Skip to contentSkip to conversation

Linux 2.6.9-ac16

DISCUSSION Alan Cox 8 participants 14 messages Dec 16, 2004 – Dec 20, 2004

lore
Share & tools
lore

Alan Cox

author
#0 lore raw

Further small fixes for different minor things. A merge of some of the small cleanups from Fedora work and also the fixes for the igmp and vc holes.

Arjan van de Ven is now building RPMS of the kernel and those can be found in the RPM subdirectory and should be yum-able. Expect the RPMS to lag the diff a little as the RPM builds and tests do take time.

The HPT366 rework project is also not ready (its gone back to the drawing board until the current panic is over if you are a volunteer and wondered what is up).

ftp://ftp.kernel.org/pub/linux/kernel/people/alan/linux-2.6/2.6.9/

Key: o - only in -ac
* - already fixed upstream
X - discarded later as wrong
+ - ac specific (fix not relevant to non -ac)

2.6.9-ac16
o Fix sys5 semaphore wakeups (Manfred Spraul)
* Avoid maths overflow on VC resize (Georgi Guninski,
Linus Torvalds)
* Add D-Link DGE-528T to the r8169 PCI idents (Francois Romieu)
* Fix cciss ioctl problems (Mike Miller)
* Fix wrong IRQ problems with DAC960 (Bjorn Helgaas)
* Minor typo fix in cdrom driver (efalk@google)
* Add 3PARdata SCSI idents for sparse lun (Castor Fu)
o Fix Paul Laufer's email address (Paul Laufer)
* More ATI IDE PCI identifiers (Enrico Scholza)
* Fix some /proc oopses (Manfred Spraul)
o Fix build of CS461x gameport (Adrian Bunk)
* Get CAPI isdn back working on SMP (Karsten Keil)
o Fix AT2701FX AMD PCnet32 on fibre (Guido Guenther)
* Fix igmp source filter counting/corruption (Chris Wright)
* Fix missing SHM_LOCK priviledge checks (Hugh Dickins)
* Clear %ebp on sysenter based syscall exit (Brad Spender)
o Suggest irqpoll when we get screaming irqs (Alan Cox)
o Fix misleading microcode message (Arjan van de Ven)
o Don't warn on scsi ioctl kmalloc fail (Arjan van de Ven)
o Acenic must use __devinitdata for hotplug (Alan Cox)
| based on an RH patch
o Add a 1620 byte slab cache for ethernet frames (Arjan van de Ven)
o Fix 64bit build of kconfig (Arjan van de Ven)
o Add more AC97 table data

2.6.9-ac15
o Kill "open failed" cdrom message. (Alan Cox)
| This is a natural event from code poking around | doing CD detection etc
o Fix crash with aacraid double complete (Mark Salyzyn, Tom Coughlan,
Alan Cox)
* Fix oops after unload ip_conntrack_standalone
o Allow cross compile of x86_32 kernel on x86_64 (Arjan van de Ven)
* Fix a small leak in the ip option code (Dave Miller)
* Fix a small leak in ip_conntrack_ftp (Dave Miller)
* Fix control message validation (Herbert Xu, Georgi Guninski)
* Backported key xfrm patches (Patrick McHardy, Dave Miller,
Herbert Xu)
* Backport vm OOM fix (Marcelo Tosatti)
o Make sx8 naming follow LANANA (Jeremy Katz)
o Fix getblk_slow hang (Chris Mason)
o Watchdog support for early cobalt ALi hardware (Mike Waychison)
o Fix cable detect on older promise (Bartlomiej Zolnierkiewicz)
o Fix vm86 irq handling (Pavel Pisa)

2.6.9-ac14
o Use LBA28 if a drive does LBA48 but rejects (Alan Cox)

	SET_MAX_EXT and friends (Maxtor 4R120L4 ..)

