Skip to contentSkip to conversation

Linux 2.6.36-rc7

PATCH New Linus Torvalds 13 participants 29 messages Oct 6, 2010 – Oct 18, 2010

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

So I decided to break my a-week-is-eight-days rut, and actually release -rc7 after a proper seven-day week instead. Wo-oo!

And yes, that's probably as exciting as it gets, which is just fine by me. This should be the last -rc, I'm not seeing any reason to keep delaying a real release. There was still more changes to drivers/gpu/drm than I really would have hoped for, but they all look harmless and good. Famous last words.

Apart from that, there's some mips updates, some late ACPI fixes, and lots of small random noise. I was looking at the dirstat, and noticed that xfs shows up at 6%, but even that is just a single commit that is mainly comment fixups. Which just goes to show that there's really a lot of small trivial one-liners: 167 files changed, 992 insertions(+), 562 deletions(-) in 137 non-merge commits really does end up being about a lot of very small changes.

[ Git hint of the day: do

    git log --stat --oneline --no-merges v2.6.36-rc6..v2.6.36-rc7

to get a dense view of the commits and visually pick out the ones
that aren't trivial one-liners)

---
Akinobu Mita (1):
powerpc/512x: fix clk_get() return value

Alex Deucher (3):
drm/radeon/kms: fix up encoder info messages for DFP6 drm/radeon/kms: fix potential segfault in r600_ioctl_wait_idle drm/radeon/kms: add quirk for MSI K9A2GM motherboard

Andrea Gelmini (2):
ACPI: Kconfig: fix typo.
MIPS: Fix a typo.

Andreas Bießmann (1):
MIPS: Octeon: Determine if helper needs to be built

Andrew Morton (3):
drivers/serial/mfd.c needs slab.h arch/m68k/mac/macboing.c: use unsigned long for irqflags drivers/serial/mrst_max3110.c needs linux/irq.h

Arnd Bergmann (1):
drm: i810/i830: fix locked ioctl variant

Axel Lin (1):
regulator: max8649 - fix setting extclk_freq

Bernhard Walle (2):
i2c-octeon: Return -ETIMEDOUT in octeon_i2c_wait() on timeout
MIPS: N32: Fix getdents64 syscall for n32

Boaz Harrosh (1):
um: Proper Fix for f25c80a4: remove duplicate structure field
initialization

Chris Wilson (5):
drm: Prune GEM vma entries
drm/i915: Fix refleak during eviction.
drm: Hold the mutex when dropping the last GEM reference (v2) drm/i915: Sanity check pread/pwrite
drm/i915: Rephrase pwrite bounds checking to avoid any potential overflow

Christoph Hellwig (1):

      writeback: always use sb->s_bdi for writeback purposes

Colin Ian King (1):
ACPI: enable repeated PCIEXP wakeup by clearing PCIEXP_WAKE_STS on resume

Cyril Chemparathy (1):
regulator: fix typo in current units

Cyrill Gorcunov (1):
perf, x86: Handle in flight NMIs on P4 platform

Damian Lukowski (1):
net-2.6: SYN retransmits: Add new parameter to retransmits_timed_out()

Dan Carpenter (1):
dma/shdma: move dereference below the NULL check

Dan Rosenberg (2):
ALSA: prevent heap corruption in snd_ctl_new()
sys_semctl: fix kernel stack leakage

Daniel J Blueman (1):
fix OMAP2 MTD build failure

Dave Airlie (2):
drm/radeon: fix PCI ID 5657 to be an RV410
drm/gem: handlecount isn't really a kref so don't make it one.

Dave Chinner (1):
xfs: force background CIL push under sustained load

David Daney (3):
MIPS: Hookup fanotify_init, fanotify_mark, and prlimit64 syscalls.
MIPS: Don't place cu2 notifiers in __cpuinitdata
MIPS: Octeon: Place cnmips_cu2_setup in __init memory.

David Howells (1):

      MN10300: Fix flush_icache_range()

David S. Miller (1):
ip_gre: Fix dependencies wrt. ipv6.

Deng-Cheng Zhu (1):
MIPS: Use generic atomic64 for 32-bit kernels

Don Mullis (1):
lib/list_sort: do not pass bad pointers to cmp callback

Eric Dumazet (2):
rcu: rcu_read_lock_bh_held(): disabling irqs also disables bh
vlan: dont drop packets from unknown vlans in promiscuous mode

Eric Millbrandt (1):
powerpc/5200: tighten up ac97 reset timing

Evgeny Kuznetsov (1):
wait: using uninitialized member of wait queue

FUJITA Tomonori (1):
MIPS: TX49xx: Rename ARCH_KMALLOC_MINALIGN to ARCH_DMA_MINALIGN

Florian Mickler (1):
iwl3945: queue the right work if the scan needs to be aborted

Frederic Weisbecker (2):
reiserfs: fix dependency inversion between inode and reiserfs mutexes
reiserfs: fix unwanted reiserfs lock recursion

Frederik Deweerdt (1):
perf ui hist browser: Fix segfault on 'a' for annotate

Geert Uytterhoeven (1):
fuse: Initialize total_len in fuse_retrieve()

Giel van Schijndel (1):
hwmon: f71882fg: use a muxed resource lock for the Super I/O port

H. Peter Anvin (1):
x86, cpu: After uncapping CPUID, re-run CPU feature detection

Heiko Carstens (1):
generic-ipi: Fix deadlock in __smp_call_function_single

Huang Ying (3):
ACPI, APEI, Fix APEI related table size checking ACPI, APEI, Fix error path for memory allocation ACPI, APEI, Fix ERST MOVE_DATA instruction implementation

Hugh Dickins (2):
ksm: fix page_address_in_vma anon_vma oops
ksm: fix bad user data when swapping

Ira W. Snyder (1):
kfifo: fix scatterlist usage

Jeff Layton (2):
cifs: set backing_dev_info on new S_ISREG inodes
cifs: prevent infinite recursion in cifs_reconnect_tcon

Jesse Barnes (1):
drm/i915: fix GMCH power reporting

Jin Dongming (2):
ACPI, APEI, Fix acpi_pre_map() return value ACPI, APEI, HEST Fix the unsuitable usage of platform_data

Jiri Olsa (2):
oprofile: Add Support for Intel CPU Family 6 / Model 29
proc: make /proc/pid/limits world readable

Joe Perches (1):
MIPS: Remove pr_<level> uses of KERN_<level>

Joel Becker (1):
ocfs2: Don't walk off the end of fast symlinks.

