Linus Torvalds [Thu, 29 Oct 2009 15:10:38 +0000 (08:10 -0700)]
Merge branch 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip
* 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
sched: Do less agressive buddy clearing
sched: Disable SD_PREFER_LOCAL for MC/CPU domains
Linus Torvalds [Thu, 29 Oct 2009 15:10:26 +0000 (08:10 -0700)]
Merge branch 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip
* 'x86-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
x86, UV: Set DELIVERY_MODE=4 for vector=NMI_VECTOR in uv_hub_send_ipi()
x86, UV: Fix and clean up bau code to use uv_gpa_to_pnode()
x86: Don't print number of MCE banks for every CPU
x86, UV: Fix information in __uv_hub_info structure
x86: Document linker script ASSERT() quirk
Pavel Herrmann [Mon, 26 Oct 2009 23:50:19 +0000 (16:50 -0700)]
lis3: better support for hp 6730x
I have learned that the 6730b and 6730s have different accelerometer
orientation, and have modified the driver accordingly (diff attached),
while dropping the wild guess for AMD based 6735 having the same
orientation as Intel based 6730 (this is not true for any other related
series/family, thus is not probable for 673x).
Kevin Hilman [Mon, 26 Oct 2009 23:50:18 +0000 (16:50 -0700)]
cpuidle: always return with interrupts enabled
In the case where cpuidle_idle_call() returns before changing state due to
a need_resched(), it was returning with IRQs disabled.
The idle path assumes that the platform specific idle code returns with
interrupts enabled (although this too is undocumented AFAICT) and on ARM
we have a WARN_ON(!(irqs_disabled()) when returning from the idle loop, so
the user-visible effects were only a warning since interrupts were
eventually re-enabled later.
On x86, this same problem exists, but there is no WARN_ON() to detect it.
As on ARM, the interrupts are eventually re-enabled, so I'm not sure of
any actual bugs triggered by this. It's primarily a
correctness/consistency fix.
This patch ensures IRQs are (re)enabled before returning.
Andy Whitcroft [Mon, 26 Oct 2009 23:50:15 +0000 (16:50 -0700)]
checkpatch: fix __attribute__ matching
In the following code,
union thread_union init_thread_union
__attribute__((__section__(".data.init_task"))) =
{ INIT_THREAD_INFO(init_task) };
There is a non-conforming declaration. It should really be like the
following,
union thread_union init_thread_union
__attribute__((__section__(".data.init_task"))) = {
INIT_THREAD_INFO(init_task)
};
However, checkpatch doesn't catch this right now because it doesn't
correctly evaluate the "__attribute__".
It is not at all clear that we care what preceeds an assignment style
attribute when we find the open brace. Relax the test so we do not need
to check the __attribute__.
Andy Whitcroft [Mon, 26 Oct 2009 23:50:13 +0000 (16:50 -0700)]
checkpatch: correctly stop scanning at the bottom of a hunk
We are allowing context scanning checks to apply against the first line of
context outside at the end of the hunk. This can lead to false matches to
patch names leading to various perl warnings. Correctly stop at the
bottom of the hunk.
Russell King [Mon, 26 Oct 2009 23:50:12 +0000 (16:50 -0700)]
mm: fix sparsemem configuration
Currently, sparsemem is only available if EXPERIMENTAL is enabled.
However, it hasn't ever been marked experimental.
It's been about four years since sparsemem was merged, and we have
platforms which depend on it; allow architectures to decide whether
sparsemem should be the default memory model.
Keith Mannthey [Mon, 26 Oct 2009 23:50:11 +0000 (16:50 -0700)]
edac: i5100 fix initialization code
Allow csrows to properly initialize when the topology only has active
channels on 2 and 3. This new check allows proper detection and
initialization in this topology. Only checking the first mrt that
represented channels 0 and 1 is not sufficient.
I also fixed up the related debug information path. I can submit as a 2nd
patch if needed.
