bpf: Overwrite the htab element atomically
Subsystems (4)
cover bpf: Overwrite the htab element atomically
Hi,
The motivation for the patch set comes from the question raised by Cody Haas [1]. He asked whether or not the update of htab of map is atomic in the perspective of lookup operation. Currently, the update is not atomic because the overwrite of existing element happens in a two-steps way, but the support of atomic update is feasible. Initiallly, I only plan to support atomic update for htab of map because array of map has already supported that. Afterwards, I think it may be reasonable to support atomic update for all kinds of hash map. However, for the BPF_F_LOCK case, although the update is protected by a spin-lock, the update is still not atomic in the perspective of the lookup operation.
Please see individual patches for details. Comments are always welcome.
---
[1] : https://lore.kernel.org/xdp-newbies/[email protected]/T/#m06fcd687c6cfdbd0f9b643b227e69b479fc8c2f6
Hou Tao (3):
rculist: add hlist_nulls_replace_rcu() helper
bpf: Overwrite the element in hash map atomically selftests/bpf: Add test case for atomic htab update
include/linux/rculist_nulls.h | 42 ++++++ kernel/bpf/hashtab.c | 14 +- .../selftests/bpf/prog_tests/htab_lookup.c | 130 ++++++++++++++++++ .../testing/selftests/bpf/progs/htab_lookup.c | 13 ++ 4 files changed, 193 insertions(+), 6 deletions(-) create mode 100644 tools/testing/selftests/bpf/prog_tests/htab_lookup.c create mode 100644 tools/testing/selftests/bpf/progs/htab_lookup.c
Signature · 1 line
2.48.1
PATCH 1/3 rculist: add hlist_nulls_replace_rcu() helper
Add hlist_nulls_replace_rcu() to replace an existing element in the hash list. For the concurrent list traversal, the replace is atomic, it will find either the old element or the new element.
--- include/linux/rculist_nulls.h | 42 +++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+)
include/linux/rculist_nulls.h +42 -0
@@ -52,6 +52,14 @@ static inline void hlist_nulls_del_init_rcu(struct hlist_nulls_node *n)#define hlist_nulls_next_rcu (node) \ (*((struct hlist_nulls_node __rcu __force **)&(node)->next)) + +/** +* hlist_nulls_pprev_rcu - returns the element of the list before @node. +* @node: element of the list. +*/ +#define hlist_nulls_pprev_rcu (node) \ + (*((struct hlist_nulls_node __rcu __force **)(node)->pprev)) +/** * hlist_nulls_del_rcu - deletes entry from hash list without re-initialization * @n: the element to delete from the hash list.
@@ -145,6 +153,40 @@ static inline void hlist_nulls_add_tail_rcu(struct hlist_nulls_node *n, } } +/** +* hlist_nulls_replace_rcu - replace an element in hash list +* @n: new element to add +* @o: old element to replace +* +* Description: +* Replace an existing element in a hash list with a new one, +* while permitting racing traversals. +* +* The caller must take whatever precautions are necessary +* (such as holding appropriate locks) to avoid racing +* with another list-mutation primitive, such as hlist_nulls_add_head_rcu() +* or hlist_nulls_del_rcu(), running on this same list. +* However, it is perfectly legal to run concurrently with +* the _rcu list-traversal primitives, such as +* hlist_nulls_for_each_entry_rcu(), used to prevent memory-consistency +* problems on Alpha CPUs. Regardless of the type of CPU, the +* list-traversal primitive must be guarded by rcu_read_lock(). +*/ +static inline void hlist_nulls_replace_rcu (struct hlist_nulls_node *n, +struct hlist_nulls_node *o) +{ +struct hlist_nulls_node *next = o->next; +struct hlist_nulls_node **pprev = o->pprev; + +WRITE_ONCE (n->next, next); +WRITE_ONCE (n->pprev, pprev); +rcu_assign_pointer (hlist_nulls_pprev_rcu (o), n); + +if (!is_a_nulls (next)) +WRITE_ONCE (next->pprev, &n->next); +WRITE_ONCE (o->pprev,LIST_POISON2 ); +} +/* after that hlist_nulls_del will work */ static inline void hlist_nulls_add_fake (struct hlist_nulls_node *n) {
Signature · 1 line
2.48.1
PATCH 2/3 bpf: Overwrite the element in hash map atomically
In reply to Hou TaoCurrently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old element
It is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.
--- kernel/bpf/hashtab.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-)
kernel/bpf/hashtab.c +8 -6
@@ -1179,12 +1179,14 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, goto err; } -/* add new element to the head of the list, so that -* concurrent search will find it before old elem -*/ -hlist_nulls_add_head_rcu (&l_new->hash_node, head); -if (l_old) { -hlist_nulls_del_rcu (&l_old->hash_node); +if (!l_old) { +hlist_nulls_add_head_rcu (&l_new->hash_node, head); + }else { +/* Replace the old element atomically, so that +* concurrent search will find either the new element or +* the old element. +*/ +hlist_nulls_replace_rcu (&l_new->hash_node, &l_old->hash_node);/* l_old has already been stashed in htab->extra_elems, free * its special fields before it is available for reuse. Also
Signature · 1 line
2.48.1
PATCH 3/3 selftests/bpf: Add test case for atomic htab update
In reply to Hou TaoAdd a test case to verify the atomic update of existing element in hash map. The test proceeds in three steps:
1) fill the map with keys in the range [0, 63]
2) create 8 threads to lookup these keys concurrently
3) create 2 threads to overwrite these keys concurrently
Without atomic update support, the lookup operation may return -ENOENT error and the test will fail. After the atomic-update change, the lookup operation will always return 0 and the test will succeed.
