This reports as a warning in linux-next along the lines of
Documentation/arch/riscv/cmodx.rst:14: WARNING: Title underline too short.
CMODX in the Kernel Space
--------------------- [docutils]
Documentation/arch/riscv/cmodx.rst:43: WARNING: Title underline too short.
CMODX in the User Space
--------------------- [docutils]
Documentation/arch/riscv/cmodx.rst:43: WARNING: Title underline too short.
CMODX in the User Space
--------------------- [docutils]
Link: https://lore.kernel.org/all/20250603154544.1602a8b5@canb.auug.org.au/
Fixes: 0e07200b2af6 ("riscv: Documentation: add a description about dynamic ftrace")
Reviewed-by: Charlie Jenkins <charlie@rivosinc.com>
Link: https://lore.kernel.org/r/20250603172856.49925-1-palmer@dabbelt.com
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
riscv patches for 6.16-rc1
* Implement atomic patching support for ftrace which finally allows to
get rid of stop_machine().
* Support for kexec_file_load() syscall
* Improve module loading time by changing the algorithm that counts the
number of plt/got entries in a module.
* Zicbop is now used in the kernel to prefetch instructions
[Palmer: There's been two rounds of surgery on this one, so as a result
it's a bit different than the PR.]
* alex-pr: (734 commits)
riscv: Improve Kconfig help for RISCV_ISA_V_PREEMPTIVE
MAINTAINERS: Update Atish's email address
riscv: hwprobe: export Zabha extension
riscv: Make regs_irqs_disabled() more clear
perf symbols: Ignore mapping symbols on riscv
RISC-V: Kconfig: Fix help text of CMDLINE_EXTEND
riscv: module: Optimize PLT/GOT entry counting
riscv: Add support for PUD THP
riscv: xchg: Prefetch the destination word for sc.w
riscv: Add ARCH_HAS_PREFETCH[W] support with Zicbop
riscv: Add support for Zicbop
riscv: Introduce Zicbop instructions
riscv/kexec_file: Fix comment in purgatory relocator
riscv: kexec_file: Support loading Image binary file
riscv: kexec_file: Split the loading of kernel and others
riscv: Documentation: add a description about dynamic ftrace
riscv: ftrace: support direct call using call_ops
riscv: Implement HAVE_DYNAMIC_FTRACE_WITH_CALL_OPS
riscv: ftrace: support PREEMPT
riscv: add a data fence for CMODX in the kernel mode
...
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
Björn Töpel <bjorn@kernel.org> says:
From: Björn Töpel <bjorn@rivosinc.com>
Hi!
For over a year ago, Daniel and I was testing the V2 of Song's series.
I also promised to take the V2, that had been sitting on the lists for
too long, to rebase it on a new kernel, and re-test it.
One year later, here's the V3! ;-)
There are no changes from V2 other, than some simple checkpatch
cleanups.
Song's original cover:
| This series makes the kexec_file_load() syscall support to load
| Image binary file. At the same time, corresponding support for
| kexec-tools had been pushed to my repo[2].
|
| Now, we can leverage that kexec-tools and this series to use the
| kexec_load() or kexec_file_load() syscall to boot both vmlinux and
| Image file, as seen in these combo tests:
|
| ```
| 1. kexec -l vmlinux
| 2. kexec -l Image
| 3. kexec -s -l vmlinux
| 4. kexec -s -l Image
| ```
Notably, kexec-tools has still not made it upstream. I've prepared a
branch on my GH [3], that I indend to post ASAP. That branch is a
collection of fixes/features, including Song's userland Image loading.
The V2 is here [2], and V1 [1].
I've tested the kexec-file/Image on qemu-rv64, with following
combinations:
* ACPI/UEFI
* DT/UEFI
* DT
both "regular" kexec (-s + -e), and crashkernels (-p).