Johannes Berg (1):

      mac80211: fix use-after-free

Jon Povey (1):
i2c-davinci: Fix race when setting up for TX

Julia Lawall (3):
powerpc/5200: efika.c: Add of_node_put to avoid memory leak drivers/gpu/drm/i915/i915_gem.c: Add missing error handling code
MIPS: kspd: Adjust confusing if indentation

Keith Packard (2):
drm/i915: vblank status not valid while training display port drm/i915: Use pipe state to tell when pipe is off

Kevin Liu (1):
mfd: Fix max8925 irq control bit incorrect setting

Kukjin Kim (1):
MAINTAINERS: update maintainer for S5P ARM ARCHITECTURES

Kumar Sanghvi (1):
Phonet: Correct header retrieval after pskb_may_pull

Kusanagi Kouichi (1):
perf tools: Fix build breakage

Len Brown (7):
intel_idle: PCI quirk to prevent Lenovo Ideapad s10-3 boot hang
ACPI: delete ZEPTO idle=nomwait DMI quirk
ACPI: expand Vista blacklist to include SP1 and SP2
ACPI: EC: add Vista incompatibility DMI entry for Toshiba Satellite L355
acpi_idle: add missing \n to printk
ACPI: acpi_pad: simplify code to avoid false gcc build warning
ACPI: invoke DSDT corruption workaround on all Toshiba Satellite

Linus Torvalds (3):
Fix up more fallout form alpha signal cleanups
modules: Fix module_bug_list list corruption race Linux 2.6.36-rc7

Lucas De Marchi (2):
ACPI: Fix typos
cpuidle: Fix typos

Luis Henriques (1):
ACPI: fan: Fix more unbalanced code block

Manuel Lauss (1):
MIPS: Alchemy: Resolve prom section mismatches

Marcin Slusarz (1):
i7core_edac: fix panic in udimm sysfs attributes registration

Mark Brown (1):
mfd: Ignore non-GPIO IRQs when setting wm831x IRQ types

Mathieu Lacage (1):
missing inline keyword for static function in linux/dmaengine.h

Matthew Garrett (1):
ACPI: Don't report current_now if battery reports in mWh

MyungJoo Ham (1):
i2c-s3c2410: fix calculation of SDA line delay

Namhyung Kim (3):
ACPI: add missing __percpu markup in arch/x86/kernel/acpi/cstate.c
intel_idle: add missing __percpu markup
[CPUFREQ] acpi-cpufreq: add missing __percpu markup

Pekka Enberg (1):
[CPUFREQ] Fix memory leaks in pcc_cpufreq_do_osc

Petr Vandrovec (1):
MAINTAINERS: update matroxfb & ncpfs status

Rafael J. Wysocki (1):
PM / ACPI: Blacklist systems known to require acpi_sleep=nonvs

Ralf Baechle (7):
MIPS: Document why RELOC_HIDE is there.
MIPS: Audit: Fix hang in entry.S.
MIPS: DMA: Fix computation of DMA flags from device's coherent_dma_mask.
MIPS: Kconfig: Fix and clarify kconfig help text for VSMP and SMTC.
MIPS: GIC: Remove dependencies from Malta files.
MIPS: PNX8550: Sort out machine halt, restart and powerdown functions.
MIPS: Fix syscall 64 bit number comments.

Ricardo Mendoza (1):
MIPS: RM7000: Symbol should be static

Robert Richter (1):
oprofile, ARM: Release resources on failure

Roel Kluin (1):
spi: spi-gpio.c tests SPI_MASTER_NO_RX bit twice, but not SPI_MASTER_NO_TX

Scott Ellis (1):
omap: McBSP: tx_irq_completion used in rx_irq_handler

Shmulik Ladkani (1):
MIPS: Calculate VMLINUZ_LOAD_ADDRESS based on the length of vmlinux.bin

Simon Guinot (1):
dmaengine: fix interrupt clearing for mv_xor

Sinan Akman (1):
of/spi: Fix OF-style driver binding of spi devices

Stefano Stabellini (2):
xen: do not set xenstored_ready before xenbus_probe on hvm
xen: do not initialize PV timers on HVM if !xen_have_vector_callback

Stephane Eranian (1):
perf trace scripting: Fix extern struct definitions

Stephen Rothwell (1):
powerpc: remove unused variable

Suresh Siddha (1):
intel_idle: Voluntary leave_mm before entering deeper

Takashi Iwai (1):
ALSA: i2c/other/ak4xx-adda: Fix a compile warning with CONFIG_PROCFS=n

Thomas Gleixner (2):
x86, irq: Plug memory leak in sparse irq x86, hpet: Fix bogus error check in hpet_assign_irq()

Thomas Hellstrom (5):
drm/vmwgfx: Fix breakage introduced by commit "drm: block
userspace under allocating buffer and having drivers overwrite it (v2)"
vmwgfx: vt-switch (master drop) fixes
vmwgfx: Enable use of the vblank system
vmwgfx: Remove initialisation of dev::devname
vmwgfx: Fix fb VRAM pinning failure due to fragmentation

Thomas Weber (1):

      intel_idle: Change mode 755 => 644

Tony Lindgren (1):
omap: Fix compile dependency to LEDS_CLASS

Vasiliy Kulikov (1):
regulator: fix device_register() error handling

Zhang Rui (3):
ACPI: fix build warnings resulting from merge window conflict ACPI video: fix a poor warning message
ACPI: Disable Windows Vista compatibility for Toshiba P305D

christophe leroy (1):
spi/mpc8xxx: fix buffer overrun on large transfers

#1 lore raw

Hi Linus,

Quoted 5 lines · On Wed, 6 Oct 2010 14:45:13 -0700 Linus Torvalds…

This should be the last -rc, I'm not seeing any reason to keep delaying a real release. There was still more changes to drivers/gpu/drm than I really would have hoped for, but they all look harmless and good. Famous last words.

I have no idea how critical any of this stuff is, but linux-next contain the following in it's "current" trees i.e. stuff that is supposed to go into 2.6.36. These are from the arm-current, scsi-rc-fixes, net-current, wireless-current, kbuild-current, input-current and ide-curent trees (contacts cc'd).

Arnaud Lacombe (1):
kconfig: delay symbol direct dependency initialization
[This one produces a lot of new warnings, but also fixes a bug ...]

Ben Hutchings (2):
Revert "ipv4: Make INET_LRO a bool instead of tristate."
netdev: Depend on INET before selecting INET_LRO

Dan Carpenter (1):
cls_u32: signedness bug

Dan Rosenberg (2):
sctp: prevent reading out-of-bounds memory
sctp: Fix out-of-bounds reading in sctp_asoc_get_hmac()

David Stevens (1):
ipv4: correct IGMP behavior on v3 query during v2-compatibility mode

Dmitry Torokhov (1):
Input: wacom - fix runtime PM related deadlock

Eric Dumazet (1):
caif: fix two caif_connect() bugs

Felix Fietkau (1):
ath9k_hw: fix regression in ANI listen time calculation

Henrik Rydberg (1):
Input: uinput - setup MT usage during device creation

Jeff Kirsher (4):
ixgbevf.txt: Update ixgbevf documentation
e1000.txt: Update e1000 documentation
e1000e.txt: Add e1000e documentation
MAINTAINERS: update Intel LAN Ethernet info

Johannes Berg (1):
mac80211: delete AddBA response timer

Kenneth Waters (1):
Input: joydev - fix JSIOCSAXMAP ioctl

Maciej Żenczykowski (1):
net: Fix IPv6 PMTU disc. w/ asymmetric routes

Martin K. Petersen (1):
[SCSI] Fix VPD inquiry page wrapper

Nagendra Tomar (1):
net: Fix the condition passed to sk_wait_event()

Neil Horman (1):
bonding: fix WARN_ON when writing to bond_master sysfs file

Santosh Shilimkar (1):
ARM: 6419/1: mmu: Fix MT_MEMORY and MT_MEMORY_NONCACHED pte flags

Sergei Shtylyov (2):
hpt366: add debounce delay to cable_detect() method
hpt366: fix clock turnaround

Stanislaw Gruszka (1):
skge: add quirk to limit DMA

Will Deacon (2):
ARM: 6416/1: errata: faulty hazard checking in the Store Buffer may lead to data corruption
ARM: 6412/1: kprobes-decode: add support for MOVW instruction

Signature · 3 lines
Cheers,
Stephen Rothwell                    [email protected]
http://www.canb.auug.org.au/~sfr/
#2 lore raw
Quoted 9 lines · Stephen Rothwell

Hi Linus,

Quoted 5 lines · On Wed, 6 Oct 2010 14:45:13 -0700 Linus Torvalds…

This should be the last -rc, I'm not seeing any reason to keep delaying a real release. There was still more changes to drivers/gpu/drm than I really would have hoped for, but they all look harmless and good. Famous last words.

