infra: add TX queue sharing support #474
Conversation
📝 WalkthroughWalkthroughAdds per-TX-queue spinlocks and propagates TX-queue-aware context and wiring through control and datapath. 📜 Recent review detailsConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro 📒 Files selected for processing (7)
🚧 Files skipped from review as they are similar to previous changes (2)
🧰 Additional context used📓 Path-based instructions (1)**/*.{c,h}⚙️ CodeRabbit configuration file
Files:
🧠 Learnings (3)📚 Learning: 2025-11-05T14:28:28.817ZApplied to files:
📚 Learning: 2025-11-03T13:28:19.489ZApplied to files:
📚 Learning: 2025-12-17T17:32:21.746ZApplied to files:
🧬 Code graph analysis (4)modules/infra/control/worker.c (3)
modules/infra/control/worker_test.c (1)
modules/infra/control/port.c (1)
modules/infra/datapath/gr_rxtx.h (1)
🪛 Cppcheck (2.19.0)modules/infra/control/worker.c[error] 231-231: Expression 'worker->txqs?free(__gr_vec_hdr(worker->txqs)) (unknownEvaluationOrder) ⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (7)
🔇 Additional comments (7)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Fix all issues with AI agents
In @modules/infra/control/port.c:
- Around line 113-121: When forcing TX queue count for drivers "net_tap" or
"net_virtio" (check info.driver_name), validate that p->n_rxq <=
info.max_tx_queues before setting p->n_txq = p->n_rxq; if it exceeds, return a
clear error (mirror the RX overflow handling) instead of assigning and proceed
only when valid, and only initialize txq_locks after the successful assignment.
Use the same error/return style as the RX overflow path and reference p->n_rxq,
p->n_txq, and info.max_tx_queues in the message so the failure is explicit (this
change affects the block that currently assigns p->n_txq and the subsequent
rte_spinlock_init loop).
In @modules/infra/control/worker.c:
- Around line 234-245: The loop that assigns tx queue IDs can divide by zero
because it uses txq_idx % port->n_txq; in gr_vec_foreach (port, ports) before
computing queue_id add a guard that checks if port->n_txq == 0 and skip mapping
for that port (or set enabled=false) so you don't perform the modulo; update the
block around STAILQ_FOREACH (worker, &workers, next) where struct queue_map is
constructed to only compute .queue_id = txq_idx % port->n_txq when port->n_txq >
0 and otherwise handle the case by continuing the worker loop or setting a safe
default, then increment txq_idx only when a mapping was actually added to
worker->txqs.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (6)
modules/infra/control/gr_port.hmodules/infra/control/graph.cmodules/infra/control/port.cmodules/infra/control/worker.cmodules/infra/datapath/gr_rxtx.hmodules/infra/datapath/port_tx.c
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{c,h}
⚙️ CodeRabbit configuration file
**/*.{c,h}: -gr_vec_*()functions cannot fail. No need to check their return value.
gr_vec_free(x)always setsx = NULL. There is no risk of double free.ec_node_*()functions consume theirec_nodearguments. No leaks on error.rte_node->ctxis an uint8_t array of size 16, not a pointer.- Never suggest to replace
assert()with graceful error checking.- We compile with
-std=gnu2x. Unnamed parameters in function definitions are valid.
Files:
modules/infra/control/worker.cmodules/infra/control/port.cmodules/infra/datapath/port_tx.cmodules/infra/datapath/gr_rxtx.hmodules/infra/control/gr_port.hmodules/infra/control/graph.c
🧠 Learnings (5)
📚 Learning: 2025-11-05T14:28:28.817Z
Learnt from: rjarry
Repo: DPDK/grout PR: 372
File: modules/infra/datapath/loop_input.c:53-56
Timestamp: 2025-11-05T14:28:28.817Z
Learning: In modules/infra/datapath/loop_input.c, the l3_edges array is intentionally indexed using network byte order (rte_be16_t) for EtherType values. Both loopback_input_add_type() registration and loopback_input_process() lookup use network byte order consistently, so no byte order conversion is needed when accessing l3_edges.
Applied to files:
modules/infra/datapath/port_tx.cmodules/infra/control/gr_port.h
📚 Learning: 2025-11-03T13:28:19.489Z
Learnt from: rjarry
Repo: DPDK/grout PR: 379
File: modules/infra/datapath/port_tx.c:36-46
Timestamp: 2025-11-03T13:28:19.489Z
Learning: In DPDK graph node process callbacks, the return value is only used for statistics and does not affect packet flow or scheduling through the graph. Nodes can return 0 when they haven't processed packets (e.g., when dropping or redirecting due to port state).