Ira W. Snyder [Mon, 26 Oct 2009 23:50:10 +0000 (16:50 -0700)]
edac: i5400 fix missing CONFIG_PCI define
When building without CONFIG_PCI the edac_pci_idx variable is unused,
causing a build-time warning. Wrap the variable in #ifdef CONFIG_PCI,
just like the rest of the PCI support.
Jeff Roberson [Mon, 26 Oct 2009 23:50:09 +0000 (16:50 -0700)]
edac: i5400 fix csrow mapping
The i5400 EDAC driver has several bugs with chip-select row computation
which most likely lead to bugs in detailed error reporting. Attempts to
contact the authors have gone mostly unanswered so I am presenting my diff
here. I do not subscribe to lkml and would appreciate being kept in the
cc.
The most egregious problem was miscalculating the addresses of MTR
registers after register 0 by assuming they are 32bit rather than 16.
This caused the driver to miss half of the memories. Most motherboards
tend to have only 8 dimm slots and not 16, so this may not have been
noticed before.
Further, the row calculations multiplied the number of dimms several
times, ultimately ending up with a maximum row of 32. The chipset only
supports 4 dimms in each of 4 channels, so csrow could not be higher than
4 unless you use a row per-rank with dual-rank dimms. I opted to
eliminate this behavior as it is confusing to the user and the error
reporting works by slot and not rank. This gives a much clearer view of
memory by slot and channel in /sys.
Mike Frysinger [Mon, 26 Oct 2009 23:50:06 +0000 (16:50 -0700)]
twl4030-gpio: remove __devexit markings from remove func
The gpio_twl4030_probe() function calls gpio_twl4030_remove(), and the
former has __devinit, so the latter cannot use __devexit. Otherwise we
hit the section mismatch warning:
WARNING: drivers/gpio/built-in.o(.devinit.text+0x71a): Section mismatch
in reference from the function _gpio_twl4030_probe() to the function
.devexit.text:_gpio_twl4030_remove()
The function __devinit _gpio_twl4030_probe() references a function
__devexit _gpio_twl4030_remove().
This is often seen when error handling in the init function uses
functionality in the exit path.
The fix is often to remove the __devexit annotation of
_gpio_twl4030_remove() so it may be used outside an exit section.
Johannes Weiner [Mon, 26 Oct 2009 23:50:00 +0000 (16:50 -0700)]
vmscan: order evictable rescue in LRU putback
Isolators putting a page back to the LRU do not hold the page lock, and if
the page is mlocked, another thread might munlock it concurrently.
Expecting this, the putback code re-checks the evictability of a page when
it just moved it to the unevictable list in order to correct its decision.
The problem, however, is that ordering is not garuanteed between setting
PG_lru when moving the page to the list and checking PG_mlocked
afterwards:
#0: #1
spin_lock()
if (TestClearPageMlocked())
if (PageLRU())
move to evictable list
SetPageLRU()
spin_unlock()
if (!PageMlocked())
move to evictable list
The PageMlocked() check may get reordered before SetPageLRU() in #0,
resulting in #0 not moving the still mlocked page, and in #1 failing to
isolate and move the page as well. The page is now stranded on the
unevictable list.
The race condition is very unlikely. The consequence currently is one
page falling off the reclaim grid and eventually getting freed with
PG_unevictable set, which triggers a warning in the page allocator.
TestClearPageMlocked() in #1 already provides full memory barrier
semantics.
This patch adds an explicit full barrier to force ordering between
SetPageLRU() and PageMlocked() so that either one of the competitors
rescues the page.
Jeff Mahoney [Mon, 26 Oct 2009 23:49:56 +0000 (16:49 -0700)]
hfs: fix oops on mount with corrupted btree extent records
A particular fsfuzzer run caused an hfs file system to crash on mount.
