[PATCH] myri10ge - Always do a dummy RDMA after loading the firmware
Always do a dummy RDMA after loading the firmware to work around
buggy PCIe chipsets which do not implement resending properly.
This is so cheap as to be almost free, and should never have been
conditional on the tx boundary != 4096.
Signed-off-by: Brice Goglin <brice@myri.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
Ron Mercer [Tue, 25 Jul 2006 07:40:21 +0000 (00:40 -0700)]
[PATCH] qla3xxx NIC driver
This is a complementary network driver for our ISP4XXX parts.
There is a concurrent effort underway to get the iSCSI driver (qla4xxx)
integrated upstream as well.
I have been through several iterations with the linux-netdev list and have had
much response from Stephen Hemminger.
- Built and tested using kernel 2.6.17-rc4.
- The chip supports two ethernet and two iSCSI functions.
- The functions ql_sem_lock, ql_sem_spinlock, ql_sem_unlock, and
ql_wait_for_drvr_lock are used to protect resources that are shared across
the network and iSCSI functions. This protection is mostly during chip
initialization and resets, but also include link management.
- The PHY/MII are not exported through ethtool due to the fact that the
iSCSI function will control the common link at least 50% of the time.
This driver has been through several iterations on the netdev list and we feel
this driver is ready for inclusion in the upstream kernel.
It has been built and tested on x86 and PPC64 platforms.
Cc: Jeff Garzik <jeff@garzik.org> Cc: Stephen Hemminger <shemminger@osdl.org> Signed-off-by: Ron Mercer <ron.mercer@qlogic.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Jeff Garzik <jeff@garzik.org>
Fix robust PI-futexes to be properly unlocked on unexpected exit.
For this to work the kernel has to know whether a futex is a PI or a
non-PI one, because the semantics are different. Since the space in
relevant glibc data structures is extremely scarce, the best solution is
to encode the 'PI' information in bit 0 of the robust list pointer.
Existing (non-PI) glibc robust futexes have this bit always zero, so the
ABI is kept. New glibc with PI-robust-futexes will set this bit.
Further fixes from Thomas Gleixner <tglx@linutronix.de>
Fix pi_state->list handling bugs: list handling mishap, locking error.
Plus add more debug checks and fix a few style issues i noticed while
debugging this.
The dwarf2 unwinder currently often gets stuck because a lot
of assembly code doesn't have proper dwarf2 annotiation yet.
This currently often happens with __down. Should fix this by
adding proper dwarf2 annotation to all inline assembly. However
until that's done we need a quick fix for 2.6.18 to avoid
incomplete backtraces.
So when this happens dump the rest of the stack with the old unwinder
instead of silently not dumping it. There was already a optional
"both" mode that dumped both, but that was too ugly.
I also clarified the headers for the different backtraces a bit.
Also add a clear error message for missing dwarf2
annotation that people can work on.
And I removed a dead variable left over from Ingo's changes.
bibo mao [Fri, 28 Jul 2006 12:44:48 +0000 (14:44 +0200)]
[PATCH] x86_64: Enlarge debug stack for nested kprobes
In x86_64 platform, INT1 and INT3 trap stack is IST stack called DEBUG_STACK,
when INT1/INT3 trap happens, system will switch to DEBUG_STACK by hardware.
Current DEBUG_STACK size is 4K, when int1/int3 trap happens, kernel will
minus current DEBUG_STACK IST value by 4k. But if int3/int1 trap is nested,
it will destroy other vector's IST stack. This patch modifies this, it sets
DEBUG_STACK size as 8K and allows two level of nested int1/int3 trap.
Kprobe DEBUG_STACK may be nested, because kprobe handler may be probed
by other kprobes.
Thanks jbeulich for pointing out error in the first patch.