Applied to files:
modules/infra/datapath/port_tx.cmodules/infra/control/graph.c
📚 Learning: 2025-12-17T17:32:21.746Z
Learnt from: christophefontaine
Repo: DPDK/grout PR: 466
File: modules/srv6/datapath/l2_encap.c:26-32
Timestamp: 2025-12-17T17:32:21.746Z
Learning: The GR_MBUF_PRIV_DATA_TYPE macro in modules/infra/datapath/gr_mbuf.h automatically adds `const struct iface *iface` as the first field to any structure defined with it. All types defined using this macro (e.g., srv6_dx2_mbuf_data, mbuf_data, queue_mbuf_data) will have the iface field available, followed by any custom fields provided as the second macro argument.
Applied to files:
modules/infra/datapath/port_tx.cmodules/infra/control/gr_port.h
📚 Learning: 2025-09-05T08:58:53.047Z
Learnt from: rjarry
Repo: DPDK/grout PR: 294
File: modules/policy/control/conntrack.c:533-539
Timestamp: 2025-09-05T08:58:53.047Z
Learning: DPDK functions rte_mempool_free() and rte_hash_free() are NULL-safe - they explicitly handle NULL pointers and do nothing when passed NULL, unlike many C standard library functions.
Applied to files:
modules/infra/control/gr_port.h
📚 Learning: 2025-09-09T12:08:36.790Z
Learnt from: rjarry
Repo: DPDK/grout PR: 294
File: modules/policy/control/conntrack.c:621-637
Timestamp: 2025-09-09T12:08:36.790Z
Learning: DPDK's rte_hash supports safe deletion during iteration via rte_hash_iterate, unlike many other hash table implementations that require two-phase deletion patterns.
Applied to files:
modules/infra/control/gr_port.h
🧬 Code graph analysis (3)
modules/infra/control/worker.c (1)
main/gr_vec.h (1)
gr_vec_len(47-51)
modules/infra/control/port.c (1)
api/gr_errno.h (1)
errno_set(10-13)
modules/infra/control/graph.c (2)
modules/infra/control/worker_test.c (1)
q(150-157)main/gr_vec.h (1)
gr_vec_len(47-51)
🪛 Cppcheck (2.19.0)
modules/infra/control/worker.c
[error] 231-231: Expression 'worker->txqs?free(__gr_vec_hdr(worker->txqs))
(unknownEvaluationOrder)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (5)
- GitHub Check: build-and-tests (clang-15, none, debugoptimized, ubuntu-22.04, false)
- GitHub Check: build-and-tests (gcc-14, address, debug, ubuntu-24.04, -Dfrr=enabled, true)
- GitHub Check: build-and-tests (gcc-13, none, debugoptimized, ubuntu-24.04, -Dfrr=enabled, false)
- GitHub Check: deb
- GitHub Check: rpm
🔇 Additional comments (1)
modules/infra/control/worker.c (1)
230-232: The macrogr_vec_free(v)uses the comma operator to sequence operations:((v) ? free(__gr_vec_hdr(v)) : (void)0, (v) = NULL). The comma operator is a sequence point, ensuring the conditional free is fully evaluated before the assignment. No undefined behavior exists when passing a member lvalue likeworker->txqs. Cppcheck's report is a false positive.Likely an incorrect or invalid review comment.
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Fix all issues with AI agents
In @modules/infra/control/worker.c:
- Around line 225-248: The assign_txqs function can divide by zero when
computing queue_id with txq_idx % port->n_txq; fix it by checking port->n_txq
before using the modulus: inside assign_txqs (and the inner STAILQ_FOREACH loop
over worker), if port->n_txq == 0 either skip creating/adding the txq for that
port (recommended) or set txq.queue_id = 0 and txq.enabled = false to avoid the
modulus; adjust the code that creates the struct queue_map (port->n_txq,
txq_idx, queue_id, enabled) accordingly so no modulo by zero occurs.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (6)
modules/infra/control/gr_port.hmodules/infra/control/graph.cmodules/infra/control/port.cmodules/infra/control/worker.cmodules/infra/datapath/gr_rxtx.hmodules/infra/datapath/port_tx.c
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{c,h}
⚙️ CodeRabbit configuration file
**/*.{c,h}: -gr_vec_*()functions cannot fail. No need to check their return value.
gr_vec_free(x)always setsx = NULL. There is no risk of double free.ec_node_*()functions consume theirec_nodearguments. No leaks on error.rte_node->ctxis an uint8_t array of size 16, not a pointer.- Never suggest to replace
assert()with graceful error checking.- We compile with
-std=gnu2x. Unnamed parameters in function definitions are valid.
