Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Robin Murphy <robin.murphy@arm.com>
To:     will@kernel.org, joro@8bytes.org
Cc:     iommu@lists.linux-foundation.org, linux-kernel@vger.kernel.org,
        hch@lst.de, john.garry@huawei.com
Subject: [PATCH v2] iommu/dma: Add config for PCI SAC address trick
Date:   Thu,  9 Jun 2022 16:12:10 +0100
Message-Id: <3f06994f9f370f9d35b2630ab75171ecd2065621.1654782107.git.robin.murphy@arm.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

For devices stuck behind a conventional PCI bus, saving extra cycles at
33MHz is probably fairly significant. However since native PCI Express
is now the norm for high-performance devices, the optimisation to always
prefer 32-bit addresses for the sake of avoiding DAC is starting to look
rather anachronistic. Technically 32-bit addresses do have shorter TLPs
on PCIe, but unless the device is saturating its link bandwidth with
small transfers it seems unlikely that the difference is appreciable.

What definitely is appreciable, however, is that the IOVA allocator
doesn't behave all that well once the 32-bit space starts getting full.
As DMA working sets get bigger, this optimisation increasingly backfires
and adds considerable overhead to the dma_map path for use-cases like
high-bandwidth networking. We've increasingly bandaged the allocator
in attempts to mitigate this, but it remains fundamentally at odds with
other valid requirements to try as hard as possible to satisfy a request
within the given limit; what we really need is to just avoid this odd
notion of a speculative allocation when it isn't beneficial anyway.

Unfortunately that's where things get awkward... Having been present on
x86 for 15 years or so now, it turns out there are systems which fail to
properly define the upper limit of usable IOVA space for certain devices
and this trick was the only thing letting them work OK. I had a similar
ulterior motive for a couple of early arm64 systems when originally
adding it to iommu-dma, but those really should be fixed with proper
firmware bindings by now. Let's be brave and default it to off in the
hope that CI systems and developers will find and fix those bugs, but
expect that desktop-focused distro configs are likely to want to turn
it back on for maximum compatibility.

Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---

v2: Tweak wording to clarify that it's not really an optimisation in
    general, remove "default X86".

 drivers/iommu/Kconfig     | 26 ++++++++++++++++++++++++++
 drivers/iommu/dma-iommu.c |  2 +-
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
index c79a0df090c0..5a225b48dd00 100644
--- a/drivers/iommu/Kconfig
+++ b/drivers/iommu/Kconfig
@@ -144,6 +144,32 @@ config IOMMU_DMA
 	select IRQ_MSI_IOMMU
 	select NEED_SG_DMA_LENGTH
 
+config IOMMU_DMA_PCI_SAC
+	bool "Enable 64-bit legacy PCI optimisation by default"
+	depends on IOMMU_DMA
+	help
+	  Enable by default an IOMMU optimisation for 64-bit legacy PCI devices,
+	  wherein the DMA API layer will always first try to allocate a 32-bit
+	  DMA address suitable for a single address cycle, before falling back
+	  to allocating from the device's full usable address range. If your
+	  system has 64-bit legacy PCI devices in 32-bit slots where using dual
+	  address cycles reduces DMA throughput significantly, this may be
+	  beneficial to overall performance.
+
+	  If you have a modern PCI Express based system, this feature mostly just
+	  represents extra overhead in the allocation path for no practical
+	  benefit, and it should usually be preferable to say "n" here.
+
+	  However, beware that this feature has also historically papered over
+	  bugs where the IOMMU address width and/or device DMA mask is not set
+	  correctly. If device DMA problems and IOMMU faults start occurring
+	  after disabling this option, it is almost certainly indicative of a
+	  latent driver or firmware/BIOS bug, which would previously have only
+	  manifested with several gigabytes worth of concurrent DMA mappings.
+
+	  If this option is not set, the feature can still be re-enabled at
+	  boot time with the "iommu.forcedac=0" command-line argument.
+
 # Shared Virtual Addressing
 config IOMMU_SVA
 	bool
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index f90251572a5d..9f9d9ba7f376 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -67,7 +67,7 @@ struct iommu_dma_cookie {
 };
 
 static DEFINE_STATIC_KEY_FALSE(iommu_deferred_attach_enabled);
-bool iommu_dma_forcedac __read_mostly;
+bool iommu_dma_forcedac __read_mostly = !IS_ENABLED(CONFIG_IOMMU_DMA_PCI_SAC);
 
 static int __init iommu_dma_forcedac_setup(char *str)
 {
-- 
2.36.1.dirty