gmane.linux.rt.user

RFC: Latency reducing TCP modifications for thin-stream interactive applications

Andreas Petlund — 2008-11-27T13:39:41

A wide range of Internet-based services that use reliable transport protocols display what we call thin-stream properties. This means that the application sends data with such a low rate that the retransmission mechanisms of the transport protocol are not fully effective. In time-dependent scenarios (like online games, control systems or some sensor networks) where the user experience depends on the data delivery latency, packet loss can be devastating for the service quality. Extreme latencies are caused by TCP's dependency on the arrival of new data from the application to trigger retransmissions effectively through fast retransmit instead of waiting for long timeouts. After analyzing a large number of time-dependent interactive applications, we have seen that they often produce thin streams (as described above) and also stay with this traffic pattern throughout its entire lifespan. The combination of time-dependency and the fact that the streams provoke high latencies when using TCP is unfo

strange getrusage timer problem with rt-kernel

2008-11-26T03:52:28

I'm testing 2.6.26.6-rt11 (is that the latest?) kernel, using tiotest from tiobench (http://sourceforge.net/projects/tiobench/), a classic multi-thread IO stress tool, it reported a strange problem, with: root:~/tiobench# time ./tiotest -R -d /dev/mapper/dg5-lv2 -t 1 -f 1000 -r 102400 tiotest: tiotest.c:609: add_timer: Assertion end_time->tv_sec >= start_time->tv_sec' failed.? this problem never happened in mainstream kernels. Regards, -- Cheng Renquan, Shenzhen, China -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo< at >vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

Strange device behavior on SMP system with CONFIG_PREEMPT_RT patch

Denis Borisevich — 2008-11-13T10:20:17

Hi everybody, I noticed a strange behavior of my PCI->RS422 adapter (MOXA CP-132 using mxer kernel module) on SMP x86 system with kernel 2.6.26.6-rt11. This PCI board has 2 RS422/RS485 ports. When I try to send cyclicaly some ammount of data more than 16 bytes (it's internal buffer is 16 bytes) from the second port to the first the data is lost at random iteration. If I try to send from first port to the second everything works fine no matter which ammount of data I write to the port. It seems that interrupts get lost. Can somebody help me with this problem? What can cause such a behavior? Thanks in advance, Denis -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo< at >vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

config_preempt_RT patch platform support question.

Amir Aharon — 2008-11-12T14:30:45

Hi, I work on a project which delivers low latency audio streaming via a Linux machine. I receive/send audio frames of 4 samples each per network frame (raw Ethernet) so I expect to have a network interrupt each 90 microseconds. I use Linux 2.6.26.3-rt7 on different platforms; unfortunately we encountered problems such as kernel panics, depend on the platform running... My questions are: How can I have more details on the tests which where used to evaluate the platform running the config_preempt_RT (specifically on networking issues)? Does Intel Q6600 2.4GHz (quad core) has been tested? Are there known issues? Does Xeon x5482 3.2GHz (x2 meaning octa core) has been tested? Are there known issues? Thanks, Amir. -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo< at >vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

kernel modules and ftrace

Juergen Beisert — 2008-11-11T17:56:30

Hi list, when I'm compiling in FTRACE support I get the following output, when I load the first of a few modules: ----------------------------------------------------------------------------- [...] ***************************************************************************** * * * REMINDER, the following debugging options are turned on in your .config: * * * * CONFIG_IRQSOFF_TRACER * * CONFIG_FTRACE * * * * they may increase runtime overhead and latencies. * * * ***************************************************************************** [...] Freeing unused kernel memory: 132k in

