bnx2: Increase max rx ring size from 1K to 2K
Subsystems (2)
PATCH 1/1 bnx2: Increase max rx ring size from 1K to 2K
A number of customers are reporting packet loss under certain workloads (e.g. heavy bursts of small packets) with flow control disabled. A larger rx ring helps to prevent these losses.
No change in default rx ring size and memory consumption.
--- drivers/net/bnx2.h | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-)
drivers/net/bnx2.h +2 -2
@@ -6502,8 +6502,8 @@ struct l2_fhdr {#define TX_DESC_CNT (BCM_PAGE_SIZE / sizeof(struct tx_bd))#define MAX_TX_DESC_CNT (TX_DESC_CNT - 1) -#define MAX_RX_RINGS 4 -#define MAX_RX_PG_RINGS 16 +#define MAX_RX_RINGS 8 +#define MAX_RX_PG_RINGS 32#define RX_DESC_CNT (BCM_PAGE_SIZE / sizeof(struct rx_bd))#define MAX_RX_DESC_CNT (RX_DESC_CNT - 1)#define MAX_TOTAL_RX_DESC_CNT (MAX_RX_DESC_CNT * MAX_RX_RINGS)
Signature · 1 line
1.6.4.GIT
David Miller
maintainerFrom: "Michael Chan" <[email protected]>
Date: Mon, 18 Oct 2010 17:30:54 -0700
Quoted 9 lines
A number of customers are reporting packet loss under certain workloads (e.g. heavy bursts of small packets) with flow control disabled. A larger rx ring helps to prevent these losses.
No change in default rx ring size and memory consumption.
Signed-off-by Andy Gospodarek <[email protected]>Acked-by John Feeney <[email protected]>Signed-off-by Michael Chan <[email protected]>
I don't see how it's any better to queue things more deeply in hardware, compared to simply using hardware flow control since that's what it's for and makes the queuing happen in the networking stack of the sender, in software, which in the end performs better and gives better feedback to the source of the data.
These huge RX queue sizes are absolutely rediculious, and I've complained about this before.
And instead of seeing less of this, I keep seeing more of this stuff. Please exert some pushback on these folks who are doing such insane things.
Thanks.
Quoted 13 lines · David Miller
From: "Michael Chan" <[email protected]>
Date: Mon, 18 Oct 2010 17:30:54 -0700Quoted 9 lines
A number of customers are reporting packet loss under certain workloads (e.g. heavy bursts of small packets) with flow control disabled. A larger rx ring helps to prevent these losses.
No change in default rx ring size and memory consumption.
Signed-off-by Andy Gospodarek <[email protected]>Acked-by John Feeney <[email protected]>Signed-off-by Michael Chan <[email protected]>
Quoted 5 lines
I don't see how it's any better to queue things more deeply in hardware, compared to simply using hardware flow control since that's what it's for and makes the queuing happen in the networking stack of the sender, in software, which in the end performs better and gives better feedback to the source of the data.
There are situations where flow control is not desirable. For example, if there are many multicast receivers in the network, you may not want a few slow receivers to slow down the entire network with flow control.
Quoted 3 lines
These huge RX queue sizes are absolutely rediculious, and I've complained about this before.
Yes you have and I was initially hesitant to post this patch. But the customer sees that many other 1G drivers in the tree can have bigger ring sizes and as a result, they have fewer packet drops using these other devices. In fact, 2K is still much smaller than many other 1G drivers in the tree. Please also note that this does not add any extra memory to the default configuration. Thanks.
Quoted 7 lines
And instead of seeing less of this, I keep seeing more of this stuff. Please exert some pushback on these folks who are doing such insane things.
Thanks.
David Miller
maintainerFrom: "Michael Chan" <[email protected]>
Date: Mon, 18 Oct 2010 17:30:54 -0700
Quoted 9 lines
A number of customers are reporting packet loss under certain workloads (e.g. heavy bursts of small packets) with flow control disabled. A larger rx ring helps to prevent these losses.
No change in default rx ring size and memory consumption.
Signed-off-by Andy Gospodarek <[email protected]>Acked-by John Feeney <[email protected]>Signed-off-by Michael Chan <[email protected]>
Ok, since the new limit is not the default, applied.
Thanks for the explanation Michael.