fix: remove 'src/' from modules install path

The move of the modules sources in the 'src/' subdirectory had the
unexpected side effect of adding 'src/' to the module install path
inside '/lib/modules/$(uname -r)/extra'. Adjust the Makefiles to return
to the initial behavior.

Change-Id: I280fe8a1e5ef367edeec0a6351351a233cc6f0be
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ifcc64ea95c0e1435567150d976d42e3add3f7523

fix: Adjust ranges for Ubuntu 5.4.0-67 kernel

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ifa0f50ffdc946d80b67bb5ae7ca4b0aa152e825b

fix: block: add a disk_uevent helper (v5.12)

See upstream commit:

  commit bc359d03c7ec1bf3b86d03bafaf6bbb21e6414fd
  Author: Christoph Hellwig <hch@lst.de>
  Date:   Sun Jan 24 11:02:39 2021 +0100

    block: add a disk_uevent helper

    Add a helper to call kobject_uevent for the disk and all partitions, and
    unexport the disk_part_iter_* helpers that are now only used in the core
    block code.

Change-Id: If6e8797049642ab382d5699660ee1dd734e92c90
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Fix: perf counters: uninitialized field

Reported by Coverity:
** CID 1452572:  Uninitialized variables  (UNINIT)
>>>     Using uninitialized value "ctx_field". Field "ctx_field.get_size_arg" is uninitialized when calling "lttng_kernel_context_append".

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I064260c6fe78af24768bb141fde2864d2abbe3a5

Cleanup: implement dedicated file operations for events and enablers

In order to simplify the code in preparation for refactoring of the
event structures, remove the "event type" field and implement specific
file operations for events and enablers.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I508bdb46fa3747112e7cfef4a62278fb566f842a

Namespace LTTng modules ABI with lttng_kernel_abi_ prefix

Standardize on lttng_kernel_abi_ prefix for the lttng modules ABI to
eliminate namespace clash between lttng-modules core code and its ABI.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I3b1c8d6f46702e57defa17d683ede576466ba3ce

Refactoring: type description structures

Refactoring:

- Type description structures: new API based on structure inheritance
  rather than union,
- Rely on compound literals,
- Rename description fields: name -> event_name,
- Rename types (lttng_kernel_ namespace):
  - struct lttng_kernel_event_desc,
  - struct lttng_kernel_event_field,
  - struct lttng_kernel_ctx.
- Move some internal definitions (not used by probes) to
  events-internal.h.
- Streamline context registration API,
- Introduce the LTTNG_CREATE_FIELD_METADATA define for the tracepoint
  code generation pass which generates the type descriptions, allowing
  to describe structures as an array of fields, and use enumerations
  from those fields.
- Adapt all tracepoint instrumentation to those changes.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I9a968a62888c52438a1f62ec24fcced4c3fd6ddf

Prefix lttng_enum_desc with lttng_kernel_

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I25b7160875c99c99c18d1e559b09698b5411dae7

Prefix lttng_enum_value and lttng_enum_entry with lttng_kernel_

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I4b4d502e2af9a1ca6425889da2697c5633200f52

Introduce internal event header

This header is to be included by the tracer, but not by tracepoint
probes.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ic9501bb58c4156d3b037b4f15251d03e25cd45f9

Cleanup: namespace string encoding

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ic7265a6617feaacfe621d895cd0bc8ce2b50d912

Cleanup: Rename abstract types to lttng_kernel_type

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Iac1810c9d9830d80109fe72b1965b7b7cf663207

Cleanup: event_notifier -> notify in comments

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Icdb4282e8b5fce231f96c56b4568555bd7a442fb

Fix: event notifier: add missing parameters validation

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Idbbbe369f64c14e7ac8914b4e7eb8e3ef75b1fb5

msgpack: use KERNEL namespace for header protection

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I8a3ee5bd935442dc5d3fcaff97859406279eb73d

Cleanup: lttng_abi_create_event{,_notifier}: use switch/case rather than if

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I2a5454521d57fdfcb0c623031b24d590aff40281

Add entryexit field to kretprobes ABI

At this stage, only "entryexit = 0" is supported. No behavior change
except for validating that ENTRY and EXIT are unsupported (-EINVAL).

It is possible to extend struct lttng_kernel_kretprobe because it is
always used within struct lttng_kernel_event which contains enough
zero-initialized padding due to char padding[LTTNG_KERNEL_EVENT_PADDING2].

