On Fri, Oct 18, 2019 at 5:21 PM David Wysochanski <dwysocha@xxxxxxxxxx> wrote:
>
> On Fri, Oct 18, 2019 at 4:59 PM Pavel Shilovskiy <pshilov@xxxxxxxxxxxxx> wrote:
> >
> > Thanks for the good news that the patch is stable in your workload!
> >
> The attached patch I ran on top of 5.4-rc3 for over 5 hrs today on the
> reboot test - before it would crash after a few minutes tops.
>
> > The extra flag may not be necessary and we may rely on a MID state but we would need to handle two states actually: MID_RETRY_NEEDED and MID_SHUTDOWN - see clean_demultiplex_info() which is doing the same things with mid as cifs_reconnect(). Please add ref counting to both functions since they both can race with system call threads.
> >
I agree that loop has the same problem. I can add that you're ok with
the mid_state approach. I think the only other option is probably a
flag like Ronnie suggested.
I will have to review the state machine more when I am more alert if
you are concerned about possible subtle regressions.
> > I also think that we need to create as smaller patch as possible to avoid hidden regressions. That's why I don't think we should change IF() to WARN_ON() in the same patch and keep it separately without the stable tag.
> >
> IMO that 'if' statement is wrong, and should be removed unless it can
> be defended. Why are we _conditionally_ setting the state to
> MID_RETRY_NEEDED in the same loop as we're putting mids on retry_list?
> What's the state machine supposed to be doing if it's ambiguous?
>
> > Another general thought is that including extra logic into the MID state may complicate the code. Having a flag like MID_QUEUED would reflect the meaning more straightforward: if mis is queued then de-queue it (aka remove it from the list), else - skip this step. This may be changed later if you think this will complicate the small stable patch.
> >
>
> You all know better than me. I'll take another look next week and
> look forward to more discussion.
>
> > --
> > Best regards,
> > Pavel Shilovsky
> >
> > -----Original Message-----
> > From: David Wysochanski <dwysocha@xxxxxxxxxx>
> > Sent: Friday, October 18, 2019 3:12 AM
> > To: Ronnie Sahlberg <lsahlber@xxxxxxxxxx>
> > Cc: Pavel Shilovskiy <pshilov@xxxxxxxxxxxxx>; linux-cifs <linux-cifs@xxxxxxxxxxxxxxx>; Frank Sorenson <sorenson@xxxxxxxxxx>
> > Subject: Re: list_del corruption while iterating retry_list in cifs_reconnect still seen on 5.4-rc3
> >
> > On Fri, Oct 18, 2019 at 5:27 AM Ronnie Sahlberg <lsahlber@xxxxxxxxxx> wrote:
> > >
> > >
> > > ----- Original Message -----
> > > > From: "David Wysochanski" <dwysocha@xxxxxxxxxx>
> > > > To: "Ronnie Sahlberg" <lsahlber@xxxxxxxxxx>
> > > > Cc: "Pavel Shilovskiy" <pshilov@xxxxxxxxxxxxx>, "linux-cifs" <linux-cifs@xxxxxxxxxxxxxxx>, "Frank Sorenson"
> > > > <sorenson@xxxxxxxxxx>
> > > > Sent: Friday, 18 October, 2019 6:16:45 PM
> > > > Subject: Re: list_del corruption while iterating retry_list in
> > > > cifs_reconnect still seen on 5.4-rc3
> > > >
> > > > On Thu, Oct 17, 2019 at 6:53 PM Ronnie Sahlberg <lsahlber@xxxxxxxxxx> wrote:
> > > > >
> > > > >
> > > > >
> > > > >
> > >
> > > Good comments.
