On 01/27/2015 01:33 PM, Bob Copeland wrote:
On Mon, Jan 26, 2015 at 02:25:18PM -0800, Peter Oh wrote:
Using ioread() to perform data sync is excessive.
Use compact API, wmb(), that intended to be used for the case.
It reduces total 14 CPU clocks per interrupt.
Hi,
ath10k_pci_write32(ar, SOC_CORE_BASE_ADDRESS + PCIE_INTR_CLR_ADDRESS,
PCIE_INTR_FIRMWARE_MASK | PCIE_INTR_CE_MASK_ALL);
- /* IMPORTANT: this extra read transaction is required to
- * flush the posted write buffer. */
- (void)ath10k_pci_read32(ar, SOC_CORE_BASE_ADDRESS +
- PCIE_INTR_ENABLE_ADDRESS);
+ /* invoke data sync barrier */
+ wmb();
}
I am no expert in arcane PCI matters, but that looks suspicious to me. I seem
to recall wmb() only enforced ordering, and maybe not even memory-IO ordering
on all platforms. If you want to disable an irq, it really seems like you
would want to flush posted writes so you know the hardware has seen it.
enforced ordering is happened by flush write buffer and wmb is commonly
used to flush write buffer.
so that wmb guarantees ordering by flush write buffer. That's why it's
called a memory barrier.
when wmb is used, it guarantees all memory accesses complete before wmb
command completes.
for instance dsb is the corresponding command for wmb in arm and arm
instruction guide says
"The DSB instruction completes when all explicit memory accesses before
it complete."
Which means DSB instruction will hold the bus (usually AXI bus) and
never returns until any memory or memory mapped I/O access complete
(dsb is for ARM and other platforms have the equivalent instruction such
as sfence, sync, mf, and dcs).
Thanks,
Peter
--
To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
