stm32mp1: util for shared resources refcount

Implement {incr|decr}_shrefcnt(refcount, secure_flag) to provide
reference counting for secure, non secure and resources used both
from secure and non se

stm32mp1: util for shared resources refcount

Implement {incr|decr}_shrefcnt(refcount, secure_flag) to provide
reference counting for secure, non secure and resources used both
from secure and non secure world.

Functions {incr|decr}_refcnt(refcount) is a secure refcount only
reference counting.

Signed-off-by: Etienne Carriere <etienne.carriere@st.com>
Acked-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

core: arm: link tee.elf with lib archives

Links tee.elf with the library archives instead of -llibname in order to
detect multiply defined symbols in several libraries.

Acked-by: Jerome Forissier <

core: arm: link tee.elf with lib archives

Links tee.elf with the library archives instead of -llibname in order to
detect multiply defined symbols in several libraries.

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

show more ...

core: user_ta: implement ASLR for TAs

Introduces CFG_TA_ASLR to enable Address Space Layout Randomization of
Trusted Applications. ASLR makes the exploitation of memory corruption
vulnerabilities ha

core: user_ta: implement ASLR for TAs

Introduces CFG_TA_ASLR to enable Address Space Layout Randomization of
Trusted Applications. ASLR makes the exploitation of memory corruption
vulnerabilities harder.
The feature is disabled by default except for the configurations I
could test (QEMU and HiKey960).
When CFG_TA_ASLR=y, the stack and subsequent ELF file(s) needed by the
TA are mapped into the user VA space with a random offset comprised
between CFG_TA_ASLR_MIN_OFFSET_PAGES and CFG_TA_ASLR_MAX_OFFSET_PAGES
pages (that is between 0 and 128 pages by default).

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Tested-by: Jerome Forissier <jerome.forissier@linaro.org> (QEMU, HiKey960)
Reviewed-by: Etienne Carriere <etienne.carriere@linaro.org>
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

core: user_ta: use consistent formatting for addresses in TA dump

Improve the layout of the TA dump message by using fixed width for
physical and virtual addresses: 0x + 8 or 16 characters, dependin

core: user_ta: use consistent formatting for addresses in TA dump

Improve the layout of the TA dump message by using fixed width for
physical and virtual addresses: 0x + 8 or 16 characters, depending on
the address size (32 or 64 bits). This makes the output more
consistent, more readable, and nicer overall.

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

show more ...

generic_boot: reserve optee_tzdram memory

Aside from reserving the shared memory, also reserve the TZDRAM OP-TEE
memory.

Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: R

generic_boot: reserve optee_tzdram memory

Aside from reserving the shared memory, also reserve the TZDRAM OP-TEE
memory.

Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: Rouven Czerwinski <r.czerwinski@pengutronix.de>

show more ...

generic_boot: retrieve address-cells and size-cells from root

If the reserved-memory subnode does not exist, retrieve address-cells
and size-cells from the root node.

The linux kernel checks whethe

generic_boot: retrieve address-cells and size-cells from root

If the reserved-memory subnode does not exist, retrieve address-cells
and size-cells from the root node.

The linux kernel checks whether these properties match between the root
and reserved-memory nodes and discards non-matching nodes.

Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: Rouven Czerwinski <r.czerwinski@pengutronix.de>

show more ...

generic_boot: rename shared reserved memory node

Rename the shared reserved memory node from "optee" to "optee_shm".
This should avoid confusion when we introduce the "optee_core" reserved
memory no

generic_boot: rename shared reserved memory node

Rename the shared reserved memory node from "optee" to "optee_shm".
This should avoid confusion when we introduce the "optee_core" reserved
memory node in later commits.

Reviewed-by: Jerome Forissier <jerome.forissier@linaro.org>
Signed-off-by: Rouven Czerwinski <r.czerwinski@pengutronix.de>

show more ...

zynqmp: fix UART1 base for zcu102, zc1751_dc1, zc1751_dc2 flavors

Fix UART1 base address for zcu102, zc1751_dc1, zc1751_dc2 flavors.
More information provided p226 of UG1085 [1].