* Make sure md uses rdev_dec properly (Neil Brown)
* Remove bogus WARN from futex v signal (Rusty Russell)
* Fix via-rhine crash on resume (Randy Dunlap)
* Fix x86+64 IRQ loss bug on uniprocessor (Petr Vandrovec)
* Don't print undefined sibling data on em64t (Venkatesh Pallipadi)
* Further overlapping vma/exec fixes for 64bit (Zou Nan hai)
boxes
* Re-enable escaped ACPI debug (Len Brown)
* SELinux hook for ia32 emulation fd pass check (Mitchell Blank Jr)
* Fix DECnet oops in accept (Patrick Caulfield)
* Fix cciss oops with config tools (Andrew Patterson)
* io_edgeport oops fix (Greg Kroah-Hartmann)
* Fix leak in bio_copy error path (Jens Axboe)
o Fix buffer sizes on ia32 emulation for x86_64 (Jeremy Fitzhardinge,
Chris Wright)
o Initial patch for ide_abort hang (Alan Cox)
o Fix serveral ide timing violations on reset (Alan Cox)
+ Fix dumb bug in ALi/ULi driver (Alan Cox)
+ Fix /proc/ide/*/capacity bug (Alan Cox)
o Support CSB6-R Serverworks raid (Alan Cox)
o Teach ide-cd to use sense data for file system (Alan Cox)
requests
- This means you get better diagonstics on CD errors
- It means a partial I/O failure will get you back the ok sectors
- It may fix the problem some users have with ISO copying and ide-cd

2.6.9-ac13
Accidentally released 2.6.9-ac12 as this

2.6.9-ac12
o Configurable 100/1Khz clock for x86 (James Bottomley)
| 100Hz is great for battery life
o Fixup subordinate bus plumbing for boxes like
the HP zv5000z (somewhat experimental)
+ Fix ACPI=n build (Chuck Ehbert)
o EDD boot options (Matt Domsch)
* Fix proc reporting of ppid (Dinakar Guniguntula)
| Backport c/o Ingo
o Improved smbfs fixes (Chuck Ebhert)
o Fix reset problems with older 3c59x/3c90x (John Linville)
o Fix oops if aic79xx is loaded with no hardware (Adam Manthei)
present
o Token ring locking fix
* Fix smc9192 (Russell King)
* Refcount scsi command queue properly (James Bottomley)
* Fix CD-ROM returns from new TUR code (James Bottomley)
* Fix scribble after close on Nvidia ethernet (Manfred Spraul)
* Fix queuecommand() behaviour for USB disk (Alan Stern)
after unplug
* Fix xtime adjustment error that could cause (Andrew Morton)
clock slew
* Fix IPv6 fragmentation problems (Herbert Xu)
* Fix mm ref leak in proc/cmdline (Prasanna Meda)
* Fix a tiny race in the timer code (Ben Herrenschmidt)
* Fix crash if security_load_policy fails (Jeff Mahoney)
* Fix scsi queue return error in ide-scsi (Jens Axboe)

	and 3w-9xxx

* Avoid overflow in sg timeouts with large user (James Bottomley)
provided timeout
* Fix statm accounting (Hugh Dickins)
| Backported by Jason Baron
* Fix mlock v VM_IO hang (Hugh Dickins)
* Fix Intel i8x0 PCM audio with ALSA (Takashi Iwai)
* Fix further visor fix problems (Roger Luethi)
o New ATI IXP PCI identifier (Pascal Lengard)
o Fix URL for lanana (Alexander Stohr)
* Fix via_irqpic to not be __devinit (Andi Kleen)
* PPP multilink fixes (Paul Mackerras)
* Backport OSS emulation fixes (Marcel Sebek)

2.6.9-ac11
o Allow users to force it8212 out of raid mode (David Howells)
o On some platforms the flashing keylights (Alan Cox)
riggers bogus keyboard warnings. The error appears from other stuff too like keyboard switches so kill it
o Add support for ALi 5228 PATA IDE (Clear Zhang)
o Add support for newer ALi AGP (Clear Zhang)
* Seqpacket and SELinux fix (James Morris)
o Fix ide /proc and legacy devices problem (Alan Cox)

