Re: circular locking splat in fs/proc/vmcore.c

[David Hildenbrand <david@redhat.com>]

On 14.02.22 16:22, Sven Schnelle wrote:
> Hi David,
> 
> i've seen the following lockdep splat in CI on one of our systems:
> 
> [   25.964518] kdump[727]: saving vmcore-dmesg.txt complete
> [   26.049877]
> [   26.049879] ======================================================
> [   26.049881] WARNING: possible circular locking dependency detected
> [   26.049883] 5.17.0-20220211.rc3.git2.2636bbc7cadf.300.fc35.s390x+debug #1 Tainted: G        W
> [   26.049885] ------------------------------------------------------
> [   26.049886] makedumpfile/730 is trying to acquire lock:
> [   26.049887] 0000000001a25720 (vmcore_cb_rwsem){.+.+}-{3:3}, at: mmap_vmcore+0x148/0x458
> [   26.049896]
> [   26.049896] but task is already holding lock:
> [   26.049897] 0000000013539d28 (&mm->mmap_lock){++++}-{3:3}, at: vm_mmap_pgoff+0x8e/0x170
> [   26.049904]
> [   26.049904] which lock already depends on the new lock.
> [   26.049904]
> [   26.049906]
> [   26.049906] the existing dependency chain (in reverse order) is:
> [   26.049907]
> [   26.049907] -> #1 (&mm->mmap_lock){++++}-{3:3}:
> [   26.049910]        __lock_acquire+0x604/0xbd8
> [   26.049914]        lock_acquire.part.0+0xe2/0x250
> [   26.049916]        lock_acquire+0xb0/0x200
> [   26.049918]        __might_fault+0x70/0xa0
> [   26.049921]        copy_to_user_real+0x8e/0xf8
> [   26.049925]        copy_oldmem_page+0xc0/0x158
> [   26.049930]        read_from_oldmem.part.0+0x14c/0x1b8
> [   26.049932]        __read_vmcore+0x116/0x1f8
> [   26.049933]        proc_reg_read+0x9a/0xf0
> [   26.049938]        vfs_read+0x94/0x1a8
> [   25.973256] kdump[729]: saving vmcore
> [   26.049941]        __s390x_sys_pread64+0x90/0xc8
> [   26.049958]        __do_syscall+0x1da/0x208
> [   26.049963]        system_call+0x82/0xb0
> [   26.049967]
> [   26.049967] -> #0 (vmcore_cb_rwsem){.+.+}-{3:3}:
> [   26.049971]        check_prev_add+0xe0/0xed8
> [   26.049972]        validate_chain+0x736/0xb20
> [   26.049974]        __lock_acquire+0x604/0xbd8
> [   26.049976]        lock_acquire.part.0+0xe2/0x250
> [   26.049978]        lock_acquire+0xb0/0x200
> [   26.049980]        down_read+0x5e/0x180
> [   26.049982]        mmap_vmcore+0x148/0x458
> [   26.049983]        proc_reg_mmap+0x8e/0xe0
> [   26.049985]        mmap_region+0x412/0x668
> [   26.049988]        do_mmap+0x3ec/0x4d0
> [   26.049989]        vm_mmap_pgoff+0xd4/0x170
> [   26.049992]        ksys_mmap_pgoff+0x1d8/0x228
> [   26.049994]        __s390x_sys_old_mmap+0xa4/0xb8
> [   26.049995]        __do_syscall+0x1da/0x208
> [   26.049997]        system_call+0x82/0xb0
> [   26.049999]
> [   26.049999] other info that might help us debug this:
> [   26.049999]
> [   26.050001]  Possible unsafe locking scenario:
> [   26.050001]
> [   26.050002]        CPU0                    CPU1
> [   26.050003]        ----                    ----
> [   26.050004]   lock(&mm->mmap_lock);
> [   26.050006]                                lock(vmcore_cb_rwsem);
> [   26.050008]                                lock(&mm->mmap_lock);
> [   26.050010]   lock(vmcore_cb_rwsem);
> [   26.050012]
> [   26.050012]  *** DEADLOCK ***
> [   26.050012]
> [   26.050013] 1 lock held by makedumpfile/730:
> [   26.050015]  #0: 0000000013539d28 (&mm->mmap_lock){++++}-{3:3}, at: vm_mmap_pgoff+0x8e/0x170
> 
> I think this was introduced with cc5f2704c934 ("proc/vmcore: convert
> oldmem_pfn_is_ram callback to more generic vmcore callbacks")
> 
> One fix might be to move the vmcore_cb_rwsem into the loop around the
> pfn_is_ram(). But this would likely slow down things. So the diff would
> look like: (UNTESTED)
> 
> diff --git a/fs/proc/vmcore.c b/fs/proc/vmcore.c
> index 702754dd1daf..4acd91507d21 100644
> --- a/fs/proc/vmcore.c
> +++ b/fs/proc/vmcore.c
> @@ -133,6 +133,7 @@ ssize_t read_from_oldmem(char *buf, size_t count,
>  	unsigned long pfn, offset;
>  	size_t nr_bytes;
>  	ssize_t read = 0, tmp;
> +	int is_ram;
>  
>  	if (!count)
>  		return 0;
> @@ -140,7 +141,6 @@ ssize_t read_from_oldmem(char *buf, size_t count,
>  	offset = (unsigned long)(*ppos % PAGE_SIZE);
>  	pfn = (unsigned long)(*ppos / PAGE_SIZE);
>  
> -	down_read(&vmcore_cb_rwsem);
>  	do {
>  		if (count > (PAGE_SIZE - offset))
>  			nr_bytes = PAGE_SIZE - offset;
> @@ -148,7 +148,10 @@ ssize_t read_from_oldmem(char *buf, size_t count,
>  			nr_bytes = count;
>  
>  		/* If pfn is not ram, return zeros for sparse dump files */
> -		if (!pfn_is_ram(pfn)) {
> +		down_read(&vmcore_cb_rwsem);
> +		is_ram = pfn_is_ram(pfn);
> +		up_read(&vmcore_cb_rwsem);
> +		if (!is_ram) {
>  			tmp = 0;
>  			if (!userbuf)
>  				memset(buf, 0, nr_bytes);
> @@ -164,10 +167,8 @@ ssize_t read_from_oldmem(char *buf, size_t count,
>  				tmp = copy_oldmem_page(pfn, buf, nr_bytes,
>  						       offset, userbuf);
>  		}
> -		if (tmp < 0) {
> -			up_read(&vmcore_cb_rwsem);
> +		if (tmp < 0)
>  			return tmp;
> -		}
>  
>  		*ppos += nr_bytes;
>  		count -= nr_bytes;
> @@ -177,7 +178,6 @@ ssize_t read_from_oldmem(char *buf, size_t count,
>  		offset = 0;
>  	} while (count);
>  
> -	up_read(&vmcore_cb_rwsem);
>  	return read;
>  }
>  
> I think we could also switch the list to an rcu protected list, but i
> don't know the code really. Any opinions how to fix this?
> 

Hi Sven,

did you stumble over

https://lkml.kernel.org/r/20220119193417.100385-1-david@redhat.com

yet?

It should be fixing the (mostly impossible to trigger) splat you've seen
--  via sleepable rcu :)

The fix is scheduled for v5.18.

-- 
Thanks,

David / dhildenb