Link: [1] https://

zynqmp: fix UART1 base for zcu102, zc1751_dc1, zc1751_dc2 flavors

Fix UART1 base address for zcu102, zc1751_dc1, zc1751_dc2 flavors.
More information provided p226 of UG1085 [1].

Link: [1] https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf
Signed-off-by: Michael Grand <michael.grand.mg@gmail.com>
[jf: move URL to a Link: tag]
Acked-by: Jerome Forissier <jerome.forissier@linaro.org>

show more ...

Add support for ultra96 ZynqMP board

Add flavor 'ultra96' to platform 'zynqmp'.
Redirect TEE console output to UART1.

Signed-off-by: Michael Grand <michael.grand.mg@gmail.com>
Acked-by: Jerome Fori

Add support for ultra96 ZynqMP board

Add flavor 'ultra96' to platform 'zynqmp'.
Redirect TEE console output to UART1.

Signed-off-by: Michael Grand <michael.grand.mg@gmail.com>
Acked-by: Jerome Forissier <jerome.forissier@linaro.org>

show more ...

core: better align output of TA dump with many or big regions

With the introduction of shared libutee/libutils/libmbedtls etc., it
is not uncommon for a TA to have more than 10 memory regions. When

core: better align output of TA dump with many or big regions

With the introduction of shared libutee/libutils/libmbedtls etc., it
is not uncommon for a TA to have more than 10 memory regions. When this
happens, the crash dump output is not properly aligned.

Similarly, since there is no width specifier when we print the region
size, misalignments can occur.

This commit makes the output look good for up to 100 regions of up to
16 MiB in size.

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 ...

core: elf_load: R_ARM_ABS32/R_AARCH64_ABS64 relocations against undefined symbol

The symbol referenced by a R_ARM_ABS32 or R_AARCH64_ABS64 relocation
may very well be external to the binary being re

core: elf_load: R_ARM_ABS32/R_AARCH64_ABS64 relocations against undefined symbol

The symbol referenced by a R_ARM_ABS32 or R_AARCH64_ABS64 relocation
may very well be external to the binary being relocated (for example,
defined in a shared library). In this case, the section table index for
the symbol is SHN_UNDEF and we need to perform process-wide symbol
resolution.

This fixes an issue I found when linking a TA against a shared version
of libutee (this configuration is introduced in a later commit). In this
case, ta_head::entry is set to __utee_entry which is in libutee.so,
hence undefined in the TA binary.

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

show more ...

core: cleanup generic traces

Remove useless newline character in few generic debug traces.

Remove argument __func__ from a FMSG trace since already
output by macro FMSG().

Remove error trace from

core: cleanup generic traces

Remove useless newline character in few generic debug traces.

Remove argument __func__ from a FMSG trace since already
output by macro FMSG().

Remove error trace from syscall_storage_obj_read() that, prior
this change, output failing error code from storage read()
handler. This is useless and not done for other storage handlers
return code.

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

show more ...

core/generic_boot: consistent DTB info traces

Use IMSG() traces when external and embedded DTB are tested.

Signed-off-by: Etienne Carriere <etienne.carriere@linaro.org>
Reviewed-by: Jerome Forissie

core/generic_boot: consistent DTB info traces

Use IMSG() traces when external and embedded DTB are tested.

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

show more ...

core/generic_boot: tone down trace for missing external DTB

Change trace message indicating absence of external DTB
from error level to debug level. Implementation and comment
clearly state the conf

core/generic_boot: tone down trace for missing external DTB

Change trace message indicating absence of external DTB
from error level to debug level. Implementation and comment
clearly state the configuration is fully legitimate.

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

show more ...

Compile user TAs with -fpic rather than -fpie

TA source files are compiled with the -fpie GCC flag in order to
generate a Position Independent Executable. This is not suitable to
produce a shared li

Compile user TAs with -fpic rather than -fpie

TA source files are compiled with the -fpie GCC flag in order to
generate a Position Independent Executable. This is not suitable to
produce a shared library as introduced by commit f8896d1301fc ("TA dev
kit: add support for creating shared libraries"). -fpic should be used
instead. Here is what the GCC man page has to say on these flags:

-fpic
Generate position-independent code (PIC) suitable for use in a
shared library [...]

