| Age | Commit message (Collapse) | Author |
|---|
|
There's a race between LOOP_CTL_GET_FREE and LOOP_SET_FD. Work around it
by just retrying if we get EBUSY on the LOOP_SET_FD call. This is what
similar code in ChromeOS already does.
Bug: http://b/135716654
|
|
marshalling for vforked subshells, fix rc returned from run_pipeline for
variable assignment. (Well, sort of: X=$(false) still needs rc = 1.)
|
|
ext stores extended attributes in a way that makes all the numbers in
the test incorrect. For now, just disable the test in that case.
|
|
|
|
A TOYFLAG_NOFORK command must run in the context of toysh, but a MAYFORK
can either run standalone or run in the toysh process. MAYFORK means it
cleans up after itself: no leaked resources (malloc, mmap, filehandles, etc),
even in error_exit() paths that would longjmp() back to the shell. It also
doesn't discard anything we need to retain (don't close stdout, change
toys.optargs[] so we can't free it, etc)...
|
|
output 5 digits of octal stat data instead of 4, due to hardwired leading 0
instead of %04a sprintf prefix.
|
|
|
|
make: *** No rule to make target 'generated/Config.probed', needed by 'Config.in'. Stop.
make: *** Waiting for unfinished jobs....
|
|
|
|
No current caller except grep needs this, but consistency seems like a
good idea.
Also change the xregcomp error message to be a bit more human-readable,
rather than mention an implementation detail.
|
|
|
|
All regcomp callers are now going through xregcomp.
|
|
We use (char *)1 and (char *)2 to indicate errors (they can never be valid
pointers because both malloc() and mmap() return aligned memory and those
align down to NULL, plus Linux maps 4k at the bottom to catch null
dereferences anyway), and then typecast it to long (trusting in LP64) to
do an integer <=2 comparison... except that needs to be UNSIGNED long or
else pointers in the top half of the virtual memory space become negative
and the <=2 false positives them as errors. (Oops.)
|
|
POSIX says there's no such thing as an empty regular expression. The
grammar excludes the possibility:
https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap09.html
BSD agrees with POSIX, and Android and macOS' BSD-based implementations
reject the empty regular expression.
GNU apparently disagrees.
Luckily, BSD does accept the empty *sub* expression `()`, despite their
error message for REG_EMPTY being "empty (sub)expression". This is
presumably a bug, except there's explicit code to support it that is at
least 26 years old:
https://github.com/freebsd/freebsd/blame/master/lib/libc/regex/regcomp.c#L383
This workaround also works fine with glibc.
If we want GNU behavior, I'm struggling to come up with another way to
fake it. If we want POSIX behavior, we could easily just add a check to
reject "" on glibc.
Also switch to xregcomp().
|
|
Just use `ASAN=1 make test_grep` or whatever.
You'll probably want to set $ASAN_SYMBOLIZER_PATH to point to
llvm-symbolizer, but Debian makes that annoying by calling the
symbolizer /usr/bin/llvm-symbolizer-4.0 or whatever, and ASan refuses to
use it:
==43370==ERROR: External symbolizer path is set to
'/usr/bin/llvm-symbolizer-4.0' which isn't a known symbolizer. Please
set the path to the llvm-symbolizer binary or other known tool.
My usual workaround for this is to drop an llvm-symbolizer symlink in
the current directory, and I'm happy to automate that in the script to
make it require no knowledge of any of this nonsense, but haven't done
so in this initial patch.
I tested that this is actually working by reverting the grep fix and
running `ASAN=1 make test_grep`.
|
|
There's probably a way to filter out the existing ones, but it's likely
to be pretty painful and isn't immediately necessary.
(I hit this case on Android cloud x86 devices.)
|
|
We need /proc/modules *and* a directory of modules.
|
|
Rather than just claim we're doing so.
Android cloud emulators don't have dummy0 like regular devices do.
|
|
Weaken the test slightly so that the extra 4KiB used for SELinux is
ignored.