This is due to a corrupted MDB extent record causing a miscalculation of
HFS_I(inode)->first_blocks for the extent tree. If the extent records are
zereod out, it won't trigger the first_blocks special case. Instead it
falls through to the extent code which we're still in the middle of
initializing.
This patch catches the 0 size extent records, reports the corruption, and
fails the mount.
Wu Fengguang [Mon, 26 Oct 2009 23:49:53 +0000 (16:49 -0700)]
vmscan: limit VM_EXEC protection to file pages
It is possible to have !Anon but SwapBacked pages, and some apps could
create huge number of such pages with MAP_SHARED|MAP_ANONYMOUS. These
pages go into the ANON lru list, and hence shall not be protected: we only
care mapped executable files. Failing to do so may trigger OOM.
mm: oom analysis: add buffer cache information to show_free_areas()
show_free_areas() is called during page allocation failures, and page
allocation failures can occur in any calling context.
But nr_blockdev_pages() takes VFS locks which should not be taken from
hard IRQ context (at least). The result is lockdep warnings (and
deadlockability) during page allocation failures.
I am not confident that I can find and fix all cases where a sector number
may be truncated. For now, avoid data loss by refusing to mount HFS+
volumes with more than 2^32 sectors (2TB).
Joe Perches [Mon, 26 Oct 2009 23:49:48 +0000 (16:49 -0700)]
MAINTAINERS: document new "K:" entry type
K: is for keyword. Syntax is perl extended regex.
Reorganized header documentation and indent the section entry descriptions
so that the first K: would not be considered a regex to match by
get_maintainer.pl
Joe Perches [Mon, 26 Oct 2009 23:49:47 +0000 (16:49 -0700)]
scripts/get_maintainer.pl: add patch/file search for keywords
Based on an idea from Wolfram Sang.
Add search for MAINTAINERS line "K:" regex pattern match in a patch or file
Matches are added after file pattern matches
Add --keywords command line switch (default 1, on)
Change version to 0.21
KOSAKI Motohiro [Mon, 26 Oct 2009 23:49:37 +0000 (16:49 -0700)]
strstrip(): mark as as must_check
strstrip() can return a modified value of its input argument, when
removing elading whitesapce. So it is surely bug for this function's
return value to be ignored. The caller is probably going to use the
incorrect original pointer.
So mark it __must_check to prevent this frm happening (as it has before).
KOSAKI Motohiro [Mon, 26 Oct 2009 23:49:35 +0000 (16:49 -0700)]
congestion_wait(): don't use WRITE
commit 8aa7e847d (Fix congestion_wait() sync/async vs read/write
confusion) replace WRITE with BLK_RW_ASYNC. Unfortunately, concurrent mm
development made the unchanged place accidentally.
The warning is
---> WARN_ON_ONCE(in_irq() || irqs_disabled());
The network code must not be called with disabled interrupts but
sys_setsid holds the tasklist_lock with spinlock_irq while calling the
connector.
After a discussion we agreed that we can move proc_sid_connector from
__set_special_pids to sys_setsid.
We also agreed that it is sufficient to change the check from
task_session(curr) != pid into err > 0, since if we don't change the
session, this means we were already the leader and return -EPERM.
One last thing:
There is also daemonize(), and some people might want to get a
notification in that case. Since daemonize() is only needed if a user
space does kernel_thread this does not look important (and there seems
to be no consensus if this connector should be called in daemonize). If
we really want this, we can add proc_sid_connector to daemonize() in an
additional patch (Scott?)
Hugh Dickins [Mon, 26 Oct 2009 23:49:32 +0000 (16:49 -0700)]
hwpoison: fix/proc/meminfo alignment
Given such a long name, the kB count in /proc/meminfo's HardwareCorrupted
line is being shown too far right (it does align with x86_64's VmallocChunk
above, but I hope nobody will ever have that much corrupted!). Align it.