--- .../selftests/bpf/prog_tests/htab_lookup.c | 130 ++++++++++++++++++ .../testing/selftests/bpf/progs/htab_lookup.c | 13 ++ 2 files changed, 143 insertions(+) create mode 100644 tools/testing/selftests/bpf/prog_tests/htab_lookup.c create mode 100644 tools/testing/selftests/bpf/progs/htab_lookup.c
tools/testing/selftests/bpf/prog_tests/htab_lookup.c new file +130 -0
@@ -0,0 +1,130 @@ +// SPDX-License-Identifier: GPL-2.0 +#define _GNU_SOURCE +#include <stdbool.h> +#include <test_progs.h> +#include "htab_lookup.skel.h" + +struct htab_op_ctx { +int fd; +int loop; +unsigned int entries; +bool stop; +}; + +static void *htab_lookup_fn (void *arg) +{ +struct htab_op_ctx *ctx = arg; +int i =0 ; + +while (i++ < ctx->loop && !ctx->stop) { +unsigned int j; + +for (j =0 ; j < ctx->entries; j++) { +unsigned long key = j, value; +int err; + + err =bpf_map_lookup_elem (ctx->fd, &key, &value); +if (err) { + ctx->stop = true; +return (void *)(long )err; + } + } + } + +return NULL ; +} + +static void *htab_update_fn (void *arg) +{ +struct htab_op_ctx *ctx = arg; +int i =0 ; + +while (i++ < ctx->loop && !ctx->stop) { +unsigned int j; + +for (j =0 ; j < ctx->entries; j++) { +unsigned long key = j, value = j; +int err; + + err =bpf_map_update_elem (ctx->fd, &key, &value,BPF_EXIST ); +if (err) { +if (err == -ENOMEM ) +continue ; + ctx->stop = true; +return (void *)(long )err; + } + } + } + +return NULL ; +} + +static int setup_htab (int fd,unsigned int entries) +{ +unsigned int i; + +for (i =0 ; i < entries; i++) { +unsigned long key = i, value = i; +int err; + + err =bpf_map_update_elem (fd, &key, &value,0 ); +if (!ASSERT_OK (err,"init update" )) +return -1 ; + } + +return 0 ; +} + +void test_htab_lookup (void ) +{ +unsigned int i, wr_nr =2 , rd_nr =8 ; +pthread_t tids[wr_nr + rd_nr]; +struct htab_lookup *skel; +struct htab_op_ctx ctx; +int err; + + skel =htab_lookup__open_and_load (); +if (!ASSERT_OK_PTR (skel,"htab_lookup__open_and_load" )) +return ; + + ctx.fd =bpf_map__fd (skel->maps.htab); + ctx.loop =50 ; + ctx.stop = false; + ctx.entries =64 ; + + err =setup_htab (ctx.fd, ctx.entries); +if (err) + goto destroy; + +memset (tids,0 ,sizeof (tids)); +for (i =0 ; i < wr_nr; i++) { + err =pthread_create (&tids[i],NULL , htab_update_fn, &ctx); +if (!ASSERT_OK (err,"pthread_create" )) { + ctx.stop = true; + goto reap; + } + } +for (i =0 ; i < rd_nr; i++) { + err =pthread_create (&tids[i + wr_nr],NULL , htab_lookup_fn, &ctx); +if (!ASSERT_OK (err,"pthread_create" )) { + ctx.stop = true; + goto reap; + } + } + +reap: +for (i =0 ; i < wr_nr + rd_nr; i++) { +void *ret =NULL ; +char desc[32 ]; + +if (!tids[i]) +continue ; + +snprintf (desc,sizeof (desc),"thread %u" , i +1 ); + err =pthread_join (tids[i], &ret); +ASSERT_OK (err, desc); +ASSERT_EQ (ret,NULL , desc); + } +destroy: +htab_lookup__destroy (skel); +}
tools/testing/selftests/bpf/progs/htab_lookup.c new file +13 -0
@@ -0,0 +1,13 @@ +// SPDX-License-Identifier: GPL-2.0 +#include <linux/bpf.h> +#include <bpf/bpf_helpers.h> + +char _license[] SEC("license") = "GPL"; + +struct { + __uint(type, BPF_MAP_TYPE_HASH); + __uint(max_entries, 64); + __type(key, unsigned long); + __type(value, unsigned long); + __uint(map_flags, BPF_F_NO_PREALLOC); +} htab SEC(".maps");
Signature · 1 line
2.48.1
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
In reply to Hou Tao+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 34 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Signed-off-by Hou Tao <[email protected]>--- kernel/bpf/hashtab.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-)kernel/bpf/hashtab.c +1
-6@@ -1179,12 +1179,14 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, goto err; } -/* add new element to the head of the list, so that -* concurrent search will find it before old elem -*/ -hlist_nulls_add_head_rcu (&l_new->hash_node, head); -if (l_old) { -hlist_nulls_del_rcu (&l_old->hash_node); +if (!l_old) {
kernel/bpf/hashtab.c static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, · line 1192
Quoted 10 lines
kernel/bpf/hashtab.c +7
-0+hlist_nulls_add_head_rcu (&l_new->hash_node, head); + }else { +/* Replace the old element atomically, so that +* concurrent search will find either the new element or +* the old element. +*/ +hlist_nulls_replace_rcu (&l_new->hash_node, &l_old->hash_node);/* l_old has already been stashed in htab->extra_elems, free * its special fields before it is available for reuse. Also
After thinking about it the second time, the atomic list replacement on the update side is enough to make lookup operation always find the existing element. However, due to the immediate reuse, the lookup may find an unexpected value. Maybe we should disable the immediate reuse for specific map (e.g., htab of maps).
