Linux Mama 2.1 - Unsorted Patches - Week 41

Author: David S. Miller
Kernel-Version: 2.1.57
Date: 971111
Type: Enhancement
Download: ftp://ftp.kernel.org/pub/linux/kernel/davem/

1) Usual assortment of sparc64 updates and fixes (me, Jakub, Eddie).

2) Ditto for PPC (Cort, Geert).

3) First shot at full merge of new sound driver code (Alan).

4) Make linux/types.h hopefully more kosher in all libc
environments (me).

5) AFFS updates/fixes by Hasn-Jaochim Widmaier (merged by Jes).

6) Lots of small multicast fixes (Alexey).

7) Things should compile with transparent proxy enabled again, fixed
a TCP doozy which could result in sockets stuck in FIN_WAIT2, and
finally clean up the arp code a bit. (me)

8) Various SRM and module symbol fixups on Alpha (Richard).

9) Latest Becker 3c59x.c driver is in our tree now (merged by Alan).

10) Do all kernel level IP autoconfiguration generically, so it is now
ripped out of the nfsroot code and can be used in other ways. See
ChangeLog for full details (Martin).

11) Fork fixes from Bill Hawes, should kill the svcthread/lockd/nfsd
and squake crashes that were reported (merged by me).

Author: Steven N. Hirsch
Date: 971010
Type: Bugfix
Download: xconfig_57.gz

This fixes a problem which occurs running xconfig as a standalone program
when sound is configured in the kernel:

Author: Bill Hawes
Kernel-Version: 2.1.57
Date: 971009
Type: Bugfix
Download: smbfs_make57-patch.gz

herman dumont wrote:
> the latest patch gives me unresolved modules:
> lib/modules/2.1.57/fs/smbfs.o: unresolved symbol(s)
> smb_add_to_cache
> smb_init_dircache
> is_bad_inode
> make_bad_inode
> smb_invalid_dir_cache
> shrink_dcache_sb
> smb_readdir
> truncate_inode_pages

OK, part of the problem is the new cache.c code isn't being compiled.
The patch should have included a change to the Makefile, as in attached
patch shows.

Some of the other symbols require previous patches posted here ..

Author: Bill Hawes
Kernel-Version: 2.1.57
Date: 971009
Type: Enhancement
Download: nfs_57-2-patch.gz

I've attached a snapshot of the 2.1.57 nfs client patch. Changes this
time include:

(1) Update dentry d_time following a successful operation, to cut down
on invalidates for proven dentries.

(2) Code to free the dircache pages when unmounting a volume. The
previous code would keep the pages until the module unloads, or forever
if nfs is built into the kernel.

(3) Changes to the nfs_entry structure to make it consistent for all
architectures. Since these are being packed into a page, we need them
to pack consistently.

(4) A check for the maximum allowable buffer size for readdir, in order
to guarantee that the worst case return will still fit into the
receiving buffer. The will guarantee that the dreaded "this shouldn't
happen" message will in fact never happen :-) The calculation is a
little too conservative now, but I want to make sure it always works.

Please give this a test and let me know of any problems.

Author: Pavel Machek
Kernel-Version: 2.1.57
Date: 971008
Type: Enhancement
Download: kbd_blank_console_57-patch.gz

This patch offers possibility to blank monitor from keyboard, which is
IMO nice and should be there. It is relative to 2.1.57. It needs patch
to kbd- package (and I do not know who should I send that to), patch
to kbd package also makes adding new features easier.

Author: Ingo Molnar
Kernel-Version: 2.1.57
Date: 971008
Type: Bugfix
Download: internal_add_timer_57-patch.gz

this is my proposed solution, patch against 2.1.57, all of sched.c is
converted to the new usage, as a demonstration.

It solves the overflow problem, has no type cast in the main code, is fast
(2 instructions) in assembly and is safe against most types of C typos.
(Personally i find it more readable as well, but thats a matter of taste.)

Author: Bill Hawes
Kernel-Version: 2.1.57
Date: 971008
Type: Bugfix
Download: fork_57-patch.gz

The attached patch for kernel/fork.c fixes a couple of problems.

