into home.osdl.org:/home/torvalds/v2.5/linux

All the users of this function in the SCSI tree call it with the host
lock held.  With the new list traversal code, it was trying to take
the lock again to traverse the list.

Fix it to use the unlocked version of list traversal and modify the
header comments to make it clear that the lock is expected to be held
on calling it.

into home.osdl.org:/home/torvalds/v2.5/linux

I've been looking at enforcing lifetime phases for SCSI devices
(primarily to try to get the mid layer to offload as much of the device
creation and hotplug pieces as it can).

I've hijacked the sdev_state field of the struct scsi_device (formerly
this was a bitmap, now it becomes an enumerated state).

I've also begun adding references sdev_gendev into the code to pin the
scsi_device---initially in the queue function, but possibly this should
also be done in the scsi_command_get/put, the idea being to prevent
scsi_device freeing while there's still device activity.

The object phases I identified are:

1. SDEV_CREATED - we've just allocated the device.  It may respond to
internally generated commands, but not to user ones (the user should
actually have no way to access a device in this state, but just in
case).

2. SDEV_RUNNING - the device is fully operational

3. SDEV_CANCEL - The device is cleanly shutting down.  It may respond to
internally generated commands (for cancellation/recovery) only; all user
commands are errored unless they have already been queued (QUEUE_FULL
handling and the like).

4. SDEV_DEL - The device is gone. *all* commands are errored out.

Ordinarily, the device should move through all four phases from creation
to destruction, but moving SDEV_RUNNING->SDEV_DEL because of surprise
ejection should work.

It's starting to look like the online flag should be absorbed into this
(offlined devices move essentially to SDEV_CANCEL and could be
reactivated by moving to SDEV_RUNNING).

I haven't altered the similar bitmap model that scsi_host has, although
this too should probably move to an enumerated state model.

I've tested this by physically yanking a module out from underneath a
running filesystem with no ill effects (other than a slew of I/O
errors).

The obvious problem is that this kills possible user error handling, but
we don't do any of that yet.

This patch is against scsi-bugfixes-2.6. I updated it based on comments
received. It breaks up the reference count initialization for scsi_device
and restores calling slave_destroy for all scsi_devices configured or
not. I ran a small regression using the scsi_debug, aic7xxx, and qla2xxx
driver. I also had a debug patch for more verbose kobject cleanup and
patch for a badness check on atomic_dec going negative (previously
provided by Linus).

The object cleanup appears to being functioning correctly. I only saw
previously reported badness output:
	- Synchronizing SCSI cache fails on cleanup.
	- scsi_debug.c missing release (I believe Doug posted a patch)
	- aic7xxx warnings on rmmod due to ahc_platform_free calling
	  scsi_remove_host with ahc_list_lock held.


This patch splits the scsi device struct device register into init and
add. It also addresses memory leak issues of not calling slave_destroy
on scsi_devices that are not configured in.

Details:
* Make scsi_device_dev_release extern for scsi_scan to use in
alloc_sdev.
* Move scsi_free_sdev code to scsi_device_dev_release. Have
previous callers of scsi_free_sdev call slave_destroy plus put_device.
* Changed name of scsi_device_register to scsi_sysfs_add_sdev to
match host call and align with split struct device init.
* Move sdev_gendev device and class init to scsi_alloc_sdev.

Thu Nov 20 22:56:11 PST 2003

 drivers/scsi/scsi_priv.h  |    4 +-
 drivers/scsi/scsi_scan.c  |   63 +++++++++++++++++++++-------------------------
 drivers/scsi/scsi_sysfs.c |   58 ++++++++++++++++++++++--------------------
 3 files changed, 62 insertions(+), 63 deletions(-)

It turns out that the SiS irq routing logic doesn't go by chipset
after all - it's just that some pirq entries are "legacy" numbers,
while others are raw offsets into PCI config space (and the legacy
numbers are more commonly used with the older chipsets, which
explains the correlations).

This simplifies the router code substantially.

A bug prevents the PnP layer from reserving some of the resources
specified by the PnPBIOS.  As a result some systems will have
unpredicable (random crashes etc.) problems because of resource
conflicts, especially when PCMCIA support is enabled.  This patch
fixes the problem by ensuring that the proper resource data is
reserved.

into home.osdl.org:/home/torvalds/v2.5/linux

into nuts.ninka.net:/disk1/davem/BK/sparc-2.5

It did not account for extension headers properly.  If we get
this length wrong, we do not determine if a multicast packet
is MLDv1 vs. MLDv2 correctly.

into nuts.ninka.net:/disk1/davem/BK/net-2.5

This caused the system call code to test for the wrong number of
system calls, with resulting exciting results. 

into home.osdl.org:/home/torvalds/v2.5/linux

into home.osdl.org:/home/torvalds/v2.5/linux

	for 2.6.0.  The proper fix is to replace ia64_ni_syscall with
	sys_ni_syscall, but that would make the patch quite large, so
	we defer that till 2.6.1.

	restore_sigcontext().  Also update ia32 subsystem accordingly.

	in practice, but it's clearly wrong and just waiting there to
	get triggered...