Hugh Dickins [Mon, 26 Oct 2009 23:49:31 +0000 (16:49 -0700)]
hwpoison: fix oops on ksm pages
Memory failure on a KSM page currently oopses on its NULL anon_vma in
page_lock_anon_vma(): that may not be much worse than the consequence of
ignoring it, but it is better to be consistent with how ZERO_PAGE and
hugetlb pages and other awkward cases are treated. Just skip it.
We could fix it for 2.6.32 at the KSM end, by putting a dummy anon_vma
pointer in there; but that would get harder next time, when KSM will put a
pointer to something else there (and I'm not currently planning to do any
work to open that up to memory_failure). So I would prefer this simple
PageKsm test, until the other exceptions are handled.
Randy Dunlap [Mon, 26 Oct 2009 23:49:29 +0000 (16:49 -0700)]
cpufreq: add cpufreq_get() stub for CONFIG_CPU_FREQ=n
When CONFIG_CPU_FREQ is disabled, cpufreq_get() needs a stub. Used by kvm
(although it looks like a bit of the kvm code could be omitted when
CONFIG_CPU_FREQ is disabled).
arch/x86/built-in.o: In function `kvm_arch_init':
(.text+0x10de7): undefined reference to `cpufreq_get'
(Needed in linux-next's KVM tree, but it's correct in 2.6.32).
Heiko Carstens [Thu, 29 Oct 2009 14:04:14 +0000 (15:04 +0100)]
[S390] smp: fix sigp sense handling
sigp sense only returns the status of a cpu if it is non zero. If the
status of the sensed cpu is all zeros condition code 0 (accpeted) is
set and no status bits are returned.
The current code however assumes that a status was returned and tests
bits in it. This means uninitalized data is accessed with random
results.
Worst case is that the code that checks if cpu is offline on cpu
hotplug assumes that the target cpu is offline while it is still
running. This leads potentially to memory corruption since resources
that are still needed by the target cpu will be freed and could be
resused while still in use.
Heiko Carstens [Thu, 29 Oct 2009 14:04:13 +0000 (15:04 +0100)]
[S390] smp: fix sigp stop handling
According to the architecture a cpu must not necessarily enter stopped
state after completion of a sigp instruction with "stop" order code.
So remove the BUG() statement after self sending sigp stop to avoid
that it ever gets reached.
Also add a sigp busy check to make sure that the order gets delivered.
The cputime_to_msecs / cputime_to_clock_t and cputime64_to_clock_t
cause fixpoint divide exceptions if the cputime is too large.
On a machine that collected 49.7 days worth of idle time reading
from /proc/stat will generate oopses like this:
Heiko Carstens [Thu, 29 Oct 2009 14:04:11 +0000 (15:04 +0100)]
[S390] call home: fix string length handling
After copying uts->nodename to the static nodename array the static
version isn't necessarily zero termininated, since the size of the
array is one byte too short.
Afterwards doing strncat(data, nodename, strlen(nodename)); may copy
an arbitrary large amount of bytes.
Fix this by getting rid of the static array and using strncat with
proper length limit.
Heiko Carstens [Thu, 29 Oct 2009 14:04:09 +0000 (15:04 +0100)]
[S390] smp: fix prefix handling of offlined cpus
Offlined cpus still have valid prefix register contents. Dumpers
will store the register contents of a cpu to the location where its
prefix register points to.
For offlined cpus the area (lowcore) has been freed and the dumper
would write the uninteresting contents of the offline cpu to a memory
location which might be in use by some other component and destroy
valueable information.
To fix this set the prefix register of offline cpus to absolute
address zero again. This prevents the current dumpers to write to
random memory locations.
If a suspended z/VM guest has been logged off before the resume the
'SET SMSG IUCV' CP command need to be repeated to reenable sending
message via SMSG. This fixes the following error:
HCPMFS057I H4214002 not receiving; SMSG off
Error: non-zero CP response for command 'SMSG H4214002 CMM SHRINK 5010': #57
Currently there is no barrier support in the block device code. That
means we cannot guarantee any sort of data integerity when using the
block device node with dis kwrite caches enabled. Using the raw block
device node is a typical use case for virtualization (and I assume
databases, too). This patch changes block_fsync to issue a cache flush
and thus make fsync on block device nodes actually useful.
