ftrace: change implementation to use binary circular buffer

The current implementation of function tracing (CFG_FTRACE_SUPPORT)
produces human-readable text into the output buffer that is passed to

ftrace: change implementation to use binary circular buffer

The current implementation of function tracing (CFG_FTRACE_SUPPORT)
produces human-readable text into the output buffer that is passed to
tee-supplicant and ultimately saved to the Linux filesystem. Two main
issues with that:

1. The string formatting code is somewhat complex. It introduces
significant overhead in the execution time of the instrumented
functions.
2. The various policies about how to handle a buffer full condition
(CFG_FTRACE_BUF_WHEN_FULL) are not very convenient. In particular,
"shift" is typically the most desirable option because it always
keeps the most recent entries, but it is very inefficient to the
point of not being usable in practice.

This commit addresses the above concerns by making the ftrace buffer
circular one, each entry being 64-bit value. The formatting code is
offloaded to a new Python script: scripts/ftrace_format.py. The
output is unchanged except for an added field showing the current
depth in the call stack.

Typical usage (captured on QEMUv8):

build$ mkdir -p ../tmp
build$ chmod a+w ../tmp
build$ make CFG_FTRACE_SUPPORT=y CFG_FTRACE_BUF_SIZE=15000 \
CFG_TA_MCOUNT=y CFG_ULIBS_MCOUNT=y CFG_SYSCALL_FTRACE=y \
QEMU_VIRTFS_AUTOMOUNT=y run
$ xtest regression_1004
...
$ cp /tmp/ftrace-cb3e5ba0-adf1-11e0-998b-0002a5d5c51b.out /mnt/host/tmp
build$ cd ..
optee$ optee_os/scripts/ftrace_format.py \
tmp/ftrace-cb3e5ba0-adf1-11e0-998b-0002a5d5c51b.out |
optee_os/scripts/symbolize.py \
-d optee_os/out/arm/core \
-d out-br/build/optee_test_ext-1.0/ta/*/out | less
TEE load address @ 0x5ab04000
Function graph for TA: cb3e5ba0-adf1-11e0-998b-0002a5d5c51b @ 80085000
| 1 | __ta_entry() {
| 2 | __utee_entry() {
43.840 us | 3 | ta_header_get_session()
7.216 us | 3 | tahead_get_trace_level()
14.480 us | 3 | trace_set_level()
| 3 | malloc_add_pool() {
| 4 | raw_malloc_add_pool() {
46.032 us | 5 | bpool()
| 5 | raw_realloc() {
166.256 us | 6 | bget()
23.056 us | 6 | raw_malloc_return_hook()
267.952 us | 5 | }
398.720 us | 4 | }
426.992 us | 3 | }
| 3 | TEE_GetPropertyAsU32() {
23.600 us | 4 | is_propset_pseudo_handle()
| 4 | __utee_check_instring_annotation() {
26.416 us | 5 | strlen()
| 5 | check_access() {
| 6 | TEE_CheckMemoryAccessRights() {
| 7 | _utee_check_access_rights() {
| 8 | syscall_check_access_rights() {
| 9 | ts_get_current_session() {
4.304 us | 10 | ts_get_current_session_may_fail()
10.976 us | 9 | }
| 9 | to_user_ta_ctx() {
2.496 us | 10 | is_user_ta_ctx()
8.096 us | 9 | }
| 9 | vm_check_access_rights() {
| 10 | vm_buf_is_inside_um_private() {
| 11 | core_is_buffer_inside() {
...

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Acked-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Sumit Garg <sumit.garg@linaro.org>
Acked-by: Etienne Carriere <etienne.carriere@foss.st.com>

show more ...

ldelf: arm: support R_ARM_TLS_DTPMOD32 and R_ARM_TLS_DTPOFF32 relocations

Preparing for C++ support in TAs.

This commit adds support for Thread Local Storage (TLS) relocation
types R_ARM_TLS_DTPMOD

ldelf: arm: support R_ARM_TLS_DTPMOD32 and R_ARM_TLS_DTPOFF32 relocations

Preparing for C++ support in TAs.

This commit adds support for Thread Local Storage (TLS) relocation
types R_ARM_TLS_DTPMOD32 and R_ARM_TLS_DTPOFF32.

OP-TEE does not support multi-threaded TAs so in principle there is no
need to handle the TLS relocations. However, this commit will allow to
run C++ TAs built with the "official" arm-linux-gnueabihf compiler
(which is built with threading support enabled), as long as no
multi-thread feature is explicitly used by the TA. In other words, it
avoids the need to re-build a toolchain with --disable-threads.

