We haven't even pretended to work on the 4.X series for quite some
time, and keeping this code around just makes things (slightly)
harder to read.
AFS_FBSD_ENV is now equivalent to AFS_FBSD50_ENV (though the
latter should not be used).
Leave the fbsd_4 sysnames in afs_sysnames.h for archival purposes.
Simon Wilkinson [Sun, 16 May 2010 22:14:11 +0000 (23:14 +0100)]
Autoconf: Use AC_CHECK_MEMBERS
AC_CHECK_MEMBERS has been available since autoconf 2.50,
which was released in 2001. Use this, rather than rolling our
own mechanism to check for structure membership.
Simon Wilkinson [Sun, 16 May 2010 21:56:13 +0000 (22:56 +0100)]
Autoconf: Use a standard test for socklen_t
Use the standard AC_CHECK_TYPES mechanism to check for the existence
of socklen_t, and to typedef it as an int if it's not found, rather
than growing our own.
Andrew Deason [Sun, 16 May 2010 04:12:33 +0000 (23:12 -0500)]
Solaris: do not call VFS_RELE with GLOCK
VFS_RELE can call afs_freevfs, which grabs GLOCK. Thus, don't call
VFS_RELE with GLOCK held or we can try to recursively acquire GLOCK
and panic. This is currently unlikely to occur (sans vfs refcount
unbalances) without support for forced unmounts, since the last vfs
ref will usually come from the caller of afs_unmount(). But it still
may be possible.
Andrew Deason [Wed, 12 May 2010 18:59:06 +0000 (13:59 -0500)]
Solaris: stop NetIfPoller on shutdown
The NetIfPoller code that is enabled on AFS_SUN510_ENV never gets
shutdown properly. The current code looks as if it was intended for
NetIfPoller to recognize the appropriate termState and just return,
but we never wait for it, and so we can complete the shutdown sequence
without NetIfPoller ever knowing that we are shutting down. This can
cause the machine to panic, as we keep attempting to run NetIfPoller
even after libafs has been unloaded.
Since NetIfPoller is fired by default every 30 seconds, we probably do
not want to wait for it to fire during shutdown. Instead, just destroy
the necessary timeout and task queue, which will wait for NetIfPoller
to complete if running, and will just prevent it from running in the
future otherwise.
Jeffrey Altman [Wed, 12 May 2010 15:27:17 +0000 (11:27 -0400)]
Windows: move afsconfig.h / param.h to afscred.h
move the inclusion of afsconfig.h and param.h to
afscred.h so that they are included everywhere
and so that the conditional use of _USE_32BIT_TIME_T
is applied consistently by afscred.h.
Jeffrey Altman [Sun, 9 May 2010 01:38:05 +0000 (21:38 -0400)]
Rx: prevent rx_rpc_stats mutex from being a global bottleneck
Prior to this patchset, the 'rx_rpc_stats' mutex was superior
to both the 'peer->peer_lock' and the 'rx_peerHashTable_lock'.
That meant that the 'rx_rpc_stats' was being held across many
operations that walk the peer hash table. For example,
rxi_ReapConnections, rx_disablePeerRPCStats, and rx_shutdown.
Since every RPC issues a call to rx_IncrementTimeAndCount, the
reap connections event would effectively bring all RPC processing
to a halt.
This patchset moves 'rx_rpc_stats' later in the hierarchy and
restructures rxi_ReapConnections, rx_disablePeerRPCStats, and
rx_shutdown so that not only doesn't the 'rx_rpc_stats' mutex
need to be held across the entire function but the
'rx_peerHashTable_lock' does not need to be held while complex
operations on the peer object are taking place.
rxi_ReceiveDebugPacket is also fixed to hold the rx_peerHashTable_lock
and peer_lock at appropriate times while completing its function.
Jeffrey Altman [Mon, 10 May 2010 13:07:50 +0000 (09:07 -0400)]
Windows: Prevent overflow during percent used calc in Explorer Shell Ext
In the Volume Info and Partition Info dialog boxes, percent used
was overflowing the variable due to using too small a variable
and multiplying before dividing.