Note that in mainline we would also need to add such code to the
->aio_write method for O_SYNC handling, but assuming that Jan's patch
series for the O_SYNC rewrite goes in it will also call into ->fsync
for 2.6.32.
Jens Axboe [Thu, 29 Oct 2009 12:59:26 +0000 (13:59 +0100)]
block: move bdi/address_space unplug functions to backing-dev.h
There's nothing block related about them, the backing device
is used by things like NFS etc as well. This gets rid of the
need to protect such calls by CONFIG_BLOCK.
Jens Axboe [Thu, 29 Oct 2009 10:46:12 +0000 (11:46 +0100)]
backing-dev: ensure that a removed bdi no longer has super_block referencing it
When the bdi is being removed, we have to ensure that no super_blocks
currently have that cached in sb->s_bdi. Normally this is ensured by
the sb having a longer life span than the bdi, but if the device is
suddenly yanked, we have to kill this reference. sb->s_bdi is pointed
to freed memory at that point.
This fixes a problem with sync(1) hanging when a USB stick is pulled
without cleanly umounting it first.
Gabor Gombas [Thu, 29 Oct 2009 10:19:11 +0000 (03:19 -0700)]
net: Fix 'Re: PACKET_TX_RING: packet size is too long'
Currently PACKET_TX_RING forces certain amount of every frame to remain
unused. This probably originates from an early version of the
PACKET_TX_RING patch that in fact used the extra space when the (since
removed) CONFIG_PACKET_MMAP_ZERO_COPY option was enabled. The current
code does not make any use of this extra space.
This patch removes the extra space reservation and lets userspace make
use of the full frame size.
Mikulas Patocka [Wed, 21 Oct 2009 08:55:28 +0000 (08:55 +0000)]
ide: Serialize CMD643 and CMD646 to fix a hardware bug with SSD
CMD646 corrupts data on concurrent transfers on both channels when IDE SSD is
connected to one of the channels.
Setup that demonstrates this hardware bug: Ultra 5, onboard CMD646, rev 3.
/dev/hda is 8GB Seagate ST38410A in MWDMA2
/dev/hdd is 32GB SSD SiliconHardDisk in MWDMA2
- When reading /dev/hdd (for example with dd or fsck), reads from /dev/hda
are corrupted, there are twiddled single bits 1->0 and some full 32-bit
words corrupted, sometimes commands fail (which switches /dev/hda to
PIO mode but the corruptions happen even in PIO).
- Reads from /dev/hdd don't seem to be corrupted (i.e. fsck passes fine).
- When I connected normal rotating harddisk to /dev/hdd, there was no
corruption, so the corruption is something specific to SSD.
- I tried the same setup on a PCI card with CMD649 and saw no corruption.
This patch serializes the operation for CMD646 and 643 (I didn't test
CMD643 but it may have the same hw bug too because it's earlier design).
CMD649 is good. I don't know anything about CMD 648.
netdev: usb: dm9601.c can drive a device not supported yet, add support for it
I found that the current version of drivers/net/usb/dm9601.c can be used to
successfully drive a low-power, low-cost network adapter with USB ID
0a46:9000, based on a DM9000E chipset. As no device with this ID is yet
present in the kernel, I have created a patch that adds support for the device
to the dm9601 driver.
Ron Mercer [Wed, 28 Oct 2009 08:39:21 +0000 (08:39 +0000)]
qlge: Fix firmware mailbox command timeout.
The mailbox command process would only process a maximum of 5 unrelated
firmware events while waiting for it's command completion status.
It should process an unlimited number of events while waiting for a maximum of 5 seconds.
