From 47f4f695cec1eb82d961fdb466cd7c99d2865b7c Mon Sep 17 00:00:00 2001 From: Alexei Starovoitov Date: Wed, 30 Sep 2026 09:59:18 +0000 Subject: [PATCH 1/3] bpf: Factor out __do_call_rcu_ttrace() Move the part of do_call_rcu_ttrace() that runs after call_rcu_ttrace_in_progress is set into __do_call_rcu_ttrace(). The next patch will call it from __free_rcu(). No functional change. Signed-off-by: Alexei Starovoitov Link: https://lore.kernel.org/bpf/20260930095920.601738-2-alexei.starovoitov@gmail.com Signed-off-by: Kumar Kartikeya Dwivedi --- kernel/bpf/memalloc.c | 27 +++++++++++++++++---------- 1 file changed, 17 insertions(+), 10 deletions(-) diff --git a/kernel/bpf/memalloc.c b/kernel/bpf/memalloc.c index 8a8f088e83e6..08e4dde66cd5 100644 --- a/kernel/bpf/memalloc.c +++ b/kernel/bpf/memalloc.c @@ -298,19 +298,10 @@ static void enque_to_free(struct bpf_mem_cache *c, void *obj) llist_add(llnode, &c->free_by_rcu_ttrace); } -static void do_call_rcu_ttrace(struct bpf_mem_cache *c) +static void __do_call_rcu_ttrace(struct bpf_mem_cache *c) { struct llist_node *llnode, *t; - if (atomic_xchg(&c->call_rcu_ttrace_in_progress, 1)) { - if (unlikely(READ_ONCE(c->draining))) { - scoped_guard(raw_spinlock_irqsave, &c->lock) - llnode = llist_del_all(&c->free_by_rcu_ttrace); - free_all(c, llnode, !!c->percpu_size); - } - return; - } - WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace)); llist_for_each_safe(llnode, t, llist_del_all(&c->free_by_rcu_ttrace)) llist_add(llnode, &c->waiting_for_gp_ttrace); @@ -328,6 +319,22 @@ static void do_call_rcu_ttrace(struct bpf_mem_cache *c) call_rcu_tasks_trace(&c->rcu_ttrace, __free_rcu); } +static void do_call_rcu_ttrace(struct bpf_mem_cache *c) +{ + struct llist_node *llnode; + + if (atomic_xchg(&c->call_rcu_ttrace_in_progress, 1)) { + if (unlikely(READ_ONCE(c->draining))) { + scoped_guard(raw_spinlock_irqsave, &c->lock) + llnode = llist_del_all(&c->free_by_rcu_ttrace); + free_all(c, llnode, !!c->percpu_size); + } + return; + } + + __do_call_rcu_ttrace(c); +} + static void free_bulk(struct bpf_mem_cache *c) { struct bpf_mem_cache *tgt = c->tgt; From fbd97dd1d3ce32d32fb7be86acc3fdb7c04fa043 Mon Sep 17 00:00:00 2001 From: Alexei Starovoitov Date: Wed, 30 Sep 2026 09:59:19 +0000 Subject: [PATCH 2/3] bpf: Fix objects stuck in free_by_rcu_ttrace do_call_rcu_ttrace() returns early when call_rcu_ttrace_in_progress is set and leaves the objects in free_by_rcu_ttrace. __free_rcu() frees waiting_for_gp_ttrace only and clears the flag. Hence the objects that free_bulk() or __free_by_rcu() added while RCU tasks trace GP was in flight stay in free_by_rcu_ttrace until free_bulk() or alloc_bulk() is called for the same bpf_mem_cache again, which may never happen. The number of such objects is not bounded. Turn call_rcu_ttrace_in_progress into three states: 0 - idle 1 - __free_rcu() is queued 2 - __free_rcu() is queued and free_by_rcu_ttrace got more objects since do_call_rcu_ttrace() sets 2. __free_rcu() does cmpxchg(1 -> 0) and starts the next GP when it fails. It cannot clear the flag first and check free_by_rcu_ttrace later, since bpf_mem_alloc_destroy() frees bpf_mem_cache without waiting for RCU callbacks when the flag is zero. Now __free_rcu() queues itself, so the one that didn't see 'draining' may do call_rcu_tasks_trace() after rcu_barrier_tasks_trace() in free_mem_alloc(). Queue it under rcu_read_lock() and do synchronize_rcu() before the barriers. Calling rcu_barrier_tasks_trace() twice works too, but creating and destroying hash maps in a loop on many cpus slows down to one free_mem_alloc() per GP and kworkers pile up. Fixes: 8d5a8011b35d ("bpf: Batch call_rcu callbacks instead of SLAB_TYPESAFE_BY_RCU.") Signed-off-by: Alexei Starovoitov Link: https://lore.kernel.org/bpf/20260930095920.601738-3-alexei.starovoitov@gmail.com Signed-off-by: Kumar