Unfortunately all the du tests suffer the same problem, and I'm less
sure what to do there. Mounting a read-only file system without the
extended attributes might be the best idea? But since this test modifies
the file system...
|
|
On Android you can't even get the hostname unless you're root, to avoid
device fingerprinting/tracking.
Also modernize by using skipnot, which fixes the incorrect test name in
the SKIP: output.
|
|
This isn't currently used in AOSP, but it is used in some of the other
codebases I can easily search, and it's a better fix for test flakiness
than inserting a sleep.
|
|
And when I say "fix", I mean "hack around in the worst possible way".
Android+mksh is often slow enough that we can do the pgrep before the
killed process has actually exited. (I'm seeing roughly 100% failure
rate on cloud x86 emulators, and about 75% on a real Pixel 2.)
|
|
It turns out some Android devices have an empty /etc/passwd and/or
/etc/group, which was defeating the previous workaround. Switch to
testing the intention more directly: we'll try the file in /etc, and if
that didn't work, we'll assume we need a workaround.
|
|
Test: no EBADF in `strace -e close ./toybox tac /proc/version`
|
|
|
|
|
|
|
|
This was basically just to help me think out loud while discussing
loopfiles_lines().
|
|
I started this last night, but thought I'd aim to send multiple small
patches rather than work through all the callers and send one big patch.
I've deliberately chosen the ugly name `allocated_length` because we've
had historical bugs where folks think this a line length in the sense of
the return value. I do wonder whether we should actually have some kind
of getline() wrapper that hides the `char *`/`size_t` pair in lib/,
which makes the function easier to use in most cases but does add the
less common gotcha that you wouldn't be able to getline() through
multiple files at once (which does happen in at least one toy).
But maybe the real fix is to look harder for places where we can just
use loopfiles_lines? Speaking of which, should we actually add two more
arguments to that? Specifically: switch it to getdelim() rather than
getline() behind the scenes, and also add a way to have the trailing
'\n' automatically removed, since that seems to be what most callers
want?
Anyway, that seemed like enough questions that it was time to send this
initial patch out before doing too much more...
|
|
There's no /etc/group on Android.
Loop devices are in different places, but that test doesn't seem to
offer anything beyond the block/char special device tests later, so
remove that rather than fix it.
Fix the block/char special tests (the expectations were missing, but
they're never normally run on the host because the mknod fails).
Likewise add the missing expectation to the "ownership" test, which
isn't normally run on the host because the chown fails.
The "sparse without overflow" test was broken because of the mksh/bash
disagreement about whether in `VAR=val shell_function` that's a
temporary assignment to VAR or one that outlives the command. Switch SUM
to being a function like LST before it. Also rewrite the TARHD stuff
because Android (a) doesn't have hd and (b) doesn't have process
substitution in its shell. (I wish I'd noticed TARHD earlier --- I'd
been manually sending the tar files to xxd for debugging instead!)
With this patch, all the tar tests now pass for me on both the device
and host.
|
|
|
|
|
|
This was embarrassing... A build script that was using xargs -I silently
started to ignore -I when we switched over to toybox xargs. The owner of
the script has rewritten it to use the shell read builtin instead, but
it's pretty unhelpful to silently ignore an option that radically
changes how xargs behaves. (The -I behavior sounds sufficiently
different from normal behavior that I'm not sure I've understood, and am
just sending this cleanup rather than actually implementing -I correctly.)
Bug: http://b/137832162
|
|
The change from ioctl to xioctl in
844f0d1aae28c54bbb672f53ba2cda182c572129 broke ifconfig for non-root on
Android. SIOCGIFHWADDR has been protected to prevent device tracking
since Nougat.
Bug: https://issuetracker.google.com/137671590
|
|
|
|
The first bug appeared as a memory overwrite, but was actually visible
without hwasan: basically any `grep -F` that let to multiple matches on
the same line was broken.
The second bug was another memory overwrite, visible when I ran the
existing grep tests.
Bug: http://b/137573082
|
|
This fixes the extended attribute settings for the ip route change
command; extended attributes are collected in a separate buffer
and added to the RTM message in a specific RTA_METRICS attribute.