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I96d86eebff66da4af186ef51c2ff13e2ed84a510

Add missing fallthrough comments

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I225a5236c108e5b0d4bd39b5f1ef6988fe8b0096

No-op instrumentation is unsupported, cleanup fallthrough comments

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Id76238ead464dda1b40c6006f26d76c8a1049698

kretprobes: rename "return" to "exit"

The kretprobes instrumentation emits <func>_entry and <func>_return.
However, pairing "entry" and "return" is not semantically right. The
term "entry" should be paired with "exit" (similarly to system call
intrumentation), since "return" should be paired with "call".

If we look at what is actually instrumented by kretprobes, we indeed
instrument the entry and exit points of a function.

Therefore, change the "<func>_return" event for "<func>_exit".

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ia541253cbb65a5afc8aaed2a6cfa77d9b296c96c

compiler warning cleanup: is_signed_type: compare -1 to 1

Comparing -1 to 0 triggers compiler warnings (gcc -Wtype-limits and
-Wbool-compare) and Coverity warning "Macro compares unsigned to 0".

Comparing -1 to 1 instead takes care of silencing those warnings while
keeping the same behavior.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Id42a51759a1c7c669e63588c05f9d4485304c541

Fix: bytecode linker: validate event and field array/sequence encoding

The bytecode linker should only allow linking filter expressions loading
fields which are string-encoded arrays and sequence for comparison
against a string, and reject arrays and sequences without encoding, so
the filter interpreter does not attempt to load non-NULL terminated
arrays/sequences as if they were strings.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ia5f33ed036e310d66aee6b682ef0a17eb5b99982

Fix: racy notifier captures update vs traversal

For captures, a new struct lttng_kernel_notification_ctx is introduced,
which is to be used as additional "context" to the notification_send()
callback. This allows passing the "eval_capture" state from the probe
to the notification callback, and fixes a bug where a sequence of:

- create notification enabler,
- enable notification enabler,
- add capture to enabler,

where a tracepoint runs concurrently with add capture happens to do a
first capture list_empty check which skips the stack preparation,
whereas the second capture list_empty check within the notification
callback finds a capture entry, and thus attempts to use an
uninitialized stack. The notification callback is also modified to use
an RCU-aware list traversal.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ib7ab673fb05b4824d0c7ba2e163dc1e88b465ee9

Fix: kretprobe: null ptr deref on session destroy

The `filter_bytecode_runtime_head` list is currently not initialized for
the return event of the kretprobe. This caused a kernel null ptr
dereference when destroying a session. It can reproduced with the
following commands:

  lttng create
  lttng enable-event -k --function=lttng_test_filter_event_write my_event
  lttng start
  lttng stop
  lttng destroy

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1162ce8b10dd7237a26331531f048346b984eee7

Fix: bytecode linker: iteration on wrong list head

lttng_enabler_link_bytecode() calls link_bytecode() passing an insertion
location (insert_loc) within the list. This insert location is meant to
be used as cursor position where to add the new element.

However, bytecode_is_linked() uses it as iteration list head, and this
is where things fall apart: it will thus consider the real list head as
being a list node, and will erroneously think that it is contained
within a struct lttng_bytecode_runtime, and thus try to perform possibly
out-of-bound read or read garbage data for the comparison.

It worked fine most of the time because in usual scenarios the insert
location is the list head. It falls apart when many bytecodes are linked
to a given event.

Fixes: 2dfda770cc6 ("Decouple `struct lttng_event` from filter code")
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I7463c7a9399b8f7f7d0e3d74e6427aae46cf56ff

counters: add coalesce_hits to ABI

This will be required by an upcoming feature in a subsequent release, so
plan ahead with a coalesce_hits field. It is currently unused by
lttng-modules.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I0dd3a5c00a89c59111b723db8e39390a16764133

fix: mm, tracing: record slab name for kmem_cache_free() (v5.12)

See upstream commit:

  commit 3544de8ee6e4817278b15fe08658de49abf58954
  Author: Jacob Wen <jian.w.wen@oracle.com>
  Date:   Wed Feb 24 12:00:55 2021 -0800

    mm, tracing: record slab name for kmem_cache_free()

    Currently, a trace record generated by the RCU core is as below.

    ... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=00000000f3b49a66

    It doesn't tell us what the RCU core has freed.

