| Message ID | 20240719010707.1319675-9-richard.henderson@linaro.org |
|---|---|
| State | Superseded |
| Headers | show |
| Series | Fixes for user-only munmap races | expand |
On Fri, 19 Jul 2024 at 02:08, Richard Henderson <richard.henderson@linaro.org> wrote: > > Mark the reserve_addr check unlikely. Use tlb_vaddr_to_host > instead of probe_write, relying on the memset itself to test > for page writability. Use set/clear_helper_retaddr so that > we can properly unwind on segfault. > > With this, a trivial loop around guest memset will spend > nearly 50% of runtime within helper_dcbz and host memset. I find this a bit difficult to interpret -- maybe add what it was before (presumably spending too much time somewhere else) ? > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Peter Maydell <peter.maydell@linaro.org> thanks -- PMM
diff --git a/target/ppc/mem_helper.c b/target/ppc/mem_helper.c index 24bae3b80c..953dd08d5d 100644 --- a/target/ppc/mem_helper.c +++ b/target/ppc/mem_helper.c @@ -280,20 +280,27 @@ static void dcbz_common(CPUPPCState *env, target_ulong addr, addr &= mask; /* Check reservation */ - if ((env->reserve_addr & mask) == addr) { + if (unlikely((env->reserve_addr & mask) == addr)) { env->reserve_addr = (target_ulong)-1ULL; } /* Try fast path translate */ +#ifdef CONFIG_USER_ONLY + haddr = tlb_vaddr_to_host(env, addr, MMU_DATA_STORE, mmu_idx); +#else haddr = probe_write(env, addr, dcbz_size, mmu_idx, retaddr); - if (haddr) { - memset(haddr, 0, dcbz_size); - } else { + if (unlikely(!haddr)) { /* Slow path */ for (int i = 0; i < dcbz_size; i += 8) { cpu_stq_mmuidx_ra(env, addr + i, 0, mmu_idx, retaddr); } + return; } +#endif + + set_helper_retaddr(retaddr); + memset(haddr, 0, dcbz_size); + clear_helper_retaddr(); } void helper_dcbz(CPUPPCState *env, target_ulong addr, int mmu_idx)
Mark the reserve_addr check unlikely. Use tlb_vaddr_to_host instead of probe_write, relying on the memset itself to test for page writability. Use set/clear_helper_retaddr so that we can properly unwind on segfault. With this, a trivial loop around guest memset will spend nearly 50% of runtime within helper_dcbz and host memset. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- target/ppc/mem_helper.c | 15 +++++++++++---- 1 file changed, 11 insertions(+), 4 deletions(-)