From: Hugh Dickins <hugh@veritas.com>
Repeated -j3 kernel builds, run in tandem on dual PIII, have been collapsing
recently on -mm with 4G/4G split, SMP and preemption. Typically 'make' fails
with Error 139 because 'as' or another got SIGSEGV; maybe within ten minutes,
maybe after ten hours.
This patch seems to fix that (ran successfully overnight on test4-mm1, will
run over the weekend on test4-mm3-1). Please cast a critical eye over it, I
expect Ingo or someone else will find it can be improved.
The problem is that a task may be preempted just after it has entered
kernelspace, while using the transitional "virtual stack" i.e. %esp pointing
to high per-cpu kmap of the kernel stack. If the task resumes on another
cpu, that %esp needs to be repointed into the new cpu's kmap.
The corresponding returns to userspace look okay to me: interrupts are
disabled over the critical points. And in general no copy is taken of %esp
while on the virtual stack e.g. setting pointer to pt_regs is and must be
done after switching to real stack. But there's one place in
__SWITCH_KERNELSPACE itself where we need to check and repeat if moved.
arch/i386/kernel/entry.S | 17 ++++++++++++++++-
arch/i386/kernel/process.c | 20 +++++++++++++++++---
2 files changed, 33 insertions(+), 4 deletions(-)
diff -puN arch/i386/kernel/entry.S~4g4g-preempt-vstack-fix arch/i386/kernel/entry.S
--- 25/arch/i386/kernel/entry.S~4g4g-preempt-vstack-fix 2003-08-30 15:42:11.000000000 -0700
+++ 25-akpm/arch/i386/kernel/entry.S 2003-08-30 15:42:11.000000000 -0700
@@ -103,6 +103,20 @@ TSS_ESP0_OFFSET = (4 - 0x200)
#ifdef CONFIG_X86_SWITCH_PAGETABLES
+#if defined(CONFIG_PREEMPT) && defined(CONFIG_SMP)
+/*
+ * If task is preempted in __SWITCH_KERNELSPACE, and moved to another cpu,
+ * __switch_to repoints %esp to the appropriate virtual stack; but %ebp is
+ * left stale, so we must check whether to repeat the real stack calculation.
+ */
+#define repeat_if_esp_changed \
+ xorl %esp, %ebp; \
+ testl $0xffffe000, %ebp; \
+ jnz 0b
+#else
+#define repeat_if_esp_changed
+#endif
+
/* clobbers ebx, edx and ebp */
#define __SWITCH_KERNELSPACE \
@@ -117,12 +131,13 @@ TSS_ESP0_OFFSET = (4 - 0x200)
movl $swapper_pg_dir-__PAGE_OFFSET, %edx; \
\
/* GET_THREAD_INFO(%ebp) intermixed */ \
- \
+0: \
movl %esp, %ebp; \
movl %esp, %ebx; \
andl $0xffffe000, %ebp; \
andl $0x00001fff, %ebx; \
orl TI_real_stack(%ebp), %ebx; \
+ repeat_if_esp_changed; \
\
movl %edx, %cr3; \
movl %ebx, %esp; \
diff -puN arch/i386/kernel/process.c~4g4g-preempt-vstack-fix arch/i386/kernel/process.c
--- 25/arch/i386/kernel/process.c~4g4g-preempt-vstack-fix 2003-08-30 15:42:11.000000000 -0700
+++ 25-akpm/arch/i386/kernel/process.c 2003-08-30 15:42:11.000000000 -0700
@@ -479,13 +479,27 @@ struct task_struct * __switch_to(struct
__kmap_atomic(next->stack_page1, KM_VSTACK1);
/*
- * Reload esp0:
- */
- /*
* NOTE: here we rely on the task being the stack as well
*/
next_p->thread_info->virtual_stack = (void *)__kmap_atomic_vaddr(KM_VSTACK0);
+
+#if defined(CONFIG_PREEMPT) && defined(CONFIG_SMP)
+ /*
+ * If next was preempted on entry from userspace to kernel,
+ * and now it's on a different cpu, we need to adjust %esp.
+ * This assumes that entry.S does not copy %esp while on the
+ * virtual stack (with interrupts enabled): which is so,
+ * except within __SWITCH_KERNELSPACE itself.
+ */
+ if (unlikely(next->esp >= TASK_SIZE)) {
+ next->esp &= THREAD_SIZE - 1;
+ next->esp |= (unsigned long) next_p->thread_info->virtual_stack;
+ }
+#endif
#endif
+ /*
+ * Reload esp0:
+ */
load_esp0(tss, virtual_esp0(next_p));
/*
_