Caught by Francois Romieu.

RFC2710 says:
1) MLD messages are never sent for multicast addresses whose scope is 0
   (reserved) or 1 (node-local).
2) MLD messages ARE sent for multicast addresses whose scope is 2
   (link-local), including Solicited-Node multicast addersses [ADDR-ARCH],
   except for the link-scope, all-nodes address (FF02::1).

      The current MLDv1 code does not send reports for link-scope addresses
and doesn't restrict scope 0. This may break switches that snoop reports for
determining which ports should receive particular addresses.  Patch below.

into nuts.ninka.net:/disk1/davem/BK/net-2.5

into nuts.ninka.net:/disk1/davem/BK/sparc-2.5

Fix the creation of the /proc/net/bonding dir.
Patch is against 2.6.

Amir

This patch (against latest 2.6.0) is also waiting for almost a month.
It's already in 2.4 but is still very much needed for 2.6.

Old ifenslave versions (like the one in Red Hat 9) don't work with
the bonding module in the latest 2.6 kernel without it.

Amir

EBUS DMA fixes:
- Add EBUS_DMA_FLAG_TCI_DISABLE that the client can set if it
  wants only to hear device interrupts, not DMA complete ones.
- When initially resetting the EBUS DMA unit, put valid burst etc.
  encodings in the CSR register.  Not doing this appears to leave
  the attached ISA device in a weird state.

PCI FLOPPY fixes:
- Do ebus_dma_enable() before ebus_dma_request()
- In sun_pci_fd_set_dma_mode() do not forget to set the direction.
- Set EBUS_DMA_FLAG_TCI_DISABLE in ebus_dma flags.
- Make sure that in error paths sun_fdc/FCD1 will be -1 (invalid).

Thanks to Soyoung Park for loaning her Ultra5 to me so I could fix this.

Original report and bug fix from: Amit Patel <patelamitv@yahoo.com>

The problem was a signed vs unsigned char problem when computing luns
for scanning.

From: Adrian Bunk <bunk@fs.tum.de>

Modular BINFMT_ELF doesn't build, and is pretty pathological anyway.

So just make it a boolean rather than a tristate.

From: Herbert Xu <herbert@gondor.apana.org.au>

The recent patch produces a message with no terminating newline on the
machine in question.  This is because one of the four bytes that you're
printing out is NUL.  The following patch avoids that problem.

From: Jeremy Higdon <jeremy@classic.engr.sgi.com>

I believe there is a bug in kernel/resource.c.

We (SGI sn2 I/O code) are using this for allocating dma map resources, and
we tracked failures we were seeing to find_resource().

The problem is that when testing bounds in the forever loop, the end bound
would be one higher than it should be if it gets set from another resource
(it's set to the proper value when it gets set from the root), causing
find_resource to return an invalid min/max when the requested size was one
greater than would fit between two existing resources.

From: Andreas Gruenbacher <agruen@suse.de>

64-bit pointer arithmetic bug in xattr code

The int offset is not enought to hold the difference between arbitraty
pointers on 64-bit machines.  Compute the offset of here and last inside
HDR(bh) instead.

From: James Cleverdon <jamesclv@us.ibm.com>

On summit-based machines the cpu_sibling_map data has been hosed for some
time.  I found out why in Intel's IA-32 Software Deveveopers' Manual Vol 2
under CPUID.  Looks like the value that cpuid returns is the one latched at
reset, and doesn't reflect any changes made by the BIOS later:

  * Local APIC ID (high byte of EBX)--this number is the 8-bit ID that
    is assigned to the local APIC on the processor during power up.  This
    field was introduced in the Pentium 4 processor.

Also, the code in init_intel was a bit overdesigned.  Until Intel releases
a chip with a non-power-of-2 sibling count on it, there's no point in all
that bit bashing.

From: Jun Sun <jsun@mvista.com>

It is needed for all those "__attribute_used__" etc to be valid.

Also, it seems that when compiling a file ending in ".S", gcc-2.95.3 does not
expand __GNUC__ at all.  This causes the compiler version check to fail when
building vsyscall.S.  So add __ASSEMBLY__ wrappers in there.