Neil Horman [Wed, 28 Oct 2009 08:59:47 +0000 (08:59 +0000)]
AF_RAW: Augment raw_send_hdrinc to expand skb to fit iphdr->ihl (v2)
Augment raw_send_hdrinc to correct for incorrect ip header length values
A series of oopses was reported to me recently. Apparently when using AF_RAW
sockets to send data to peers that were reachable via ipsec encapsulation,
people could panic or BUG halt their systems.
I've tracked the problem down to user space sending an invalid ip header over an
AF_RAW socket with IP_HDRINCL set to 1.
Basically what happens is that userspace sends down an ip frame that includes
only the header (no data), but sets the ip header ihl value to a large number,
one that is larger than the total amount of data passed to the sendmsg call. In
raw_send_hdrincl, we allocate an skb based on the size of the data in the msghdr
that was passed in, but assume the data is all valid. Later during ipsec
encapsulation, xfrm4_tranport_output moves the entire frame back in the skbuff
to provide headroom for the ipsec headers. During this operation, the
skb->transport_header is repointed to a spot computed by
skb->network_header + the ip header length (ihl). Since so little data was
passed in relative to the value of ihl provided by the raw socket, we point
transport header to an unknown location, resulting in various crashes.
This fix for this is pretty straightforward, simply validate the value of of
iph->ihl when sending over a raw socket. If (iph->ihl*4U) > user data buffer
size, drop the frame and return -EINVAL. I just confirmed this fixes the
reported crashes.
Jiri Bohac [Thu, 29 Oct 2009 05:23:54 +0000 (22:23 -0700)]
bonding: fix a race condition in calls to slave MII ioctls
In mii monitor mode, bond_check_dev_link() calls the the ioctl
handler of slave devices. It stores the ndo_do_ioctl function
pointer to a static (!) ioctl variable and later uses it to call the
handler with the IOCTL macro.
If another thread executes bond_check_dev_link() at the same time
(even with a different bond, which none of the locks prevent), a
race condition occurs. If the two racing slaves have different
drivers, this may result in one driver's ioctl handler being
called with a pointer to a net_device controlled with a different
driver, resulting in unpredictable breakage.
Unless I am overlooking something, the "static" must be a
copy'n'paste error (?).
Rusty Russell [Thu, 29 Oct 2009 14:56:19 +0000 (08:56 -0600)]
param: fix setting arrays of bool
We create a dummy struct kernel_param on the stack for parsing each
array element, but we didn't initialize the flags word. This matters
for arrays of type "bool", where the flag indicates if it really is
an array of bools or unsigned int (old-style).
Rusty Russell [Thu, 29 Oct 2009 14:56:16 +0000 (08:56 -0600)]
param: fix lots of bugs with writing charp params from sysfs, by leaking mem.
e180a6b7759a "param: fix charp parameters set via sysfs" fixed the case
where charp parameters written via sysfs were freed, leaving drivers
accessing random memory.
Unfortunately, storing a flag in the kparam struct was a bad idea: it's
rodata so setting it causes an oops on some archs. But that's not all:
1) module_param_array() on charp doesn't work reliably, since we use an
uninitialized temporary struct kernel_param.
2) there's a fundamental race if a module uses this parameter and then
it's changed: they will still access the old, freed, memory.
The simplest fix (ie. for 2.6.32) is to never free the memory. This
prevents all these problems, at cost of a memory leak. In practice, there
are only 18 places where a charp is writable via sysfs, and all are
root-only writable.
On SMP guests, reads from the ring might bypass used index reads. This
causes guest crashes because host writes to used index to signal ring
data readiness. Fix this by inserting rmb before used ring reads.
Commit f68d24082e22ccee3077d11aeb6dc5354f0ca7f1
in 2.6.32-rc1 broke requesting IRQs for per-VQ MSI-X vectors:
- vector number was used instead of the vector itself
- we try to request an IRQ for VQ which does not
have a callback handler
This is a regression that causes warnings in kernel log,
potentially lower performance as we need to scan vq list,
and might cause system failure if the interrupt
requested is in fact needed by another system.