I traced the cause of two different oops reports(one in /proc, the other
with squake in vm_area_pages) back to the small window in time when a
new task has been created but hasn't yet allocated new page tables. If
the mm->pgd field is left NULL here, some code that looks at other
tasks' fields may do a pgd_offset() without checking for NULL pgd. This
was happening in /proc get_phys_address and in vmalloc when it updates
the page dir for every task.

The fix is very simple -- mm->pgd is left at the parent task's value
until the new_page_table call, and set to NULL on failure so that
mmput() stays happy.

The other fix is a test in copy_files for NULL task->files; this allows
a background task to call exit_files() and later fork a new process.

Author: Bill Hawes
Kernel-Version: 2.1.57
Date: 971008
Type: Enhancement
Download: smbfs_57-2-patch.gz

The attached patch is a snapshot of the latest 2.1.57 smbfs. In
addition to some bugfixes, I've completely rewritten the directory
cacheing code, and it now uses the page cache to hold the cached
directories. This has some major advantages, such as (1) ability to
cache as many directories as you need, (2) automatic invalidation on
inode changes, (3) system-friendly memory management, and (4) guaranteed
invalidation at unmount time.

I've written the code with the intent to generalize it for other
(network) filesystems, (nfs, ncpfs, ??). It's not quite at the
"generic_dir_read" point, but should prove to be a good start.

Another noteworthy change is a new "dentry renew" routine to update the
dentry timestamp after successful operations. This greatly cuts down on
attempted invalidations. I've also added this to the nfs client code.

This patch incorporates only the changed smbfs files, plus one small
patch for mm/filemap.c to handle some of the page cache issues. If
you're already using smbfs successfully, this patch should apply cleanly
to 2.1.57 and not give an new undefined symbols (I hope).

Author: Kurt Garloff
Kernel-Version: 2.1.57
Date: 971008
Type: New Driver
Download: dc390-1.11.tar.gz

I sent these patches to the Linux-SCSI list before, but I'm not sure if
somebody picks them up there and merges them into the kernel.

I recognized that the Tekram DC390 SCSI driver is in both 2.1.5x and
2.0.30-preX kernels, so it would be better to include the latest version
which is from Feb97(!). It fixes a problem with large disks.
The patch is applicable to both 2.0.31preX and 2.1.57 kernels.
(This is the first file attached to the message.)

I used to package the DC390 driver together with the DC390WUF drivers and
wrote some lines about not confusing those two and using the right driver.

This no longer holds, as we do not need the DC390WUF driver any more.
The latest ncr5x8xx drivers (2.4 and up) perfectly work together with
DC390WUF, so the documentation should say that and not to use an
unexisting driver.
(The second file ...)

Author: David S. Miller
Date: 971007
Type: Bugfix
Download: trans_proxy_snapshot_fix.gz

This makes it compile once again. Things work without any
redirections enabled, as to whether those work I don't know yet.

Author: Matthew Kirkwood
Date: 971007
Type: Enhancement
Download: mem-dead-code-remove.gz

Having lost my net access for the summer, I was reduced to
reading through various bits of kernel source, and I found
the following two little opportunities to shrink the kernel
(though not by very much at all ;-)

Author: Terje Malmedal
Kernel-Version: 2.1.57
Date: 971007
Type: Enhancement
Download: disk_stat_increase_57.gz

Linux will by default not report statistics for more than the four
first disks in the system. This patch increases that to eight. Which
is what I need. If anybody needs more all they need to do is to change
DK_NDRIVE in kernel_stat.h

I'm also including a patch against ftp.feral.com:/pub/linux/iostat.c
to make it understand the new format.

Author: David Woodhouse
Kernel-Version: 2.1.57-vger-971005
Date: 971007
Type: Enhancement
Download: dev_alloc-clear-vger57.gz

Given that the only place dev_alloc seems to be used in the entire kernel
source is in net/tunnel.c, which is what I'm working on, I feel justified in
changing the semantics a bit.
OK?