Skip site navigation (1)Skip section navigation (2) 
Date:      Tue, 26 Apr 2016 14:49:01 +0200
From:      Sebastian Huber <sebastian.huber@embedded-brains.de>
To:        freebsd-hackers@freebsd.org
Subject:   UMA alloc with max items < potential items in per processor buckets
Message-ID:  <571F63BD.1090001@embedded-brains.de>
next in thread | raw e-mail | index | archive | help 
Hello,

I don't use FreeBSD directly, but instead a port of the FreeBSD network=20
stack to RTEMS. So, this problem may not apply to FreeBSD. I observed=20
the following problem during an UDP socket create. They are allocated=20
from the following zone:

void
udp_init(void)
{

     in_pcbinfo_init(&V_udbinfo, "udp", &V_udb, UDBHASHSIZE, UDBHASHSIZE,
         "udp_inpcb", udp_inpcb_init, NULL, UMA_ZONE_NOFREE,
         IPI_HASHFIELDS_2TUPLE);
     V_udpcb_zone =3D uma_zcreate("udpcb", sizeof(struct udpcb),
         NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
     uma_zone_set_max(V_udpcb_zone, maxsockets);
     EVENTHANDLER_REGISTER(maxsockets_change, udp_zone_change, NULL,
         EVENTHANDLER_PRI_ANY);
}

In my setup maxsockets is 32. This is probably artificially small=20
compared to a real FreeBSD machine. The system has 24 processors, so we=20
need 128 * 24 items for the per processor cache buckets. This is=20
considerably lager than the single keg of the zone can deliver. Thus in=20
case a processor without a per processor bucket tries to do a=20
uma_zalloc(V_udpcb_zone, M_NOWAIT | M_ZERO), then it will get no item if=20
other processors already consumed all items for their per processor=20
cache buckets. I adjusted the uma_zone_set_max() like this

diff --git a/freebsd/sys/vm/uma_core.c b/freebsd/sys/vm/uma_core.c
index 845c433..f88d3d3 100644
--- a/freebsd/sys/vm/uma_core.c
+++ b/freebsd/sys/vm/uma_core.c
@@ -2903,6 +2903,10 @@ uma_zone_set_max(uma_zone_t zone, int nitems)
         uma_keg_t keg;

         ZONE_LOCK(zone);
+#ifdef SMP
+       if (nitems < zone->uz_count * mp_maxid)
+               nitems =3D zone->uz_count * mp_maxid;
+#endif
         keg =3D zone_first_keg(zone);
         keg->uk_maxpages =3D (nitems / keg->uk_ipers) * keg->uk_ppera;
         if (keg->uk_maxpages * keg->uk_ipers < nitems)

and this seems to fix the problem. Is there an implicit assumption in=20
FreeBSD that the zone max is always large enough to provide items for=20
the per processor cache buckets?

--=20
Sebastian Huber, embedded brains GmbH

Address : Dornierstr. 4, D-82178 Puchheim, Germany
Phone   : +49 89 189 47 41-16
Fax     : +49 89 189 47 41-09
E-Mail  : sebastian.huber@embedded-brains.de
PGP     : Public key available on request.

Diese Nachricht ist keine gesch=C3=A4ftliche Mitteilung im Sinne des EHUG=
.

Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?571F63BD.1090001>