Josh Stone [Mon, 2 Mar 2015 23:18:37 +0000 (15:18 -0800)]
tapset: tag many inline embedded-C blocks as pure
If embedded-C is not explicitly marked pure, then it is assumed to have
side effects, and can never be optimized away. In many cases, we're
simply referencing a constant value, so this is certainly pure, and
should be tagged so.
This patch doesn't bother updating instances that are in impure contexts
anyway, like the many cases of @__syscall_gate et al. which test the
constant to perhaps issue probe "next".
Josh Stone [Mon, 2 Mar 2015 21:06:06 +0000 (13:06 -0800)]
tapset/linux: Introduce and use @task macro
The @task macro performs the very common @cast to a task_struct.
The embedded-C bodies of task_current() and pid2task() are now wrapped
by @task, which gives them a debuginfo type on the return value. With
autocast type propagation, this removes the need for any explicit @cast
in many places.
Other places which take untyped task pointers as parameters, for
instance, now use @task as well to simplify their code.
Tangentially related, task_start_time() was broken for a while by the
change of real_start_time from timespec to u64, commit 57e0be041d9e2.
This is fixed and added to buildok/task_test.stp.
Abegail Jakop [Mon, 2 Mar 2015 16:13:21 +0000 (11:13 -0500)]
PR10488: add proctop.stp example script
proctop.stp: a top-like script that periodically prints out process
information such as state, mem usage and time, while keeping track of
processes that exited while the script was running.
Jonathan Lebon [Mon, 23 Feb 2015 16:10:03 +0000 (11:10 -0500)]
stmt_rel.exp and vars.exp: adapt to latest kernel
The bio.c file was moved from the fs directory to the block directory.
The two test cases stmt_rel.exp and vars.exp used functions from that
file to make some sanity checks. Update the testcases to be more
flexible.
Jonathan Lebon [Fri, 20 Feb 2015 18:24:18 +0000 (13:24 -0500)]
nearest probes: allow ENUMERATED and WILDCARD linenos
This commit reverses commit 2ff2868 which added the restriction of only
allowing .nearest probes on ABSOLUTE and RELATIVE linenos. We now allow
this for all lineno types.
WILDCARD lineno types do not need any special handling for .nearest
probes since the resulting linenos are already known to be correct. For
ENUMERATED linenos however, we now must add .nearest handling: for each
lineno enumerated, if we fail to find an address, we check if there is a
nearby lineno we can instead probe.
This commit also fixes failures in the unprivileged_myproc.exp testcase.
Related PRs: PR17906, PR17986.
Abegail Jakop [Wed, 25 Feb 2015 21:48:30 +0000 (16:48 -0500)]
PR16974: wrap if, try and for tmpvar decls in unions
translate.cxx: in c_tmpcounter::visit_block, the tmpvar declarations
in the generated c code are wrapped in a union, with each
block->statements[i] wrapped in a struct, to reduce context memory
allocation. following ::visit_block, the tmpvar declarations in
::visit_if_statement, ::visit_try_block and ::visit_for_loop are
wrapped in a union. since this causes quite a lot of nesting, comments
are added to the generated c code to indicate the source file and
line the struct/union tmpvars correspond to.
Abegail Jakop [Wed, 25 Feb 2015 21:32:17 +0000 (16:32 -0500)]
PR16974: move struct+visit outputting to new fn(s)
translate.cxx: from c_tmpcounter::visit_block(), move the portion
to wrap the statement->visit() part in a struct declaration to a
function, ::wrap_visit_in_struct()
Frank Ch. Eigler [Mon, 23 Feb 2015 18:12:00 +0000 (13:12 -0500)]
dtrace.in python3 support, testsuite parts
Prepare for fuller /usr/bin/dtrace python3-compatibility by extending
the dtrace.exp test case to run with -each- python interpreter.
Various embedded input files were also factored out into separate
files for easier diagnostics.
David Smith [Mon, 23 Feb 2015 19:04:48 +0000 (13:04 -0600)]
Add minor syscall testsuite test case tweaks.
* testsuite/systemtap.syscall/chmod.c: Avoid nesting by looking for return
values. Test old 16-bit syscall variants on all 32-bit platforms.
* testsuite/systemtap.syscall/uid16.c: Ditto.
* testsuite/systemtap.syscall/getgroups.c: Test old 16-bit syscall
variants on all 32-bit platforms.
* testsuite/systemtap.syscall/poll.c: Add 'epoll_create1' to the covered
list.
Josh Stone [Fri, 20 Feb 2015 19:40:56 +0000 (11:40 -0800)]
PR18000: Always use new tracequery objects directly
When new tracequery objects are built, always use them directly from the
build directory, regardless of whether the systemtap cache is enabled.
This allow traceponts to be found even under --disable-cache.
Martin Cermak [Tue, 17 Feb 2015 18:19:00 +0000 (19:19 +0100)]
PR16716 partial fix: Fix types and nesting for 'syscall.{mbind,move_pages,migrate_pages}'.