Jeffrey Altman [Mon, 10 May 2010 00:49:05 +0000 (20:49 -0400)]
Windows: use system CreateUuid instead of afs variant
The afs_create_uuid function on Windows results in
duplicate uuids being produced. Windows provides its own
CreateUuid function. Use it when afs_create_uuid is called.
Jeffrey Altman [Sun, 9 May 2010 14:15:30 +0000 (10:15 -0400)]
Make afsio be useful on Windows and fix bugs
On Windows, the stdin and stdout handles default to text
mode which corrupts data piped through them. At the start
of readFile and writeFile, the mode is now set to binary.
The ProbeUuid callback is supposed to determine if this
client is the same as the one being requested. Since we
are advertising a new uuid in TellMeAboutYourself and
WhoAmI, we should perform the uuid check and return a
valid answer in order to avoid confusing the file server.
Add support for TellMeAboutYourself. On Windows, return
valid interface information.
Always print verbose output to stderr.
Display MB/sec to three decimal places.
In readFile, always perform a AFSFetchStatus operation
otherwise we have no idea what the file size is.
Jeffrey Altman [Thu, 6 May 2010 21:12:48 +0000 (17:12 -0400)]
Windows: RXAFS_InlineBulkStat errors must be processed via cm_Analyze
RXAFS_InlineBulkStatus does not return errors such as EACCES,
VNOVOL, VNOVNODE, VOFFLINE, VBUSY, VIO, VMOVED, etc. as an RPC return
code. Instead they are returned in the status info errorCode field
for each file.
Traditionally, the error associated with the first FID in the query
list has been returned to the caller of cm_TryBulkStatRPC().
However, the error has never been processed through cm_Analyze()
which means that the per-vnode processing for VNOVNODE and the volume
global processing for VMOVED, VNOVOL, etc. has never been performed.
As a result, failover to other .readonly volume instances cannot occur,
volume moves will not be handled, and files that have been deleted
are not detected.
This patchset makes the following changes:
1. If an inline bulk operation has been performed and the inline
errorCode is a volume global error, then that error replaces
the RPC return code within the cm_Analyze() processing for
the RPC. This will affect whether or not a retry operation
is performed.
2. The variable 'inlinebulk' is reset to 0 at the top of the
cm_Analyze() loop in case failover from an inlinebulk capable
file to an inlinebulk incapable file server takes place.
3. The FID that is passed into cm_Analyze() is not a real fid.
Instead it consists of the cell and volume but vnode = 0.
This ensures that the error (if any) is not applied to the
directory object.
4. If an inline bulk operation was performed, prior to performing
the cm_MergeStatus() operation a vnode a check is made to
determine if an error was returned for that vnode. If so,
cm_Analyze() is called with no connection, a fake cm_req_t,
the fid, and the error. This permits cm_Analyze() processing
to be performed on the file.
Michael Meffie [Tue, 11 May 2010 19:54:53 +0000 (15:54 -0400)]
Linux: missing configure test
Add the kmem_cache_create test for dtor to the configure to
build again on older versions of linux. Fixes,
osi_vfsops.c:264: error: too few arguments to function
‘kmem_cache_create’
Simon Wilkinson [Thu, 6 May 2010 21:22:50 +0000 (22:22 +0100)]
Remove intptr.m4
We now require an autoconf version of at least 2.60. This version
supplies AC_TYPE_INTPTR_T and AC_TYPE_UINTPTR_T, so we no longer
need to have our own.
Andrew Deason [Wed, 5 May 2010 18:51:23 +0000 (13:51 -0500)]
vlserver: prevent duplicate IPs via ChangeAddr
Add a safety check in VL_ChangeAddr/ChangeIPAddr to check if we are
adding a duplicate server IP in the database. Return VL_MULTIPADDR
when we try to do that.
Jeffrey Altman [Mon, 10 May 2010 00:46:11 +0000 (20:46 -0400)]
Windows: define HAVE_SSIZE_T
The windows param.*.h files define the ssize_t type.
