On Tue, 18 Mar 1997, David S. Miller wrote:
>
> For other Posix-ish features like PID->task relations we could use hash
> tables ... i'm not sure yet how important this is ... if you've got an
> application that uses this heavily, tell me. [or if you know some part of
> the kernel that frequently uses it]
>
> This seems to suggest that we should hash processes when they are
> created based upon pid, and perhaps hashed in other tables as well
> based upon pgrp etc.
>
> That would be the real fix to these problems.
yup, here is PID hashing implemented for 2.1.30.
trace of kill() from lmbench lat_sig without the patch [166MHz P5]:
------------------------------------------->
Trace: c011bb97 <*SYS_ENTRY*+17/20> (0.72)
Trace: c0118693 (0.64)
Trace: c0118632 (16.45) <---- walks full task list
Trace: c01181cf (1.39)
Trace: c011bba8 <*SYS_EXIT*+8/10> (0.88)
<------------------------------------------
with the patch:
------------------------------------------->
Trace: c011bbc7 <*SYS_ENTRY*+17/20> (0.13)
Trace: c01187cf (0.11)
Trace: c011877f (0.10)
Trace: c01181d9 (0.22) <---- doesnt have to walk
Trace: c0118303 (0.49)
Trace: c011bbd8 <*SYS_EXIT*+8/10> (1.44)
<------------------------------------------
lookup_pid() is used in kill() and fork() currently, but there might be
other places too which could use it.
Comments, reports welcome.