From: Alexander Duyck <alexander.h.duyck@intel.com>
Subject: [PATCH v2 0/7] Improve swiotlb performance by using physical addresses
To: konrad.wilk@oracle.com, tglx@linutronix.de, mingo@redhat.com,
        hpa@zytor.com, rob@landley.net, akpm@linux-foundation.org,
        joerg.roedel@amd.com, bhelgaas@google.com, shuahkhan@gmail.com,
        fujita.tomonori@lab.ntt.co.jp
Cc: linux-kernel@vger.kernel.org, x86@kernel.org
Date: Thu, 11 Oct 2012 13:34:03 -0700
Message-ID: <20121011203010.12444.15503.stgit@gitlad.jf.intel.com>
User-Agent: StGIT/0.14.2
MIME-Version: 1.0
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3376
Lines: 69

While working on 10Gb/s routing performance I found a significant amount of
time was being spent in the swiotlb DMA handler. Further digging found that a
significant amount of this was due to virtual to physical address translation
and calling the function that did it. It accounted for nearly 60% of the
total swiotlb overhead.

This patch set works to resolve that by replacing the io_tlb_start and
io_tlb_end virtual addresses with a physical addresses. In addition it changes
the io_tlb_overflow_buffer from a virtual to a physical address. I followed
through with the cleanup to the point that the only functions that really
require the virtual address for the DMA buffer are the init, free, and
bounce functions.

In the case of devices that are using the bounce buffers these patches should
result in only a slight performance gain if any. This is due to the locking
overhead required to map and unmap the buffers.

In the case of devices that are not making use of bounce buffers these patches
can significantly reduce their overhead. In the case of an ixgbe routing test
for example, these changes result in 7 fewer calls to __phys_addr and
allow is_swiotlb_buffer to become inlined due to a reduction in the number of
instructions. When running a routing throughput test using small packets I
saw roughly a 6% increase in packets rates after applying these patches. This
appears to match up with the CPU overhead reduction I was tracking via perf.

Before:
Results 10.0Mpps

After:
Results 10.6Mpps

Finally, I updated the parameter names for several of the core function calls
as there was some ambiguity in naming. Specifically virtual address pointers
were named dma_addr. When I changed these pointers to physical I instead used
the name tlb_addr as this value represented a physical address in the
io_tlb_start region and is less likely to be confused with a bus address.

v2:
I reviewed the changes and realized that the first patch that was dropping
io_tlb_end and calculating the value didn't actually gain me much once I had
gone through and translated the rest of the addresses to physical addresses.
As such I have updated the patch so that it instead is converting io_tlb_end
from a virtual address to a physical address.  This actually helps to reduce
the overhead for is_swiotlb_buffer and swiotlb_dma_supported by several
instructions.

---

Alexander Duyck (7):
      swiotlb:  Do not export swiotlb_bounce since there are no external consumers
      swiotlb: Use physical addresses instead of virtual in swiotlb_tbl_sync_single
      swiotlb: Use physical addresses for swiotlb_tbl_unmap_single
      swiotlb: Return physical addresses when calling swiotlb_tbl_map_single
      swiotlb: Make io_tlb_overflow_buffer a physical address
      swiotlb: Make io_tlb_start a physical address instead of a virtual one
      swiotlb: Make io_tlb_end a physical address instead of a virtual one


 drivers/xen/swiotlb-xen.c |   25 ++--
 include/linux/swiotlb.h   |   20 ++-
 lib/swiotlb.c             |  269 +++++++++++++++++++++++----------------------
 3 files changed, 163 insertions(+), 151 deletions(-)

-- 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/