This was not noticed earlier because in most cases
we were falling back on shared interrupt for all vqs.
Jiri Kosina [Wed, 28 Oct 2009 15:26:00 +0000 (00:26 +0900)]
sched: move rq_weight data array out of .percpu
Commit 34d76c41 introduced percpu array update_shares_data, size of which
being proportional to NR_CPUS. Unfortunately this blows up ia64 for large
NR_CPUS configuration, as ia64 allows only 64k for .percpu section.
Fix this by allocating this array dynamically and keep only pointer to it
percpu.
The per-cpu handling doesn't impose significant performance penalty on
potentially contented path in tg_shares_up().
Jiri Kosina [Wed, 28 Oct 2009 15:25:59 +0000 (00:25 +0900)]
percpu: allow pcpu_alloc() to be called with IRQs off
pcpu_alloc() and pcpu_extend_area_map() perform a series of
spin_lock_irq()/spin_unlock_irq() calls, which make them unsafe
with respect to being called from contexts which have IRQs off.
This patch converts the code to perform save/restore of flags instead,
making pcpu_alloc() (or __alloc_percpu() respectively) to be called
from early kernel startup stage, where IRQs are off.
This is needed for proper initialization of per-cpu rq_weight data from
sched_init().
tj: added comment explaining why irqsave/restore is used in alloc path.
virtio net used to unlink skbs from send queues on error,
but ever since 48925e372f04f5e35fec6269127c62b2c71ab794
we do not do this. This causes guest data corruption and crashes
with vhost since net core can requeue the skb or free it without
it being taken off the list.
This patch fixes this by queueing the skb after successful
transmit.
Ben Hutchings [Wed, 28 Oct 2009 10:43:49 +0000 (03:43 -0700)]
sfc: Set ip_summed correctly for page buffers passed to GRO
Page buffers containing packets with an incorrect checksum or using a
protocol not handled by hardware checksum offload were previously not
passed to LRO. The conversion to GRO changed this, but did not set
the ip_summed value accordingly.
Jeff Moyer [Fri, 2 Oct 2009 22:57:36 +0000 (18:57 -0400)]
aio: implement request batching
Hi,
Some workloads issue batches of small I/O, and the performance is poor
due to the call to blk_run_address_space for every single iocb. Nathan
Roberts pointed this out, and suggested that by deferring this call
until all I/Os in the iocb array are submitted to the block layer, we
can realize some impressive performance gains (up to 30% for sequential
4k reads in batches of 16).
Jeff Moyer [Fri, 2 Oct 2009 22:56:53 +0000 (18:56 -0400)]
block: get rid of the WRITE_ODIRECT flag
Hi,
The WRITE_ODIRECT flag is only used in one place, and that code path
happens to also call blk_run_address_space. The introduction of this
flag, then, could result in the device being unplugged twice for every
I/O.
Further, with the batching changes in the next patch, we don't want an
O_DIRECT write to imply a queue unplug.
Corrado Zoccolo [Mon, 26 Oct 2009 21:45:29 +0000 (22:45 +0100)]
cfq-iosched: fairness for sync no-idle queues
Currently no-idle queues in cfq are not serviced fairly:
even if they can only dispatch a small number of requests at a time,
they have to compete with idling queues to be serviced, experiencing
large latencies.
We should notice, instead, that no-idle queues are the ones that would
benefit most from having low latency, in fact they are any of:
* processes with large think times (e.g. interactive ones like file
managers)
* seeky (e.g. programs faulting in their code at startup)
* or marked as no-idle from upper levels, to improve latencies of those
requests.
