On Tue, Mar 26, 2013 at 10:38:27AM +0800, Asias He wrote:
> On Mon, Mar 25, 2013 at 01:13:39PM +0200, Michael S. Tsirkin wrote:
> > On Mon, Mar 25, 2013 at 03:39:42PM +0800, Asias He wrote:
> > > On Sun, Mar 24, 2013 at 05:11:37PM +0200, Michael S. Tsirkin wrote:
> > > > On Fri, Mar 22, 2013 at 02:55:27PM +0800, Asias He wrote:
> > > > > This patch makes vhost_scsi_flush() wait for all the pending requests
> > > > > issued before the flush operation to be finished.
> > > > >
> > > > > Signed-off-by: Asias He <[email protected]>
> > > > > ---
> > > > > drivers/vhost/tcm_vhost.c | 117
> > > > > ++++++++++++++++++++++++++++++++++++++--------
> > > > > drivers/vhost/tcm_vhost.h | 4 ++
> > > > > 2 files changed, 101 insertions(+), 20 deletions(-)
> > > > >
> > > > > diff --git a/drivers/vhost/tcm_vhost.c b/drivers/vhost/tcm_vhost.c
> > > > > index e734ead..dc0af52 100644
> > > > > --- a/drivers/vhost/tcm_vhost.c
> > > > > +++ b/drivers/vhost/tcm_vhost.c
> > > > > @@ -82,6 +82,15 @@ struct vhost_scsi {
> > > > >
> > > > > bool vs_events_dropped; /* any missed events, protected by
> > > > > dev.mutex */
> > > > > u64 vs_events_nr; /* num of pending events, protected by
> > > > > dev.mutex */
> > > > > +
> > > > > + /*
> > > > > + * vs_inflight[0]/[1] are used to track requests issued
> > > > > + * before/during the flush operation
> > > > > + */
> > > > > + u64 vs_inflight[2];
> > > > > + wait_queue_head_t vs_flush_wait; /* wait queue for flush
> > > > > operation */
> > > > > + spinlock_t vs_flush_lock; /* lock to protect vs_during_flush */
> > > > > + int vs_during_flush; /* flag to indicate if we are in flush
> > > > > operation */
> > > >
> > > > So this adds a spinlock on data path and I'm not sure
> > > > I understand why this is correct (see also comment below).
> > >
> > > vs_flush_lock is accessed in:
> > >
> > > 1) tcm_vhost_allocate_evt() and tcm_vhost_free_evt()
> > > 2) vhost_scsi_allocate_cmd() and vhost_scsi_free_cmd().
> > >
> > > The former is not on data path. The later is always executed in the
> > > vhost thread. So we can almost always take the lock with no contention.
> > >
> > > And I am not seeing any real perf differences.
> > >
> > > > And generally we should try to avoid reimplementing refcounting. How
> > > > about we use a kref pointer instead? Access can use RCU (or maybe put
> > > > it in vq->private and use the tricky vhost version of RCU). Initialize
> > > > it to 1. To flush you replace the pointer, decrement then wait for
> > > > refcount to reach 0.
> > >
> > > This makes the code even more tricky and hard to understand. I am not
> > > sure this is a place where kref is the right choice.
> >
> > Point is, this homegrown reference-counting implementation seems to be
> > buggy, see the comment below. And it is doing flushes which is similar
> > to RCU anyway. Direct use of RCU will at least be well documented.
>
> If we do full flush. We do no bothering the complex at all.
Hmm parse error. What do you mean?
> > > > This still adds atomics on data path so maybe worth benchmarking to
> > > > verify performance overhead is not measureable, but at least it's
> > > > one atomic and not a full lock.
> > > > Hmm?
> > > >
> > > >
> > > > > };
> > > > >
> > > > > /* Local pointer to allocated TCM configfs fabric module */
> > > > > @@ -99,6 +108,46 @@ static int iov_num_pages(struct iovec *iov)
> > > > > ((unsigned long)iov->iov_base & PAGE_MASK)) >>
> > > > > PAGE_SHIFT;
> > > > > }
> > > > >
> > > > > +static int tcm_vhost_inc_inflight(struct vhost_scsi *vs)
> > > > > +{
> > > > > + int during_flush;
> > > > > +
> > > > > + spin_lock(&vs->vs_flush_lock);
> > > > > + during_flush = vs->vs_during_flush;
> > > > > + vs->vs_inflight[during_flush]++;
> > > > > + spin_unlock(&vs->vs_flush_lock);
> > > > > +
> > > > > + return during_flush;
> > > > > +}
> > > > > +
> > > > > +static void tcm_vhost_dec_inflight(struct vhost_scsi *vs, int
> > > > > during_flush)
> > > > > +{
> > > > > + u64 inflight;
> > > > > +
> > > > > + spin_lock(&vs->vs_flush_lock);
> > > > > + inflight = vs->vs_inflight[during_flush]--;
> > > > > + /*
> > > > > + * Wakeup the waiter when all the requests issued before the
> > > > > flush
> > > > > + * operation are finished and we are during the flush operation.