2.6.9-ac10
o Fix tiny ide-cd race (Alan Cox)
* Remove an apparently bogus IDE CD blacklist (Srihari Vijayaraghavan)
for a drive confirmed to work
o Correct vendor of some Cyrix chipsets in docs
* Error path locking fix for appletalk (Andries Brouwer)
* Remove a suspicious __initdata (Andries Brouwer)
* Further binfmt_elf work (Jakub Jelinek)
o SMBfs overflow fixes (Stefan Esser)
| Collated together by Juan Quintela
* Don't spew debug on bioses with ACPI (Len Brown)
breakpoints left in them.
* Further NFS disconnected dentry fix (Neil Brown)
* tmpfs inode accounting leak fix (Hugh Dickins)
* x86-64: Fix user triggerable oops on debugger (Andi Kleen)
vsyscall page access
* Serialize datagram read on AF_UNIX (Dave Miller)
* Disable PnP BIOS when using ACPI (Adam Belay)
* Fix oops in visor driver caused by DoS fixes (Roger Luethi)
o Fix several IDE drivers that assumed > 0 was (Alan Cox)
also an error return for pci probe functions
* Backport netlink updates/fixes from 10rc2 (Herbert Xu,
Dave Miller)
* Backport multihop routing oops fix (Christian Ehrhardt)
* Backport TSO fixes (Herbert Xu)

2.6.9-ac9
* Linus moved the remap_page_range flag fixes (Linus Torvalds)
into the function. Now this has had some
testing do the same in -ac and shrink the
diff a lot
* Fix low memory oops in device mapper (N Cunningham)
* Fix duplicate kfree in dm-target error path (Alasdair Kergon)
* Use a new bio on a md retry (Neil Brown)
+ Fix mediabay compile (Alan Cox)
* Increase EDD array size (Matt Domsch)
* Fix locking error/DoS in k15kusb105 (Greg Kroah-Hartmann)
* Fix locking hang on error path of whiteheat (Greg Kroah-Hartmann)
* Use sector_t for md (fixes some large raids) (Neil Brown)
* Fix further USB locking errors (Greg Kroah-Hartmann)
* Report the right thing on a pnpbios fault (Andy Whitcroft)
* Add PCI quirk for VIA audio (David Shaohua Li)
* Fix neighbour table counter atomicity (Herbert Xu)
* Error out on early exec before rootfs (Chris Wright)
* Fix a.out crash with junk binary and (Chris Wright)
virtual memory limits disabled
* Import atomic_int_return for the neighbour fix (KaiKai Kohei)

2.6.9-ac8
* Fix binfmt_exec partial read problem (Chris Wright)
* Fix E820 overflow on x86-64 as per x86-32 (Andi Kleen)

2.6.9-ac7
o Redo the fixups in siimage/it8212 so they (Alan Cox)
always actually work
o Fix up both drives on an IT8212 raid (Alan Cox)
* Remove a debug printk/2 sec wait from CS5520 (Alan Cox)
* Move partial decode test to ide-cs/delkin only (Alan Cox)
* Fix partial decode test for no serial number (Alan Cox)
o Add support for disks on early rev IT821x (Alan Cox)
+ Allow ide-disk to be modular again (Tomas Szepe)
* Further fixup fixes (Bartlomiej Zolnierkiewicz)
* Apple Ipod-mini size reporting fix (Avi Kivity)
* Initial (non SMP) cdu31a driver rescue (Ondrej Zary)
o Allow READ_BUFFER_CAPACITY to SG_IO users (Daniel Drake)

2.6.9-ac6
+ Fix problem with -ac5 msdos changes (Vojtech Pavlik)

2.6.9-ac5
o Fix oops in and enable IT8212 driver (Alan Cox)
o Minor delkin driver fix (Mark Lord)
* Fix NFS mount hangs with long FQDN (Jan Kasprzak)
| I've used this version as its clearly correct for 2.6.9 | although it might not be the right future solution
o Fix overstrict FAT checks stopping reading of (Vojtech Pavlik)
some devices like Nokia phones
* Fix misdetection of some drives as MRW capable (Peter Osterlund)
* Fix promise 20267 hang with very long I/O's (Krzysztof Chmielewski)
* Fix a case where serial break was not sent for (Paul Fulghum)
the right time.
* Fix S/390 specific SACF hole (Martin Schwidefsky)
* NVidia ACPI timer override (Andi Kleen)
o Correct VIA PT880 PCI ident (and AGP ident) (Dave Jones)
* Fix EDID/E820 corruption (Venkatesh Pallipadi)
* Tighten security on TIOCCONS ([email protected])
* Fix incorrect __init s that could cause crash (Randy Dunlap)

