Fix: Add granularity to RHEL kernel version checks

Our RHEL kernel version macros were based on the major and minor version
of the distribution version like "7.2" instead of the full kernel version
like "327.4.4". This prevented us from adding compatibility changes with
sufficient granularity.

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

Version 2.8.0-rc1

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

Bump minor tracer ABI version

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

Fix: build against out-of-tree kernel build

Fix following system header include modifications.

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

probes/Kbuild: remove upper level header path

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

Move leftover relative include paths to system paths

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

syscall instrumentation: use system headers

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

tracepoint event instrumentation: use system headers

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

Use system include path in wrappers

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

libs: use system include paths

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

Use system header paths in core implementation files

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

Use system include paths in root directory headers

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

Use system include paths in probes

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

Update gitignore

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

Use system include paths in lttng-types.h

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

Use system include paths in lttng-tracepoint-event-impl.h

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

Use system include path in probes/define_trace.h

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

Use system include paths in probes/lttng-tracepoint-event-impl.h

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

Rename probes/lttng-events.h to probes/lttng-tracepoint-event-impl.h

Ensure we don't have a duplicated header file name in the tree.

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

Add makefile directory to include path

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

Fix: event ctx get size should be after chan ctx

Match the record functions, and the CTF spec. This has no impact
currently because event contexts are not implemented.

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

Fix: filter interpreter with userspace strings

The filter interpreter is populating the "str" register field, but the
string comparison functions expect to use the "user_str" register field
whenever it deals with a userspace string ("user" flag set to 1).

Since we are using __copy_from_user_inatomic(), this is typically
bailing out on anything that requires to read the string (except the '*'
wildcard).

Fixes #943

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

Fix: rename kvm x86 TRACE_SYSTEM to avoid name clash

The lttng-probe-kvm and lttng-probe-kvm-x86 have the same TRACE_SYSTEM
string. Because of that, both modules can't be loaded at the same time.
The second module fails to load silently. Rename the TRACE_SYSTEM of
lttng-probe-kvm-x86 to avoid the clash.

We need to use the LTTNG_TRACEPOINT_EVENT_MAP to perform this
namespacing while still hooking onto the kernel tracepoints.

Fixes #895

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

Fix: format address fields as hexadecimal

Fixes #947

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

PowerPC-64 ABIv1: add build error if KALLSYMS_ALL is missing

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

Fix: system call instrumentation overrides

* All architectures

- For mmap, print the "ret" value as an hexadecimal integer (rather than
  base 10), which better suits a pointer.
- Add missing "clone" override define, which ensures the clone
  system call override is used rather than leaving it as an
  "unknown" system call.

* ARM32

- Add missing output return values to arm_fadvise64_64,
  sync_file_range2.
- Handle the 64-bit parameters of both arm_fadvise64_64 and
  sync_file_range2 correctly on arm32, on big and little endian builds.
- Move the mmap2 system call instrumentation from "pointers" to
  "integers" override header, and add the missing return value
  (sc_exit).

* PPC32

- Add sync_file_range2 system call instrumentation. Handle the
  64-bit parameters on big and little endian builds.
- Move the mmap2 system call instrumentation from "pointers" to
  "integers" override header, and add the missing return value
  (sc_exit).

* PPC64 compat

- Include compatibility (ppc32) header in ppc64 build to trace compat
  system calls on PPC64 big endian.

* x86-32

- Add mmap2 system call instrumentation override.
- Add sync_file_range instrumentation override, correctly combining
  the low/high parts of offset and nbytes into 64-bit fields.

Tested-by: Jonathan Rajotte <jonathan.rajotte-julien@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: page_alloc wrapper incorrect parenthesis

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

Fix: system call tracing for arm-64 compat !CONFIG_UID16

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

Fix: add struct user_msghdr forward declaration for older kernels

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

Cleanup: Remove unused lttng-types module

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

Fix: add missing versions and init call for page_alloc wrapper

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

Fix: Define lttng-tracer metadata only once

Fixes sysfs trying to create the lttng-tracer module version node
multiple times.

