When OpenSBI is built with a relatively new compiler (gcc-13 and greater)
I observed that GDB is unable to produce proper backtraces and some
variable values appear corrupted (even if the associated DWARF
location descriptor is correct).
Turns out that to properly work with debug information, debuggers often
need to unwind the stack. They generally rely on Call Frame Information
(CFI) records provided by the compiler to facilitate this task.
Currently, the GCC compiler offers two mechanisms:
- `.debug_frame` section (as described in the DWARF specification).
- `.eh_frame` sections (as described in LSB documents).
The latter (`.eh_frame`) supports stack unwinding at runtime, providing
a framework for C++ exceptions or enabling backtrace generation using
libraries like libunwind. However, the downside of this approach is that
these sections should be part of loadable segments.
The former (`.debug_frame`) is simply an ordinary debug section.
Starting from GCC 13, Linux targets enable the `-fasynchronous-unwind-tables`
and `-funwind-tables` flags by default. Relevant commit:
https://github.com/gcc-mirror/gcc/commit/3cd08f7168
When these flags are active, the compiler generates `.eh_frame` sections
instead of `.debug_frame`. Since OpenSBI is built using the **Linux
toolchain**, this behavior applies to OpenSBI as well.
The problem arises because the SBI build system uses `-Wl,--gc-sections`,
which discards the `.eh_frame` section.
Possible Fixes:
1. Enforce `.debug_frame` generation – Modify compiler flags to generate
`.debug_frame` instead of `.eh_frame`.
2. Preserve `.eh_frame` in the linker script – Add `KEEP(*(.eh_frame))`
to ensure the section is not discarded.
I chose Option 1 because it avoids any runtime overhead.
Signed-off-by: Parshintsev Anatoly <anatoly.parshintsev@syntacore.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20250421124729.36364-1-anatoly.parshintsev@syntacore.com
Signed-off-by: Anup Patel <anup@brainfault.org>
If we hotplug a core and then perform a suspend-to-RAM operation on a
multi-core platform, the hotplugged CPU may be woken up along with the rest
of the system, particularly on platforms that wake all cores from the
deepest sleep state. When this happens, the hotplugged CPU enters the
sbi_hsm_wait WFI wait loop instead of transitioning into a
platform-specific low-power state. To address this, we add a HSM stop call
within the wait loop. This allows platforms that support HSM stop to enter
a low-power state when the CPU is unexpectedly woken up.
Signed-off-by: Nick Hu <nick.hu@sifive.com>
Reviewed-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20250418064506.15771-1-nick.hu@sifive.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Most CSRs are XLEN bits wide, but some are 64 bit, so rv32 needs two
accesses, plaguing the code with ifdefs.
Add new helpers that split 64 bit operation into two operations on rv32.
The helpers don't use "csr + 0x10", but append "H" at the end of the csr
name to get a compile-time error when accessing a non 64 bit register.
This has the downside that you have to use the name when accessing them.
e.g. csr_read64(0x1234) or csr_read64(CSR_SATP) won't compile and the
error messages you get for these bugs are not straightforward.
Reviewed-by: Anup Patel <anup@brainfault.org>
Signed-off-by: Radim Krčmář <rkrcmar@ventanamicro.com>
Link: https://lore.kernel.org/r/20250429142549.3673976-3-rkrcmar@ventanamicro.com
Signed-off-by: Anup Patel <anup@brainfault.org>
In current implementation, the length of hartindex_to_context_table[]
array is fixed as SBI_HARTMASK_MAX_BITS. However, the number of harts
supported by the platform might not be SBI_HARTMASK_MAX_BITS and is
usually smaller than SBI_HARTMASK_MAX_BITS. This means it is unnecessary
to allocate such fixed-length array here.
Precisely, current implementation always allocates 1024 bytes for
hartindex_to_context_table[128] on RV64 platform. However, a platform
supports two harts only needs hartindex_to_context_table[2], which only
needs 16 bytes.
This commit calculates needed size of hartindex_to_context_table[]
according to supported number of harts on the platform when registering
per-domain data, so that memory usage of per-domain context data can be
reduced.
Signed-off-by: Alvin Chang <alvinga@andestech.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20250326062051.3763530-1-alvinga@andestech.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Switch all existing platform overrides to use the helper pattern instead
of the platform hooks. After this commit, only the .match_table and
.init members of struct platform_override are used.
There are two minor behavioral differences:
- For Allwinner D1, fdt_add_cpu_idle_states() is now called before the
body of generic_final_init(). This should have no functional impact.
- For StarFive JH7110, if the /chosen/starfive,boot-hart-id property is
missing, the code now falls back to using generic_coldboot_harts,
instead of accepting any hart.
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20250325234342.711447-7-samuel.holland@sifive.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Currently the generic platform follows the middleware pattern: it
implements the sbi_platform hooks, while providing its own set of hooks
for further customization. This has a few disadvantages: each location
where customization is needed requires a separate platform_override
hook, including places where the generic function does nothing except
forward to a platform_override hook, and the extra layer of function
pointers adds runtime overhead.