2.6.9-ac4
* Fix minor DoS bug in visor USB driver (Greg Kroah-Hartmann)
o Delkin cardbus IDE support (Mark Lord)
+ Fix SMP hang with IDE unregister (Mark Lord)
* Fix proc file removal with IDE unregister (Mark Lord)
* Fix aic7xxx sleep with locks held and debug (Luben Tuikov)
spew
o First take at HPT372N problem fixing (Alan Cox)

2.6.9-ac3
* Fix syncppp/async ppp problems with new hangup (Paul Fulghum)
* Fix broken parport_pc unload (Andrea Arcangeli)
* Security fix for smbfs leak/overrun (Urban Widmark)
* Stop i8xx_tco making some boxes reboot on load (wim@iguana)
* Fix cpia/module tools deadlock (Peter Pregler)
+ Fix missing suid_dumpable export (Alan Cox)

2.6.9-ac2
+ Fix invalid kernel version stupidity (Adrian Bunk)
* Compiler ICE workaround/fixup (Linus Torvalds)
* Fix network DoS bug in 2.6.9 (Herbert Xu)
| Suggested by Sami Farin
o Flash lights on panic as in 2.4 (Andi Kleen)

2.6.9-ac1

Security Fixes
* Set VM_IO on areas that are temporarily (Alan Cox)
marked PageReserved (Serious bug)
o Lock ide-proc against driver unload (Alan Cox)
(very low severity)

Bug Fixes
o Working IDE locking (Alan Cox)
| And a great deal of review by Bartlomiej
o Handle E7xxx boxes with USB legacy flaws (Alan Cox)

Functionality
o Allow booting with "irqpoll" or "irqfixup" (Alan Cox)
on systems with broken IRQ tables.
o Support for setuid core dumping in some (Alan Cox)
environments (off by default)
* Support for drives that don't report geometry
o IT8212 support (raid and passthrough) (Alan Cox)
o Allow IDE to grab all unknown generic IDE (Alan Cox)
devices (boot with "all-generic-ide")
o Restore PWC driver (Luc Saillard)

Other
* Small pending tty clean-up to moxa (Alan Cox)
* Put VIA Velocity (tm) adapters under gigabit (VIA)

--
Subliminal URLS: www.no2id.org, www.magnatune.com, www.ntk.net
www.machinaesupremacy.com www.bzflag.org

#1 lore raw
Quoted 2 lines · Alan Cox

o Acenic must use __devinitdata for hotplug (Alan Cox)
| based on an RH patch

diff -u --new-file --recursive --exclude-from /usr/src/exclude linux.vanilla-2.6.9/drivers/net/acenic.c linux-2.6.9/drivers/net/acenic.c
--- linux.vanilla-2.6.9/drivers/net/acenic.c 2004-10-20
23:16:54.000000000 +0100
+++ linux-2.6.9/drivers/net/acenic.c 2004-12-16 17:13:10.799818288
+0000
@@ -444,7 +444,7 @@
MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used
for TX/RX descriptors (range 0-63)");

-static char version[] __initdata =
+static char version[] __devinitdata =
"acenic.c: v0.92 08/05/2002 Jes Sorensen, [email protected]\n" "
http://home.cern.ch/~jes/gige/acenic.html\n";

you broke this one... :-)
the version var *cannot* be initdata of any kind, since the ethtool ioctl uses the variable. End Of Story(tm)

#2 lore raw

Arjan van de Ven <[email protected]> :

Quoted 3 lines · Alan Cox
Quoted 2 lines

o Acenic must use __devinitdata for hotplug (Alan Cox)
| based on an RH patch

[...]

Quoted 9 lines

MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used
for TX/RX descriptors (range 0-63)");

(quoted diff) +1 -1
-static char version[] __initdata =
+static char version[] __devinitdata =
   "acenic.c: v0.92 08/05/2002  Jes Sorensen, [email protected]\n"
   "http://home.cern.ch/~jes/gige/acenic.html\n";
Quoted 3 lines

you broke this one... :-)
the version var *cannot* be initdata of any kind, since the ethtool ioctl uses the variable. End Of Story(tm)

Episode 2

Actually ethtool_drvinfo.get_drvinfo() does not use _this_ version var in the acenic driver. Btw it would be waaaaay too long for the 32 bytes allowed in the misc fields of struct ethtool_drvinfo) (I messed the r8169 driver this way).

