Re: [PATCH next V2] 9p: Correct the session info

[Eric Sandeen <sandeen@redhat.com>]

On 9/24/25 9:17 PM, Dominique Martinet wrote:
> Eric Sandeen wrote on Wed, Sep 24, 2025 at 05:19:11PM -0500:
>>> What I'm not following now is how the v9ses is created/handled around
>>> the new mount API:
>>> - in v9fs_get_tree a v9ses is allocated and passed along in
>>> fc->s_fs_info (that this patches now uses)
>>
>> This one should exist for the lifetime of the mount, and is the real,
>> live session info.
>>
>>> - but in v9fs_init_fs_context then a `v9fs_context` is created that
>>> also embeds (not a pointer) a v9ses struct, which is accessed through
>>> fc->fs_private as the code before this patch.
>>
>> This one should be ephemeral, used to hold options during mount/remount.
>> (see the patch description for
>> "9p: create a v9fs_context structure to hold parsed options")
>>
>> The idea is that the v9ses within the v9fs_context is used to hold
>> state/options while parsing or reconfiguring. Its lifetime is only during
>> mount/remount, and freed in v9fs_free_fc when that is done.
>>
>> Before it is freed, at the end of mount/remount, the values stored in it
>> are copied into the "live" v9ses in v9fs_apply_options().
> 
> Thanks for clarifying this;
> 
> I had which is a pointer and which is embedded wrong on the first
> reading so I first wrote that this should be ok if we could just null
> the pointer in fs_private to avoid access after the copy, but the
> temporary values are the embedded ones we can't "unaccess"...
> Hmm, do we need anything else in fs_private after v9fs_apply_options()?
> Would it make sense to call v9fs_free_fc() early on and null the full
> fs_private in or after v9fs_apply_options()?
> 
> Using a different struct tailored for mount options is certainly more
> appropriate if you have the time to do it, but as long as the risk of
> accessing "the wrong one" goes away I'm fine either way, so if you think
> nulling fs_private is possible without too much churn I think that's
> good enough.
> 
> What do you think?

I think that in retrospect, (ab)using the full v9ses was a poor choice by
me, it clearly caused confusion (for me!)

>>>> diff --git a/fs/9p/vfs_super.c b/fs/9p/vfs_super.c
>>>> index f6065b5e0e5d..bcb6ebdb8037 100644
>>>> --- a/fs/9p/vfs_super.c
>>>> +++ b/fs/9p/vfs_super.c
>>>> @@ -49,8 +49,7 @@ static int v9fs_set_super(struct super_block *s, struct fs_context *fc)
>>>>  static int v9fs_fill_super(struct super_block *sb, struct fs_context *fc)
>>>>  {
>>>>  	int ret;
>>>> -	struct v9fs_context	*ctx = fc->fs_private;
>>>> -	struct v9fs_session_info *v9ses = &ctx->v9ses;
>>>> +	struct v9fs_session_info *v9ses = sb->s_fs_info;
>>
>> This seems correct (though we can get rid of *fc too right?) 
> 
> (Sounds right as nothing else seems to use fc)
> 
>> To get here (v9fs_fill_super), we've gone through:
>>
>> v9fs_get_tree(fc)
>> 	// this will be our real live v9ses for the mount
>> 	v9ses = kzalloc(sizeof(struct v9fs_session_info), GFP_KERNEL);
>>
>>         v9fs_session_init(v9ses, fc)
>> 		// copies fc/context options in context v9ses
>> 		// into the "real" v9ses
>>                 v9fs_apply_options(v9ses, fc)
>> 		// inits maxdata on the real, live v9ses
>> 		v9ses->maxdata = v9ses->clnt->msize - P9_IOHDRSZ;
>>
>> 	fc->s_fs_info = v9ses;	// real, live v9ses on fc->s_fs_info now.
>>         sb = sget_fc(fc, NULL, v9fs_set_super);
>> 		s->s_fs_info = fc->s_fs_info; // live v9ses on s->fs_info now
>>                 err = set(s, fc)
>> 			v9fs_set_super(s, fc)
>> 				// assigns s->s_fs_info to fc->s_fs_info
>> 				// but I think this is redundant
>> 				// and should be removed
> 
> My reading is that we need it, because the super block isn't the fs
> context, and we need it for v9fs_umount_begin (it doesn't help that the
> field name is the same between both structs, and that some super_block
> variables are just 's' and others are 'sb', but I think that's the only
> place where it's used)
> 
> At this point these are both the "real live" v9ses so that should be
> fine as far as I can see.

I said remove because sget_fc() does s->s_fs_info = fc->s_fs_info, and
then v9fs_set_super essentially does s->s_fs_info = fc->s_fs_info again,
so I think it's redundant but I'll look again when I'm less sleepy. 
>> 				*v9ses = fc->s_fs_info;
>> 	                        s->s_fs_info = v9ses;
>> 				return set_anon_super(s, NULL);
>>
>>         v9fs_fill_super(sb, fc)
>>
>> and yes, with my patch were were using the option-holding "v9ses" here,
>> not the actual, fully initialized, legitimate v9ses for this mount, and
>> because v9ses->maxdata is not something that gets options parsed into it,
>> it was unset, and it so hit this bug. :(
>>
>> Going back to the prior explanation of the fc context, it's probably
>> a better idea to not re-use the v9ses struct "for convenience" and
>> make a new dedicated structure just for option parsing, so unset
>> structure members aren't even present and therefore inaccessible.
>> I'm sorry for the mistake.
>>
>> I should spin up a new version of the series fixing this and the
>> uint32/int32 difference you spotted as well
> 
> That would be great!
> 
> I'm not sure the unsigned makes much different ultimately, but might as
> well make it a fsparam_s32 as you suggested...
> Now I understand the lifecycle a bit better I can have another read with
> that in mind before merging, and we can do the nulling fs_private or
> other "make sure this bug doesn't come back" later if you don't have
> time, I'm leaving this up to you.

Sounds good. Thanks again, and sorry for somehow completely missing this
thread earlier.

(Assume we've missed this merge window by now, so I probably won't rush
on this but will try to do it sooner than later.)

Thanks,
-Eric

>> and credit Edward for his fix in the new commit log somehow?
> 
> I've never really been sure what kind of tag is appropriate in this
> case, but I assume any mention would be welcome :)
> 
> 
> Thanks,