If the TSC is constant and non-stop, also set it reliable.
(We will turn this off in DMI quirks for multi-chassis systems)
The performance number on a 16-way Nehalem system running
32 tasks that context-switch between each other is significant:
sched_clock_stable=0 sched_clock_stable=1
.................... ....................
22.456925 million/sec 24.306972 million/sec [+8.2%]
lmbench's "lat_ctx -s 0 2" goes from 0.63 microseconds to
0.59 microseconds - a 6.7% increase in context-switching
performance.
Perfstat of 1 million pipe context switches between two tasks:
Performance counter stats for './pipe-test-1m':
[before] [after]
............ ............
37621.421089 36436.848378 task clock ticks (msecs)
0 0 CPU migrations (events)
2000274 2000189 context switches (events)
194 193 pagefaults (events)
8433799643 8171016416 CPU cycles (events) -3.21%
8370133368 8180999694 instructions (events) -2.31%
4158565 3895941 cache references (events) -6.74%
44312 46264 cache misses (events)
2349.287976 2279.362465 wall-time (msecs) -3.06%
The speedup comes straight from the reduction in the instruction
count. sched_clock_cpu() got simpler and the whole workload thus
executes faster.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Allow CONFIG_HAVE_UNSTABLE_SCHED_CLOCK architectures to still specify
that their sched_clock() implementation is reliable.
This will be used by x86 to switch on a faster sched_clock_cpu()
implementation on certain CPU types.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'for-linus' of git://git.kernel.dk/linux-2.6-block:
xen/blkfront: use blk_rq_map_sg to generate ring entries
block: reduce stack footprint of blk_recount_segments()
cciss: shorten 30s timeout on controller reset
block: add documentation for register_blkdev()
block: fix bogus gcc warning for uninitialized var usage
It needs to happen before any firewire driver actually registers itself,
and that was previously handled by having the Makefile list the core
ieee1394 files before the drivers.
But now there are firewire drivers in drivers/media, and the Makefile
games aren't enough. So just make ieee1394_init happen earlier in the
init sequence, the way all other bus layers already do.
Reported-and-tested-by: Ingo Molnar <mingo@elte.hu>
Cc: Stefan Richter <stefanr@s5r6.in-berlin.de>
Cc: Henrik Kurelid <henrik@kurelid.se>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: Ben Backx <ben@bbackx.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Impact: implement new tracing timestamp APIs
Add three trace clock variants, with differing scalability/precision
tradeoffs:
- local: CPU-local trace clock
- medium: scalable global clock with some jitter
- global: globally monotonic, serialized clock
Make the ring-buffer use the local trace clock internally.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
make sure we dont execute more complex sched_clock() code in NMI context.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: add new tracepoints
Add them to the generic IRQ code, that way every architecture
gets these new tracepoints, not just x86.
Using Steve's new 'TRACE_FORMAT', I can get function graph
trace as follows using the original two IRQ tracepoints:
3) | handle_IRQ_event() {
3) | /* (irq_handler_entry) irq=28 handler=eth0 */
3) | e1000_intr_msi() {
3) 2.460 us | __napi_schedule();
3) 9.416 us | }
3) | /* (irq_handler_exit) irq=28 handler=eth0 return=handled */
3) + 22.935 us | }
Signed-off-by: Jason Baron <jbaron@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Masami Hiramatsu <mhiramat@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Mathieu Desnoyers <compudj@krystal.dyndns.org>
Cc: "Frank Ch. Eigler" <fche@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: restructure the VFS layout of per CPU trace buffers
The per cpu trace files are all in a single directory:
/debug/tracing/per_cpu. In case of a large number of cpu, the
content of this directory becomes messy so we create now one
directory per cpu inside /debug/tracing/per_cpu which contain
each their own trace_pipe and trace files.
