Re: [PATCH net-next 2/7] dma: avoid expensive redundant calls for sync operations
From: Robin Murphy
Date: Fri Jan 26 2024 - 10:49:17 EST
On 26/01/2024 1:54 pm, Alexander Lobakin wrote:
From: Eric Dumazet <edumazet@xxxxxxxxxx>
Quite often, NIC devices do not need dma_sync operations on x86_64
at least.
Indeed, when dev_is_dma_coherent(dev) is true and
dev_use_swiotlb(dev) is false, iommu_dma_sync_single_for_cpu()
and friends do nothing.
However, indirectly calling them when CONFIG_RETPOLINE=y consumes about
10% of cycles on a cpu receiving packets from softirq at ~100Gbit rate.
Even if/when CONFIG_RETPOLINE is not set, there is a cost of about 3%.
Add dev->skip_dma_sync boolean which is set during the device
initialization depending on the setup: dev_is_dma_coherent() for direct
DMA, !(sync_single_for_device || sync_single_for_cpu) or positive result
from the new callback, dma_map_ops::can_skip_sync for non-NULL DMA ops.
Then later, if/when swiotlb is used for the first time, the flag
is turned off, from swiotlb_tbl_map_single().
I think you could probably just promote the dma_uses_io_tlb flag from
SWIOTLB_DYNAMIC to a general SWIOTLB thing to serve this purpose now.
Similarly I don't think a new op is necessary now that we have
dma_map_ops.flags. A simple static flag to indicate that sync may be
skipped under the same conditions as implied for dma-direct - i.e.
dev_is_dma_coherent(dev) && !dev->dma_use_io_tlb - seems like it ought
to suffice.
Thanks,
Robin.
On iavf, the UDP trafficgen with XDP_DROP in skb mode test shows
+3-5% increase for direct DMA.
Signed-off-by: Eric Dumazet <edumazet@xxxxxxxxxx>
Co-developed-by: Alexander Lobakin <aleksander.lobakin@xxxxxxxxx>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@xxxxxxxxx>
---
include/linux/device.h | 5 +++++
include/linux/dma-map-ops.h | 17 +++++++++++++++++
include/linux/dma-mapping.h | 12 ++++++++++--
drivers/base/dd.c | 2 ++
kernel/dma/mapping.c | 34 +++++++++++++++++++++++++++++++---
kernel/dma/swiotlb.c | 14 ++++++++++++++
6 files changed, 79 insertions(+), 5 deletions(-)
diff --git a/include/linux/device.h b/include/linux/device.h
index 97c4b046c09d..f23e6a32bea0 100644
--- a/include/linux/device.h
+++ b/include/linux/device.h
@@ -686,6 +686,8 @@ struct device_physical_location {
* other devices probe successfully.
* @dma_coherent: this particular device is dma coherent, even if the
* architecture supports non-coherent devices.
+ * @dma_skip_sync: DMA sync operations can be skipped for coherent non-SWIOTLB
+ * buffers.
* @dma_ops_bypass: If set to %true then the dma_ops are bypassed for the
* streaming DMA operations (->map_* / ->unmap_* / ->sync_*),
* and optionall (if the coherent mask is large enough) also
@@ -800,6 +802,9 @@ struct device {
defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL)
bool dma_coherent:1;
#endif
+#ifdef CONFIG_DMA_NEED_SYNC
+ bool dma_skip_sync:1;
+#endif
#ifdef CONFIG_DMA_OPS_BYPASS
bool dma_ops_bypass : 1;
#endif
diff --git a/include/linux/dma-map-ops.h b/include/linux/dma-map-ops.h
index 4abc60f04209..937c295e9da8 100644
--- a/include/linux/dma-map-ops.h
+++ b/include/linux/dma-map-ops.h
@@ -78,6 +78,7 @@ struct dma_map_ops {
int nents, enum dma_data_direction dir);
void (*cache_sync)(struct device *dev, void *vaddr, size_t size,
enum dma_data_direction direction);
+ bool (*can_skip_sync)(struct device *dev);
int (*dma_supported)(struct device *dev, u64 mask);
u64 (*get_required_mask)(struct device *dev);
size_t (*max_mapping_size)(struct device *dev);
@@ -111,6 +112,22 @@ static inline void set_dma_ops(struct device *dev,
}
#endif /* CONFIG_DMA_OPS */
+#ifdef CONFIG_DMA_NEED_SYNC
+
+static inline void dma_set_skip_sync(struct device *dev, bool skip)
+{
+ dev->dma_skip_sync = skip;
+}
+
+void dma_setup_skip_sync(struct device *dev);
+
+#else /* !CONFIG_DMA_NEED_SYNC */
+
+#define dma_set_skip_sync(dev, skip) do { } while (0)
+#define dma_setup_skip_sync(dev) do { } while (0)
+