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Quoted 46 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 41 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Signed-off-by Hou Tao <[email protected]>--- kernel/bpf/hashtab.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-)kernel/bpf/hashtab.c +8
-6@@ -1179,12 +1179,14 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, goto err; } -/* add new element to the head of the list, so that -* concurrent search will find it before old elem -*/ -hlist_nulls_add_head_rcu (&l_new->hash_node, head); -if (l_old) { -hlist_nulls_del_rcu (&l_old->hash_node); +if (!l_old) { +hlist_nulls_add_head_rcu (&l_new->hash_node, head); + }else { +/* Replace the old element atomically, so that +* concurrent search will find either the new element or +* the old element. +*/ +hlist_nulls_replace_rcu (&l_new->hash_node, &l_old->hash_node);
Quoted 9 lines
kernel/bpf/hashtab.c static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, · line 1192
Quoted 3 lines
kernel/bpf/hashtab.c +0
-0/* l_old has already been stashed in htab->extra_elems, free * its special fields before it is available for reuse. Also After thinking about it the second time, the atomic list replacement on the update side is enough to make lookup operation always find the existing element. However, due to the immediate reuse, the lookup may find an unexpected value. Maybe we should disable the immediate reuse for specific map (e.g., htab of maps).
Hmm, in an RCU-protected data structure, reusing the memory before an RCU grace period has elapsed is just as wrong as freeing it, isn't it? I.e., the reuse logic should have some kind of call_rcu redirection to be completely correct?
-Toke
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Hi Toke,
Quoted 56 lines · Toke Høiland-Jørgensen
Quoted 46 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 41 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Signed-off-by Hou Tao <[email protected]>--- kernel/bpf/hashtab.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-)kernel/bpf/hashtab.c +8
-6@@ -1179,12 +1179,14 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, goto err; } -/* add new element to the head of the list, so that -* concurrent search will find it before old elem -*/ -hlist_nulls_add_head_rcu (&l_new->hash_node, head); -if (l_old) { -hlist_nulls_del_rcu (&l_old->hash_node); +if (!l_old) { +hlist_nulls_add_head_rcu (&l_new->hash_node, head); + }else { +/* Replace the old element atomically, so that +* concurrent search will find either the new element or +* the old element. +*/ +hlist_nulls_replace_rcu (&l_new->hash_node, &l_old->hash_node);Quoted 8 lines
kernel/bpf/hashtab.c static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, · line 1192
Quoted 3 lines
kernel/bpf/hashtab.c +0
-0/* l_old has already been stashed in htab->extra_elems, free * its special fields before it is available for reuse. Also After thinking about it the second time, the atomic list replacement on the update side is enough to make lookup operation always find the existing element. However, due to the immediate reuse, the lookup may find an unexpected value. Maybe we should disable the immediate reuse for specific map (e.g., htab of maps).
Quoted 4 lines
Hmm, in an RCU-protected data structure, reusing the memory before an RCU grace period has elapsed is just as wrong as freeing it, isn't it? I.e., the reuse logic should have some kind of
call_rcuredirection to be completely correct?
Not for all cases. There is SLAB_TYPESAFE_BY_RCU-typed slab. For hash map, the reuse is also tricky (e.g., the goto again case in lookup_nulls_elem_raw), however it can not prevent the lookup procedure from returning unexpected value. I had post a patch set [1] to "fix" that, but Alexei said it is "a known quirk". Here I am not sure about whether it is reasonable to disable the reuse for htab of maps only. I will post a v2 for the patch set.
[1]:
https://lore.kernel.org/bpf/[email protected]/
Quoted 1 line
-Toke
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Quoted 23 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.
Quoted 50 lines
Quoted 37 lines
Quoted 29 lines
Quoted 29 lines
Signed-off-by Hou Tao <[email protected]>--- kernel/bpf/hashtab.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-)kernel/bpf/hashtab.c +8
-6@@ -1179,12 +1179,14 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, goto err; } -/* add new element to the head of the list, so that -* concurrent search will find it before old elem -*/ -hlist_nulls_add_head_rcu (&l_new->hash_node, head); -if (l_old) { -hlist_nulls_del_rcu (&l_old->hash_node); +if (!l_old) { +hlist_nulls_add_head_rcu (&l_new->hash_node, head); + }else { +/* Replace the old element atomically, so that +* concurrent search will find either the new element or +* the old element. +*/ +hlist_nulls_replace_rcu (&l_new->hash_node, &l_old->hash_node);Quoted 8 lines
kernel/bpf/hashtab.c static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, · line 1192
Quoted 3 lines
kernel/bpf/hashtab.c +0
-0/* l_old has already been stashed in htab->extra_elems, free * its special fields before it is available for reuse. Also After thinking about it the second time, the atomic list replacement on the update side is enough to make lookup operation always find the existing element. However, due to the immediate reuse, the lookup may find an unexpected value. Maybe we should disable the immediate reuse for specific map (e.g., htab of maps).
Quoted 4 lines
Hmm, in an RCU-protected data structure, reusing the memory before an RCU grace period has elapsed is just as wrong as freeing it, isn't it? I.e., the reuse logic should have some kind of
call_rcuredirection to be completely correct?Not for all cases. There is SLAB_TYPESAFE_BY_RCU-typed slab. For hash map, the reuse is also tricky (e.g., the goto again case in
lookup_nulls_elem_raw), however it can not prevent the lookup procedure from returning unexpected value. I had post a patch set [1] to "fix" that, but Alexei said it is "a known quirk". Here I am not sure about whether it is reasonable to disable the reuse for htab of maps only. I will post a v2 for the patch set.
Quoted 2 lines
yes. we still have to keep prealloc as default for now :( Eventually bpf_mem_alloc is replaced with fully re-entrant and safe kmalloc, then we can do fully re-entrant and safe kfree_rcu. Then we can talk about closing this quirk. Until then the prog has to deal with immediate reuse. That was the case for a decade already.
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Hi,
Quoted 23 lines · Alexei Starovoitov
Quoted 22 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.
Quoted 6 lines
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.
The point is that overwriting an existed element may lead to concurrent lookups return ENOENT as demonstrated by the added selftest and the patch tried to "fix" that. However, it seems using hlist_nulls_replace_rcu() for the overwriting update is still not enough. Because when the lookup procedure found the old element, the old element may be reusing, the comparison of the map key may fail, and the lookup procedure may still return ENOENT.
Quoted 49 lines
Quoted 49 lines
Quoted 37 lines
Quoted 29 lines
Quoted 29 lines
Signed-off-by Hou Tao <[email protected]>--- kernel/bpf/hashtab.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-)kernel/bpf/hashtab.c +8
-6@@ -1179,12 +1179,14 @@ static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, goto err; } -/* add new element to the head of the list, so that -* concurrent search will find it before old elem -*/ -hlist_nulls_add_head_rcu (&l_new->hash_node, head); -if (l_old) { -hlist_nulls_del_rcu (&l_old->hash_node); +if (!l_old) { +hlist_nulls_add_head_rcu (&l_new->hash_node, head); + }else { +/* Replace the old element atomically, so that +* concurrent search will find either the new element or +* the old element. +*/ +hlist_nulls_replace_rcu (&l_new->hash_node, &l_old->hash_node);Quoted 8 lines
kernel/bpf/hashtab.c static long htab_map_update_elem(struct bpf_map *map, void *key, void *value, · line 1192
Quoted 3 lines
kernel/bpf/hashtab.c +0
-0/* l_old has already been stashed in htab->extra_elems, free * its special fields before it is available for reuse. Also After thinking about it the second time, the atomic list replacement on the update side is enough to make lookup operation always find the existing element. However, due to the immediate reuse, the lookup may find an unexpected value. Maybe we should disable the immediate reuse for specific map (e.g., htab of maps).
Quoted 4 lines
Hmm, in an RCU-protected data structure, reusing the memory before an RCU grace period has elapsed is just as wrong as freeing it, isn't it? I.e., the reuse logic should have some kind of
call_rcuredirection to be completely correct?Not for all cases. There is SLAB_TYPESAFE_BY_RCU-typed slab. For hash map, the reuse is also tricky (e.g., the goto again case in
lookup_nulls_elem_raw), however it can not prevent the lookup procedure from returning unexpected value. I had post a patch set [1] to "fix" that, but Alexei said it is "a known quirk". Here I am not sure about whether it is reasonable to disable the reuse for htab of maps only. I will post a v2 for the patch set.
Quoted 8 lines
Quoted 2 lines
yes. we still have to keep prealloc as default for now :( Eventually
bpf_mem_allocis replaced with fully re-entrant and safe kmalloc, then we can do fully re-entrant and safekfree_rcu. Then we can talk about closing this quirk. Until then the prog has to deal with immediate reuse. That was the case for a decade already.
I see. In v2 I will fallback to the original idea: adding a standalone update procedure for htab of maps in which it will atomically overwrite the map_ptr just like array of maps does.
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Quoted 34 lines · Hou Tao
Hi,
Quoted 23 lines · Alexei Starovoitov
Quoted 22 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Quoted 6 lines
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.
Quoted 7 lines
The point is that overwriting an existed element may lead to concurrent lookups return ENOENT as demonstrated by the added selftest and the patch tried to "fix" that. However, it seems using
hlist_nulls_replace_rcu()for the overwriting update is still not enough. Because when the lookup procedure found the old element, the old element may be reusing, the comparison of the map key may fail, and the lookup procedure may still return ENOENT.
you mean l_old == l_new ? I don't think it's possible
within htab_map_update_elem(),
but htab_map_delete_elem() doing hlist_nulls_del_rcu()
then free_htab_elem, htab_map_update_elem, alloc, hlist_nulls_add_head_rcu and just deleted elem is reused to be inserted
into another bucket.
I'm not sure whether this new hlist_nulls_replace_rcu() primitive works with nulls logic.
So back to the problem statement..
Are you saying that adding new to a head while lookup is in the middle is causing it to miss an element that
is supposed to be updated assuming atomicity of the update? While now update_elem() is more like a sequence of delete + insert?
Hmm.
Quoted 3 lines
I see. In v2 I will fallback to the original idea: adding a standalone update procedure for htab of maps in which it will atomically overwrite the
map_ptrjust like array of maps does.
hold on. is this only for hash-of-maps ?
How that non-atomic update manifested in real production?
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Quoted 54 lines · Alexei Starovoitov
Quoted 34 lines · Hou Tao
Hi,
Quoted 23 lines · Alexei Starovoitov
Quoted 22 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Quoted 6 lines
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.Quoted 7 lines
The point is that overwriting an existed element may lead to concurrent lookups return ENOENT as demonstrated by the added selftest and the patch tried to "fix" that. However, it seems using
hlist_nulls_replace_rcu()for the overwriting update is still not enough. Because when the lookup procedure found the old element, the old element may be reusing, the comparison of the map key may fail, and the lookup procedure may still return ENOENT.you mean
l_old==l_new? I don't think it's possible
withinhtab_map_update_elem(),
buthtab_map_delete_elem()doinghlist_nulls_del_rcu()
thenfree_htab_elem,htab_map_update_elem, alloc,hlist_nulls_add_head_rcuand just deleted elem is reused to be inserted
into another bucket.I'm not sure whether this new
hlist_nulls_replace_rcu()primitive works with nulls logic.
Quoted 7 lines
So back to the problem statement..
Are you saying that adding new to a head while lookup is in the middle is causing it to miss an element that
is supposed to be updated assuming atomicity of the update? While nowupdate_elem()is more like a sequence of delete + insert?Hmm.
Yes, exactly that. Because update_elem is actually a delete + insert (actually an insert + delete, I think?), it is possible for a concurrent lookup to see no element instead of either the old or new value.