-fpie
-fPIE
These options are similar to -fpic and -fPIC, but generated
position independent code can be only linked into executables.

So, it is quite clear that -fpie is wrong for a shared library. It is
not very clear however if -fpic can be used when generating code for an
executable. I think it can, and there is a bug report against the GCC
documentation that would confirm this [1]. Therefore we can simply use
-fpic in all cases. This is quite convenient because we currently make
no difference in the compile flags when we are building an executable,
a static library or a shared library.

The difference between -fpie and -fpic has to do with the kinds of
relocations that the compiler is allowed to emit. I stumbled upon this
issue when experimenting with shared libraries and the code proposed
by Jens to share read-only pages between TAs [2]. In my test case, a
shared library already loaded by one TA, is used by another TA. During
the load phase of the second TA, the TEE core crashed with a data-abort
(write permission fault) when trying to apply an R_ARM_REL32 relocation
to some literal pool data in the .text section of the library. The
whole .text being mapped read-only, there should be no relocation to do
here. And indeed the cause was the wrong flag (-fpie) used when
compiling the shared library.

Link: [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70419
Link: [2] https://github.com/OP-TEE/optee_os/pull/2801
Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
CC: Jens Wiklander <jens.wiklander@linaro.org>
Tested-by: Jerome Forissier <jerome.forissier@linaro.org> (QEMU)
Tested-by: Jerome Forissier <jerome.forissier@linaro.org> (HiKey960 32 & 64-bit TA)
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Tested-by: Etienne Carriere <etienne.carriere@linaro.org> (stm32mp1 Armv7)

show more ...

core: use PTA as acronym for pseudo TA

Make inline comments and trace messages more consistent by
using PTA as acronym for pseudo TA, rather than using pTA, PTA
and pta at various places.

Signed-of

core: use PTA as acronym for pseudo TA

Make inline comments and trace messages more consistent by
using PTA as acronym for pseudo TA, rather than using pTA, PTA
and pta at various places.

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

show more ...

core: pta/gprof.c: remove redundant access check

There is no need to call tee_mmu_check_access_rights() to check the
parameters of gprof_start_pc_sampling(), because they have been checked
already b

core: pta/gprof.c: remove redundant access check

There is no need to call tee_mmu_check_access_rights() to check the
parameters of gprof_start_pc_sampling(), because they have been checked
already by utee_param_to_param() in core/tee/tee_svc.c.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reviewed-by: Joakim Bech <joakim.bech@linaro.org>
Reviewed-by: Etienne Carriere <etienne.carriere@linaro.org>

show more ...

core: REE FS TAs: add option to verify signature before processing

Adds configuration flag CFG_REE_FS_TA_BUFFERED, default enabled.

A new TA store is introduced which depends on the TEE FS TA store

core: REE FS TAs: add option to verify signature before processing

Adds configuration flag CFG_REE_FS_TA_BUFFERED, default enabled.

A new TA store is introduced which depends on the TEE FS TA store to
load the whole binary into a temporary buffer in secure DDR and
authenticate it before being processed further.

This reduces the attack surface of the TEE core in case of a
vulnerability in the ELF loader, at the expense of increased memory
usage at load time.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [3.6]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Etienne Carriere <etienne.carriere@linaro.org>

show more ...

libutils: remove buf_compare_ct()

Now that we have consttime_memcmp(), buf_compare_ct() is redundant.
Every time buf_compare_ct() is used, consttime_memcmp() may be used
instead.

This commit remove

libutils: remove buf_compare_ct()

Now that we have consttime_memcmp(), buf_compare_ct() is redundant.
Every time buf_compare_ct() is used, consttime_memcmp() may be used
instead.

This commit removes buf_compare_ct(). A compatibility wrapper is kept
in <string_ext.h> to avoid knowingly breaking the build of any TA that
may use it.