Ie:
/debug/tracing$ ls -R per_cpu
per_cpu:
cpu0 cpu1
per_cpu/cpu0:
trace trace_pipe
per_cpu/cpu1:
trace trace_pipe
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
On occasion, the request will apparently have more segments than we
fit into the ring. Jens says:
> The second problem is that the block layer then appears to create one
> too many segments, but from the dump it has rq->nr_phys_segments ==
> BLKIF_MAX_SEGMENTS_PER_REQUEST. I suspect the latter is due to
> xen-blkfront not handling the merging on its own. It should check that
> the new page doesn't form part of the previous page. The
> rq_for_each_segment() iterates all single bits in the request, not dma
> segments. The "easiest" way to do this is to call blk_rq_map_sg() and
> then iterate the mapped sg list. That will give you what you are
> looking for.
> Here's a test patch, compiles but otherwise untested. I spent more
> time figuring out how to enable XEN than to code it up, so YMMV!
> Probably the sg list wants to be put inside the ring and only
> initialized on allocation, then you can get rid of the sg on stack and
> sg_init_table() loop call in the function. I'll leave that, and the
> testing, to you.
[Moved sg array into info structure, and initialize once. -J]
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
If reset_devices is set for kexec, then cciss will delay 30 seconds
since the old 5i controller _may_ need that long to recover. Replace
the long sleep with incremental sleep and tests to reduce the 30 seconds
to worst case for 5i, so that other controllers will proceed quickly.
Reviewed-by: Mike Miller <mike.miller@hp.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Add documentation for register_blkdev() function and for the parameters.
Signed-off-by: Márton Németh <nm127@freemail.hu>
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Newer gcc throw this warning:
fs/bio.c: In function ?bio_alloc_bioset?:
fs/bio.c:305: warning: ?p? may be used uninitialized in this function
since it cannot figure out that 'p' is only ever used if 'bs' is non-NULL.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
This fixes a regression introduced by commit
a4e22f02f5 ("powerpc: Update 64bit
__copy_tofrom_user() using CPU_FTR_UNALIGNED_LD_STD").
The same bug that existed in the 64bit memcpy() also exists here so fix
it here too. The fix is the same as that applied to memcpy() with the
addition of fixes for the exception handling code required for
__copy_tofrom_user().
This stops us reading beyond the end of the source region we were told
to copy.
Signed-off-by: Mark Nelson <markn@au1.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This fixes a regression introduced by commit
25d6e2d7c5 ("powerpc: Update 64bit memcpy()
using CPU_FTR_UNALIGNED_LD_STD").
This commit allowed CPUs that have the CPU_FTR_UNALIGNED_LD_STD CPU
feature bit present to do the memcpy() with unaligned load doubles. But,
along with this came a bug where our final load double would read bytes
beyond a page boundary and into the next (unmapped) page. This was caught
by enabling CONFIG_DEBUG_PAGEALLOC,
The fix was to read only the number of bytes that we need to store rather
than reading a full 8-byte doubleword and storing only a portion of that.
In order to minimise the amount of existing code touched we use the
original do_tail for the src_unaligned case.