Only added setter support for the RTAX_MTU (per-route MTU), as that
was the only option currently parsed.
Also added support for displaying extended attributes, even those
not currently settable.
|
|
Pointed out during code review of the recent refactor.
|
|
Used by some ART tests and also some LLVM tests. (The motivating example
is the latter, but I noticed the former when looking for other users.)
Bug: http://b/137298656
|
|
Before this patch, we're effectively doing `pidof -x` all the time. This
patch changes names_to_pid() to allow us to say whether or not we want to
include scripts, and adjusts the callers appropriately.
Also add tests for `pidof` versus `pidof -x` which pass after this
patch, without regressing the existing killall tests.
|
|
On BSD these are actually the same, and there's a -S that you need in
addition. So strictly this is a behavior change for Android (which is
going from BSD grep to toybox grep), but it's a behavior preserving
change for the AOSP build (which is going from GNU grep to toybox grep),
and the latter actually has a checked-in use of -R where the former
doesn't.
|
|
|
|
Rename the existing xbind/xconnect to xbindany/xconnectany, to make room
for new xbind/xconnect that are more like 'x' versions of the regular
bind and connect. Move explicit bind/connect callers over to
xbind/xconnect.
Of the affected commands, only netcat is actually used by Android. It
was the most recent patch to netcat that made the lack of a more
traditional xbind/xconnect apparent.
|
|
The other xargs implementations in the wild don't seem to count the
space taken by each `char *`. Go for bug compatibility unless future
experience proves that to be a bad idea.
|
|
The test relied on the non-builtins not being toybox (or another
multicall binary). A multicall binary will actually look at the
symlinked name (in this case "true") and do what it says on the tin.
Use a tiny shell script instead.
|
|
Investigating why the toybox tar tests fail on Android with toybox dd, I
realized I was rewriting a part of dd I'd rewritten before!
This is a re-send of my 2019-02-22 patch, rebased against the current
ToT...
This patch was originally motivated because after suggesting to the author of
https://stackoverflow.com/questions/17157820/access-vdsolinux/54797221#54797221
that he could tell dd to work in bytes rather than blocks, I realized
that our dd doesn't actually support that. But the rewrite of the main
loop is necessary to fix the incorrect output from the dd calls in the
tar test.
Without this patch, `yes | dd bs=65536 count=1 > fweep` basically gives
random output, based on how many bytes the pipe feels like giving you in
your first read. (As far as I know, dd *without* bs= was fine, but I
can't guarantee that that's true, just that I haven't seen it fail.)
Also switch to TAGGED_ARRAY and comma_* for conv rather than add two more
copies of an undesired idiom. It turned out -- contrary to the belief of
cp(1) -- that comma_scan isn't suitable for this because of its magic
handling of "no" prefixes. (It's actually harmless in cp because none
of the --preserve options begin with "no", but some dd options do.) To
this end, comma_remove is a less-magic comma_scan.
I've also changed an `if` to a `while` because other implementations
allow things like `--preserve=mode,mode` or `conv=sync,sync`. (If we
decide this is a bug rather than a feature, we should at least fix the
error message to be clear that we're rejecting the *duplication*, not
the option itself.)
I've also fixed the ^C behavior by simply adding a direct SIGINT handler
rather than trying to be clever inside the read loop (which is why we
weren't handling the SIGINT until the read returned).
I've also removed `strstarteq` and just added the '=' to each literal
when calling regular `strstart`.
Plus basic tests.
|
|
|
|
|
|
"xz compressed data" was missing a newline. Also switch a couple of
other unnecessary (but correct) xprintf calls over to xputs.
Also add .otf fonts, perf/simpleperf data files, and Android boot
images, sparse images, and device tree blobs/overlays.
|
|
Change names_to_pid() so that we can actually match shell scripts with
long names (the code to get the shell script's name was correct, but
there was an extra test preventing us from actually comparing it to the
sought name).
In kill.c itself, remove a dead test for -l and switch to the FLAG()
macro.
Also extend the tests to explicitly cover long and short names.
|