Note that there are two purgatory patches that has to be present (part
of -rc1, so all good):
commit 28093cfef5 ("riscv/kexec_file: Handle R_RISCV_64 in purgatory relocator")
commit 3f7023171d ("riscv/purgatory: 4B align purgatory_start")
[1] https://lore.kernel.org/linux-riscv/20230914020044.1397356-1-songshuaishuai@tinylab.org/
[2] https://lore.kernel.org/linux-riscv/20231016092006.3347632-1-songshuaishuai@tinylab.org/
[3] https://github.com/bjoto/kexec-tools/tree/rv-on-master
* patches from https://lore.kernel.org/r/20250409193004.643839-1-bjorn@kernel.org:
riscv: kexec_file: Support loading Image binary file
riscv: kexec_file: Split the loading of kernel and others
Link: https://lore.kernel.org/r/20250409193004.643839-1-bjorn@kernel.org
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
perf reports that 99.63% of the cycles from `modprobe amdgpu` are spent
inside module_frob_arch_sections(). This is because amdgpu.ko contains
about 300000 relocations in its .rela.text section, and the algorithm in
count_max_entries() takes quadratic time.
Apply two optimizations from the arm64 code, which together reduce the
total execution time by 99.58%. First, sort the relocations so duplicate
entries are adjacent. Second, reduce the number of relocations that must
be sorted by filtering to only relocations that need PLT/GOT entries, as
done in commit d4e0340919 ("arm64/module: Optimize module load time by
optimizing PLT counting").
Unlike the arm64 code, here the filtering and sorting is done in a
scratch buffer, because the HI20 relocation search optimization in
apply_relocate_add() depends on the original order of the relocations.
This allows accumulating PLT/GOT relocations across sections so sorting
and counting is only done once per module.
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
Link: https://lore.kernel.org/r/20250409171526.862481-3-samuel.holland@sifive.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
Andy Chiu <andybnac@gmail.com> says:
This series makes atomic code patching in ftrace possible and eliminates
the need of the stop_machine dance. The major difference of this version
is that we merge the CALL_OPS support from Puranjay [1] and make direct
calls available for practical uses such as BPF. Thanks for the time
reviewing the series and suggestions, we hope this version gets a step
closer to happening in the upstream.
Please reference the link to v3 below for more introductory view of the
implementation [2]
Added patch: 2, 4, 10, 11, 12
Modified patch: 5, 6
Unchanged patch: 1, 3, 7, 8, 9
(1, 8 has commit msg modified)
Special thanks to Björn for his efforts on testing and guiding the
series!
[1]: https://lore.kernel.org/lkml/20240306165904.108141-1-puranjay12@gmail.com/
[2]: https://lore.kernel.org/linux-riscv/20241127172908.17149-1-andybnac@gmail.com/
* patches from https://lore.kernel.org/r/20250407180838.42877-1-andybnac@gmail.com:
riscv: Documentation: add a description about dynamic ftrace
riscv: ftrace: support direct call using call_ops
riscv: Implement HAVE_DYNAMIC_FTRACE_WITH_CALL_OPS
riscv: ftrace: support PREEMPT
riscv: add a data fence for CMODX in the kernel mode
riscv: vector: Support calling schedule() for preemptible Vector
riscv: ftrace: do not use stop_machine to update code
riscv: ftrace: prepare ftrace for atomic code patching
kernel: ftrace: export ftrace_sync_ipi
riscv: ftrace: align patchable functions to 4 Byte boundary
riscv: ftrace factor out code defined by !WITH_ARG
riscv: ftrace: support fastcc in Clang for WITH_ARGS
Link: https://lore.kernel.org/r/20250407180838.42877-1-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Apparently sec_base doesn't mean relocated symbol value, which seems a
copy-pasting error in the comment. Assigned with the address of section
indexed by sym->st_shndx, it should represent base address of the
relevant section. Let's fix the comment to avoid possible confusion.
Fixes: 838b3e2848 ("RISC-V: Load purgatory in kexec_file")
Signed-off-by: Yao Zi <ziyao@disroot.org>
Reviewed-by: Björn Töpel <bjorn@rivosinc.com>
Link: https://lore.kernel.org/r/20250326073450.57648-2-ziyao@disroot.org
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
This is the preparative patch for kexec_file_load Image support.
It separates the elf_kexec_load() as two parts:
- the first part loads the vmlinux (or Image)
- the second part loads other segments (e.g. initrd,fdt,purgatory)
And the second part is exported as the load_extra_segments() function
which would be used in both kexec-elf.c and kexec-image.c.
No functional change intended.
Signed-off-by: Song Shuai <songshuaishuai@tinylab.org>
Signed-off-by: Björn Töpel <bjorn@rivosinc.com>
Link: https://lore.kernel.org/r/20250409193004.643839-2-bjorn@kernel.org
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
This patch enables support for DYNAMIC_FTRACE_WITH_CALL_OPS on RISC-V.