> > > New version of the patch, please test and see comments inline below
> > >
> > > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c index
> > > bdea4b3e8005..8a78358693a5 100644
> > > --- a/fs/cifs/connect.c
> > > +++ b/fs/cifs/connect.c
> > > @@ -564,8 +564,13 @@ cifs_reconnect(struct TCP_Server_Info *server)
> > > spin_lock(&GlobalMid_Lock);
> > > list_for_each_safe(tmp, tmp2, &server->pending_mid_q) {
> > > mid_entry = list_entry(tmp, struct mid_q_entry, qhead);
> > > - if (mid_entry->mid_state == MID_REQUEST_SUBMITTED)
> > > - mid_entry->mid_state = MID_RETRY_NEEDED;
> > > + kref_get(&mid_entry->refcount);
> > > + WARN_ON(mid_entry->mid_state != MID_REQUEST_SUBMITTED);
> > > + /*
> > > + * Set MID_RETRY_NEEDED to prevent the demultiplex loop from
> > > + * removing us, or our neighbours, from the linked list.
> > > + */
> > > + mid_entry->mid_state = MID_RETRY_NEEDED;
> > > list_move(&mid_entry->qhead, &retry_list);
> > > }
> > > spin_unlock(&GlobalMid_Lock);
> > > @@ -575,7 +580,9 @@ cifs_reconnect(struct TCP_Server_Info *server)
> > > list_for_each_safe(tmp, tmp2, &retry_list) {
> > > mid_entry = list_entry(tmp, struct mid_q_entry, qhead);
> > > list_del_init(&mid_entry->qhead);
> > > +
> > > mid_entry->callback(mid_entry);
> > > + cifs_mid_q_entry_release(mid_entry);
> > > }
> > >
> > > if (cifs_rdma_enabled(server)) { @@ -895,7 +902,7 @@
> > > dequeue_mid(struct mid_q_entry *mid, bool malformed)
> > > if (mid->mid_flags & MID_DELETED)
> > > printk_once(KERN_WARNING
> > > "trying to dequeue a deleted mid\n");
> > > - else
> > > + else if (mid->mid_state != MID_RETRY_NEEDED)
> >
> > I'm just using an 'if' here not 'else if'. Do you see any issue with that?
> >
> > Actually this section needed a little of reorganizing due to the setting of the mid_state. So I have this now for this hunk:
> >
> > mid->when_received = jiffies;
> > #endif
> > spin_lock(&GlobalMid_Lock);
> > - if (!malformed)
> > - mid->mid_state = MID_RESPONSE_RECEIVED;
> > - else
> > - mid->mid_state = MID_RESPONSE_MALFORMED;
> > /*
> > * Trying to handle/dequeue a mid after the send_recv()
> > * function has finished processing it is a bug.
> > @@ -895,8 +893,14 @@ static inline int reconn_setup_dfs_targets(struct cifs_sb_info *cifs_sb,
> > if (mid->mid_flags & MID_DELETED)
> > printk_once(KERN_WARNING
> > "trying to dequeue a deleted mid\n");
> > - else
> > + if (mid->mid_state != MID_RETRY_NEEDED)
> > list_del_init(&mid->qhead);
> > +
> > + if (!malformed)
> > + mid->mid_state = MID_RESPONSE_RECEIVED;
> > + else
> > + mid->mid_state = MID_RESPONSE_MALFORMED;
> > +
> > spin_unlock(&GlobalMid_Lock);
> > }
> >
> >
> >
> > > list_del_init(&mid->qhead);
> > > spin_unlock(&GlobalMid_Lock);
> > > }
> > > diff --git a/fs/cifs/transport.c b/fs/cifs/transport.c index
> > > 308ad0f495e1..17a430b58673 100644
> > > --- a/fs/cifs/transport.c
> > > +++ b/fs/cifs/transport.c
> > > @@ -173,7 +173,8 @@ void
> > > cifs_delete_mid(struct mid_q_entry *mid) {
> > > spin_lock(&GlobalMid_Lock);
> > > - list_del_init(&mid->qhead);
> > > + if (mid->mid_state != MID_RETRY_NEEDED)
> > > + list_del_init(&mid->qhead);
> > > mid->mid_flags |= MID_DELETED;
> > > spin_unlock(&GlobalMid_Lock);
> > >
> > > @@ -872,7 +873,8 @@ cifs_sync_mid_result(struct mid_q_entry *mid, struct TCP_Server_Info *server)
