AMD General Hi All,
Quick update, implementing CPU_ALLOC instead of rte_cpu_set is turing out to eat a lot of memory. I am working on logic to use CPU_ALLOC when CPU_SIZE > rte_cpu_Set capacity or find smarter ways not to pre-alloc extra. Note: for accommodating CPU alloc we need to do it for all valid l1, l2, l3 and l4. On 384 single socket machine it exhausts memory (heap) like anything. > -----Original Message----- > From: Varghese, Vipin <[email protected]> > Sent: Thursday, May 28, 2026 12:48 PM > To: Stephen Hemminger <[email protected]> > Cc: [email protected]; Tummala, Sivaprasad <[email protected]>; > [email protected]; [email protected]; > [email protected]; [email protected]; [email protected] > Subject: RE: [PATCH v5 0/3] eal/topology: introduce topology-aware lcore > grouping > > Caution: This message originated from an External Source. Use proper caution > when opening attachments, clicking links, or responding. > > > AMD General > > Sharing v6 with fixes and cache-id fetch shortly. > > > -----Original Message----- > > From: Stephen Hemminger <[email protected]> > > Sent: Wednesday, April 15, 2026 1:52 AM > > To: Varghese, Vipin <[email protected]> > > Cc: [email protected]; Tummala, Sivaprasad <[email protected]>; > > [email protected]; [email protected]; > > [email protected]; [email protected]; > > [email protected] > > Subject: Re: [PATCH v5 0/3] eal/topology: introduce topology-aware > > lcore grouping > > > > Caution: This message originated from an External Source. Use proper > > caution when opening attachments, clicking links, or responding. > > > > > > On Wed, 15 Apr 2026 01:08:18 +0530 > > Vipin Varghese <[email protected]> wrote: > > > > > This series introduces a topology library that groups DPDK lcores > > > based on CPU cache hierarchy and NUMA topology. The goal is to > > > provide a stable and explicit API that allows applications to select > > > lcores with better locality and cache sharing characteristics. > > > > > > The series includes: > > > - EAL support for topology discovery using hwloc and domain-based lcore > > > grouping (L1/L2/L3/L4/NUMA) > > > - Topology-aware test cases validating API behavior and edge conditions > > > - Programmer’s guide describing the topology library and APIs > > > > > > The API is marked experimental and does not change existing lcore > > > behavior unless explicitly used by the application. > > > > > > Changes in v5: > > > - Addressed review comments from v4 > > > - Fixed ARM cross-compilation issues > > > - Cleaned up domain iteration and error handling > > > - Updated tests to cover domain edge cases > > > - Documentation refinements and API usage clarification > > > > > > Changes in v4: > > > - Corrected domain selection semantics > > > - Updated example usage > > > - Fixed naming and typo issues > > > > > > Changes in v3: > > > - Fixed macro naming (USE_NO_TOPOLOGY) > > > - Minor cleanups based on early feedback > > > > > > Tested on: > > > - AMD EPYC (Milan, Genoa, Siena, Turin, Turin-Dense, Sorano) > > > - Intel Xeon (SPR-SP, GNR-SP) > > > - ARM Ampere > > > - NVIDIA Grace Superchip > > > > > > Dependencies: > > > - hwloc-dev (tested with 2.10.0) > > > > > > Patch breakdown: > > > 1/3 eal/topology: add topology grouping for lcores > > > 2/3 app: add topology-aware test cases > > > 3/3 doc: add topology library documentation > > > > > > Future Work: > > > - integrate into examples > > > -- hellowrld: ready > > > -- pkt-distributor: in-progress > > > -- l2fwd: ready > > > -- l3fwd: to start > > > -- eventdevpipeline: PoC