+#endif /* !CONFIG_DMA_NEED_SYNC */
+
#ifdef CONFIG_DMA_CMA
extern struct cma *dma_contiguous_default_area;
diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index 9dd7e1578bf6..bc9f67e0c139 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -365,9 +365,17 @@ __dma_sync_single_range_for_device(struct device *dev, dma_addr_t addr,
#ifdef CONFIG_DMA_NEED_SYNC
-#define dma_skip_sync(dev) false
+static inline bool dma_skip_sync(const struct device *dev)
+{
+ return dev->dma_skip_sync;
+}
+
+bool __dma_need_sync(struct device *dev, dma_addr_t dma_addr);
-bool dma_need_sync(struct device *dev, dma_addr_t dma_addr);
+static inline bool dma_need_sync(struct device *dev, dma_addr_t dma_addr)
+{
+ return dma_skip_sync(dev) ? false : __dma_need_sync(dev, dma_addr);
+}
#else /* !CONFIG_DMA_NEED_SYNC */
diff --git a/drivers/base/dd.c b/drivers/base/dd.c
index 85152537dbf1..67ad3e1d51f6 100644
--- a/drivers/base/dd.c
+++ b/drivers/base/dd.c
@@ -642,6 +642,8 @@ static int really_probe(struct device *dev, struct device_driver *drv)
goto pinctrl_bind_failed;
}
+ dma_setup_skip_sync(dev);
+
ret = driver_sysfs_add(dev);
if (ret) {
pr_err("%s: driver_sysfs_add(%s) failed\n",
diff --git a/kernel/dma/mapping.c b/kernel/dma/mapping.c
index a30f37f9d4db..8fa464b3954e 100644
--- a/kernel/dma/mapping.c
+++ b/kernel/dma/mapping.c
@@ -842,15 +842,43 @@ size_t dma_opt_mapping_size(struct device *dev)
EXPORT_SYMBOL_GPL(dma_opt_mapping_size);
#ifdef CONFIG_DMA_NEED_SYNC
-bool dma_need_sync(struct device *dev, dma_addr_t dma_addr)
+bool __dma_need_sync(struct device *dev, dma_addr_t dma_addr)
{
const struct dma_map_ops *ops = get_dma_ops(dev);
if (dma_map_direct(dev, ops))
+ /*
+ * dma_skip_sync could've been set to false on first SWIOTLB
+ * buffer mapping, but @dma_addr is not necessary an SWIOTLB
+ * buffer. In this case, fall back to more granular check.
+ */
return dma_direct_need_sync(dev, dma_addr);
- return ops->sync_single_for_cpu || ops->sync_single_for_device;
+
+ return true;
+}
+EXPORT_SYMBOL_GPL(__dma_need_sync);
+
+void dma_setup_skip_sync(struct device *dev)
+{
+ const struct dma_map_ops *ops = get_dma_ops(dev);
+ bool skip;
+
+ if (dma_map_direct(dev, ops))
+ /*
+ * dma_skip_sync will be set to false on first SWIOTLB buffer
+ * mapping, if any. During the device initialization, it's
+ * enough to check only for DMA coherence.
+ */
+ skip = dev_is_dma_coherent(dev);
+ else if (!ops->sync_single_for_device && !ops->sync_single_for_cpu)
+ skip = true;
+ else if (ops->can_skip_sync)
+ skip = ops->can_skip_sync(dev);
+ else
+ skip = false;
+
+ dma_set_skip_sync(dev, skip);
}
-EXPORT_SYMBOL_GPL(dma_need_sync);
#endif /* CONFIG_DMA_NEED_SYNC */
unsigned long dma_get_merge_boundary(struct device *dev)
diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
index b079a9a8e087..b62ea0a4f106 100644
--- a/kernel/dma/swiotlb.c
+++ b/kernel/dma/swiotlb.c
@@ -1286,6 +1286,16 @@ static unsigned long mem_used(struct io_tlb_mem *mem)
#endif /* CONFIG_DEBUG_FS */
+static inline void swiotlb_disable_dma_skip_sync(struct device *dev)
+{
+ /*
+ * If dma_skip_sync was set, reset it to false on first SWIOTLB buffer
+ * mapping/allocation to always sync SWIOTLB buffers.
+ */
+ if (unlikely(dma_skip_sync(dev)))
+ dma_set_skip_sync(dev, false);
+}
+
phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
size_t mapping_size, size_t alloc_size,
unsigned int alloc_align_mask, enum dma_data_direction dir,
@@ -1323,6 +1333,8 @@ phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
return (phys_addr_t)DMA_MAPPING_ERROR;
}
+ swiotlb_disable_dma_skip_sync(dev);
+
/*
* Save away the mapping from the original address to the DMA address.
* This is needed when we sync the memory. Then we sync the buffer if
@@ -1640,6 +1652,8 @@ struct page *swiotlb_alloc(struct device *dev, size_t size)
if (index == -1)
return NULL;
+ swiotlb_disable_dma_skip_sync(dev);
+
tlb_addr = slot_addr(pool->start, index);
return pfn_to_page(PFN_DOWN(tlb_addr));