[PATCH][2.6-mm] fix any_online_cpu() with cpumask_t and NR_CPUS < BITS_PER_LONG

[PATCH][2.6-mm] fix any_online_cpu() with cpumask_t and NR_CPUS < BITS_PER_LONG

[Zwane Mwaikambo]

I have no idea how this slipped in, all my boxes couldn't boot, the thing 
is it'll boot on 2way but as soon as you get to cpu2 you're shafted. 
Tested on 3x and 8x. The problem is that we really want to find the first 
cpu in the map and not whether there are processors in the map.

checking TSC synchronization across 8 CPUs: passed.
Starting migration thread for cpu 0
Bringing up 1
CPU 1 IS NOW UP!
Starting migration thread for cpu 1
Bringing up 2
CPU 2 IS NOW UP!
Starting migration thread for cpu 2
------------[ cut here ]------------
kernel BUG at kernel/softirq.c:334!
invalid operand: 0000 [#1]
PREEMPT SMP DEBUG_PAGEALLOC
CPU:    1
EIP:    0060:[<c012b187>]    Not tainted VLI
EFLAGS: 00010297
EIP is at ksoftirqd+0x47/0xf0
eax: cc3fd000   ebx: 00000002   ecx: c0329718   edx: 00000004
esi: cc3c8000   edi: 00000000   ebp: 00000000   esp: cc3c9fe8
ds: 007b   es: 007b   ss: 0068
Process ksoftirqd/2 (pid: 7, threadinfo=cc3c8000 task=cc3fd000)
Stack: c012b140 00000000 c010abd9 00000002 00000000 00000000
Call Trace:
 [<c012b140>] ksoftirqd+0x0/0xf0
 [<c010abd9>] kernel_thread_helper+0x5/0xc

Code: 50 e8 1e 6e ff ff 8b 06 88 d9 83 c4 10 81 48 0c 00 80 00 00 b8 01 00 
00 00 d3 e0 50 8b 0e 51 e8 c0 80 ff ff 58 5a 39 5e 10 74 08 <0f> 0b 4e 01 
43 99 32 c0 8b 06 c7 00 01 00 00 00 f0 83 44 24 00

--- linux-2.6.0-test2-mm4/include/asm-generic/cpumask_arith.h.fix	2003-08-06 06:05:28.810903144 -0400
+++ linux-2.6.0-test2-mm4/include/asm-generic/cpumask_arith.h	2003-08-06 06:06:16.364673872 -0400
@@ -38,9 +38,6 @@
 #define cpus_shift_right(dst, src, n)	do { dst = (src) >> (n); } while (0)
 #define cpus_shift_left(dst, src, n)	do { dst = (src) >> (n); } while (0)
 
-#define any_online_cpu(map)		(!cpus_empty(map))
-
-
 #define CPU_MASK_ALL	(~((cpumask_t)0) >> (8*sizeof(cpumask_t) - NR_CPUS))
 #define CPU_MASK_NONE	((cpumask_t)0)
 
@@ -58,4 +55,6 @@
 #define next_cpu(cpu, map)		1
 #endif /* CONFIG_SMP */
 
+#define any_online_cpu(map)		first_cpu(map)
+
 #endif /* __ASM_GENERIC_CPUMASK_ARITH_H */

Re: [PATCH][2.6-mm] fix any_online_cpu() with cpumask_t and NR_CPUS < BITS_PER_LONG

[William Lee Irwin III]

On Wed, Aug 06, 2003 at 06:23:20AM -0400, Zwane Mwaikambo wrote:
> I have no idea how this slipped in, all my boxes couldn't boot, the thing 
> is it'll boot on 2way but as soon as you get to cpu2 you're shafted. 
> Tested on 3x and 8x. The problem is that we really want to find the first 
> cpu in the map and not whether there are processors in the map.

There's more where that came from. I've not done quite as good a job of
tracking the fixes as sending them upstream.

Basically:
(a) any_online_cpu()
(b) cpus_shift_left()
(c) cpu_set()/cpu_clear() need atomicity
(d) something else that's really obvious from looking

all needed fixes, and should have all had stuff sent upstream/to akpm.


-- wli