Quoted 5 lines

I have no idea how critical any of this stuff is, but linux-next contain the following in it's "current" trees i.e. stuff that is supposed to go into 2.6.36. These are from the arm-current, scsi-rc-fixes, net-current, wireless-current, kbuild-current, input-current and ide-curent trees (contacts cc'd).

The SCSI rc-fixes stuff is critical if you run into the bugs, but the bugs are fairly rare cases for most people. I'd still like to get them in, though (and I have another 3 rc fixes candidates going through the test pipeline).

James

#3 lore raw
In reply to Linus Torvalds

On Wednesday 06 Oct 2010 22:45:13 Linus Torvalds wrote:
[snip]

Quoted 5 lines

And yes, that's probably as exciting as it gets, which is just fine by me. This should be the last -rc, I'm not seeing any reason to keep delaying a real release. There was still more changes to drivers/gpu/drm than I really would have hoped for, but they all look harmless and good. Famous last words.

Hi Linus,

Please see http://marc.info/?l=linux-kernel&m=128618485204253&w=2

I have sent this proposed bugfix several times now but no one is picking it up and Eric seems to have disappeared. It would be suboptimal to release 2.6.36 with a core fanotify feature non-functional.

Thanks,

Tvrtko

Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon, OX14 3YP, United Kingdom. Company Reg No 2096520. VAT Reg No GB 348 3873 20.

#4 lore raw
Quoted 9 lines · On Thursday 07 Oct 2010 17:10:46 Tvrtko Ursulin wrote:

On Wednesday 06 Oct 2010 22:45:13 Linus Torvalds wrote:
[snip]

Quoted 5 lines

And yes, that's probably as exciting as it gets, which is just fine by me. This should be the last -rc, I'm not seeing any reason to keep delaying a real release. There was still more changes to drivers/gpu/drm than I really would have hoped for, but they all look harmless and good. Famous last words.

Quoted 7 lines

Hi Linus,

Please see http://marc.info/?l=linux-kernel&m=128618485204253&w=2

I have sent this proposed bugfix several times now but no one is picking it up and Eric seems to have disappeared. It would be suboptimal to release 2.6.36 with a core fanotify feature non-functional.

Unfortunately I have another showstopper. Sadly I missed it until now because internally we were more worried of issues which were kind of direct problems for us and I went to deep instead of spending more time reviewing it breath first.

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=08ae89380a8210a9965d04083e1de78cb8bca4b1

Priority argument was dropped from the fanotify_init syscall, and since it is a syscall once released it is set in stone. Without the priority argument, how are multiple clients supposed to be ordered?

Co-existence between multiple clients was something which was supposed to be designed in from the start. Use cases like hierarchical storage management, anti-malware and content indexing should all be able to co-exist. Without a priority argument I do not see how it can be assured HSM sees the perm event before anti-malware, and content indexing after both of them? If there was any discussion about dropping priority I missed it. :(

Tvrtko

Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon, OX14 3YP, United Kingdom. Company Reg No 2096520. VAT Reg No GB 348 3873 20.

#5 lore raw
Quoted 7 lines · Tvrtko Ursulin
Quoted 1 line · On Thursday 07 Oct 2010 17:10:46 Tvrtko Ursulin wrote:

On Wednesday 06 Oct 2010 22:45:13 Linus Torvalds wrote:

Priority argument was dropped from the fanotify_init syscall, and since it is a syscall once released it is set in stone. Without the priority argument, how are multiple clients supposed to be ordered?

Quoted 6 lines

Co-existence between multiple clients was something which was supposed to be designed in from the start. Use cases like hierarchical storage management, anti-malware and content indexing should all be able to co-exist. Without a priority argument I do not see how it can be assured HSM sees the perm event before anti-malware, and content indexing after both of them? If there was any discussion about dropping priority I missed it. :(

Shit. I'm trying to remember the logic. hrmph.... You could have a
real interface issue.... Shit. Let me think about it for an hour or two.

Original idea of priorities was to allow multiple permissions decision makers to co-exist without having the livelock problem of each trying to grant and deny access to each other. That was solved with the O_NONOTIFY hack and I think the priority was then thought to be useless. But you're absolutely right, it isn't useless if we consider that an HSM might need to run first to make sure data exists on disk before an indexer looks at the data.

I see two possibilities off the top of my head:

I could just slap an (unused) priority field onto the end of the fanotify_init() syscall (assuming Linus doesn't murder me) so we can build that support out with explicit priorities down the line, which I think might be overkill, or

The other option (without breaking ABI as it stands today) is to define some set of the fanotify_init() flags to be a priority field, we've got 32 bits and only use 2 of them so giving 4-8 bits of that as a priority (next cycle) isn't an issue and can be easily backwards compatible.

-Eric

#6 lore raw
Quoted 11 lines

I see two possibilities off the top of my head:

I could just slap an (unused) priority field onto the end of the fanotify_init() syscall (assuming Linus doesn't murder me) so we can build that support out with explicit priorities down the line, which I think might be overkill, or

The other option (without breaking ABI as it stands today) is to define some set of the fanotify_init() flags to be a priority field, we've got 32 bits and only use 2 of them so giving 4-8 bits of that as a priority (next cycle) isn't an issue and can be easily backwards compatible.

Except you've then got a magic release that works differently to every other release afterwards and which sods law says will get shipped by some big vendor. Both your proposals are "we got the API wrong, lets have one kernel that is special", that isn't good in the bigger scheme of things and will have if kernel 2.6.blah then crap forced into bits of app/lib code forever.

Given two chunks of "oh dear" last minute stuff would it be safer to simply punt and just pull the syscall/prototype itself (leaving the rest) for the release. That can go into the first pass of the next kernel tree, and if it the fixes and priority bits all work out may well then be tiny enough for -stable.

Alan

#7 lore raw
Quoted 7 lines · Alan Cox
Quoted 7 lines

I see two possibilities off the top of my head:

I could just slap an (unused) priority field onto the end of the fanotify_init() syscall (assuming Linus doesn't murder me) so we can build that support out with explicit priorities down the line, which I think might be overkill, or

Quoted 8 lines
Quoted 4 lines

The other option (without breaking ABI as it stands today) is to define some set of the fanotify_init() flags to be a priority field, we've got 32 bits and only use 2 of them so giving 4-8 bits of that as a priority (next cycle) isn't an issue and can be easily backwards compatible.

