Re: [PATCH V3 2/5] mailbox: Add support for QTI CPUCP mailbox controller

[Bjorn Andersson]

On Thu, Apr 18, 2024 at 01:01:50PM +0530, Sibi Sankar wrote:
> On 4/18/24 02:56, Bjorn Andersson wrote:
> > On Wed, Apr 17, 2024 at 06:58:53PM +0530, Sibi Sankar wrote:
> > > diff --git a/drivers/mailbox/qcom-cpucp-mbox.c b/drivers/mailbox/qcom-cpucp-mbox.c
[..]
> > 
> > > +		if (status & BIT(i)) {
> > 
> > Can't you combine the for loop and this conditional into a
> > for_each_bit_set()?
> 
> The only drawback I see here is if the number of channels increase to
> it's full capacity of 64 since for_each_set_bit expects unsigned long.
> 

It takes a unsigned long * and it can take a size > BITS_PER_LONG. But
I've not convinced myself that the bit order across two of those matches
the u64 bits.

> > 
> > > +			val = readl(cpucp->rx_base + APSS_CPUCP_RX_MBOX_CMD(i) + APSS_CPUCP_MBOX_CMD_OFF);
> > > +			chan = &cpucp->chans[i];
> > > +			spin_lock_irqsave(&chan->lock, flags);
> > 
> > Can you please add a comment here to document that the lock is taken
> > here to deal with races against client registration? (It wasn't obvious
> > to me...)
> 
> This is was put in to handle irqs after channel closure. Meaning we
> don't want to send data on a closed/empty channel.
> 

You're dealing with that through the chan->cl check below, not the lock
itself. So the lock here would be for synchronizing this code with
potentially concurrent execution of __mbox_bind_client() or e.g.
mbox_free_channel().

But if this is indeed the problem, then we're locking here to ensure
that mbox_chan_received_data() will not dereference chan->cl while it's
being modified elsewhere int he mailbox core.

If that's the case, I think this needs to be strongly documented in the
API, or perhaps better yet the lock being moved into
mbox_chan_received_data().

Regards,
Bjorn