ready > > > - integrate topology test > > > -- crypto: yet to start > > > -- compression: yet to start > > > -- dma: PoC ready > > > - add new features for > > > -- PQoS: yet to start > > > -- Data Injection: PoC with BRDCM Thor-2 ready > > > > > > Tested OS: Linux only, need help with BSD and Windows > > > > > > Tested with and without hwloc-dev library for > > > - Ampere, aarch64, Neoverse-N1, NUMA-2, 256 CPU threads > > > - Grace superchip, aarch64, Neoverse-V2, NUMA-2, 144 CPU threads > > > - Intel GNR-SP, 6767P, NUMA-2, 256 Threads > > > - AMD EPYC Siena, 8534P, NUMA-1, 128 Threads > > > - AMD EPYC Sorano, 8635P, NUMA-1, 168 Threads > > > > > > Signed-off-by: Vipin Varghese <[email protected]> `` > > > > > > Vipin Varghese (3): > > > eal/topology: add Topology grouping for lcores > > > app: add topology aware test case > > > doc: add new section topology > > > > > > app/test/meson.build | 1 + > > > app/test/test_ring_perf.c | 416 +++++++++++++- > > > app/test/test_stack_perf.c | 409 ++++++++++++++ > > > app/test/test_topology.c | 676 ++++++++++++++++++++++ > > > config/meson.build | 18 + > > > doc/api/doxy-api-index.md | 1 + > > > doc/guides/prog_guide/index.rst | 3 +- > > > doc/guides/prog_guide/topology_lib.rst | 155 +++++ > > > lib/eal/common/eal_private.h | 74 +++ > > > lib/eal/common/eal_topology.c | 747 +++++++++++++++++++++++++ > > > lib/eal/common/meson.build | 1 + > > > lib/eal/freebsd/eal.c | 10 +- > > > lib/eal/include/meson.build | 1 + > > > lib/eal/include/rte_topology.h | 255 +++++++++ > > > lib/eal/linux/eal.c | 7 + > > > lib/eal/meson.build | 4 + > > > 16 files changed, 2773 insertions(+), 5 deletions(-) create mode > > > 100644 app/test/test_topology.c create mode 100644 > > > doc/guides/prog_guide/topology_lib.rst > > > create mode 100644 lib/eal/common/eal_topology.c create mode > > > 100644 lib/eal/include/rte_topology.h > > > > > > > AI review reported lots of possible correctness problems: > > > > > > # DPDK Patch Review - bundle-1843-topology-v5.mbox > > > > ## Overview > > > > This patch series introduces topology awareness to DPDK, allowing > > applications to select logical cores based on CPU cache and I/O > > topology. The series adds a new `rte_topology` library with APIs to > > query and enumerate logical cores within topology domains > > (L1/L2/L3/L4/NUMA). > > > > --- > > > > ## Correctness Bugs (HIGHEST PRIORITY - report at >=50% confidence) > > > > ### Patch 1/3: lib/eal/common/eal_topology.c > > > > #### **Error: Resource leak on `eal_topology_map_layer()` failure** > > > > In `rte_eal_topology_init()`, if any call to > > `eal_topology_map_layer()` fails (returns < 0), the code calls > > `rte_eal_topology_release()` which frees previously allocated layers. > > However, the `hwloc_topology_t` handle `topo_cnfg.topology` is **NOT** > destroyed before returning. This leaks the hwloc topology handle. > > > > **Location:** `lib/eal/common/eal_topology.c:685-712` > > > > ```c > > for (int i = 0; i < 5; i++) { > > *layers[i].count = hwloc_get_nbobjs_by_depth(topo_cnfg.topology, > > layers[i].depth); > > if (eal_topology_map_layer(topo_cnfg.topology, layers[i].depth, > > layers[i].count, > > layers[i].ptr, layers[i].total_cores, layers[i].name) < 0) { > > rte_eal_topology_release(); /* frees layer memory */ > > return -1; /* BUG: topo_cnfg.topology NOT destroyed here */ > > } > > } > > > > hwloc_topology_destroy(topo_cnfg.topology); /* only reached on > > success */ topo_cnfg.topology = NULL; ``` > > > > **Fix:** Destroy the topology before returning on error: > > > > ```c > > if (eal_topology_map_layer(...) < 0) { > > hwloc_topology_destroy(topo_cnfg.topology); > > topo_cnfg.topology = NULL; > > rte_eal_topology_release(); > > return -1; > > } > > ``` > > > > --- > > > > #### **Error: Potential use-after-free in `eal_topology_map_layer()` > > on partial allocation failure** > > > > In `eal_topology_map_layer()`, the code allocates `dm->cores` for each > > domain. If a later allocation in the same loop iteration fails (e.g., > > for domain `j+1`), the function returns -1 immediately **without > > freeing `dm->cores` already allocated for earlier domains**. The > > caller (`rte_eal_topology_init()`) then calls > > `rte_eal_topology_release()`, which expects `layer_ptr[j]` to be > > non-NULL but `layer_ptr[j]->cores` may be uninitialized or garbage if > > the allocation failed before that point. This can cause a use-after-free or > > double- > free when `rte_eal_topology_release()` calls `rte_free(d->map[i]->cores)`. > > > > **Location:** `lib/eal/common/eal_topology.c:537-557` > > > > ```c > > for (uint16_t j = 0; j < *layer_cnt; j++) { > > hwloc_obj_t obj = hwloc_get_obj_by_depth(topology, depth, j); > > int cpu_count = hwloc_bitmap_weight(obj->cpuset); > > if (cpu_count == -1) > > continue; > > > > struct core_domain_mapping *dm = > > rte_zmalloc(NULL, sizeof(struct core_domain_mapping), 0); > > if (!dm) > > return -1; /* BUG: leaks layer_ptr array allocated by caller > > */ > > > > (*layer_ptr)[j] = dm; > > CPU_ZERO(&dm->core_set); > > dm->core_count = 0; > > > > dm->cores = rte_malloc(NULL, sizeof(uint16_t) * cpu_count, 0); > > if (!dm->cores) > > return -1; /* BUG: leaks dm (just allocated) and previous > > entries */ } ``` > > > > **Fix:** On allocation failure, free all previously allocated entries > > before returning: > > > > ```c > > if (!dm) { > > /* Free all previously allocated entries */ > > for (uint16_t k = 0; k < j; k++) { > > if ((*layer_ptr)[k]) { > > rte_free((*layer_ptr)[k]->cores); > > rte_free((*layer_ptr)[k]); > > } > > } > > rte_free(*layer_ptr); > > return -1; > > } > > > > /* Same cleanup for dm->cores allocation failure */ if (!dm->cores) { > > rte_free(dm); > > for (uint16_t k = 0; k < j; k++) { > > if ((*layer_ptr)[k]) { > > rte_free((*layer_ptr)[k]->cores); > > rte_free((*layer_ptr)[k]); > > } > > } > > rte_free(*layer_ptr); > > return -1; > > } > > ``` > > > > --- > > > > #### **Error: NULL pointer dereference in `eal_topology_map_layer()` > > when `hwloc_get_obj_by_depth()` returns NULL** > > > > In the second loop of `eal_topology_map_layer()` (lines 560-620), the > > code calls `hwloc_get_obj_by_depth(topology, depth, j)` and checks `if > > (!obj || hwloc_bitmap_iszero(obj->cpuset))` to skip NULL objects. > > However, in the **inner > > loop** (lines 579-618), the code calls > > `hwloc_get_obj_by_depth(topology, depth, k)` and assigns it to > > `obj_core` but does **NOT** check if `obj_core` is NULL before calling > > `hwloc_bitmap_weight(obj_core->cpuset)` on line 581. If > > `hwloc_get_obj_by_depth()` returns NULL for domain `k`, this will > > dereference a > NULL pointer. > > > > **Location:** `lib/eal/common/eal_topology.c:579-582` > > > > ```c > > for (uint16_t k = 0; k < *layer_cnt; k++) { > > hwloc_obj_t obj_core = hwloc_get_obj_by_depth(topology, depth, k); > > int cpu_count_core = hwloc_bitmap_weight(obj_core->cpuset); /* > > NULL deref if obj_core == NULL */ > > if (cpu_count_core == -1) > > continue; > > ``` > > > > **Fix:** Check `obj_core` before dereferencing: > > > > ```c > > hwloc_obj_t obj_core = hwloc_get_obj_by_depth(topology, depth, k); if > > (!obj_core) > > continue; > > int cpu_count_core = hwloc_bitmap_weight(obj_core->cpuset); > > if (cpu_count_core == -1) > > continue; > > ``` > > > > --- > > > > #### **Error: Incorrect second argument to > > `rte_topo_get_nth_lcore_from_domain()` in `get_same_l1_domains()` > > (test_ring_perf.c and test_stack_perf.c)** > > > > In both `app/test/test_ring_perf.c:290` and > > `app/test/test_stack_perf.c:258`, the function `get_same_l1_domains()` > > calls: > > > > ```c > > id2 = rte_topo_get_nth_lcore_from_domain(domain, 0, 0, > > RTE_TOPO_DOMAIN_L1); ``` > > > > The second argument (`lcore_pos`) is `0`, which is the same as for > > `id1`. This will assign **the same lcore** to both `id1` and `id2`, > > causing the subsequent check `if > > (id1 == id2) return 3;` to always trigger. This is a logic error: the > > intent is clearly to get two **different** lcores from the same domain. > > > > **Location:** `app/test/test_ring_perf.c:287-290` and > > `app/test/test_stack_perf.c:255-258` > > > > **Fix:** Use position `1` for the second lcore: > > > > ```c > > id1 = rte_topo_get_nth_lcore_from_domain(domain, 0, 0, > > RTE_TOPO_DOMAIN_L1); > > id2 = rte_topo_get_nth_lcore_from_domain(domain, 1, 0, > > RTE_TOPO_DOMAIN_L1); ``` > > > > --- > > > > #### **Error: Iteration condition in `test_main_lcore_in_domain()` > > uses wrong domain type for lookup** > > > > In `app/test/test_topology.c:211`, the loop iterates over > > `domain_count` for `domain_types[d]`, but the call to > > `rte_topo_is_main_lcore_in_domain()` uses `RTE_TOPO_DOMAIN_NUMA` > > instead of `domain_types[d]`. This means the test only checks the NUMA > > domain regardless of which domain type `d` selects (L1/L2/L3/L4). > > > > **Location:** `app/test/test_topology.c:206-216` > > > > ```c > > for (unsigned int d = 0; d < RTE_DIM(domain_types); d++) { > > bool main_lcore_found = false; > > unsigned int domain_count = rte_topo_get_domain_count(domain_types[d]); > > for (unsigned int dmn_idx = 0; dmn_idx < domain_count; dmn_idx++) { > > main_lcore_found = > > rte_topo_is_main_lcore_in_domain(RTE_TOPO_DOMAIN_NUMA, /* BUG: > should > > be domain_types[d] */ > > dmn_idx); > > if (main_lcore_found) > > break; > > } > > ``` > > > > **Fix:** > > > > ```c > > main_lcore_found = rte_topo_is_main_lcore_in_domain(domain_types[d], > > dmn_idx); ``` > > > > --- > > > > #### **Error: Infinite loop risk in > > `rte_topo_get_nth_lcore_from_domain()` when `ptr- > > >core_count` is 0** > > > > In `lib/eal/common/eal_topology.c:296-318`, the function enters a > > `while (1)` loop that increments `new_lcore_pos`. If `ptr->core_count` > > is 0 (which the code checks earlier but does not return immediately), > > the loop will wrap `new_lcore_pos` back to > > 0 indefinitely, never breaking. While the function returns > > `RTE_MAX_LCORE` if `ptr- > > >core_count == 0` before the loop, the logic flow is unclear and the > > >loop body does > > not have a clear termination condition if the core count is 0. > > > > **Location:** `lib/eal/common/eal_topology.c:283-318` > > > > **Fix:** Add a sanity check inside the loop to prevent infinite iteration: > > > > ```c > > unsigned int iterations = 0; > > while (1) { > > if (iterations++ > ptr->core_count * 2) /* safety limit */ > > return RTE_MAX_LCORE; > > /* ... rest of loop ... */ > > } > > ``` > > > > However, the real issue is that the code already returns > > `RTE_MAX_LCORE` if `ptr->core_count == 0` on line 287, so this is more > > of a defensive-programming note. The