Quoted 6 lines
Quoted 4 lines
I see. In v2 I will fallback to the original idea: adding a standalone update procedure for htab of maps in which it will atomically overwrite the
map_ptrjust like array of maps does.hold on. is this only for hash-of-maps ?
I believe this was also replicated for hash as well as hash-of-maps. Cody can confirm, or use the reproducer he has to test that case.
Quoted 2 lines
How that non-atomic update manifested in real production?
See [1] (in the cover letter for this series, also replicated below).
[1] : https://lore.kernel.org/xdp-newbies/[email protected]/T/#m06fcd687c6cfdbd0f9b643b227e69b479fc8c2f6
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Hi,
Quoted 42 lines · Zvi Effron
Quoted 40 lines · Alexei Starovoitov
Quoted 39 lines · Hou Tao
Hi,
Quoted 23 lines · Alexei Starovoitov
Quoted 22 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Quoted 6 lines
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.The point is that overwriting an existed element may lead to concurrent lookups return ENOENT as demonstrated by the added selftest and the patch tried to "fix" that. However, it seems using
hlist_nulls_replace_rcu()for the overwriting update is still not enough. Because when the lookup procedure found the old element, the old element may be reusing, the comparison of the map key may fail, and the lookup procedure may still return ENOENT.
Quoted 6 lines
Quoted 6 lines
you mean
l_old==l_new? I don't think it's possible
withinhtab_map_update_elem(),
buthtab_map_delete_elem()doinghlist_nulls_del_rcu()
thenfree_htab_elem,htab_map_update_elem, alloc,hlist_nulls_add_head_rcuand just deleted elem is reused to be inserted
into another bucket.
No. I mean the following procedure in which the lookup procedure finds the old element and tries to match its key, one update procedure has already deleted the old element, and another update procedure is reusing the old element:
lookup procedure A
A: find the old element (instead of the new old)
update procedure B
B: delete the old element
update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in the same bucket)
A: the key is mismatched and return -ENOENT.
It can be reproduced by increasing ctx.loop in the selftest.
Quoted 10 lines
Quoted 10 lines
I'm not sure whether this new
hlist_nulls_replace_rcu()primitive works with nulls logic.So back to the problem statement..
Are you saying that adding new to a head while lookup is in the middle is causing it to miss an element that
is supposed to be updated assuming atomicity of the update? While nowupdate_elem()is more like a sequence of delete + insert?
Quoted 4 lines
Quoted 1 line
Hmm.
Yes, exactly that. Because
update_elemis actually a delete + insert (actually an insert + delete, I think?), it is possible for a concurrent lookup to see no element instead of either the old or new value.
Yep.
Quoted 7 lines
Quoted 5 lines
Quoted 4 lines
I see. In v2 I will fallback to the original idea: adding a standalone update procedure for htab of maps in which it will atomically overwrite the
map_ptrjust like array of maps does.hold on. is this only for hash-of-maps ?
I believe this was also replicated for hash as well as hash-of-maps. Cody can confirm, or use the reproducer he has to test that case.
The fix for hash-of-maps will be much simpler and the returned map pointer will be correct. For other types of hash map, beside switching to hlist_nulls_replace_rcu(), I think we also need to detect the reuse of element during the loop: if the element has been reused, the lookup should restart from the head again. However, even withing the above two fixes, the value returned by lookup procedure for other types of hash map may still be incorrect (as shown long time ago [1]), so I think maybe for other types of map, the atomic update doesn't matter too much.
[1]:
https://lore.kernel.org/bpf/[email protected]/
Quoted 6 lines
Quoted 3 lines
How that non-atomic update manifested in real production?
See [1] (in the cover letter for this series, also replicated below).
[1] : https://lore.kernel.org/xdp-newbies/[email protected]/T/#m06fcd687c6cfdbd0f9b643b227e69b479fc8c2f6
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Quoted 61 lines · Hou Tao
Hi,
Quoted 42 lines · Zvi Effron
Quoted 40 lines · Alexei Starovoitov
Quoted 39 lines · Hou Tao
Hi,
Quoted 23 lines · Alexei Starovoitov
Quoted 22 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Quoted 6 lines
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.The point is that overwriting an existed element may lead to concurrent lookups return ENOENT as demonstrated by the added selftest and the patch tried to "fix" that. However, it seems using
hlist_nulls_replace_rcu()for the overwriting update is still not enough. Because when the lookup procedure found the old element, the old element may be reusing, the comparison of the map key may fail, and the lookup procedure may still return ENOENT.Quoted 6 lines
Quoted 6 lines
you mean
l_old==l_new? I don't think it's possible
withinhtab_map_update_elem(),
buthtab_map_delete_elem()doinghlist_nulls_del_rcu()
thenfree_htab_elem,htab_map_update_elem, alloc,hlist_nulls_add_head_rcuand just deleted elem is reused to be inserted
into another bucket.No. I mean the following procedure in which the lookup procedure finds the old element and tries to match its key, one update procedure has already deleted the old element, and another update procedure is reusing the old element:
lookup procedure A
A: find the old element (instead of the new old)
Quoted 7 lines
update procedure B
B: delete the old element
update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in
the same bucket)A: the key is mismatched and return
-ENOENT.
This is fine. It's just normal reuse.
Orthogonal to 'update as insert+delete' issue.
Quoted 21 lines
It can be reproduced by increasing ctx.loop in the selftest.
Quoted 10 lines
Quoted 10 lines
I'm not sure whether this new
hlist_nulls_replace_rcu()primitive works with nulls logic.So back to the problem statement..
Are you saying that adding new to a head while lookup is in the middle is causing it to miss an element that
is supposed to be updated assuming atomicity of the update? While nowupdate_elem()is more like a sequence of delete + insert?Quoted 4 lines
Quoted 1 line
Hmm.
