Re: [PATCH 6/6] xive: Don't use CPU_FOREACH() to perform CAM line matching

[David Gibson]

On Thu, Oct 24, 2019 at 02:33:27PM +0200, Greg Kurz wrote:
> On Thu, 24 Oct 2019 14:05:36 +1100
> David Gibson <address@hidden> wrote:
> 
> > On Wed, Oct 23, 2019 at 04:52:27PM +0200, Greg Kurz wrote:
> > > Now that the TCTX objects are children of the XIVE router, stop
> > > using CPU_FOREACH() when looking for a matching VCPU target.
> > > 
> > > Signed-off-by: Greg Kurz <address@hidden>
> > > ---
> > >  hw/intc/xive.c |  100 
> > > +++++++++++++++++++++++++++++++++++---------------------
> > >  1 file changed, 62 insertions(+), 38 deletions(-)
> > > 
> > > diff --git a/hw/intc/xive.c b/hw/intc/xive.c
> > > index 40ce43152456..ec5e7d0ee39a 100644
> > > --- a/hw/intc/xive.c
> > > +++ b/hw/intc/xive.c
> > > @@ -1403,55 +1403,79 @@ typedef struct XiveTCTXMatch {
> > >      uint8_t ring;
> > >  } XiveTCTXMatch;
> > >  
> > > -static bool xive_presenter_match(XiveRouter *xrtr, uint8_t format,
> > > -                                 uint8_t nvt_blk, uint32_t nvt_idx,
> > > -                                 bool cam_ignore, uint8_t priority,
> > > -                                 uint32_t logic_serv, XiveTCTXMatch 
> > > *match)
> > > +typedef struct XivePresenterMatch {
> > > +    uint8_t format;
> > > +    uint8_t nvt_blk;
> > > +    uint32_t nvt_idx;
> > > +    bool cam_ignore;
> > > +    uint8_t priority;
> > > +    uint32_t logic_serv;
> > > +    XiveTCTXMatch *match;
> > > +    int count;
> > > +} XivePresenterMatch;
> > > +
> > > +static int do_xive_presenter_match(Object *child, void *opaque)
> > >  {
> > > -    CPUState *cs;
> > > +    XiveTCTX *tctx = XIVE_TCTX(child);
> > > +    XivePresenterMatch *xpm = opaque;
> > > +    int ring;
> > >  
> > >      /*
> > >       * TODO (PowerNV): handle chip_id overwrite of block field for
> > >       * hardwired CAM compares
> > >       */
> > >  
> > > -    CPU_FOREACH(cs) {
> > > -        XiveTCTX *tctx = xive_router_get_tctx(xrtr, cs);
> > > -        int ring;
> > > +    /*
> > > +     * HW checks that the CPU is enabled in the Physical Thread
> > > +     * Enable Register (PTER).
> > > +     */
> > >  
> > > -        /*
> > > -         * Skip partially initialized vCPUs. This can happen when
> > > -         * vCPUs are hotplugged.
> > > -         */
> > > -        if (!tctx) {
> > > -            continue;
> > > +    /*
> > > +     * Check the thread context CAM lines and record matches. We
> > > +     * will handle CPU exception delivery later
> > > +     */
> > > +    ring = xive_presenter_tctx_match(tctx, xpm->format, xpm->nvt_blk,
> > > +                                     xpm->nvt_idx, xpm->cam_ignore,
> > > +                                     xpm->logic_serv);
> > > +
> > > +    /*
> > > +     * Save the context and follow on to catch duplicates, that we
> > > +     * don't support yet.
> > > +     */
> > > +    if (ring != -1) {
> > > +        if (xpm->match->tctx) {
> > > +            qemu_log_mask(LOG_GUEST_ERROR, "XIVE: already found a thread 
> > > "
> > > +                          "context NVT %x/%x\n", xpm->nvt_blk, 
> > > xpm->nvt_idx);
> > > +            return -1;
> > >          }
> > >  
> > > -        /*
> > > -         * HW checks that the CPU is enabled in the Physical Thread
> > > -         * Enable Register (PTER).
> > > -         */
> > > +        xpm->match->ring = ring;
> > > +        xpm->match->tctx = tctx;
> > > +        xpm->count++;
> > > +    }
> > >  
> > > -        /*
> > > -         * Check the thread context CAM lines and record matches. We
> > > -         * will handle CPU exception delivery later
> > > -         */
> > > -        ring = xive_presenter_tctx_match(tctx, format, nvt_blk, nvt_idx,
> > > -                                         cam_ignore, logic_serv);
> > > -        /*
> > > -         * Save the context and follow on to catch duplicates, that we
> > > -         * don't support yet.
> > > -         */
> > > -        if (ring != -1) {
> > > -            if (match->tctx) {
> > > -                qemu_log_mask(LOG_GUEST_ERROR, "XIVE: already found a 
> > > thread "
> > > -                              "context NVT %x/%x\n", nvt_blk, nvt_idx);
> > > -                return false;
> > > -            }
> > > -
> > > -            match->ring = ring;
> > > -            match->tctx = tctx;
> > > -        }
> > > +    return 0;
> > > +}
> > > +
> > > +static bool xive_presenter_match(XiveRouter *xrtr, uint8_t format,
> > > +                                 uint8_t nvt_blk, uint32_t nvt_idx,
> > > +                                 bool cam_ignore, uint8_t priority,
> > > +                                 uint32_t logic_serv, XiveTCTXMatch 
> > > *match)
> > > +{
> > > +    XivePresenterMatch xpm = {
> > > +        .format     = format,
> > > +        .nvt_blk    = nvt_blk,
> > > +        .nvt_idx    = nvt_idx,
> > > +        .cam_ignore = cam_ignore,
> > > +        .priority   = priority,
> > > +        .logic_serv = logic_serv,
> > > +        .match      = match,
> > > +        .count      = 0,
> > > +    };
> > > +
> > > +    if (object_child_foreach_type(OBJECT(xrtr), TYPE_XIVE_TCTX,
> > > +                                  do_xive_presenter_match, &xpm) < 0) {
> > > +        return false;
> > 
> > Hrm... xive_presenter_match() is potentially a pretty hot path, it's
> > called on every interrupt delivery - especially since we don't have a
> > usable KVM irq chip for Boston machines.  I'm concerned that using
> > something as heavyweight as object_child_foreach() might have a
> > noticeable performance impact.
> > 
> 
> Well, the XiveRouter _only_ has 3 extra children (XiveSource,
> XiveENDSource and TIMA) but indeed object_child_foreach() might
> cost more than QTAILQ_FOREACH_RCU().

Right, it's not so much the redundant children, but whatever we have
to do to delve into the children data structure I'm concerned about
here.

> A possible option could be
> to have a QTAILQ of presenters under the machine for sPAPR or
> under the chip for PNV, in order to avoid the need to filter out
> VCPUs that we don't want to consider, ie. partly realized with
> sPAPR and from another chip with PNV.

If we can manage it, I'd actually suggest a plain old array/vector of
them, rather than a QTAILQ.

> But as said in another mail, the safer for 4.2 is probably to
> fix the CPU_FOREACH() users, which is already the case here for
> sPAPR.
> 
> > >      }
> > >  
> > >      if (!match->tctx) {
> > > 
> > 
> 



-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature