On 09/15, Ivan T. Ivanov wrote:

On Mon, 2015-09-14 at 18:28 -0700, Stephen Boyd wrote:

quoted

On 09/14/2015 02:54 PM, Stephen Boyd wrote:

quoted

The current driver implementation supports only 128 peripherals.
Add support for more than 128 peripherals by taking a lazy
caching approach to the mapping tables. Instead of reading the
tables at boot given some fixed size, read them on an as needed
basis and cache the results. We still assume a max number of 512
peripherals, trading off some space for simplicity.

Based on a patch by Gilad Avidov [off-list ref] and
Sagar Dharia [off-list ref].

Signed-off-by: Stephen Boyd <redacted>
---

Hi Ivan,

This patch causes 8916 to crash, because there isn't a mapping for ppid
257 in the ppid to channel table. It seems that we're reading the revid
from the slave id 1 pmic by going through channel 0, which seems to be
setup for ppid 9 (slave id 0 and the peripheral starting at 0x900). Can
we stop reading the revid registers from non-zero slave id pmic devices?
That would be one solution to fix this problem. Or maybe we need to
special case this in the pmic arbiter code to fold ppid 0xN01 (slave id
N and address 0x100) onto channel 0 all the time?

Yes, we can. We are not using this information at the moment.
Right now, revision read is more or less for debug purposes.

Would following patch work for you? Of course it will be difficult

Yes the patch works fine. Feel free to add a

Tested-by: Stephen Boyd <redacted>

to guaranties that some other driver misbehave and touch non-existing
register, right?

Yeah some driver could misbehave and try to read/write something
that doesn't exist. As in this case, the device will reset. Given
that all this stuff is in the kernel though, it seems fine to
blow up instead of add a bunch of checks that aren't ever going
to matter in practice.

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project