Skip to contentSkip to conversation

Kernel 2.4.23 on Cobalt Qube2 - area of problem

DISCUSSION Peter Horton 3 participants 4 messages Dec 12, 2003 – Dec 13, 2003 MIPS

lore
Share & tools
lore
#0 lore raw

More info on the random segmentation faults and data corruption on my Qube2.

2.4.21 from CVS is the first kernel to exhibit the problem. I tracked it down to the cache handling changes that went in between 2.4.20 and 2.4.21.

By (not very scientifically) removing flush_dcache_page() and re-instating flush_page_to_ram() I managed to get the 2.4.21 kernel stable (see attached patch). Applying a similiar patch to 2.4.23 (CVS HEAD) allows me to run 2.4.23 too.

I don't know how to track the problem any further - the kernel's cache handling is a bit out of my league.

Anyone got a clue stick they can point me in the right direction with ?

P.

arch/mips/mm/c-r4k.c +0 -13
diff -urN linux-2.4.21-xxx/arch/mips/mm/c-r4k.c linux-2.4.21/arch/mips/mm/c-r4k.c
@@ -1037,16 +1037,6 @@
 	return 1;
 }
 
-static void r4k_flush_page_to_ram_d16(struct page *page)
-{
-	blast_dcache16_page((unsigned long)page_address(page));
-}
-
-static void r4k_flush_page_to_ram_d32(struct page *page)
-{
-	blast_dcache32_page((unsigned long)page_address(page));
-}
-
 static void __init setup_noscache_funcs(void)
 {
 	unsigned int prid;
@@ -1059,7 +1049,6 @@
 			_clear_page = r4k_clear_page32_d16;
 		_copy_page = r4k_copy_page_d16;
 
-		_flush_page_to_ram = r4k_flush_page_to_ram_d16;
 		break;
 	case 32:
 		prid = read_c0_prid() & 0xfff0;
@@ -1076,8 +1065,6 @@
 				_clear_page = r4k_clear_page32_d32;
 			_copy_page = r4k_copy_page_d32;
 		}
-
-		_flush_page_to_ram = r4k_flush_page_to_ram_d32;
 		break;
 	}
 }
arch/mips/mm/cache.c +0 -4
diff -urN linux-2.4.21-xxx/arch/mips/mm/cache.c linux-2.4.21/arch/mips/mm/cache.c
@@ -20,8 +20,6 @@
 	return 0;
 }
 
-#if 0
-
 void flush_dcache_page(struct page *page)
 {
 	unsigned long addr;
@@ -67,5 +65,3 @@
 }
 
 EXPORT_SYMBOL(flush_dcache_page);
-
-#endif
arch/mips/mm/loadmmu.c +0 -2
diff -urN linux-2.4.21-xxx/arch/mips/mm/loadmmu.c linux-2.4.21/arch/mips/mm/loadmmu.c
@@ -40,8 +40,6 @@
 void (*_flush_data_cache_page)(unsigned long addr);
 void (*_flush_icache_all)(void);
 
-void (*_flush_page_to_ram)(struct page * page);
-
 #ifdef CONFIG_NONCOHERENT_IO
 
 /* DMA cache operations. */
include/asm-mips/cacheflush.h +1 -5
diff -urN linux-2.4.21-xxx/include/asm-mips/cacheflush.h linux-2.4.21/include/asm-mips/cacheflush.h
@@ -46,16 +46,12 @@
 extern void (*_flush_icache_all)(void);
 extern void (*_flush_data_cache_page)(unsigned long addr);
 
-extern void (*_flush_page_to_ram)(struct page * page);
-
-#define flush_dcache_page(page)		do { } while(0)
-
 #define flush_cache_all()		_flush_cache_all()
 #define __flush_cache_all()		___flush_cache_all()
 #define flush_cache_mm(mm)		_flush_cache_mm(mm)
 #define flush_cache_range(mm,start,end)	_flush_cache_range(mm,start,end)
 #define flush_cache_page(vma,page)	_flush_cache_page(vma, page)
-#define flush_page_to_ram(page)		_flush_page_to_ram(page)
+#define flush_page_to_ram(page)		do { } while (0)
 
 #define flush_icache_range(start, end)	_flush_icache_range(start,end)
 #define flush_icache_user_range(vma, page, addr, len) \
include/asm-mips/pgtable.h +1 -1
diff -urN linux-2.4.21-xxx/include/asm-mips/pgtable.h linux-2.4.21/include/asm-mips/pgtable.h
@@ -261,7 +261,7 @@
 	unsigned long address, pte_t pte)
 {
 	__update_tlb(vma, address, pte);
-//	__update_cache(vma, address, pte);
+	__update_cache(vma, address, pte);
 }
 
 /* Swap entries must have VALID and GLOBAL bits cleared. */
#1 lore raw
Quoted 10 lines · Peter Horton

More info on the random segmentation faults and data corruption on my Qube2.

