[lkml]   [2018]   [Oct]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[PATCH 0/4] MIPS: Ordering enforcement fixes for MMIO accessors

This patch series is a follow-up to my earlier consideration about MMIO
access ordering recorded here: <>.

As I have learnt in a recent Alpha/Linux discussion starting here:
related to MMIO accessor ordering barriers ports are actually required to
follow the x86 strongly ordered semantics. As the ordering is not
specified in the MIPS architecture except for the SYNC instruction we do
have to put explicit barriers in MMIO accessors as otherwise ordering may
not be guaranteed.

Fortunately on strongly ordered systems SYNC is expected to be as cheap
as a NOP, and on weakly ordered ones it is needed anyway. As from
revision 2.60 of the MIPS architecture specification however we have a
number of SYNC operations defined, and SYNC 0 has been upgraded from an
ordering to a completion barrier. We currently don't make use of these
extra operations and always use SYNC 0 instead, which this means that we
may be doing too much synchronisation with the barriers we have already

This patch series does not make an attempt to optimise for SYNC operation
use, which belongs to a separate improvement. Instead it focuses on
fixing MMIO accesses so that drivers can rely on our own API definition.

Following the original consideration specific MMIO barrier operations are
added. As they have turned out to be required to be implied by MMIO
accessors there is no immediate need to make them form a generic
cross-architecture internal Linux API. Therefore I defined them for the
MIPS architecture only, using the names originally coined by mostly taking
them from the PowerPC port.

Then I have used them to fix `mmiowb', and then `readX' and `writeX'
accessors. Finally I have updated the `_relaxed' accessors to avoid
unnecessary synchronisation WRT DMA.

See individual commit descriptions for further details.

As a follow-up clean-up places across the architecture tree could be
reviewed for barrier use that is actually related to MMIO rather than
memory and updated to use the new names of the MMIO barrier operations. I
plan to do this for the DECstation and possibly the SiByte platform,
however I am leaving it for someone else to do it elsewhere.

Similarly I think the DMA barrier in `readX' and `inX' should be using
`dma_rmb' rather than `rmb', but I'm leaving it for someone else to

These changes have been verified at run time with an R3000 (MIPS I)
DECstation machine (32-bit kernel, little endianness), an R4400 (MIPS III)
DECstation machine (64-bit kernel, little endianness) and an SB-1 (MIPS64)
SWARM machine (64-bit kernel, big endianness), by booting them into the
multiuser mode and running them for a couple of hours.

Please apply.


 \ /
  Last update: 2018-10-08 02:37    [W:0.068 / U:0.388 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site