Below is an example of the regression, as reported by Sachin Sant:
Unable to handle kernel paging request for data at address 0xc00000003f380000
Faulting instruction address: 0xc000000000039574
cpu 0x1: Vector: 300 (Data Access) at [c00000003baf3020]
pc: c000000000039574: .memcpy+0x74/0x244
lr: d00000000244916c: .ext3_xattr_get+0x288/0x2f4 [ext3]
sp: c00000003baf32a0
msr: 8000000000009032
dar: c00000003f380000
dsisr: 40000000
current = 0xc00000003e54b010
paca = 0xc000000000a53680
pid = 1840, comm = readahead
enter ? for help
[link register ] d00000000244916c .ext3_xattr_get+0x288/0x2f4 [ext3]
[c00000003baf32a0] d000000002449104 .ext3_xattr_get+0x220/0x2f4 [ext3]
(unreliab
le)
[c00000003baf3390] d00000000244a6e8 .ext3_xattr_security_get+0x40/0x5c [ext3]
[c00000003baf3400] c000000000148154 .generic_getxattr+0x74/0x9c
[c00000003baf34a0] c000000000333400 .inode_doinit_with_dentry+0x1c4/0x678
[c00000003baf3560] c00000000032c6b0 .security_d_instantiate+0x50/0x68
[c00000003baf35e0] c00000000013c818 .d_instantiate+0x78/0x9c
[c00000003baf3680] c00000000013ced0 .d_splice_alias+0xf0/0x120
[c00000003baf3720] d00000000243e05c .ext3_lookup+0xec/0x134 [ext3]
[c00000003baf37c0] c000000000131e74 .do_lookup+0x110/0x260
[c00000003baf3880] c000000000134ed0 .__link_path_walk+0xa98/0x1010
[c00000003baf3970] c0000000001354a0 .path_walk+0x58/0xc4
[c00000003baf3a20] c000000000135720 .do_path_lookup+0x138/0x1e4
[c00000003baf3ad0] c00000000013645c .path_lookup_open+0x6c/0xc8
[c00000003baf3b70] c000000000136780 .do_filp_open+0xcc/0x874
[c00000003baf3d10] c0000000001251e0 .do_sys_open+0x80/0x140
[c00000003baf3dc0] c00000000016aaec .compat_sys_open+0x24/0x38
[c00000003baf3e30] c00000000000855c syscall_exit+0x0/0x40
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When we introduced VSX, we changed the way FPRs are stored in the
thread_struct. Unfortunately we missed the load/store float double
alignment handler code when updating how we access FPRs in the
thread_struct.
Below fixes this and merges the little/big endian case.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Peter Zijlstra warned that TPPROTO and TPARGS might become something
other than a simple copy of itself. To prevent this from having
side effects in the TRACE_FORMAT macro in tracepoint.h, we add a
PARAMS() macro to be defined as just a wrapper.
Reported-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
There's been a bit confusion to whether DEFINE/DECLARE_TRACE_FMT should
be a DEFINE or a DECLARE. Ingo Molnar suggested simply calling it
TRACE_FORMAT.
Reported-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
* 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6:
[IA64] Don't go beyond iosapic_intr_info's arraysize
[IA64] Do not go beyond ARRAY_SIZE of unw.hash
[IA64] enable setting DMAR on by default
* 'upstream-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jgarzik/libata-dev:
[libata] pata_legacy: for VLB 32bit PIO don't try tricks with slop
[libata] pata_amd: program FIFO
sata_mv: fix SoC interrupt breakage
pata_it821x: resume from hibernation fails with RAID volume
These devices are generally used with ATA anyway and it seems that some
ATAPI will need us to issue the right number of words. Therefore as we
can't switch mid burst on VLB devices we should only use 32bit I/O for
suitable block sizes.
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
With 32bit PIO we can use the posted write buffers, but only for 32bit I/O
cycles. This means we must disable the FIFO for ATAPI where a final 16bit
cycle may occur.
Rework the FIFO logic so that we disable the FIFO then selectively
re-enable it when we set the timings on AMD devices. Also fix a case
where we scribbled on PCI config 0x41 of Nvidia chips when we shouldn't.
Signed-off-by: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
For some reason, sata_mv doesn't clear interrupt status during init
when it's running on an SoC host adapter. If the bootloader has
touched the SATA controller before starting Linux, Linux can end up
enabling the SATA interrupt with events pending, which will cause the
interrupt to be marked as spurious and then be disabled, which then
breaks all further accesses to the controller.
This patch makes the SoC path clear interrupt status on init like in
the non-SoC case.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Signed-off-by: Mark Lord <mlord@pobox.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Hibernation didn't work for me since I started to use IT8212 controller.
I did some debugging (booting with no_console_suspend init=/bin/sh).
Found that resume fails (2.6.28) with "serial number mismatch 'some
garbage' != 'some other garbage'" and "revalidation failed" messages.
That's because the controller firmware fills different serial number in
the IDENTIFY every boot.
The patch below fixes the resume simply clearing the serial number. The
proper fix would be probably to fill in the serial number of the RAID
volume instead. I assume that there must be something like that stored on
the drives but I don't know where.
