REVISED: Experimentation with Athlon and fast_page_copy
Quoted 10 lines
---
Quoted 1 line
__asm__ __volatile__ (158c157
< "3: movw $0x1AEB, 1b\n"
---Quoted 1 line
"3: movw $0x1AEB, 1b\n" /* jmp on 26 bytes */166c165
< */
---Quoted 1 line
Quoted 6 lines
170c169
< "1: nop\n" /* prefetch 320(%0)\n" */
---Quoted 1 line
"1: prefetch 320(%0)\n"-------------------------
Please let me know if that makes sense :).
Very interesting.
You've removed only the prefetch 320(%0), not the other prefetch instructions?
prefetch 320(%0) can fetch memory behind the end of the source page. Perhaps it accesses memory in the ISA hole, or beyond the end of memory? Could you post the e820 map from dmesg?
It's possible to build manually a memory map.
Could you build one with wide margins from "dangerous" areas? (untested:
mem=exactmap mem=620k@0 mem=<your mem in MB-2>M@1M)
Then boot with prefetch enabled.
--
Manfred
On Fri, 4 May 2001, Manfred Spraul wrote:
| > --- | > > __asm__ __volatile__ ( | > 158c157 | > < "3: movw $0x1AEB, 1b\n" | > --- | > > "3: movw $0x1AEB, 1b\n" /* jmp on 26 bytes */ | > 166c165 | > < */ | > --- | > > | > 170c169 | > < "1: nop\n" /* prefetch 320(%0)\n" */ | > --- | > > "1: prefetch 320(%0)\n" | > ------------------------- | > Please let me know if that makes sense :). | | Very interesting. | You've removed only the prefetch 320(%0), not the other prefetch | instructions?
No, I have removed them all -- I just commented the block above it
completely out :). (maybe you didn't see the whole patch?)
--Seth
Quoted 9 lines
prefetch 320(%0) can fetch memory behind the end of the source page. Perhaps it accesses memory in the ISA hole, or beyond the end of memory? Could you post the e820 map from dmesg?
It's possible to build manually a memory map.
Could you build one with wide margins from "dangerous" areas? (untested:
mem=exactmap mem=620k@0 mem=<your mem in MB-2>M@1M)Then boot with prefetch enabled.
That might not be the actual bug but for rev 1 Athlon it is a real bug. The first step athlons have an unfortunate problem in that will prefetch memory marked uncachable and corrupt their caches with it.
Arjan - care to unroll the tail 320 bytes of copying from the main loop ?
Quoted 1 line · In article <[email protected]> you wrote:
Arjan - care to unroll the tail 320 bytes of copying from the main loop ?
I'll see what I can do to make us not loose too much speed.