Date: Sun, 7 Jul 2002 15:26:52 -0700 (PDT) From: Matthew Dillon <dillon@apollo.backplane.com> To: Alan Cox <alc@cs.rice.edu> Cc: Peter Wemm <peter@wemm.org>, Jake Burkholder <jake@locore.ca>, Alan Cox <alc@FreeBSD.ORG>, cvs-committers@FreeBSD.ORG, cvs-all@FreeBSD.ORG Subject: Re: cvs commit: src/sys/vm vm_zeroidle.c Message-ID: <200207072226.g67MQqoT017816@apollo.backplane.com> References: <20020707211530.GZ11391@cs.rice.edu> <20020707212707.BD3713808@overcee.wemm.org> <20020707213541.GA11391@cs.rice.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
:> Thats what I did to avoid reentrancy problems. The zeroidle process uses
:> CMAP3 on x86. (ugh! why not allocate some kvm? hmmm).
:>
:> However, with preemption, we can end up on another cpu. Or, we get
:> scheduled on a different cpu next time. If we had a way to pin a khread to
:> a particular cpu, now that would be nice. We could get in, pin ourselves,
:> do our mappings, zero the page, unmap, unpin and done. No preemption
:> problems to worry about then, and no need to worry about coherency on other
:> cpus since this is explicitly single threaded.
:
:Let me suggest another way to deal with preemption: associate a flag
:with the pmap that says flush the TLB on activation even if you
:appear to be switching to the same page table, in this case the kernel's
:page table. pmap_zero_page() would set and unset this flag around
:the bzero().
:
:Alan
You mean CMAP2, not CMAP3, I think.
...
*CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M;
invltlb_1pg((vm_offset_t)CADDR2);
#if defined(I686_CPU)
if (cpu_class == CPUCLASS_686)
i686_pagezero(CADDR2);
else
#endif
bzero(CADDR2, PAGE_SIZE);
*CMAP2 = 0;
...
Hmm. pmap_copy_page() also uses CMAP2, wouldn't they collide with
each other? pmap_copy_page() and pmap_zero_page()? Am I missing
something here or does it work simply because all of this stuff is
still under Giant?
In anycase, an obvious but unpopular answer would be to use a
critical section. One page == 4K @ 200 MBytes/sec = 20 uS latency.
That would work as a temporary solution. Or you could break the
zeroing operation down into four 1K segments within the critical
section, invltlb'ing the page between each section giving you four
5 uS critical sections.
-Matt
Matthew Dillon
<dillon@backplane.com>
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe cvs-all" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200207072226.g67MQqoT017816>