Align XDP Queues with RSS Better #4484

nibanks · 2024-08-22T16:32:51Z

Description

Updates the XDP datapath layer (on Windows) to try to query the RSS processor before fully creating the queues, so that the real queue may be correctly aligned with RSS.

Testing

CI/CD

Documentation

N/A

codecov · 2024-08-23T14:00:35Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.60%. Comparing base (733da8d) to head (3b66352).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4484      +/-   ##
==========================================
+ Coverage   86.37%   86.60%   +0.22%     
==========================================
  Files          56       56              
  Lines       15537    17354    +1817     
==========================================
+ Hits        13420    15029    +1609     
- Misses       2117     2325     +208

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

src/platform/datapath_raw_xdp_win.c

nibanks · 2024-08-29T14:47:42Z

src/platform/datapath_raw_xdp_win.c

+        }
+        if (QueueIndex == UINT32_MAX) {
+            Status = QUIC_STATUS_INVALID_STATE;
+            QuicTraceEvent(


@mtfriesen I'm hitting this error in test (definitely over duonic, but also netperf is failing, but I don't have logs). I'm assuming that means we're trying to bind a queue to a processor that doesn't have an RSS queue? I'm trying to decide how to handle this. Thoughts?

More queues than processors in the indirection table is completely legal. I'd just leave any leftover queues unaffinitized and hope for the best.

I'm actually thinking it might be the opposite. We have more processors than queues. I actually don't know what duonic does here. Do you?

Hmm, ok, I'm not that familiar with this QUIC code. NICs just indicate packets on the processors in the indirection table - the rest are idle.

nibanks · 2024-08-29T15:58:52Z

src/platform/datapath_raw_xdp_win.c

+            }
+        }
+        if (!Found) {
+            CXPLAT_FRE_ASSERT(FALSE); // TODO - What do we do if there is no partition for this processor?


@mtfriesen thoughts here? We allow an app to specify the CPUs they want us to run on. What do we do if there is an RSS queue that doesn't have a CPU being run on?

I think the question you are asking here is the question I answered in the incorrect location above - a queue without an assigned processor is effectively not a queue.

mtfriesen · 2024-09-11T15:32:56Z

src/platform/datapath_raw_xdp_win.c

-        if (0 == uReturn) {
+        uint32_t Flags = XSK_BIND_FLAG_TX;
+        Status = Xdp->XdpApi->XskBind(TxXsk, InterfaceIndex, i, Flags);
+        if (QUIC_FAILED(Status)) { // No more queues. Break out.


there could be other errors than just queue count exceeded. probably worth logging the error code and current index explicitly.

You want me to just log this one?

this and the "activate" are the only two that might fail in nontrivial and unpredictable ways. it's probably worth logging them all if you have time.

mtfriesen · 2024-09-11T15:36:36Z

src/platform/datapath_raw_xdp_win.c

-            free(rssTable);
-        }
+        XSK_NOTIFY_RESULT_FLAGS OutFlags;
+        Status = Xdp->XdpApi->XskNotifySocket(TxXsk, XSK_NOTIFY_FLAG_POKE_TX|XSK_NOTIFY_FLAG_WAIT_TX, 1000, &OutFlags);


1000 magic number - should be #define ?

mtfriesen · 2024-09-11T15:38:15Z

src/platform/datapath_raw_xdp_win.c


-Cleanup:
+        const CXPLAT_PROCESSOR_GROUP_INFO* Group = &CxPlatProcessorGroupInfo[ProcNumber.Group];
+        Queues[i] = Group->Offset + (ProcNumber.Number % Group->Count);


why is the modulo necessary? i probably just need to go read more of the QUIC code, but i would have expected all ProcNumber.Numbers to be less than Group->Count.

I think it shouldn't be necessary.

is it for hot add? where QUIC has a fixed set of CPUs per group that might see more CPUs added to the group later?