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

Fix: Add CONFIG_LTTNG to modules_install target

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

Fix: Use kbuild env instead of a custom var KERNELDIR

We currently use the custom variables KERNELDIR and LTTNG_KERNELDIR
to pass the kernel source tree directory to helper scripts. We should
instead use the built-in Kbuild variable CURDIR which is available in
all sub makefiles and always points to the kernel sources.

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

Update clock offset comment

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

clock offset: print negative value in metadata

In the upcoming 2.8, we can print the negative value of clock offset in
the metadata, and expect users to update their viewers to correct any
bug that could prevent them from handling those negative values.

This happens only in cases where the system sets CLOCK_REALTIME to 0
after boot, thus leading to a negative clock offset value between
realtime and monotonic clocks.

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

Fix: handle negative clock offset

In the unlikely situation where a system sets its hardware clock
(CLOCK_REALTIME) to 0 (Epoch) after boot, the difference

  monotonic - realtime

becomes negative.

Fixup this situation by returning a 0 offset in this case.

This ensures that trace viewer implementations (e.g. babeltrace) which
currently cannot handle the negative offset (known bug) still work with
the generated traces.

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

Fix: sched instrumentation for 4.1-rt11

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

Add RT patch version macros

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

Expect filter context identifiers starting with $ctx.

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

ARM64: wire up compat system calls

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

Add support for arm64 syscalls

[ Edit by Mathieu Desnoyers: rename "x86_64" strings to "arm64" in
  comments. ]

Signed-off-by: Jan Willeke <jan.willeke@harman.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: Ubuntu kernel range for block_rq_complete

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

Fix: Ubuntu kernel range for mm_page_alloc_extfrag

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

Fix: update vmscan instrumentation for kernel 4.5

The following tracepoint changes made their way into Linux 4.5-rc1:

commit 3aa2385111168187f24a6db04697c6fab0fab9b4
Author: yalin wang <yalin.wang2010@gmail.com>
Date:   Thu Jan 14 15:18:30 2016 -0800

    mm/vmscan.c: change trace_mm_vmscan_writepage() proto type

and

commit ba5e9579433aefcdccdec207601e124d3bdf2a71
Author: yalin wang <yalin.wang2010@gmail.com>
Date:   Thu Jan 14 15:18:48 2016 -0800

    mm: change mm_vmscan_lru_shrink_inactive() proto types

Update the lttng-modules instrumentation to match those changes.

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

Fix: check reference counts for overflow

Linux kernel CVE-2016-0728 is a use-after-free based on overflow of the
reference counting mechanism.

Implement a kref wrapper in lttng that validates overflows, and use it
instead of kref_get(). Also check explicitly for overflows on file
fcount counters.

This should not be an issue in practice in lttng-modules because the ABI
is only exposed to root, but let's err on the safe side.

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

Fix: RHEL 7.2 scsi instrumentation

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

Fix: RHEL 7.2 kvm instrumentation

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

Kconfig: describe both module and in-kernel compilation options

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

Add comments and indentation to Makefile

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

Add support for built-in kernel build

To build the LTTng modules as builtins in a kernel image, simply run
the "built-in.sh" script with the path to you kernel source directory
as an argument. Then configure your kernel build and add the "CONFIG_LTTNG"
option. Build as usual and voila!

This is an early implementation with a single configuration option that
enables or disables everything, We could add fine grained config options
for each modules or group of modules.

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

Rename sub makefiles to Kbuild

This follows kbuild's convention and makes it clear that the make
command has to be run from the base directory.

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

Cleanup sub makefiles

Remove unsused directives which are part of the main makefile, add
indentation and comments to conditions.

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

Make goals configurable with CONFIG_LTTNG

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

Fix: Use generic raw_irqs_disabled_flags

instead of arch_irqs_disabled_flags, this fixes the build for kernel < 2.6.37

See linux commit df9ee29270c11dba7d0fe0b83ce47a4d8e8d2101 for more details.

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

Fix: Use MAX_RT_PRIO offset in sched_wakeup_template

All other sched_* events with prio fields offset the task_struct's
prio value by subtracting MAX_RT_PRIO. This way the value corresponds
to that given by the kernel's task_prio function, and thus with the
prio context field.

