If anyone cares, a copy of this build is at:

http://huchra.bufferbloat.net/~cerowrt/cerowrt-wndr3700-dbg/

and the kernel is at:

http://huchra.bufferbloat.net/~cerowrt/vmlinux



On Fri, Jun 3, 2011 at 9:49 PM, Dave Taht <dave.taht@gmail.com> wrote:

> So I built a fresh build of openwrt from git head, patched in my
> preliminary patch to the switch, compiled it with oprofile support, and
> turned on -O2 rather than -Os, in my usual approach to all-up testing that
> makes it hard to track down what's really going on without backtracking.
>
> The net result with iperf was better than before,
> with iptables and qos turned entirely off, no nat, etc.
> pings stayed very flat and well below 8ms, down from 100ms.
>
> elara: [  3]  0.0-60.0 sec   435 MBytes  60.7 Mbits/sec
> io: [  3]  0.0-60.0 sec   495 MBytes  69.2 Mbits/sec
> leda: [  3]  0.0-60.9 sec   226 MBytes  31.1 Mbits/sec
> thebe: [  3]  0.0-60.0 sec   105 MBytes  14.7 Mbits/sec
>
> (how all this stuff is routed is beyond the scope of this, but only
> partially explains the discrepancies above)
>
> Oprofiling a later run, we're spending a lot of time in an unaligned trap,
> and I have no idea why iptables and conntrack even register.
>
> I am not sure at the moment how to track down the source of the alignment
> fault, I assume I can look at the performance counter somehow, and backtrack
> that to a given function or functions. It's late...
>
> samples  cum. samples  %        cum. %     app name                 symbol
> name
> 14992    14992         22.7158  22.7158    vmlinux                  do_ade
> 10054    25046         15.2338  37.9496    ip_tables
> /ip_tables
> 7257     32303         10.9958  48.9454    nf_conntrack
> /nf_conntrack
> 3398     35701          5.1486  54.0941    vmlinux
> handle_adel_int
> 2613     38314          3.9592  58.0533    vmlinux                  ip_rcv
> 2356     40670          3.5698  61.6231    vmlinux
> nf_iterate
> 1869     42539          2.8319  64.4550    iptable_nat
> /iptable_nat
> 1626     44165          2.4637  66.9187    vmlinux
> r4k_dma_cache_inv
> 1544     45709          2.3395  69.2582    nf_conntrack_ipv4
> /nf_conntrack_ipv4
> 1331     47040          2.0167  71.2749    vmlinux
> ag71xx_poll
> 1011     48051          1.5319  72.8068    vmlinux
> ip_route_input_common
> 988      49039          1.4970  74.3038    vmlinux
> ret_from_exception
>
>
>
> asmlinkage void do_ade(struct pt_regs *regs)
> {
>         unsigned int __user *pc;
>         mm_segment_t seg;
>
>         perf_sw_event(PERF_COUNT_SW_ALIGNMENT_FAULTS,
>                         1, 0, regs, regs->cp0_badvaddr);
>         /*
>          * Did we catch a fault trying to load an instruction?
>          * Or are we running in MIPS16 mode?
>          */
>         if ((regs->cp0_badvaddr == regs->cp0_epc) || (regs->cp0_epc & 0x1))
>                 goto sigbus;
>
>         pc = (unsigned int __user *) exception_epc(regs);
>         if (user_mode(regs) && !test_thread_flag(TIF_FIXADE))
>                 goto sigbus;
>         if (unaligned_action == UNALIGNED_ACTION_SIGNAL)
>                 goto sigbus;
>         else if (unaligned_action == UNALIGNED_ACTION_SHOW)
>                 show_registers(regs);
>
>         /*
>          * Do branch emulation only if we didn't forward the exception.
>          * This is all so but ugly ...
>          */
>         seg = get_fs();
>         if (!user_mode(regs))
>                 set_fs(KERNEL_DS);
>         emulate_load_store_insn(regs, (void __user *)regs->cp0_badvaddr,
> pc);
>         set_fs(seg);
>
>         return;
>
>
>
> --
> Dave Täht
> SKYPE: davetaht
> US Tel: 1-239-829-5608
> http://the-edge.blogspot.com
>



-- 
Dave Täht
SKYPE: davetaht
US Tel: 1-239-829-5608
http://the-edge.blogspot.com