Files:
modules/infra/control/worker.cmodules/infra/datapath/port_tx.cmodules/infra/control/port.cmodules/infra/control/gr_port.hmodules/infra/datapath/gr_rxtx.hmodules/infra/control/graph.c
🧠 Learnings (5)
📚 Learning: 2025-11-05T14:28:28.817Z
Learnt from: rjarry
Repo: DPDK/grout PR: 372
File: modules/infra/datapath/loop_input.c:53-56
Timestamp: 2025-11-05T14:28:28.817Z
Learning: In modules/infra/datapath/loop_input.c, the l3_edges array is intentionally indexed using network byte order (rte_be16_t) for EtherType values. Both loopback_input_add_type() registration and loopback_input_process() lookup use network byte order consistently, so no byte order conversion is needed when accessing l3_edges.
Applied to files:
modules/infra/datapath/port_tx.cmodules/infra/control/gr_port.h
📚 Learning: 2025-11-03T13:28:19.489Z
Learnt from: rjarry
Repo: DPDK/grout PR: 379
File: modules/infra/datapath/port_tx.c:36-46
Timestamp: 2025-11-03T13:28:19.489Z
Learning: In DPDK graph node process callbacks, the return value is only used for statistics and does not affect packet flow or scheduling through the graph. Nodes can return 0 when they haven't processed packets (e.g., when dropping or redirecting due to port state).
Applied to files:
modules/infra/datapath/port_tx.cmodules/infra/control/graph.c
📚 Learning: 2025-12-17T17:32:21.746Z
Learnt from: christophefontaine
Repo: DPDK/grout PR: 466
File: modules/srv6/datapath/l2_encap.c:26-32
Timestamp: 2025-12-17T17:32:21.746Z
Learning: The GR_MBUF_PRIV_DATA_TYPE macro in modules/infra/datapath/gr_mbuf.h automatically adds `const struct iface *iface` as the first field to any structure defined with it. All types defined using this macro (e.g., srv6_dx2_mbuf_data, mbuf_data, queue_mbuf_data) will have the iface field available, followed by any custom fields provided as the second macro argument.
Applied to files:
modules/infra/datapath/port_tx.c
📚 Learning: 2025-09-05T08:58:53.047Z
Learnt from: rjarry
Repo: DPDK/grout PR: 294
File: modules/policy/control/conntrack.c:533-539
Timestamp: 2025-09-05T08:58:53.047Z
Learning: DPDK functions rte_mempool_free() and rte_hash_free() are NULL-safe - they explicitly handle NULL pointers and do nothing when passed NULL, unlike many C standard library functions.
Applied to files:
modules/infra/control/gr_port.h
📚 Learning: 2025-09-09T12:08:36.790Z
Learnt from: rjarry
Repo: DPDK/grout PR: 294
File: modules/policy/control/conntrack.c:621-637
Timestamp: 2025-09-09T12:08:36.790Z
Learning: DPDK's rte_hash supports safe deletion during iteration via rte_hash_iterate, unlike many other hash table implementations that require two-phase deletion patterns.
Applied to files:
modules/infra/control/gr_port.h
🧬 Code graph analysis (2)
modules/infra/control/port.c (1)
api/gr_errno.h (1)
errno_set(10-13)
modules/infra/control/graph.c (2)
modules/infra/control/worker_test.c (1)
q(150-157)main/gr_vec.h (1)
gr_vec_len(47-51)
🪛 Cppcheck (2.19.0)
modules/infra/control/worker.c
[error] 231-231: Expression 'worker->txqs?free(__gr_vec_hdr(worker->txqs))
(unknownEvaluationOrder)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (5)
- GitHub Check: build-and-tests (gcc-13, none, debugoptimized, ubuntu-24.04, -Dfrr=enabled, false)
- GitHub Check: build-and-tests (gcc-14, address, debug, ubuntu-24.04, -Dfrr=enabled, true)
- GitHub Check: build-and-tests (clang-15, none, debugoptimized, ubuntu-22.04, false)
- GitHub Check: deb
- GitHub Check: rpm
🔇 Additional comments (1)
modules/infra/control/port.c (1)
94-126: No fix needed. DPDK's ethdev layer guarantees thatinfo.max_tx_queueswill not exceedRTE_MAX_QUEUES_PER_PORT, sop->n_txqcannot exceed the array bounds oftxq_locks[RTE_MAX_QUEUES_PER_PORT].
