]> git.baikalelectronics.ru Git - kernel.git/commitdiff
powerpc/64: context tracking remove _TIF_NOHZ
authorNicholas Piggin <npiggin@gmail.com>
Sat, 30 Jan 2021 13:08:40 +0000 (23:08 +1000)
committerMichael Ellerman <mpe@ellerman.id.au>
Mon, 8 Feb 2021 13:02:12 +0000 (00:02 +1100)
Add context tracking to the system call handler explicitly, and remove
_TIF_NOHZ.

This improves system call performance when nohz_full is enabled. On a
POWER9, gettid scv system call cost on a nohz_full CPU improves from
1129 cycles to 1004 cycles and on a housekeeping CPU from 550 cycles
to 430 cycles.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210130130852.2952424-31-npiggin@gmail.com
arch/powerpc/Kconfig
arch/powerpc/include/asm/thread_info.h
arch/powerpc/kernel/ptrace/ptrace.c
arch/powerpc/kernel/signal.c
arch/powerpc/kernel/syscall_64.c

index ef64d96bf31664b44941662cc516ff26d97d41ef..686829dd1c665686ad34c5a1c31971d2783ccb4b 100644 (file)
@@ -196,7 +196,6 @@ config PPC
        select HAVE_STACKPROTECTOR              if PPC64 && $(cc-option,-mstack-protector-guard=tls -mstack-protector-guard-reg=r13)
        select HAVE_STACKPROTECTOR              if PPC32 && $(cc-option,-mstack-protector-guard=tls -mstack-protector-guard-reg=r2)
        select HAVE_CONTEXT_TRACKING            if PPC64
-       select HAVE_TIF_NOHZ                    if PPC64
        select HAVE_DEBUG_KMEMLEAK
        select HAVE_DEBUG_STACKOVERFLOW
        select HAVE_DYNAMIC_FTRACE
index 3d8a47af7a25885cab19caa77b683fc95ed9eba4..386d576673a1df56b6e18c73477d73e09ecf0eae 100644 (file)
@@ -94,7 +94,6 @@ void arch_setup_new_exec(void);
 #define TIF_PATCH_PENDING      6       /* pending live patching update */
 #define TIF_SYSCALL_AUDIT      7       /* syscall auditing active */
 #define TIF_SINGLESTEP         8       /* singlestepping active */
-#define TIF_NOHZ               9       /* in adaptive nohz mode */
 #define TIF_SECCOMP            10      /* secure computing */
 #define TIF_RESTOREALL         11      /* Restore all regs (implies NOERROR) */
 #define TIF_NOERROR            12      /* Force successful syscall return */
@@ -128,11 +127,10 @@ void arch_setup_new_exec(void);
 #define _TIF_UPROBE            (1<<TIF_UPROBE)
 #define _TIF_SYSCALL_TRACEPOINT        (1<<TIF_SYSCALL_TRACEPOINT)
 #define _TIF_EMULATE_STACK_STORE       (1<<TIF_EMULATE_STACK_STORE)
-#define _TIF_NOHZ              (1<<TIF_NOHZ)
 #define _TIF_SYSCALL_EMU       (1<<TIF_SYSCALL_EMU)
 #define _TIF_SYSCALL_DOTRACE   (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | \
                                 _TIF_SECCOMP | _TIF_SYSCALL_TRACEPOINT | \
-                                _TIF_NOHZ | _TIF_SYSCALL_EMU)
+                                _TIF_SYSCALL_EMU)
 
 #define _TIF_USER_WORK_MASK    (_TIF_SIGPENDING | _TIF_NEED_RESCHED | \
                                 _TIF_NOTIFY_RESUME | _TIF_UPROBE | \
index 3d44b73adb8333bc0828f827efaf6b59d331a7c3..4f3d4ff3728cb05a7d3e713b916fdb2737da02a6 100644 (file)
@@ -262,8 +262,6 @@ long do_syscall_trace_enter(struct pt_regs *regs)
 {
        u32 flags;
 
-       user_exit();
-
        flags = READ_ONCE(current_thread_info()->flags) &
                (_TIF_SYSCALL_EMU | _TIF_SYSCALL_TRACE);
 
@@ -340,8 +338,6 @@ void do_syscall_trace_leave(struct pt_regs *regs)
        step = test_thread_flag(TIF_SINGLESTEP);
        if (step || test_thread_flag(TIF_SYSCALL_TRACE))
                tracehook_report_syscall_exit(regs, step);
-
-       user_enter();
 }
 
 void __init pt_regs_check(void);
index 53782aa60ade678543231c3483e122a65103472f..9ded046edb0e16aa850c423a084465b8abfa42bb 100644 (file)
@@ -282,8 +282,6 @@ static void do_signal(struct task_struct *tsk)
 
 void do_notify_resume(struct pt_regs *regs, unsigned long thread_info_flags)
 {
-       user_exit();
-
        if (thread_info_flags & _TIF_UPROBE)
                uprobe_notify_resume(regs);
 