2.6.27.5 RT patch status (26rt11)

Paul Gortmaker — 2008-11-10T23:47:40

I've updated from the 2.6.27-rc8 patch series based on 26rt9 to be based on 2.6.27.5 kernel.org and 26rt11 preempt_rt. I've done a successful boot to X11 on basic old P4 and it survived a make -j20 of kernel.org git. There is still work to be done in order to stabilize more complex hardware (e.g. SMP etc) -- but I see this still being the baseline for that 2.6.27.x based work. The trees are where they were before, with the individual patches at: git://opensource1.windriver.com/people/paulg/rt/patches and the fully patched 2.6.27.5 git tree can be had from here: git://opensource1.windriver.com/people/paulg/rt/linux-2.6 Note the patched tree is on the branch v2.6.27.5-26rt11 and that master is just kernel.org master. The patched trees are just the result of applying all the patches with guilt, and so for those interested in the details of what changed in the patches, looking at the patch repo is probably the most educational, i.e: git whatchanged v2.6.27-rc8-26rt9..v2.6.27.5-26rt11 There are gitweb

[PATCH][RT] Trivial: Correctly dereference when clearing unusedvariable

Sven-Thorsten Dietrich — 2008-11-07T21:19:42

From: Sven-Thorsten Dietrich suse.de> Subject: Correctly dereference flags when clearing unused variable. Its probably unsafe to set the flags pointer to 0, since this will oops, if it is dereferenced elsewhere for some odd reason. Signed-off-by: Sven-Thorsten Dietrich suse.de> --- mm/page_alloc.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) --- a/mm/page_alloc.c +++ b/mm/page_alloc.c < at >< at > -170,7 +170,7 < at >< at > static inline void __lock_cpu_pcp(unsign { #ifdef CONFIG_PREEMPT_RT spin_lock(&__get_cpu_lock(pcp_locks, cpu)); -flags = 0; +*flags = 0; #else local_irq_save(*flags); #endif < at >< at > -180,7 +180,7 < at >< at > static inline void lock_cpu_pcp(unsigned { #ifdef CONFIG_PREEMPT_RT (void)get_cpu_var_locked(pcp_locks, this_cpu); -flags = 0; +*flags = 0; #else local_irq_save(*flags); *this_cpu = smp_processor_id(); -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo< at >vger.kernel.org More majordomo inf

[PATCH][RT] Dereference pointer to cpu id, not to address of CPUID

Sven-Thorsten Dietrich — 2008-11-07T21:13:27

This patch applies to 2.6.25-rt, 2.6.26-rt and 2.6.27-rt From: Sven-Thorsten Dietrich suse.de> Subject: Dereference pointer to cpu id, when evaluating condition. Without dereferencing, the condition always evaluates to true. Signed-off-by: Sven-Thorsten Dietrich suse.de> --- mm/slab.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) --- a/mm/slab.c +++ b/mm/slab.c < at >< at > -2033,7 +2033,7 < at >< at > slab_destroy(struct kmem_cache *cachep, } else { kmem_freepages(cachep, addr); if (OFF_SLAB(cachep)) { -if (this_cpu) +if (*this_cpu) __cache_free(cachep->slabp_cache, slabp, this_cpu); else kmem_cache_free(cachep->slabp_cache, slabp); -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo< at >vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

2.6.26.6-rt11 nanosleep() does have accurate behavior

Gautam Thaker — 2008-11-07T15:56:02

I recently started to study how nanosleep() behaves under 2.6.26.6-rt11 compiled in in FULL PREEMPT, uniprocessor mode. I find that between 2.6.20-rt8 and this new kernel something has changed and nanosleep() is back to its old behavior or sleeping for next higher number of msec rather than carefully following the "y=x" line. I don't follow the RT world closely enough, but is this a regression or an expected change in behavior? The link below show the results very clearly. http://www.atl.external.lmco.com/projects/QoS/documents/nanosleep_1.png Gautam -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo< at >vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html

nohz=off

Sergio Monteiro Basto — 2008-11-03T16:27:23

Hi, When I have to boot with nohz=off , this could be a problem with rt isn't it ? Who I could debug this problem ? This is a laptop with linux x86_64 ( I had try fedora 10 beta ) and others kernel like 2.6.22 and 2.6.24 . Can you point me some documentation about nohz=off , I try search a bit but don't found anything interesting. Thanks,

