linuxsh-dev

[linuxsh-dev] Unaligned access crash in signal handlers

[Denis D. <ddo...@po...>]

Hello Niibe-san,

I have been getting rare crashes in Linux 2.4.0 when handling SIGALRM
signals. I have managed to trace the problem down and it seems that the
sigreturn trap code is not getting flushed to the signal frame
sometimes. The problem goes away completely if I do a flush_cache_all
instead of flush_icache_range. I have attached a patch that does this. I
cannot see anything wrong with the current flush_icache_range usage but
suspect it is broken. The flush_cache_all is complete overkill so would
be very keen to hear if you have a fix for flush_icache_range.

The problem is very difficult to reproduce in a test program. It only
happens when parts of the stack have already been mapped into the
address cache and the signal handler is very short. I have a compiled
version of lrz from the lrzsz package that consistently crashes 2.4.0
everytime it is run. I can send this to you if you want.

Regards,
Denis

Re: [linuxsh-dev] Unaligned access crash in signal handlers

[SUGIOKA T. <su...@it...>]

At 18:45 01/01/16 +1100, Denis Dowling wrote:
>I have been getting rare crashes in Linux 2.4.0 when handling SIGALRM
>signals. I have managed to trace the problem down and it seems that the
>sigreturn trap code is not getting flushed to the signal frame
>sometimes. The problem goes away completely if I do a flush_cache_all
>instead of flush_icache_range. I have attached a patch that does this. I
>cannot see anything wrong with the current flush_icache_range usage but
>suspect it is broken. The flush_cache_all is complete overkill so would
>be very keen to hear if you have a fix for flush_icache_range.

Currently flush_icache_range uses A-bit. but A-bit works only when
ITLB entry exists. UTLB entry should not work.
So every I-cache entry should be compared with physical address by hand
or all I-cache should be invalidated.


----
SUGIOKA Toshinobu

Re: [linuxsh-dev] Unaligned access crash in signal handlers

[Denis D. <ddo...@po...>]

SUGIOKA Toshinobu wrote:
> 
> At 18:45 01/01/16 +1100, Denis Dowling wrote:
> >I have been getting rare crashes in Linux 2.4.0 when handling SIGALRM
> >signals. I have managed to trace the problem down and it seems that the
> >sigreturn trap code is not getting flushed to the signal frame
> >sometimes. The problem goes away completely if I do a flush_cache_all
> >instead of flush_icache_range. I have attached a patch that does this. I
> >cannot see anything wrong with the current flush_icache_range usage but
> >suspect it is broken. The flush_cache_all is complete overkill so would
> >be very keen to hear if you have a fix for flush_icache_range.
> 
> Currently flush_icache_range uses A-bit. but A-bit works only when
> ITLB entry exists. UTLB entry should not work.
> So every I-cache entry should be compared with physical address by hand
> or all I-cache should be invalidated.

At the moment I just completely invalidate the I and U caches. Do you
have any ideas on how to do this more efficiently? I still have problems
when I try to restart processes in the background. 

sh-2.03# xterm

(here I type ^Z)

[2]+  Stopped                 xterm
sh-2.03# bg
[2]+ xterm &
xterm: Error 50, errno 14: Bad address
sh-2.03# 
[2]+  Done(50)                xterm


I am not sure if this problem is related?

Regards,
Denis.

Re: [linuxsh-dev] Unaligned access crash in signal handlers

[SUGIOKA T. <su...@it...>]

At 12:48 01/01/19 +1100, Denis Dowling <ddo...@po...> wrote:
>
>sh-2.03# xterm
>
>(here I type ^Z)
>
>[2]+  Stopped                 xterm
>sh-2.03# bg
>[2]+ xterm &
>xterm: Error 50, errno 14: Bad address
>sh-2.03# 
>[2]+  Done(50)                xterm
>
>
>I am not sure if this problem is related?

I hear that xterm triggers binutils's bug.
Kaz Kojima's latest patch may fix this problem.

Please check
http://dodo.nurs.or.jp/~kkojima/index.html/temporary-README-new-toolchain

If you update binutils, you should rebuild gcc with new binutils 
glibc-2.2, and then all other shared libraries in this order.

Rpms of new tool chains are at http://www.sh-linux.org/rpm-index/index.html
If you are RedHat user, please test them.


----
SUGIOKA Toshinobu

Re: [linuxsh-dev] Unaligned access crash in signal handlers

[Denis D. <ddo...@po...>]

I don't think this is a binutils bug. I had to apply all of Kaz Kojima's
patches at the start of this year to get shared libraries to work. I
have just rechecked the patches and I am running the latest copies of
each.

SUGIOKA Toshinobu wrote:
> 
> At 12:48 01/01/19 +1100, Denis Dowling <ddo...@po...> wrote:
> >
> >sh-2.03# xterm
> >
> >(here I type ^Z)
> >
> >[2]+  Stopped                 xterm
> >sh-2.03# bg
> >[2]+ xterm &
> >xterm: Error 50, errno 14: Bad address
> >sh-2.03#
> >[2]+  Done(50)                xterm
> >
> >
> >I am not sure if this problem is related?
> 
> I hear that xterm triggers binutils's bug.
> Kaz Kojima's latest patch may fix this problem.
> 
> Please check
> http://dodo.nurs.or.jp/~kkojima/index.html/temporary-README-new-toolchain
> 
> If you update binutils, you should rebuild gcc with new binutils
> glibc-2.2, and then all other shared libraries in this order.
> 
> Rpms of new tool chains are at http://www.sh-linux.org/rpm-index/index.html
> If you are RedHat user, please test them.

