sfc: reduce the number of requested xdp ev queues
Subsystems (4)
PATCH 1/1 sfc: reduce the number of requested xdp ev queues
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
--- drivers/net/ethernet/sfc/efx_channels.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/net/ethernet/sfc/efx_channels.c +4 -2
@@ -17,6 +17,7 @@#include "rx_common.h" #include "nic.h" #include "sriov.h" +#include "workarounds.h" /* This is the first interrupt mode to try out of: * 0 => MSI-X
@@ -137,6 +138,7 @@ static int efx_allocate_msix_channels(struct efx_nic *efx, {unsigned int n_channels = parallelism;int vec_count; +int tx_per_ev;int n_xdp_tx;int n_xdp_ev;
@@ -149,9 +151,9 @@ static int efx_allocate_msix_channels(struct efx_nic *efx,* multiple tx queues, assuming tx and ev queues are both * maximum size. */ - + tx_per_ev =EFX_MAX_EVQ_SIZE /EFX_TXQ_MAX_ENT (efx); n_xdp_tx =num_possible_cpus (); - n_xdp_ev =DIV_ROUND_UP (n_xdp_tx,EFX_MAX_TXQ_PER_CHANNEL ); + n_xdp_ev =DIV_ROUND_UP (n_xdp_tx, tx_per_ev); vec_count =pci_msix_vec_count (efx->pci_dev);if (vec_count <0 )
Signature · 1 line
2.29.2
Quoted 9 lines · On Mon, 14 Dec 2020 17:29:06 -0800 Ivan Babrou <[email protected]> wrote:
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
I have a similar QA report with XDP_REDIRECT:
sfc 0000:05:00.0 ens1f0np0: XDP redirect failed (-22)
Here we are back to the issue we discussed with ixgbe, that NIC / msi-x interrupts hardware resources are not enough on machines with many logical cpus.
After this fix, what will happen if (cpu >= efx->xdp_tx_queue_count) ? (Copied efx_xdp_tx_buffers code below signature)
The question leads to, does this driver need a fallback mechanism when HW resource or systems logical cpus exceed the one TX-queue per CPU assumption?
Quoted 38 lines
Signed-off-by Ivan Babrou <[email protected]>--- drivers/net/ethernet/sfc/efx_channels.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/drivers/net/ethernet/sfc/efx_channels.cindexa4a626e9cd9a..1bfeee283ea9100644 ---a/drivers/net/ethernet/sfc/efx_channels.c+++b/drivers/net/ethernet/sfc/efx_channels.c@@ -17,6 +17,7 @@
#include "rx_common.h"
#include "nic.h"
#include "sriov.h"drivers/net/ethernet/sfc/efx_channels.c +4
-2+#include "workarounds.h" /* This is the first interrupt mode to try out of: * 0 => MSI-X @@ -137,6 +138,7 @@ static int efx_allocate_msix_channels(struct efx_nic *efx, {unsigned int n_channels = parallelism;int vec_count; +int tx_per_ev;int n_xdp_tx;int n_xdp_ev;@@ -149,9 +151,9 @@ static int efx_allocate_msix_channels(structefx_nic *efx, * multiple tx queues, assuming tx and ev queues are both * maximum size. */ - + tx_per_ev =EFX_MAX_EVQ_SIZE /EFX_TXQ_MAX_ENT (efx); n_xdp_tx =num_possible_cpus (); - n_xdp_ev =DIV_ROUND_UP (n_xdp_tx,EFX_MAX_TXQ_PER_CHANNEL ); + n_xdp_ev =DIV_ROUND_UP (n_xdp_tx, tx_per_ev); vec_count =pci_msix_vec_count (efx->pci_dev);if (
vec_count< 0)
Signature · 85 lines
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer
/* Transmit a packet from an XDP buffer
*
* Returns number of packets sent on success, error code otherwise.
* Runs in NAPI context, either in our poll (for XDP TX) or a different NIC
* (for XDP redirect).
*/
int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
bool flush)
{
struct efx_tx_buffer *tx_buffer;
struct efx_tx_queue *tx_queue;
struct xdp_frame *xdpf;
dma_addr_t dma_addr;
unsigned int len;
int space;
int cpu;
int i;
cpu = raw_smp_processor_id();
if (!efx->xdp_tx_queue_count ||
unlikely(cpu >= efx->xdp_tx_queue_count))
return -EINVAL;
tx_queue = efx->xdp_tx_queues[cpu];
if (unlikely(!tx_queue))
return -EINVAL;
if (unlikely(n && !xdpfs))
return -EINVAL;
if (!n)
return 0;
/* Check for available space. We should never need multiple
* descriptors per frame.