    This patch adds the slab name to trace_kmem_cache_free().
    The new format is as follows.

    ... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=0000000037f79c8d name=dentry
    ... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=00000000f78cb7b5 name=sock_inode_cache
    ... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=0000000018768985 name=pool_workqueue
    ... kmem_cache_free: call_site=rcu_core+0x1fd/0x610 ptr=000000006a6cb484 name=radix_tree_node

    We can use it to understand what the RCU core is going to free. For
    example, some users maybe interested in when the RCU core starts
    freeing reclaimable slabs like dentry to reduce memory pressure.

Link: https://lkml.kernel.org/r/20201216072804.8838-1-jian.w.wen@oracle.com
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1ee2fc476614cadcc8d3ac5d8feddc7910e1aa3a

Fix: filter interpreter early-exits on uninitialized value

I observed that syscall filtering on string arguments wouldn't work on
my development machines, both running 5.11.2-arch1-1 (Arch Linux).

For instance, enabling the tracing of the `openat()` syscall with the
'filename == "/proc/cpuinfo"' filter would not produce events even
though matching events were present in another session that had no
filtering active. The same problem occurred with `execve()`.

I tried a couple of kernel versions before (5.11.1 and 5.10.13, if
memory serves me well) and I had the same problem. Meanwhile, I couldn't
reproduce the problem on various Debian machines (the LTTng CI) nor on a
fresh Ubuntu 20.04 with both the stock kernel and with an updated 5.11.2
kernel.

I built the lttng-modules with the interpreter debugging printout and
saw the following warning:
  LTTng: [debug bytecode in /home/jgalar/EfficiOS/src/lttng-modules/src/lttng-bytecode-interpreter.c:bytecode_interpret@1508] Bytecode warning: loading a NULL string.

After a shedload (yes, a _shed_load) of digging, I figured that the
problem was hidden in plain sight near that logging statement.

In the `BYTECODE_OP_LOAD_FIELD_REF_USER_STRING` operation, the 'ax'
register's 'user_str' is initialized with the stack value (the user
space string's address in our case). However, a NULL check is performed
against the register's 'str' member.

I initialy suspected that both members would be part of the same union
and alias each-other, but they are actually contiguous in a structure.

On the unaffected machines, I could confirm that the `str` member was
uninitialized to a non-zero value causing the condition to evaluate to
false.

Francis Deslauriers reproduced the problem by initializing the
interpreter stack to zero.

I am unsure of the exact kernel configuration option that reveals this
issue on Arch Linux, but my kernel has the following option enabled:

CONFIG_GCC_PLUGIN_STRUCTLEAK_BYREF_ALL:
   Zero-initialize any stack variables that may be passed by reference
   and had not already been explicitly initialized. This is intended to
   eliminate all classes of uninitialized stack variable exploits and
   information exposures.

I have not tried to build without this enabled as, anyhow, this seems
to be a legitimate issue.

I have spotted what appears to be an identical problem in
`BYTECODE_OP_LOAD_FIELD_REF_USER_SEQUENCE` and corrected it. However,
I have not exercised that code path.

The commit that introduced this problem is 5b4ad89.

The debug print-out of the `BYTECODE_OP_LOAD_FIELD_REF_USER_STRING`
operation is modified to print the user string (truncated to 31 chars).

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I2da3c31b9e3ce0e1b164cf3d2711c0893cbec273

Fix: memory leaks on event notifier destroy

Both filter runtime and event enabler ref objects are owned by the
event notifier, but are not freed upon destruction of the event notifier
object, thus leaking memory.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I511569f56a38f670549a93cb6179b77861245712

Fix: memory leaks on event destroy

Both filter runtime and event enabler ref objects are owned by the
event, but are not freed upon destruction of the event object, thus
leaking memory.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ice9b1c18b47584838aea2b965494d3c8391f4c84

fix: Adjust ranges for Ubuntu 5.8.0-44 kernel

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I419904dc9da316b38c2c16a08b6c17625b19b305

Fix: do not use bdi_unknown_name symbol

Use the GPL-exported bdi_dev_name introduced in kernel 5.7. Do not use
static inline bdi_dev_name in prior kernels because it uses the bdi_unknown_name
symbol which is not exported to GPL modules.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I8b4e4fd84ecacef7942b308e615ca88db8dce7b6

fix: memcg: fix a crash in wb_workfn when a device disappears (5.6)

See upstream commit:

commit 68f23b89067fdf187763e75a56087550624fdbee
("memcg: fix a crash in wb_workfn when a device disappears")

It is currently backported into stable branches 5.4 and 5.5, but appears
to be missing from the 4.4, 4.9, 4.14, 4.19 LTS branches.

Implement our own lttng_bdi_dev_name wrapper to provide this fix on
builds against stable kernels which do not have this fix.

There is one user-visible change with this commit: for builds against
kernels < 4.4.0, the writeback_work_class events did use the
default_backing_dev_info to handle cases where the device is NULL,
writing "default" into the trace. This behavior is now aligned to
match what is done in kernels >= 4.4.0, which is to write "(unknown)"
into the name field.

Link: https://lore.kernel.org/r/537870616.15400.1612973059419.JavaMail.zimbra@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I0823643aa2f9d4c2b9f2005748a2adfd4457979a

Fix: writeback: out-of-bound reads

Use ctf_string rather than ctf_array_text for name fields, because the
source strings are not guaranteed to be at least 32 bytes.

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

fix: Add one digit to RHEL major release version

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I4e8b10ee985db59f7795e026486b3d47b76b5728

fix: Add one digit to SLES minor release version

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ia4b67b38377bdd997b754b92b09b96934940a013

fix: RT_PATCH_VERSION is close to overflow

We allocated only 8bits for RT_PATCH_VERSION in LTTNG_RT_VERSION_CODE,
the current RT patch version for the 4.4 branch is currently 214 which
is getting close to 256. Bump it to 16bits to avoid breakage in the
future.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I0666bbe996854696ac98e025eb02e5fced0540b1

fix: cast LTTNG_KERNEL_VERSION/LTTNG_LINUX_VERSION_CODE to uint64_t

Cast our version macros to an unsigned 64bits value to prevent
overflowing when we append distro specific version information.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ia42a5dc0dfddf64515aea144283af5cc0c3b97e0

fix: double defined LTTNG_KERNEL_VERSION

Cleanup a rebase error which resulted in double defined macros.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Id9ddce968c0d142c2c1013e0f51f649caa23ee4c

fix: UTS_UBUNTU_RELEASE_ABI is close to overflow

We allocated only 8bits for UTS_UBUNTU_RELEASE_ABI in
LTTNG_UBUNTU_KERNEL_VERSION, the current Xenial kernel has an ABI of 207
which is getting close to 256. Bump it to 16bits to avoid breakage in
the future.

Change-Id: Iee99757bb28cdd958b044b31df3232b9f8816873
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

fix: sublevel version overflow in LINUX_VERSION_CODE

The 4.4.256 and 4.9.256 stable release overflow the 8bits allocated to
the sublevel in LINUX_VERSION_CODE which ends means they report
themselves as 4.5.0 and 4.10.0 respectively. The next releases in these
stables branches will have sublevel clamped at 255 and will thus report
themselves as 4.4.255 and 4.9.255 for all subsequent releases.

We need a way to way to properly detect these release since I doubt they
will stop breaking tracepoints declarations. As a workaround, extract
the version information from the Makefile in the kernel headers and use
this information to generate a version code when the sublevel is equal
or greater than 256.

Change-Id: I96ae9f22c0c1ba8c619643946a5311c767fbcf8c
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

Namespace kernel version macros

This patch replaces all uses of the LINUX_VERSION_CODE and
KERNEL_VERSION macros by an 'LTTNG_' prefixed version, this will allow
us to override them.

Change-Id: If8fab7fe341dd35e2211d30427519a48dfb708f9
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

aarch64: blacklist gcc prior to 5.1

Linux aarch64 requires GCC 5.1 or better because prior versions perform
unsafe access to deallocated stack.

Some Linux distributions may have backported the fix, but it was never
released into earlier upstream gcc versions.

Link: https://lwn.net/Articles/842122/
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63293
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I72993e446f7f54f39d0f360273b68f194be8c13a

fix: missing include for 'task_struct' in fdtable.h

In some kernel versions, linux/fdtable.h dereferences a pointer in a
forward declared 'struct task_struct' without an include of 'linux/sched.h'.

Add this missing include to the wrapper.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I06a6a9f7efd456843d1e4fdcc9003b839d7f3329

Fix: counter-api: always inline counter add function

The counter add function uses cmpxchg() and cmpxchg_local() on 1, 2, 4,
and 8 bytes types.