* runtime/linux/compat_unistd.h: Define __NR_move_pages for rhel5 powerpc platform
* tapset/linux/nd_syscalls.stp: Fix types and nesting for 'syscall.{mbind,move_pages,migrate_pages}'.
* tapset/linux/syscalls.stp: Fix types and nesting for 'syscall.{mbind,move_pages,migrate_pages}'.
* testsuite/buildok/nd_syscalls-detailed.stp: Test all convenience variables.
* testsuite/buildok/syscalls-detailed.stp: Test all convenience variables.
* testsuite/systemtap.syscall/numa.c: New testcase.
Frank Ch. Eigler [Mon, 16 Feb 2015 20:50:55 +0000 (15:50 -0500)]
testsuite listing_mode: Adapt to changed line numbers
commit 2538abbc94 inserted some blank lines to separate function
declarations and bodies, but the .exp file needs correponding updates
for the pushed-down line numbers.
Frank Ch. Eigler [Fri, 13 Feb 2015 15:31:32 +0000 (10:31 -0500)]
BZ1128209: add a configure --{enable,disable}-virt option
It was desirable that a systemtap rpm be buildable in a predictable
manner (in this case, without virthost support), even if prerequisites
(libvirt etc.) were accidentally installed.
Josh Stone [Fri, 13 Feb 2015 02:06:54 +0000 (18:06 -0800)]
PR17966: never generate .call for an unresolved .return $var
The saved-return code already skipped .call for any $var that generated
an error, but variables attempted in a symtab context don't generate
errors until later. This means a @defined-guarded $var might create an
unguarded access in .call that causes unrecoverable errors further on.
Now the .call will be skipped if the target_symbol persists for any
reason, regardless of its error state. Any resolved target_symbol
should become something quite different, usually a loc2c function call.
The minidebuginfo.exp test is extended to exercise @defined and .return
scenarios as well.
Josh Stone [Thu, 12 Feb 2015 21:45:15 +0000 (13:45 -0800)]
runtime/linux: block utrace-via-tracepoints without tracepoints
The code in autoconf-utrace-via-tracepoints.c will compile just fine on
kernels which do have those tracepoints, but not CONFIG_TRACEPOINTS=y.
That would only result in ENOSYS at runtime.
To get this autoconf working at compile time, check the config directly.
Reported-by: Crestez Dan Leonard <cdleonard@gmail.com>
Frank Ch. Eigler [Tue, 10 Feb 2015 22:42:35 +0000 (17:42 -0500)]
gcc5 -Wformat=2 tweaks
-Wformat=2 now includes -Wformat=signedness, which sadly some of our
violates. Correct this throughout with some casts. In the case of
stapdyn, we downgrade to -Wformat=1 due to problems with
dyninst-devel.
Regenerate auto* configury on fedora rawhide, since automake was
complaining about needing "subdir-objects" in AUTOMAKE_OPTIONS.
William Cohen [Tue, 10 Feb 2015 12:36:51 +0000 (07:36 -0500)]
Make all socketcall.connect probes optional.
Both 32-bit and 64-bit arm do not implement syscall socketcall. A
number of tests were failing on those architectures because two of the
syscall socketcall return probes were not optional. Made those two
syscall socketcall variants optional like all the others.
William Cohen [Sat, 7 Feb 2015 05:42:07 +0000 (00:42 -0500)]
The 32-bit arm does not remap epoll_wait or poll syscalls
The poll test was failing because the test was expecting the 32-bit
arm processor to remap the epoll_wait syscall to epoll_pwait and to
remap the poll syscall to ppoll. Only the arm64 is remapping those
syscalls.
William Cohen [Sat, 7 Feb 2015 04:15:17 +0000 (23:15 -0500)]
Do not expect that alarm(-1) to work for the syscall alarm.c test
Setting a alarm for a time in the past (alarm(-1)) doesn't make much
sense. Some kernels allow this and do not flag an error. Others such
as the Fedora 20 32-bit ARM kernel return -EINVAL. Allow the
syscall alarm test to pass with kernels that do return an error.
Abegail Jakop [Thu, 5 Feb 2015 21:30:27 +0000 (16:30 -0500)]
PR12276: example usages of symfileline() to profile
EXAMPLE/profiling/periodic.stp: for the case of printing out symdata for a
kernel address, if file:line information is available, then print out the
symbol@file:line information.
EXAMPLE/profiling/fileline-profile.*: use symfileline() and symname(),
or symdata if file:line info is not available, to print out information
on the top 20 addresses sampled from when the stap script was running.
Abegail Jakop [Mon, 2 Feb 2015 23:05:47 +0000 (18:05 -0500)]
PR12276: check all sequences in _stp_linenumber_lookup
runtime/sym.c: addresses can only increase within a sqeunce, but at the
start of the new sequence, the starting address can be lower than the
previous'. _stp_linenumber_lookup() was changed to iterate through all
sequences until a suitable address is found, instead of giving up when
the line data address passes the desired address.
git://sourceware.org / systemtap.git / log