modules/infra/datapath/port_tx.c
Outdated
| } | ||
|
|
||
| tx_ok = rte_eth_tx_burst(ctx->port_id, ctx->queue_id, mbufs, nb_objs); | ||
| if (ctx->lock != NULL) |
There was a problem hiding this comment.
it's not possible to have a dedicated function tx_process_locked to avoid the if (ctx->lock != NULL).
TX_NODE_BASE_LOCKED et TX_NODE_BASE.
It will avoid two branch in the hot path.
| // force number of TX queues equal to requested RX queues | ||
| p->n_txq = p->n_rxq; | ||
| if (p->n_txq > info.max_tx_queues) | ||
| return errno_set(EOVERFLOW); |
There was a problem hiding this comment.
Why not a tx queue with N virtio ring ?
| } | ||
|
|
||
| if (p->n_rxq > info.max_rx_queues) | ||
| return errno_set(EOVERFLOW); |
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
modules/infra/control/graph.c (1)
196-213: TX queue user count may include disabled queues.The loop counting
txq_usersdoesn't checkq->enabled, so disabled queue mappings are counted. This won't cause data races (overcounting is safe—just adds an unnecessary lock), but the warning log may be misleading.If this is intentional (defensive locking), no action needed. Otherwise:
🔧 Optional: filter by enabled queues
gr_vec_foreach_ref (struct queue_map *q, w->txqs) { - if (q->port_id == qmap->port_id && q->queue_id == qmap->queue_id) + if (q->enabled && q->port_id == qmap->port_id && q->queue_id == qmap->queue_id) txq_users++; }
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (9)
modules/infra/control/gr_port.hmodules/infra/control/graph.cmodules/infra/control/port.cmodules/infra/control/worker.cmodules/infra/control/worker_test.cmodules/infra/datapath/gr_rxtx.hmodules/infra/datapath/port_tx.csubprojects/dpdk.wrapsubprojects/packagefiles/dpdk/iavf-fix-reported-max-TX-and-RX-queues-in-ethdev-inf.patch
✅ Files skipped from review due to trivial changes (1)
- subprojects/dpdk.wrap
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{c,h}
⚙️ CodeRabbit configuration file
**/*.{c,h}: -gr_vec_*()functions cannot fail. No need to check their return value.
gr_vec_free(x)always setsx = NULL. There is no risk of double free.ec_node_*()functions consume theirec_nodearguments. No leaks on error.rte_node->ctxis an uint8_t array of size 16, not a pointer.- Never suggest to replace
assert()with graceful error checking.- We compile with
-std=gnu2x. Unnamed parameters in function definitions are valid.
Files:
modules/infra/control/gr_port.hmodules/infra/datapath/port_tx.cmodules/infra/control/port.cmodules/infra/control/graph.cmodules/infra/control/worker_test.cmodules/infra/control/worker.cmodules/infra/datapath/gr_rxtx.h
🧠 Learnings (5)
📚 Learning: 2025-09-05T08:58:53.047Z
Learnt from: rjarry
Repo: DPDK/grout PR: 294
File: modules/policy/control/conntrack.c:533-539
Timestamp: 2025-09-05T08:58:53.047Z
Learning: DPDK functions rte_mempool_free() and rte_hash_free() are NULL-safe - they explicitly handle NULL pointers and do nothing when passed NULL, unlike many C standard library functions.
Applied to files:
modules/infra/control/gr_port.h
📚 Learning: 2025-11-05T14:28:28.817Z
Learnt from: rjarry
Repo: DPDK/grout PR: 372
File: modules/infra/datapath/loop_input.c:53-56
Timestamp: 2025-11-05T14:28:28.817Z
Learning: In modules/infra/datapath/loop_input.c, the l3_edges array is intentionally indexed using network byte order (rte_be16_t) for EtherType values. Both loopback_input_add_type() registration and loopback_input_process() lookup use network byte order consistently, so no byte order conversion is needed when accessing l3_edges.
