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/[email protected]
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
Hi David,
David Hildenbrand <[email protected]> writes:
> On 14.02.22 16:22, Sven Schnelle wrote:
>> 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?
>>
> did you stumble over
>
> https://lkml.kernel.org/r/[email protected]
>
> yet?
>
> It should be fixing the (mostly impossible to trigger) splat you've seen
> -- via sleepable rcu :)
>
> The fix is scheduled for v5.18.
No, i missed that. Thank you very much!
/Sven