The following patch does some "micro optimizations" for 2.4, that are
partly not so micro. It tries to avoid segment register reloads for
interrupts taken from kernel mode (e.g. the idle task), because they
require expensive locked cycles to read the GDT. It also changes
switch_to() not to misalign the call/ret stack during context switch,
which helps somewhat on the Athlon.
They have been running well for some time on my workstation.
-Andi
--- arch/i386/kernel/entry.S 2000/11/10 04:49:56 1.92
+++ arch/i386/kernel/entry.S 2000/12/22 18:16:38
@@ -106,10 +106,14 @@
popl %edi; \
popl %ebp; \
popl %eax; \
+ cmpl $__KERNEL_DS,(%esp); \
+ je 4f ; \
1: popl %ds; \
2: popl %es; \
addl $4,%esp; \
3: iret; \
+4: addl $12,%esp; \
+ iret; \
.section .fixup,"ax"; \
4: movl $0,(%esp); \
jmp 1b; \
@@ -271,13 +275,13 @@
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?
+ 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
-
+
ALIGN
handle_softirq:
call SYMBOL_NAME(do_softirq)
@@ -313,9 +317,11 @@
pushl %esi # push the error code
pushl %edx # push the pt_regs pointer
movl $(__KERNEL_DS),%edx
+ cmpl %edx,%ecx
+ jz 1f # should only happen for *_user or kernel bugs
movl %edx,%ds
movl %edx,%es
- GET_CURRENT(%ebx)
+1: GET_CURRENT(%ebx)
call *%edi
addl $8,%esp
jmp ret_from_exception
--- include/asm-i386/system.h 2000/09/26 05:25:56 1.44
+++ include/asm-i386/system.h 2000/12/22 18:17:26
@@ -20,7 +20,9 @@
"movl %3,%%esp\n\t" /* restore ESP */ \
"movl $1f,%1\n\t" /* save EIP */ \
"pushl %4\n\t" /* restore EIP */ \
- "jmp __switch_to\n" \
+ "call __switch_to\n\t" \
+ "popl %%esi\n\t" \
+ "jmp *%%esi\n\t" \
"1:\t" \
"popl %%ebp\n\t" \
"popl %%edi\n\t" \