Hi Finn, On 13/09/21 13:27, Finn Thain wrote:
On Sun, 12 Sep 2021, Finn Thain wrote:... I've now done as you did, that is, diff --git a/arch/m68k/kernel/irq.c b/arch/m68k/kernel/irq.c index 9ab4f550342e..b46d8a57f4da 100644 --- a/arch/m68k/kernel/irq.c +++ b/arch/m68k/kernel/irq.c @@ -20,10 +20,13 @@ asmlinkage void do_IRQ(int irq, struct pt_regs *regs) { struct pt_regs *oldregs = set_irq_regs(regs); + unsigned long flags; + local_irq_save(flags); irq_enter(); generic_handle_irq(irq); irq_exit(); + local_irq_restore(flags); set_irq_regs(oldregs); } There may be a better way to achieve that. If the final IPL can be found in regs then it doesn't need to be saved again. I haven't looked for a possible entropy pool improvement from correct locking in random.c -- it would not surprise me if there was one. But none of this explains the panics I saw so I went looking for potential race conditions in the irq_enter_rcu() and irq_exit_rcu() code. I haven't found the bug yet.Turns out that the panic bug was not affected by that patch... running --mmap -1 --mmap-odirect --mmap-bytes 100% -t 60 --timestamp --no-rand-seed --times stress-ng: 17:06:09.62 info: [1241] setting to a 60 second run per stressor stress-ng: 17:06:09.62 info: [1241] dispatching hogs: 1 mmap [ 807.270000] Kernel panic - not syncing: Aiee, killing interrupt handler! [ 807.270000] CPU: 0 PID: 1243 Comm: stress-ng Not tainted 5.14.0-multi-00002-g69f953866c7e #2 [ 807.270000] Stack from 00bcbde4: [ 807.270000] 00bcbde4 00488d85 00488d85 000c0000 00bcbe00 003f3708 00488d85 00bcbe20 [ 807.270000] 003f270e 000c0000 418004fc 00bca000 009f8a80 00bca000 00a06fc0 00bcbe5c [ 807.270000] 000317f6 0048098b 00000009 418004fc 00bca000 00000000 07408000 00000009 [ 807.270000] 00000008 00bcbf38 00a06fc0 00000006 00000000 00000001 00bcbe6c 000319ac [ 807.270000] 00000009 01438a20 00bcbeb8 0003acf0 00000009 0000000f 0000000e c043c000 [ 807.270000] 00000000 07408000 00000003 00bcbf98 efb2c944 efb2b8a8 00039afa 00bca000 [ 807.270000] Call Trace: [<000c0000>] insert_vmap_area.constprop.91+0xbc/0x15a [ 807.270000] [<003f3708>] dump_stack+0x10/0x16 [ 807.270000] [<003f270e>] panic+0xba/0x2bc [ 807.270000] [<000c0000>] insert_vmap_area.constprop.91+0xbc/0x15a [ 807.270000] [<000317f6>] do_exit+0x87e/0x9d6 [ 807.270000] [<000319ac>] do_group_exit+0x28/0xb6 [ 807.270000] [<0003acf0>] get_signal+0x126/0x720 [ 807.270000] [<00039afa>] send_signal+0xde/0x16e [ 807.270000] [<00004f70>] do_notify_resume+0x38/0x61c [ 807.270000] [<0003abaa>] force_sig_fault_to_task+0x36/0x3a [ 807.270000] [<0003abc6>] force_sig_fault+0x18/0x1c [ 807.270000] [<000074f4>] send_fault_sig+0x44/0xc6 [ 807.270000] [<00006a62>] buserr_c+0x2c8/0x6a2 [ 807.270000] [<00002cfc>] do_signal_return+0x10/0x1a [ 807.270000] [<0018800e>] ext4_htree_fill_tree+0x7c/0x32a [ 807.270000] [<0010800a>] d_absolute_path+0x18/0x6a [ 807.270000] [ 807.270000] ---[ end Kernel panic - not syncing: Aiee, killing interrupt handler! ]--- On the Quadra 630, the panic almost completely disappeared when I enabled the relevant CONFIG_DEBUG_* options. After about 7 hours of stress testing I got this: [23982.680000] list_add corruption. next->prev should be prev (00b51e98), but was 00bb22d8. (next=00b75cd0).
