Message ID | cover.1590474856.git.greentime.hu@sifive.com (mailing list archive) |
---|---|
Headers | show |
Series | riscv: Add vector ISA support | expand |
Hi Greentime & Vincent, Thx for the dynamic vlen implementation. I've two suggestions: - Please give out glibc patches mail URL, we need to review them together. - We need to consider that not all processes need vectors. Most system processes do not have vector features, and we should not force save/restore vector for every process. On Tue, May 26, 2020 at 3:03 PM Greentime Hu <greentime.hu@sifive.com> wrote: > > This patchset is based on Guo Ren's v3 patchset to add dynamic vlen vector > support for all different kinds of vector length in riscv. To make this > happened we defined a new __riscv_v_state in sigcontext to save the vector > related registers. The datap pointer will be allocated dynamically in > kernel space and it will be put right after the __riscv_v_state data > structure to save all the vector registers in signal handler stack for user > space. So does the implementation in ptrace, they will be saved in ubuf > in which we put the __riscv_v_state data structure and datap pointer right > after it for vector registers. This patchset also fixes several bugs for > vector lazy save/restore and vtype not saving issue. It also adds new CSR > support for vector based on the 0.9 vector spec and clean up some unused > macros. > > This patchset is rebased to v5.7-rc4 and it is tested by running several > vector programs simultaneously. It also can get the correct ucontext_t in > signal handler and restore correct context after sigreturn. It is also > tested with ptrace() syscall to use PTRACE_GETREGSET/PTRACE_SETREGSET to > get/set the vector registers. I have tested vlen=128 and vlen=256 cases in > virt machine of qemu-system-riscv32 and qemu-system-riscv64 provided by > Zhiwei Lui. > > Since the vector spec is under developing, there might still need some > changes. For example the vle.v/vse.v instructions will be replaced with > proper instructions. The reason that I don't replace the instruction in > this patchset is because that the Qemu doesn't fully support 0.9 spec yet. > I have no simulator to test. We also like to discuss the default setting of > MINSIGSTKSZ and SIGSTKSZ. They might also need to set a proper number. They > are 2048 and 8096 now. Since the stack in signal will be reserved for > ucontext and the vector registers might be larger and larger someday, these > two macros will need to be defined as a proper value or maybe we should > provide a better mechanism to provide user to get a better default signal > stack size. > > > [1] https://github.com/romanheros/qemu/tree/linux-vector-dev > [2] https://blog.linuxplumbersconf.org/2017/ocw/sessions/4671.html > [3] https://github.com/riscv/riscv-v-spec/blob/master/v-spec.adoc > > --- > Changelog V4 > - Support dynamic vlen > - Fix bugs: lazy save/resotre, not saving vtype > - Update VS bit offset based on latest vector spec > - Add new vector csr based on latest vector spec > - Code refine and removed unused macros > > Changelog V3 > - Rebase linux-5.6-rc3 and tested with qemu > - Seperate patches with Anup's advice > - Give out a ABI puzzle with unlimited vlen > > Changelog V2 > - Fixup typo "vecotr, fstate_save->vstate_save". > - Fixup wrong saved registers' length in vector.S. > - Seperate unrelated patches from this one. > > Greentime Hu (1): > ptrace: Use regset_size() for dynamic regset size. > > Guo Ren (11): > riscv: Separate patch for cflags and aflags > riscv: Rename __switch_to_aux -> fpu > riscv: Extending cpufeature.c to detect V-extension > riscv: Add new csr defines related to vector extension > riscv: Add vector feature to compile > riscv: Add has_vector/riscv_vsize to save vector features. > riscv: Reset vector register > riscv: Add vector struct and assembler definitions > riscv: Add task switch support for vector > riscv: Add ptrace vector support > riscv: Add sigcontext save/restore for vector > > Vincent Chen (1): > riscv: signal: Report signal frame size to userspace via auxv > > arch/riscv/Kconfig | 9 ++ > arch/riscv/Makefile | 19 ++-- > arch/riscv/include/asm/csr.h | 16 +++- > arch/riscv/include/asm/elf.h | 17 +++- > arch/riscv/include/asm/processor.h | 3 + > arch/riscv/include/asm/switch_to.h | 77 ++++++++++++++- > arch/riscv/include/uapi/asm/auxvec.h | 2 + > arch/riscv/include/uapi/asm/elf.h | 1 + > arch/riscv/include/uapi/asm/hwcap.h | 1 + > arch/riscv/include/uapi/asm/ptrace.h | 13 +++ > arch/riscv/include/uapi/asm/sigcontext.h | 2 + > arch/riscv/kernel/Makefile | 1 + > arch/riscv/kernel/asm-offsets.c | 8 ++ > arch/riscv/kernel/cpufeature.c | 15 ++- > arch/riscv/kernel/entry.S | 2 +- > arch/riscv/kernel/head.S | 49 +++++++++- > arch/riscv/kernel/process.c | 40 ++++++++ > arch/riscv/kernel/ptrace.c | 114 +++++++++++++++++++++++ > arch/riscv/kernel/setup.c | 5 + > arch/riscv/kernel/signal.c | 108 ++++++++++++++++++++- > arch/riscv/kernel/vector.S | 84 +++++++++++++++++ > include/uapi/linux/elf.h | 1 + > kernel/ptrace.c | 11 ++- > 23 files changed, 573 insertions(+), 25 deletions(-) > create mode 100644 arch/riscv/kernel/vector.S > > -- > 2.26.2 > >
Guo Ren <guoren@kernel.org> 於 2020年5月31日 週日 下午11:52寫道: > > Hi Greentime & Vincent, > > Thx for the dynamic vlen implementation. I've two suggestions: > - Please give out glibc patches mail URL, we need to review them together. > - We need to consider that not all processes need vectors. Most > system processes do not have vector features, and we should not force > save/restore vector for every process. > Hi Guo, Thanks for reviewing the patch. We are still cooking the glibc patch, we will add the glibc link address once we post it. For the save/restore mechanism in signal, it is basically the same with FPU porting, we can optimize it when setup_sigcontext() for both FPU and VECTOR in the future. /* Save the floating-point state. */ if (has_fpu) err |= save_fp_state(regs, &sc->sc_fpregs); /* Save the vector state. */ if (has_vector) err |= save_v_state(regs, sc); There should be a better way to detect whether this task use fpu/vector or not. Might be elf attributes or something else. > On Tue, May 26, 2020 at 3:03 PM Greentime Hu <greentime.hu@sifive.com> wrote: > > > > This patchset is based on Guo Ren's v3 patchset to add dynamic vlen vector > > support for all different kinds of vector length in riscv. To make this > > happened we defined a new __riscv_v_state in sigcontext to save the vector > > related registers. The datap pointer will be allocated dynamically in > > kernel space and it will be put right after the __riscv_v_state data > > structure to save all the vector registers in signal handler stack for user > > space. So does the implementation in ptrace, they will be saved in ubuf > > in which we put the __riscv_v_state data structure and datap pointer right > > after it for vector registers. This patchset also fixes several bugs for > > vector lazy save/restore and vtype not saving issue. It also adds new CSR > > support for vector based on the 0.9 vector spec and clean up some unused > > macros. > > > > This patchset is rebased to v5.7-rc4 and it is tested by running several > > vector programs simultaneously. It also can get the correct ucontext_t in > > signal handler and restore correct context after sigreturn. It is also > > tested with ptrace() syscall to use PTRACE_GETREGSET/PTRACE_SETREGSET to > > get/set the vector registers. I have tested vlen=128 and vlen=256 cases in > > virt machine of qemu-system-riscv32 and qemu-system-riscv64 provided by > > Zhiwei Lui. > > > > Since the vector spec is under developing, there might still need some > > changes. For example the vle.v/vse.v instructions will be replaced with > > proper instructions. The reason that I don't replace the instruction in > > this patchset is because that the Qemu doesn't fully support 0.9 spec yet. > > I have no simulator to test. We also like to discuss the default setting of > > MINSIGSTKSZ and SIGSTKSZ. They might also need to set a proper number. They > > are 2048 and 8096 now. Since the stack in signal will be reserved for > > ucontext and the vector registers might be larger and larger someday, these > > two macros will need to be defined as a proper value or maybe we should > > provide a better mechanism to provide user to get a better default signal > > stack size. > > > > > > [1] https://github.com/romanheros/qemu/tree/linux-vector-dev > > [2] https://blog.linuxplumbersconf.org/2017/ocw/sessions/4671.html > > [3] https://github.com/riscv/riscv-v-spec/blob/master/v-spec.adoc > > > > --- > > Changelog V4 > > - Support dynamic vlen > > - Fix bugs: lazy save/resotre, not saving vtype > > - Update VS bit offset based on latest vector spec > > - Add new vector csr based on latest vector spec > > - Code refine and removed unused macros > > > > Changelog V3 > > - Rebase linux-5.6-rc3 and tested with qemu > > - Seperate patches with Anup's advice > > - Give out a ABI puzzle with unlimited vlen > > > > Changelog V2 > > - Fixup typo "vecotr, fstate_save->vstate_save". > > - Fixup wrong saved registers' length in vector.S. > > - Seperate unrelated patches from this one. > > > > Greentime Hu (1): > > ptrace: Use regset_size() for dynamic regset size. > > > > Guo Ren (11): > > riscv: Separate patch for cflags and aflags > > riscv: Rename __switch_to_aux -> fpu > > riscv: Extending cpufeature.c to detect V-extension > > riscv: Add new csr defines related to vector extension > > riscv: Add vector feature to compile > > riscv: Add has_vector/riscv_vsize to save vector features. > > riscv: Reset vector register > > riscv: Add vector struct and assembler definitions > > riscv: Add task switch support for vector > > riscv: Add ptrace vector support > > riscv: Add sigcontext save/restore for vector > > > > Vincent Chen (1): > > riscv: signal: Report signal frame size to userspace via auxv > > > > arch/riscv/Kconfig | 9 ++ > > arch/riscv/Makefile | 19 ++-- > > arch/riscv/include/asm/csr.h | 16 +++- > > arch/riscv/include/asm/elf.h | 17 +++- > > arch/riscv/include/asm/processor.h | 3 + > > arch/riscv/include/asm/switch_to.h | 77 ++++++++++++++- > > arch/riscv/include/uapi/asm/auxvec.h | 2 + > > arch/riscv/include/uapi/asm/elf.h | 1 + > > arch/riscv/include/uapi/asm/hwcap.h | 1 + > > arch/riscv/include/uapi/asm/ptrace.h | 13 +++ > > arch/riscv/include/uapi/asm/sigcontext.h | 2 + > > arch/riscv/kernel/Makefile | 1 + > > arch/riscv/kernel/asm-offsets.c | 8 ++ > > arch/riscv/kernel/cpufeature.c | 15 ++- > > arch/riscv/kernel/entry.S | 2 +- > > arch/riscv/kernel/head.S | 49 +++++++++- > > arch/riscv/kernel/process.c | 40 ++++++++ > > arch/riscv/kernel/ptrace.c | 114 +++++++++++++++++++++++ > > arch/riscv/kernel/setup.c | 5 + > > arch/riscv/kernel/signal.c | 108 ++++++++++++++++++++- > > arch/riscv/kernel/vector.S | 84 +++++++++++++++++ > > include/uapi/linux/elf.h | 1 + > > kernel/ptrace.c | 11 ++- > > 23 files changed, 573 insertions(+), 25 deletions(-) > > create mode 100644 arch/riscv/kernel/vector.S > > > > -- > > 2.26.2 > > > > > > > -- > Best Regards > Guo Ren > > ML: https://lore.kernel.org/linux-csky/
On Tue, Jun 2, 2020 at 10:21 AM Greentime Hu <greentime.hu@sifive.com> wrote: > > Guo Ren <guoren@kernel.org> 於 2020年5月31日 週日 下午11:52寫道: > > > > Hi Greentime & Vincent, > > > > Thx for the dynamic vlen implementation. I've two suggestions: > > - Please give out glibc patches mail URL, we need to review them together. > > - We need to consider that not all processes need vectors. Most > > system processes do not have vector features, and we should not force > > save/restore vector for every process. > > > > Hi Guo, > > Thanks for reviewing the patch. We are still cooking the glibc patch, > we will add the glibc link address once we post it. > For the save/restore mechanism in signal, it is basically the same > with FPU porting, we can optimize it when setup_sigcontext() for both > FPU and VECTOR in the future. > > /* Save the floating-point state. */ > if (has_fpu) > err |= save_fp_state(regs, &sc->sc_fpregs); > /* Save the vector state. */ > if (has_vector) > err |= save_v_state(regs, sc); > > There should be a better way to detect whether this task use > fpu/vector or not. Might be elf attributes or something else. Ok, we could improve it in future patches.