> > > > > + */
> > > > > + if (!inflight && !during_flush && vs->vs_during_flush)
> > > > > + wake_up(&vs->vs_flush_wait);
> > > > > + spin_unlock(&vs->vs_flush_lock);
> > > > > +}
> > > > > +
> > > > > +static bool tcm_vhost_done_inflight(struct vhost_scsi *vs)
> > > > > +{
> > > > > + bool ret = false;
> > > > > +
> > > > > + /* The requests issued before the flush operation are finished
> > > > > ? */
> > > > > + spin_lock(&vs->vs_flush_lock);
> > > > > + if (!vs->vs_inflight[0])
> > > > > + ret = true;
> > > > > + spin_unlock(&vs->vs_flush_lock);
> > > > > +
> > > > > + return ret;
> > > > > +}
> > > > > +
> > > > > static bool tcm_vhost_check_feature(struct vhost_scsi *vs, int
> > > > > feature)
> > > > > {
> > > > > bool ret = false;
> > > > > @@ -391,6 +440,8 @@ static int tcm_vhost_queue_tm_rsp(struct se_cmd
> > > > > *se_cmd)
> > > > >
> > > > > static void tcm_vhost_free_evt(struct vhost_scsi *vs, struct
> > > > > tcm_vhost_evt *evt)
> > > > > {
> > > > > +
> > > > > + tcm_vhost_dec_inflight(vs, evt->during_flush);
> > > > > mutex_lock(&vs->dev.mutex);
> > > > > vs->vs_events_nr--;
> > > > > kfree(evt);
> > > > > @@ -412,6 +463,7 @@ static struct tcm_vhost_evt
> > > > > *tcm_vhost_allocate_evt(struct vhost_scsi *vs,
> > > > > if (evt) {
> > > > > evt->event.event = event;
> > > > > evt->event.reason = reason;
> > > > > + evt->during_flush = tcm_vhost_inc_inflight(vs);
> > > > > vs->vs_events_nr++;
> > > > > }
> > > > > mutex_unlock(&vs->dev.mutex);
> > > > > @@ -422,6 +474,7 @@ static struct tcm_vhost_evt
> > > > > *tcm_vhost_allocate_evt(struct vhost_scsi *vs,
> > > > > static void vhost_scsi_free_cmd(struct tcm_vhost_cmd *tv_cmd)
> > > > > {
> > > > > struct se_cmd *se_cmd = &tv_cmd->tvc_se_cmd;
> > > > > + struct vhost_scsi *vs = tv_cmd->tvc_vhost;
> > > > >
> > > > > /* TODO locking against target/backend threads? */
> > > > > transport_generic_free_cmd(se_cmd, 1);
> > > > > @@ -434,13 +487,16 @@ static void vhost_scsi_free_cmd(struct
> > > > > tcm_vhost_cmd *tv_cmd)
> > > > > kfree(tv_cmd->tvc_sgl);
> > > > > }
> > > > >
> > > > > + tcm_vhost_dec_inflight(vs, tv_cmd->during_flush);
> > > > > +
> > > > > kfree(tv_cmd);
> > > > > }
> > > > >
> > > > > static void tcm_vhost_do_evt_work(struct vhost_scsi *vs,
> > > > > - struct virtio_scsi_event *event)
> > > > > + struct tcm_vhost_evt *evt)
> > > > > {
> > > > > struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT];
> > > > > + struct virtio_scsi_event *event = &evt->event;
> > > > > struct virtio_scsi_event __user *eventp;
> > > > > unsigned out, in;
> > > > > int head, ret;
> > > > > @@ -503,7 +559,7 @@ static void tcm_vhost_evt_work(struct vhost_work
> > > > > *work)
> > > > > while (llnode) {
> > > > > evt = llist_entry(llnode, struct tcm_vhost_evt, list);
> > > > > llnode = llist_next(llnode);
> > > > > - tcm_vhost_do_evt_work(vs, &evt->event);
> > > > > + tcm_vhost_do_evt_work(vs, evt);
> > > > > tcm_vhost_free_evt(vs, evt);
> > > > > }
> > > > > }
> > > > > @@ -521,8 +577,8 @@ static void vhost_scsi_complete_cmd_work(struct
> > > > > vhost_work *work)
> > > > > struct virtio_scsi_cmd_resp v_rsp;
> > > > > struct tcm_vhost_cmd *tv_cmd;
> > > > > struct llist_node *llnode;
> > > > > - struct se_cmd *se_cmd;
> > > > > int ret, vq;
> > > > > + struct se_cmd *se_cmd;
> > > > >