In libcounter, the 8 bytes type is only supported on 64-bit
architectures, but the 1, 2, 4 byte type code is present for all
architectures, even though only the 4 byte code is currently used by
lttng-modules.

The ARM implementation of cmpxchg uses the "__bad_cmpxchg" linker error
to report use of cmpxchg on an unsupported size.

Considering that "inline" does not strictly mean always inline (depends
on CONFIG_OPTIMIZE_INLINING on some kernels, and does not mean forced
inlining in recent kernels), the compiler is free to generate a function
rather than perform inlining. If that happens, then the __bad_cmpxchg
linker error is generated even if the 1 and 2 bytes types are unused.

Therefore, use __always_inline for functions in counter-api.h to force
inlining, and therefore removal of unused code before linking, which is
required by this Linux kernel __bad_cmpxchg linker error trick.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1adccd1382e71abc5880e0351d976b779245468a

fix: genirq: Restrict export of irq_to_desc() (v5.11)

See upstream commit:

  commit 64a1b95bb9fe3ec76e1a2cd803eff06389341ae4
  Author: Thomas Gleixner <tglx@linutronix.de>
  Date:   Thu Dec 10 20:26:06 2020 +0100

    genirq: Restrict export of irq_to_desc()

    No more (ab)use in drivers finally. There is still the modular build of
    PPC/KVM which needs it, so restrict it to this case which still makes it
    unavailable for most drivers.

Link: https://lore.kernel.org/r/20201210194045.551428291@linutronix.de
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ie5a2af2f02ade07c73e1c7a8aa0fb155280b3d8b

fix: block: merge struct block_device and struct hd_struct (v5.11)

See upstream commit :

  commit 0d02129e76edf91cf04fabf1efbc3a9a1f1d729a
  Author: Christoph Hellwig <hch@lst.de>
  Date:   Fri Nov 27 16:43:51 2020 +0100

    block: merge struct block_device and struct hd_struct

    Instead of having two structures that represent each block device with
    different life time rules, merge them into a single one.  This also
    greatly simplifies the reference counting rules, as we can use the inode
    reference count as the main reference count for the new struct
    block_device, with the device model reference front ending it for device
    model interaction.

Change-Id: I47702d1867fda0d8fc0754d761aa4d1ae702cdeb
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

fix: kprobes: Remove kretprobe hash (v5.11)

See upstream commit:

  commit d741bf41d7c7db4898bacfcb020353cddc032fd8
  Author: Peter Zijlstra <peterz@infradead.org>
  Date:   Sat Aug 29 22:03:24 2020 +0900

    kprobes: Remove kretprobe hash

    The kretprobe hash is mostly superfluous, replace it with a per-task
    variable.

    This gets rid of the task hash and it's related locking.

    Note that this may change the kprobes module-exported API for kretprobe
    handlers. If any out-of-tree kretprobe user uses ri->rp, use
    get_kretprobe(ri) instead.

Link: https://lore.kernel.org/r/159870620431.1229682.16325792502413731312.stgit@devnote2
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I855765f390ad7caf481ef5fea334645e852f5b0f

fix: file: Rename fcheck lookup_fd_rcu (v5.11)

See upstream commit:

  commit 460b4f812a9d473d4b39d87d37844f9fc30a9eb3
  Author: Eric W. Biederman <ebiederm@xmission.com>
  Date:   Fri Nov 20 17:14:27 2020 -0600

    file: Rename fcheck lookup_fd_rcu

    Also remove the confusing comment about checking if a fd exists.  I
    could not find one instance in the entire kernel that still matches
    the description or the reason for the name fcheck.

    The need for better names became apparent in the last round of
    discussion of this set of changes[1].

    [1] https://lkml.kernel.org/r/CAHk-=wj8BQbgJFLa+J0e=iT-1qpmCRTbPAJ8gd6MJQ=kbRPqyQ@mail.gmail.com

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I33d2c912ebfdf82a68e506c9b6be17c51d2f254e

fix: block: remove the request_queue argument to the block_bio_remap tracepoint (v5.11)

See upstream commit:

  commit 1c02fca620f7273b597591065d366e2cca948d8f
  Author: Christoph Hellwig <hch@lst.de>
  Date:   Thu Dec 3 17:21:38 2020 +0100

    block: remove the request_queue argument to the block_bio_remap tracepoint

    The request_queue can trivially be derived from the bio.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ic49a9c9ebeea37e4ec79c736382293a6c9ce86d3

fix: block: remove the request_queue argument to the block_split tracepoint (v5.11)

