Re: NULL pointer dereference using pnfs with block layout

[Kinglong Mee <kinglongmee@gmail.com>]

On 10/14/2015 18:53, Trond Myklebust wrote:
> On Wed, Oct 14, 2015 at 3:18 AM, Kinglong Mee <kinglongmee@gmail.com> wrote:
>> On 10/13/2015 21:45, Trond Myklebust wrote:
>>> On Tue, Oct 13, 2015 at 8:45 AM, Kinglong Mee <kinglongmee@gmail.com> wrote:
>>>> ping ...
>>>>
>>>> What's your opinion about this problem ?
>>>>
>>>> If read/write of block layout file with bad length (res.count != arg.count),
>>>> should nfs retry?  NFS try to call rpc_restart_call_prepare() right now,
>>>> that cause a panic with uninitialized task.
>>>
>>> The client should not be attempting to read more data than what was
>>> requested by the O_DIRECT read request. It should be strictly
>>> respecting the boundaries of the user buffer that was supplied.
>>
>> Yes, that's right.
>>
>>> Any idea why this is happening?
>>
>> As post before, bl_read_pagelist() return a longer result that causes the panic.
>>
>>>>> [ 1004.001842] bl_read_pagelist enter nr_pages 1 offset 2048 count 2048
>>>>> [ 1004.002110] bl_read_pagelist: pg_offset 2048
>>>>> [ 1004.002370] bl_read_pagelist: pg_len 2048 is_dio
>>>>> [ 1004.002617] bl_read_pagelist: pg_len 2048 after do_add_page_to_bio
>>>>> [ 1004.002853] bl_read_pagelist: 2048 4096 "(isect << SECTOR_SHIFT) < header->inode->i_size"
>>>>> [ 1004.003774] NFS: nfs_pgio_result:     0, (status 0), tk_ops      (null)
>>>>> [ 1004.003989] --> nfs4_read_done
>>>>> [ 1004.004224] nfs_readpage_done: 0
>>>>> [ 1004.004459] nfs_pgio_result: 0
>>>>> [ 1004.004691] nfs_readpage_result: eof 0, res.count 4096, args.count 2048
>>>>> [ 1004.004926] nfs_readpage_retry: tk_ops           (null)
> 
> Right, but that means one of two things: Either we need to fix
> bl_read_pagelist, or we need to fall back to read-through-MDS in this
> case.

I don't know the restrict of calling bl_read_pagelist,
Should the offset or count be aligning to PAGE_SIZE or not?

If not, there maybe some problem exist in bl_read_pagelist.

Otherwise, if bl_read_pagelist success but with res.count 
that not equal to args.count, nfs should fall back to read-through-MDS.

So, both need be fixed.

thanks,
Kinglong Mee