Applied to files:
modules/infra/control/gr_port.hmodules/infra/datapath/port_tx.c
📚 Learning: 2025-09-09T12:08:36.790Z
Learnt from: rjarry
Repo: DPDK/grout PR: 294
File: modules/policy/control/conntrack.c:621-637
Timestamp: 2025-09-09T12:08:36.790Z
Learning: DPDK's rte_hash supports safe deletion during iteration via rte_hash_iterate, unlike many other hash table implementations that require two-phase deletion patterns.
Applied to files:
modules/infra/control/gr_port.h
📚 Learning: 2025-11-03T13:28:19.489Z
Learnt from: rjarry
Repo: DPDK/grout PR: 379
File: modules/infra/datapath/port_tx.c:36-46
Timestamp: 2025-11-03T13:28:19.489Z
Learning: In DPDK graph node process callbacks, the return value is only used for statistics and does not affect packet flow or scheduling through the graph. Nodes can return 0 when they haven't processed packets (e.g., when dropping or redirecting due to port state).
Applied to files:
modules/infra/datapath/port_tx.cmodules/infra/control/graph.c
📚 Learning: 2025-12-17T17:32:21.746Z
Learnt from: christophefontaine
Repo: DPDK/grout PR: 466
File: modules/srv6/datapath/l2_encap.c:26-32
Timestamp: 2025-12-17T17:32:21.746Z
Learning: The GR_MBUF_PRIV_DATA_TYPE macro in modules/infra/datapath/gr_mbuf.h automatically adds `const struct iface *iface` as the first field to any structure defined with it. All types defined using this macro (e.g., srv6_dx2_mbuf_data, mbuf_data, queue_mbuf_data) will have the iface field available, followed by any custom fields provided as the second macro argument.
Applied to files:
modules/infra/datapath/port_tx.c
🧬 Code graph analysis (2)
modules/infra/control/port.c (1)
api/gr_errno.h (1)
errno_set(10-13)
modules/infra/control/worker_test.c (1)
modules/infra/control/worker.c (2)
worker_rxq_assign(251-312)worker_count(125-133)
🪛 Cppcheck (2.19.0)
modules/infra/control/worker.c
[error] 231-231: Expression 'worker->txqs?free(__gr_vec_hdr(worker->txqs))
(unknownEvaluationOrder)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (7)
- GitHub Check: build-and-tests (clang-16, none, debugoptimized, ubuntu-24.04, false)
- GitHub Check: build-and-tests (clang-18, none, debugoptimized, ubuntu-24.04, false)
- GitHub Check: build-and-tests (gcc-13, none, debugoptimized, ubuntu-24.04, -Dfrr=enabled, false)
- GitHub Check: build-and-tests (clang-15, none, debugoptimized, ubuntu-22.04, false)
- GitHub Check: build-and-tests (gcc-14, address, debug, ubuntu-24.04, -Dfrr=enabled, true)
- GitHub Check: deb
- GitHub Check: rpm
🔇 Additional comments (10)
modules/infra/datapath/port_tx.c (1)
57-63: LGTM! Conditional locking for shared TX queues is correctly implemented.The lock-only-when-shared pattern avoids spinlock overhead for exclusively-owned queues while properly protecting concurrent access when multiple workers share a TX queue.
subprojects/packagefiles/dpdk/iavf-fix-reported-max-TX-and-RX-queues-in-ethdev-inf.patch (1)
38-44: LGTM! Correctly caps max queue reporting based on VF capability.The fix properly checks
VIRTCHNL_VF_LARGE_NUM_QPAIRScapability before reporting 256 queues, falling back to 16 when the PF doesn't support large VF.modules/infra/control/gr_port.h (1)
36-36: LGTM! Per-TX-queue spinlock array for shared queue synchronization.The array size matches DPDK's
RTE_MAX_QUEUES_PER_PORT, ensuring coverage for all possible queue configurations.modules/infra/control/port.c (1)
94-118: LGTM! Device-aware TX queue configuration with proper driver quirks handling.The implementation correctly:
- Caps TX queues to device maximum
- Handles net_tap/net_virtio requirement for equal RX/TX counts
- Logs a notice when queue count is reduced
- Initializes per-queue spinlocks for the sharing mechanism
modules/infra/datapath/gr_rxtx.h (1)
21-35: LGTM! Clean separation of TX node context with optional locking.The
tx_node_ctxstructure properly encapsulates the TX queue info and lock pointer, enabling conditional locking only for shared queues (whenlock != NULL).modules/infra/control/worker_test.c (1)
408-412: LGTM! Test correctly validates TX queue sharing via modulo distribution.With 5 workers and 4 TX queues per port, the modulo assignment causes workers to share TXQ 0 (w3 and w4), which is the expected behavior protected by the spinlocks in port_tx.c.