This allows each ftrace callsite to provide an ftrace_ops to the common
ftrace trampoline, allowing each callsite to invoke distinct tracer
functions without the need to fall back to list processing or to
allocate custom trampolines for each callsite. This significantly speeds
up cases where multiple distinct trace functions are used and callsites
are mostly traced by a single tracer.
The idea and most of the implementation is taken from the ARM64's
implementation of the same feature. The idea is to place a pointer to
the ftrace_ops as a literal at a fixed offset from the function entry
point, which can be recovered by the common ftrace trampoline.
We use -fpatchable-function-entry to reserve 8 bytes above the function
entry by emitting 2 4 byte or 4 2 byte nops depending on the presence of
CONFIG_RISCV_ISA_C. These 8 bytes are patched at runtime with a pointer
to the associated ftrace_ops for that callsite. Functions are aligned to
8 bytes to make sure that the accesses to this literal are atomic.
This approach allows for directly invoking ftrace_ops::func even for
ftrace_ops which are dynamically-allocated (or part of a module),
without going via ftrace_ops_list_func.
We've benchamrked this with the ftrace_ops sample module on Spacemit K1
Jupiter:
Without this patch:
baseline (Linux rivos 6.14.0-09584-g7d06015d936c #3 SMP Sat Mar 29
+-----------------------+-----------------+----------------------------+
| Number of tracers | Total time (ns) | Per-call average time |
|-----------------------+-----------------+----------------------------|
| Relevant | Irrelevant | 100000 calls | Total (ns) | Overhead (ns) |
|----------+------------+-----------------+------------+---------------|
| 0 | 0 | 1357958 | 13 | - |
| 0 | 1 | 1302375 | 13 | - |
| 0 | 2 | 1302375 | 13 | - |
| 0 | 10 | 1379084 | 13 | - |
| 0 | 100 | 1302458 | 13 | - |
| 0 | 200 | 1302333 | 13 | - |
|----------+------------+-----------------+------------+---------------|
| 1 | 0 | 13677833 | 136 | 123 |
| 1 | 1 | 18500916 | 185 | 172 |
| 1 | 2 | 22856459 | 228 | 215 |
| 1 | 10 | 58824709 | 588 | 575 |
| 1 | 100 | 505141584 | 5051 | 5038 |
| 1 | 200 | 1580473126 | 15804 | 15791 |
|----------+------------+-----------------+------------+---------------|
| 1 | 0 | 13561000 | 135 | 122 |
| 2 | 0 | 19707292 | 197 | 184 |
| 10 | 0 | 67774750 | 677 | 664 |
| 100 | 0 | 714123125 | 7141 | 7128 |
| 200 | 0 | 1918065668 | 19180 | 19167 |
+----------+------------+-----------------+------------+---------------+
Note: per-call overhead is estimated relative to the baseline case with
0 relevant tracers and 0 irrelevant tracers.
With this patch:
v4-rc4 (Linux rivos 6.14.0-09598-gd75747611c93 #4 SMP Sat Mar 29
+-----------------------+-----------------+----------------------------+
| Number of tracers | Total time (ns) | Per-call average time |
|-----------------------+-----------------+----------------------------|
| Relevant | Irrelevant | 100000 calls | Total (ns) | Overhead (ns) |
|----------+------------+-----------------+------------+---------------|
| 0 | 0 | 1459917 | 14 | - |
| 0 | 1 | 1408000 | 14 | - |
| 0 | 2 | 1383792 | 13 | - |
| 0 | 10 | 1430709 | 14 | - |
| 0 | 100 | 1383791 | 13 | - |
| 0 | 200 | 1383750 | 13 | - |
|----------+------------+-----------------+------------+---------------|
| 1 | 0 | 5238041 | 52 | 38 |
| 1 | 1 | 5228542 | 52 | 38 |
| 1 | 2 | 5325917 | 53 | 40 |
| 1 | 10 | 5299667 | 52 | 38 |
| 1 | 100 | 5245250 | 52 | 39 |
| 1 | 200 | 5238459 | 52 | 39 |
|----------+------------+-----------------+------------+---------------|
| 1 | 0 | 5239083 | 52 | 38 |
| 2 | 0 | 19449417 | 194 | 181 |
| 10 | 0 | 67718584 | 677 | 663 |
| 100 | 0 | 709840708 | 7098 | 7085 |
| 200 | 0 | 2203580626 | 22035 | 22022 |
+----------+------------+-----------------+------------+---------------+
Note: per-call overhead is estimated relative to the baseline case with
0 relevant tracers and 0 irrelevant tracers.
