Bug#733826: crazy loop "xhci_hcd Too many fragments"
On Mon, Jan 06, 2014 at 03:52:24PM +0000, David Laight wrote:
> > From: Alan Stern
> > Subject: Re: Bug#733826: crazy loop "xhci_hcd Too many fragments"
> >
> > On Mon, 6 Jan 2014, Ben Hutchings wrote:
> >
> > > On Sat, 2014-01-04 at 05:44 +0800, jidanni@jidanni.org wrote:
> ...
> > > > # cat /var/log/syslog
> > > >
> > > > Jan 1 06:57:38 jidanni5 ntpd[2822]: Listen normally on 5 lo ::1 UDP 123
> > > > Jan 1 06:57:38 jidanni5 ntpd[2822]: Listen normally on 6 eth0 fe80::2289:84ff:fe28:ad9 UDP 123
> > > > Jan 1 06:57:38 jidanni5 ntpd[2822]: peers refreshed
> > > > Jan 1 06:57:38 jidanni5 ntpd[2822]: Listening on routing socket on fd #23 for interface updates
> > > > Jan 1 07:04:49 jidanni5 kernel: [ 559.624680] xhci_hcd 0000:00:14.0: Too many fragments 79, max
> > 63
> > > > Jan 1 07:04:49 jidanni5 kernel: [ 559.624695] xhci_hcd 0000:00:14.0: Too many fragments 79, max
> > 63
> > > > Jan 1 07:04:49 jidanni5 kernel: [ 559.624704] xhci_hcd 0000:00:14.0: Too many fragments 79, max
> > 63
> > > >
> > > > 100000 lines later... oops I mean an actual MILLION lines later
> > >
> > > Assuming my fix for the repetition is correct, the remaining problem is
> > > why usb-storage is generating such large/fragmented urbs.
> >
> > usb-storage doesn't generate large or fragmented anything. It merely
> > passes on the scatter-gather information it gets from the block layer.
>
> Although not a real fix to the underlying problem, it seems that the default
> ring size is far too small.
Did you mean ring segment size?
> Any amount of network traffic also activates the ring expansion code.
> IIRC each ring entry is 16 bytes, so increasing the ring size to 256
> still keeps the rings to a single 4k page.
>
> Whether anything regularly exceeds 255 fragments is a another matter.
If so, yes, changing the segment size makes sense. TRBS_PER_SEGMENT
could be increased to 256. I'm not sure if we should switch to using
dma_alloc_coherent instead of a DMA pool. Some systems could be using
bigger than 4K pages, so we should probably still stick with DMA pools.
Ben, can you change your patch to increase TRBS_PER_SEGMENT to 256?
Sarah Sharp
Reply to: