Date: Thu, 24 Jul 2003 17:00:13 -0400 (EDT) From: John Baldwin <jhb@FreeBSD.org> To: Marcel Moolenaar <marcel@xcllnt.net> Cc: cvs-all@FreeBSD.org Subject: Re: cvs commit: src/lib/libthr/arch/alpha/alpha _curthread.c Message-ID: <XFMail.20030724170013.jhb@FreeBSD.org> In-Reply-To: <20030724191943.GA1028@dhcp01.pn.xcllnt.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On 24-Jul-2003 Marcel Moolenaar wrote:
> On Thu, Jul 24, 2003 at 03:08:32PM -0400, John Baldwin wrote:
>>
>> On 24-Jul-2003 Marcel Moolenaar wrote:
>> > marcel 2003/07/24 00:51:49 PDT
>> >
>> > FreeBSD src repository
>> >
>> > Modified files:
>> > lib/libthr/arch/alpha/alpha _curthread.c
>> > Log:
>> > Implement _get_curthread and _set_curthread. We use GCCs builtin
>> > function this, which expands to PAL calls (rduniq and wruniq).
>> > This needs adjustment when TLS is implemented.
>>
>> This trashes SMP since we use the per-CPU unique value already
>> to maintain the per-cpu data pointer.
>
> Writing to the per-CPU unique value is an unprivileged operation.
> You cannot possibly use that for in-kernel use without saving and
> restoring it on kernel exit/entry. Which is what needs to be done.
There is a chicken and egg problem. We use the unique value to
initialize the per-cpu pointer on kernel entry. We only have one
such beast, and in the kernel we cache it in a register that userland
gets to clobber (and frequently does). If you can think of a better
way to store the per-cpu pointer such that kernel entry can load it
into the register go for it.
Read some code:
syscall(...)
{
...
/*
* Find our per-cpu globals.
*/
#ifdef SMP
s = intr_disable();
#endif
pcpup = (struct pcpu *) alpha_pal_rdval();
td = curthread;
#ifdef SMP
td->td_md.md_kernnest++;
intr_restore(s);
#endif
...
}
--
John Baldwin <jhb@FreeBSD.org> <>< http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve!" - http://www.FreeBSD.org/
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?XFMail.20030724170013.jhb>