Hi all,
There are various PPAs relevant to linaro, but they can be hard to
find when you don't work with them actively. Do we have a list
somewhere? If not, it could be worth listing them on the wiki
somewhere, or documenting how to find them.
Cheers
---Dave
Hi all,
I've hit some problems in trying to get perf record/report to
work correctly on systems with separated debug images (Ubuntu,
in particular). I worked on some patches to ensure that
separate debug images do actually get loaded, when present --
these commits have now been merged in linux-2.6-tip/master.
(See below for a list of the commits.)
Now that these are in place, I'm observing a new problem which
can mess up symbol locations -- though I think there might be a
practical workaround, I can't see a trivial solution, so I'm
keen to hear if anyone has any ideas.
The problem:
perf makes some incorrect assumptions, which mean that the
symbol locations seen by perf report and friends can be
incorrect.
My analysis:
a) perf represents symbols as offsets from the start of
the mmap'd code region which contains each
symbol.
b) For shared libs (ET_DYN objects), the desired offset
is (usually) equal to the VMA of each given
symbol in the image. Perf assumes that these
offsets are always equal, and this normally
works.
Atypical link configurations could cause this
to fail, but this would probably be unusual in
the real world.
c) For executables (ET_EXEC objects), the desired
offset is:
st_addr - (mmap_start + sh_offset - mmap_offset)
Where st_addr is the symbol VMA, sh_offset is
the offset of the parent *section* (not
segment) from the start of the image,
mmap_start is the base VMA of the corresponding
mmap'd *segment* (or part of a segment) in
memory, and mmap_offset is the file offset at
which the mapping starts (as in mmap(2))
(As a minor note, the pgoff field in the perf
mmap event data (from which mmap_offset should
be deducible) is probably also incorrect at
present -- see snippet from util/event.c
below.)
d) perf assumes that the above is equivalent to:
st_addr - (sh_addr - sh_offset)
Where sh_addr is the VMA of the start of the
symbol's parent section.
This is true only given certain assumptions,
namely that each symbol's parent section is
mapped in such a way that mmap_offset is zero.
In practice, this is true for the main
executable segment of ET_EXEC images in typical
link configurations (i.e., using the default
GNU ld scripts).
e) The assumptions in (d) break down when using
separated debug images, because although
separated debug images contain faithful VMA
information, the image is stripped of loadable
sections' contents, resulting in radically
different sh_offset values from the "real"
image.
Because sh_offset bears no relationship to the
layout of the data mmap'd at runtime, this
produces bogus symbol adjustments which results
in the wrong symbol offsets (a) being
calculated.
Solutions:
(e) needs to be solved in order for perf report etc. to produce
sensible output when using separated debug images.
I think that a comprehensive solution should also fix the
assumptions in (b) and (d) so that any valid ELF images will be
processed correctly.
Solving (e) doesn't appear trivial, since it requires section
layout information from the image that was mmap'd at _runtime_
to be correlated with the symbol information collected maybe
from other files at _report_ time. Currently, runtime and
report time are treated rather separately, so this isn't
straightforward to achieve.
A simple workaround in the meantime would be to assume that
.text is mmap'd with mmap_offset=0 for ET_EXEC images, and fix
the adjustment calculation appropriately. This is not a full
fix but is probably worth doing if (as I guess) it gets the
tools working correctly in the common case: see
[PATCH 2/2] perf symbols: work around incorrect ET_EXEC symbol adjustment
-------------------------------------------------------------------------------
My current list of patches in linux-2.6-tip/master:
commit 6da80ce8c43ddda153208cbb46b75290cf566fac
perf symbols: Improve debug image search when loading symbols
commit 8b1389ef93b36621c6acdeb623bd85aee3c405c9
perf tools: remove extra build-id check factored into dso__load
commit 21916c380d93ab59d6d07ee198fb31c8f1338e26
perf tools: Factor out buildid reading and make it implicit in dso__load
commit 88ca895dd4e0e64ebd942adb7925fa60ca5b2a98
perf tools: Remove unneeded code for tracking the cwd in perf sessions
commit 361d13462585474267a0c41e956f1a1c19a93f17
perf report: Don't abbreviate file paths relative to the cwd
Illustrative code snippets (from v2.6.35):
util/symbol.c:
968:static int dso__load_sym(struct dso *self, struct map *map, const char *name,
969: int fd, symbol_filter_t filter, int kmodule,
970: int want_symtab)
971:{
[...]