[AK: nested kprobes are pretty dubious. Hopefully one nest
will be enough. This will cost 8K per CPU (4K more than before)]
[PATCH] x86_64: Don't clobber r8-r11 in int 0x80 handler
When int 0x80 is called from long mode r8-r11 would leak out of the
kernel (or rather they would be filled with some values from
the kernel stack). I don't think it's a security issue because
the values come from the fixed stack frame which should be near
always user registers from a previous interrupt.
Still better fix it.
Longer term the register save macros need to be cleaned up
to avoid such mistakes in the future.
Original analysis from Richard Brunner, fix by me.
[PATCH] i386/x86-64: Add user_mode checks to profile_pc for oprofile
Fixes a obscure user space triggerable crash during oprofiling.
Oprofile calls profile_pc from NMIs even when user_mode(regs) is not true and
the program counter is inside the kernel lock section. This opens
a race - when a user program jumps to a kernel lock address and
a NMI happens before the illegal page fault exception is raised
and the program has a unmapped esp or ebp then the kernel could
oops. NMIs have a higher priority than exceptions so that could
happen.
Add user_mode checks to i386/x86-64 profile_pc to prevent that.
* git://oss.sgi.com:8090/nathans/xfs-rc-2.6:
[XFS] Ensure bulkstat from an invalid inode number gets caught always with
[XFS] Fix a barrier related forced shutdown on mounts with quota enabled.
[XFS] Fix remount vs no/barrier options by ensuring we clear unwanted
[XFS] All xfs_disk_dquot_t values are (as the name says) disk endian.
[PATCH] ide: option to disable cache flushes for buggy drives
Some drives claim they support cache flushing, but get seriously
confused if you try. Add this option to be able to boot with
barriers enabled by default.
David S. Miller [Thu, 27 Jul 2006 23:49:21 +0000 (16:49 -0700)]
[SPARC64]: Fix quad-float multiply emulation.
Something is wrong with the 3-multiply (vs. 4-multiply) optimized
version of _FP_MUL_MEAT_2_*(), so just use the slower version
which actually computes correct values.
Noticed by Rene Rebe
Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Drake [Tue, 18 Jul 2006 21:00:25 +0000 (22:00 +0100)]
[PATCH] softmac: Add MAINTAINERS entry
Signed-off-by: Daniel Drake <dsd@gentoo.org> Acked-by: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: John W. Linville <linville@tuxdriver.com>
This patch changes ieee80211_rx_any to always claim the skb, which avoids
further driver complexity and the possibility of leaking management frames.
It also exports the function so that people can actually use it.
Signed-off-by: Daniel Drake <dsd@gentoo.org> Acked-by: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Daniel Drake [Tue, 18 Jul 2006 20:34:56 +0000 (21:34 +0100)]
[PATCH] softmac: export highest_supported_rate function
zd1211 needs this functionality, no point duplicating it.
Signed-off-by: Daniel Drake <dsd@gentoo.org> Acked-by: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Daniel Drake [Tue, 18 Jul 2006 20:33:27 +0000 (21:33 +0100)]
[PATCH] softmac: ERP handling and driver-level notifications
This patch implements ERP handling in softmac so that the drivers can support
protection and preambles properly.
I added a new struct, ieee80211softmac_bss_info, which is used for
BSS-dependent variables like these.
A new hook has been added (bssinfo_change), which allows the drivers to be
notified when anything in bssinfo changes.
I modified the txrates_change API to match the bssinfo_change API. The
existing one is a little messy and the usefulness of providing the old rates
is questionable (and can be implemented at driver level if really necessary).
No drivers are using this API (yet), so this should be safe.
Signed-off-by: Daniel Drake <dsd@gentoo.org> Acked-by: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Daniel Drake [Tue, 18 Jul 2006 20:30:34 +0000 (21:30 +0100)]
[PATCH] ieee80211: small ERP handling additions
This adds a flag to the ieee80211_network structure which indicates whether
the stored erp_value is valid (a check against 0 is not enough, since an ERP
of 0 is valid and very meaningful).
I also added the ERP IE bit-definitions to ieee80211.h.
This is needed by some upcoming softmac patches.