See upstream commit:

  commit eb6f7f7cd3af0f67ce57b21fab1bc64beb643581
  Author: Christoph Hellwig <hch@lst.de>
  Date:   Thu Dec 3 17:21:37 2020 +0100

    block: remove the request_queue argument to the block_split tracepoint

    The request_queue can trivially be derived from the bio.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I93e8c4c51ba36d22b587841e95ff4be8d5224230

fix: block: simplify and extend the block_bio_merge tracepoint class (v5.11)

See upstream commit:

  commit e8a676d61c07eccfcd9d6fddfe4dcb630651c29a
  Author: Christoph Hellwig <hch@lst.de>
  Date:   Thu Dec 3 17:21:36 2020 +0100

    block: simplify and extend the block_bio_merge tracepoint class

    The block_bio_merge tracepoint class can be reused for most bio-based
    tracepoints.  For that it just needs to lose the superfluous q and rq
    parameters.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I90a1da20ab07605ed88f29b02f63134fa4aee6a8

fix: block: remove the request_queue to argument request based tracepoints (v5.11)

See upstream commit :

  commit a54895fa057c67700270777f7661d8d3c7fda88a
  Author: Christoph Hellwig <hch@lst.de>
  Date:   Thu Dec 3 17:21:39 2020 +0100

    block: remove the request_queue to argument request based tracepoints

    The request_queue can trivially be derived from the request.

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ibe4a13f06ed57955fa3e0b77b87a44b7e6b57775

fix: remove floating-point arguments in msgpack

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I6366ac303157e4ab66045b09c6f03aadee0717c8

abi: use 0, 1 for counter enumerations

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

Cleanup: use tabs in abi.h

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

Bump LTTNG_MODULES_ABI_MINOR_VERSION to 6

New operations were added to the lttng-modules ABI as part of the 2.13
release cycle to support event notifiers.

This will allow future LTTng-tools versions to check for those
capabilities.

Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Idfd58f3914b87be3688a10a06239a1b0ab556a7c

Fix: handle default switch case with -EINVAL in lttng-syscalls

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I7768ea3d0da259976548f6b9417350227390b3be

fix: use wrapper for hlist_for_each_entry_rcu

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I4a0fe4abea70a4cae605cebc3c51fe5c47b24715

fix: asm/barrier.h was introduced in v3.4

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1aa81452fd4c08c151f0f87cfa1b24018ad4fc30

Fix: handle default switch case with -EINVAL in __lttng_counter_add

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

Introduce lttng smp_store_release smp_load_acquire wrappers

Kernels prior to 3.14 do not implement smp_store_release nor
smp_load_acquire. Implement our own wrappers with smp_mb instead for
those older kernels.

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

fix: include byteorder.h for 'cpu_to_*'

Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ie9da5b1d17b7311ac0d6898c3d23a6c4d592927a

Fix: notifier: use store-release/load-acquire for error counter

The "record_error" operation is executed concurrently with setting the
error counter in the notifier group without locking, so we need to
explicitly provide existance guarantees.

The only visible transition is from NULL -> !NULL, because the only
situation reverting the error counter back to NULL is on destruction of
the notification group, after an RCU synchronisation guarantees that no
record_error can observe this pointer anymore.

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

Fix: event-notifier: Groups may not have an error counter

Issue
=====
If no error counter is attached to an event notifier group, calling
the `record_error()` function leads to a NULL pointer dereference.

Solution
========
Check if the error_counter field is allocated before dereferencing it.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I7ba68f9ae5b7b66cea538cf030bc5e1287d6eace

Bump LTTNG_KERNEL_EVENT_NOTIFIER_PADDING to 32 bytes

Allows using 3 more 64-bit integers, and one extension pointer.

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

Fix: remove dead code in msgpack.c

Identified by Coverity. Comparing unsigned types < 0 is never true.

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

Fix: counter: cast UINT*_MAX to 64-bit signed type before negative

(long long)-UINT32_MAX leads to value 1 which is not what we expect.
This is due to implicit type promotion from unsigned to signed 32-bit
integer.

Apply this to 8-bit and 16-bit types as well even though they are
not affected by this issue to keep things regular.