1055: elf_symtab__for_each_symbol(syms, nr_syms, idx, sym) {
1133: if (curr_dso->adjust_symbols) {
[...]
1134: pr_debug4("%s: adjusting symbol: st_value: %#Lx "
1135: "sh_addr: %#Lx sh_offset: %#Lx\n", __func__,
1136: (u64)sym.st_value, (u64)shdr.sh_addr,
1137: (u64)shdr.sh_offset);
1138: sym.st_value -= shdr.sh_addr - shdr.sh_offset;
1139: }
Probable incorrect symbol adjustment which currently gets used
for ET_EXEC images. Explanation above.
util/event.c:
109:static int event__synthesize_mmap_events(pid_t pid, pid_t tgid,
110: event__handler_t process,
111: struct perf_session *session)
112:{
[...]
153: if (*pbf == 'x') { /* vm_exec */
[...]
166: /* pgoff is in bytes, not pages */
167: if (n >= 0)
168: ev.mmap.pgoff = vm_pgoff << getpagesize();
169: else
170: ev.mmap.pgoff = 0;
Unless I've misunderstood, getpagesize(2) returns the size in
bytes of a page; in this case using it as a shift count is
wrong.
Also, the code appears to be trying to convert a page count
into an address for storage in the profile data. This would
mean that when interpreting the data, knowledge of the page
size is not needed. However, since I can't see any code which
actually uses the pgoff information from the profile data, I'm
not sure whether this was the intention.
In any case, the offset field from /proc/*/maps is already a
byte count, not a page offset, so it does not look necessary to
do convert it at all.
See [PATCH 1/2] perf events: Fix mmap offset determination.
Dave Martin (2):
perf events: Fix mmap offset determination
perf symbols: work around incorrect ET_EXEC symbol adjustment
tools/perf/util/event.c | 8 +-------
tools/perf/util/symbol.c | 10 +++++++++-
2 files changed, 10 insertions(+), 8 deletions(-)
The weekly report for the Linaro Infrastructure team may be found at:
Status report: https://wiki.linaro.org/Platform/Infrastructure/Status/2010-08-12
Burndown chart: http://people.canonical.com/~pitti/workitems/maverick/arm-infrastructure.ht…
Progress:
* LEP/DerivativeDistributions: Backend work under way. User
testing of UI mockups complete. Feedback being incorporated into
mockups for a second round of user testing.
* arm-m-private-archive-hosting-infrastructure: Changes to allow LP
archive UI to be reskinned should land soon. Some existing bugs fixed
to allow easier testing of vostock. Several blueprint items in review
and waiting to land. Launchpad monthly release freeze delaying reviews
and landing slightly.
* arm-m-derived-archive-rebuild: Working together with LP team to
consider how to consolidate Linaro/LP changes. Working on archive
deletion. Analysis and planning complete, work about to begin on
actual coding. Groundwork necessary is not as complete as first
thought. Archive deletion is estimated to take longer than originally
thought
* arm-m-image-building-tool: Blueprint has one remaining work item, met
with Cody, Ian about it. Will look into documentation changes Cody
proposed.
* arm-m-image-building-console: Progress made on getting several bugs
fixes landed. Two of the three branches should make it into next
LexBuilder release for dogfooding by OEM. Continuing work on other
bugs. First security review complete, no surprises. Started process
of choosing a new name for release.
* arm-m-automated-testing-framework: Reworking authentication code.
Pushed a branch for doing browser startup time benchmarking. Added
software and hardware profile gathering code. Evaluating OpenPosix and
Phoronix test suites on ARM for correctness. Awaiting Linaro board
approval for license.
