Age | Commit message (Collapse) | Author |
|
|
|
|
|
|
|
|
|
This reverts commit 49d4d35918d457b0e9206679ecad3b9c84f11e66.
|
|
The cached data values were being collected when an OP_EXEC call was made.
This is incorrect as the values are only for logging purposes. It's believed
this caused an occasional crash in certain instances.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
It'd be handy for the WR build system if new state directories could
be created as needed. It is made so. And to answer the first
question everyone, including me, has on reading this: You can't
do system("mkdir -p ...") because the invoked shell would need to
run under pseudo, so it'd have to check for a server, and...
|
|
|
|
The problem is that path_by_ino could end up being the same pointer
as cache_path, after which, if cache_path were freed (or kept around
for later), there would be malloc arena problems.
Also, fix the calculation for pathlen to increase cache hits. The
IPC messages use length of path *plus one* as the length, because
the buffer is defined to include its terminating null byte.
|
|
The pathlen that is cached could be wrong in certain operations (RENAME).
Fix this by resetting it to the proper path length.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
When pseudo is disabled, we skip a bunch of the prefix, localstate, etc
processing. This allows pseudo to run with a directory that does not yet
exist.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
before pseudo_tables.h existed. Fixed.
|
|
|
|
Changing the file journal to OFF, this makes things a bit more susceptible
to failure in cases where pseudo crashes or the system crashes.. however, this
is believed to be unlikely.
Timing:
Before:
real 6m42.093s
user 0m34.321s
sys 2m46.086s
(with journal set to MEMORY)
real 6m33.037s
user 0m33.133s
sys 2m48.668s
After (journal set to OFF)
real 6m17.313s
user 0m32.757s
sys 2m47.654s
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
We change the index from simply "path" to "path, dev, ino". This allows for
slightly faster searches for the exact file information, and does not cause
any penalty for a simple "path" based search.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
Add a cache of the last object found in pseudo_op. Profiling has indiciated
that many operations come in clusters. So instead of doing select, operation
for each item in the cluster, we check to see if we already know the item and
perform the op..
Performance improvement when processing 500k or so files:
Previous:
real 7m11.778s
user 0m35.929s
sys 2m46.723s
This commit:
real 6m42.093s
user 0m34.321s
sys 2m46.086s
Also validation of the component can be added by compiling with NVALIDATE.
This verifies the result of the cache is the same as what would have come
from the database. Differences are logged to the standard pseudo.log.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
We restructure the pseudo_op file identification, which involves a fairly
significant performance increase.
The old method would do:
if pdb_find_file_path:
found_path = 1;
if found_path && pdb_find_file_exact:
found_ino = 1;
else if pdb_find_file_dev:
found_ino = 1;
This resulted in at least two select calls for each file. One for "path"
and one for exact or dev.
The new method instead does:
if pdb_find_file_exact:
found_path = 1;
found_ino = 1;
else
if pdb_find_file_path:
found_path = 1;
if pdb_find_file_dev:
found_ino = 1;
This shrinks the number of selects to either one or three. Potentially cutting
the number of selects in half -- or increasing the number to three on an empty
set... (Profiling has shown this is a net win)
Timing numbers when manipulating a large number (500k) of files in a ramdisk:
real 7m48.354s
user 0m32.895s
sys 2m50.274s
After this change:
real 7m11.778s
user 0m35.929s
sys 2m46.723s
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
Add sqlite call profiling, this allows us to see the sqlite calls
that are being made as the system runs, via the pseudo log.
It was noted that by this profiling that a small change to pseudo.c,
when a file was found, reduced the sqlite SELECT calls by about 1/3.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
Restructure wrapfuncs.c in an attempt to improve PSEUDO_DISABLED=1
performance.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
2010-12-09:
* (mhatle) Add doc/program_flow to attempt to explain startup/running
* (mhatle) guts/* minor cleanup
* (mhatle) Reorganize into a new constructor for libpseudo ONLY
pseudo main() now manually calls the util init
new / revised init for client, wrappers and utils
* (mhatle) Add central "reinit" function
* (mhatle) Add manul execv* functions
* (mhatle) rename pseudo_populate_wrappers to pseudo_check_wrappers
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
* (mhatle) Add guts/clone.c to cleanup the clone support
* (mhatle) guts/clone.c only run setupenv and reinit when NOT PSEUDO_RELOADED
* (mhatle) guts/execve.c whitespace fixe
* (mhatle) guts/fork.c similar to guts/clone.c change
* (mhatle) pseudo_client.c add reinit function
* (mhatle) pseudo_client.c revise client reset, include code from pseudo_wrappers.c
* (mhatle) pseudo_server.c move the pid writing to the parent
* (mhatle) pseudo_wrappers.c clone cleanup and populate cleanup
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
be out of sync in a very inconvenient way.
Changes include:
* Some whitespace fixes, also move the pseudo_variables definition
into pseudo_util.c since it's not used anywhere else.
* Further improvements in the fork() support:
We now recognize both positive and negative forms of PSEUDO_DISABLED,
so we can distinguish between "it was removed from the environment
by env -i" (restore the old value) and "it was intentionally turned
off" (the new value wins).
* clone(2) support. This is a little primitive, and programs might still
fail horribly due to clone's semantics, but at least it's there and
passes easy test cases.
Plus a big patch from Mark Hatle:
Cleanup fork/clone and PSEUDO_DISABLED
guts/fork.c:
* cleanup function and make it more robust
* be sure to call pseudo_setupenv prior to pseudo_client_reset
to match exec behavior
pseudo_wrappers.c:
* fix mismatched type in execl_to_v call via typecast
* Simplify fork call via single call to wrap_fork()
* be sure to save pseudo_disabled
* be sure to call pseudo_setupenv prior to pseudo_client_reset
to match exec behavior
tests:
* Add a test of whether pseudo can be disabled/enabled on a fork.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
an environment variable across fork().