Trouble compiling 2.6.26.6-rt11 for ARM

Bo Hansen — 2008-11-03T15:02:56

Hello, I am trying to compile 2.6.26.6-rt11 for an ARM-based board using AT91RM9200, but get the following error message when I compile: LD .tmp_vmlinux1 kernel/built-in.o(.text+0xdf08): In function `$a': : undefined reference to `wrong_size_cmpxchg' kernel/built-in.o(.text+0xe05c): In function `$a': : undefined reference to `wrong_size_cmpxchg' kernel/built-in.o(.text+0xe32c): In function `$a': : undefined reference to `wrong_size_cmpxchg' make: *** [.tmp_vmlinux1] Error 1 I get the message both using my custom board and the EK-board from Atmel. The cross-compiler is version 3.4.1. Any ideas of what has gone wrong? Best ragards, Bo Hansen

rt-tests version 0.28 available

Clark Williams — 2008-10-22T20:22:49

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 All, The latest version of rt-tests (cyclictest, signaltest, pi_stress) is available at: http://www.kernel.org/pub/linux//kernel/people/clrkwllms/rt-tests/rt-tests-0.28.tar.gz as either a gziped or bziped tarball. This version has two new options added: --histogramwhich dumps a histogram of latencies after the run --duration which allows the run to be of a specific duration Thanks to Sven for the very useful histogram code! If you prefer, you can access the git tree at: git://git.kernel.org/pub/scm/linux/kernel/git/clrkwllms/rt-tests.git Summary of git commits: Clark Williams (3): add option to run for specified duration changed release target in Makefile told git to ignore tarballs and tmpdir Sven-Thorsten Dietrich (1): Subject: Add histogram support to cyclictest .gitignore | 4 +- Makefile | 6 +- src/cyclictest/cyclictest.8 | 11 +++- src/cyclictest/cyclict

[PATCH 2/2] netrate - network rate/rtt measurement utility

Clark Williams — 2008-10-22T03:40:29

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 - From 1ecdeb7f48302a2ea8a37fd8aecd90081637e0dc Mon Sep 17 00:00:00 2001 From: Clark Williams redhat.com> Date: Thu, 2 Oct 2008 17:26:05 -0500 Subject: [PATCH] consolidate into common consolidated both rtt and transmit to use same socket setup function; added exchange code and daemon mode; removed --server and --receive; added versioning Signed-off-by: Clark Williams redhat.com> - --- common.c | 82 +++++++++++++++++++++++++++++++++ main.c | 146 +++++++++++++++++++++++++++++++++++++++++++++++++++++------- netrate.h | 28 +++++++++--- receive.c | 71 +----------------------------- rtt.c | 51 +-------------------- server.c | 74 +------------------------------ transmit.c | 50 +-------------------- 7 files changed, 237 insertions(+), 265 deletions(-) diff --git a/common.c b/common.c index 48cd4e9..8ed0590 100644 - --- a/common.c +++ b/common.c < at >< at > -433,3 +433,85 < at >< at > free_histogram(struc

[PATCH 0/2] netrate - network rate/rtt measurement utility

Clark Williams — 2008-10-22T03:39:50

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Thomas, The patch series that follows is for a program called 'netrate'. It's a utility I wrote recently to prove to a customer (and myself) that we were receiving data on a TCP socket at the same rate as it was being sent from another system. Since I did that, I've added a round-trip-time measurement as well. Nothing in netrate is specific to the -rt kernel, since I tried to ensure I only used posix interfaces. So if you spot something which breaks that, please let me know. The default host address is localhost, so you can test this quite easily on the same system. $ netrate --help usage: netrate {--transmit|--rtt|--listen} [options] where options are: --rate= --size= --tcp --udp --sctp --nodelay --multicast --port= --host= --debug --histogram --help Curre