As can be seen from the above:
a) Whenever there is a single relevant tracer function associated with a
tracee, the overhead of invoking the tracer is constant, and does not
scale with the number of tracers which are *not* associated with that
tracee.
b) The overhead for a single relevant tracer has dropped to ~1/3 of the
overhead prior to this series (from 122ns to 38ns). This is largely
due to permitting calls to dynamically-allocated ftrace_ops without
going through ftrace_ops_list_func.
Signed-off-by: Puranjay Mohan <puranjay12@gmail.com>
[update kconfig, asm, refactor]
Signed-off-by: Andy Chiu <andybnac@gmail.com>
Tested-by: Björn Töpel <bjorn@rivosinc.com>
Link: https://lore.kernel.org/r/20250407180838.42877-10-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
RISC-V spec explicitly calls out that a local fence.i is not enough for
the code modification to be visble from a remote hart. In fact, it
states:
To make a store to instruction memory visible to all RISC-V harts, the
writing hart also has to execute a data FENCE before requesting that all
remote RISC-V harts execute a FENCE.I.
Although current riscv drivers for IPI use ordered MMIO when sending IPIs
in order to synchronize the action between previous csd writes, riscv
does not restrict itself to any particular flavor of IPI. Any driver or
firmware implementation that does not order data writes before the IPI
may pose a risk for code-modifying race.
Thus, add a fence here to order data writes before making the IPI.
Signed-off-by: Andy Chiu <andybnac@gmail.com>
Reviewed-by: Björn Töpel <bjorn@rivosinc.com>
Link: https://lore.kernel.org/r/20250407180838.42877-8-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
Each function entry implies a call to ftrace infrastructure. And it may
call into schedule in some cases. So, it is possible for preemptible
kernel-mode Vector to implicitly call into schedule. Since all V-regs
are caller-saved, it is possible to drop all V context when a thread
voluntarily call schedule(). Besides, we currently don't pass argument
through vector register, so we don't have to save/restore V-regs in
ftrace trampoline.
Signed-off-by: Andy Chiu <andy.chiu@sifive.com>
Link: https://lore.kernel.org/r/20250407180838.42877-7-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
We use an AUIPC+JALR pair to jump into a ftrace trampoline. Since
instruction fetch can break down to 4 byte at a time, it is impossible
to update two instructions without a race. In order to mitigate it, we
initialize the patchable entry to AUIPC + NOP4. Then, the run-time code
patching can change NOP4 to JALR to eable/disable ftrcae from a
function. This limits the reach of each ftrace entry to +-2KB displacing
from ftrace_caller.
Starting from the trampoline, we add a level of indirection for it to
reach ftrace caller target. Now, it loads the target address from a
memory location, then perform the jump. This enable the kernel to update
the target atomically.
The new don't-stop-the-world text patching on change only one RISC-V
instruction:
| -8: &ftrace_ops of the associated tracer function.
| <ftrace enable>:
| 0: auipc t0, hi(ftrace_caller)
| 4: jalr t0, lo(ftrace_caller)
|
| -8: &ftrace_nop_ops
| <ftrace disable>:
| 0: auipc t0, hi(ftrace_caller)
| 4: nop
This means that f+0x0 is fixed, and should not be claimed by ftrace,
e.g. kprobe should be able to put a probe in f+0x0. Thus, we adjust the
offset and MCOUNT_INSN_SIZE accordingly.
[ alex: Fix build errors with !CONFIG_DYNAMIC_FTRACE ]
Co-developed-by: Björn Töpel <bjorn@rivosinc.com>
Signed-off-by: Björn Töpel <bjorn@rivosinc.com>
Signed-off-by: Andy Chiu <andy.chiu@sifive.com>
Link: https://lore.kernel.org/r/20250407180838.42877-5-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
The following ftrace patch for riscv uses a data store to update ftrace
function. Therefore, a romote fence is required to order it against
function_trace_op updates. The mechanism is similar to the fence between
function_trace_op and update_ftrace_func in the generic ftrace, so we
leverage the same ftrace_sync_ipi function.