Signed-off-by: Daniel Drake <dsd@gentoo.org> Acked-by: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Dan Williams [Tue, 18 Jul 2006 01:21:47 +0000 (21:21 -0400)]
[PATCH] prism54: update to WE-19 for WPA support
Add WE-19 capabilities to prism54 fullmac driver so that it supports the
necessary Wireless Extensions WPA calls. Convert reporting of WPA/RSN
Generic Information Elements to IWEVGENIE rather than pre-WE-19
IWEVCUSTOM as well.
Signed-off-by: Dan Williams <dcbw@redhat.com> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Dan Williams [Fri, 14 Jul 2006 15:41:47 +0000 (11:41 -0400)]
[PATCH] orinoco: fix setting transmit key only
When determining whether there's a key to set or not, orinoco should be
looking at the key length, not the key data. Otherwise confusion reigns
when trying to set TX key only, passing in zero-length key, but non-NULL
pointer. Key length takes precedence over non-NULL key data.
Signed-off-by: Dan Williams <dcbw@redhat.com> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Daniel Drake [Tue, 11 Jul 2006 22:16:34 +0000 (23:16 +0100)]
[PATCH] softmac: do shared key auth in workqueue
Johann Uhrmann reported a bcm43xx crash and Michael Buesch tracked
it down to a problem with the new shared key auth code (recursive
calls into the driver)
This patch (effectively Michael's patch with a couple of small
modifications) solves the problem by sending the authentication
challenge response frame from a workqueue entry.
I also removed a lone \n from the bcm43xx messages relating to
authentication mode - this small change was previously discussed but
not patched in.
Signed-off-by: Daniel Drake <dsd@gentoo.org> Acked-by: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: Michael Buesch <mb@bu3sch.de> Signed-off-by: John W. Linville <linville@tuxdriver.com>
Robert Schulze [Wed, 5 Jul 2006 20:52:43 +0000 (22:52 +0200)]
[PATCH] airo: should select crypto_aes
The driver airo (for Cisco Wlan-Cards) complains about "failed to load
transform for AES", when it is loaded and CRYPTO_AES is not selected
in Kconfig.
Signed-off-by: John W. Linville <linville@tuxdriver.com>
[S390] permanent subchannel busy conditions may cause I/O stall
In special conditions where a subchannel rejects the HALT I/O-
instruction with a busy indication (cc 2), I/O may stall.
I/O request termination logic retries HALT I/O indefinitely
because it expects HALT I/O to alter the subchannel status which
is not true when cc 2 is returned.
In case of a busy indication, try CLEAR I/O instruction immediately.
Signed-off-by: Peter Oberparleiter <peter.oberparleiter@de.ibm.com> Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
[PATCH] fix compile regression for a few scsi drivers
This fixes three drivers to compile again after my patch that removes
the data_cmnd member from struct scsi_cmnd.
The fas216 change is trivial, it should have been using ->cmnd all the
time.
NCR53C9 (which seem to be mostly duplicate driver with esp.c!) is doing
something odd, it should only have looked at ->cmnd before not the saved
copy that is kept for the error handlers sake. Note that it really
should deal with the sync setting themselves but use the generic domain
validation code that get this right - but that's for later let's push
this simple compile fix for now.
And sorry for the late fix for this, I have been busy with OLS and
associated activities last week.
Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
* master.kernel.org:/pub/scm/linux/kernel/git/davem/sparc-2.6:
[SCSI] esp: Fix build.
[SPARC]: Fix SA_STATIC_ALLOC value.
[SPARC64]: Explicitly print return PC when the kernel fault PC is bogus.
Arjan van de Ven [Wed, 26 Jul 2006 13:40:07 +0000 (15:40 +0200)]
[PATCH] Reorganize the cpufreq cpu hotplug locking to not be totally bizare
The patch below moves the cpu hotplugging higher up in the cpufreq
layering; this is needed to avoid recursive taking of the cpu hotplug
lock and to otherwise detangle the mess.
