Hi Ohad,
On 03/14/2014 03:10 PM, Ohad Ben-Cohen wrote:
Hi Suman, Mark,
On Mon, Feb 24, 2014 at 8:14 PM, Suman Anna <s-anna@xxxxxx> wrote:
Mark, Ohad,
...
Gentle reminder, can you provide your acks/comments?
Sorry for the late jump in.
I have a few comments:
Thanks for the comments. It probably covers few topics that are slightly
beyond the scope of the series, but nevertheless are good discussion
points for finalizing the series.
- Hardware spinlocks must have global and system-wide id numbers;
these numbers cannot be maintained internally by the Linux Kernel.
Think of an SoC with several asynchronous heterogeneous processors,
each of which is running a different OS, and they all need to use a
specific hardware spinlock in order to share some resource. For that
to happen, every hwlock must have a predefined and deterministic id
number which is global in the system. We can't have those id numbers
be relative to an hwlock "controller", and we can't have two hwlock
"controllers" share the same id numbers.
The series doesn't change the semantics of hwspinlock registration or
adds a new OF controller registration function. Implementations would
still need to register a controller using a base_id and number of locks.
The series rather adds a DT-friendly function _ONLY_ for requesting a
specific hwlock, and there are no restrictions on the args specifier
being relative id numbers. Though this is what the simple default xlate
helper does (most common usage), the added xlate ops and #hwlock-cells
should allow individual implementation drivers to adjust any variations,
and return a relative lock w.r.t its registered base_id, as this is how
a lock gets registered in the first place.
- I suspect the simplest and most straight forward way to achieve this
is by (a) bringing back the concept of the base_id property, and
I actually started out this series with the base_id property, and
dropped it in v3 based on comments looking at it from the
request-specific-lock semantics with DT. That said, the drivers still
need to manage a 'base_id' needed for registration when they get probed
for multiple controllers. Getting the base_id from DT _may_ be useful
just for registration purposes, but for requesting a hwlock, a
controller phandle and an implementation defined args-specifier should
suffice IMHO.
(b)
letting the global hwlock id be the DT identifier (plus the base_id)
and then providing it directly to the drivers when needed.The latter
is required in order to support dynamically allocation of hwlocks, in
which case Linux must know the global system-wide id number, and then
share it with the other asynchronous OSes via some IPC.
Each lock still retains a global lock id value, and you can retrieve it
using the existing hwspin_lock_get_id(). Why is the latter required for
dynamic allocation, isn't it the other way around, allocate a lock, and
you will be able to get the lock id. If wanting to request a specific
lock received across, the regular hwspin_lock_request_specific should be
used.
- If we feel there's no way any system is going to have more than a
single hwlock controller, then we can live without a base_id property,
but then this needs to be clearly documented and prohibited. Today
both the hwlock DT representation, and the coupled kernel code,
implicitly allow this anomaly to exist.
I haven't removed the concept of base_id, it is just not defined in the
DT-bindings, and am currently expecting the drivers to manage it and use
it for registration.
- Hwlock controller nodes should have a list of reserved locks (those
locks for which other nodes have a phandle+identifier pointing at) to
prevent those locks from being dynamically allocated by eager drivers.
The exact notion of informing the hwspinlock core about a list of
reserved locks is missing at the moment (even in the non-DT case). I am
not sure if this got lost during the conversion of the registration from
per lock to registering a bank of locks together, or if it is implied by
the base_id + num_locks combination. The core today supports requesting
only those locks that were actually registered, whether allocating a
free one dynamically or giving a specific one.
There were some slightly similar comments from Kumar earlier on the v2
series, please see the thread in [1].
Most of these issues were discussed Arnd, Benoit and myself back then,
please see below:
http://lists.infradead.org/pipermail/linux-arm-kernel/2011-September/064265.html
Thanks for the pointer to the previous discussion, I wasn't aware of an
earlier attempt. The primary approach on requesting locks is actually no
different from what Arnd suggested originally.
regards
Suman
[1] http://marc.info/?l=linux-omap&m=138031002012191&w=2
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