[ alex: Fix build warning when !CONFIG_DYNAMIC_FTRACE ]
Signed-off-by: Andy Chiu <andybnac@gmail.com>
Link: https://lore.kernel.org/r/20250407180838.42877-4-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
We are changing ftrace code patching in order to remove dependency from
stop_machine() and enable kernel preemption. This requires us to align
functions entry at a 4-B align address.
However, -falign-functions on older versions of GCC alone was not strong
enoungh to align all functions. In fact, cold functions are not aligned
after turning on optimizations. We consider this is a bug in GCC and
turn off guess-branch-probility as a workaround to align all functions.
GCC bug id: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88345
The option -fmin-function-alignment is able to align all functions
properly on newer versions of gcc. So, we add a cc-option to test if
the toolchain supports it.
Suggested-by: Evgenii Shatokhin <e.shatokhin@yadro.com>
Signed-off-by: Andy Chiu <andy.chiu@sifive.com>
Reviewed-by: Björn Töpel <bjorn@rivosinc.com>
Link: https://lore.kernel.org/r/20250407180838.42877-3-andybnac@gmail.com
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@dabbelt.com>
Document the support for sifive vendor extensions using the key
RISCV_HWPROBE_KEY_VENDOR_EXT_SIFIVE_0 and two vendor extensions for SiFive
Int8 Matrix Multiplication Instructions using
RISCV_HWPROBE_VENDOR_EXT_XSFVQMACCDOD and
RISCV_HWPROBE_VENDOR_EXT_XSFVQMACCQOQ.
Signed-off-by: Cyan Yang <cyan.yang@sifive.com>
Link: https://lore.kernel.org/r/20250418053239.4351-4-cyan.yang@sifive.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
When RISC-V borned, DT_GNU_HASH had already became the de-facto
standard so DT_HASH is just wasting storage space. Remove the explicit
--hash-style=both setting and rely on the distro toolchain default,
which is most likely "gnu" (i.e. generating only DT_GNU_HASH, no
DT_HASH).
Following the logic of commit 48f6430505
("arm64/vdso: Remove --hash-style=sysv").
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
Link: https://lore.kernel.org/r/20250224112042.60282-2-xry111@xry111.site
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
Cyril Bur <cyrilbur@tenstorrent.com> says:
This series tries to optimize riscv uaccess by allowing the use of
user_access_begin() and user_access_end() which permits grouping user accesses
and avoiding the CSR write penalty for each access.
The error path can also be optimised using asm goto which patches 3 and 4
achieve. This will speed up jumping to labels by avoiding the need of an
intermediary error type variable within the uaccess macros
I did read the discussion this series generated. It isn't clear to me
which direction to take the patches, if any.
* b4-shazam-merge:
riscv: uaccess: use 'asm_goto_output' for get_user()
riscv: uaccess: use 'asm goto' for put_user()
riscv: uaccess: use input constraints for ptr of __put_user()
riscv: implement user_access_begin() and families
riscv: save the SR_SUM status over switches
Link: https://lore.kernel.org/r/20250410070526.3160847-1-cyrilbur@tenstorrent.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
With 'asm goto' we don't need to test the error etc, the exception just
jumps to the error handling directly.
Because there are no output clobbers which could trigger gcc bugs [1]
the use of asm_goto_output() macro is not necessary here. Not using
asm_goto_output() is desirable as the generated output asm will be
cleaner.
Use of the volatile keyword is redundant as per gcc 14.2.0 manual section
6.48.2.7 Goto Labels:
> Also note that an asm goto statement is always implicitly considered
volatile.
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113921 # 1
Signed-off-by: Jisheng Zhang <jszhang@kernel.org>
[Cyril Bur: Rewritten commit message]
Signed-off-by: Cyril Bur <cyrilbur@tenstorrent.com>
Reviewed-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Link: https://lore.kernel.org/r/20250410070526.3160847-5-cyrilbur@tenstorrent.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
Palmer Dabbelt
Miquel Sabaté Solà
Atish Patra
Alexandre Ghiti
Tiezhu Yang
Haibo Xu
谢致邦 (XIE Zhibang)
Samuel Holland
Guo Ren
Yao Zi
Song Shuai
Andy Chiu
Puranjay Mohan
Andy Chiu
Palmer Dabbelt
Cyan Yang
Xi Ruoyao
Jisheng Zhang