Except you've then got a magic release that works differently to every other release afterwards and which sods law says will get shipped by some big vendor. Both your proposals are "we got the API wrong,

Correct.

Quoted 4 lines

lets have one
kernel that is special", that isn't good in the bigger scheme of things and will have if kernel 2.6.blah then crap forced into bits of app/lib code forever.

Not sure I understand this logic completely. I see both of those options as: we'd have a 2.6.36 kernel which don't have a priority feature (and would reject apps that try to use it) and that feature could be built into 2.6.37. Apps built against 2.6.36 kernels would still work on 2.6.37 (with a priority of 0 since that's all they could set in 2.6.36)

Quoted 5 lines

Given two chunks of "oh dear" last minute stuff would it be safer to simply punt and just pull the syscall/prototype itself (leaving the rest) for the release. That can go into the first pass of the next kernel tree, and if it the fixes and priority bits all work out may well then be tiny enough for -stable.

The safest thing would probably be to punt the syscalls to 2.6.37. Which is sad since I know a number of people are already working against them, but maybe that proves it's the best approach?

-Eric

#8 lore raw
Quoted 3 lines · On Thursday 07 October 2010 19:49:28 Eric Paris wrote:

The safest thing would probably be to punt the syscalls to 2.6.37. Which is sad since I know a number of people are already working against them, but maybe that proves it's the best approach?

I agree with removing the syscalls from 2.6.36 because of the following reasons:

* Reviewers have complained that the feature was not ready to be merged, yet.

* At least some of the criticism did not get addressed (neither discussed nor fixed).

* Some weaknesses in the interface design were only identified and fixed late in the -rc phase, changing the ABI. There may be more issues, like the priority discussion. This might leave us with a broken ABI we would need to support forever.

(Making fanotify fit for HSM hasn't been thought through at all AFAIK but fanotify has legitimate use cases apart from HSM, so I don't necessarily consider this a blocker.)

* The code has shown to contain the kinds of bugs which show that it was not tested very well before merging.

Thanks,
Andreas

#9 lore raw
Quoted 7 lines · Andreas Gruenbacher
Quoted 3 lines · On Thursday 07 October 2010 19:49:28 Eric Paris wrote:

The safest thing would probably be to punt the syscalls to 2.6.37. Which is sad since I know a number of people are already working against them, but maybe that proves it's the best approach?

I agree with removing the syscalls from 2.6.36 because of the following reasons:

How would the mechanics of this be achieved?

Is it enough to just unconditionally return -ENOSYS from the sys_*() functions? Or should all the patches be reverted?

David Daney

#10 lore raw
Quoted 6 lines · On Friday 08 October 2010 18:33:55 David Daney wrote:
Quoted 5 lines · Andreas Gruenbacher
Quoted 3 lines · On Thursday 07 October 2010 19:49:28 Eric Paris wrote:

The safest thing would probably be to punt the syscalls to 2.6.37. Which is sad since I know a number of people are already working against them, but maybe that proves it's the best approach?

Quoted 7 lines
Quoted 2 lines

I agree with removing the syscalls from 2.6.36 because of the following reasons:

How would the mechanics of this be achieved?

Is it enough to just unconditionally return -ENOSYS from the sys_*() functions? Or should all the patches be reverted?

Whatever works I guess ... they would get reactivated pretty soon, anyway.

Andreas

#11 lore raw
Quoted 10 lines · Andreas Gruenbacher
Quoted 9 lines · On Friday 08 October 2010 18:33:55 David Daney wrote:
Quoted 7 lines · Andreas Gruenbacher
Quoted 3 lines · On Thursday 07 October 2010 19:49:28 Eric Paris wrote:

The safest thing would probably be to punt the syscalls to 2.6.37. Which is sad since I know a number of people are already working against them, but maybe that proves it's the best approach?

I agree with removing the syscalls from 2.6.36 because of the following reasons:

Quoted 7 lines
Quoted 4 lines

How would the mechanics of this be achieved?

Is it enough to just unconditionally return -ENOSYS from the sys_*() functions? Or should all the patches be reverted?

Whatever works I guess ... they would get reactivated pretty soon, anyway.

Returning -ENOSYS should be sufficient (that's what a non-system-call does); it would *also* be good to block any headers from getting exported to userspace so people don't end up with compiling against the wrong version of the kernel headers and then wonder why their code doesn't work in the future.

-hpa

#12 lore raw

fanotify: disable fanotify syscalls

This patch disables the fanotify syscalls by just not building them and letting the cond_syscall() statements in kernel/sys_ni.c redirect them to sys_ni_syscall().

It was pointed out by Tvrtko Ursulin that the fanotify interface did not include an explicit prioritization between groups. This is necessary for fanotify to be usable for hierarchical storage management software, as they must get first access to the file, before inotify-like notifiers see the file.

This feature can be added in an ABI compatible way in the next release (by using a number of bits in the flags field to carry the info) but it was suggested by Alan that maybe we should just hold off and do it in the next cycle, likely with an (new) explicit argument to the syscall. I don't like this approach best as I know people are already starting to use the current interface, but Alan is all wise and noone on list backed me up with just using what we have. I feel this is needlessly ripping the rug out from under people at the last minute, but if others think it needs to be a new argument it might be the best way forward.

Three choices:
Go with what we got (and implement the new feature next cycle). Add a new field right now (and implement the new feature next cycle). Wait till next cycle to release the ABI (and implement the new feature next cycle). This is number 3.

---

 fs/notify/Kconfig    |    2 +-
 include/linux/Kbuild |    1 -
 2 files changed, 1 insertion(+), 2 deletions(-)
fs/notify/Kconfig +1 -1
index 22c629e..b388443 100644
@@ -3,4 +3,4 @@ config FSNOTIFY
 
 source "fs/notify/dnotify/Kconfig"
 source "fs/notify/inotify/Kconfig"
-source "fs/notify/fanotify/Kconfig"
+#source "fs/notify/fanotify/Kconfig"
include/linux/Kbuild +0 -1
index 626b629..4e8ea8c 100644
@@ -118,7 +118,6 @@ header-y += eventpoll.h
 header-y += ext2_fs.h
 header-y += fadvise.h
 header-y += falloc.h
-header-y += fanotify.h
 header-y += fb.h
 header-y += fcntl.h
 header-y += fd.h
#13 lore raw
In reply to Andreas Gruenbacher
Quoted 7 lines · Andreas Gruenbacher
Quoted 3 lines · On Thursday 07 October 2010 19:49:28 Eric Paris wrote:

The safest thing would probably be to punt the syscalls to 2.6.37. Which is sad since I know a number of people are already working against them, but maybe that proves it's the best approach?

I agree with removing the syscalls from 2.6.36 because of the following reasons:

I disagree with all of your reasons and have argued my position on this topic repeatedly and don't see the need to refute your claims again.

