svn commit: r278977 - in head/sys: dev/cxgb dev/cxgb/sys dev/cxgb/ulp/tom dev/xen/netfront sys

[Navdeep Parhar]

On 02/19/15 12:02, Gleb Smirnoff wrote:
> On Thu, Feb 19, 2015 at 07:47:18PM +0100, Roger Pau Monné wrote:
> R> El 19/02/15 a les 2.19, Gleb Smirnoff ha escrit:
> R> > Author: glebius
> R> > Date: Thu Feb 19 01:19:42 2015
> R> > New Revision: 278977
> R> > URL: https://svnweb.freebsd.org/changeset/base/278977
> R> >
> R> > Log:
> R> >   Provide a set of inline functions to manage simple mbuf(9) queues, based
> R> >   on queue(3)'s STAILQ.  Utilize them in cxgb(4) and Xen, deleting home
> R> >   grown implementations.
> R> >
> R> >   Sponsored by:	Netflix
> R> >   Sponsored by:	Nginx, Inc.
> R>
> R> Have you tested this commit on Xen? I'm getting the following:
> R>
> R> xn0: <Virtual Network Interface> at device/vif/0 on xenbusb_front0
> R> xn0: Ethernet address: 00:16:3e:51:85:e3
> R> xenbusb_back0: <Xen Backend Devices> on xenstore0
> R> xbd0: Back-end specified ring-pages of 15 is not a power of 2. Limited to 8.
> R> xn0: backend features: feature-sg feature-gso-tcp4
> R> panic: no mbufs processed
> R> cpuid = 0
> R> KDB: stack backtrace:
> R> db_trace_self_wrapper() at db_trace_self_wrapper+0x2b/frame 0xfffffe007adc3920
> R> vpanic() at vpanic+0x189/frame 0xfffffe007adc39a0
> R> kassert_panic() at kassert_panic+0x132/frame 0xfffffe007adc3a10
> R> network_alloc_rx_buffers() at network_alloc_rx_buffers+0x439/frame 0xfffffe007adc3ac0
> R> network_connect() at network_connect+0xac1/frame 0xfffffe007adc3b50
> R> netfront_backend_changed() at netfront_backend_changed+0xed/frame 0xfffffe007adc3b90
> R> xenwatch_thread() at xenwatch_thread+0x1a2/frame 0xfffffe007adc3bb0
> R> fork_exit() at fork_exit+0x84/frame 0xfffffe007adc3bf0
> R> fork_trampoline() at fork_trampoline+0xe/frame 0xfffffe007adc3bf0
> R> --- trap 0, rip = 0, rsp = 0xfffffe007adc3cb0, rbp = 0 ---
> R> KDB: enter: panic
> R> [ thread pid 15 tid 100038 ]
> R> Stopped at      kdb_enter+0x3e: movq    $0,kdb_why
>
> I guess the problem is that the queue limit isn't initialized. Please
> try the attached patch.
>
> The problem with older mbufq was that it doesn't have any sane limit
> on it. So, converting to new, I simply put INT_MAX, which is ugly.

Is mq_len supposed to count the number of mbufs in the mbufq?   Then it 
doesn't work for m_next linked chains.  It pretends to enforce a cap on 
the number of mbufs in the mbufq while doing no such thing.  If it's not 
trying to count the number of mbufs then I'd like to know exactly what 
it does?

Regards,
Navdeep

>
> I'd appreciate if you fix mbufq_init()s in netfront.c to some appropriate
> values.
>