Hi Linus,
It turns out that segment register reloading is rather expensive on modern
CPUs -- it needs several locked cycles to access the segment table which
are especially expensive on SMP.
This patch against 2.4.0test11-pre5 tries to avoid the segment register
reload for interrupts if possible, when the interrupt was entered
from kernel mode.
Only subtle point is the changed retint_with_reschedule test, the comment
should explain my reasoning. Testing against the kernel data segment
is equivalent to testing the EFLAGS/cs (unless some kernel user messes
with segments, which I consider not likely)
It also removes a bogus optimization that was added during 2.3 -- using
pushl $ret_from_intr ; jmp do_IRQ for interrupts. This one usually is slower on
Athlon (and probably P3 too) because it misaligns the call/return stack.
I changed it back to call do_IRQ ; jmp ret_from_intr.
Ideas for this patch came from Jeff Merkey and from starring at CPU simulator
logs for too long.
It is not a bug fix, but I think it is a nice simple optimization and I hope
you consider it for 2.4.
-Andi
Index: include/asm-i386/hw_irq.h
===================================================================
RCS file: /cvs/linux/include/asm-i386/hw_irq.h,v
retrieving revision 1.11
diff -u -u -r1.11 hw_irq.h
--- include/asm-i386/hw_irq.h 2000/05/16 16:08:15 1.11
+++ include/asm-i386/hw_irq.h 2000/11/18 10:33:48
@@ -92,6 +92,10 @@
#define __STR(x) #x
#define STR(x) __STR(x)
+/*
+ * A segment register reload is rather expensive. Try to avoid it
+ * if possible.
+ */
#define SAVE_ALL \
"cld\n\t" \
"pushl %es\n\t" \
@@ -103,9 +107,12 @@
"pushl %edx\n\t" \
"pushl %ecx\n\t" \
"pushl %ebx\n\t" \
- "movl $" STR(__KERNEL_DS) ",%edx\n\t" \
- "movl %edx,%ds\n\t" \
- "movl %edx,%es\n\t"
+ "movl $" STR(__KERNEL_DS),%eax\n\t" \
+ "cmpl %eax,7*4(%esp)\n\t" \
+ "je 1f\n\t" \
+ "movl %eax,%ds\n\t" \
+ "movl %eax,%es\n\t" \
+ "1:\n\t"
#define IRQ_NAME2(nr) nr##_interrupt(void)
#define IRQ_NAME(nr) IRQ_NAME2(IRQ##nr)
@@ -155,9 +162,9 @@
"\n" __ALIGN_STR"\n" \
"common_interrupt:\n\t" \
SAVE_ALL \
- "pushl $ret_from_intr\n\t" \
SYMBOL_NAME_STR(call_do_IRQ)":\n\t" \
- "jmp "SYMBOL_NAME_STR(do_IRQ));
+ "call " SYMBOL_NAME_STR(do_IRQ) "\n\t" \
+ "jmp ret_from_intr\n");
/*
* subtle. orig_eax is used by the signal code to distinct between
Index: arch/i386/kernel/entry.S
===================================================================
RCS file: /cvs/linux/arch/i386/kernel/entry.S,v
retrieving revision 1.92
diff -u -u -r1.92 entry.S
--- arch/i386/kernel/entry.S 2000/11/10 04:49:56 1.92
+++ arch/i386/kernel/entry.S 2000/11/18 10:33:48
@@ -97,6 +97,17 @@
movl $(__KERNEL_DS),%edx; \
movl %edx,%ds; \
movl %edx,%es;
+
+#define RESTORE_ALL_KERNEL \
+ popl %ebx; \
+ popl %ecx; \
+ popl %edx; \
+ popl %esi; \
+ popl %edi; \
+ popl %ebp; \
+ popl %eax; \
+ addl $12,%esp; \
+ iret
#define RESTORE_ALL \
popl %ebx; \
@@ -271,12 +282,13 @@
jne handle_softirq
ENTRY(ret_from_intr)
- GET_CURRENT(%ebx)
- movl EFLAGS(%esp),%eax # mix EFLAGS and CS
- movb CS(%esp),%al
- testl $(VM_MASK | 3),%eax # return to VM86 mode or non-supervisor?
- jne ret_with_reschedule
- jmp restore_all
+ GET_CURRENT(%ebx)
+ /* Following test implies that the beginning of SAVE_ALL or the end of RESTORE_ALL
+ could be rescheduled in kernel. This is ok because they're interruptible. It also
+ decreases the missing reschedule race on syscall exit by a few instructions. -AK */
+ cmpl $__KERNEL_DS,DS(%esp)
+ jne ret_with_reschedule
+ RESTORE_ALL_KERNEL
ALIGN
handle_softirq:
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Thu Nov 23 2000 - 21:00:15 EST