function should be refactored for > > clarity. > > > > --- > > > > #### **Error: Missing NULL check after `get_domain_lcore_mapping()` in > > `rte_topo_get_next_lcore()`** > > > > In `rte_topo_get_next_lcore()`, the code calls > > `get_domain_lcore_mapping(flag, lcore_domain)` and checks if `ptr` is > > NULL on line 350. However, if `ptr` is NULL, the function returns > > `RTE_MAX_LCORE`. This is correct, but the subsequent logic on line 381 > > calls `rte_topo_is_main_lcore_in_domain(flag, lcore_domain)`, which > > internally may call `get_domain_lcore_mapping()` again. If that call > > also returns NULL (which it will if the domain is invalid), the > > function `rte_topo_is_main_lcore_in_domain()` will return `false`, > > which is safe. However, the logic is fragile and should explicitly handle > > the NULL > case to avoid relying on transitive safety. > > > > **Location:** `lib/eal/common/eal_topology.c:381` > > > > **Recommendation:** The code is technically safe but could be clearer. > > No change required, but consider restructuring for maintainability. > > > > --- > > > > ### Patch 2/3: app/test Topology Tests > > > > #### **Error: Macro > `RTE_TOPO_FOREACH_WORKER_LCORE_IN_DOMAIN` > > declares variable in macro expansion (shadowing risk)** > > > > In `lib/eal/include/rte_topology.h:243-248`, the macro > > `RTE_TOPO_FOREACH_WORKER_LCORE_IN_DOMAIN` declares a local > variable > > `main_lcore` inside the macro expansion: > > > > ```c > > #define RTE_TOPO_FOREACH_WORKER_LCORE_IN_DOMAIN(lcore, > > domain_indx, flag) \ > > lcore = rte_topo_get_nth_lcore_from_domain(domain, 0, 0, flag); \ > > uint16_t main_lcore = rte_get_main_lcore(); \ > > for (lcore = (lcore != main_lcore) ? \ > > lcore : rte_topo_get_next_lcore(lcore, 1, 0, flag); \ > > lcore < RTE_MAX_LCORE; \ > > lcore = rte_topo_get_next_lcore(lcore, 1, 0, flag)) ``` > > > > This can cause a compiler error or shadowing if the caller already has > > a variable named `main_lcore` in scope. Additionally, the macro uses > > `domain` (line 244) but the parameter is `domain_indx`, which is a typo and > > will > cause a compilation error. > > > > **Location:** `lib/eal/include/rte_topology.h:243-248` > > > > **Fix:** Wrap in a `do { } while (0)` and use a uniquely-named > > variable, or document that the macro must not be used if `main_lcore` > > is already declared. Also fix the > > typo: > > > > ```c > > #define RTE_TOPO_FOREACH_WORKER_LCORE_IN_DOMAIN(lcore, > > domain_indx, flag) \ > > lcore = rte_topo_get_nth_lcore_from_domain(domain_indx, 0, 0, flag); \ > > uint16_t __topo_main_lcore = rte_get_main_lcore(); \ > > for (lcore = (lcore != __topo_main_lcore) ? \ > > lcore : rte_topo_get_next_lcore(lcore, 1, 0, flag); \ > > lcore < RTE_MAX_LCORE; \ > > lcore = rte_topo_get_next_lcore(lcore, 1, 0, flag)) ``` > > > > --- > > > > ### Summary of Correctness Bugs > > > > 1. **hwloc topology leak** on `eal_topology_map_layer()` failure 2. > > **Resource > > leak** in `eal_topology_map_layer()` on partial allocation failure 3. > > **NULL pointer > > dereference** in `eal_topology_map_layer()` inner loop 4. **Logic > > error** in `get_same_l1_domains()` (same lcore assigned to `id1` and > > `id2`) 5. **Wrong domain type** in `test_main_lcore_in_domain()` (uses > > `RTE_TOPO_DOMAIN_NUMA` instead of `domain_types[d]`) 6. **Macro typo** > > in `RTE_TOPO_FOREACH_WORKER_LCORE_IN_DOMAIN` (uses `domain` > instead of > > `domain_indx`) 7. **Potential infinite loop** in > > `rte_topo_get_nth_lcore_from_domain()` if `ptr->core_count == 0` > > (mitigated by early
