Fix: ust-app: per-PID app unregister vs tracing stop races

There are various races with UST application unregister performed
concurrently with tracing stop operation when tracing with per-pid
buffers. This randomly affects availability of data shortly after the
data pending check returns that no more data is available.

ust_app_stop_trace_all() iterates on all applications in the ust_app_ht
hash table to issue a flush on all buffers. This is needed to ensure
that the sub-buffers being written to are made available to the
consumer, for both data consumption, and for the data pending check.
Failure to execute the sub-buffer flush makes following data pending
check return that there is no data in the buffers too early, thus
resulting in an incomplete trace.

It is therefore important that an application flushes all its buffers
before it is removed from the ust_app_ht.

This is where ust_app_unregister() needs to be fixed. Note that
ust_app_unregister() executes concurrently with
ust_app_stop_trace_all(), only taking the per-session lock. The order of
flush vs hash table removal therefore matters:

We need to push the metadata before removing application from
ust_app_ht. We also need to issue a flush for all application buffers
before removing the application from ust_app_ht.

Once this is fixed, there is yet another race, this time in
ust_app_flush_trace() (now renamed ust_app_flush_session()). It is
caused by the use of ustctl_sock_flush_buffer() which asks the
application to perform the buffer flush. Unfortunately, if the
application vanishes (not reachable anymore), but its unregistration has
not yet been processed by sessiond, then ust_app_stop_trace_all() will
fail to flush the application buffers, because
ustctl_sock_flush_buffer() will fail.

This final issue is fixed by asking the consumer daemon to flush the
associated channel rather than relying on the application.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: data pending: allow empty streams

There are cases where a stream can be completely empty (no packet to
write) with UST: for instance, if a traced application is either
preempted for a long time, terminated, or stopped, between reserve and
commit. This will make the consumer consider that this stream has no
data ready. If this situation occurs in the first sub-buffer of a
stream, this stream will have no data at all (0 bytes).

Therefore, we need to let the data pending check consider that no data
is pending in this situation, otherwise it can make the data pending
check always return that there is data pending.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: ust snapshot: cleanup after error

The "break" statement on error skips the rest of the functions, thus
leaving test applications running after the end of the test, which is a
side-effect on the following tests.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Mismatching return type in ust_app_ht_alloc() stub

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: lock file already taken is error, not warning

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: ust-app null pointer check needed for main refactoring

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: relayd: centralize thread stopping function

Rather than relying on having main.c and live.c threads both using the
same notification pipe from different stop_thread implementations,
centralize thread stop in one central function exposed to both main.c
and live.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: Replace all perror() uses by the PERROR macro

PERROR prints the perror() message in our own logging format rather
than going straight to stderr.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Refactor relayd main/set_options/cleanup

- Enforce symmetry between allocation and teardown,
- Handle all errors,
- Return all errors as EXIT_FAILURE,
- Standardize on zero being success, nonzero being error,
  (rather than < 0 being error),
- Fix pthread PERROR: we need to store ret into errno before
  calling PERROR, since pthread API does not set errno,
- Join errors now fall-through, rather than rely on the OS
  to teardown the rest.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Refactor consumerd main/cleanup

- Enforce symmetry between allocation and teardown,
- Handle all errors,
- Return all errors as EXIT_FAILURE,
- Standardize on zero being success, nonzero being error,
  (rather than < 0 being error),
- Fix pthread PERROR: we need to store ret into errno before
  calling PERROR, since pthread API does not set errno,
- Join errors now fall-through, rather than rely on the OS
  to teardown the rest.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Refactor sessiond main/cleanup/ht-cleanup

Main reason for this entire patchset: move teardown of ht-cleanup thread
_after_ the sessiond cleanup which needs to destroy hash tables. This
fixes leaks at sessiond teardown that makes the valgrind output hard to
use. As this has been done, various other issues with error handling,
leaks, and symmetry of allocation and teardown have been fixed, which
makes this a refactoring.

- Enforce symmetry between allocation and teardown,
- Handle all errors,
- Return all errors as EXIT_FAILURE,
- Standardize on zero being success, nonzero being error,
  (rather than < 0 being error),
- Fix pthread PERROR: we need to store ret into errno before
  calling PERROR, since pthread API does not set errno,
- Join errors now fall-through, rather than rely on the OS
  to teardown the rest.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Missing error handling: consumer_signal_init should return its error status

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: consumerd: lines over 80 columns

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: consumerd: use PERROR rather than perror

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Cleanup: lttng-consumerd: remove stale TODO comment

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Misleading error handling: utils_create_pid_file() should return 0 on success

fprintf returns a positive value on success, which is used as return
value for utils_create_pid_file(), which is odd. This is in preparation
for main() refactoring.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Missing error handling: consumerd parse_args should return errors to caller

In preparation for having main() handle the return value.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Missing error handling: ust_app_ht_alloc should return error status

In preparation for having sessiond main() handle the error.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: ust app leak on UST buffer creation error

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: leak on UST buffer registry creation error

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: agents ht leaks on destroy session

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: add missing RCU read unlock

Master commit 24eb8569dc841f318d7964ca9a0ad0a4f9508a8e
"Fix: Missing RCU read locks in syscall_list_channel()"

Introduces unbalanced RCU read-side lock/unlock. Should be applied to
master and stable-2.6.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: help message for Python agent mentions Java

Bad copy/paste.

