On Wednesday, May 25, 2016 4:15:31 PM CEST Sricharan wrote:

quoted

Any operation that could trigger a DMA from a device is required
to have a barrier preceding it (usually wmb() one implied by writel()),
so this is clearly not about a driver that installs a DMA mapping
before starting a DMA, but I don't see what else it would be.

Ok, so i was doing this from the idea that, other iommu drivers
 where polling on a status bit in their sync call to ensure completion
of pending TLB invalidations. But in this case, there is no status bit.
 So added a barrier to have no ordering issues before the client
triggers the dma operation. But as you say above that it is implicit that
the device would have a barrier before starting the trigger, then the
barrier here becomes redundant.

Ok. There are two more things to note here:

* On other platforms, polling the register is likely required because
  an MMIO write is "posted", meaning that a sync after writel() will
  only ensure that it has left the CPU write queue, but it may still be
  on the bus fabric and whatever side-effects are triggered by the
  write are normally not guaranteed to be completed even after the
  'sync'. You need to check the datasheet for your IOMMU to find out
  whether the 'dsb' instruction actually has any effect on the IOMMU.
  If not, then neither the barrier that you add here nor the barrier
  in the following writel() is sufficient.

* The one barrier that is normally required in an IOMMU is between
  updating the in-memory page tables and the following MMIO store
  that triggers the TLB flush for that entry. This barrier is
  implied by writel() but not writel_relaxed(). If you don't have
  a hardware page table walker in your IOMMU, you don't need to worry
  about this.

	Arnd