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

show more ...

core: get_elf_segments(): use memmove on overlapping memory

get_elf_segments() final stage aggregates ELF segments. In the while
loop, the logic to remove the current index is to use memcpy() to shi

core: get_elf_segments(): use memmove on overlapping memory

get_elf_segments() final stage aggregates ELF segments. In the while
loop, the logic to remove the current index is to use memcpy() to shift
down everything beyond that point. This is incorrect; memmove() should
be used instead.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [2.8]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Joakim Bech <joakim.bech@linaro.org>

show more ...

core: load_elf_from_store(): check stack size

Inside load_elf_from_store(), the ta_head structure is retrieved from
un-authenticated area, and contains the stack size. The stack size could
either al

core: load_elf_from_store(): check stack size

Inside load_elf_from_store(), the ta_head structure is retrieved from
un-authenticated area, and contains the stack size. The stack size could
either already be 0, or could be large enough so it becomes 0 when rounded
up to STACK_ALIGNMENT. This could result in vm_map() returning a virtual
address for a 0-size memory block or other issues.

Check the rounded-up stack_size value before using it.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [2.7]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Joakim Bech <joakim.bech@linaro.org>

show more ...

core: add overflow check in mobj_reg_shm_alloc()

In function mobj_reg_shm_alloc(), the macro MOBJ_REG_SHM_SIZE() could
overflow depending on 'nr_pages'. In such case, the mobj_reg_shm memory
would b

core: add overflow check in mobj_reg_shm_alloc()

In function mobj_reg_shm_alloc(), the macro MOBJ_REG_SHM_SIZE() could
overflow depending on 'nr_pages'. In such case, the mobj_reg_shm memory
would be a small memory block, while num_pages would be large, which could
lead to a generous memcpy() when copying the pages in internal memory, the
outcome of this depends on memory mapping.

Note: no attack path are identified to exploit this overflow, however it
is error prone and could lead to a future vulnerability.

This commit replaces the MOBJ_REG_SHM_SIZE() macro with a static
function that performs the same computation, but returns 0 in case of
integer overflow. The call site is updated to return an error status
should this situation happen.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [2.3]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Joakim Bech <joakim.bech@linaro.org>

show more ...

core: do not use virtual addresses as session identifier

Session context virtual address is returned to the REE in
entry_open_session(); it is then used back in entry_close_session() and
entry_invok

core: do not use virtual addresses as session identifier

Session context virtual address is returned to the REE in
entry_open_session(); it is then used back in entry_close_session() and
entry_invoke_command(). Sharing virtual addresses with the REE leads to
virtual memory addresses disclosure that could be leveraged to defeat
ASLR (if/when implemented) and/or mount an attack.

Similarly, syscall_open_ta_session() returns a session ID directly
derived from the session virtual address to the caller TA.

This commit introduces a 32-bit identifier field in struct tee_ta_session.
The ID is generated when the session is created, starting from the id of
the last session in the queue, and counting up until a number that is not
used in the session queue is found.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [2.1]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Joakim Bech <joakim.bech@linaro.org>

show more ...

core: ELF relocation: use ADD_OVERFLOW()

The ELF relocation functions e32_process_rel() and e64_process_rel()
can experience integer overflows which could result in invalid memory
access. Use ADD_OV

core: ELF relocation: use ADD_OVERFLOW()

The ELF relocation functions e32_process_rel() and e64_process_rel()
can experience integer overflows which could result in invalid memory
access. Use ADD_OVERFLOW() to prevent these.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [1.8]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>

show more ...

core: elf_load_body(): use MUL_OVERFLOW() to get size of section headers

At the end of elf_load_body(), section headers are copied in a system heap
memory block, associated to state->shdr. As the co

core: elf_load_body(): use MUL_OVERFLOW() to get size of section headers

At the end of elf_load_body(), section headers are copied in a system heap
memory block, associated to state->shdr. As the computed size is the
result of an uncontrolled multiplication (ehdr.e_shnum * ehdr.e_shentsize),
it could have overflowed and result in allocating a small memory block.

Use an overflow checking macro to prevent this case.

Signed-off-by: Jerome Forissier <jerome.forissier@linaro.org>
Reported-by: Bastien Simondi <bsimondi@netflix.com> [1.7]
Reviewed-by: Jens Wiklander <jens.wiklander@linaro.org>
Reviewed-by: Joakim Bech <joakim.bech@linaro.org>

show more ...