This patch improves the fairness and latency for those queues, by:
* separating sync idle, sync no-idle and async queues in separate
service_trees, for each priority
* service all no-idle queues together
* and idling when the last no-idle queue has been serviced, to
anticipate for more no-idle work
* the timeslices allotted for idle and no-idle service_trees are
computed proportionally to the number of processes in each set.
Servicing all no-idle queues together should have a performance boost
for NCQ-capable drives, without compromising fairness.
Corrado Zoccolo [Mon, 26 Oct 2009 21:45:11 +0000 (22:45 +0100)]
cfq-iosched: enable idling for last queue on priority class
cfq can disable idling for queues in various circumstances.
When workloads of different priorities are competing, if the higher
priority queue has idling disabled, lower priority queues may steal
its disk share. For example, in a scenario with an RT process
performing seeky reads vs a BE process performing sequential reads,
on an NCQ enabled hardware, with low_latency unset,
the RT process will dispatch only the few pending requests every full
slice of service for the BE process.
The patch solves this issue by always performing idle on the last
queue at a given priority class > idle. If the same process, or one
that can pre-empt it (so at the same priority or higher), submits a
new request within the idle window, the lower priority queue won't
dispatch, saving the disk bandwidth for higher priority ones.
Note: this doesn't touch the non_rotational + NCQ case (no hardware
to test if this is a benefit in that case).
Corrado Zoccolo [Tue, 27 Oct 2009 18:16:03 +0000 (19:16 +0100)]
cfq-iosched: reimplement priorities using different service trees
We use different service trees for different priority classes.
This allows a simplification in the service tree insertion code, that no
longer has to consider priority while walking the tree.
Corrado Zoccolo [Mon, 26 Oct 2009 21:44:33 +0000 (22:44 +0100)]
cfq-iosched: preparation to handle multiple service trees
We embed a pointer to the service tree in each queue, to handle multiple
service trees easily.
Service trees are enriched with a counter.
cfq_add_rq_rb is invoked after putting the rq in the fifo, to ensure
that all fields in rq are properly initialized.
Corrado Zoccolo [Mon, 26 Oct 2009 21:44:04 +0000 (22:44 +0100)]
cfq-iosched: adapt slice to number of processes doing I/O
When the number of processes performing I/O concurrently increases,
a fixed time slice per process will cause large latencies.
This patch, if low_latency mode is enabled, will scale the time slice
assigned to each process according to a 300ms target latency.
In order to keep fairness among processes:
* The number of active processes is computed using a special form of
running average, that quickly follows sudden increases (to keep latency low),
and decrease slowly (to have fairness in spite of rapid decreases of this
value).
To safeguard sequential bandwidth, we impose a minimum time slice
(computed using 2*cfq_slice_idle as base, adjusted according to priority
and async-ness).
Dave Airlie [Fri, 23 Oct 2009 08:49:03 +0000 (18:49 +1000)]
drm/kms: fix kms/fbdev colormap support properly.
This sets the fbcon to use TRUECOLOR by default, it then
only modifies the pseudo palette for fbcon, and only touches
the real palette when in 8-bit pseudo color mode.
Zhao Yakui [Wed, 14 Oct 2009 01:11:25 +0000 (09:11 +0800)]
drm: Add the basic check for the detailed timing in EDID
Sometimes we will get the incorrect display modeline when parsing the detailed
timing in EDID. For example:
>hsync/vsync width is zero
>sync is beyond the blank.
So add the basic check for the detailed timing in EDID to avoid the incorrect
display modeline.
Dave Airlie [Wed, 28 Oct 2009 01:09:58 +0000 (11:09 +1000)]
drm/radeon/kms: ignore vga arbiter return.
Since we register all radeon devices, and the arbiter only cares about
VGA class ones, we will fail to startup on display controller class devices.
We don't gain anything by using the return value here.
xen_setup_stackprotector() ends up trying to set page protections,
so we need to have vm_mmu_ops set up before trying to do so.
Failing to do so causes an early boot crash.
projects / linux.git / log