Yes, exactly that. Because
update_elemis actually a delete + insert (actually an insert + delete, I think?), it is possible for a concurrent lookup to see no element instead of either the old or new value.Yep.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
I see. In v2 I will fallback to the original idea: adding a standalone update procedure for htab of maps in which it will atomically overwrite the
map_ptrjust like array of maps does.
Quoted 7 lines
Quoted 3 lines
Quoted 1 line
hold on. is this only for hash-of-maps ?
I believe this was also replicated for hash as well as hash-of-maps. Cody can confirm, or use the reproducer he has to test that case.
The fix for hash-of-maps will be much simpler and the returned map pointer will be correct. For other types of hash map, beside switching to
hlist_nulls_replace_rcu(),
It's been a long time since I looked into rcu_nulls details. Pls help me understand that this new replace_rcu_nulls()
is correct from nulls pov,
If it is then this patch set may be the right answer to non-atomic update.
And for the future, please please focus on "why" part in the cover letter and commit logs instead of "what".
Since the only thing I got from the log was:
"Currently, the update is not atomic
because the overwrite of existing element happens in a two-steps way, but the support of atomic update is feasible".
"is feasible" doesn't explain "why".
Link to xdp-newbie question is nice for additional context, but reviewers should not need to go and read some thread somewhere to understand "why" part.
All of it should be in the commit log.
Quoted 5 lines
map may still be incorrect (as shown long time ago [1]), so I think maybe for other types of map, the atomic update doesn't matter too much.
A thread from 3 years ago ?! Sorry, it's not helpful to ask people to page-in such an old context with lots of follow ups that may or may not be relevant today.
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Hi,
Quoted 66 lines · Alexei Starovoitov
Quoted 65 lines · Hou Tao
Hi,
Quoted 42 lines · Zvi Effron
Quoted 40 lines · Alexei Starovoitov
Quoted 39 lines · Hou Tao
Hi,
Quoted 23 lines · Alexei Starovoitov
Quoted 22 lines · Hou Tao
Hi Toke,
Quoted 19 lines · Toke Høiland-Jørgensen
Quoted 17 lines · Hou Tao
+cc Cody Haas
Sorry for the resend. I sent the reply in the HTML format.
Quoted 12 lines · Hou Tao
Currently, the update of existing element in hash map involves two steps:
1) insert the new element at the head of the hash list
2) remove the old elementIt is possible that the concurrent lookup operation may fail to find either the old element or the new element if the lookup operation starts before the addition and continues after the removal.
Therefore, replacing the two-step update with an atomic update. After the change, the update will be atomic in the perspective of the lookup
operation: it will either find the old element or the new element.Quoted 6 lines
I'm missing the point.
This "atomic" replacement doesn't really solve anything. lookup will see one element.
That element could be deleted by another thread. bucket lock and either two step update or single step don't change anything from the pov of bpf prog doing lookup.The point is that overwriting an existed element may lead to concurrent lookups return ENOENT as demonstrated by the added selftest and the patch tried to "fix" that. However, it seems using
hlist_nulls_replace_rcu()for the overwriting update is still not enough. Because when the lookup procedure found the old element, the old element may be reusing, the comparison of the map key may fail, and the lookup procedure may still return ENOENT.Quoted 6 lines
Quoted 6 lines
you mean
l_old==l_new? I don't think it's possible
withinhtab_map_update_elem(),
buthtab_map_delete_elem()doinghlist_nulls_del_rcu()
thenfree_htab_elem,htab_map_update_elem, alloc,hlist_nulls_add_head_rcuand just deleted elem is reused to be inserted
into another bucket.No. I mean the following procedure in which the lookup procedure finds the old element and tries to match its key, one update procedure has already deleted the old element, and another update procedure is reusing the old element:
lookup procedure A
A: find the old element (instead of the new old)update procedure B
B: delete the old element
update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in
the same bucket)
Quoted 3 lines
Quoted 1 line
A: the key is mismatched and return
-ENOENT.This is fine. It's just normal reuse.
Orthogonal to 'update as insert+delete' issue.
OK. However, it will break the lookup procedure because it expects it will return an valid result instead of -ENOENT.
Quoted 22 lines
Quoted 22 lines
It can be reproduced by increasing ctx.loop in the selftest.
Quoted 9 lines
Quoted 9 lines
I'm not sure whether this new
hlist_nulls_replace_rcu()primitive works with nulls logic.So back to the problem statement..
Are you saying that adding new to a head while lookup is in the middle is causing it to miss an element that
is supposed to be updated assuming atomicity of the update? While nowupdate_elem()is more like a sequence of delete + insert?Quoted 4 lines
Quoted 1 line
Hmm.
Yes, exactly that. Because
update_elemis actually a delete + insert (actually an insert + delete, I think?), it is possible for a concurrent lookup to see no element instead of either the old or new value.Yep.
Quoted 6 lines
Quoted 4 lines
Quoted 3 lines
I see. In v2 I will fallback to the original idea: adding a standalone update procedure for htab of maps in which it will atomically overwrite the
map_ptrjust like array of maps does.hold on. is this only for hash-of-maps ?
I believe this was also replicated for hash as well as hash-of-maps. Cody can confirm, or use the reproducer he has to test that case.
Quoted 7 lines
Quoted 3 lines
The fix for hash-of-maps will be much simpler and the returned map pointer will be correct. For other types of hash map, beside switching to
hlist_nulls_replace_rcu(),It's been a long time since I looked into
rcu_nullsdetails. Pls help me understand that this newreplace_rcu_nulls()
is correct from nulls pov,
If it is then this patch set may be the right answer to non-atomic update.
If I understand correctly, only the manipulations of ->first pointer and ->next pointer need to take care of nulls pointer.
Quoted 9 lines
And for the future, please please focus on "why" part in the cover letter and commit logs instead of "what".