This adds the same offset to sched_wakeup, sched_wakeup_new, and
sched_waking, via the sched_wakeup_template event class. This way
their value is comparable to other sched_* events, and the prio
context.

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

Fix: update sched instrumentation for kernel 4.4.0

Due to upstream Linux kernel commit:

commit c73464b1c8434ad4cbfd5369c3e724f3e8ffe5a4
Author: Peter Zijlstra <peterz@infradead.org>
Date:   Mon Sep 28 18:06:56 2015 +0200

    sched/core: Fix trace_sched_switch()

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

Bump version number for development branch

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

Fix: interruptible context: reversed logic

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

Contexts for RT debugging

Add the interruptible, preemptible, need_resched and migratable
contexts.

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

Fix: instrumentation: v4lv2 missing include

Starting at least with 4.3 kernel, we need to specifically include
media/videobuf2-core.h.

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

Fix: header size larger than 256 bytes

When adding large context (e.g. callstack), headers larger than 256
bytes cause discrepancy between calculated size and size written into
the trace buffers. This generates a corrupted trace and triggers a
warning in ring buffer backend, which triggers a safety net disabling
tracing for the current channel.

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

Instrument x86 page faults

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

Fix: irq_vectors TRACE_SYSTEM name

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

Instrument x86_irq_vectors

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

Fix: tracepoint listing misses last loaded probe

The seqfile iteration listing tracepoints operates on the probe list
directly, without going through the lttng_get_probe_list_head()
accessor which ensures that the lazy list of probes is moved to the
actual list of probes.

This causes an issue when loading a probe and then listing the
tracepoints when no tracing sessions are active: the last probe loaded
is missing from the listing.

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

Fix: lttng trace-clock needs to disable preemption

We use a per-cpu data structure, and some contexts (e.g. buffer create)
call this function with preemption enabled. This is the case also for
buffer flush operation.

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

Tracepoint event: add "TP_code_post()"

Useful for probes that need to perform teardown after serialization,
e.g. allocate temporary memory in TP_code_pre() and free it in
TP_code_post(). Be careful to only use memory allocation from contexts
allowing it.

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

Implement array and sequence bitfields

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

Fix: kmem: add mm.h include

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

Add page frame number (pfn) to kmem page events

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

Use pfn rather than struct page in ring buffer array

First step to support permanent memory (e.g. pmem) which does not have
struct page.

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

Implement clock plugin support

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

Command to regenerate the metadata of a session

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

Fix: writeback instrumentation update for 4.3 Linux kernel

Linux upstream commit:

commit 5634cc2aa9aebc77bc862992e7805469dcf83dac
Author: Tejun Heo <tj@kernel.org>
Date:   Tue Aug 18 14:54:56 2015 -0700

    writeback: update writeback tracepoints to report cgroup

Updated the tracepoint signatures for writeback probes. Follow this
change. Don't add extra data (cgroup information) yet.

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

Fix: update sched wakeup instrumentation for 4.3 kernel

Change introduced in upstream Linux:

commit fbd705a0c6184580d0e2fbcbd47a37b6e5822511
Author: Peter Zijlstra <peterz@infradead.org>
Date:   Tue Jun 9 11:13:36 2015 +0200

    sched: Introduce the 'trace_sched_waking' tracepoint

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

Fix: lttng-test build failure for kernels < 3.0

Fixes an undefined reference to kstrtouint_from_user.

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

Use 3.18 syscall instrumentation for MIPS32

Signed-off-by: Lars Persson <larper@axis.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Update the MIPS32 syscall instrumentation to 3.18

Signed-off-by: Lars Persson <larper@axis.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: Move pipe override to the arch specific pointers_override.h

The MIPS architecture has a non-standard calling convention for pipe()
and cannot use the provided override for the pipe event.

Signed-off-by: Lars Persson <larper@axis.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Implement lttng test module

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

Cleanup: remove logically dead code

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

Fix: filter validator use after free

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

Fix: use after free in channel release

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

Fix: non-enabler events should be disabled by default

The session daemon expects that all events (enablers and events) are
disabled when created, so we can attach a filter before enabling them.

Also fix the kretprobe to ensure we can enable/disable both the entry
and return probe (--function instrumentation).