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

Fix: include limits.h wrapper from libcounter

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

Fix: include kernel.h in wrapper/limits.h

linux/kernel.h is present in older kernels, but linux/limits.h is not.

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

Introduce limits wrapper

Older kernels (v3.13.11 and older on the CI) miss the U32_MAX
macro. Implement them in a wrapper.

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

Introduce lazy system call event creation

Introduce lazy system call event creation. If only specific system call
events are enabled, only create the metadata for those.

Align implementation of system call event probes for recording events
with their notifier counterpart.

This is part of the background work required for a future feature (trace
hit counters and map support) which brings benefits on its own.

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

Implement event notifier error counter

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I566d34ab92aaa642f50407d913b16c4a6b8479a2

Implement lib counter

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I47424817ca874357d5ae4349a048cce889988a76

capture: Replace FIXME by explanation of the limit

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

Implement capturing payload on event notifier

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ibdaa07ecfe36ecd1c340942821503ea4cb494e2c

Fix: comment related to filter bytecode list

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

Fix: use vmalloc for filter bytecode allocation

Given that the filter bytecode can reach 64kB, use vmalloc rather than
kmalloc to minimize the chances of failure on systems with fragmented
memory.

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

Implement enum and sequence capture serialization functions

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I2a342e5e1d74e7c2f676b21822d8b7f171d68ebd

bytecode: Add `OBJECT_TYPE_{UN,}SIGNED_ENUM` type

Enumerations are currently converted to their integer counterparts as
soon as they are encountered.

In order to use them in captures, we need to differentiate the
enumerations from integers for the entirety of the interpretation.

This commit adds the `OBJECT_TYPE_SIGNED_ENUM` and
`OBJECT_TYPE_UNSIGNED_ENUM` to keep track of these objects

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ic0ab518588d7250190e42576d9baba2e8d8ce94a

Generalize `lttng_enabler_link_bytecode()` bytecode list

So it can be used for both filter and capture bytecode runtimes

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I14dd5eb1eb018d358446d68ba14aa376d1647545

Cleanup: rename `_lttng_filter_link_bytecode()` -> `link_bytecode()`

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I8a9b5d97eae502181ae8673caf39e4f54fdbdeb6

Add `interpreter_funcs` to `lttng_bytecode_runtime`

Background
==========
The current filter interpreter function signature looks like this:

uint64_t lttng_bytecode_filter_interpret(void *filter_data,
              const char *filter_stack_data);

The upcoming capture interpreter function will need an output parameter
to extract the top of stack register. It will look like this:

uint64_t lttng_bytecode_capture_interpret(void *filter_data,
              const char *filter_stack_data, struct output_register *output);

Problems
========
We can't reuse the same function pointer field in `struct
lttng_bytecode_runtime` as both interpreter functions will have
different signatures.

We can't change the signature of this existing filter function because
it's used in the tracepoint probes.

Solution
========
Add a union of callbacks to hold both interpreter functions. This also
doesn't change the layout of the `struct lttng_bytecode_runtime`
objects.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: Ic4423100e1f97654fe43c0927d3b81de2d1d724f

bytecode: rename `lttng_filter_sync_state()` -> `lttng_bytecode_filter_sync_state()`

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I2d63001ed633236a98eba6398ad06fccf83fde2c

Rename filter bytecode types and files

File renames:
- filter-bytecode.h -> bytecode.h
- lttng-filter-interpreter.c -> lttng-bytecode-interpreter.c
- lttng-filter-specialize.c -> lttng-bytecode-specialize.c
- lttng-filter-validator.c -> lttng-bytecode-validator.c
- lttng-filter.c -> lttng-bytecode.c
- lttng-filter.h -> lttng-bytecode.h

Function renames:
- `lttng_filter_interpret_bytecode_false()` -> `lttng_bytecode_filter_interpret_false()`
- `lttng_filter_interpret_bytecode()` -> `lttng_bytecode_filter_interpret()`
- `lttng_filter_specialize_bytecode()` -> `lttng_bytecode_specialize()`
- `lttng_filter_validate_bytecode()` -> `lttng_bytecode_validate()`

Type renames
- `filter_opcode_t` to `bytecode_opcode_t`

Enum renames:
- `enum filter_op` to `enum bytecode_op`
- `FILTER_OP_*` to `BYTECODE_OP_*`
- `enum lttng_filter_ret` -> `enum lttng_bytecode_interpreter_ret`
- `LTTNG_FILTER_DISCARD` -> `LTTNG_INTERPRETER_DISCARD`
- `LTTNG_FILTER_RECORD_FLAG` -> `LTTNG_INTERPRETER_RECORD_FLAG`

