diff options
author | David Arinzon <darinzon@amazon.com> | 2023-06-12 15:14:48 +0300 |
---|---|---|
committer | Jakub Kicinski <kuba@kernel.org> | 2023-06-16 08:45:47 +0300 |
commit | f7d625adeb7bc6a9ec83d32d9615889969d64484 (patch) | |
tree | 6771c7b35455f5bde6a5ee4c43aa9c5627b2dba6 /Documentation/networking/device_drivers | |
parent | e1d001fa5b477c4da46a29be1fcece91db7c7c6f (diff) | |
download | linux-f7d625adeb7bc6a9ec83d32d9615889969d64484.tar.xz |
net: ena: Add dynamic recycling mechanism for rx buffers
The current implementation allocates page-sized rx buffers.
As traffic may consist of different types and sizes of packets,
in various cases, buffers are not fully used.
This change (Dynamic RX Buffers - DRB) uses part of the allocated rx
page needed for the incoming packet, and returns the rest of the
unused page to be used again as an rx buffer for future packets.
A threshold of 2K for unused space has been set in order to declare
whether the remainder of the page can be reused again as an rx buffer.
As a page may be reused, dma_sync_single_for_cpu() is added in order
to sync the memory to the CPU side after it was owned by the HW.
In addition, when the rx page can no longer be reused, it is being
unmapped using dma_page_unmap(), which implicitly syncs and then
unmaps the entire page. In case the kernel still handles the skbs
pointing to the previous buffers from that rx page, it may access
garbage pointers, caused by the implicit sync overwriting them.
The implicit dma sync is removed by replacing dma_page_unmap() with
dma_unmap_page_attrs() with DMA_ATTR_SKIP_CPU_SYNC flag.
The functionality is disabled for XDP traffic to avoid handling
several descriptors per packet.
Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: Shay Agroskin <shayagr@amazon.com>
Signed-off-by: David Arinzon <darinzon@amazon.com>
Link: https://lore.kernel.org/r/20230612121448.28829-1-darinzon@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Diffstat (limited to 'Documentation/networking/device_drivers')
-rw-r--r-- | Documentation/networking/device_drivers/ethernet/amazon/ena.rst | 32 |
1 files changed, 32 insertions, 0 deletions
diff --git a/Documentation/networking/device_drivers/ethernet/amazon/ena.rst b/Documentation/networking/device_drivers/ethernet/amazon/ena.rst index 8bcb173e0353..491492677632 100644 --- a/Documentation/networking/device_drivers/ethernet/amazon/ena.rst +++ b/Documentation/networking/device_drivers/ethernet/amazon/ena.rst @@ -205,6 +205,7 @@ Adaptive coalescing can be switched on/off through `ethtool(8)`'s More information about Adaptive Interrupt Moderation (DIM) can be found in Documentation/networking/net_dim.rst +.. _`RX copybreak`: RX copybreak ============ The rx_copybreak is initialized by default to ENA_DEFAULT_RX_COPYBREAK @@ -315,3 +316,34 @@ Rx - The new SKB is updated with the necessary information (protocol, checksum hw verify result, etc), and then passed to the network stack, using the NAPI interface function :code:`napi_gro_receive()`. + +Dynamic RX Buffers (DRB) +------------------------ + +Each RX descriptor in the RX ring is a single memory page (which is either 4KB +or 16KB long depending on system's configurations). +To reduce the memory allocations required when dealing with a high rate of small +packets, the driver tries to reuse the remaining RX descriptor's space if more +than 2KB of this page remain unused. + +A simple example of this mechanism is the following sequence of events: + +:: + + 1. Driver allocates page-sized RX buffer and passes it to hardware + +----------------------+ + |4KB RX Buffer | + +----------------------+ + + 2. A 300Bytes packet is received on this buffer + + 3. The driver increases the ref count on this page and returns it back to + HW as an RX buffer of size 4KB - 300Bytes = 3796 Bytes + +----+--------------------+ + |****|3796 Bytes RX Buffer| + +----+--------------------+ + +This mechanism isn't used when an XDP program is loaded, or when the +RX packet is less than rx_copybreak bytes (in which case the packet is +copied out of the RX buffer into the linear part of a new skb allocated +for it and the RX buffer remains the same size, see `RX copybreak`_). |