modules/infra/control/worker.c (1)
225-249: LGTM! Clean TX queue distribution with proper sharing semantics.The implementation correctly:
- Clears all TXQ assignments before redistribution
- Only assigns TXQs to workers that have RXQs (workers without RXQs won't transmit)
- Uses modulo assignment to distribute workers across available TXQs
The static analysis warning on line 231 is a false positive - per coding guidelines,
gr_vec_free()is safe and sets the pointer to NULL.modules/infra/control/graph.c (3)
77-87: LGTM!Helper function cleanly encapsulates port lookup logic. The
assertis appropriate here since callers should only pass valid port IDs from the existing queue mappings.
161-168: LGTM!Proper initialization of TX node contexts to invalid values prevents accidental transmission on unassigned queues.
215-223: LGTM!Edge wiring correctly maps
port_idto the corresponding TX node edge. The index is bounded by DPDK's port limits.
This fix has been submitted upstream. Link: https://patches.dpdk.org/project/dpdk/patch/20260109142129.1550058-2-rjarry@redhat.com/ Signed-off-by: Robin Jarry <rjarry@redhat.com>
Replace worker_txq_id() with worker_txq_distribute() which clears all TX queue assignments and reassigns them only to workers that have at least one RX queue. Workers without RX queues do not process packets and therefore have no need for TX queues. Call worker_txq_distribute() from both worker_queue_distribute() during initial setup and worker_rxq_assign() when RX queues are moved at runtime. Since worker_rxq_assign() now calls worker_graph_reload_all() instead of reloading individual workers, remove the now-redundant worker_graph_reload() call from worker_queue_distribute(). To prevent two workers from transmitting on the same TX queue during graph reloads, modify worker_graph_reload_all() to first stop all workers by reloading them with NULL ports, then reload with the actual configuration. Adjust worker_graph_new() to treat NULL ports as a request to create a NULL graph. Update unit tests to expect empty TX queue assignments for workers without RX queues. Signed-off-by: Robin Jarry <rjarry@redhat.com>
Replace direct use of port_queue as the TX node context with a new tx_node_ctx structure that wraps port_queue as a member. This allows extending the TX node context with additional fields in future commits. Convert port_queue from a GR_NODE_CTX_TYPE macro to a plain struct since it is no longer used directly as a node context. It remains used for packet tracing where only port_id and queue_id are recorded. Signed-off-by: Robin Jarry <rjarry@redhat.com>
Extract the port lookup loop into a dedicated find_port() function. This will be reused in the next commit when configuring TX node context. Signed-off-by: Robin Jarry <rjarry@redhat.com>
Some network devices report fewer TX queues than the number of datapath workers. Cap the TX queue count to the device maximum and distribute workers across available queues using modulo assignment. When multiple workers share the same TX queue, protect rte_eth_tx_burst() with a per-queue spinlock stored in iface_info_port. The lock pointer in tx_node_ctx is only set for shared queues to avoid locking overhead when a worker has exclusive access. Update unit tests to set max_tx_queues in the mock device info and adjust expected TX queue assignments for modulo wrapping. Signed-off-by: Robin Jarry <rjarry@redhat.com>
Unlike TX queues which are silently capped to the device maximum, RX queues are explicitly requested by the user. Return EOVERFLOW if the requested count exceeds what the device supports. Signed-off-by: Robin Jarry <rjarry@redhat.com>
The net_tap and net_virtio drivers require equal RX and TX queue counts. Set TX queue count to match RX queues for these drivers. This may result in fewer TX queues than workers, which is now handled by the TX queue sharing support. Signed-off-by: Robin Jarry <rjarry@redhat.com>
2 participants
Some devices report fewer TX queues than the number of datapath workers. Cap the TX queue count to the device maximum and distribute workers across available queues using modulo assignment.
When multiple workers are assigned to the same TX queue, protect rte_eth_tx_burst() with a per-queue spinlock. The lock pointer in tx_node_ctx is only set for queues that are actually shared, avoiding locking overhead when each worker has exclusive access to its queue.
Only assign TX queues to workers that have at least one RX queue since workers without RX queues will never transmit packets.
Add a workaround for the iavf driver which reports 256 max queues but fails to configure more than 16 when the PF does not support large VF.
Summary by CodeRabbit
New Features
Refactor
Tests
Chores
✏️ Tip: You can customize this high-level summary in your review settings.