> > > rc = -EHOSTDOWN;
> > > break;
> > > default:
> > > - list_del_init(&mid->qhead);
> > > + if (mid->mid_state != MID_RETRY_NEEDED)
> > > + list_del_init(&mid->qhead);
> > > cifs_server_dbg(VFS, "%s: invalid mid state mid=%llu state=%d\n",
> > > __func__, mid->mid, mid->mid_state);
> > > rc = -EIO;
> > >
> > >
> > >
> > >
> > >
> > >
> > > > >
> > > > > ----- Original Message -----
> > > > > > From: "Pavel Shilovskiy" <pshilov@xxxxxxxxxxxxx>
> > > > > > To: "Ronnie Sahlberg" <lsahlber@xxxxxxxxxx>, "David Wysochanski"
> > > > > > <dwysocha@xxxxxxxxxx>
> > > > > > Cc: "linux-cifs" <linux-cifs@xxxxxxxxxxxxxxx>, "Frank Sorenson"
> > > > > > <sorenson@xxxxxxxxxx>
> > > > > > Sent: Friday, 18 October, 2019 8:02:23 AM
> > > > > > Subject: RE: list_del corruption while iterating retry_list in
> > > > > > cifs_reconnect still seen on 5.4-rc3
> > > > > >
> > > > > > Ok, looking at cifs_delete_mid():
> > > > > >
> > > > > > 172 void
> > > > > > 173 cifs_delete_mid(struct mid_q_entry *mid)
> > > > > > 174 {
> > > > > > 175 >-------spin_lock(&GlobalMid_Lock);
> > > > > > 176 >-------list_del_init(&mid->qhead);
> > > > > > 177 >-------mid->mid_flags |= MID_DELETED;
> > > > > > 178 >-------spin_unlock(&GlobalMid_Lock);
> > > > > > 179
> > > > > > 180 >-------DeleteMidQEntry(mid);
> > > > > > 181 }
> > > > > >
> > > > > > So, regardless of us taking references on the mid itself or not,
> > > > > > the mid might be removed from the list. I also don't think
> > > > > > taking GlobalMid_Lock would help much because the next mid in
> > > > > > the list might be deleted from the list by another process while
> > > > > > cifs_reconnect is calling callback for the current mid.
> > > > > >
> > > >
> > > > Yes the above is consistent with my tracing the crash after the first
> > > > initial refcount patch was applied.
> > > > After the simple refcount patch, when iterating the retry_loop, it was
> > > > processing an orphaned list with a single item over and over and
> > > > eventually ran itself down to refcount == 0 and crashed like before.
> > > >
> > > >
> > > > > > Instead, shouldn't we try marking the mid as being reconnected? Once we
> > > > > > took
> > > > > > a reference, let's mark mid->mid_flags with a new flag MID_RECONNECT
> > > > > > under
> > > > > > the GlobalMid_Lock. Then modify cifs_delete_mid() to check for this flag
> > > > > > and
> > > > > > do not remove the mid from the list if the flag exists.
> > > > >
> > > > > That could work. But then we should also use that flag to suppress the
> > > > > other places where we do a list_del*, so something like this ?
> > > > >
> > > > > diff --git a/fs/cifs/cifsglob.h b/fs/cifs/cifsglob.h
> > > > > index 50dfd9049370..b324fff33e53 100644
> > > > > --- a/fs/cifs/cifsglob.h
> > > > > +++ b/fs/cifs/cifsglob.h
> > > > > @@ -1702,6 +1702,7 @@ static inline bool is_retryable_error(int error)
> > > > > /* Flags */
> > > > > #define MID_WAIT_CANCELLED 1 /* Cancelled while waiting for response
> > > > > */
> > > > > #define MID_DELETED 2 /* Mid has been dequeued/deleted */
> > > > > +#define MID_RECONNECT 4 /* Mid is being used during reconnect
> > > > > */
> > > > >
> > > > Do we need this extra flag? Can just use mid_state ==
> > > > MID_RETRY_NEEDED in the necessary places?