Re: [linuxsh-dev] Unaligned access crash in signal handlers

[NIIBE Y. <gn...@m1...>]

SUGIOKA Toshinobu wrote:
 > Currently flush_icache_range uses A-bit. but A-bit works only when
 > ITLB entry exists. UTLB entry should not work.
 > So every I-cache entry should be compared with physical address by hand
 > or all I-cache should be invalidated.

Umm...  I didn't know that.  I've checked the hardware manual, and
you're right.  We need ITLB entry to flush I-cache with A-bit.

While The Right Thing may be fixing flush_icache_range, I'm not sure
about the semantics well, right now.  It seems for me that we don't
need to purge out the I-cache, but just write back D-cache, with newer
scheme (Documentation/cachetlb.txt), but I'm not sure.

Besides, currently, the arguments (START, END) to flush_icache_range
is large one (i.e. mmap ranges) except signal handling, so we need to
re-write the function anyway.

Could you please try this out?  This purges the entry regardless it's
the page mapped or not.

This also includes a bug fix of SH-3 shared page.  X server worked
quite badly in SH-3 implementation because of that.


2001-01-23  NIIBE Yutaka  <gn...@m1...>

	* arch/sh/kernel/signal.c (setup_frame, setup_rt_frame): 
	Use flush_cache_sigtramp.

	* arch/sh/mm/cache.c (flush_cache_sigtramp): Implemented.

	* include/asm-sh/pgtable.h (_PAGE_SHARED): Always _PAGE_U0_SHARED.
	(was conditionally _PAGE_HW_SHARED on SH-3).  With _PAGE_HW_SHARED,
	all processes share the page, while proper semantics is "some
	processes share the page".
	(flush_cache_sigtramp): New function.

Index: arch/sh/kernel/signal.c
===================================================================
RCS file: /cvsroot/linuxsh/kernel/arch/sh/kernel/signal.c,v
retrieving revision 1.13
diff -u -r1.13 signal.c
--- arch/sh/kernel/signal.c	2000/09/04 06:41:45	1.13
+++ arch/sh/kernel/signal.c	2001/01/23 07:46:38
@@ -433,7 +433,7 @@
 		current->comm, current->pid, frame, regs->pc, regs->pr);
 #endif
 
-	flush_icache_range(regs->pr, regs->pr+4);
+	flush_cache_sigtramp(regs->pr);
 	return;
 
 give_sigsegv:
@@ -507,7 +507,7 @@
 		current->comm, current->pid, frame, regs->pc, regs->pr);
 #endif
 
-	flush_icache_range(regs->pr, regs->pr+4);
+	flush_cache_sigtramp(regs->pr);
 	return;
 
 give_sigsegv:
Index: arch/sh/mm/cache.c
===================================================================
RCS file: /cvsroot/linuxsh/kernel/arch/sh/mm/cache.c,v
retrieving revision 1.21
diff -u -r1.21 cache.c
--- arch/sh/mm/cache.c	2000/09/18 05:15:35	1.21
+++ arch/sh/mm/cache.c	2001/01/23 07:46:38
@@ -244,6 +244,22 @@
 }
 
 /*
+ * Write back the D-cache and purge the I-cache for signal trampoline. 
+ */
+void flush_cache_sigtramp(unsigned long addr)
+{
+	unsigned long v, index;
+
+	v = addr & ~(L1_CACHE_BYTES-1);
+	asm volatile("ocbwb	%0"
+		     : /* no output */
+		     : "m" (__m(v)));
+
+	index = CACHE_IC_ADDRESS_ARRAY| (v&CACHE_IC_ENTRY_MASK);
+	ctrl_outl(0, index);	/* Clear out Valid-bit */
+}
+
+/*
  * Invalidate the I-cache of the page (don't need to write back D-cache).
  *
  * Called from kernel/ptrace.c, mm/memory.c after flush_page_to_ram is called.
Index: include/asm-sh/pgtable.h
===================================================================
RCS file: /cvsroot/linuxsh/kernel/include/asm-sh/pgtable.h,v
retrieving revision 1.16
diff -u -r1.16 pgtable.h
--- include/asm-sh/pgtable.h	2000/11/22 07:05:18	1.16
+++ include/asm-sh/pgtable.h	2001/01/23 07:46:41
@@ -39,6 +39,7 @@
 #define flush_dcache_page(page)			do { } while (0)
 #define flush_icache_range(start, end)		do { } while (0)
 #define flush_icache_page(vma,pg)		do { } while (0)
+#define flush_cache_sigtramp(vaddr)		do { } while (0)
 #elif defined(__SH4__)
 /*
  *  Caches are broken on SH-4, so we need them.
@@ -52,6 +53,7 @@
 extern void flush_dcache_page(struct page *pg);
 extern void flush_icache_range(unsigned long start, unsigned long end);
 extern void flush_icache_page(struct vm_area_struct *vma, struct page *pg);
+extern void flush_cache_sigtramp(unsigned long addr);
 #endif
 
 /*
@@ -125,11 +127,7 @@
 /* Hardware flags: SZ=1 (4k-byte) */
 #define _PAGE_FLAGS_HARD		0x00000010
 
-#if defined(__sh3__)
-#define _PAGE_SHARED	_PAGE_HW_SHARED
-#elif defined(__SH4__)
 #define _PAGE_SHARED	_PAGE_U0_SHARED
-#endif
 
 #define _PAGE_TABLE	(_PAGE_PRESENT | _PAGE_RW | _PAGE_USER | _PAGE_ACCESSED | _PAGE_DIRTY)
 #define _KERNPG_TABLE	(_PAGE_PRESENT | _PAGE_RW | _PAGE_ACCESSED | _PAGE_DIRTY)
--