On Thu, Sep 14, 2023 at 2:54 AM Palmer Dabbelt <palmer@xxxxxxxxxxxx> wrote: > > On Sun, 06 Aug 2023 22:23:34 PDT (-0700), sorear@xxxxxxxxxxxx wrote: > > On Wed, Aug 2, 2023, at 12:46 PM, guoren@xxxxxxxxxx wrote: > >> From: Guo Ren <guoren@xxxxxxxxxxxxxxxxx> > >> > >> According to qspinlock requirements, RISC-V gives out a weak LR/SC > >> forward progress guarantee which does not satisfy qspinlock. But > >> many vendors could produce stronger forward guarantee LR/SC to > >> ensure the xchg_tail could be finished in time on any kind of > >> hart. T-HEAD is the vendor which implements strong forward > >> guarantee LR/SC instruction pairs, so enable qspinlock for T-HEAD > >> with errata help. > >> > >> T-HEAD early version of processors has the merge buffer delay > >> problem, so we need ERRATA_WRITEONCE to support qspinlock. > >> > >> Signed-off-by: Guo Ren <guoren@xxxxxxxxxxxxxxxxx> > >> Signed-off-by: Guo Ren <guoren@xxxxxxxxxx> > >> --- > >> arch/riscv/Kconfig.errata | 13 +++++++++++++ > >> arch/riscv/errata/thead/errata.c | 24 ++++++++++++++++++++++++ > >> arch/riscv/include/asm/errata_list.h | 20 ++++++++++++++++++++ > >> arch/riscv/include/asm/vendorid_list.h | 3 ++- > >> arch/riscv/kernel/cpufeature.c | 3 ++- > >> 5 files changed, 61 insertions(+), 2 deletions(-) > >> > >> diff --git a/arch/riscv/Kconfig.errata b/arch/riscv/Kconfig.errata > >> index 4745a5c57e7c..eb43677b13cc 100644 > >> --- a/arch/riscv/Kconfig.errata > >> +++ b/arch/riscv/Kconfig.errata > >> @@ -96,4 +96,17 @@ config ERRATA_THEAD_WRITE_ONCE > >> > >> If you don't know what to do here, say "Y". > >> > >> +config ERRATA_THEAD_QSPINLOCK > >> + bool "Apply T-Head queued spinlock errata" > >> + depends on ERRATA_THEAD > >> + default y > >> + help > >> + The T-HEAD C9xx processors implement strong fwd guarantee LR/SC to > >> + match the xchg_tail requirement of qspinlock. > >> + > >> + This will apply the QSPINLOCK errata to handle the non-standard > >> + behavior via using qspinlock instead of ticket_lock. > >> + > >> + If you don't know what to do here, say "Y". > > > > If this is to be applied, I would like to see a detailed explanation somewhere, > > preferably with citations, of: > > > > (a) The memory model requirements for qspinlock > > (b) Why, with arguments, RISC-V does not architecturally meet (a) > > (c) Why, with arguments, T-HEAD C9xx meets (a) > > (d) Why at least one other architecture which defines ARCH_USE_QUEUED_SPINLOCKS > > meets (a) > > I agree. > > Just having a magic fence that makes qspinlocks stop livelocking on some > processors is going to lead to a mess -- I'd argue this means those > processors just don't provide the forward progress guarantee, but we'd > really need something written down about what this new custom > instruction aliasing as a fence does. The "magic fence" is not related to the LR/SC forward progress guarantee, and it's our processors' store buffer hardware problem that needs to be fixed. Not only is qspinlock suffering on it, but also kernel/locking/osq_lock.c. This is about this patch: https://lore.kernel.org/linux-riscv/20230910082911.3378782-10-guoren@xxxxxxxxxx/ I've written down the root cause of that patch: The "new custom fence instruction" triggers the store buffer flush. Only the normal store instruction has the problem, and atomic stores (including LR/SC) are okay. > > > As far as I can tell, the RISC-V guarantees concerning constrained LR/SC loops > > (livelock freedom but no starvation freedom) are exactly the same as those in > > Armv8 (as of 0487F.c) for equivalent loops, and xchg_tail compiles to a > > constrained LR/SC loop with guaranteed eventual success (with -O1). Clearly you > > disagree; I would like to see your perspective. > > It sounds to me like this processor might be quite broken: if it's > permanently holding stores in a buffer we're going to have more issues > than just qspinlock, pretty much anything concurrent is going to have > issues -- and that's not just in the kernel, there's concurrent > userspace code as well. Yes, the userspace scenarios are our worries because modifying the userspace atomic library is impossible. Now, we are stress-testing various userspace parallel applications, especially userspace queued spinlock, on the 128-core hardware platform. We haven't detected the problem in the userspace, and it could be because of timer interrupts, which breaks the store buffer waiting. > > > -s > > > >> + > >> endmenu # "CPU errata selection" > >> diff --git a/arch/riscv/errata/thead/errata.c b/arch/riscv/errata/thead/errata.c > >> index 881729746d2e..d560dc45c0e7 100644 > >> --- a/arch/riscv/errata/thead/errata.c > >> +++ b/arch/riscv/errata/thead/errata.c > >> @@ -86,6 +86,27 @@ static bool errata_probe_write_once(unsigned int stage, > >> return false; > >> } > >> > >> +static bool errata_probe_qspinlock(unsigned int stage, > >> + unsigned long arch_id, unsigned long impid) > >> +{ > >> + if (!IS_ENABLED(CONFIG_ERRATA_THEAD_QSPINLOCK)) > >> + return false; > >> + > >> + /* > >> + * The queued_spinlock torture would get in livelock without > >> + * ERRATA_THEAD_WRITE_ONCE fixup for the early versions of T-HEAD > >> + * processors. > >> + */ > >> + if (arch_id == 0 && impid == 0 && > >> + !IS_ENABLED(CONFIG_ERRATA_THEAD_WRITE_ONCE)) > >> + return false; > >> + > >> + if (stage == RISCV_ALTERNATIVES_EARLY_BOOT) > >> + return true; > >> + > >> + return false; > >> +} > >> + > >> static u32 thead_errata_probe(unsigned int stage, > >> unsigned long archid, unsigned long impid) > >> { > >> @@ -103,6 +124,9 @@ static u32 thead_errata_probe(unsigned int stage, > >> if (errata_probe_write_once(stage, archid, impid)) > >> cpu_req_errata |= BIT(ERRATA_THEAD_WRITE_ONCE); > >> > >> + if (errata_probe_qspinlock(stage, archid, impid)) > >> + cpu_req_errata |= BIT(ERRATA_THEAD_QSPINLOCK); > >> + > >> return cpu_req_errata; > >> } > >> > >> diff --git a/arch/riscv/include/asm/errata_list.h > >> b/arch/riscv/include/asm/errata_list.h > >> index fbb2b8d39321..a696d18d1b0d 100644 > >> --- a/arch/riscv/include/asm/errata_list.h > >> +++ b/arch/riscv/include/asm/errata_list.h > >> @@ -141,6 +141,26 @@ asm volatile(ALTERNATIVE( \ > >> : "=r" (__ovl) : \ > >> : "memory") > >> > >> +static __always_inline bool > >> +riscv_has_errata_thead_qspinlock(void) > >> +{ > >> + if (IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) { > >> + asm_volatile_goto( > >> + ALTERNATIVE( > >> + "j %l[l_no]", "nop", > >> + THEAD_VENDOR_ID, > >> + ERRATA_THEAD_QSPINLOCK, > >> + CONFIG_ERRATA_THEAD_QSPINLOCK) > >> + : : : : l_no); > >> + } else { > >> + goto l_no; > >> + } > >> + > >> + return true; > >> +l_no: > >> + return false; > >> +} > >> + > >> #endif /* __ASSEMBLY__ */ > >> > >> #endif > >> diff --git a/arch/riscv/include/asm/vendorid_list.h > >> b/arch/riscv/include/asm/vendorid_list.h > >> index 73078cfe4029..1f1d03877f5f 100644 > >> --- a/arch/riscv/include/asm/vendorid_list.h > >> +++ b/arch/riscv/include/asm/vendorid_list.h > >> @@ -19,7 +19,8 @@ > >> #define ERRATA_THEAD_CMO 1 > >> #define ERRATA_THEAD_PMU 2 > >> #define ERRATA_THEAD_WRITE_ONCE 3 > >> -#define ERRATA_THEAD_NUMBER 4 > >> +#define ERRATA_THEAD_QSPINLOCK 4 > >> +#define ERRATA_THEAD_NUMBER 5 > >> #endif > >> > >> #endif > >> diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c > >> index f8dbbe1bbd34..d9694fe40a9a 100644 > >> --- a/arch/riscv/kernel/cpufeature.c > >> +++ b/arch/riscv/kernel/cpufeature.c > >> @@ -342,7 +342,8 @@ void __init riscv_fill_hwcap(void) > >> * spinlock value, the only way is to change from queued_spinlock to > >> * ticket_spinlock, but can not be vice. > >> */ > >> - if (!force_qspinlock) { > >> + if (!force_qspinlock && > >> + !riscv_has_errata_thead_qspinlock()) { > >> set_bit(RISCV_ISA_EXT_XTICKETLOCK, isainfo->isa); > >> } > >> #endif > >> -- > >> 2.36.1 > >> > >> > >> _______________________________________________ > >> linux-riscv mailing list > >> linux-riscv@xxxxxxxxxxxxxxxxxxx > >> http://lists.infradead.org/mailman/listinfo/linux-riscv -- Best Regards Guo Ren