> > >
> > > That is a good point.
> > > It saves us a redundant flag.
> > >
> > > >
> > > >
> > > > > /* Types of response buffer returned from SendReceive2 */
> > > > > #define CIFS_NO_BUFFER 0 /* Response buffer not returned */
> > > > > diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
> > > > > index bdea4b3e8005..b142bd2a3ef5 100644
> > > > > --- a/fs/cifs/connect.c
> > > > > +++ b/fs/cifs/connect.c
> > > > > @@ -564,6 +564,8 @@ cifs_reconnect(struct TCP_Server_Info *server)
> > > > > spin_lock(&GlobalMid_Lock);
> > > > > list_for_each_safe(tmp, tmp2, &server->pending_mid_q) {
> > > > > mid_entry = list_entry(tmp, struct mid_q_entry, qhead);
> > > > > + kref_get(&mid_entry->refcount);
> > > > > + mid_entry->mid_flags |= MID_RECONNECT;
> > > > > if (mid_entry->mid_state == MID_REQUEST_SUBMITTED)
> > > > > mid_entry->mid_state = MID_RETRY_NEEDED;
> > > >
> > > > What happens if the state is wrong going in there, and it is not set
> > > > to MID_RETRY_NEEDED, but yet we queue up the retry_list and run it
> > > > below?
> > > > Should the above 'if' check for MID_REQUEST_SUBMITTED be a WARN_ON
> > > > followed by unconditionally setting the state?
> > > >
> > > > WARN_ON(mid_entry->mid_state != MID_REQUEST_SUBMITTED);
> > > > /* Unconditionally set MID_RETRY_NEEDED */
> > > > mid_etnry->mid_state = MID_RETRY_NEEDED;
> > >
> > > Yepp.
> > >
> > > >
> > > >
> > > > > list_move(&mid_entry->qhead, &retry_list);
> > > > > @@ -575,7 +577,9 @@ cifs_reconnect(struct TCP_Server_Info *server)
> > > > > list_for_each_safe(tmp, tmp2, &retry_list) {
> > > > > mid_entry = list_entry(tmp, struct mid_q_entry, qhead);
> > > > > list_del_init(&mid_entry->qhead);
> > > > > +
> > > > > mid_entry->callback(mid_entry);
> > > > > + cifs_mid_q_entry_release(mid_entry);
> > > > > }
> > > > >
> > > > > if (cifs_rdma_enabled(server)) {
> > > > > @@ -895,7 +899,7 @@ dequeue_mid(struct mid_q_entry *mid, bool malformed)
> > > > > if (mid->mid_flags & MID_DELETED)
> > > > > printk_once(KERN_WARNING
> > > > > "trying to dequeue a deleted mid\n");
> > > > > - else
> > > > > + else if (!(mid->mid_flags & MID_RECONNECT))
> > > >
> > > > Instead of the above,
> > > >
> > > > - else
> > > > + else if (mid_entry->mid_state == MID_RETRY_NEEDED)
> > >
> > > Yes, but mid_state != MID_RETRY_NEEDED
> > >
> >
> > Yeah good catch on that - somehow I reversed the logic, and when I
> > tested the former it blew up spectacularly almost instantaenously!
> > Doh!
> >
> > So far the latest patch has been running for about 25 minutes, which
> > is I think the longest this test has survived.
> > I need a bit more runtime to be sure it's good, but if it keeps going
> > I'll plan to create a patch header and submit to list by end of today.
> > Thanks Ronnie and Pavel for the help tracking this down.
> >
> >
> >
> >
> >
> >