Therefore, we must define HAVE_SSIZE_T in order to avoid
a collision with the ssize_t typedef in afs/stds.h.
Simon Wilkinson [Sun, 9 May 2010 18:40:58 +0000 (19:40 +0100)]
Add a list of files to import from Heimdal
Add an initial list of files to import into OpenAFS from Heimdal.
Currently this is just the list that are required to support
our new configuration file parser.
Simon Wilkinson [Sun, 9 May 2010 18:37:52 +0000 (19:37 +0100)]
Add a tool to import external repositories
This commit adds a tool to aid with the importing of external git
repositories. import-external-git.pl allows a developer to easily
update the src/external portion of the tree to the latest release
of externally developed software. It's currently intended to support
our use of Heimdal functions, but could be extended to support any
project whose code we wish to use in the kernel, and which has a
git repo.
Russ Allbery [Thu, 6 May 2010 19:30:23 +0000 (12:30 -0700)]
Include common param file for Linux alpha builds
The sysname for Linux Alpha is alpha_linux_26 (or 24 or 22) with an
extra underscore before the version, unlike all the other Linux sysnames.
Allow for that case when deciding whether to include the common Linux
param header.
Simon Wilkinson [Thu, 6 May 2010 14:00:00 +0000 (15:00 +0100)]
Modernise use of AC_CHECK_TYPE
Prior to autoconf 2.13, AC_CHECK_TYPE took two arguments, the type
to check and the a default type to use if that type wasn't defined.
This usage has been deprecated since 2.13, and the AC_CHECK_TYPE
macro itself now behaves differently. Whilst there is a compatibility
mode, the autoconf documentation discourages its use.
We also have an occurence of a slightly more modern AC_CHECK_TYPE,
where we explicitly #define a default value if one isn't provided.
The autoconf manual also discourages this form, however, in favour
of using AC_CHECK_TYPES, and placing suitable typedefs in an
external header file.
Modify our code so we do things in the recommended way.
Simon Wilkinson [Thu, 6 May 2010 13:58:52 +0000 (14:58 +0100)]
Always include afsconfig.h in autogenerated files
Modify rxgen and compile_et so that the files they output always
contain afsconfig.h before afs/param.h. This avoids problems where
afs/param.h, or headers included from it, rely on having the results
of configure tests available.
OpenBSD: Use osi_obsd_Free() for all releases of OBSD
Use osi_obsd_Free() in afs_osi_Free() for all releases of OpenBSD,
not just releases from 4.4 onward. Otherwise, afs_osi_Free() will
use AFS_KFREE, which does not exist, since OBSD/osi_machdep.h is
not included, resulting in unresolved symbols when trying to load
the kernel module.
In OpenBSD releases prior to 4.2, use the system (kernel-specific)
FREE macro to release space in the kernel module, not KFREE. The
latter does not exist and results in unresolved symbols when trying
to load the kernel module. (This was undoubtedly a typo.)
Simon Wilkinson [Thu, 6 May 2010 13:55:59 +0000 (14:55 +0100)]
Always include afsconfig.h
Our coding standards say that OpenAFS source files should always have
#include <afsconfig.h>
#include <afs/param.h>
at the start of the file. Including just param.h, or having these
includes in a different order can produce unexpected effects, because
param.h includes sysnames.h, which in turn includes stds.h, which may
rely upon having the results of configure tests available.
Fix the obvious places in the Unix build which get this wrong.
Windows: Secure C String usage in src\WINNT\afsd\fs.c
The use of strlen, strcat, strcpy, strncpy and sprintf in
src\WINNT\afsd\fs.c does not effectively protect against buffer
overruns and string truncation errors. This patchset replaces
their use with StringCbLen, StringCbCat, StringCbCopy,
StringCbCopyN, StringCbCopyEx and StringCbPrintf respectively from
strsafe.h.
The functions sscanf, fscanf and scanf do not check for the size of
the parameters so it is prone to buffer overruns. These functions
can be replaced by sscanf_s, fscanf_s, scanf_s when the Visual Studio
compiler version is 1400 or greater. Also memcpy will be replaced
by memcpy_s which will check for the size of the destination buffer.
