If two CPUs run mod_timer against the same not-pending timer then they
have no locking relationship.  They can both see the timer as
not-pending and they both add the timer to their cpu-local list.  The
CPU which gets there second corrupts the first CPU's lists.

This was causing Dave Hansen's 8-way to oops after a couple of minutes
of specweb testing.

I believe that to fix this we need locking which is associated with the
timer itself.  The easy fix is hashed spinlocking based on the timer's
address.  The hard fix is a lock inside the timer itself.

It is hard because init_timer() becomes compulsory, to initialise that
spinlock.  An unknown number of code paths in the kernel just wipe the
timer to all-zeroes and start using it.

I chose the hard way - it is cleaner and more idiomatic.  The patch
also adds a "magic number" to the timer so we can detect when a timer
was not correctly initialised.  A warning and stack backtrace is
generated and the timer is fixed up.  After 16 such warnings the
warning mechanism shuts itself up until a reboot.

It took six patches to my kernel to stop the warnings from coming out.
The uninitialised timers are extremely easy to find and fix.  But it
will take some time to weed them all out.  Maybe we should go for
the hashed locking...

Note that the new timer->lock means that we can clean up some awkward
"oh we raced, let's try again" code in timer.c.  But to do that we'd
also need to take timer->lock in the commonly-called del_timer(), so I
left it as-is.

The lock is not needed in add_timer() because concurrent
add_timer()/add_timer() and concurrent add_timer()/mod_timer() are
illegal.

Final act, from Manfred:

The attached patch removes 'event' entirely from the kernel: it's not
used anymore.

All event users [vfat dentry revalidation; ext2/3 inode generation;
readdir() file position revalidation in several filesystems] were
converted to local counters.

Patch from Manfred Spraul

Several filesystems compare f_version and i_version to validate
directory positions in readdir(): The directory position is revalidated
if i_version is not equal f_version.  Operations that could invalidate
the cached position set i_version or f_version to '++event', event is a
global variable.  Global uniqueness is not needed, 'i_version++' and
'f_version=0' is sufficient to guarantee that the next readdir() will
revalidate the directory position, and that avoids the need for an ugly
global variable.

The attached patch converts all filesystems except ext2, which was
converted with a seperate patch.

Patch from Manfred Spraul

Use a local counter instead of the global 'event' variable for the
readdir() optimization.

Depends on patch-event-II

Background:
  The only user of i_version and f_version in ext2 is
  ext2_readdir(). As an optimization, ext2 performs the
  validation of the start position for readdir() only if
        flip->f_version != inode->i_version.
  If there was no llseek and no directory change since the
  last readdir() call, then f_pos can be trusted.
  f_version is set to 0 in get_empty_flip and during llseek.
  Right now, i_version set to ++event during ext2_read_inode
  and commit_chunk, i.e. at inode creation and if a directory
  is changed.
  Initializing i_version to 1, and updating with i_version++
  achieves the same effect, without the need of a global variable.
  Global uniqueness is not required, there are no other uses
  of [if]_version in ext2.

Change relative to the patch you have right now:
i_version is initialized to 1 instead of 0. For ext2 it's doesn't
matter [there is always a valid 'len' value at the beginning of a
directory data block], but it's cleaner.

Patch from Manfred Spraul

f_version and i_version are used by filesystems to check if it can
reuse the f_pos position across readdir calls without validation.

Right now f_version and i_version are modified by
    f_version = ++event;
    i_version = ++event;
    if (f_version != i_version) goto revalidate
and event is a global, exported variable.

But that's not needed,
    f_version  = 0;
    i_version++;
    if (f_version != i_version) goto revalidate
works too, without the ugly 'event' variable.

I got an ok from viro, and I had notified the fs maintainers, no
complaints either

- block_dev.c, block_llseek updates f_version to '++event'.
   grep showed that no device driver uses f_version, this is dead
   code copied from the default llseek implementation.

- the llseek implementations and get_empty_flip set f_version
   to '++event'
   This is not dead code, but
            filp->f_version = 0
   achieves the same effect:
   f_version is used by the readdir() implementation of several
   filesystems to skip the revalidation of f_pos at the beginning
   of a readdir call: If llseek was not called and the filesystem
   did not change since the last readdir call, then the value in
   f_pos can be trusted.
   The implementation (for example in ext2) is
     inode->i_version = ++event;
   in all operations that change a directory
   At the beginning of file_operation->readdir():
     if(inode->i_version != flip->f_version)
                revalidate();
     filp->f_version = inode->i_version;
   There are other users of f_version, but none of them use the
   default llseek implementation (e.g. fs/pipe.c)

Both the /proc/sys/cpu/ and /proc/cpufreq interface can safely be
enabled in the same kernel. This simplifies the transition to the newer
interface. Only minor updates are needed in order to allow this to be
done.