I chased a similar list corruption bug (shadow LRU list corrupt in mm/workingset.c:shadow_lru_isolate()) in 4.10. I believe that was related to an out of bounds memory access - maybe get_reg() from drivers/char/random.c but it might have been something else.
That bug had disappeared in 4.12, haven't seen it ever since.
[23982.690000] kernel BUG at lib/list_debug.c:25! [23982.700000] *** TRAP #7 *** FORMAT=0 [23982.710000] Current process id is 15489 [23982.720000] BAD KERNEL TRAP: 00000000 [23982.740000] Modules linked in: [23982.750000] PC: [<00261e62>] __list_add_valid+0x62/0xc0 [23982.760000] SR: 2000 SP: e2fb938b a2: 00bcba80 [23982.770000] d0: 00000022 d1: 00000002 d2: 008c4e40 d3: 00b7a9c0 [23982.780000] d4: 00b51e98 d5: 000da3c0 a0: 00067f00 a1: 00b51d2c [23982.790000] Process stress-ng (pid: 15489, task=35ee07ca) [23982.800000] Frame format=0 [23982.810000] Stack from 00b51e80: [23982.810000] 004cbab9 004ea3a1 00000019 004ea34f 00b51e98 00bb22d8 00b75cd0 008c4e38 [23982.810000] 00b51ecc 000da3f2 008c4e40 00b51e98 00b75cd0 00b51e5c 000f5d40 00b75cd0 [23982.810000] 00b7a9c0 00bb22d0 00b7a9c0 00b51f04 000dc346 00b51e5c 008c4e38 00b7a9c0 [23982.810000] c4c97000 00000000 c4c96000 00102073 00b14960 c4c97000 00b51e5c 00b75c94 [23982.810000] 00000001 00b51f24 000d5628 00b51e5c 00b75c94 00102070 00000000 00b75c94 [23982.810000] 00b75c94 00b51f3c 000d5728 00b14960 00b75c94 c4c97000 00000000 00b51f78 [23982.830000] Call Trace: [<000da3f2>] anon_vma_chain_link+0x32/0x80 [23982.840000] [<000f5d40>] kmem_cache_alloc+0x0/0x200 [23982.850000] [<000dc346>] anon_vma_clone+0xc6/0x180 [23982.860000] [<00102073>] cdev_get+0x33/0x80 [23982.870000] [<000d5628>] __split_vma+0x68/0x140 [23982.880000] [<00102070>] cdev_get+0x30/0x80 [23982.890000] [<000d5728>] split_vma+0x28/0x40 [23982.900000] [<000d83ba>] mprotect_fixup+0x13a/0x200 [23982.910000] [<00102070>] cdev_get+0x30/0x80 [23982.920000] [<000d8280>] mprotect_fixup+0x0/0x200 [23982.930000] [<000d85b2>] sys_mprotect+0x132/0x1c0 [23982.940000] [<00102070>] cdev_get+0x30/0x80 [23982.950000] [<00001000>] kernel_pg_dir+0x0/0x1000 [23982.960000] [<000071df>] flush_icache_range+0x1f/0x40 [23982.970000] [<00002ca4>] syscall+0x8/0xc [23982.980000] [<00001000>] kernel_pg_dir+0x0/0x1000 [23982.990000] [<00001000>] kernel_pg_dir+0x0/0x1000 [23983.000000] [<00002000>] _start+0x0/0x40 [23983.010000] [<0018800e>] ext4_ext_remove_space+0x20e/0x1540 [23983.030000] [23983.040000] Code: 4879 004e a3a1 4879 004c bab9 4e93 4e47 <b089> 6704 b088 661c 2f08 2f2e 000c 2f00 4879 004e a404 47f9 0043 d16c 4e93 4878 [23983.060000] Disabling lock debugging due to kernel taint I am still unable to reproduce this in Aranym or QEMU. (Though I did find a QEMU bug in the attempt.) I suppose list pointer corruption could have resulted in the above panic had it gone undetected. So it's tempting to blame the panic on bad DRAM --
Yes, such list corruption may well cause a kernel panic. I'd expect bad DRAM to manifest other ways before a corrupted linked list causes a kernel panic though. Filesystem corruption, for instance.
especially if this anon_vma_chain struct always gets placed at the same physical address (?)
Does it? I think this would be part of the per-process VMA data, not something like page tables ...
Incidentally - have you ever checked whether Al Viro's signal handling fixes have an impact on these bugs?
Cheers, Michael