Since the only thing I got from the log was:
"Currently, the update is not atomic
because the overwrite of existing element happens in a two-steps way, but the support of atomic update is feasible".
Quoted 6 lines
"is feasible" doesn't explain "why".
Link to xdp-newbie question is nice for additional context, but reviewers should not need to go and read some thread somewhere to understand "why" part.
All of it should be in the commit log.
OK. My original thought is that is a reported problem, so an extra link will be enough. Will try to add more context next time.
Quoted 9 lines
Quoted 6 lines
map may still be incorrect (as shown long time ago [1]), so I think maybe for other types of map, the atomic update doesn't matter too much.
A thread from 3 years ago ?! Sorry, it's not helpful to ask people to page-in such an old context with lots of follow ups that may or may not be relevant today.
Let me reuse part of the diagram above to explain how does the lookup procedure may return a incorrect value:
lookup procedure A
A: find the old element (instead of the new element)
update procedure B
B: delete the old element
update procedure C on the same CPU:
A: the key is matched and return the value in the element
C: reuse the old element (overwrite its key and value)
A: read the value (it is incorrect, because it has been reused and overwritten)
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Quoted 11 lines · Hou Tao
Quoted 11 lines
Quoted 11 lines
lookup procedure A
A: find the old element (instead of the new old)update procedure B
B: delete the old element
update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in
the same bucket)
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
A: the key is mismatched and return
-ENOENT.This is fine. It's just normal reuse.
Orthogonal to 'update as insert+delete' issue.OK. However, it will break the lookup procedure because it expects it will return an valid result instead of
-ENOENT.
What do you mean 'breaks the lookup' ? lookup_elem_raw() matches hash, and then it memcmp(key), if the element is reused anything can happen. Either it succeeds in memcmp() and returns an elem, or miscompares in memcmp().
Both are expected, because elems are reused in place.
And this behavior is expected and not-broken, because bpf prog that does lookup on one cpu and deletes the same element on the other cpu is asking for trouble. bpf infra guarantees the safety of the kernel. It doesn't guarantee that bpf progs are sane.
Quoted 7 lines
Quoted 4 lines
It's been a long time since I looked into
rcu_nullsdetails. Pls help me understand that this newreplace_rcu_nulls()
is correct from nulls pov,
If it is then this patch set may be the right answer to non-atomic update.If I understand correctly, only the manipulations of ->first pointer and ->next pointer need to take care of nulls pointer.
hmm. I feel we're still talking past each other.
See if (get_nulls_value() == ...) in lookup_nulls_elem_raw(). It's there because of reuse. The lookup can start in one bucket and finish in another.
Quoted 39 lines
Quoted 9 lines
And for the future, please please focus on "why" part in the cover letter and commit logs instead of "what".
Since the only thing I got from the log was:
"Currently, the update is not atomic
because the overwrite of existing element happens in a two-steps way, but the support of atomic update is feasible".Quoted 6 lines
"is feasible" doesn't explain "why".
Link to xdp-newbie question is nice for additional context, but reviewers should not need to go and read some thread somewhere to understand "why" part.
All of it should be in the commit log.OK. My original thought is that is a reported problem, so an extra link will be enough. Will try to add more context next time.
Quoted 9 lines
Quoted 6 lines
map may still be incorrect (as shown long time ago [1]), so I think maybe for other types of map, the atomic update doesn't matter too much.
A thread from 3 years ago ?! Sorry, it's not helpful to ask people to page-in such an old context with lots of follow ups that may or may not be relevant today.
Let me reuse part of the diagram above to explain how does the lookup procedure may return a incorrect value:
lookup procedure A
A: find the old element (instead of the new element)update procedure B
B: delete the old element
update procedure C on the same CPU:
Quoted 6 lines
A: the key is matched and return the value in the element
C: reuse the old element (overwrite its key and value)
A: read the value (it is incorrect, because it has been reused and overwritten)
... and it's fine. It's by design. It's an element reuse behavior.
Long ago hashmap had two modes: prealloc (default) and NO_PREALLOC (call_rcu + kfree)
The call_rcu part was there to make things safe.
The memory cannot be kfree-ed to the kernel until RCU GP. With bpf_mem_alloc hashmap elements are freed back to bpf_ma right away. Hashmap is doing bpf_mem_cache_free()
(instead of bpf_mem_cache_free_rcu()) because users need speed. So since 2022 both prealloc and no_prealloc reuse elements. We can consider a new flag for the hash map like F_REUSE_AFTER_RCU_GP that will use _rcu() flavor of freeing into bpf_ma,
but it has to have a strong reason.
And soon as we add it the default with prealloc would need to use call_rcu() too, right?
and that becomes nightmare, since bpf prog can easily DoS the system. Even if we use bpf_mem_cache_free_rcu() only, the DoS is a concern. Unlike new things like bpf_obj_new/obj_drop the hashmap
is unpriv, so concerns are drastically different.
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
Hi,
Quoted 22 lines · Alexei Starovoitov
Quoted 9 lines · Hou Tao
Quoted 9 lines
Quoted 9 lines
lookup procedure A
A: find the old element (instead of the new old)update procedure B
B: delete the old element
update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in
the same bucket)Quoted 5 lines
Quoted 3 lines
Quoted 1 line
A: the key is mismatched and return
-ENOENT.This is fine. It's just normal reuse.
Orthogonal to 'update as insert+delete' issue.OK. However, it will break the lookup procedure because it expects it will return an valid result instead of
-ENOENT.What do you mean 'breaks the lookup' ?
lookup_elem_raw()matches hash, and then it memcmp(key), if the element is reused anything can happen. Either it succeeds inmemcmp()and returns an elem, or miscompares inmemcmp().
Both are expected, because elems are reused in place.