2.4.21 from CVS is the first kernel to exhibit the problem. I tracked it down to the cache handling changes that went in between 2.4.20 and 2.4.21.

By (not very scientifically) removing flush_dcache_page() and re-instating flush_page_to_ram() I managed to get the 2.4.21 kernel stable (see attached patch). Applying a similiar patch to 2.4.23 (CVS HEAD) allows me to run 2.4.23 too.

Quoted 4 lines

I don't know how to track the problem any further - the kernel's cache handling is a bit out of my league.

Anyone got a clue stick they can point me in the right direction with ?

Can you put a printk into c-r4k.c and print the value of the shm_align_mask variable? I want to make sure it's got a sane value on your box. Also the first few lines of your bootup messages with the processor and cache stuff would be useful.

Ralf

#2 lore raw
Quoted 15 lines · Ralf Baechle
Quoted 13 lines · Peter Horton

More info on the random segmentation faults and data corruption on my Qube2.

2.4.21 from CVS is the first kernel to exhibit the problem. I tracked it down to the cache handling changes that went in between 2.4.20 and 2.4.21.

By (not very scientifically) removing flush_dcache_page() and re-instating flush_page_to_ram() I managed to get the 2.4.21 kernel stable (see attached patch). Applying a similiar patch to 2.4.23 (CVS HEAD) allows me to run 2.4.23 too.

I don't know how to track the problem any further - the kernel's cache handling is a bit out of my league.

Quoted 7 lines
Quoted 1 line

Anyone got a clue stick they can point me in the right direction with ?

Can you put a printk into c-r4k.c and print the value of the shm_align_mask variable? I want to make sure it's got a sane value on your box. Also the first few lines of your bootup messages with the processor and cache stuff would be useful.

See below. All the cache settings and the shm_align_mask look fine according to the RM5231 data sheet I have here.

At the moment the only change I make from the linux_2_4 HEAD kernel is to update include/asm-mips/cacheflush.h so:

-#define flush_page_to_ram(page) do { } while(0)
+#define flush_page_to_ram(page) flush_dcache_page(page)

This single change gives me a rock solid kernel, so it masks the problem somehow.

P.

CPU revision is: 000028a0
FPU revision is: 000028a0
D-cache:

  size    32768
  linesz  32
  sets    512
  ways    2
  waysize 16384
  waybit  14
I-cache:
  size    32768
  linesz  32
  sets    512
  ways    2

waysize 16384
waybit 14
Primary instruction cache 32kB, physically tagged, 2-way, linesize 32 bytes.
Primary data cache 32kB 2-way, linesize 32 bytes.
shm_align_mask 0x3fff
Linux version 2.4.23 (pdh@skeleton-jack) (gcc version 2.95.4 20011002 (Debian prerelease)) #6 Sat Dec 13 18:13:09 GMT 2003 Determined physical RAM map:
memory: 02000000 @ 00000000 (usable)
On node 0 totalpages: 8192
zone(0): 8192 pages.
zone(1): 0 pages.
zone(2): 0 pages.
Kernel command line: console=ttyS0,115200 ide1=noprobe root=/dev/hda2
ide_setup: ide1=noprobe
Calibrating delay loop... 249.03 BogoMIPS
Memory: 30380k/32768k available (1147k kernel code, 2388k reserved, 168k data, 100k init, 0k highmem)
...

#3 lore raw
In reply to Ralf Baechle
Quoted 15 lines · Ralf Baechle
Quoted 13 lines · Peter Horton

More info on the random segmentation faults and data corruption on my Qube2.

2.4.21 from CVS is the first kernel to exhibit the problem. I tracked it down to the cache handling changes that went in between 2.4.20 and 2.4.21.

By (not very scientifically) removing flush_dcache_page() and re-instating flush_page_to_ram() I managed to get the 2.4.21 kernel stable (see attached patch). Applying a similiar patch to 2.4.23 (CVS HEAD) allows me to run 2.4.23 too.

I don't know how to track the problem any further - the kernel's cache handling is a bit out of my league.

Quoted 7 lines
Quoted 1 line

Anyone got a clue stick they can point me in the right direction with ?

Can you put a printk into c-r4k.c and print the value of the shm_align_mask variable? I want to make sure it's got a sane value on your box. Also the first few lines of your bootup messages with the processor and cache stuff would be useful.

I've just tried changing pages_do_alias() to always return non-zero. The resulting kernel is still unstable - segmentation faults when loading shared libraries, signal 11 when compiling etc.

P.