However, THIS is potentially a real ABI problem and something which deals with the interface. Alan seemed to lean towards pulling the syscalls. It is relatively easily solved without changing the interface or breaking userspace in 2.6.37. We use some set of the flags bits as a priority (we only use 2 of the 32 bits today so we have plenty) and order groups with highest priority first, 0 priority last, and 2+ groups with the same priority have unpredictable ordering. I'd then call priorities other than 0 a 2.6.37 feature. If we do it in flags I think that leaves us with say 8 bits and thus 255 priorities. Maybe people want more, if so, that's a interface change to add a new argument.

Now if Alan would still like me to pull, if anyone has any other 11th hour interface problems, or if 255 priorities doesn't seem like enough to someone I am wondering what the best way to unhook is. Just make the functions return -ENOSYS as the first line or actually troll through all of the arches and explicily unhook and rehook to sys_ni_syscall? I started on the latter, but it seems to be a rather large patch at this point...

-Eric

------

I said I wouldn't refute your claims but I can't help myself on one account which I think might mislead people.

* Some weaknesses in the interface design were only identified and fixed late in the -rc phase, changing the ABI. There may be more issues, like the priority discussion. This might leave us with a broken ABI we would need to support forever.

Between rc2 and rc3 we switched the order and size of a couple of fields to help alignment, it did break ABI, but it wasn't an interface failing.
See: 0fb85621df4f. It also lead to an interesting idea about a new type for linux/types.h which both fanotify and the networking could make use of (but isn't picked up, I'm not sure we know who is in charge of types.h)

I concede the interface may in fact not be perfect for every user. I have ask for ideas and feedback on proposals for literally (not figuratively) years. It has gone through many iterations. The interface has been in Linus's kernel for some months and I know of numerous people who are starting to code to it and haven't (until now) found interface failings. We can talk about how problems might be out there but if they haven't been found after all this time I doubt just waiting longer is going to change anything.

#14 lore raw
Quoted 9 lines · On Friday 08 October 2010 18:38:28 Eric Paris wrote:

However, THIS is potentially a real ABI problem and something which deals with the interface. Alan seemed to lean towards pulling the syscalls. It is relatively easily solved without changing the interface or breaking userspace in 2.6.37. We use some set of the flags bits as a priority (we only use 2 of the 32 bits today so we have plenty) and order groups with highest priority first, 0 priority last, and 2+ groups with the same priority have unpredictable ordering. I'd then call priorities other than 0 a 2.6.37 feature. If we do it in flags I think that leaves us with say 8 bits and thus 255 priorities.

That's a possibility but it seems quite messy for a brand new system call. I'd still pull the system call and work out the few remaining quirkses.

Thanks.

Quoted 5 lines

------

I said I wouldn't refute your claims but I can't help myself on one account which I think might mislead people.

Quoted 8 lines

* Some weaknesses in the interface design were only identified and fixed late in the -rc phase, changing the ABI. There may be more issues, like the priority discussion. This might leave us with a broken ABI we would need to support forever.

Between rc2 and rc3 we switched the order and size of a couple of fields to help alignment, it did break ABI, but it wasn't an interface failing.
See: 0fb85621df4f.

See, a weakness, not a failure. What I said.

The main issue that Andreas Schwab has pointed out (and which also lead to this commit) was the packing of the structs which leads to inefficient code, though. This hasn't been fixed but it still can be, in a backwards compatible way.

Thanks,
Andreas

#15 lore raw
In reply to Eric Paris
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Eric" == Eric Paris <[email protected]> writes:

Quoted 7 lines · Tvrtko Ursulin
Quoted 7 lines · On Thursday 07 Oct 2010 17:10:46 Tvrtko Ursulin wrote:
Quoted 1 line · On Wednesday 06 Oct 2010 22:45:13 Linus Torvalds wrote:

Priority argument was dropped from the fanotify_init syscall, and since it is a syscall once released it is set in stone. Without the priority argument, how are multiple clients supposed to be ordered?

Quoted 6 lines
Quoted 6 lines

Co-existence between multiple clients was something which was supposed to be designed in from the start. Use cases like hierarchical storage management, anti-malware and content indexing should all be able to co-exist. Without a priority argument I do not see how it can be assured HSM sees the perm event before anti-malware, and content indexing after both of them? If there was any discussion about dropping priority I missed it. :(

Eric> Shit. I'm trying to remember the logic. hrmph.... You could Eric> have a real interface issue.... Shit. Let me think about it Eric> for an hour or two.

Eric> Original idea of priorities was to allow multiple permissions Eric> decision makers to co-exist without having the livelock problem Eric> of each trying to grant and deny access to each other. That was Eric> solved with the O_NONOTIFY hack and I think the priority was Eric> then thought to be useless. But you're absolutely right, it Eric> isn't useless if we consider that an HSM might need to run first Eric> to make sure data exists on disk before an indexer looks at the Eric> data.

Eric> I see two possibilities off the top of my head:

Eric> I could just slap an (unused) priority field onto the end of the Eric> fanotify_init() syscall (assuming Linus doesn't murder me) so we Eric> can build that support out with explicit priorities down the Eric> line, which I think might be overkill, or

Eric> The other option (without breaking ABI as it stands today) is to Eric> define some set of the fanotify_init() flags to be a priority Eric> field, we've got 32 bits and only use 2 of them so giving 4-8 Eric> bits of that as a priority (next cycle) isn't an issue and can Eric> be easily backwards compatible.

So what happens when you try to register a priority level and someone else has already gotten that level? Does the call fail? Do you get bumped down to the next open level? Can you *tell* what level you're at and whether or not some other decision maker is ahead of you?

So if I register an HSM module for /home, with a priority of 1, and then register a content indexer for /home/john at priority 1, will they clash? Who wins? The one registered first?

I tried looking in Documentation/fs/fanotify.txt but I couldn't find it anywhere. So I had to grep around looking for the file which held fanotify_init() so I could look it over... and then my brain started bleeding from the lack of any comments on the various functions on WTF they were supposed to do.

But hey, I admit I'm not a kernel programmer at all, nor a low level FS guy, so I probably just don't have the indepth understanding of Linux kernel internals. I just need to spend six months hacking on the code to come upto speed.

But I'd really like some docs in the next release which tells me as a poor dumb sysadmin how it can and should be used and what the gotchas are.

Thanks,
John

#16 lore raw
Quoted 4 lines · John Stoffel

So what happens when you try to register a priority level and someone else has already gotten that level? Does the call fail? Do you get bumped down to the next open level? Can you *tell* what level you're at and whether or not some other decision maker is ahead of you?

Well it hasn't been discussed and implemented so I can't answer that. *smile*

I will tell you that the way I envision it working (and being backwards compatible) is that priority 0 is the last thing to be serviced. If 2 things register at the same priority the order between them getting events is unpredictable. So when an HSM uses the interface it would use the highest priority. An AV vendor might use (highest priority / 2) while normal inotify like listeners would all be happy using priority 0.

Quoted 3 lines

But I'd really like some docs in the next release which tells me as a poor dumb sysadmin how it can and should be used and what the gotchas are.