> > > > > bitmap_zero(signal, VHOST_SCSI_MAX_VQ);
> > > > > llnode = llist_del_all(&vs->vs_completion_list);
> > > > > @@ -560,6 +616,7 @@ static void vhost_scsi_complete_cmd_work(struct
> > > > > vhost_work *work)
> > > > > }
> > > > >
> > > > > static struct tcm_vhost_cmd *vhost_scsi_allocate_cmd(
> > > > > + struct vhost_scsi *vs,
> > > > > struct tcm_vhost_tpg *tv_tpg,
> > > > > struct virtio_scsi_cmd_req *v_req,
> > > > > u32 exp_data_len,
> > > > > @@ -584,6 +641,8 @@ static struct tcm_vhost_cmd
> > > > > *vhost_scsi_allocate_cmd(
> > > > > tv_cmd->tvc_exp_data_len = exp_data_len;
> > > > > tv_cmd->tvc_data_direction = data_direction;
> > > > > tv_cmd->tvc_nexus = tv_nexus;
> > > > > + tv_cmd->tvc_vhost = vs;
> > > > > + tv_cmd->during_flush = tcm_vhost_inc_inflight(vs);
> > > > >
> > > > > return tv_cmd;
> > > > > }
> > > > > @@ -824,7 +883,7 @@ static void vhost_scsi_handle_vq(struct
> > > > > vhost_scsi *vs,
> > > > > for (i = 0; i < data_num; i++)
> > > > > exp_data_len += vq->iov[data_first + i].iov_len;
> > > > >
> > > > > - tv_cmd = vhost_scsi_allocate_cmd(tv_tpg, &v_req,
> > > > > + tv_cmd = vhost_scsi_allocate_cmd(vs, tv_tpg, &v_req,
> > > > > exp_data_len, data_direction);
> > > > > if (IS_ERR(tv_cmd)) {
> > > > > vq_err(vq, "vhost_scsi_allocate_cmd failed
> > > > > %ld\n",
> > > > > @@ -834,7 +893,6 @@ static void vhost_scsi_handle_vq(struct
> > > > > vhost_scsi *vs,
> > > > > pr_debug("Allocated tv_cmd: %p exp_data_len: %d,
> > > > > data_direction"
> > > > > ": %d\n", tv_cmd, exp_data_len, data_direction);
> > > > >
> > > > > - tv_cmd->tvc_vhost = vs;
> > > > > tv_cmd->tvc_vq = vq;
> > > > >
> > > > > if (unlikely(vq->iov[out].iov_len !=
> > > > > @@ -887,6 +945,7 @@ static void vhost_scsi_handle_vq(struct
> > > > > vhost_scsi *vs,
> > > > > * tcm_vhost_queue_data_in() and
> > > > > tcm_vhost_queue_status()
> > > > > */
> > > > > tv_cmd->tvc_vq_desc = head;
> > > > > +
> > > > > /*
> > > > > * Dispatch tv_cmd descriptor for cmwq execution in
> > > > > process
> > > > > * context provided by tcm_vhost_workqueue. This also
> > > > > ensures
> > > > > @@ -956,6 +1015,34 @@ static void vhost_scsi_handle_kick(struct
> > > > > vhost_work *work)
> > > > > vhost_scsi_handle_vq(vs, vq);
> > > > > }
> > > > >
> > > > > +static void vhost_scsi_flush_vq(struct vhost_scsi *vs, int index)
> > > > > +{
> > > > > + vhost_poll_flush(&vs->dev.vqs[index].poll);
> > > > > +}
> > > > > +
> > > > > +static void vhost_scsi_flush(struct vhost_scsi *vs)
> > > > > +{
> > > > > + int i;
> > > > > +
> > > > > + /* Flush operation is started */
> > > > > + spin_lock(&vs->vs_flush_lock);
> > > > > + vs->vs_during_flush = 1;
> > > > > + spin_unlock(&vs->vs_flush_lock);
> > > > > +
> > > > > + for (i = 0; i < VHOST_SCSI_MAX_VQ; i++)
> > > > > + vhost_scsi_flush_vq(vs, i);
> > > > > + vhost_work_flush(&vs->dev, &vs->vs_completion_work);
> > > > > + vhost_work_flush(&vs->dev, &vs->vs_event_work);
> > > > > +
> > > > > + /* Wait until all requests issued before the flush to be
> > > > > finished */
> > > > > + wait_event(vs->vs_flush_wait, tcm_vhost_done_inflight(vs));
> > > > > +
> > > > > + /* Flush operation is finished */
> > > > > + spin_lock(&vs->vs_flush_lock);
> > > > > + vs->vs_during_flush = 0;
> > > > > + spin_unlock(&vs->vs_flush_lock);
> > > > > +}
> > > > > +
> > > >
> > > > Weird.
> > > > What about requests issued during flush operation?
>
> We can do:
>
> 1) flush all the requests (before and during the flush operation)
> 2) flush the requests before the flush operation.
>
> 1) might take unbounded time to finish the flushes. You suggested me not
> to do this when I was doing it for vhost-blk. But it is simpler to
> implement. The before and during machinery is gone.
>
Right. But classifying requests to ones submitted during flush
and ones submitted before also might not work well.
> > > > When next flush starts shouldn't we flush them?
>
> Yes. we flush them.
Hmm, I don't see where. They are counted in s->vs_inflight[1], while
tcm_vhost_done_inflight only checks s->vs_inflight[0].
> > Missed this comment?
>
> Yes.
>
> > > > > /*
> > > > > * Called from vhost_scsi_ioctl() context to walk the list of
> > > > > available
> > > > > * tcm_vhost_tpg with an active struct tcm_vhost_nexus
> > > > > @@ -1042,6 +1129,7 @@ static int vhost_scsi_clear_endpoint(
> > > > > u8 target;
> > > > >
> > > > > mutex_lock(&vs->dev.mutex);
> > > > > +
> > > > > /* Verify that ring has been setup correctly. */
> > > > > for (index = 0; index < vs->dev.nvqs; ++index) {
> > > > > if (!vhost_vq_access_ok(&vs->vqs[index])) {
> > > > > @@ -1109,6 +1197,10 @@ static int vhost_scsi_open(struct inode
> > > > > *inode, struct file *f)
> > > > > vhost_work_init(&s->vs_event_work, tcm_vhost_evt_work);
> > > > >
> > > > > s->vs_events_nr = 0;
> > > > > + s->vs_inflight[0] = 0;
> > > > > + s->vs_inflight[1] = 0;
> > > > > + spin_lock_init(&s->vs_flush_lock);
> > > > > + init_waitqueue_head(&s->vs_flush_wait);
> > > > >
> > > > > s->vqs[VHOST_SCSI_VQ_CTL].handle_kick =
> > > > > vhost_scsi_ctl_handle_kick;
> > > > > s->vqs[VHOST_SCSI_VQ_EVT].handle_kick =
> > > > > vhost_scsi_evt_handle_kick;
> > > > > @@ -1139,21 +1231,6 @@ static int vhost_scsi_release(struct inode
> > > > > *inode, struct file *f)
> > > > > return 0;
> > > > > }
> > > > >
> > > > > -static void vhost_scsi_flush_vq(struct vhost_scsi *vs, int index)
> > > > > -{
> > > > > - vhost_poll_flush(&vs->dev.vqs[index].poll);
> > > > > -}
> > > > > -
> > > > > -static void vhost_scsi_flush(struct vhost_scsi *vs)
> > > > > -{
> > > > > - int i;
> > > > > -
> > > > > - for (i = 0; i < VHOST_SCSI_MAX_VQ; i++)
> > > > > - vhost_scsi_flush_vq(vs, i);
> > > > > - vhost_work_flush(&vs->dev, &vs->vs_completion_work);
> > > > > - vhost_work_flush(&vs->dev, &vs->vs_event_work);
> > > > > -}
> > > > > -
> > > > > static int vhost_scsi_set_features(struct vhost_scsi *vs, u64
> > > > > features)
> > > > > {
> > > > > if (features & ~VHOST_SCSI_FEATURES)
> > > > > diff --git a/drivers/vhost/tcm_vhost.h b/drivers/vhost/tcm_vhost.h
> > > > > index 94e9ee53..dd84622 100644
> > > > > --- a/drivers/vhost/tcm_vhost.h
> > > > > +++ b/drivers/vhost/tcm_vhost.h
> > > > > @@ -37,6 +37,8 @@ struct tcm_vhost_cmd {
> > > > > unsigned char tvc_sense_buf[TRANSPORT_SENSE_BUFFER];
> > > > > /* Completed commands list, serviced from vhost worker thread */
> > > > > struct llist_node tvc_completion_list;
> > > > > + /* Indicate this command is issued during the flush operaton */
> > > > > + int during_flush;
> > > > > };
> > > > >
> > > > > struct tcm_vhost_nexus {
> > > > > @@ -91,6 +93,8 @@ struct tcm_vhost_evt {
> > > > > struct virtio_scsi_event event;
> > > > > /* virtio_scsi event list, serviced from vhost worker thread */
> > > > > struct llist_node list;
> > > > > + /* Indicate this event is issued during the flush operaton */
> > > > > + int during_flush;
> > > > > };
> > > > >
> > > > > /*
> > > > > --
> > > > > 1.8.1.4
> > >
> > > --
> > > Asias
>
> --
> Asias
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