In all cases, failure conditions are handled.
Patchset development was mentored by Asanka Herath and Jeffrey Altman.
Change-Id: I01e9dfc616641c220b57d2871e3e140406df3653
Reviewed-on: http://gerrit.openafs.org/1736 Tested-by: Jeffrey Altman <jaltman@openafs.org> Reviewed-by: Tharidu Fernando <tharidufernando@gmail.com> Tested-by: Tharidu Fernando <tharidufernando@gmail.com> Reviewed-by: Jeffrey Altman <jaltman@openafs.org>
Andrew Deason [Mon, 3 May 2010 19:57:28 +0000 (14:57 -0500)]
Solaris: shutdown in freevfs, not unmount
Call the afs shutdown sequence (afs_shutdown()) in the VFS_FREEVFS
handler, instead of in the VFS_UNMOUNT handler. This way we shutdown
when the last reference to our vfs goes away, instead of when we are
unmounted. That doesn't make much of a difference right now, but doing
so will be required for supporting forced unmounts.
Andrew Deason [Thu, 29 Apr 2010 22:47:15 +0000 (17:47 -0500)]
Solaris: prevent AFS umount while busy
Return EBUSY from unmount if someone still references stuff in AFS.
This prevents kernel panics that can occur on shutdown if we umount
while there is a file in AFS open. Normally a process can hold a file
in AFS open, AFS is unmounted, and the file is closed, triggering our
code which explodes if called after we're unmounted.
This adds VFS_HOLD/VFS_RELE calls whenever we 'create' a vcache, or
retire an old one, to keep track if anyone has an open reference to
us.
Simon Wilkinson [Sat, 7 Nov 2009 00:16:28 +0000 (00:16 +0000)]
Refactor afs_NewVCache
afs_NewVCache was a twisty turny maze of #ifdefs and duplicated code.
This makes a number of sweeping changes to simplify this code, and to
move platform specific elements out into their own directories.
*) ShakeLooseVCaches is refactored so that the same code can be used
both for platforms that support dynamic vcaches, and those which
don't.
*) afs_NewVCache, ShakeLooseVCaches, and afs_AllocVCache are all
modified to remove platform specific code, and to call platform
specific functions.
*) A new platform file 'osi_vcache.c' is created to hold a number of
platform specific vcache operations:
*) osi_TryEvictVCache handles the decision of whether a vcache can
be evicted or not, and does so if required
*) osi_NewVnode allocates a new vnode
*) osi_PrePopulateVCache does the necessary population of the
vcache, before it's threaded onto the VLRUQ and associated hash
tables.
*) osi_AttachVnode handles attaching an OS vnode to our vcache,
where that is necessary
*) osi_PostPopulateVCache handles the vcache population that must
occur after we're on the VLRUQ and have a vnode attached.
OpenBSD: Fix bug in setpag() when group list is empty
In OpenBSD, the PAG uses the 2nd and 3rd group slots in the task's
group list. If an application sets en empty group list (such as Samba
does), any existing PAG is lost and any new one is NOT set because
the existing code will set the new group count to 2 instead of 3, and
it leaves the first group entry as garbage (whatever random value the
memory contained), thereby totally messing up the task's group list.
This patch fixes it so that it behaves as expected.
Fix pattern to detect i386-based OpenBSD system type
When the pattern to detect OpenBSD system types (via MKAFS_OSTYPE) was split
into two separate ones for amd64 and i386 architectures, the pattern for
i386-based systems was not set quite right. As a result, i386-based OpenBSD
systems are no longer detected properly at this stage, resulting in an error
of "cannot find input file: src/libafs/MakefileProto..in" during configure.
This patch corrects the pattern.
Windows: Show configuration pages for all types of MSI installations
The OpenAFS MSI installer wizard used to not show any configuration
pages for "Typical" and "Complete" installations. Setting the
workstation cell and logon options during installation required
selecting the "Custom" option. Many users choose the "Typical" option
during installation, and thus would never see the configuration pages.
Therefore, for these users, the workstation cell was being set to the
default.