[PATCH] Cyclictest Hiistogram Support

Sven-Thorsten Dietrich — 2008-10-21T20:19:56

Subject: Add histogram support to cyclictest From: Sven-Thorsten Dietrich suse.de> Add -h parameter and functionality to log histograms of latencies in cyclictest. Signed-off-by: Sven-Thorsten Dietrich suse.de> diff --git a/src/cyclictest/cyclictest.8 b/src/cyclictest/cyclictest.8 index 9827458..82c2c43 100644 --- a/src/cyclictest/cyclictest.8 +++ b/src/cyclictest/cyclictest.8 < at >< at > -16,7 +16,7 < at >< at > cyclictest \- High resolution test program .SH SYNOPSIS .B cyclictest -.RI "[ \-hfmnqrsv ] [\-a " proc " ] [\-b " usec " ] [\-c " clock " ] [\-d " dist " ] [\-i " intv " ] [\-l " loop " ] [\-o " red " ] [\-p " prio " ] [\-t " num " ]" +.RI "[ \-hfmnqrsv ] [\-a " proc " ] [\-b " usec " ] [\-c " clock " ] [\-d " dist " ] [\-h " histogram " ] [\-i " intv " ] [\-l " loop " ] [\-o " red " ] [\-p " prio " ] [\-t " num " ]" .\" .SH DESCRIPTION .\" This manual page documents briefly the .\" .B cyclictest commands. < at >< at > -81,6 +81,9 < at >< at > Set the distance of thread intervals in microseconds (d

[-rt] 2.6.26.6-rt11: Another sleeping function called from invalid context

Jürgen Mell — 2008-10-18T20:17:40

Hi, when I boot the 2.6.26.6-rt11 kernel, I get another of these 'sleeping function called from invalid context' error messages: BUG: sleeping function called from invalid context IRQ-15(75) at arch/x86/mm/highmem_32.c:8 in_atomic():0 [00000000], irqs_disabled():1 Pid: 75, comm: IRQ-15 Not tainted 2.6.26.6-rt11-HWS #2 [] __might_sleep+0xd4/0xdb [] kmap+0x42/0x55 [] ata_sff_hsm_move+0x3ca/0x671 [libata] [] ata_sff_interrupt+0x141/0x1c0 [libata] [] handle_IRQ_event+0x45/0xb6 [] do_irqd+0x126/0x223 [] ? do_irqd+0x0/0x223 [] kthread+0x39/0x60 [] ? kthread+0x0/0x60 [] kernel_thread_helper+0x7/0x10 Hardware is a Pentium-M Processor on an industrial PC mainboard with Intel chip set, 1GB RAM, HDD + DVD ROM drives. Software is OpenSUSE 10.3 with KDE 3.5. The message comes up twice during start of the system and will be repeated indefintely if there is a CD or DVD in the drive. Bye, Jürgen -

[PATCH -rt] sched: remove unneeded #if directive

Hiroshi Shimamoto — 2008-10-18T01:21:29

From: Hiroshi Shimamoto ct.jp.nec.com> __cond_resched_spinlock() is in #ifdef CONFIG_PREEMPT_RT. So the following line always true. #if (defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)) || defined(CONFIG_PREEMPT_RT) Signed-off-by: Hiroshi Shimamoto ct.jp.nec.com> --- kernel/sched.c | 3 +-- 1 files changed, 1 insertions(+), 2 deletions(-) diff --git a/kernel/sched.c b/kernel/sched.c index 0d68e79..2b34f00 100644 --- a/kernel/sched.c +++ b/kernel/sched.c < at >< at > -5841,7 +5841,6 < at >< at > EXPORT_SYMBOL(__cond_resched_raw_spinlock); int __cond_resched_spinlock(spinlock_t *lock) { -#if (defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)) || defined(CONFIG_PREEMPT_RT) if (lock->break_lock) { lock->break_lock = 0; spin_unlock_no_resched(lock); < at >< at > -5849,7 +5848,7 < at >< at > int __cond_resched_spinlock(spinlock_t *lock) spin_lock(lock); return 1; } -#endif + return 0; } EXPORT_SYMBOL(__cond_resched_spinlock);

New rt-tests maintainer

Thomas Gleixner — 2008-10-13T20:23:18

Hi everyone, I'm handing over the maintainership of rt-tests (cyclictest, signaltest, pi_stress) to Clark Williams. rt-tests started with cyclictest, a test program I wrote for validating the high resolution timers implementation. In hindsight I'm still surprised that cyclictest became a defacto standard for benchmarking preempt-rt, but on the technical side it's quite clear as it tests the whole chain: interrupt -> wakeup -> scheduler -> userspace. I never expected that cyclictest might ship with distros and I'm very grateful to all the people who contributed and helped to bring it and the whole rt-tests package into shape. I know that I'm a lousy user space programmer :) < at >Clark, thanks for taking over. It takes away my permanent guilty conscience of not caring about rt-tests enough. < at >all, please support Clark with his new project as you supported me. Thanks for all the support and help again ! Thanks, tglx P.S.: If you really need to add autocrap^H^H^H^Htools please do it in a way so I still