* arm-m-validation-dashboard: JSON schema for dashboard data bundle
published. Changes from sprint still being reviewed. Should be landed
next week. Initial UI work started. Work on test tool for interacting
with test results repository coming along. Awaiting Linaro board
approval for license. Please let me know if you have any questions.
Cheers,
Scott
--
Scott Bambrough <scott.bambrough(a)linaro.org>
Linaro Infrastructure Team Lead
A u-boot-linaro package is available in the linaro maintainers kernel ppa:
ppa:linaro-maintainers/kernel
The package is u-boot-linaro and the version is v2010.06-265-gcdb7497-linaro-1.
If you just want to download the .deb and extract it by hand:
https://launchpad.net/~linaro-maintainers/+archive/kernel/+files/u-boot-lin…
>From the changelog:
u-boot-linaro (v2010.06-265-gcdb7497-linaro-1) maverick; urgency=low
* Initial Linaro u-boot tree based on
* Upstream v2010.06-265-gcdb7497
* A series of patches from ssakomans exp branch that add
support for beagle C4 and xM. These have been rebased
onto current upstream. The gitorious beagle trees are
still based on v2010.03. These patches will be replaced
with patches from ssakoman.
* Versatile express patches from Matt Waddel.
* Debian packaging by John Rigby based on the lucid
u-boot-omap3 package.
* Currently builds u-boot binaries for these platforms
ca9x4_ct_vxp aka Versatile Express
mx51evk aka Babbage
omap4_panda
omap3_beagle
If you have access to any of these platforms your help with testing
would be appreciated. If you would like other platforms added please
let me know.
Thanks
John
Hi there!
I unpacked our minimal release image and ran an xdiskusage on it,
mostly to see what we're shipping -- and I was surprised to see that a
fourth of the image is actually apt package caches and lists. Can we
put into the image generation script something to strip them out before
generating the image?
The untarring also suggests a number of places where we could further
trim the image, some of which are probably pretty hard to do:
* stripping /usr/share/doc out (but everybody knew that)
* dropping charmaps, zones and locale info that will never really be
used
* stripping out modules for devices that won't ever be on
this ARM device
* stripping out firmware for peripherals that won't ever be on this
ARM device
I've included the xdiskusage output so you can see what I mean. And good
job on the release!
--
Christian Robottom Reis | [+55 16] 3376 0125 | http://launchpad.net/~kikoLinaro.org | [+55 16] 9112 6430 | http://async.com.br/~kiko
On Wed, 11 Aug 2010, Will Deacon wrote:
> Hi Nicolas,
>
> > The Linaro freeze is imminent, and at that point a stable branch that
> > won't be rebased anymore will be forked. The stable branch should only
> > see bug fixes and no new features. Therefore I'd like to know if there
> > is anything that should be added or removed from this tree before the
> > freeze is in place. Please tell me ASAP.
>
> The only thing I can think of from my end is the hardware breakpoint patches
> I have:
>
> http://www.linux-arm.org/git?p=linux-2.6-wd.git;a=shortlog;h=refs/heads/hw-…
>
> I've not merged these into my Linaro branch because I'm not happy with the
> amount of testing they've received [it's a bit of a catch 22 because one
> good way to get them tested is to merge them!].
My criteria would be:
1) Could their addition break something else even when not used?
If there is no risk for regression then we could as well just include
them and add fixes later if need be.
2) Is there some ABI implied with gdb or the like?
If there are ABI issues then I'd prefer having the assurance that it
won't change by the time this gets merged into the mainline kernel.
Nicolas
I had a quick go at benchmarking the Linaro GCC compiler and was
presently surprised. The workload was decoding 10 s of h.264 video
using ffmpeg with assembly optimisations disabled. This gives the
compiler a good workout.
Compared to FSF GCC 4.4.4, Linaro GCC 4.4 is 6 % faster, FSF GCC 4.5.0
is 12 % faster, and Linaro GCC 4.5 is 24 % faster.
Now this is with a sample size of one so the results are not
statistically valid, and I haven't told you how to reproduce it so
they're not scientifically valid, but the numbers are pleasing none
the less.
-- Michael