Fix:
 - Makefile using obj-m directly, rather than subsituting the "m"
 - compiler breakage against older versions of gcc
 - adds some lacking return statements that gcc didn't catch
 - removes some 2.4 compatibility stuff
 - RTT calculation
 - puts the timeouts in terms of HZ rather than assuming HZ==100

A supplement to manfred's patch; remove copy_segments, release_segments
and even forget_segments from all architectures (except x86-64 since
Andi wants to do that seperately):

Did the changes that created this line actually build for anyone?
:-)

mbcache should be too. 

into home.transmeta.com:/home/torvalds/v2.5/linux

into home.transmeta.com:/home/torvalds/v2.5/linux

This patch changes how media-change is detected to use a mechanism more
similar to what 'popular' OSes use.

The motive for this change is that more and more emulated-SCSI devices
(sbp2, usb-storage, etc) don't support START_STOP unless they need it.
Other OSes won't send a START_STOP unless the device reports NOT_READY,
indicating that it needs an explicit command to start up.

One could argue that these devices are out-of-spec, but they are so common
that they basically define the de facto spec.

Basically, any device that needs a START_STOP should still get it, but ones
that don't need it won't.

The i386 LDT code had it's own set of arch hooks (??_segments), I've
replaced most of them with the mmu context hooks in a previous patch.
The attached patch completes that change: replace release_segments with
destroy_context.

The patch is part of the -ac kernels in 2.4.  The patch breaks x86-64,
Andi Kleen promised to send you the corresponding
s/release_segments/destroy_context/ patch.

Cleaning simplified by descending down in aicasm when cleaning.

Fixed firmware build, someone made a spelling mistakei (aix -> aic).

No longer include Rules.make.

Made the *docs targets work after the restructuring of the kbuild files.

Fix cleaning up after make htmldocs, there are some directories
that needs to be deleted in that case.

This patch updates the p4-clockmod.c driver to correctly manage
HyperThreading-enabled Pentium IVs as well as those models which do
not support HyperThreading - thanks to Venkatesh Pallipadi for
explaining cpu_sibling_map to me. Additionally, an EXPORT_SYMBOL
was missing. (spotted by Marc-Christian Petersen - thanks!)

Update to the latest parport_gsc in the PA tree.

Some misc updates:
 - adapt our config.in changes to arch/parisc/Kconfig
 - general Makefile updaes for the new build system
 - remove asm-parisc/gsc.h in favour of asm-parisc/io.h
 - Alan-approved fix for the loff_t problem
 - Define POLLREMOVE like other architectures.
 - irq handling updates from Grant Grundler

x86 doesn't evaluate the argument to flush_dcache_page so this compilation
problem has gone overlooked:

SOM is the file format used for HPUX binaries.

This huge patch moves a bunch of drivers from arch/parisc/kernel to
drivers/parisc and adds some new drivers in drivers/parisc.

Some might argue that HPUX already has quite enough of a personality.

Build fixes, PA64 fixes, some new hpux syscalls, VFS fixes.

into home.transmeta.com:/home/torvalds/v2.5/linux

M68k: Remove obsolete cruft

Add Sun-3 DVMA debugging code (currently disabled, from Sam Creasey)

Zorro: trivial patch to use loff_t and not int (from Silvio Cesare
<silvio@qualys.com>)

Add Sun-3 VME support (by Sam Creasey):
  - Added VME version of the Sun-3 NCR5380 scsi driver.
  - Modified the Sun-3 dvma routines to support VME.
  - Added sun3_map_test() -- uses the ex_table to trap faults on VME mappings
    in order to determine if a card is present.  This seems a little hackish to
    be, but:
      1) All changes are entirely within Sun3-only code paths,
      2) netbsd, sunos, and mach all use this mechanism for VME probes anyway.

Update m68k for the change of xtime from struct timeval to struct timespec,
which has a higher resolution (ns vs. us), in 2.5.35.

Fix rwsemtrace() message

Add ElBox Computer FastATA 1200/4000/ZIV to the Zorro ID database (from
zorroutils)

General cleanup/updating of purely sun3/3x code (from Sam Creasey)
  - arch/m68k/sun3(x):
      o Cleaned up prom code slightly
      o Added reporting of prom revision in /proc/hardware
      o Removed some warnings
  - Use common idprom code for both Sun-3 and Sun-3x
  - Sun-3x idprom updates
  - drivers/scsi/sun3*: updated scsi drivers to reflect new struct scatterlist.