Fix resume on pata_it821x RAID volume by clearing the serial number in
IDENTIFY data, which is otherwise different on each boot.
Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Acked-by: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Each time I exit Firefox, /proc/meminfo's Committed_AS goes down almost
400 kB: OVERCOMMIT_NEVER would be allowing overcommits it should
prohibit.
Commit fc8744adc8 "Stop playing silly
games with the VM_ACCOUNT flag" changed shmem_file_setup() to set the
shmem file's VM_ACCOUNT flag according to VM_NORESERVE not being set in
the vma flags; but did so only _after_ the shmem_acct_size(flags, size)
call which is expected to pre-account a shared anonymous object.
It's all clearer if we switch shmem.c over to use VM_NORESERVE
throughout in place of !VM_ACCOUNT.
But I very nearly sent in a patch which mistakenly removed the
accounting from tmpfs files: shmem_get_inode()'s memset was good for not
setting VM_ACCOUNT, but now it needs to set VM_NORESERVE.
Rather than setting that by default, then perhaps clearing it again in
shmem_file_setup(), let's pass it as a flag to shmem_get_inode(): that
allows us to remove the #ifdef CONFIG_SHMEM from shmem_file_setup().
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
vi arch/ia64/kernel/iosapic.c +142
static struct iosapic_intr_info {
...
} iosapic_intr_info[NR_IRQS];
But at line 510 we have:
for (i = 0; i <= NR_IRQS; i++) {
s/<=/</
Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
static struct {
... :114
unsigned short hash[UNW_HASH_SIZE];
... :2152
for (index = 0; index <= UNW_HASH_SIZE; ++index) {
This is a bug, isn't it?
s/<=/</
Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
The previous commit which introduced the DMAR_DEFAULT_ON setting in
drivers/pci/dmar.c neglected to add the ability for ia64 to enable
the IOMMU by default. Rectify that mistake, doh!
Signed-off-by: Kyle McMartin <kyle@redhat.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
During host driver module removal del_gendisk() results in a final
put on drive->gendev and freeing the drive by drive_release_dev().
Convert device drivers from using struct kref to use struct device
so device driver's object holds reference on ->gendev and prevents
drive from prematurely going away.
Also fix ->remove methods to not erroneously drop reference on a
host driver by using only put_device() instead of ide*_put().
Reported-by: Stanislaw Gruszka <stf_xl@wp.pl>
Tested-by: Stanislaw Gruszka <stf_xl@wp.pl>
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
Just copy the comment from drivers/scsi/sr.c::sr_done()
(from which the capacity hack has been originated).
Cc: Borislav Petkov <petkovbb@gmail.com>
Acked-by: Sergei Shtylyov <sshtylyov@ru.mvista.com>
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
Fix missing parentheses so PIO/DMA timings for master device on the
second channel are programmed correctly (IOW "8 0 24 16" offset values
should be used instead of the current "8 0 16 16").
[ The bug went unnoticed because after PIO/DMA timings get programmed
incorrectly for the third device they are overwritten with timings
for the fourth device and since BIOS should also program timings for
the third device everything should work fine until suspend/resume
cycle or user requested transfer mode changes. ]
Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
Cc: Sergei Shtylyov <sshtylyov@ru.mvista.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
[bart: update patch description]
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
Documentation/kernel-parameters.txt
- ide=nodma is no longer valid.
drivers/ide/Kconfig
- The module is ide-core.ko not ide.
drivers/ide/ide.c
- It took me a while to figure out what the arguments %d.%d:%d to nodma
module parameter ment, so I added a comment to each.
- Added a comment to each of the sscanf lines.
- There is a bug, if j is 0 it would previously clear all the other bits
except the current device, changed in three different places.
mask &= (1 << i) should be mask &= ~(1 << i).