We have example man-like pages in the commit logs which I expected to be used as the basis for man pages once the interface was accepted. They aren't perfect but they are

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=52c923dd079df49f58016a9e56df184b132611d6
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=2a3edf86040a7e15684525a2aadc29f532c51325

You'll also find an example program which shows all of the features at

http://git.kernel.org/?p=linux/kernel/git/agruen/fanotify-example.git;a=summary

I don't think digging around in kernel code is the right way :)

-Eric

#17 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Eric" == Eric Paris <[email protected]> writes:

Quoted 4 lines · John Stoffel
Quoted 4 lines

So what happens when you try to register a priority level and someone else has already gotten that level? Does the call fail? Do you get bumped down to the next open level? Can you *tell* what level you're at and whether or not some other decision maker is ahead of you?

Eric> Well it hasn't been discussed and implemented so I can't answer that. Eric> *smile*

Ah, but you have implemented multiple notifiers registered at the same time, without priorities, right? So that's just the degenerate case which I assume is handled already.

Eric> I will tell you that the way I envision it working (and being Eric> backwards compatible) is that priority 0 is the last thing to be Eric> serviced. If 2 things register at the same priority the order Eric> between them getting events is unpredictable. So when an HSM Eric> uses the interface it would use the highest priority. An AV Eric> vendor might use (highest priority / 2) while normal inotify Eric> like listeners would all be happy using priority 0.

Ugh, I'd prefer that priority 0 is first (highest), but I can see how that would make ABI problems, assuming we don't punt on releasing the ABI now.

So if the order is undefined now, that's not good. Well... maybe it's acceptable, but I can see all kinds of problems cropping up. Hopefully they're processed in order of notifier creation instead. So if I insert a notifier, anyone who inserts a notifier after me gets serviced after my notifier gets run. That would seem to be the logical and sane semantics to use here.

Again, I'm really approaching this as a SysAdmin who'd love to use this in an HSM environment, so ordering is a *key* requirement.

Which is why I'm going to harp of the documentation in the Kernel from day one too.

Quoted 3 lines
Quoted 3 lines

But I'd really like some docs in the next release which tells me as a poor dumb sysadmin how it can and should be used and what the gotchas are.

Eric> We have example man-like pages in the commit logs which I expected to be Eric> used as the basis for man pages once the interface was accepted. They Eric> aren't perfect but they are

Eric> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=52c923dd079df49f58016a9e56df184b132611d6
Eric> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=2a3edf86040a7e15684525a2aadc29f532c51325

Eric> You'll also find an example program which shows all of the Eric> features at

Eric> http://git.kernel.org/?p=linux/kernel/git/agruen/fanotify-example.git;a=summary

Eric> I don't think digging around in kernel code is the right way :)

Sure, but putting some design documentation which explains how it's expected to be used, with pointers to the canonical site holding docs and such *is* expected. Just see the docs for inotify and such.

Thanks for the pointers, I'll try to find the time to look them over and make some more comments.

John

#18 lore raw
Quoted 20 lines · On Friday 08 Oct 2010 16:42:06 John Stoffel wrote:
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Eric" == Eric Paris <[email protected]> writes:

Eric> I will tell you that the way I envision it working (and being Eric> backwards compatible) is that priority 0 is the last thing to be Eric> serviced. If 2 things register at the same priority the order Eric> between them getting events is unpredictable. So when an HSM Eric> uses the interface it would use the highest priority. An AV Eric> vendor might use (highest priority / 2) while normal inotify Eric> like listeners would all be happy using priority 0.

Ugh, I'd prefer that priority 0 is first (highest), but I can see how that would make ABI problems, assuming we don't punt on releasing the ABI now.

So if the order is undefined now, that's not good. Well... maybe it's acceptable, but I can see all kinds of problems cropping up. Hopefully they're processed in order of notifier creation instead. So if I insert a notifier, anyone who inserts a notifier after me gets serviced after my notifier gets run. That would seem to be the logical and sane semantics to use here.

Quoted 2 lines

Again, I'm really approaching this as a SysAdmin who'd love to use this in an HSM environment, so ordering is a *key* requirement.

Ordering by time of registration? I thought of that but it falls flat once any service needs to restart or be restarted.

Priority is also not that great concept. I may have proposed classes or something similar at some point, don't remember any more. It would be equivalent to having allocated priority ranges, like:

Quoted 2 lines

1000 - pre-content
=100 - access-control

<100 - content

Doesn't really solve ordering inside groups so maybe we do not need priorities at all just these three classes?

Tvrtko

Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon, OX14 3YP, United Kingdom. Company Reg No 2096520. VAT Reg No GB 348 3873 20.

#19 lore raw
Quoted 6 lines · Tvrtko Ursulin

On Friday 08 Oct 2010 16:42:06 John Stoffel wrote:

Priority is also not that great concept. I may have proposed classes or something similar at some point, don't remember any more. It would be equivalent to having allocated priority ranges, like:

Quoted 6 lines
Quoted 2 lines

1000 - pre-content
=100 - access-control

<100 - content

Doesn't really solve ordering inside groups so maybe we do not need priorities at all just these three classes?

I originally thought of trying to enumerate the types of users and came up with the same 3 you did. Then I thought it better to give a general priority field which we could indicate in documentation something like those 3 classes (exactly like you did above). I don't want to hard code some limited number of types of users into the interface. (ok it's going to limited, I was thinking 8 bits, but maybe others think we need more?)

As an extreme example going with 3 fixed type of users (and thus equivalently only 3 priorities) would not allow for hierarchies of hierarchical storage managers. What if priority MAX only brought in enough info for priority MAX-1 to bring in the real file? If they had to share the single 'pre-content' priority we have another ordering problem.

-Eric

#20 lore raw
Quoted 22 lines · On Friday 08 Oct 2010 17:41:46 Eric Paris wrote:
Quoted 5 lines · Tvrtko Ursulin

On Friday 08 Oct 2010 16:42:06 John Stoffel wrote:

Priority is also not that great concept. I may have proposed classes or something similar at some point, don't remember any more. It would be

Quoted 8 lines

equivalent to having allocated priority ranges, like:

Quoted 2 lines

1000 - pre-content
=100 - access-control

<100 - content

Doesn't really solve ordering inside groups so maybe we do not need priorities at all just these three classes?

I originally thought of trying to enumerate the types of users and came up with the same 3 you did. Then I thought it better to give a general priority field which we could indicate in documentation something like those 3 classes (exactly like you did above). I don't want to hard code some limited number of types of users into the interface. (ok it's going to limited, I was thinking 8 bits, but maybe others think we need more?)

Quoted 6 lines

As an extreme example going with 3 fixed type of users (and thus equivalently only 3 priorities) would not allow for hierarchies of hierarchical storage managers. What if priority MAX only brought in enough info for priority MAX-1 to bring in the real file? If they had to share the single 'pre-content' priority we have another ordering problem.

I think I am fine with priorities if we make these three ranges documented.

Ordering between ranges, like in your chained HSM setup, would then be sysadmin's responsibility but those setups look esoteric enough to think we could get away with it. More importantly, I really do not see an automatic solution which would solve all imaginable (more or less crazy) setups one could come up with and priorities solve all normal ones plus giving some more options than having only three classes.

If you go with 8 bits, you could even have classes plus priorities. Like having low bits for in-class priority and then high two or three for class (three if with a single bit set in class field implementation would be more efficient). I do not see any usability advantage of this scheme though, it just may enable an implementation with less decisions based on arbitrary numbers.