Define renames:
- `FILTER_STACK_EMPTY` -> `INTERPRETER_STACK_EMPTY`
- `FILTER_STACK_LEN`-> `INTERPRETER_STACK_LEN`
- `FILTER_MAX_DATA_LEN` -> `BYTECODE_MAX_DATA_LEN`

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I8cd3a15c8ba19aa1d66e611913a1b08e51d94a43

bytecode: generalize `struct lttng_filter_bytecode_node`

Rename `struct lttng_filter_bytecode_node` to `struct
lttng_bytecode_node` so it can be used by capture bytecode as well.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I27162522bb20d0fdce6af551fbd982a791d1067c

Add msgpack implementation for serializing captures

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I7d9cfd4906c5c047cfb4fc9467b293c4e895523d

bytecode: allow interpreter to return any type

The bytecode interpreter when used by capture bytecode needs to return
types other than an integer or dynamic type.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I099749183fbd0622f258f9c38e37fdb167493a0b

bytecode: propagate `rev_bo` of element

When specializing and executing bytecode.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I86aea5b5733f92c56564c6352bd6dcb85f6a2d30

bytecode: set register type to `REG_PTR` even if not used

There was no need to set the field when using filter as the next
instruction would assume that the top of stack is a `REG_PTR`.

With the upcoming capture feature, we need to ensure this field is
consistent for extraction.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I9c60416dd452949e584fadd70b15cdc3d402aa46

Add `lttng_bytecode_interpret_format_output()` for top of stack extraction

This new static function will be used to extract the register on the top of
stack after the execution of the bytecode. This is currently not used by the
filter bytecode but will be used by capture bytecode.

The returned value is saved in a tagged union struct named `struct
lttng_interpreter_output` and can be used by the caller of the interpreter
function.

Typically, this struct will be allocated on the stack to avoid dynamic
allocation inside the tracepoint probes.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1cfd3ab6e84b7e308c48ed7a8a9555a3e338eea7

bytecode: add `REG_U64` interpreter register type

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I53c12a54cbd416617834982bbd2b7cf528d41a76

Fix: filter validator: refuse string and star glob input to bitwise operation

The validator refuses input ax=string,bx=unknown, but accepts input
ax=unknown,bx=string. Both inputs should be refused.

The same goes for the error glob input.

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

Fix: bytecode: Validate register type for instructions expecting unknown type

The bytecode validator allows unknown type as input for some
instructions which are not specialized. The interpreter therefore needs
to check the register type for their input.

Thie requires that every instruction in the interpreter sets the
register type of the output it populates (unless it is unchanged).

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I3339c36340645937b801f6bf6dbf517d06416a14

Cleanup: Rename filter functions/fields to mention "filter"

This will be cleaner when we introduce the capture bytecode functions
and fields.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I04aca5bfd31f2526b24fe3a4b2e8f2b1c1b482f9

Implement event notifiers for syscalls

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I31e60d0d758b93ada11d208f583d71f05168c014

Fix: syscalls: address of statically allocated element never null

This check is intended to confirm that the table element for that syscall
is indeed populated but checked that the element is NULL. This was never
the case because the address of an element of a statically allocated
array cannot be NULL.

Fix this by check if the function pointer is NULL instead. This means
that the element is not populated.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1d769d6609fa4517199f022e1a262c4494c8f63a

Rename LTTNG_SYSCALL_MATCH_ -> LTTNG_KERNEL_SYSCALL_MATCH_

This is done to keep the same name scheme used to all ABI enums.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I8e1010ab21a47b7f1e519df498acd230315cdc26

Allow LTTNG_KERNEL_SYSCALL_{ENTRY, EXIT}

Signed-off-by: Francis Deslauriers <fdeslaur@gmail.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I1ea097797da5db474f2f33d779f66254b7979c46

syscalls: extract `lttng_syscall_filter_enable()` for reuse

The syscall event notifiers will reuse the concept of syscall filtering
to avoid needlessly preparing arguments for disabled syscalls.

Signed-off-by: Francis Deslauriers <francis.deslauriers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Change-Id: I64395a031e526e8485e10b4b72f653058c8d0a38