[PATCH 2/2] add function_trace_stop

Steven Rostedt — 2008-10-13T19:16:34

2.6.26.6-rt10 added /debug/tracing/function_trace_stop which is a very quick way of starting and stopping the tracer. It does not disable the mcount call, but just sets a variable in the mcount call to return. This patch tries to open the function_trace_stop file and if it succeeds it will use the trace stop to start and stop the tracer an keep the ftrace enabled the entire time. Signed-off-by: Steven Rostedt redhat.com> --- src/cyclictest/cyclictest.c | 37 +++++++++++++++++++++++++++++++++---- 1 file changed, 33 insertions(+), 4 deletions(-) Index: rt-tests.git/src/cyclictest/cyclictest.c =================================================================== --- rt-tests.git.orig/src/cyclictest/cyclictest.c2008-10-13 15:04:11.000000000 -0400 +++ rt-tests.git/src/cyclictest/cyclictest.c2008-10-13 15:13:36.000000000 -0400 < at >< at > -134,6 +134,7 < at >< at > static int tracetype; static int lockall = 0; static int traceenable_fd = -1; +static int funcstop_fd = -1; /* Backup of kernel variables that

[PATCH 1/2] cyclictest: preopen trace_enabled

Steven Rostedt — 2008-10-13T19:09:14

The conversion to use kernvar caused an open and close at every iteration, which is very expensive. This patch opens the file at start up and closes it at exit. Signed-off-by: Steven Rostedt redhat.com> --- src/cyclictest/cyclictest.c | 17 ++++++++++++++--- 1 file changed, 14 insertions(+), 3 deletions(-) Index: rt-tests.git/src/cyclictest/cyclictest.c =================================================================== --- rt-tests.git.orig/src/cyclictest/cyclictest.c2008-10-13 13:01:59.000000000 -0400 +++ rt-tests.git/src/cyclictest/cyclictest.c2008-10-13 15:01:00.000000000 -0400 < at >< at > -133,6 +133,8 < at >< at > static int oscope_reduction = 1; static int tracetype; static int lockall = 0; +static int traceenable_fd = -1; + /* Backup of kernel variables that we modify */ static struct kvars { char name[KVARNAMELEN]; < at >< at > -276,20 +278,25 < at >< at > static inline long calcdiff(struct times return diff; } +void set_trace_enable(char val) +{ +write(traceenable_fd, &val, 1); +} + void tracing(int on)

2.6.26.6-rt10

Steven Rostedt — 2008-10-12T19:30:14

We are pleased to announce the 2.6.26.6-rt10 tree, which can be downloaded from the location: http://rt.et.redhat.com/download/ Information on the RT patch can be found at: http://rt.wiki.kernel.org/index.php/Main_Page Changes since 2.6.26.5-rt9 - ported to 2.6.26.6 - rwlock to handle unnested order unlocks (Steven Rostedt) - rwlock update torture test to test unnested unlocks (Steven Rostedt) - remove rt push comment (Gregory Haskins) - rt sched fix to handle dequeue (Gregory Haskins) - PPC: remove duplicate stack trace code (Robert Schwebel) - ARM: ftrace event IRQ call fix (Alexander Pazdnikov) - fair sched not to resched RT tasks (Peter Zijlstra) - ftrace quick stop (Steven Rostedt) - event trace of resched (Steven Rostedt) to build a 2.6.26.6-rt10 tree, the following patches should be applied: http://www.kernel.org/pub/linux/kernel/v2.6/linux-2.6.26.tar.bz2 http://kernel.org/pub/linux/kernel/v2.6/patch-2.6.26.6.bz2 http://rt.et.redhat.com/download/patch-2.6