*/
space = efx->txq_entries +
tx_queue->read_count - tx_queue->insert_count;
for (i = 0; i < n; i++) {
xdpf = xdpfs[i];
if (i >= space)
break;
/* We'll want a descriptor for this tx. */
prefetchw(__efx_tx_queue_get_insert_buffer(tx_queue));
len = xdpf->len;
/* Map for DMA. */
dma_addr = dma_map_single(&efx->pci_dev->dev,
xdpf->data, len,
DMA_TO_DEVICE);
if (dma_mapping_error(&efx->pci_dev->dev, dma_addr))
break;
/* Create descriptor and set up for unmapping DMA. */
tx_buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
tx_buffer->xdpf = xdpf;
tx_buffer->flags = EFX_TX_BUF_XDP |
EFX_TX_BUF_MAP_SINGLE;
tx_buffer->dma_offset = 0;
tx_buffer->unmap_len = len;
tx_queue->tx_packets++;
}
/* Pass mapped frames to hardware. */
if (flush && i > 0)
efx_nic_push_buffers(tx_queue);
if (i == 0)
return -EIO;
efx_xdp_return_frames(n - i, xdpfs + i);
return i;
}Quoted 13 lines · Jesper Dangaard Brouer
Quoted 9 lines · On Mon, 14 Dec 2020 17:29:06 -0800 Ivan Babrou <[email protected]> wrote:
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
Quoted 8 lines
I have a similar QA report with
XDP_REDIRECT:
sfc 0000:05:00.0 ens1f0np0: XDP redirect failed (-22)Here we are back to the issue we discussed with ixgbe, that NIC / msi-x interrupts hardware resources are not enough on machines with many logical cpus.
After this fix, what will happen if (cpu >=
efx->xdp_tx_queue_count) ?
Same as happened before: the "failed -22". But this fix will make that
less likely to happen, because it ties more TXQs to each EVQ, and it's the EVQs that are in short supply.
(Strictly speaking, I believe the limitation is a software one, that
comes from the driver's channel structures having been designed a decade ago when 32 cpus ought to be enough for anybody... AFAIR the hardware is capable of giving us something like 1024 evqs if we ask for them, it just might not have that many msi-x vectors for us.)
Anyway, the patch looks correct, so
-ed
Quoted 11 lines · Ivan Babrou
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
Signed-off-by Ivan Babrou <[email protected]>
Quoted 38 lines
--- drivers/net/ethernet/sfc/efx_channels.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)drivers/net/ethernet/sfc/efx_channels.c +4
-2@@ -17,6 +17,7 @@#include "rx_common.h" #include "nic.h" #include "sriov.h" +#include "workarounds.h" /* This is the first interrupt mode to try out of: * 0 => MSI-X @@ -137,6 +138,7 @@ static int efx_allocate_msix_channels(struct efx_nic *efx, {unsigned int n_channels = parallelism;int vec_count; +int tx_per_ev;int n_xdp_tx;int n_xdp_ev;@@ -149,9 +151,9 @@ static int efx_allocate_msix_channels(struct efx_nic *efx,* multiple tx queues, assuming tx and ev queues are both * maximum size. */ - + tx_per_ev =EFX_MAX_EVQ_SIZE /EFX_TXQ_MAX_ENT (efx); n_xdp_tx =num_possible_cpus (); - n_xdp_ev =DIV_ROUND_UP (n_xdp_tx,EFX_MAX_TXQ_PER_CHANNEL ); + n_xdp_ev =DIV_ROUND_UP (n_xdp_tx, tx_per_ev); vec_count =pci_msix_vec_count (efx->pci_dev);if (vec_count <0 )Signature · 1 line
2.29.2
Signature · 1 line
Martin Habets <[email protected]>
Quoted 21 lines · On Tue, 15 Dec 2020 18:49:55 +0000 Edward Cree <[email protected]> wrote:
Quoted 20 lines · Jesper Dangaard Brouer
Quoted 9 lines · On Mon, 14 Dec 2020 17:29:06 -0800 Ivan Babrou <[email protected]> wrote:
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
I have a similar QA report with
XDP_REDIRECT:
sfc 0000:05:00.0 ens1f0np0: XDP redirect failed (-22)Here we are back to the issue we discussed with ixgbe, that NIC / msi-x interrupts hardware resources are not enough on machines with many logical cpus.