--
Ueimor

#3 lore raw
In reply to Alan Cox

Alan Cox escreveu:

Quoted 2 lines

Further small fixes for different minor things. A merge of some of the small cleanups from Fedora work and also the fixes for the igmp and vc holes.

This kernel still suffers from the faulty OOM killing troubles of vanilla 2.6.9. Could you please pick up at least one of the recent fixes for this problem, such as as Rik van Riel's? As it stands -ac16 is unusable on a machine with as "little" as 64MB RAM, and bigger machines will just run into the same problems later.

For example the following happens when trying to compile UML on a 64MB
machine, whereas the same build completes without incident on 2.6.10-rc2-mm4

LD init/built-in.o
LD .tmp_vmlinux1
collect2: ld terminated with signal 9 [Killed]
make[2]: *** [.tmp_vmlinux1] Error 1
make[2]: Leaving directory `/home/chris/src/umlsim-65/kernel/linux-2.6.9' make[1]: *** [linux-2.6.9/vmlinux] Error 2
make[1]: Leaving directory `/home/chris/src/umlsim-65/kernel'
make: *** [all] Error 1

The entry in the system log for this is

Dec 17 13:46:39 sleepy oom-killer: gfp_mask=0xd2
Dec 17 13:46:40 sleepy DMA per-cpu:
Dec 17 13:46:40 sleepy cpu 0 hot: low 2, high 6, batch 1
Dec 17 13:46:40 sleepy cpu 0 cold: low 0, high 2, batch 1
Dec 17 13:46:40 sleepy Normal per-cpu:
Dec 17 13:46:40 sleepy cpu 0 hot: low 4, high 12, batch 2
Dec 17 13:46:40 sleepy cpu 0 cold: low 0, high 4, batch 2
Dec 17 13:46:40 sleepy HighMem per-cpu: empty
Dec 17 13:46:40 sleepy
Dec 17 13:46:40 sleepy Free pages:         244kB (0kB HighMem)
Dec 17 13:46:40 sleepy Active:12016 inactive:499 dirty:0 writeback:0 unstable:0
free:61 slab:1349 mapped:12003 pagetables:130
Dec 17 13:46:40 sleepy DMA free:60kB min:60kB low:120kB high:180kB active:11960k
B inactive:256kB present:16384kB
Dec 17 13:46:40 sleepy protections[]: 0 0 0
Dec 17 13:46:40 sleepy Normal free:184kB min:188kB low:376kB high:564kB active:3
6104kB inactive:1740kB present:49144kB
Dec 17 13:46:40 sleepy protections[]: 0 0 0
Dec 17 13:46:40 sleepy HighMem free:0kB min:128kB low:256kB high:384kB active:0k
B inactive:0kB present:0kB
Dec 17 13:46:40 sleepy protections[]: 0 0 0
Dec 17 13:46:40 sleepy DMA: 1*4kB 1*8kB 1*16kB 1*32kB 0*64kB 0*128kB 0*256kB 0*5
12kB 0*1024kB 0*2048kB 0*4096kB = 60kB
Dec 17 13:46:40 sleepy Normal: 0*4kB 1*8kB 1*16kB 1*32kB 0*64kB 1*128kB 0*256kB
0*512kB 0*1024kB 0*2048kB 0*4096kB = 184kB
Dec 17 13:46:40 sleepy HighMem: empty
Dec 17 13:46:40 sleepy Swap cache: add 2428, delete 2095, find 168/191, race 0+0
Dec 17 13:46:40 sleepy Out of Memory: Killed process 11648 (ld).

Regards,
Chris R.

#4 lore raw
Quoted 8 lines · Chris Ross

Alan Cox escreveu:

Quoted 2 lines

Further small fixes for different minor things. A merge of some of the small cleanups from Fedora work and also the fixes for the igmp and vc holes.

This kernel still suffers from the faulty OOM killing troubles of vanilla 2.6.9. Could you please pick up at least one of the recent fixes for this problem, such as as Rik van Riel's?

Can someone point me to his patch? I've been working on and off to try and get reasonable OOM behaviour. As it stands, 2.6.10-rc2-mm4 still shows nasty behaviour in OOM conditions, killing off more tasks than strictly required, and locking up the system for 10-15secs while doing it.