This patch makes the workstation cell and logon option configuration
pages visible to all types of installations (except silent
installations which show no UI).
Jeffrey Altman [Sun, 2 May 2010 19:30:27 +0000 (15:30 -0400)]
Windows: Restructure cm_LookupInternal addition of Freelance entries
cm_LookupInternal creates Freelance mount points and symlinks
when queries cannot be found in the Freelance root.afs directory.
If the search name is a full cell name for which vldb information
can be obtained, then a mount point is added. If the search name
is a left-most substring or the full cell name with a dot appended
to it, then a symlink was created. This approach created a very
poluted Freelance name space.
This patchset makes the following changes:
1. Do not create symlinks with a dot appended to the cellname
2. Do not create symlinks where the left-most substring is not
a full dot separated component of the cellname.
3. Permit lookups to succeed when we would have created a
symlink in the past without creating the symlink.
Jeffrey Altman [Sun, 2 May 2010 19:26:39 +0000 (15:26 -0400)]
Windows: Permit BPlus tree lookups within cm_ApplyDir
BPlus tree lookups are much faster than searching through
the native directory format on Windows because the case sensitive
hash tables cannot be used successfully. Permit BPlus trees
to be used except when called with cm_BPlusDirFoo as the action
function because cm_BPlusDirFoo is used to build the BPlus trees
from the native directory format.
Jeffrey Altman [Sun, 2 May 2010 18:48:00 +0000 (14:48 -0400)]
Windows: Force rebuilding Freelance directory in AddMount and AddSymlink
cm_FreelanceAddMount and cm_FreelanceAddSymlink is supposed to
return the allocated FID of the entry that was added. However,
cm_NameI is called to perform the lookup without forcing an update
of the Freelance fake directory. As a result the entry may not be
found.
Force an update prior to calling cm_NameI() by using
cm_clearLocalMountPointChange() and cm_reInitLocalMountPoints()
if required.
Jeffrey Altman [Sun, 2 May 2010 18:03:23 +0000 (14:03 -0400)]
Windows: use cm_noteLocalMountPointChange whenever fakeDirVersion is changed
cm_noteLocalMountPointChange() is meant to be used when fakeDirVersion
is changed. It previously wasn't used because cm_noteLocalMountPointChange()
would obtain the cm_Freelance_Lock. Now that cm_noteLocalMountPointChange()
can be called while holding the lock, use it everywhere.
Jeffrey Altman [Sun, 2 May 2010 17:56:24 +0000 (13:56 -0400)]
Windows: add locked parameter to cm_noteLocalMountPointChange
If cm_Freelance_Lock is held set the 'locked' parameter to TRUE.
This avoids a requirement to drop the lock only to have it be
obtained when cm_noteLocalMountPointChange.
The validation check for the response from the GetVolumeStatus
pioctl is incorrect. The response is not simply a VolumeStatus
structure but also several C strings appended to it.
Jeffrey Altman [Sun, 2 May 2010 13:16:29 +0000 (09:16 -0400)]
Windows: getAFSServer wrong variable used in uncompiled code
In the uncompiled implementation of getAFSServer, the unmodified
'cellname' is used in the send_DNS_AFSDB_Query call instead of
the properly constructed 'query' variable.
Windows: FlushFile ioctl should call cm_FSync, not buf_CleanVnode
When flushing a file, we need to commit the file length changes
as well as the dirty buffers. Call cm_FSync instead of buf_CleanVnode
which is called by cm_FSync.
Windows: add 'locked' flag to cm_FSync and call when dropping write locks
cm_FSync should be called when releasing file server write locks.
This ensures that all dirty buffers are stored before the lock
can be acquired by another client.
Since cm_Unlock holds the cm_scache_t->rw exclusively when calling
cm_FSync, add a 'locked' parameter to cm_FSync to avoid an unnecessary
release and reacquire of the rwlock.
Matt Benjmain [Wed, 28 Apr 2010 10:19:16 +0000 (06:19 -0400)]
netbsd: rebase cm at NetBSD 4.0
Rebases the NetBSD client port at OpenBSD, which was originally
based on an original NetBSD client port by John Kohl. The
platforms remain closely connected.