Let's restructure the generic platform to follow the helper pattern.
Allow platform overrides to treat the generic platform as a template,
adding or replacing the sbi_platform_operations as needed. Export the
generic implementations, so they can be called as helpers from inside
the override functions. With this pattern, the platform_override
function pointers are replaced by direct calls, and the forwarding
functions can be removed.
The forwarding functions are not exported, since there is no reason for
an override to call them. generic_vendor_ext_check() must be rewritten,
since now there is a new way to override vendor_ext_provider.
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20250325234342.711447-6-samuel.holland@sifive.com
Signed-off-by: Anup Patel <anup@brainfault.org>
OpenSBI supports multiple supervisor domains run on same platform. When
these supervisor domains want to communicate with OpenSBI through MPXY
channels, they will allocate MPXY shared memory from their own memory
regions. Therefore, the MPXY state data structure must be per-domain and
per-hart data structure.
This commit registers per-domain MPXY state data in sbi_mpxy_init(). The
original MPXY state allocated in scratch region is also removed. We also
replace sbi_scratch_thishart_offset_ptr() macro as new
sbi_domain_mpxy_state_thishart_ptr() macro which gets MPXY state from
per-domain data.
Signed-off-by: Alvin Chang <alvinga@andestech.com>
Reviewed-by: Yu-Chien Peter Lin <peter.lin@sifive.com>
Reviewed-by: Samuel Holland <samuel.holland@sifive.com>
Link: https://lore.kernel.org/r/20250325071314.3113941-1-alvinga@andestech.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Recursively expanded variables (defined with '=') are expanded at
evaluation time. These version information variables are evaluated
inside a recipe as part of GENFLAGS. As a result, the shell commands
are executed separately for each compiler invocation. Convert the
version information variables to be simply expanded, so the shell
commands are executed only once, at Makefile evaluation time. This
speeds up the build by as much as 75%.
A separate check is needed to maintain the behavior of preferring the
value of OPENSBI_BUILD_TIME_STAMP from the environment.
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Xiang W <wxjstz@126.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Link: https://lore.kernel.org/r/20250313035755.3796610-1-samuel.holland@sifive.com
Signed-off-by: Anup Patel <anup@brainfault.org>
Similarly to what is done for SPELP, handle SSTATUS.SDT upon event
injection. In order to mimick an interrupt, set SDT to 1 for injection and
save its previous value in interrupted_flags[5:5]. Restore it upon
completion.
Signed-off-by: Clément Léger <cleger@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
As raised during the ARC review, SPELP was not handled during the event
injection process. Save it as part of the interrupted flags, clear it
before injecting the event and restore it after completion.
Signed-off-by: Clément Léger <cleger@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
For some reason, there was a pair of useless parenthesis around MSTATUS_*
value usage. Remove them.
Signed-off-by: Clément Léger <cleger@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
As raised by Andrew on the kvm-unit-test review, this flags are meant to
hold SSTATUS bits in the specification. Rename them to match that.
Signed-off-by: Clément Léger <cleger@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
The SSE specification did specified that read only parameters should
return SBI_EBADRANGE but was modified recently to return SBI_EDENIED.
Signed-off-by: Clément Léger <cleger@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
The latest specification added new high priority RAS events and renamed
the PMU to PMU_OVERFLOW.
Signed-off-by: Clément Léger <cleger@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
Harts associated with an ACLINT_MSWI need not have sequential hartids.
It is insufficient to use first_hartid and hart_count. To account for
non-sequential hart ids, include the empty hart-ids' generate hart-count.
Signed-off-by: Raj Vishwanathan <Raj.Vishwanathan@gmail.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
Simplify the code and improve consistency by using the new macros where
possible. sbi_hart_count() obsoletes sbi_scratch_last_hartindex().
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
There is currently no helper for iterating through the harts in a
system, and code must choose between sbi_scratch_last_hartindex() and
sbi_platform_hart_count() for the loop condition.
sbi_scratch_last_hartindex() has unusual semantics, leading to the
likelihood of off-by-one errors, and sbi_platform_hart_count() is
provided by the platform and so may not be properly bounded.
Add a new helper which definitively reports the number of harts managed
by this OpenSBI instance, i.e. the number of valid hart indexes, and a
convenient iterator macro.
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
The compiler generates much better code for sbi_hartindex_to_hartid()
and sbi_hartindex_to_scratch() when using a constant for the bounds
check. This works out nicely because the underlying arrays are already
a constant size, so the only change needed is to fill the remainder of
each array with the appropriate default/out-of-bounds value. The
ellipsis in the designated initializer is a GCC extension (also
supported by Clang), but avoids runtime initialization of the array.
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Anup Patel <anup@brainfault.org>