I'd be much happier doing a quick and dirty scan and knocking off something *now* rather than locking up the system. Surely it can't take 60 billion cycles of cpu time to pick a task to kill.

Chris

#5 lore raw

Hi Chris,

Chris Friesen escreveu:

Quoted 6 lines

As it stands, 2.6.10-rc2-mm4 still shows nasty behaviour in OOM conditions, killing off more tasks than strictly required, and locking up the system for 10-15secs while doing it.

I'd be much happier doing a quick and dirty scan and knocking off something *now* rather than locking up the system. Surely it can't take 60 billion cycles of cpu time to pick a task to kill.

Thomas Gleixner has been particularly interested the algorithms for deciding which task to kill (like me he got fed up with it picking the ssh daemon first).

See for example the thread at http://marc.theaimsgroup.com/?t=110189482200001&r=1&w=2

Some of the delay is by design: when OOM is reached we kill something off, wait a bit for the memory to be freed and become available to the system again, check whether now have enough memory, if not rinse and repeat. However, as I recall this is compounded by 2.6.9 having some nasty rentrancy problems causing the OOM killer to be called something like 100 times instead of once.

Perhaps Thomas could enlighten us as to the current state of play here?

Regards,
Chris R.

#6 lore raw
Quoted 18 lines · Chris Ross

Hi Chris,

Chris Friesen escreveu:

Quoted 7 lines

As it stands, 2.6.10-rc2-mm4 still shows nasty behaviour in OOM conditions, killing off more tasks than strictly required, and locking up the system for 10-15secs while doing it.

I'd be much happier doing a quick and dirty scan and knocking off something *now* rather than locking up the system. Surely it can't take 60 billion cycles of cpu time to pick a task to kill.

Thomas Gleixner has been particularly interested the algorithms for deciding which task to kill (like me he got fed up with it picking the ssh daemon first).

See for example the thread at
http://marc.theaimsgroup.com/?t=110189482200001&r=1&w=2

Quoted 8 lines

Some of the delay is by design: when OOM is reached we kill something off, wait a bit for the memory to be freed and become available to the system again, check whether now have enough memory, if not rinse and repeat. However, as I recall this is compounded by 2.6.9 having some nasty rentrancy problems causing the OOM killer to be called something like 100 times instead of once.

Perhaps Thomas could enlighten us as to the current state of play here?

Andrea fixed the invocation problem, which also handles the reentrancy problem in a clean way. It get's us rid of the ugly count, time, whatever mechanisms in out_of_memory which was designed to cover the invocation problem but was not able to prevent reentrancy and the resulting overkill (kill a random amount of processes even if enough memory is available).

I added the "Take child processes into account" modification for the whom to kill selection on top of that and I was not able to make it missbehave with my different test scenarios.

The patches are available in parts in this thread and the final combined patch is there:
http://marc.theaimsgroup.com/?l=linux-kernel&m=110269783227867&w=2

2.6.10-rc3 contains a partial fix for the erroneous invocation problem, but it is not as effective as Andrea's solution and it still runs into overkill once the oom mechanism is invoked.

Andrea's fix and the selection changes should go into 2.6.10, but I suspect that the VM gurus havent still reached a point, where they agree. I also have the feeling that the problem is partially ignored. Obviously has everybody plenty of memory in his boxes. </rant off>

Andrea's fix revealed some GFP_ flag related problems, which should be addressed seperately. Detailed explanation is in the mail thread.

tglx

#7 lore raw
Quoted 15 lines · Thomas Gleixner

Andrea fixed the invocation problem, which also handles the reentrancy problem in a clean way. It get's us rid of the ugly count, time, whatever mechanisms in out_of_memory which was designed to cover the invocation problem but was not able to prevent reentrancy and the resulting overkill (kill a random amount of processes even if enough memory is available).

I added the "Take child processes into account" modification for the whom to kill selection on top of that and I was not able to make it missbehave with my different test scenarios.

The patches are available in parts in this thread and the final combined patch is there:
http://marc.theaimsgroup.com/?l=linux-kernel&m=110269783227867&w=2

Quoted 8 lines

2.6.10-rc3 contains a partial fix for the erroneous invocation problem, but it is not as effective as Andrea's solution and it still runs into overkill once the oom mechanism is invoked.