Tvrtko

Sophos Plc, The Pentagon, Abingdon Science Park, Abingdon, OX14 3YP, United Kingdom. Company Reg No 2096520. VAT Reg No GB 348 3873 20.

#21 lore raw
In reply to Tvrtko Ursulin
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Tvrtko" == Tvrtko Ursulin <[email protected]> writes:

Quoted 1 line · Tvrtko> On Friday 08 Oct 2010 16:42:06 John Stoffel wrote:
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Eric" == Eric Paris <[email protected]> writes:

Eric> I will tell you that the way I envision it working (and being Eric> backwards compatible) is that priority 0 is the last thing to be Eric> serviced. If 2 things register at the same priority the order Eric> between them getting events is unpredictable. So when an HSM Eric> uses the interface it would use the highest priority. An AV Eric> vendor might use (highest priority / 2) while normal inotify Eric> like listeners would all be happy using priority 0.

Quoted 12 lines
Quoted 12 lines

Ugh, I'd prefer that priority 0 is first (highest), but I can see how that would make ABI problems, assuming we don't punt on releasing the ABI now.

So if the order is undefined now, that's not good. Well... maybe it's acceptable, but I can see all kinds of problems cropping up. Hopefully they're processed in order of notifier creation instead. So if I insert a notifier, anyone who inserts a notifier after me gets serviced after my notifier gets run. That would seem to be the logical and sane semantics to use here.

Quoted 2 lines
Quoted 2 lines

Again, I'm really approaching this as a SysAdmin who'd love to use this in an HSM environment, so ordering is a *key* requirement.

Tvrtko> Ordering by time of registration? I thought of that but it Tvrtko> falls flat once any service needs to restart or be restarted.

Yup, it can fail, esp if you have indexing before the HSM retrieval... :] That could lead to all kinds of silliness.

But then again, how *serious* is it when it does fail? For example, we use CommVault at work on Netapps to do HSM for some data. CV leaves a little stub file holding the key to the content to be restored, and a note for the end user saying "Don't mess with this file, it's a pointer to your data. Call your SysAdmin if you see this..."

So that's one way of approaching the error handling case.

Tvrtko> Priority is also not that great concept. I may have proposed Tvrtko> classes or something similar at some point, don't remember any Tvrtko> more. It would be equivalent to having allocated priority Tvrtko> ranges, like:

Tvrtko> 1000 - pre-content
Tvrtko> =100 - access-control
Tvrtko> <100 - content

Umm, why is there only one value for the access-control part? Are you saying there should only be one program to decide on access control?
Shouldnt' there be more? So I think you're classes are just passed
in by the notifier setup call:

   flags = FANOTIFY_PRE_CONTENT & FANOTIFY_EXCLUSIVE;
   add_notifier(filesystem,flags,&hook);

   flags = ACCESS & SUFFICIENT;
   add_notififier(filesystem,flags,&hook);

And you'd have flags like:

PRE_ACCESS 1
ACCESS 2
POST_ACCESS 4
EXCLUSIVE 8
SUFFICIENT 16 /* Passing this means all ACCESS is allowed */
/* even with other ACCESS notifiers in the chain */

so say where in the process you need to be. The EXCLUSIVE flag means we want to be the only one doing this and to bail if someone else is already registered.

Tvrtko> Doesn't really solve ordering inside groups so maybe we do not Tvrtko> need priorities at all just these three classes?

Ordering is a key thing. In most cases, you don't care, but for some cases you really do care.

Maybe we should take some of the ideas used by PAM for ordering to used as a basis?

John

#22 lore raw
In reply to Tvrtko Ursulin
Quoted 2 lines · On Friday 08 October 2010 18:17:25 Tvrtko Ursulin wrote:

Doesn't really solve ordering inside groups so maybe we do not need priorities at all just these three classes?

Applications can easily make the priority they register with configurable (and default to something in their "range" if we define such ranges). Wouldn't this be enough?

Andreas

#23 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Andreas" == Andreas Gruenbacher <[email protected]> writes:

Quoted 2 lines · Andreas> On Friday 08 October 2010 18:17:25 Tvrtko Ursulin wrote:
Quoted 2 lines

Doesn't really solve ordering inside groups so maybe we do not need priorities at all just these three classes?

Andreas> Applications can easily make the priority they register with Andreas> configurable (and default to something in their "range" if we Andreas> define such ranges). Wouldn't this be enough?

What about conflicts, when two applications register with the same priority? And then one of the them exits, dies, fails, and re-registers itself.

Does it get the same slot back? Does it goto the end of the line? If it's an HSM handling (or other content manipulation tool) what happens?

These are the questions which need to be answered, and the ABI updated to reflect these questions. Even if it's just "We don't care..." it needs to be spelled out.

John

maintainer
#24 lore raw
In reply to Linus Torvalds

Hello, Linus.

Quoted 8 lines · Linus Torvalds

So I decided to break my a-week-is-eight-days rut, and actually release -rc7 after a proper seven-day week instead. Wo-oo!

And yes, that's probably as exciting as it gets, which is just fine by me. This should be the last -rc, I'm not seeing any reason to keep delaying a real release. There was still more changes to drivers/gpu/drm than I really would have hoped for, but they all look harmless and good. Famous last words.

I'm afraid there is a possibly workqueue related deadlock under high memory pressure. It happens on dm-crypt + md raid1 configuration. I'm not yet sure whether this is caused by workqueue failing to kick rescuers under memory pressure or the shared workqueue is making an already existing problem more visible and in the process of setting up an environment to reproduce the problem.

http://thread.gmane.org/gmane.comp.file-systems.xfs.general/34922/focus=1044784

Thanks.

Signature · 1 line
tejun
#25 lore raw
Quoted 8 lines · Tejun Heo

I'm afraid there is a possibly workqueue related deadlock under high memory pressure. It happens on dm-crypt + md raid1 configuration. I'm not yet sure whether this is caused by workqueue failing to kick rescuers under memory pressure or the shared workqueue is making an already existing problem more visible and in the process of setting up an environment to reproduce the problem.

http://thread.gmane.org/gmane.comp.file-systems.xfs.general/34922/focus=1044784

Yes, XFS is very good to show up problems in dm-crypt:)

But there was no change in dm-crypt which can itself cause such problem, planned workqueue changes are not in 2.6.36 yet.
Code is basically the same for the last few releases.

So it seems that workqueue processing really changed here under memory pressure.

Milan

p.s.
Anyway, if you are able to reproduce it and you think that there is problem in per-device dm-crypt workqueue, there are patches from Andi for shared per-cpu workqueue, maybe it can help here. (But this is really not RC material.)

Unfortunately not yet in dm-devel tree, but I have them here ready for review:
http://mbroz.fedorapeople.org/dm-crypt/2.6.36-devel/
(all 4 patches must be applied, I hope Alasdair will put them in dm quilt soon.)

maintainer
#26 lore raw

Hello, again.

Quoted 10 lines · Milan Broz

Yes, XFS is very good to show up problems in dm-crypt:)

But there was no change in dm-crypt which can itself cause such problem, planned workqueue changes are not in 2.6.36 yet.
Code is basically the same for the last few releases.

So it seems that workqueue processing really changed here under memory pressure.