@@ -299,8 +297,6 @@ void do_notify_resume(struct pt_regs *regs, unsigned long thread_info_flags)
                tracehook_notify_resume(regs);
                rseq_handle_notify_resume(NULL, regs);
        }
-
-       user_enter();
 }
 
 static unsigned long get_tm_stackpointer(struct task_struct *tsk)
index 316c3ba165546d925b52a84f154572f2b6ae0ce0..45c4420fe3391dfe24102d4e3effc769262cbce2 100644 (file)
@@ -1,9 +1,11 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#include <linux/context_tracking.h>
 #include <linux/err.h>
 #include <asm/asm-prototypes.h>
 #include <asm/kup.h>
 #include <asm/cputime.h>
+#include <asm/interrupt.h>
 #include <asm/hw_irq.h>
 #include <asm/interrupt.h>
 #include <asm/kprobes.h>
@@ -28,6 +30,9 @@ notrace long system_call_exception(long r3, long r4, long r5,
        if (IS_ENABLED(CONFIG_PPC_IRQ_SOFT_MASK_DEBUG))
                BUG_ON(irq_soft_mask_return() != IRQS_ALL_DISABLED);
 
+       CT_WARN_ON(ct_state() == CONTEXT_KERNEL);
+       user_exit_irqoff();
+
        trace_hardirqs_off(); /* finish reconciling */
 
        if (IS_ENABLED(CONFIG_PPC_BOOK3S))
@@ -144,7 +149,7 @@ notrace long system_call_exception(long r3, long r4, long r5,
  * enabled when the interrupt handler returns (indicating a process-context /
  * synchronous interrupt) then irqs_enabled should be true.
  */
-static notrace inline bool prep_irq_for_enabled_exit(bool clear_ri, bool irqs_enabled)
+static notrace inline bool __prep_irq_for_enabled_exit(bool clear_ri)
 {
        /* This must be done with RI=1 because tracing may touch vmaps */
        trace_hardirqs_on();
@@ -161,29 +166,6 @@ static notrace inline bool prep_irq_for_enabled_exit(bool clear_ri, bool irqs_en
                trace_hardirqs_off();
                local_paca->irq_happened |= PACA_IRQ_HARD_DIS;
 
-               /*
-                * Must replay pending soft-masked interrupts now. Don't just
-                * local_irq_enabe(); local_irq_disable(); because if we are
-                * returning from an asynchronous interrupt here, another one
-                * might hit after irqs are enabled, and it would exit via this
-                * same path allowing another to fire, and so on unbounded.
-                *
-                * If interrupts were enabled when this interrupt exited,
-                * indicating a process context (synchronous) interrupt,
-                * local_irq_enable/disable can be used, which will enable
-                * interrupts rather than keeping them masked (unclear how
-                * much benefit this is over just replaying for all cases,
-                * because we immediately disable again, so all we're really
-                * doing is allowing hard interrupts to execute directly for
-                * a very small time, rather than being masked and replayed).
-                */
-               if (irqs_enabled) {
-                       local_irq_enable();
-                       local_irq_disable();
-               } else {
-                       replay_soft_interrupts();
-               }
-
                return false;
        }
        local_paca->irq_happened = 0;
@@ -192,6 +174,37 @@ static notrace inline bool prep_irq_for_enabled_exit(bool clear_ri, bool irqs_en
        return true;
 }
 
+static notrace inline bool prep_irq_for_enabled_exit(bool clear_ri, bool irqs_enabled)
+{
+       if (__prep_irq_for_enabled_exit(clear_ri))
+               return true;
+
+       /*
+        * Must replay pending soft-masked interrupts now. Don't just
+        * local_irq_enabe(); local_irq_disable(); because if we are
+        * returning from an asynchronous interrupt here, another one
+        * might hit after irqs are enabled, and it would exit via this
+        * same path allowing another to fire, and so on unbounded.
+        *
+        * If interrupts were enabled when this interrupt exited,
+        * indicating a process context (synchronous) interrupt,
+        * local_irq_enable/disable can be used, which will enable
+        * interrupts rather than keeping them masked (unclear how
+        * much benefit this is over just replaying for all cases,
+        * because we immediately disable again, so all we're really
+        * doing is allowing hard interrupts to execute directly for
+        * a very small time, rather than being masked and replayed).
+        */
+       if (irqs_enabled) {
+               local_irq_enable();
+               local_irq_disable();
+       } else {
+               replay_soft_interrupts();
+       }
+
+       return false;
+}
+
 /*
  * This should be called after a syscall returns, with r3 the return value
  * from the syscall. If this function returns non-zero, the system call
@@ -209,6 +222,8 @@ notrace unsigned long syscall_exit_prepare(unsigned long r3,
        unsigned long ti_flags;
        unsigned long ret = 0;
 
+       CT_WARN_ON(ct_state() == CONTEXT_USER);
+
        kuap_check_amr();
 
        regs->result = r3;
@@ -240,9 +255,9 @@ notrace unsigned long syscall_exit_prepare(unsigned long r3,
                ret |= _TIF_RESTOREALL;
        }
 
+again:
        local_irq_disable();
 
-again:
        ti_flags = READ_ONCE(*ti_flagsp);
        while (unlikely(ti_flags & (_TIF_USER_WORK_MASK & ~_TIF_RESTORE_TM))) {
                local_irq_enable();
@@ -286,9 +301,14 @@ again:
                }
        }
 
+       user_enter_irqoff();
+
        /* scv need not set RI=0 because SRRs are not used */
-       if (unlikely(!prep_irq_for_enabled_exit(!scv, true)))
+       if (unlikely(!__prep_irq_for_enabled_exit(!scv))) {
+               user_exit_irqoff();
+               local_irq_enable();
                goto again;
+       }
 
 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
        local_paca->tm_scratch = regs->msr;