Andrea's fix and the selection changes should go into 2.6.10, but I suspect that the VM gurus havent still reached a point, where they agree. I also have the feeling that the problem is partially ignored. Obviously has everybody plenty of memory in his boxes. </rant off>

As someone who runs most new versions first on a 96MB (slow) machine, I would agree that this is a desirable change. I'm not sure yet if the selection is optimal, but it's better than the stock kernel, which seems to follow the "kill them all, let god sort it out" principle.

Signature · 3 lines
bill davidsen <[email protected]>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979
#8 lore raw
Quoted 4 lines
Quoted 4 lines

Andrea's fix and the selection changes should go into 2.6.10, but I suspect that the VM gurus havent still reached a point, where they agree. I also have the feeling that the problem is partially ignored. Obviously has everybody plenty of memory in his boxes. </rant off>

Well you can always take out your VMware and cut it down to <few RAM> MB by hand, just to get an experience how "low-end" users feel.

Quoted 4 lines

As someone who runs most new versions first on a 96MB (slow) machine, I would agree that this is a desirable change. I'm not sure yet if the selection is optimal, but it's better than the stock kernel, which seems to follow the "kill them all, let god sort it out" principle.

Jan Engelhardt

#9 lore raw

Jan Engelhardt escreveu:

Quoted 1 line

Well you can always take out your VMware and cut it down to <few RAM> MB by hand, just to get an experience how "low-end" users feel.

It's not just "low-end users", my interest is in the embedded space where it's normal to dimension things as closely (read, cheaply) as possible. Random processes being killed off is not helpful, especially when the system hasn't actually run out of memory which is the case with the 2.6.9+

Regards,
Chris R.

author
#10 lore raw
In reply to Jan Engelhardt
Quoted 7 lines · Jan Engelhardt
Quoted 4 lines
Quoted 4 lines

Andrea's fix and the selection changes should go into 2.6.10, but I suspect that the VM gurus havent still reached a point, where they agree. I also have the feeling that the problem is partially ignored. Obviously has everybody plenty of memory in his boxes. </rant off>

Well you can always take out your VMware and cut it down to <few RAM> MB by hand, just to get an experience how "low-end" users feel.

Or boot with mem=. I actually do test runs with -ac on a 128Mb box with about 16Mb of that stolen as video ram. 2.6.9 isn't behaving perfectly but seems reasonably ok for most loads except brokenoffice

#11 lore raw

Alan Cox escreveu:

Quoted 3 lines

Or boot with mem=. I actually do test runs with -ac on a 128Mb box with about 16Mb of that stolen as video ram. 2.6.9 isn't behaving perfectly but seems reasonably ok for most loads except brokenoffice

What sort of reasonably OK? My experience on my 64MB P2-350 Dell Optiplex is that 2.6.9-acXX will kill things off at random even when the machine isn't out of memory. If you have any particular test cases you would like run just ask, I understand that some of the difficulty is that the VM developers have machines plenty big enough not to suffer the problems.

The embedded market is different, we're not likely to see multi-gigabyte mobile phones, PDAs, PVRs or whatever for a while at least. 64MB is not *that* small.

Regards,
Chris R.

author
#12 lore raw
Quoted 6 lines · Chris Ross

What sort of reasonably OK? My experience on my 64MB P2-350 Dell Optiplex is that 2.6.9-acXX will kill things off at random even when the machine isn't out of memory. If you have any particular test cases you would like run just ask, I understand that some of the difficulty is that the VM developers have machines plenty big enough not to suffer the problems.

I normally test down to about 100Mbyte but not lower. The 2.6.9-ac15 tree introduced Marcelo's suggested fix for the VM writeout oom cases.

#13 lore raw
In reply to Bill Davidsen
Quoted 4 lines · Bill Davidsen

As someone who runs most new versions first on a 96MB (slow) machine, I would agree that this is a desirable change. I'm not sure yet if the selection is optimal, but it's better than the stock kernel, which seems to follow the "kill them all, let god sort it out" principle.

The main fix IMHO is Andrea's correct solution of the invocation including the removal of the evidential useless protection timers and counters in out_of_memory().

The selection mechanism is now nearly matching the comment above badness () much better: "this algorithm has been meticulously tuned to meet the principle of least surprise ... (be careful when you change it)".

It can always be further optimized, but it will never meet the expectation of all users

tglx