The new rules are:
1. you must do lock_cpu_hotplug() around the following functions:
__cpufreq_driver_target
__cpufreq_governor (for CPUFREQ_GOV_LIMITS operation only)
__cpufreq_set_policy
2. governer methods (.governer) must NOT take the lock_cpu_hotplug()
lock in any way; they are called with the lock taken already
3. if your governer spawns a thread that does things, like calling
__cpufreq_driver_target, your thread must honor rule #1.
4. the policy lock and other cpufreq internal locks nest within
the lock_cpu_hotplug() lock.
I'm not entirely happy about how the __cpufreq_governor rule ended up
(conditional locking rule depending on the argument) but basically all
callers pass this as a constant so it's not too horrible.
The patch also removes the cpufreq_governor() function since during the
locking audit it turned out to be entirely unused (so no need to fix it)
The patch works on my testbox, but it could use more testing
(otoh... it can't be much worse than the current code)
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The recvmsg() for raw socket seems to return random u16 value
from the kernel stack memory since port field is not initialized.
But I'm not sure this patch is correct.
Does raw socket return any information stored in port field?
[ BSD defines RAW IP recvmsg to return a sin_port value of zero.
This is described in Steven's TCP/IP Illustrated Volume 2 on
page 1055, which is discussing the BSD rip_input() implementation. ]
Acked-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Signed-off-by: David S. Miller <davem@davemloft.net>
IP multicast route code was reusing an skb which causes use after free
and double free.
From: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
Note, it is real skb_clone(), not alloc_skb(). Equeued skb contains
the whole half-prepared netlink message plus room for the rest.
It could be also skb_copy(), if we want to be puristic about mangling
cloned data, but original copy is really not going to be used.
Acked-by: Stephen Hemminger <shemminger@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
[PATCH] cciss: fix stall with softirq handling and CFQ
We need to postpone the queue startup until after the softirq
handler has actually finished some requests, otherwise we could
be racing with cciss_softirq_done() and not actually restart
the queue handling.
Patrick McHardy [Tue, 25 Jul 2006 05:54:55 +0000 (22:54 -0700)]
[NETFILTER]: bridge netfilter: add deferred output hooks to feature-removal-schedule
Add bridge netfilter deferred output hooks to feature-removal-schedule
and disable them by default. Until their removal they will be
activated by the physdev match when needed.
Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
Phil Oester [Tue, 25 Jul 2006 05:54:14 +0000 (22:54 -0700)]
[NETFILTER]: xt_pkttype: fix mismatches on locally generated packets
Locally generated broadcast and multicast packets have pkttype set to
PACKET_LOOPBACK instead of PACKET_BROADCAST or PACKET_MULTICAST. This
causes the pkttype match to fail to match packets of either type.
The below patch remedies this by using the daddr as a hint as to
broadcast|multicast. While not pretty, this seems like the only way
to solve the problem short of just noting this as a limitation of the
match.
This resolves netfilter bugzilla #484
Signed-off-by: Phil Oester <kernel@linuxace.com> Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
Patrick McHardy [Tue, 25 Jul 2006 05:52:47 +0000 (22:52 -0700)]
[NETFILTER]: nf_queue: handle NF_STOP and unknown verdicts in nf_reinject
In case of an unknown verdict or NF_STOP the packet leaks. Unknown verdicts
can happen when userspace is buggy. Reinject the packet in case of NF_STOP,
drop on unknown verdicts.
Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Tue, 25 Jul 2006 05:47:14 +0000 (22:47 -0700)]
[SCSI] esp: Fix build.
The data_cmd[] member got deleted, so do not use it any more. Scsi
commands do not have their ->cmd[] overwritten temporary to probe for
status after an error before retrying.
Signed-off-by: David S. Miller <davem@davemloft.net>
skbuff.h has an #ifndef CONFIG_HAVE_ARCH_DEV_ALLOC_SKB to allow
architectures to reimplement __dev_alloc_skb. It's not set on any
architecture and now that we have an architecture-overrideable
NET_SKB_PAD there is not point at all to have one either.
Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: David S. Miller <davem@davemloft.net>
Stefan Rompf [Mon, 24 Jul 2006 20:52:13 +0000 (13:52 -0700)]
[VLAN]: Fix link state propagation
When the queue of the underlying device is stopped at initialization time
or the device is marked "not present", the state will be propagated to the
vlan device and never change. Based on an analysis by Patrick McHardy.
Signed-off-by: Stefan Rompf <stefan@loplof.de> ACKed-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller [Mon, 24 Jul 2006 20:49:06 +0000 (13:49 -0700)]
[IPV6] xfrm6_tunnel: Delete debugging code.
It doesn't compile, and it's dubious in several regards:
1) is enabled by non-Kconfig controlled CONFIG_* value
(noted by Randy Dunlap)
2) XFRM6_TUNNEL_SPI_MAGIC is defined after it's first use
3) the debugging messages print object pointer addresses
which have no meaning without context
So let's just get rid of it.
Signed-off-by: David S. Miller <davem@davemloft.net>
[Bluetooth] Enable SCO support for Broadcom HID proxy dongle
The Broadcom dongles with HID proxy support actually support SCO over
HCI if the SCO buffer size values are corrected. So instead of disabling
the SCO support, mark this dongle with the quirk for the Bluetooth core
to correct the wrong buffer size values.
[Bluetooth] Add quirk for another broken RTX Telecom based dongle
This patch disables the ISOC transfers for another broken RTX Telecom
based USB dongle. Starting the USB ISOC transfers only ends in a burst
of error messages for invalid SCO packets on connection handle 0.
[Bluetooth] Correct SCO buffer size for Belkin devices
The Belkin F8T012 and F8T013 devices are both based on a Bluetooth chip
from Broadcom and their SCO buffer size values are wrong. The Bluetooth
core should correct these values.
[Bluetooth] Correct SCO buffer size for another Broadcom chip
The SCO buffer size values on IBM/Lenovo ThinkPad laptops with a
Bluetooth chip from Broadcom are wrong. The USB Bluetooth driver
has to set a quirk to correct the SCO buffer size values.
[Bluetooth] Correct RFCOMM channel MTU for broken implementations
Some Bluetooth RFCOMM implementations try to negotiate a bigger channel
MTU than we can support for a particular session. The maximum MTU for
a RFCOMM session is limited through the L2CAP layer. So if the other
side proposes a channel MTU that is bigger than the underlying L2CAP
MTU, we should reduce it to the L2CAP MTU of the session minus five
bytes for the RFCOMM headers.
In PSCHED_TADD and PSCHED_TADD2, if delta is less than tv.tv_usec (so,
less than USEC_PER_SEC too) then tv_res will be smaller than tv. The
affectation "(tv_res).tv_usec = __delta;" is wrong. The fix is to
revert to the original code before 4ee303dfeac6451b402e3d8512723d3a0f861857 and change the 'if' in
'while'.
I went to alter the default sysctl and it didn't take for new sockets.
Below patch fixes this.
I think the default is too low but it is what the DCCP spec specifies.
As a side effect of this my rx speed using iperf goes from about 2.8 Mbits/sec
to 3.5. This is still far too slow but it is a step in the right direction.
Compile tested only for IPv6 but not particularly complex change.
Signed off by: Ian McDonald <ian.mcdonald@jandi.co.nz> Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com> Signed-off-by: David S. Miller <davem@davemloft.net>
IB/ipoib: Fix packet loss after hardware address update
The neighbour ha field may get updated without destroying the
neighbour. In this case, the ha field gets out of sync with the
address handle stored in ipoib_neigh->ah, with the result that
the ah field would point to an incorrect path, resulting in all
packets being lost.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Sean Hefty [Thu, 20 Jul 2006 08:25:50 +0000 (11:25 +0300)]
IB/mad: Validate MADs for spec compliance
Validate MADs sent by userspace clients for spec compliance with
C13-18.1.1 (prevent duplicate requests and responses sent on the
same port). Without this, RMPP transactions get aborted because
of duplicate packets.
