v2 pseudo 1.7 (now 1.7.1)
cover v2 pseudo 1.7 (now 1.7.1)
So, this is just like the previous pseudo 1.7.0 patch, except that the very strange and sporadic failures that were showing up in the RPM backend are now fixed.
The actual issue: realpath(foo, NULL) is supposed to allocate storage. When I stopped having path canonicalization allocate new storage, I fixed the points where the returned values were being freed, but I did not spot this particular instance of a place where the value was being returned with the expectation that the caller would free it.
Also changed the lstat() calls in path canonicalization so they will use lstat64 where appropriate (such as all the time on Linux), because otherwise lstat() can fail if a file has an *inode value* out of range, which is distinct from the "size out of range" issues previously corrected. This may have broken path lookups in some cases on XFS.
The xattrdb stuff has not been otherwise changed, so it's still considered experimental.
The following changes since commit bdeb32b4cdbe316f17c2fd854d59e05e8f2e8ffc:
rt-tests: drop unnecessary added-missing-dependencies.patch (2015-09-01 11:43:38 +0100)
are available in the git repository at:
git://git.yoctoproject.org/poky-contrib seebs/pseudo171
http://git.yoctoproject.org/cgit.cgi/poky-contrib/log/?h=seebs/pseudo171
Peter Seebach (1):
pseudo_1.7.1.bb: New version of pseudo
meta/recipes-devtools/pseudo/pseudo_1.7.1.bb | 19 +++++++++++++++++++ meta/recipes-devtools/pseudo/pseudo_git.bb | 4 ++-- 2 files changed, 21 insertions(+), 2 deletions(-) create mode 100644 meta/recipes-devtools/pseudo/pseudo_1.7.1.bb
Signature · 1 line
2.3.1
PATCH 1/1 pseudo_1.7.1.bb: New version of pseudo
Pseudo 1.7 adds an experimental feature (which I think needs more testing before it becomes the default) allowing the pseudo client to store modes and uid/gid values in extended attributes rather than using the sqlite database. On most Linux-like systems, this works only if the underlying file is a plain file or a directory.
Also added is a profiling feature to allow some amount of reporting on the wall-clock time the client spends in wrappers, processing operations, or in IPC. This feature is not intendeded to be precisely accurate, but gives a good overview of where time is going.
Based on the results from the profiling feature, the client now suppresses OP_OPEN and OP_EXEC messages if the server is not logging messages, and no longer uses constant dynamic allocation and free cycles for canonicalized paths.
There's a few other likely-looking optimizations being considered, but this seemed like a good cutoff for now.
(1.7.1 fixes two bugs, one affecting mostly XFS systems with 64-bit inode values, and one affecting code that called realpath(x, NULL), such as the RPM backend.)
--- meta/recipes-devtools/pseudo/pseudo_1.7.1.bb | 19 +++++++++++++++++++ meta/recipes-devtools/pseudo/pseudo_git.bb | 4 ++-- 2 files changed, 21 insertions(+), 2 deletions(-) create mode 100644 meta/recipes-devtools/pseudo/pseudo_1.7.1.bb
meta/recipes-devtools/pseudo/pseudo_1.7.1.bb new file +19 -0
@@ -0,0 +1,19 @@ +require pseudo.inc + +SRC_URI = " \ + http://downloads.yoctoproject.org/releases/pseudo/${BPN}-${PV}.tar.bz2 \ + file://fallback-passwd \ + file://fallback-group \ +" + +SRC_URI[md5sum] = "f7ee16ebf6ab0f2bf80d4ca64862de09" +SRC_URI[sha256sum] = "aac9659cb15b7a40ff27d5de29ef37f033890ec1671d3fa41abbf2d57fdc0358" + +PSEUDO_EXTRA_OPTS ?= "--enable-force-async --without-passwd-fallback" + +do_install_append_class-native () { + install -d ${D}${sysconfdir} + # The fallback files should never be modified + install -m 444 ${WORKDIR}/fallback-passwd ${D}${sysconfdir}/passwd + install -m 444 ${WORKDIR}/fallback-group ${D}${sysconfdir}/group +}
meta/recipes-devtools/pseudo/pseudo_git.bb +2 -2
@@ -1,7 +1,7 @@ require pseudo.inc -SRCREV = "db758fb11167c79d9682a17d359568e2a3c4acd5" -PV = "1.6.5+git${SRCPV}" +SRCREV = "a49c56fbf1fdff5a3050d30b47851c52b4e35fc2" +PV = "1.7.1+git${SRCPV}" DEFAULT_PREFERENCE = "-1"
Signature · 1 line
2.3.1
Re: [PATCH 1/1] pseudo_1.7.1.bb: New version of pseudo
Quoted 16 lines · Peter Seebach
Pseudo 1.7 adds an experimental feature (which I think needs more testing before it becomes the default) allowing the pseudo client to store modes and uid/gid values in extended attributes rather than using the sqlite database. On most Linux-like systems, this works only if the underlying file is a plain file or a directory.
Also added is a profiling feature to allow some amount of reporting on the wall-clock time the client spends in wrappers, processing operations, or in IPC. This feature is not intendeded to be precisely accurate, but gives a good overview of where time is going.
Based on the results from the profiling feature, the client now suppresses
OP_OPENandOP_EXECmessages if the server is not logging messages, and no longer uses constant dynamic allocation and free cycles for canonicalized paths.
Quoted 6 lines
There's a few other likely-looking optimizations being considered, but this seemed like a good cutoff for now.
(1.7.1 fixes two bugs, one affecting mostly XFS systems with 64-bit inode values, and one affecting code that called
realpath(x, NULL), such as the RPM backend.)
This is better however opkg is still unhappy. This is build with this pseudo patch applied:
https://autobuilder.yoctoproject.org/main/builders/nightly-ipk/builds/469
and this is a build with the same patches expect for the pseudo one:
https://autobuilder.yoctoproject.org/main/builders/nightly-ipk/builds/470
so something is still unhappy :(. The rpm builds were a lot better though.
Cheers,
Richard
Re: [PATCH 1/1] pseudo_1.7.1.bb: New version of pseudo
Quoted 4 lines · On Thu, 3 Sep 2015 09:43:50 +0100 Richard Purdie…
This is better however opkg is still unhappy. This is build with this pseudo patch applied:
https://autobuilder.yoctoproject.org/main/builders/nightly-ipk/builds/469
I've found the cause of that particular bit. During my attempts to reproduce this, I ran into segfaults in cross-localedef that I cannot now reproduce.
The cause of this was fixing up filesystem modes for the case where umask contains 0200, which wasn't originally expected to work, but which was breaking the xattrdb code. (There are a few things in the system which create files with umask 0777, it turns out; this is arguably bad, but now there's a workaround.)
But when I put in the workaround, I used 0600 as the "definitely at least these bits" mode, instead of the requested mode, and that breaks scripts which were expecting, say, 0700. So the things which have a specified mode now use that | 0600 instead of just 0600.
I want to run a couple more things to see whether I can get any of the other failures to come back again.
I also partially know why I wasn't seeing this on one of my systems, but only partially. When I do my builds on CentOS 6.5, the .postinst files in question either don't come into existence, or get run successfully anyway. So far as I can tell, opkg is doing "sh -c /path/to/script" on those, which requires 0700 to run. I'm not sure what's causing the difference between the hosts.
-s
Signature · 1 line
Listen, get this. Nobody with a good compiler needs to be justified.