On 04/07/2022 09:24, Greg KH wrote:
CAUTION: Email originated externally, do not click links or open attachments unless you recognize the sender and know the content is safe.
On Mon, Jul 04, 2022 at 09:04:07AM +0200, Łukasz Spintzyk wrote:
DisplayLink ethernet devices require NTB buffers larger then 32kb in order to run with highest performance.
Signed-off-by: Łukasz Spintzyk <lukasz.spintzyk@xxxxxxxxxxxxx>
---
include/linux/usb/cdc_ncm.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/include/linux/usb/cdc_ncm.h b/include/linux/usb/cdc_ncm.h
index f7cb3ddce7fb..2d207cb4837d 100644
--- a/include/linux/usb/cdc_ncm.h
+++ b/include/linux/usb/cdc_ncm.h
@@ -53,8 +53,8 @@
#define USB_CDC_NCM_NDP32_LENGTH_MIN 0x20
/* Maximum NTB length */
-#define CDC_NCM_NTB_MAX_SIZE_TX 32768 /* bytes */
-#define CDC_NCM_NTB_MAX_SIZE_RX 32768 /* bytes */
+#define CDC_NCM_NTB_MAX_SIZE_TX 65536 /* bytes */
+#define CDC_NCM_NTB_MAX_SIZE_RX 65536 /* bytes */
Does this mess with the throughput of older devices that are not on
displaylink connections?
What devices did you test this on, and what is the actual performance
changes? You offer no real information here at all, and large buffer
sizes does have other downsides, so determining how you tested this is
key.
Also, please wrap your changelogs at 72 columns like git asks you to do.
thanks,
greg k-h
Hi Greg,
To my best knowledge that patch should not affect other devices because:
- tx,rx buffers size is initialized to 16kb with
CDC_NCM_NTB_DEF_SIZE_RX, and CDC_NCM_NTB_DEF_SIZE_TX
So all existing devices should not be affected by default.
- In order to change tx and rx buffer max size you need to
additionally modify cdc_ncm/tx_max and cdc_ncm/rx_max parameters. This
can be done with udev rules and ethtool. So if you want to use higher
buffer sizes you need to specially request that.
For DisplayLink devices this will be done with udev rule that will be
installed with other DisplayLink drivers.
- This tx,rx buffer sizes are always capped to dwNtbMaxInMaxSize and
dwNtbMaxOutMaxSize that are advertised by the device itself. So in
theory that values should be acceptable by the device.
Here is summary of my tests I have done on such devices:
- DisplayLink DL-3xxx family device
- DisplayLink DL-6xxx family device
- ASUS USB-C2500 2.5G USB3 eth adapter
- Plugable USB3 1G USB3 adapter
- EDIMAX EU-4307 USB-C adapter
- Dell DBQBCBC064 USB-C adapter
Unfortunately I was not able to find more or older than USB-3 device on
company's shelf.
I was doing measurements with:
- iperf3 between two linux boxes
- http://openspeedtest.com/ instance running on local testing machine
I can provide you with detailed results, but I think they are quite
verbose so I will stay with some high level results:
- All except one from third party usb adapters were not affected by
increased buffer size. (I have forced them to use tx,rx size as big as
their advertised dwNtbOutMaxSize and dwNtbInMaxSize).
They were generally reaching 912 - 940Mbps (download/upload)
Only Edimax adapter experienced decreased download size from 929Mbps
to 827 with iper3. In openspeedtest this was decrease from 968Mbps to
886Mbps
- DisplayLink DL-3xxx family devices experienced increase of performance
Iperf3:
Download from (300Mbps to 870Mbps), also from historical
measurements on other setup this was (90Mbps to 500Mbps)
Upload from 782Mbps to 844Mbps
Openspeedtest:
Download from 556Mbps to 873Mbps
Upload from 727Mbps to 973Mbps
- DiplayLink DL-6xxx family devices are not affected greatly by buffer
size. It is more affected by patch enabling ZLP which prevents device
from temporary network dropouts when playing video from web and network
traffic going through it is high.
thanks
Łukasz Spintzyk
