On Jun 14, 2016, at 2:46 PM, J . Bruce Fields wrote:
> On Tue, Jun 14, 2016 at 11:56:20AM -0400, Oleg Drokin wrote:
>>
>> On Jun 14, 2016, at 11:46 AM, J . Bruce Fields wrote:
>>
>>> On Sun, Jun 12, 2016 at 09:26:27PM -0400, Oleg Drokin wrote:
>>>> It used to be the case that state had an rwlock that was locked for write
>>>> by downgrades, but for read for upgrades (opens). Well, the problem is
>>>> if there are two competing opens for the same state, they step on
>>>> each other toes potentially leading to leaking file descriptors
>>>> from the state structure, since access mode is a bitmap only set once.
>>>>
>>>> Extend the holding region around in nfsd4_process_open2() to avoid
>>>> racing entry into nfs4_get_vfs_file().
>>>> Make init_open_stateid() return with locked stateid to be unlocked
>>>> by the caller.
>>>>
>>>> Now this version held up pretty well in my testing for 24 hours.
>>>> It still does not address the situation if during one of the racing
>>>> nfs4_get_vfs_file() calls we are getting an error from one (first?)
>>>> of them. This is to be addressed in a separate patch after having a
>>>> solid reproducer (potentially using some fault injection).
>>>>
>>>> Signed-off-by: Oleg Drokin <gr...@linuxhacker.ru>
>>>> ---
>>>> fs/nfsd/nfs4state.c | 47 +++++++++++++++++++++++++++--------------------
>>>> fs/nfsd/state.h | 2 +-
>>>> 2 files changed, 28 insertions(+), 21 deletions(-)
>>>>
>>>> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
>>>> index f5f82e1..fa5fb5a 100644
>>>> --- a/fs/nfsd/nfs4state.c
>>>> +++ b/fs/nfsd/nfs4state.c
>>>> @@ -3487,6 +3487,10 @@ init_open_stateid(struct nfs4_ol_stateid *stp,
>>>> struct nfs4_file *fp,
>>>> struct nfs4_openowner *oo = open->op_openowner;
>>>> struct nfs4_ol_stateid *retstp = NULL;
>>>>
>>>> + /* We are moving these outside of the spinlocks to avoid the warnings */
>>>> + mutex_init(&stp->st_mutex);
>>>> + mutex_lock(&stp->st_mutex);
>>>
>>> A mutex_init_locked() primitive might also be convenient here.
>>
>> I know! I would be able to do it under spinlock then without moving this
>> around too.
>>
>> But alas, not only there is not one, mutex documentation states this is
>> disallowed.
>
> You're just talking about this comment?:
>
> * It is not allowed to initialize an already locked mutex.
>
> That's a weird comment. You're proably right that what they meant was
> something like "It is not allowed to initialize a mutex to locked
> state". But, I don't know, taken literally that comment doesn't make
> sense (how could you even distinguish between an already-locked mutex
> and an uninitialized mutex?), so maybe it'd be worth asking.
I think this is because of the strict ownership tracking or something.
I guess I can ask.
>>> You could also take the two previous lines from the caller into this
>>> function instead of passing in stp, that might simplify the code.
>>> (Haven't checked.)
>>
>> I am not really sure what do you mean here.
>> These lines are moved from further away in this function )well, just the
>> init, anyway).
>>
>> Having half initialisation of stp here and half in the caller sounds kind of
>> strange
>> to me.
>
> I was thinking of something like the following--so init_open_stateid
> hides more of the details of the swapping. Untested. Does it look like
> an improvement to you?
>
> There's got to be a way to make this code a little less convoluted....
>
> --b.
>
> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> index fa5fb5aa4847..41b59854c40f 100644
> --- a/fs/nfsd/nfs4state.c
> +++ b/fs/nfsd/nfs4state.c
> @@ -3480,13 +3480,15 @@ alloc_init_open_stateowner(unsigned int strhashval,
> struct nfsd4_open *open,
> }
>
> static struct nfs4_ol_stateid *
> -init_open_stateid(struct nfs4_ol_stateid *stp, struct nfs4_file *fp,
> - struct nfsd4_open *open)
> +init_open_stateid(struct nfs4_file *fp, struct nfsd4_open *open)
> {
>
> struct nfs4_openowner *oo = open->op_openowner;
> struct nfs4_ol_stateid *retstp = NULL;
> + struct nfs4_ol_stateid *stp;
>
> + stp = open->op_stp;
> + open->op_stp = NULL;
> /* We are moving these outside of the spinlocks to avoid the warnings */
> mutex_init(&stp->st_mutex);
> mutex_lock(&stp->st_mutex);
> @@ -3512,9 +3514,12 @@ init_open_stateid(struct nfs4_ol_stateid *stp, struct
> nfs4_file *fp,
> out_unlock:
> spin_unlock(&fp->fi_lock);
> spin_unlock(&oo->oo_owner.so_client->cl_lock);
> - if (retstp)
> - mutex_lock(&retstp->st_mutex);
> - return retstp;
> + if (retstp) {
> + nfs4_put_stid(&stp->st_stid);
So as I am trying to integrate this into my patchset,
do we really need this?
We don't if we took the other path and left this one
hanging off the struct nfsd4_open (why do we need to
assign it NULL before the search?) I imagine then
we'd save some free/realloc churn as well?
I assume struct nfsd4_open cannot be shared between threads?
Otherwise we have bigger problems at hand like mutex init on a locked
mutex from another thread and stuff.
I'll try this theory I guess.
> + stp = retstp;
> + mutex_lock(&stp->st_mutex);
> + }
> + return stp;
> }
>
> /*
> @@ -4310,7 +4315,6 @@ nfsd4_process_open2(struct svc_rqst *rqstp, struct
> svc_fh *current_fh, struct nf
> struct nfs4_client *cl = open->op_openowner->oo_owner.so_client;
> struct nfs4_file *fp = NULL;
> struct nfs4_ol_stateid *stp = NULL;
> - struct nfs4_ol_stateid *swapstp = NULL;
> struct nfs4_delegation *dp = NULL;
> __be32 status;
>
> @@ -4347,16 +4351,9 @@ nfsd4_process_open2(struct svc_rqst *rqstp, struct
> svc_fh *current_fh, struct nf
> goto out;
> }
> } else {
> - stp = open->op_stp;
> - open->op_stp = NULL;
> - /*
> - * init_open_stateid() either returns a locked stateid
> - * it found, or initializes and locks the new one we passed in
> - */
> - swapstp = init_open_stateid(stp, fp, open);
> - if (swapstp) {
> - nfs4_put_stid(&stp->st_stid);
> - stp = swapstp;
> + /* stp is returned locked: */
> + stp = init_open_stateid(fp, open);
> + if (stp->st_access_bmap == 0) {
> status = nfs4_upgrade_open(rqstp, fp, current_fh,
> stp, open);
> if (status) {