Kartikeya Dwivedi --- kernel/bpf/memalloc.c | 34 ++++++++++++++++++++++++++++++++-- 1 file changed, 32 insertions(+), 2 deletions(-) diff --git a/kernel/bpf/memalloc.c b/kernel/bpf/memalloc.c index 08e4dde66cd5..15684d0fc883 100644 --- a/kernel/bpf/memalloc.c +++ b/kernel/bpf/memalloc.c @@ -118,6 +118,11 @@ struct bpf_mem_cache { struct llist_head free_by_rcu_ttrace; struct llist_head waiting_for_gp_ttrace; struct rcu_head rcu_ttrace; + /* + * 0 - idle + * 1 - __free_rcu() is queued + * 2 - __free_rcu() is queued and free_by_rcu_ttrace got more objects since + */ atomic_t call_rcu_ttrace_in_progress; raw_spinlock_t lock; }; @@ -276,6 +281,8 @@ static int free_all(struct bpf_mem_cache *c, struct llist_node *llnode, bool per return cnt; } +static void __do_call_rcu_ttrace(struct bpf_mem_cache *c); + static void __free_rcu(struct rcu_head *head) { struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu_ttrace); @@ -285,7 +292,19 @@ static void __free_rcu(struct rcu_head *head) llnode = llist_del_all(&c->waiting_for_gp_ttrace); free_all(c, llnode, !!c->percpu_size); - atomic_set(&c->call_rcu_ttrace_in_progress, 0); + + /* + * do_call_rcu_ttrace() that ran while GP was in flight left its objects + * in free_by_rcu_ttrace. This cache may never free or alloc in bulk + * again, so start the next GP from here. + * 'c' can be freed as soon as call_rcu_ttrace_in_progress is zero. + */ + if (atomic_cmpxchg(&c->call_rcu_ttrace_in_progress, 1, 0) == 1) + return; + + /* Pairs with synchronize_rcu() in free_mem_alloc() */ + guard(rcu)(); + __do_call_rcu_ttrace(c); } static void enque_to_free(struct bpf_mem_cache *c, void *obj) @@ -302,6 +321,12 @@ static void __do_call_rcu_ttrace(struct bpf_mem_cache *c) { struct llist_node *llnode, *t; + /* + * Must be done before llist_del_all(). Objects that it misses were + * added by do_call_rcu_ttrace() that will set 2 after this store. + */ + atomic_set(&c->call_rcu_ttrace_in_progress, 1); + WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace)); llist_for_each_safe(llnode, t, llist_del_all(&c->free_by_rcu_ttrace)) llist_add(llnode, &c->waiting_for_gp_ttrace); @@ -323,7 +348,7 @@ static void do_call_rcu_ttrace(struct bpf_mem_cache *c) { struct llist_node *llnode; - if (atomic_xchg(&c->call_rcu_ttrace_in_progress, 1)) { + if (atomic_xchg(&c->call_rcu_ttrace_in_progress, 2)) { if (unlikely(READ_ONCE(c->draining))) { scoped_guard(raw_spinlock_irqsave, &c->lock) llnode = llist_del_all(&c->free_by_rcu_ttrace); @@ -707,7 +732,12 @@ static void free_mem_alloc(struct bpf_mem_alloc *ma) * to wait for the pending __free_by_rcu(), and __free_rcu(). RCU Tasks * Trace grace period implies RCU grace period, so all __free_rcu don't * need extra call_rcu() (and thus extra rcu_barrier() here). + * + * __free_rcu() queues itself again unless it sees 'draining'. After + * synchronize_rcu() it either did that already or will not do it, so + * rcu_barrier_tasks_trace() cannot miss it. */ + synchronize_rcu(); rcu_barrier(); /* wait for __free_by_rcu */ rcu_barrier_tasks_trace(); /* wait for __free_rcu */ free_mem_alloc_no_barrier(ma); From aeb709c767aeca996e6011133e4f7bdb2a4af652 Mon Sep 17 00:00:00 2001 From: Alexei Starovoitov Date: Wed, 30 Sep 2026 09:59:20 +0000 Subject: [PATCH 3/3] selftests/bpf: Add a test for objects stuck in free_by_rcu_ttrace Delete all elements of BPF_F_NO_PREALLOC hash map in one batch. The first free_bulk() starts RCU tasks trace GP and the rest of the elements are freed while it's in flight. Wait for call_rcu_ttrace_in_progress to clear in bpf_mem_cache of every cpu and check that free_by_rcu_ttrace and waiting_for_gp_ttrace lists are empty. Signed-off-by: Alexei Starovoitov Link: https://lore.kernel.org/bpf/20260930095920.601738-4-alexei.starovoitov@gmail.com Signed-off-by: Kumar Kartikeya Dwivedi --- .../selftests/bpf/prog_tests/bpf_ma_ttrace.c | 60 +++++++++++++++++++ .../selftests/bpf/progs/bpf_ma_ttrace.c | 50 ++++++++++++++++ 2 files changed, 110 insertions(+) create mode 100644 tools/testing/selftests/bpf/prog_tests/bpf_ma_ttrace.c create mode 100644 tools/testing/selftests/bpf/progs/bpf_ma_ttrace.c diff --git a/tools/testing/selftests/bpf/prog_tests/bpf_ma_ttrace.c b/tools/testing/selftests/bpf/prog_tests/bpf_ma_ttrace.c new file mode 100644 index 000000000000..a1d41b109940 --- /dev/null +++ b/tools/testing/selftests/bpf/prog_tests/bpf_ma_ttrace.c @@ -0,0 +1,60 @@ +// SPDX-License-Identifier: GPL-2.0 +#include +#include "bpf_ma_ttrace.skel.h" + +#define NR_ELEMS 4096 + +/* + * The first free_bulk() starts RCU tasks trace GP. The rest of the elements are + * deleted while it's in flight. They should be freed without further alloc or + * free from this map. + */ +void test_bpf_ma_ttrace(void) +{ + LIBBPF_OPTS(bpf_test_run_opts, opts); + struct bpf_ma_ttrace *skel; + __u32 cnt = NR_ELEMS; + long *vals = NULL; + int *keys = NULL; + int i, err, fd, nr_cpus; + + skel = bpf_ma_ttrace__open_and_load(); + if (!ASSERT_OK_PTR(skel, "open_and_load")) + return; + nr_cpus = libbpf_num_possible_cpus(); + if (!ASSERT_GT(nr_cpus, 0, "nr_cpus")) + goto out; + skel->bss->nr_cpus = nr_cpus; + + keys = calloc(NR_ELEMS, sizeof(*keys)); + vals = calloc(NR_ELEMS, sizeof(*vals)); + if (!ASSERT_OK_PTR(keys, "keys") || !ASSERT_OK_PTR(vals, "vals")) + goto out; + for (i = 0; i < NR_ELEMS; i++) + keys[i] = i; + + fd = bpf_map__fd(skel->maps.htab); + err = bpf_map_update_batch(fd, keys, vals, &cnt, NULL); + if (!ASSERT_OK(err, "update_batch") || !ASSERT_EQ(cnt, NR_ELEMS, "update_cnt")) + goto out; + err = bpf_map_delete_batch(fd, keys, &cnt, NULL); + if (!ASSERT_OK(err, "delete_batch") || !ASSERT_EQ(cnt, NR_ELEMS, "delete_cnt")) + goto out; + + /* Wait for all __free_rcu() callbacks to finish */ + for (i = 0; i < 300; i++) { + err = bpf_prog_test_run_opts(bpf_program__fd(skel->progs.check_ttrace), &opts); + if (!ASSERT_OK(err, "test_run") || !ASSERT_OK(opts.retval, "retval")) + goto out; + if (!skel->bss->in_progress) + break; + usleep(100000); + } + ASSERT_EQ(skel->bss->nr_caches, nr_cpus, "nr_caches"); + ASSERT_EQ(skel->bss->in_progress, 0, "in_progress"); + ASSERT_EQ(skel->bss->not_freed, 0, "not_freed"); +out: + free(keys); + free(vals); + bpf_ma_ttrace__destroy(skel); +} diff --git a/tools/testing/selftests/bpf/progs/bpf_ma_ttrace.c b/tools/testing/selftests/bpf/progs/bpf_ma_ttrace.c new file mode 100644 index 000000000000..31b2a962e339 --- /dev/null +++ b/tools/testing/selftests/bpf/progs/bpf_ma_ttrace.c @@ -0,0 +1,50 @@ +// SPDX-License-Identifier: GPL-2.0 +#include +#include +#include +#include "bpf_kfuncs.h" + +struct { + __uint(type, BPF_MAP_TYPE_HASH); + __uint(map_flags, BPF_F_NO_PREALLOC); + __uint(max_entries, 4096); + __type(key, int); + __type(value, long); +} htab SEC(".maps"); + +extern const void __per_cpu_offset __ksym; + +int nr_cpus; +int nr_caches; +int in_progress; +int not_freed; + +/* Look at bpf_mem_cache of every cpu that htab allocates its elements from */ +SEC("syscall") +int check_ttrace(void *ctx) +{ + struct bpf_htab *h = bpf_core_cast(&htab, struct bpf_htab); + unsigned long cache = (unsigned long)BPF_CORE_READ(h, ma.cache); + const unsigned long *offsets = &__per_cpu_offset; + struct bpf_mem_cache *c; + unsigned long off; + int cpu; + + nr_caches = 0; + in_progress = 0; + not_freed = 0; + bpf_for(cpu, 0, nr_cpus) { + if (bpf_probe_read_kernel(&off, sizeof(off), offsets + cpu)) + return -1; + c = bpf_core_cast((void *)(cache + off), struct bpf_mem_cache); + if (c->unit_size) + nr_caches++; + if (c->call_rcu_ttrace_in_progress.counter) + in_progress++; + if (c->free_by_rcu_ttrace.first || c->waiting_for_gp_ttrace.first) + not_freed++; + } + return 0; +} + +char _license[] SEC("license") = "GPL";