Signed-off-by: David Fries <david@fries.net>
[bart: s/disk/device/ in ide.c, beautify patch description]
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
* 'for-linus' of git://neil.brown.name/md:
md: avoid races when stopping resync.
md/raid10: Don't call bitmap_cond_end_sync when we are doing recovery.
md/raid10: Don't skip more than 1 bitmap-chunk at a time during recovery.
When iwlan runs on IOMMU, IOMMU generates a lot of PTE write faults
because PTE write bit is not set on some of PTE's. This is because
iwlan driver calls DMA mapping with PCI_DMA_TODEVICE which is read only
in mapping PTE. But iwlan device actually writes to the mapped page to
update its contents. This issue is not exposed in swiotlb. But VT-d
hardware can capture this fault and stop the fault transaction.
The following patch fixes the issue.
Signed-off-by: Fenghua Yu <fenghua.yu@intel.com>
Reviewed-by: Bhavesh Davda <bhavesh@vmware.com>
Tested-by: Chris Wright <chrisw@sous-sol.org>
Acked-by: Tomas Winkler <tomas.winkler@intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now that several per-cpu files can be read or spliced at the
same, we want the read/splice callbacks for tracing files to be
reentrants.
Until now, a single global mutex (trace_types_lock) serialized
the access to tracing_read_pipe(), tracing_splice_read_pipe(),
and the seq helpers.
Ie: it means that if a user tries to read trace_pipe0 and
trace_pipe1 at the same time, the access to the function
tracing_read_pipe() is contended and one reader must wait for
the other to finish its read call.
The trace_type_lock mutex is mostly here to serialize the access
to the global current tracer (current_trace), which can be
changed concurrently. Although the iter struct keeps a private
pointer to this tracer, its callbacks can be changed by another
function.
The method used here is to not keep anymore private reference to
the tracer inside the iterator but to make a copy of it inside
the iterator. Then it checks on subsequents read calls if the
tracer has changed. This is not costly because the current
tracer is not expected to be changed often, so we use a branch
prediction for that.
Moreover, we add a private mutex to the iterator (there is one
iterator per file descriptor) to serialize the accesses in case
of multiple consumers per file descriptor (which would be a
silly idea from the user). Note that this is not to protect the
ring buffer, since the ring buffer already serializes the
readers accesses. This is to prevent from traces weirdness in
case of concurrent consumers. But these mutexes can be dropped
anyway, that would not result in any crash. Just tell me what
you think about it.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: split up tracing output per cpu
Currently, on the tracing debugfs directory, three files are
available to the user to let him extracting the trace output:
- trace is an iterator through the ring-buffer. It's a reader
but not a consumer It doesn't block when no more traces are
available.
- trace pretty similar to the former, except that it adds more
informations such as prempt count, irq flag, ...
- trace_pipe is a reader and a consumer, it will also block
waiting for traces if necessary (heh, yes it's a pipe).
The traces coming from different cpus are curretly mixed up
inside these files. Sometimes it messes up the informations,
sometimes it's useful, depending on what does the tracer
capture.
The tracing_cpumask file is useful to filter the output and
select only the traces captured a custom defined set of cpus.
But still it is not enough powerful to extract at the same time
one trace buffer per cpu.
So this patch creates a new directory: /debug/tracing/per_cpu/.
Inside this directory, you will now find one trace_pipe file and
one trace file per cpu.
Which means if you have two cpus, you will have:
trace0
trace1
trace_pipe0
trace_pipe1
And of course, reading these files will have the same effect
than with the usual tracing files, except that you will only see
the traces from the given cpu.
The original all-in-one cpu trace file are still available on
their original place.
Until now, only one consumer was allowed on trace_pipe to avoid
racy consuming on the ring-buffer. Now the approach changed a
bit, you can have only one consumer per cpu.
Which means you are allowed to read concurrently trace_pipe0 and
trace_pipe1 But you can't have two readers on trace_pipe0 or
trace_pipe1.
Following the same logic, if there is one reader on the common
trace_pipe, you can not have at the same time another reader on
trace_pipe0 or in trace_pipe1. Because in trace_pipe is already
a consumer in all cpu buffers in essence.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
