Re: [PATCH] powerpc/64s: Add support for ASB_Notify on POWER9

On Sat, 2017-08-05 at 14:28 +1000, Benjamin Herrenschmidt wrote:

On Fri, 2017-08-04 at 16:56 +0200, Christophe Lombard wrote:

quoted

The POWER9 core supports a new feature: ASB_Notify which requires the
support of the Special Purpose Register: TIDR.
=20
The ASB_Notify command, generated by the AFU, will attempt to
wake-up the host thread identified by the particular LPID:PID:TID.
=20
The special register TIDR has to be updated to with the same value
present in the process element.
=20
If the length of the register TIDR is 64bits, the CAPI Translation
Service Layer core (XSL9) for Power9 systems limits the size (16bits) o=

f

quoted

the Thread ID when it generates the ASB_Notify message adding
PID:LPID:TID information from the context.
=20
The content of the internal kernel Thread ID (32bits) can not therefore
be used to fulfill the register TIDR.
=20
This patch allows to avoid this limitation by adding a new interface
for the user. The instructions mfspr/mtspr SPRN_TIDR are emulated,
save/restore SPRs (context switch) are updated and a new feature
(CPU_FTR_TIDR) is added to POWER9 system.

=20
Those CPU_FTR_* are internal to the kernel. You probably also need a
feature in AT_HWCAP2 to indicate to userspace that this is supported.
=20
Also you put the onus of allocating the TIDs onto userspace which is a
bit tricky. What happens if there are duplicate TIDs for example ? (ie,
userspace doesn't allocate it or uses a library that spawns a thread)

I tend to agree.  I don't want userspace knowing anything about TIDR
allocations.  If we want userspace to receive one of these as_notifys, ther=
e
should be some abstract handle (like a file descriptor) that the kernel giv=
es
out. That handle should be associated with an LPID/PID/TID tuple by the ker=
nel.

This is similar to the PE number in CAPI. Most of the time userspace doesn'=
t
need to know its PE number (there is some cases were a master needs to know=
 a
slave PE, but that's the exception, not the rule).  Similarly, guests don't=
 know
their LPID.  Processes don't know their PID. Threads shouldn't know the TID=
.

Also, Suka has posted a patch that does TID allocation in the kernel...=C2=
=A0
http://patchwork.ozlabs.org/patch/799494/

Regards,
Mikey