No, we fix this list at process start.

right, so if the group is fixed at size 1 during process start, and the OS later adds proc 2 to the group, then that's why you have the modulo?

mtfriesen · 2024-09-11T15:38:41Z

src/platform/datapath_raw_xdp_win.c

@@ -480,6 +372,9 @@ CxPlatDpRawInterfaceInitialize(
    Interface->OffloadStatus.Transmit.NetworkLayerXsum = Xdp->SkipXsum;
    Interface->Xdp = Xdp;

+    uint32_t Processors[256]; // TODO - Use max processor count


unfinished TODO

mtfriesen · 2024-09-11T15:39:24Z

src/platform/datapath_raw_xdp_win.c

@@ -521,6 +427,7 @@ CxPlatDpRawInterfaceInitialize(
    for (uint8_t i = 0; i < Interface->QueueCount; i++) {
        XDP_QUEUE* Queue = &Interface->Queues[i];

+        Queue->RssProcessor = (uint16_t)Processors[i]; // TODO - Should memory be aligned with this?


as in, should all allocations for this queue be aligned to that processor? ideally yes, for sure.

Exactly, this is a larger issue too. For all memory allocation that is meant to be used on a particular process, should we temporarily affinitize the thread, allocate the memory and then revert the affinitization?

in user mode, there are fancy APIs to allocate from a specific node, even if you're not currently running on that node. in kernel mode, it's more difficult to allocate pool from an arbitrary node, though i think newer OSes have better APIs. one sec.

that said, XDP itself treats the calling thread as the "ideal" thread when it performs its own allocations for queues, so temporarily setting the affinity would probably solve the problem the most universally and with the least fuss.

yeah, PoolExtendedParameterNumaNode is supported for ExAllocatePool3. i'd still just set the thread affinity and be done.

So, if we're all on a single NUMA, how much would this actually matter?

if you have UMA (idk if that's a proper term) then yeah, just do nothing.

For VMs that are currently affixed to just one of the NUMA nodes, are we sure there's no per-proc memory perf involved? Also, can we be guaranteed that the VM actually knows the NUMA of the HW?

AFAIK the VM can't do anything to optimize that scenario. It's all hidden and controlled by the host/hypervisor.

nibanks added 5 commits August 22, 2024 11:15

Initial WIP function

70e299c

Successfully compiles

4f1a414

Find right processor queue

47eb702

Fix build

b1ff11c

nits

14330bf

nibanks added the Area: Performance label Aug 22, 2024

nibanks added 3 commits August 23, 2024 09:12

Merge branch 'main' into xdp-get-queue-procs

5586c62

Works on my machine now

d6fbdec

No sleep

25150a2

nibanks marked this pull request as ready for review August 23, 2024 18:26

nibanks requested a review from a team as a code owner August 23, 2024 18:26

nibanks added 2 commits August 23, 2024 14:55

Merge branch 'main' into xdp-get-queue-procs

b115b40

Merge branch 'main' into xdp-get-queue-procs

1fc88e5

nibanks commented Aug 29, 2024

View reviewed changes

src/platform/datapath_raw_xdp_win.c Show resolved Hide resolved

nibanks commented Aug 29, 2024

View reviewed changes

Let's try this instead

ca5d8e6

nibanks commented Aug 29, 2024

View reviewed changes

nibanks added 4 commits August 29, 2024 13:27

Logs and round robin

4a6929c

clog again

984d142

Missed the sidecar

9936166

Merge branch 'main' into xdp-get-queue-procs

3b66352

mtfriesen reviewed Sep 11, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Align XDP Queues with RSS Better #4484

Align XDP Queues with RSS Better #4484

nibanks commented Aug 22, 2024

codecov bot commented Aug 23, 2024 •

edited

Loading

nibanks Aug 29, 2024

mtfriesen Aug 29, 2024

nibanks Aug 29, 2024

mtfriesen Aug 29, 2024

nibanks Aug 29, 2024

mtfriesen Aug 29, 2024

mtfriesen Sep 11, 2024 •

edited

Loading

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

mtfriesen Sep 11, 2024

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

mtfriesen Sep 11, 2024

mtfriesen Sep 11, 2024

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

mtfriesen Sep 11, 2024 •

edited

Loading

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

nibanks Sep 11, 2024

mtfriesen Sep 11, 2024

Align XDP Queues with RSS Better #4484

Are you sure you want to change the base?

Align XDP Queues with RSS Better #4484

Conversation

nibanks commented Aug 22, 2024

Description

Testing

Documentation

codecov bot commented Aug 23, 2024 • edited Loading

Codecov Report

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mtfriesen Sep 11, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mtfriesen Sep 11, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

codecov bot commented Aug 23, 2024 •

edited

Loading

mtfriesen Sep 11, 2024 •

edited

Loading

mtfriesen Sep 11, 2024 •

edited

Loading