This does a couple of things. One is, move the fork/vfork and execl*
functions out of the standard wrapper environment. Having done that,
the execl*() wrappers are rewritten so they can use real_execv*()
when running in antimagic mode.
This allows us to run pseudo entirely in antimagic mode. Which is
handy, because we also add support for enabling/disabling antimagic
mode in the child process on a fork, using an additional hook for
this in the pseudo_client_init() routine. The redone fork() wrapper
now calls pseudo_client_init() in the child process after a
successful fork even if it's in antimagic mode already.
This is not yet well-tested.
|
|
database to preserve the meaning of log messages in old databases
in case of renumbering... but these tables were never used, and
the tables used to create them were, in fact, out of date and
inconsistent with the range of operations now supported. Remove
the vestigial remains of the never-implemented feature.
Also, update .gitignore to stop git from complaining about some
more generated files.
|
|
are generated from text files and templates, making it now (we hope)
impossible for the list of strings to get out of sync with the
enum.
|
|
in an incoming "maketables".)
|
|
|
|
one device to another, for instance.
|
|
|
|
|
|
|
|
one.
There's a long story here, but to abbreviate it:
The shell script was annoying at best to maintain and starting to
show signs of not really being the right tool for the job. For
various reasons, we have some other Python stuff in our build system,
so we picked Python as the language we were already using for other
stuff. We think this works with anything reasonably recent (around
Python 2.4 through 2.6).
There's a little bit of cleanup, also, of the wrapper templates.
|
|
Note that this replaces an existing getxattr() wrapper which canonicalized
pathnames for use in a chroot() environment. The need for that to work
in some possible future may require us to revisit this.
|
|
Current versions of gnu coreutils 'cp' were attempting to set permissions
using fsetxattr. As a workaround, we need to return ENOTSUP so that it falls
back to using fchmod, which pseudo does know how to wrap.
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|
|
argument -- which turns out to be a problem, as this controls the
canonicalization. The upshot is that some invocations of symlinks
were corrupting $0.
|
|
in one place, but then often reuse them in other project directories.
As a result, the --with-sqlite path is no longer the right path to
use for RPATH later in the program's life. In the past, we were
using chrpath to change the path to something more suitable, but this
doesn't work if the new path is longer.
Instead, allow an explicit --with-rpath setting which our build
system can use. The logic is:
* If you don't specify --with-sqlite or --with-rpath, you get
no -Wl,-R
* If you specify an empty path in --with-rpath, or --without-rpath,
you get no -Wl,-R
* If you specify --with-rpath with a non-empty path, you get
-Wl,-R<path>
* If you specify --with-sqlite, but not --with-rpath, you get
-Wl,-R$(SQLITE)/lib
|
|
mknodat() uses a dev_t * argument, not a dev_t argument, so the
0 passed into wrap___xmknodat() by wrap_mkfifoat() is an invalid
argument rather than an unused but valid zero. Change mkfifoat
to create a dummy dev_t and pass its address.
|
|
in the default location.
|
|
... Also, make that error message a debugging message, since that's
a normal and legitimate use case.
|
|
|
|
flaw in the code. Yes, Virginia, your printf arguments must match
their format strings, and size_t is not the same thing as int.
|
|
that for a macro that does it correctly. Why not just use strcmp,
you ask? Because we aren't doing a string compare, we're looking
for a prefix.
|
|
an LD_LIBRARY_PATH that included the pseudo library directory
and some other directories, the other directories would get
wiped out. Also a couple of whitespace rationalizatoins.
|
|
all) noticed, that offsets.c was busted. Fixed.
|
|
pseudo_saved_sigmask occurred outside of the check against
antimagic, calls to wrapped functions made by the internals
of a wrapper resulted on overwriting pseudo_saved_sigmask,
so we restored the already-blocked sigmask. This could
result in programs getting their sigmask permanently changed
to block various signals, most crucially SIGCHLD.
We now stash the value immediately before calling a wrapper,
and never stash it if we're not calling a wrapper. Now
the anti-magic stuff happening inside wrappers isn't trashing
the signal mask. YAY!
|
|
in processes started under pseudo.
|
|
For reasons that may never be known, the /usr/bin/find on
RHEL5 contains its own local copies of regcomp() and
regexec(). Thus, when pseudo makes calls to these functions,
it gets the local copies in the binary instead of the ones in
libc.
But wait! That's not all. There's also the fascinating
detail that, for reasons unknown, these local copies have
an incompatible API, such that:
regexec(pattern, list, 1, pmatch, 0);
can write to more than one element of pmatch, and since
that's a local array of one member, that means that they
can crush something on the stack, such that a couple of
function calls later you get a segfault in Nothing In
Particular.
So. We try to grab the real regcomp/regexec from libc,
using dlsym, and if we can't, we fall back on whatever the
defaults were.
Inexplicably, this code actually made pseudo crash less often
on at least one target. This is madness within madness, and
I really have no idea why on earth /usr/bin/find, on a Linux
system, would have its own local copies of regcomp/regexec,
let alone why it would have copies that had substantially
different semantics.
|
|
bug in the speculative-unlink operation.
The intent is to mark and then confirm or cancel the delete. This
removes the quirk where we tried to stash old database entries,
which didn't handle directories anyway; "rmdir non-empty-directory"
is a bit too common a case to dismiss as unthinkable.
|
|
Fixed a couple of allocation issues, corrected an off-by-one
error in environment setup.
|
|
Fix an obvious ld_preload/ld_library_path mixup in pseudo_util.c
Signed-off-by: Richard Purdie <richard.purdie@intel.com>
Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
|