At latest milestone, the port builds as a NetBSD LKM, which was
loadable and can mount /afs (but much work remains past this
point).
Simon Wilkinson [Wed, 28 Apr 2010 00:13:30 +0000 (01:13 +0100)]
Linux: Register number sysctls only when desperate
The AFS sysctl interface was trying to register numeric system
calls whenever the kernel would let it. This is anti-social, as
we've never had sysctl numbers allocated to us. In kernels newer
than 2.6.24, this misbehaviour is detected and punished. In recent
kernels, the binary interface has gone away entirely.
Since 2.6.19, the kernel has supported allocated unnumbered system
calls. These are only available via /proc/sys (and not the using
the legacy, binary, API), but they're really the only interface
available to us.
Change the sysctl registration code to use unnumbered calls whenever
they are available. Unnumbered calls aren't available in 2.4.x, so
completely remove that code there.
Marc Dionne [Sun, 25 Apr 2010 02:12:27 +0000 (22:12 -0400)]
Linux: Remove unnecessary crhold on incomplete kernel credentials
When credentials debugging is active in the kernel, sanity checks
in various credentials functions such as get_cred or put_cred
require that they are called with full fledged kernel credentials.
This can be a problem with recent kernels in osi_Init where we
build our own credentials and try to "crhold" them.
Getting a reference doesn't make much sense anyway since we're
using a statically allocated structure that we never try to free
or modify.
As a minimal fix, just don't call crhold when we're using
the kernel credentials structure.
right now, flock with LOCK_EX, on unlock, triggers an async store.
make it sync, but also, make sure to ask to have any in-memory
data sent to us (e.g. VM_StoreAllSegments); unlike Solaris
VMSYNC_INVAL, we don't want to invalidate the pages, just get copies
written back.
LOCK_SH did not, does not and AFAICT should not trigger writes.
Simon Wilkinson [Tue, 27 Apr 2010 18:31:55 +0000 (19:31 +0100)]
Linux: Reorder the tests
As a final change to the Linux autoconf stuff, reorder the list of
tests so that it's more easy to see what's being checked for, and to
prevent multiple copies of the same thing being added.
Simon Wilkinson [Tue, 27 Apr 2010 17:57:30 +0000 (18:57 +0100)]
Linux: Don't preprocess osi_vfs.h
Before we did inodes properly, we needed to pre-process osi_vfs.h,
in order to merge our inode structure with the Linux one. Ever since
we moved to native inodes, that preprocessing has been disabled, and
we've just copied osi_vfs.hin to osf_vfs.h Skip this pointless step,
permanently rename osi_vfs.hin to osi_vfs.h, and remove the unused
support scripts.
Simon Wilkinson [Tue, 27 Apr 2010 16:41:22 +0000 (17:41 +0100)]
Linux: Add general autoconf macro for Linux kernel
Add AC_CHECK_LINUX_BUILD() to do a standardised Linux build, which
takes a "checking" message, the autoconf variable to use to cache
the results, headers and code to run, preprocessor variable to
define, and a description of that variable.
Reimplement all of our existing check macros in terms of this one,
resolving many typos along the way.
Simon Wilkinson [Tue, 27 Apr 2010 12:02:20 +0000 (13:02 +0100)]
Linux: Add autoconf macro for structure checks
Add a new autoconf macro for doing structure element checks, and
modify all simple structure checks to use it. This introduces a
standard name form - STRUCT_structure_HAS_element, so there are
some changes in the code to make use of this standard form.
Simon Wilkinson [Mon, 26 Apr 2010 22:54:05 +0000 (23:54 +0100)]
Linux: Macroise kernel function autoconf tests
We spend a lot of time in autoconf checking to see if kernel
functions are available. Rather than copying and pasting the same
code everytime we do so, create a macro to do function tests, and
use that macro. This results in standardised naming, so change all
of the places that use autoconf results to use the same standardised
names.
projects / packages / o / openafs.git / log