This patch is similar to that provided by Jack Morgenstein.
Signed-off-by: Sean Hefty <sean.hefty@intel.com> Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il> Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Ralph Campbell [Tue, 18 Jul 2006 01:21:24 +0000 (18:21 -0700)]
IB/ipath: ipath_skip_sge() can break if num_sge > 1
ipath_skip_sge() doesn't exactly duplicate the side effects of
ipath_copy_sge() if num_sge > 1 since it doesn't decrement ss->num_sge.
This could result in the sg_list being accessed out of bounds.
Since ipath_skip_sge() is almost always called with num_sge == 1,
the original "optimization" is almost never used.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Ralph Campbell [Tue, 18 Jul 2006 01:19:54 +0000 (18:19 -0700)]
IB/ipath: Fix ib_ipath driver to work with SRP
I am still working on a proposal to remove the phys_to_virt() calls
in the ib_ipath driver. In the mean time, this patch allows SRP
to work by fixing the R_Key check and conversion from IB address
to kernel virtual address. It also returns the correct page size
for FMRs.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Ralph Campbell [Tue, 18 Jul 2006 01:18:36 +0000 (18:18 -0700)]
IB/ipath: Fix a data corruption
This patch fixes a problem where certain error packets are passed
to the InfiniBand layer for processing even though the packet
actually was received with an error.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Mem-free HCAs always keep one spare SRQ WQE, so the SRQ limit cannot
be set beyond srq->max - 1.
Signed-off-by: Dotan Barak <dotanb@mellanox.co.il> Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Roland Dreier [Sun, 23 Jul 2006 22:16:04 +0000 (15:16 -0700)]
IB/uverbs: Fix lockdep warnings
Lockdep warns because uverbs is trying to take uobj->mutex when it
already holds that lock. This is because there are really multiple
types of uobjs even though all of their locks are initialized in
common code.
ib_uverbs_create_ah() and ib_uverbs_create_srq() did not release the
PD's read lock in their error paths, which lead to deadlock when
destroying the PD.
Signed-off-by: Michael S. Tsirkin <mst@mellanox.co.il> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Paul Jackson [Sun, 23 Jul 2006 18:36:08 +0000 (11:36 -0700)]
[PATCH] Cpuset: fix ABBA deadlock with cpu hotplug lock
Fix ABBA deadlock between lock_cpu_hotplug() and the cpuset
callback_mutex lock.
It only happens on cpu_exclusive cpusets, due to the dynamic
sched domain code trying to take the cpu hotplug lock inside
the cpuset callback_mutex lock.
This bug has apparently been here for several months, but didn't
get hit until the right customer load on a large system.
This fix appears right from inspection, but it will take a few
more days running it on that customers workload to be confident
we nailed it. We don't have any other reproducible test case.
The cpu_hotplug_lock() tends to cover large runs of code.
The other places that hold both that lock and the cpuset callback
mutex lock always nest the cpuset lock inside the hotplug lock.
This place tries to do the reverse, risking an ABBA deadlock.
This is in the cpuset_rmdir() code, where we:
* take the callback_mutex lock
* mark the cpuset CS_REMOVED
* call update_cpu_domains for cpu_exclusive cpusets
* in that call, take the cpu_hotplug lock if the
cpuset is marked for removal.
Thanks to Jack Steiner for identifying this deadlock.
The fix is to tear down the dynamic sched domain before we grab
the cpuset callback_mutex lock. This way, the two locks are
serialized, with the hotplug lock taken and released before
trying for the cpuset lock.
I suspect that this bug was introduced when I changed the
cpuset locking from one lock to two. The dynamic sched domain
dependency on cpu_exclusive cpusets and its hotplug hooks were
added to this code earlier, when cpusets had only a single lock.
It may well have been fine then.
Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
projects / linux-beck.git / log