Signed-off-by: Julien Desfossez <jdesfossez@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: don't destroy the sockets if the snapshot was successful

Missing a goto to skip the error condition that was destroying the
relayd sockets even if a snapshot was successful. We want to keep them
open to reuse them for the next snapshots.

Signed-off-by: Julien Desfossez <jdesfossez@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing RCU read locks in syscall_list_channel()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing RCU read lock in kernel_destroy_session()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: channel names are not validated

This patch ensures:

  1. A channel name does not contain any '/' character, since
     relative paths may be injected in the channel name
     otherwise (knowing that the channel name is eventually
     part of a file name)
  2. A channel name does not start with a '.' character, since
     trace readers (Babeltrace is one of them) could interpret
     files starting with a dot as hidden files and ignore
     them when opening the CTF trace

Fixes: #751
Signed-off-by: Philippe Proulx <eeppeliteloop@gmail.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: Missing whitespace

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Apply consumer URI changes to all domains

See the associated bug report for a lenghty explanation of the issue
and of this fix. It fixes an issue when saving a live session's
configuration that was created by loading an .lttng file.

Fixes #866

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: Fix a typo in the MI tests

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Test fix: Don't invoke the system's lttng client

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Tests: Add a test to check disabled events are loaded in the correct state

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Don't leave events enabled if they were saved in a disabled state

Events are enabled by default on creation. The session configuration
loader must make sure to disable them after creation when restoring an
event that was saved in the "disabled" state.

Reported-by: Marcel Hamer <marcel.hamer@ericsson.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: "Any" loglevel's value is -1 and not 0

lttng_disable_event_ext() was erroring-out on loglevel != 0.
loglevel 0 is a valid level whereas "-1" is used an unknown/unspecified
in the rest of the code.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Test fix: LTTNG_SESSION_CONFIG_XSD_PATH expects an absolute path

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Memory leak on error in alloc_argv_from_user_opts()

*** CID 1256181:  Resource leak  (RESOURCE_LEAK)

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: Unchecked close() return value in utils_create_lock_file()

*** CID 1226188:  Unchecked return value  (CHECKED_RETURN)

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: No check of the return value of lttng_ht_del()

*** CID 1230590:  Unchecked return value  (CHECKED_RETURN)

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: Mixed enums used for return code in send_sessiond_channel()

*** CID 1132649:  Inferred misuse of enum  (MIXED_ENUMS)

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: Cast poll() return value to void

*** CID 1019969:  Unchecked return value from library  (CHECKED_RETURN)

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Unreachable error logging in set_option()

*** CID 1256136:  Logically dead code  (DEADCODE)

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: OOM leaks in sessiond modprobe.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: sessiond modprobe.c: coding style

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: test_utils_expand_path.c prototypes

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: test_utils_expand_path.c: out of memory error handling

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: test test_ust_data.c: check OOM

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: test libc-wrapper: check malloc OOM

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: test libc-wrapper: main parameters

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: lttng-ctl: use zmalloc(), missing OOM check

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: filter-parser.y: use zmalloc(), missing OOM check

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: utils.c: check str*dup OOM

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: implement zmalloc as static inline

Whenever we can implement as static inline rather than macro, it's
better (type checking of arguments, less chances of hitting preprocessor
aweful semantic corner-cases....).

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: rculfhash: use zmalloc()

Follow lttng-tools rules.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: ini.c: use zmalloc()

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: lttng lttng.c: check strdup OOM

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: lttng conf.c: use zmalloc()

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: lttng view.c missing strdup OOM check

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: use zmalloc() in lttng view.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: use zmalloc() in lttng list.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: oom error check, realloc oom leak

print_exclusions can return a NULL pointer on OOM, should be checked.
Also, realloc always need to store result in a different variable than
its input in case it fails (and its returned pointer needs to be checked
for NULL).