Signed-off-by: Jerome Forissier <jerome@forissier.org>
Acked-by: Etienne Carriere <etienne.carriere@linaro.org>
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

ftrace: Add support for syscall function tracer

This patch adds support for syscall tracing in TEE core. It complements
existing ftrace support for user TAs via adding trace for syscalls that
are in

ftrace: Add support for syscall function tracer

This patch adds support for syscall tracing in TEE core. It complements
existing ftrace support for user TAs via adding trace for syscalls that
are invoked by user TAs into the TEE core.

And after this patch ftrace will cover both TA and TEE core code. So lets
rename config option from CFG_TA_FTRACE_SUPPORT to CFG_FTRACE_SUPPORT.

It is optional to enable syscall trace via CFG_SYSCALL_FTRACE=y config
option in addition to CFG_FTRACE_SUPPORT=y config option.

Signed-off-by: Sumit Garg <sumit.garg@linaro.org>
Reviewed-by: Jerome Forissier <jerome@forissier.org>

show more ...

core, ldelf: add support for runtime loading of shared libraries

This commit prepares the introduction of libdl, a dynamic linking
library which will allow TAs to load shared libraries at run time,

core, ldelf: add support for runtime loading of shared libraries

This commit prepares the introduction of libdl, a dynamic linking
library which will allow TAs to load shared libraries at run time,
and resolve symbols on demand. It adds the following function to the
system PTA, inspired from the POSIX dlopen() and dlsym():

- system_dlopen(): takes a UUID and flags. Performs an upcall into
ldelf which then uses the usual system PTA functions to load an map the
requested library into the address space of the calling TA.
- system_dlsym(): takes a UUID and a symbol name. The symbol is
looked up in the library specified by UUID by calling into ldelf. If
UUID is all zeros, all the mapped binaries are searched.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Acked-by: Etienne Carriere <etienne.carriere@linaro.org>

show more ...

ftrace: Add function execution time support

Add support to dump function execution time at corresponding function
exit points in output ftrace buffer.

To achieve proper function execution time we n

ftrace: Add function execution time support

Add support to dump function execution time at corresponding function
exit points in output ftrace buffer.

To achieve proper function execution time we need to exclude TA suspend
time from timestamps, so add corresponding support in TEE core.

Also user mapping must be active to access ftrace buffer, so do that
during TA resume.

Signed-off-by: Sumit Garg <sumit.garg@linaro.org>
Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Tested-by: Jerome Forissier <jerome.forissier@linaro.org> (QEMU)
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

ldelf: ftrace: pass ftrace buffer address to kernel

Enable support to pass ftrace buffer address to kernel during loading
of particular TA. It is required to support function execution time
feature

ldelf: ftrace: pass ftrace buffer address to kernel

Enable support to pass ftrace buffer address to kernel during loading
of particular TA. It is required to support function execution time
feature since kernel needs to update timestamps in ftrace buffer during
TA suspends for proper execution time.

Signed-off-by: Sumit Garg <sumit.garg@linaro.org>
Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

ldelf: fix warning in ftrace_init()

Fixes warning in ftrace_init():

ldelf/ftrace.c: In function ‘ftrace_init’:
ldelf/ftrace.c:47:9: warning: cast to pointer from integer of different size [-Wint-to

ldelf: fix warning in ftrace_init()

Fixes warning in ftrace_init():

ldelf/ftrace.c: In function ‘ftrace_init’:
ldelf/ftrace.c:47:9: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]
fbuf = (struct ftrace_buf *)finfo->buf_start.ptr64;
^
Fixes: c96d7091b566 ("ftrace: Enable support for 32 bit apps")
Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

ftrace: Enable support for 32 bit apps

By default 32 bit trusted applications are compiled in thumb mode but
thumb mode doesn't support function graph tracing due to missing frame
pointer support re

ftrace: Enable support for 32 bit apps

By default 32 bit trusted applications are compiled in thumb mode but
thumb mode doesn't support function graph tracing due to missing frame
pointer support required to trace function call chain. So rather compile
trusted applications in ARM mode in case function tracing is enabled.

Signed-off-by: Sumit Garg <sumit.garg@linaro.org>
Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>

show more ...

ldelf: support TA ftrace

Adds support in ldelf to dump ftrace data.

Reviewed-by: Sumit Garg <sumit.garg@linaro.org>
Acked-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: Jens Wikl

ldelf: support TA ftrace

Adds support in ldelf to dump ftrace data.

Reviewed-by: Sumit Garg <sumit.garg@linaro.org>
Acked-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...