Milan

Quoted 8 lines

p.s.
Anyway, if you are able to reproduce it and you think that there is problem in per-device dm-crypt workqueue, there are patches from Andi for shared per-cpu workqueue, maybe it can help here. (But this is really not RC material.)

Unfortunately not yet in dm-devel tree, but I have them here ready for review:
http://mbroz.fedorapeople.org/dm-crypt/2.6.36-devel/
(all 4 patches must be applied, I hope Alasdair will put them in dm quilt soon.)

Okay, spent the whole day reproduing the problem and trying to determine what's going on. In the process, I've found a bug and a potential issue (not sure whether it's an actual issue which should be fixed for this release yet) but the hang doesn't seem to have anything to do with workqueue update. All the queues are behaving exactly as expected during hang.

Also, it isn't a regression. I can reliably trigger the same deadlock on v2.6.35.

Here's the setup, which should be mostly similar to Torsten's setup I used to trigger the problem.

The machine is dual quad-core Opteron (8 phys cores) w/ 4GiB memory.

* 80GB raid1 of two SATA disks
* On top of that, luks encrypted device w/ twofish-cbc-essiv:sha256
* In the encrypted device, xfs filesystem which hosts 8GiB swapfile
* 12GiB tmpfs

The workload is v2.6.35 allyesconfig -j 128 build in the tmpfs. Not too long after swap starts being used (several tens of secs), the system hangs. IRQ handling and all are fine but no IO gets through with a lot of tasks stuck in bio allocation somewhere.

I suspected that with md and dm stacked together, something in the upper layer ended up exhausting a shared bio pool and tried a couple of things but haven't succeeded at finding where the culprit is. It probably would be best to run blktrace together and analyze how IO gets stuck.

So, well, we seem to be broken the same way as before. No need to delay release for this one.

Thanks.

Signature · 1 line
tejun
#27 lore raw
Quoted 47 lines · Tejun Heo

Hello, again.

Quoted 10 lines · Milan Broz

Yes, XFS is very good to show up problems in dm-crypt:)

But there was no change in dm-crypt which can itself cause such problem, planned workqueue changes are not in 2.6.36 yet.
Code is basically the same for the last few releases.

So it seems that workqueue processing really changed here under memory pressure.

Milan

Quoted 8 lines

p.s.
Anyway, if you are able to reproduce it and you think that there is problem in per-device dm-crypt workqueue, there are patches from Andi for shared per-cpu workqueue, maybe it can help here. (But this is really not RC material.)

Unfortunately not yet in dm-devel tree, but I have them here ready for review:
http://mbroz.fedorapeople.org/dm-crypt/2.6.36-devel/
(all 4 patches must be applied, I hope Alasdair will put them in dm quilt soon.)

Okay, spent the whole day reproduing the problem and trying to determine what's going on.  In the process, I've found a bug and a potential issue (not sure whether it's an actual issue which should be fixed for this release yet) but the hang doesn't seem to have anything to do with workqueue update.  All the queues are behaving exactly as expected during hang.

Also, it isn't a regression.  I can reliably trigger the same deadlock on v2.6.35.

Here's the setup, which should be mostly similar to Torsten's setup I used to trigger the problem.

The machine is dual quad-core Opteron (8 phys cores) w/ 4GiB memory.

* 80GB raid1 of two SATA disks
* On top of that, luks encrypted device w/ twofish-cbc-essiv:sha256
* In the encrypted device, xfs filesystem which hosts 8GiB swapfile
* 12GiB tmpfs

The workload is v2.6.35 allyesconfig -j 128 build in the tmpfs.  Not too long after swap starts being used (several tens of secs), the system hangs.  IRQ handling and all are fine but no IO gets through with a lot of tasks stuck in bio allocation somewhere.

Quoted 8 lines

I suspected that with md and dm stacked together, something in the upper layer ended up exhausting a shared bio pool and tried a couple of things but haven't succeeded at finding where the culprit is.  It probably would be best to run blktrace together and analyze how IO gets stuck.

So, well, we seem to be broken the same way as before.  No need to delay release for this one.

I instrument mm/mempool.c, trying to find what shared pool gets exhausted. On the last run, it seemed that the fs_bio_set from fs/bio.c runs dry.

As far as I can see, that pool is used by bio_alloc() and bio_clone(). Above bio_alloc() a dire warning says, that any bio allocated that way needs to be submitted from IO, otherwise the system could livelock. bio_clone() does not have this warning, but as it uses the same pool in the same way, I would expect the same rule applies.

Looking for uses of bio_allow() and bio_clone() in drivers/md it looks like dm-crypt uses its own pools and not the fs_bio_set.
But drivers/md/raid1.c uses this pool, and in my eyes it does it wrong.

When writing to a RAID1 array the function make_request() in raid1.c does a bio_clone() for each drive (lines 967-1001 in 2.6.36-rc7) and only after all bios are allocates they will be merged into the pending_bio_list.

So a RAID1 with 3 mirrors is a sure way to lock up a system as soon as the mempool is needed?
(The fs_bio_set pool only allocates BIO_POOL_SIZE entries and that is defined as 2)

From the use of atomic_inc(&r1_bio->remaining) and the use of the spin_lock_irqsave(&conf->device_lock, flags) when merging the bio list, I would suspect that its even possible that multiple CPUs concurrently get into this allocation loop, or that the use of multiple RAID1 devices each with only 2 drives could lock up the same way.

What am I missing, or is the use of bio_clone() really the wrong thing?

Torsten

maintainer
#28 lore raw

PATCH 1/1 workqueue: fix HIGHPRI handling in keep_working()

In reply to Tejun Heo

The policy function keep_working() didn't check GCWQ_HIGHPRI_PENDING and could return %false with highpri work pending. This could lead to late execution of a highpri work which was delayed due to @max_active throttling if other works are actively consuming CPU cycles.

For example, the following could happen.

1. Work W0 which burns CPU cycles.

2. Two works W1 and W2 are queued to a highpri wq w/ @max_active of 1.

3. W1 starts executing and W2 is put to delayed queue. W0 and W1 are both runnable.

4. W1 finishes which puts W2 to pending queue but keep_working() incorrectly returns %false and the worker goes to sleep.

5. W0 finishes and W2 starts execution.

With this patch applied, W2 starts execution as soon as W1 finishes.

---
This is the workqueue bug I've found while trying to debug the dm/raid hang. Although the bug may introduce unexpected delay in scheduling a highpri work, the delay can only be as long as the combined length of CPU cycle burns of the already running works. Given that HIGHPRI is currently only used by xfs and its usage, I don't think it's likely to cause an actual issue. I'll queue it for #for-next.

Thank you.

 kernel/workqueue.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)
kernel/workqueue.c +3 -1
index f77afd9..d355278 100644
@@ -604,7 +604,9 @@ static bool keep_working(struct global_cwq *gcwq)
 {
 	atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
 
-	return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1;
+	return !list_empty(&gcwq->worklist) &&
+		(atomic_read(nr_running) <= 1 ||
+		 gcwq->flags & GCWQ_HIGHPRI_PENDING);
 }
 
 /* Do we need a new worker?  Called from manager. */
Signature · 1 line
1.7.1