Hi Eric,
On 12/1/20 4:48 PM, Auger Eric wrote:
> Hi Alexandru,
>
> On 11/25/20 4:51 PM, Alexandru Elisei wrote:
>> The GICv3 driver executes a DSB barrier before sending an IPI, which
>> ensures that memory accesses have completed. This removes the need to
>> enforce ordering with respect to stats_reset() in the IPI handler.
>>
>> For GICv2, we still need the DMB to ensure ordering between the read of the
>> GICC_IAR MMIO register and the read from the acked array. It also matches
>> what the Linux GICv2 driver does in gic_handle_irq().
>>
>> Signed-off-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
>> ---
>> arm/gic.c | 9 +++++++--
>> 1 file changed, 7 insertions(+), 2 deletions(-)
>>
>> diff --git a/arm/gic.c b/arm/gic.c
>> index 4e947e8516a2..7befda2a8673 100644
>> --- a/arm/gic.c
>> +++ b/arm/gic.c
>> @@ -60,7 +60,6 @@ static void stats_reset(void)
>> bad_sender[i] = -1;
>> bad_irq[i] = -1;
>> }
>> - smp_wmb();
> Here we are (pair removed). Still the one in check_acked still exists.
The smp_rmb() from check_acked() is there to implement the message passing pattern
wrt the writes from the ipi_handler() function, not the writes from stats_reset().
See the next patch where I try to explain how the barriers should work.
Thanks,
Alex
>> }
>>
>> static void check_acked(const char *testname, cpumask_t *mask)
>> @@ -150,7 +149,13 @@ static void ipi_handler(struct pt_regs *regs __unused)
>>
>> if (irqnr != GICC_INT_SPURIOUS) {
>> gic_write_eoir(irqstat);
>> - smp_rmb(); /* pairs with wmb in stats_reset */
>> + /*
>> + * Make sure data written before the IPI was triggered is
>> + * observed after the IAR is read. Pairs with the smp_wmb
>> + * when sending the IPI.
>> + */
>> + if (gic_version() == 2)
>> + smp_rmb();
>> ++acked[smp_processor_id()];
>> check_ipi_sender(irqstat);
>> check_irqnr(irqnr);
>>
> Thanks
>
> Eric
>
