On Wed, Oct 28, 2020 at 5:47 AM Joel Stanley <j...@jms.id.au> wrote:
>
> On Thu, 22 Oct 2020 at 07:41, Benjamin Herrenschmidt
> <b...@kernel.crashing.org> wrote:
> > > >
> > > > + /* Ensure the descriptor config is visible before setting the tx
> > > > + * pointer.
> > > > + */
> > > > + smp_wmb();
> > > > +
> > >
> > > Shouldn't these be paired with smp_rmb() on the reader side?
> >
> > (Not near the code right now) I *think* the reader already has them
> > where it matters but I might have overlooked something when I quickly
> > checked the other day.
>
> Do we need a read barrier at the start of ftgmac100_tx_complete_packet?
>
> pointer = priv->tx_clean_pointer;
> <--- here
> txdes = &priv->txdes[pointer];
>
> ctl_stat = le32_to_cpu(txdes->txdes0);
> if (ctl_stat & FTGMAC100_TXDES0_TXDMA_OWN)
> return false;
>
> This was the only spot I could see that might require one.
No, I don't think this is the one, since tx_clean_pointer is not updated
in the other CPU, only this one. From what I can tell, you have
a communication between three concurrent threads in the TX
path:
a) one CPU runs start_xmit. It reads tx_clean_pointer
from the second CPU, writes the descriptors for the hardware,
notifies the hardware and updates tx_pointer.
b) the hardware gets kicked by start_xmit, it reads the
descriptors, reads the data and updates the descriptors.
it may send an interrupt, which is not important here.
c) a second CPU runs the poll() function. It reads the
tx_pointer, reads the descriptors, updates the descriptors
and updates tx_clean_pointer.
Things get a bit confusing because the tx_pointer and
tx_clean_pointer variables are hidden behind macros
and both sides repeatedly read and write them, with no
serialization.
This is what I would try to untangle it:
- mark both tx_pointer and tx_clean_pointer as
____cacheline_aligned_in_smp to avoid the cacheline
pingpong between the two CPUs
- Use smp_load_acquire() to read a pointer that may have
been updated by the other thread, and smp_store_release()
to update the one the other side will read.
- pull these accesses into the callers and only do them
once if possible
- reconsider the "keep poll() going as long as tx
packets are queued" logic, and instead serialize
against the packets that are actually completed
by the hardware.
Arnd
