On Tue, 1 Apr 2003 14:27:35 +0530,
"Vamsi Krishna S ." <vamsi@xxxxxxxxxx> wrote:
>Capturing the state in spin loops when a flag say kdb_enter_debugger
>is set seems like a good idea. There doesn't seem to be much point
>in doing it at other places as if a cpu is not in a tight spin loop,
>it would have got the (NMI class) IPI. But, adding test for a flag
>and call to save register state within spinlock's code may
>be considered 'bloat' and unacceptable by many people :-(.
kdb v4.1 does not test a flag in i386 spinlocks because the NMI
"alwasy" gets through on i386, the test is only done on ia64. I have
rewritten the spinlock code on ia64[*] to put only the uncontended path
inline. There is a single copy of the code for the contended path, so
adding a kdb test to that single copy is not an issue.
[*] The current ia64 spinlock code uses 5 bundles for each spinlock()
call. By moving the contended code out of line I got the
spinlock() code down to 3 bundles for McKinley, 4 bundles for
Itanium. The new spinlock code is smaller, it is measurably faster
for uncontended locks and it gives better debugging for contended
locks.
|