Re: [PATCH v2] tcp: input header length, prediction, and timestampbugs
From: William Allen Simpson
Date: Tue Jan 19 2010 - 12:35:15 EST
Over the weekend, I've been reviewing the .lst output that Andi
explained, and I've found a few interesting things.
1) The 4.4 compiler isn't very smart about shifts:
static inline unsigned int tcp_header_len_th(const struct tcphdr *th)
return th->doff * 4;
c04ea93e: 0f b6 47 0c movzbl 0xc(%edi),%eax
c04ea942: c0 e8 04 shr $0x4,%al
c04ea945: 0f b6 c0 movzbl %al,%eax
c04ea948: c1 e0 02 shl $0x2,%eax
That could easily be done as shr $0x2 instead.
2) This "fast path" code is under quite a bit of register pressure on
the 32-bit i386. There's a lot of saving and re-loading.
3) Particularly, the seldom used *th and len parameters are saved and
reloaded in the very beginning of the fast path, really wasting time.
4) Since both *th and len are actually indexed loads from *skb (which
the compiler keeps in a register most of the time), doing indexed loads
from the stack (%ebp) is the same, so they shouldn't be sent as
parameters at all!
5) There's already code, added back in 2006 for DMA, that directly
references skb->len instead of the len parameter. Probably lack of
header documentation, so the coder failed to notice:
c04ead5d: 8b 56 50 mov 0x50(%esi),%edx
c04ead60: 89 d8 mov %ebx,%eax
c04ead62: e8 fc ff ff ff call c04ead63 <tcp_rcv_established+0x633>
Therefore, I'll resubmit this patch, removing the existing len parameter.
And maybe *th, too....
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/