Date: Sat, 7 Dec 2019 00:28:09 +0000 (UTC) From: Konstantin Belousov <kib@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r355475 - head/sys/x86/x86 Message-ID: <201912070028.xB70S9c7018775@repo.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: kib Date: Sat Dec 7 00:28:08 2019 New Revision: 355475 URL: https://svnweb.freebsd.org/changeset/base/355475 Log: x86: Restore the critical section around whole ipi_bitmap_handler() if hardclock IPI is delivered. In the current code after r355311, critical section is taken only around hardclockintr() call, and sched_preempt() is called after the section is exited. If we reschedule after exit, as we typically would due to conditions that caused IPI, in ULE the runq tdq_ipipending is not cleared, which blocks generation of further preempt IPIs. Since all relatively modern (10 years) hardware has per-cpu event timers, restoring the critical section conditionally does not affect it. Reported and tested by: cy Diagnosed and reviewed by: jeff (previous version) Sponsored by: The FreeBSD Foundation Differential revision: https://reviews.freebsd.org/D22716 Modified: head/sys/x86/x86/mp_x86.c Modified: head/sys/x86/x86/mp_x86.c ============================================================================== --- head/sys/x86/x86/mp_x86.c Sat Dec 7 00:23:19 2019 (r355474) +++ head/sys/x86/x86/mp_x86.c Sat Dec 7 00:28:08 2019 (r355475) @@ -1262,10 +1262,28 @@ ipi_bitmap_handler(struct trapframe frame) u_int ipi_bitmap; td = curthread; + ipi_bitmap = atomic_readandclear_int(&cpuid_to_pcpu[cpu]-> + pc_ipi_bitmap); + + /* + * sched_preempt() must be called to clear the pending preempt + * IPI to enable delivery of further preempts. However, the + * critical section will cause extra scheduler lock thrashing + * when used unconditionally. Only critical_enter() if + * hardclock must also run, which requires the section entry. + */ + if (ipi_bitmap & (1 << IPI_HARDCLOCK)) + critical_enter(); + td->td_intr_nesting_level++; oldframe = td->td_intr_frame; td->td_intr_frame = &frame; - ipi_bitmap = atomic_readandclear_int(&cpuid_to_pcpu[cpu]->pc_ipi_bitmap); + if (ipi_bitmap & (1 << IPI_PREEMPT)) { +#ifdef COUNT_IPIS + (*ipi_preempt_counts[cpu])++; +#endif + sched_preempt(td); + } if (ipi_bitmap & (1 << IPI_AST)) { #ifdef COUNT_IPIS (*ipi_ast_counts[cpu])++; @@ -1273,23 +1291,15 @@ ipi_bitmap_handler(struct trapframe frame) /* Nothing to do for AST */ } if (ipi_bitmap & (1 << IPI_HARDCLOCK)) { - critical_enter(); #ifdef COUNT_IPIS (*ipi_hardclock_counts[cpu])++; #endif hardclockintr(); - critical_exit(); } - - /* Run preempt after clock handlers since it may switch. */ - if (ipi_bitmap & (1 << IPI_PREEMPT)) { -#ifdef COUNT_IPIS - (*ipi_preempt_counts[cpu])++; -#endif - sched_preempt(td); - } td->td_intr_frame = oldframe; td->td_intr_nesting_level--; + if (ipi_bitmap & (1 << IPI_HARDCLOCK)) + critical_exit(); } /*
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201912070028.xB70S9c7018775>