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: use zmalloc in lttng enable_events.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: missing strdup oom check in lttng create.c

Introduce tmp_path to ensure that no code path can possibly try to free
the return value of utils_get_home_dir(). Re-using alloc_path for both
static and dynamically allocated pointer is error-prone.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: lttng add_context.c: use zmalloc

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: add missing static to function modprobe_remove_lttng()

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: add missing str*dup oom check in sessiond main.c

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: add missing sessiond kernel-consumer missing str*dup oom checks

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: add missing strdup OOM check

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: add missing synchronization point for before app test case

Fixes a race where the application could generate all its events before
trace start.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: tests: wait output hide Terminate errors

Also: Don't hide kill errors.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: tests: remove killall, add missing SIGTERM handlers

Applications may change name and, thus, be missed by using
killall.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: high throughput test: reset bw limit on sigterm

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: tests: add missing wait, document missing synchro

Move all wait ${!} that target a single process to "wait", to minimize
the chances to forget some background process in the future.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Document test anti-patterns

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: test flaky sleep and wait patterns

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: tests: don't use pidof to wait for test apps

Use the bash shell "wait" to wait for all background tasks rather than
the racy "pidof". Indeed, it's possible that applications have been
forked, but not executed yet, when pidof is done, which would therefore
miss applications. Using "wait" from the shell solves this.

If we want to be really strict, we should have sessiond, consumerd, and
relayd export a file containing their own PID, and wait for this instead
of using pidof. But this will be for another fix.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Live tracing does not honor live timer after first tracefile with tracefile rotation

When we pass to the 2nd sub-file (or following sub-files) of a stream in
relayd, the live timer has no visible effect from a live reader
perspective, and then everything is flushed when we reach the following
sub-file.

This is caused by the reset of stream->total_index_received after each
tracefile rotation. It should keep on incrementing to match what is
expected by check in check_index_status():

  (rstream->total_index_received <= vstream->last_sent_index)

Fixes #863

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Handle empty daemon configuration file lines

Empty daemon configuration file lines such as
my_option=

will cause set_option to be called with an arg value of length 0
which should display a warning.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Docs: Add -f/--config option to lttng-sessiond man page

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Cleanup: spaghetti function return path

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: relayd: unbalanced RCU read-side lock/unlock

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: relayd: unbalanced RCU read lock/unlock

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: lttng-relayd: unhandled out of memory error

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: relayd viewer stream: unhandled out of memory error

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Build fix: Apps defining _LGPL_SOURCE must link to urcu-bp

Defining _LGPL_SOURCE causes lttng-ust tracepoint headers to
inline uses of userspace-rcu, which make linking against urcu
necessary.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Define _LGPL_SOURCE in test apps' tracepoint provider files

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Define _LGPL_SOURCE in test applications

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in cmd_list_events()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in list_lttng_agent_events()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in cmd_list_channels()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_locks in cmd_list_domains()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in cmd_snapshot_list_outputs()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in save_domains()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in ust_app_get_nb_stream()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: UST subbuffers silently dropped on moderate trace traffic

Well, it looks like we really screwed up on this one.

lttng-tools commit 02b3d1769d5f8a33e4109b1e681141c9295dfda6 introduced
an important regression for lttng-ust tracing in the consumer daemon:
after reading a sub-buffer, a check has been added to see whether there
are more sub-buffers available to read, and if it is the case, it
ensures the wakeup pipe will be awakened again.

The issue lies in the use of ustctl_put_next_subbuf() in this check.
This acts as if the sub-buffer has been read, when in reality it has not
been read. It therefore trashes the data contained by this sub-buffer.

This check should use ustctl_put_subbuf(), which does not move the
consumer position.

This is a severe bug, and the fix needs to be applied to stable-2.6,
stable-2.5, and stable-2.4.

Fixes #861

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Missing rcu_read_lock in get_session_max_subbuf_size()

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: create/destroy a splice_pipe per stream

We had a per-thread splice_pipe (one for data and one for metadata), but
in case of error, we would end up filling the write side of the pipe and
never emptying it. This could lead to leaking data from one session to
the other, but also to stall the consumer trying to splice into a full
pipe.

Now we create a splice_pipe per-stream, so it is destroyed when the
session is destroyed.

Fixes: #726
Signed-off-by: Julien Desfossez <jdesfossez@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Note find_ust_app_context must be called with RCU read lock

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: empty indexes_ht before destroying it.

Fixes: #722
Signed-off-by: Julien Desfossez <jdesfossez@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: removed useless key from relay_index

This field was never used except in debug messages (instead of using the
good value that is part of index_n).

Signed-off-by: Julien Desfossez <jdesfossez@efficios.com>
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>

Fix: Ambiguous agent event filter bytecode ownership

A shared ownership of the filter bytecode between UST events
and Agent event led to a leak in cmd_enable_event() which was
fixed previously. However, that fix introduced a bug which
passed NULL to the agent event creation function.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>