Quoted 6 lines
Quoted 1 line
After this fix, what will happen if (cpu >=
efx->xdp_tx_queue_count) ?Same as happened before: the "failed -22". But this fix will make that
less likely to happen, because it ties more TXQs to each EVQ, and it's the EVQs that are in short supply.
So, what I hear is that this fix is just pampering over the real issue.
I suggest that you/we detect the situation, and have a code path that will take a lock (per 16 packets bulk) and solve the issue.
If you care about maximum performance you can implement this via changing the ndo_xdp_xmit pointer to the fallback function when needed, to avoid having a to check for the fallback mode in the fast-path.
Quoted 8 lines
(Strictly speaking, I believe the limitation is a software one, that
comes from the driver's channel structures having been designed a decade ago when 32 cpus ought to be enough for anybody... AFAIR the hardware is capable of giving us something like 1024 evqs if we ask for them, it just might not have that many msi-x vectors for us.)
Anyway, the patch looks correct, soAcked-by Edward Cree <[email protected]>
Signature · 4 lines
Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer
Edward Cree
maintainerQuoted 1 line · Jesper Dangaard Brouer
So, what I hear is that this fix is just pampering over the real issue.
Yes, it is, but it's better than nothing in the meantime while we work
out the complete fix.
Quoted 2 lines
I suggest that you/we detect the situation, and have a code path that will take a lock (per 16 packets bulk) and solve the issue.
Imho that would _also_ paper over the issue, because it would mean the
system degraded to a lower performance mode of operation, while still appearing to support XDP_TX. I think that that in general should not happen unless there is a way for the user to determine at runtime whether it has/should happen. Perhaps Marek's new XDP feature flags could include a "tx-lockless" flag to indicate this?
-ed
Jakub Kicinski
maintainerQuoted 11 lines · On Mon, 14 Dec 2020 17:29:06 -0800 Ivan Babrou wrote:
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
Signed-off-by Ivan Babrou <[email protected]>
Looks like the discussion may have concluded, but we don't take -next patches during the merge window, so please repost when net-next reopens.
Thanks!
--
# Form letter - net-next is closed
We have already sent the networking pull request for 5.11 and therefore net-next is closed for new drivers, features, code refactoring and optimizations. We are currently accepting bug fixes only.
Please repost when net-next reopens after 5.11-rc1 is cut.
Look out for the announcement on the mailing list or check:
http://vger.kernel.org/~davem/net-next.html
RFC patches sent for review only are obviously welcome at any time.
Quoted 21 lines · Jakub Kicinski
Quoted 11 lines · On Mon, 14 Dec 2020 17:29:06 -0800 Ivan Babrou wrote:
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
Signed-off-by Ivan Babrou <[email protected]>Looks like the discussion may have concluded, but we don't take -next patches during the merge window, so please repost when net-next reopens.
Thanks!
--
# Form letter - net-next is closed
Quoted 5 lines
We have already sent the networking pull request for 5.11 and therefore net-next is closed for new drivers, features, code refactoring and optimizations. We are currently accepting bug fixes only.
Please repost when net-next reopens after 5.11-rc1 is cut.
Should I resend my patch now that the window is open or is bumping this thread enough?
Quoted 4 lines
Look out for the announcement on the mailing list or check:
http://vger.kernel.org/~davem/net-next.htmlRFC patches sent for review only are obviously welcome at any time.
Jakub Kicinski
maintainerQuoted 21 lines · On Tue, 19 Jan 2021 15:43:43 -0800 Ivan Babrou wrote:
Quoted 20 lines · Jakub Kicinski
Quoted 11 lines · On Mon, 14 Dec 2020 17:29:06 -0800 Ivan Babrou wrote:
Without this change the driver tries to allocate too many queues, breaching the number of available msi-x interrupts on machines with many logical cpus and default adapter settings:
Insufficient resources for 12 XDP event queues (24 other channels, max 32)
Which in turn triggers EINVAL on XDP processing:
sfc 0000:86:00.0 ext0: XDP TX failed (-22)
Signed-off-by Ivan Babrou <[email protected]>Looks like the discussion may have concluded, but we don't take -next patches during the merge window, so please repost when net-next reopens.
Thanks!
--
# Form letter - net-next is closed
Quoted 8 lines
Quoted 5 lines
We have already sent the networking pull request for 5.11 and therefore net-next is closed for new drivers, features, code refactoring and optimizations. We are currently accepting bug fixes only.
Please repost when net-next reopens after 5.11-rc1 is cut.
Should I resend my patch now that the window is open or is bumping this thread enough?
You need to resend, thanks!