Re: [RFC PATCH v4] ceph: prevent a client from exceeding the MDS maximum xattr size

[Xiubo Li <xiubli@redhat.com>]

On 6/3/22 7:52 PM, Luís Henriques wrote:
> Xiubo Li <xiubli@redhat.com> writes:
>
>> On 6/2/22 10:36 PM, Luís Henriques wrote:
>>> The MDS tries to enforce a limit on the total key/values in extended
>>> attributes.  However, this limit is enforced only if doing a synchronous
>>> operation (MDS_OP_SETXATTR) -- if we're buffering the xattrs, the MDS
>>> doesn't have a chance to enforce these limits.
>>>
>>> This patch adds support for decoding the xattrs maximum size setting that is
>>> distributed in the mdsmap.  Then, when setting an xattr, the kernel client
>>> will revert to do a synchronous operation if that maximum size is exceeded.
>>>
>>> While there, fix a dout() that would trigger a printk warning:
>>>
>>> [   98.718078] ------------[ cut here ]------------
>>> [   98.719012] precision 65536 too large
>>> [   98.719039] WARNING: CPU: 1 PID: 3755 at lib/vsprintf.c:2703 vsnprintf+0x5e3/0x600
>>> ...
>>>
>>> URL: https://tracker.ceph.com/issues/55725
>>> Signed-off-by: Luís Henriques <lhenriques@suse.de>
>>> ---
>>>    fs/ceph/mdsmap.c            | 28 ++++++++++++++++++++++++----
>>>    fs/ceph/xattr.c             | 12 ++++++++----
>>>    include/linux/ceph/mdsmap.h |  1 +
>>>    3 files changed, 33 insertions(+), 8 deletions(-)
>>>
>>> * Changes since v3
>>>
>>> As per Xiubo review:
>>>     - Always force a (sync) SETXATTR Op when connecting to an old cluster
>>>     - use '>' instead of '>='
>>> Also fixed the warning detected by 0day.
>>>
>>> * Changes since v2
>>>
>>> Well, a lot has changed since v2!  Now the xattr max value setting is
>>> obtained through the mdsmap, which needs to be decoded, and the feature
>>> that was used in the previous revision was dropped.  The drawback is that
>>> the MDS isn't unable to know in advance if a client is aware of this xattr
>>> max value.
>>>
>>> * Changes since v1
>>>
>>> Added support for new feature bit to get the MDS max_xattr_pairs_size
>>> setting.
>>>
>>> Also note that this patch relies on a patch that hasn't been merged yet
>>> ("ceph: use correct index when encoding client supported features"),
>>> otherwise the new feature bit won't be correctly encoded.
>>>
>>> diff --git a/fs/ceph/mdsmap.c b/fs/ceph/mdsmap.c
>>> index 30387733765d..c6ce83a48175 100644
>>> --- a/fs/ceph/mdsmap.c
>>> +++ b/fs/ceph/mdsmap.c
>>> @@ -13,6 +13,12 @@
>>>      #include "super.h"
>>>    +/*
>>> + * Maximum size of xattrs the MDS can handle per inode by default.  This
>>> + * includes the attribute name and 4+4 bytes for the key/value sizes.
>>> + */
>>> +#define MDS_MAX_XATTR_SIZE (1<<16) /* 64K */
>> This macro makes no sense anymore, right ?
> Ah, yeah.  That's still leftovers from the previous version.
>
>> I will merge this patch by removing this then.
> Thanks!  Also... maybe you could also fix the (embarrassing) issue
> detected by 0day that I _tried_ to fix but failed (see below).
>
>> -- Xiubo
>>
>>
>>> +
>>>    #define CEPH_MDS_IS_READY(i, ignore_laggy) \
>>>    	(m->m_info[i].state > 0 && ignore_laggy ? true : !m->m_info[i].laggy)
>>>    @@ -352,12 +358,10 @@ struct ceph_mdsmap *ceph_mdsmap_decode(void **p, void
>>> *end, bool msgr2)
>>>    		__decode_and_drop_type(p, end, u8, bad_ext);
>>>    	}
>>>    	if (mdsmap_ev >= 8) {
>>> -		u32 name_len;
>>>    		/* enabled */
>>>    		ceph_decode_8_safe(p, end, m->m_enabled, bad_ext);
>>> -		ceph_decode_32_safe(p, end, name_len, bad_ext);
>>> -		ceph_decode_need(p, end, name_len, bad_ext);
>>> -		*p += name_len;
>>> +		/* fs_name */
>>> +		ceph_decode_skip_string(p, end, bad_ext);
>>>    	}
>>>    	/* damaged */
>>>    	if (mdsmap_ev >= 9) {
>>> @@ -370,6 +374,22 @@ struct ceph_mdsmap *ceph_mdsmap_decode(void **p, void *end, bool msgr2)
>>>    	} else {
>>>    		m->m_damaged = false;
>>>    	}
>>> +	if (mdsmap_ev >= 17) {
>>> +		/* balancer */
>>> +		ceph_decode_skip_string(p, end, bad_ext);
>>> +		/* standby_count_wanted */
>>> +		ceph_decode_skip_32(p, end, bad_ext);
>>> +		/* old_max_mds */
>>> +		ceph_decode_skip_32(p, end, bad_ext);
>>> +		/* min_compat_client */
>>> +		ceph_decode_skip_8(p, end, bad_ext);
>>> +		/* required_client_features */
>>> +		ceph_decode_skip_set(p, end, 64, bad_ext);
>>> +		ceph_decode_64_safe(p, end, m->m_max_xattr_size, bad_ext);
>>> +	} else {
>>> +		/* This forces the usage of the (sync) SETXATTR Op */
>>> +		m->m_max_xattr_size = 0;
>>> +	}
>>>    bad_ext:
>>>    	dout("mdsmap_decode m_enabled: %d, m_damaged: %d, m_num_laggy: %d\n",
>>>    	     !!m->m_enabled, !!m->m_damaged, m->m_num_laggy);
>>> diff --git a/fs/ceph/xattr.c b/fs/ceph/xattr.c
>>> index 8c2dc2c762a4..1be415e9220b 100644
>>> --- a/fs/ceph/xattr.c
>>> +++ b/fs/ceph/xattr.c
>>> @@ -1086,7 +1086,7 @@ static int ceph_sync_setxattr(struct inode *inode, const char *name,
>>>    			flags |= CEPH_XATTR_REMOVE;
>>>    	}
>>>    -	dout("setxattr value=%.*s\n", (int)size, value);
>>> +	dout("setxattr value size: %lu\n", size);
> This should be "%zu" and not "%lu".  I'm OK sending v5 to fix these 2
> issues if you prefer.

Sure, please send a v5 for it.


>
> [ And then there's also the MDS-side which hasn't been merged yet.  But
>    this patch is probably worth in itself as it will change the client
>    behavior to do the sync operation. ]

It's okay, this will fix this issue when we are connecting to old cephs, 
and we could force to do the sync request always.

No need to wait the ceph PR.

-- Xiubo

>
> Cheers,