Fixes: #926
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: build failure on 2.6.36

Fixes build failure on the 2.6.36 kernel branch which misses the
vzalloc() function.

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

Cleanup: split bdi_dirty_ratelimit with ifdef into separate macros

Makes the code easier to maintain in the long run.

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

Fix: writeback instrumentation for 4.2 kernels

Bandwidth related fields were moved by this commit:

commit a88a341a73be4ef035ca26170c849f002797da27
Author: Tejun Heo <tj@kernel.org>
Date:   Fri, 22 May 2015 17:13:28 -0400

    writeback: move bandwidth related fields from backing_dev_info into
    bdi_writeback

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

Fix: mm_page_alloc_extfrag instrumentation for kernel 3.18.10

The following commit has been backported from kernel 3.19.2 into
3.18.10, introducing a change to mm_page_alloc_extfrag and therefore
provoking a build failure on a 3.18.10 kernel:

commit a4f3f96f850db69890f47984c60606a2362bee89
Author: Vlastimil Babka <vbabka@suse.cz>
Date:   Wed Feb 11 15:28:15 2015 -0800

    mm: when stealing freepages, also take pages created by splitting buddy page

    commit 99592d598eca62bdbbf62b59941c189176dfc614 upstream.

This patch adds the proper version check to allow building on 3.18.y.

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

Add stream instance id to the packet header

This new field allows the viewer to distinguish between trace files
belonging to the same packet stream (in LTTng: the same CPU in the same
channel).

Signed-off-by: Julien Desfossez <jdesfossez@efficios.com>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Add a packet sequence number

This allows the viewer to identify the gaps between trace packets.

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

Fix: kmem probe with Ubuntu 3.13 kernels

Fixes an undefined reference to get_pageblock_flags_mask.

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

Fix: disable kvm probe if lapic.h isn't found

In a typical distribution if just linux-headers is installed, lapic.h
will not be installed. We should check if that file exists and not build
the module.

Signed-off-by: Chris J Arges <chris.j.arges@canonical.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: timer instrumentation for 4.2 kernels

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

Fix: ext4 instrumentation for 4.2 kernels

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

Fix: kvm instrumentation for 4.2 kernels

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

Fix: allow LTTng to be built within kernel tree

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

Version 2.7.0-rc1

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

Fix: Building the event list fails on fragmented memory

On a small arm imx6 solo with 256MB RAM it often happens that memory
becomes fragmented rather quickly, so that kmalloc will not be able to
get enough consecutive pages (enocuntered for example if you enable all
kernel events: lttng enable-event -k syscall --all).

This patch switches the allocation to vmalloc. Tested for x86 on Ubuntu
12.04 Lts and on imx6 solo 256MB RAM

If this patch is not applied, you can identify low and/or fragmented
memory failures by looking at the kernel ring buffer (please ignore DMA,
it is due to some memory setup misconfiguration, should read Normal):

...
[  321.993820] lttng-sessiond: page allocation failure: order:4, mode:0x1040d0
...
[  321.994711] lowmem_reserve[]: 0 0 0
[  321.994727] DMA: 801*4kB (UEMC) 424*8kB (EMC) 355*16kB (UEMC) 344*32kB (MC) 340*64kB (C) 8*128kB (C) 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB 0*8192kB 0*16384kB 0*32768kB = 46068kB

[ Edit by Mathieu: use vzalloc() rather than vmalloc() + memset 0. ]

Signed-off-by: Martin Leisener <martin@leisener.de>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: use after free on metadata cache reallocation

When the metadata cache is expanded (reallocated) by
lttng_metadata_printf(), the metadata cache reader
(lttng_metadata_output_channel()) may use freed memory, because the
metadata cache is not protected from concurrent read accesses. The
metadata cache updates are protected from each other by the sessions
mutex, but metadata cache reads do not hold the sessions mutex.
Actually, the comment on top of lttng_metadata_output_channel() stating
"We have exclusive access to our metadata buffer (protected by the
sessions_mutex)" is simply wrong, because this mutex is never held when
calling lttng_metadata_output_channel().

Promote the per-stream lock to the metadata cache used by each of those
metadata streams, thus ensuring mutual exclusion between metadata cache
reallocation and readers.

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