Quoted 3 lines
And this behavior is expected and not-broken, because bpf prog that does lookup on one cpu and deletes the same element on the other cpu is asking for trouble.
It seems I mislead you in the above example. It is also possible for doing lookup in one CPU and updating the same element in other CPU. So does such concurrence also look insane ?
lookup procedure A
A: find the old element (instead of the new old)
update procedure B
B: overwrite the same element and free the old element update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in the same bucket)
A: the key is mismatched and return -ENOENT.
Quoted 7 lines
bpf infra guarantees the safety of the kernel.
It doesn't guarantee that bpf progs are sane.Quoted 4 lines
Quoted 4 lines
It's been a long time since I looked into
rcu_nullsdetails. Pls help me understand that this newreplace_rcu_nulls()
is correct from nulls pov,
If it is then this patch set may be the right answer to non-atomic update.
Quoted 6 lines
Quoted 2 lines
If I understand correctly, only the manipulations of ->first pointer and ->next pointer need to take care of nulls pointer.
hmm. I feel we're still talking past each other.
See if (get_nulls_value()== ...) inlookup_nulls_elem_raw(). It's there because of reuse. The lookup can start in one bucket and finish in another.
Yes. I noticed that. However, it happens when the deleted element is reused and inserted to a different bucket, right ? For replace_rcu_nulls(), the reuse of old element is possible only after replace_rcu_nulls() completes, so I think it is irrelevant.
Quoted 1 line
If it is then this patch set may be the right answer to non-atomic update.
I also didn't follow that. Due to the immediate reuse, the lookup procedure may still return -ENOENT when there is concurrent overwriting of the same element as show above, so I think it only reduce the possibility. I still prefer to implement a separated update procedure of htab of maps to fix the atomic update problem completely for htab of maps first.
Quoted 64 lines
Quoted 37 lines
Quoted 9 lines
And for the future, please please focus on "why" part in the cover letter and commit logs instead of "what".
Since the only thing I got from the log was:
"Currently, the update is not atomic
because the overwrite of existing element happens in a two-steps way, but the support of atomic update is feasible".Quoted 6 lines
"is feasible" doesn't explain "why".
Link to xdp-newbie question is nice for additional context, but reviewers should not need to go and read some thread somewhere to understand "why" part.
All of it should be in the commit log.OK. My original thought is that is a reported problem, so an extra link will be enough. Will try to add more context next time.
Quoted 8 lines
Quoted 5 lines
map may still be incorrect (as shown long time ago [1]), so I think maybe for other types of map, the atomic update doesn't matter too much.
A thread from 3 years ago ?! Sorry, it's not helpful to ask people to page-in such an old context with lots of follow ups that may or may not be relevant today.
Let me reuse part of the diagram above to explain how does the lookup procedure may return a incorrect value:
lookup procedure A
A: find the old element (instead of the new element)update procedure B
B: delete the old element
update procedure C on the same CPU:Quoted 6 lines
A: the key is matched and return the value in the element
C: reuse the old element (overwrite its key and value)
A: read the value (it is incorrect, because it has been reused and overwritten)
... and it's fine. It's by design. It's an element reuse behavior.
Long ago hashmap had two modes: prealloc (default) and
NO_PREALLOC(call_rcu+ kfree)The
call_rcupart was there to make things safe.
The memory cannot be kfree-ed to the kernel until RCU GP. Withbpf_mem_allochashmap elements are freed back tobpf_maright away. Hashmap is doingbpf_mem_cache_free()
(instead ofbpf_mem_cache_free_rcu()) because users need speed. So since 2022 both prealloc andno_preallocreuse elements. We can consider a new flag for the hash map likeF_REUSE_AFTER_RCU_GPthat will use_rcu()flavor of freeing intobpf_ma,
but it has to have a strong reason.
And soon as we add it the default with prealloc would need to usecall_rcu()too, right?
and that becomes nightmare, since bpf prog can easily DoS the system. Even if we usebpf_mem_cache_free_rcu()only, the DoS is a concern. Unlike new things like bpf_obj_new/obj_drop the hashmap
is unpriv, so concerns are drastically different.
Quoted 1 line
.
I see. Thanks for the detailed explanation. And that is part of the reason why I proposed adding a seq-counter in the htab-element and checking the seq-counter during lookup, so at least the lookup will not return -ENOENT for the concurrent overwriting procedure.
Re: [RESEND] [PATCH bpf-next 2/3] bpf: Overwrite the element in hash map atomically
In reply to Alexei StarovoitovQuoted 96 lines · Alexei Starovoitov
Quoted 11 lines · Hou Tao
Quoted 11 lines
Quoted 11 lines
lookup procedure A
A: find the old element (instead of the new old)update procedure B
B: delete the old element
update procedure C on the same CPU:
C: reuse the old element (overwrite its key and insert in
the same bucket)Quoted 6 lines
Quoted 3 lines
Quoted 1 line
A: the key is mismatched and return
-ENOENT.This is fine. It's just normal reuse.
Orthogonal to 'update as insert+delete' issue.OK. However, it will break the lookup procedure because it expects it will return an valid result instead of
-ENOENT.What do you mean 'breaks the lookup' ?
lookup_elem_raw()matches hash, and then it memcmp(key), if the element is reused anything can happen. Either it succeeds inmemcmp()and returns an elem, or miscompares inmemcmp().
Both are expected, because elems are reused in place.And this behavior is expected and not-broken, because bpf prog that does lookup on one cpu and deletes the same element on the other cpu is asking for trouble. bpf infra guarantees the safety of the kernel. It doesn't guarantee that bpf progs are sane.
Quoted 7 lines
Quoted 4 lines
It's been a long time since I looked into
rcu_nullsdetails. Pls help me understand that this newreplace_rcu_nulls()
is correct from nulls pov,
If it is then this patch set may be the right answer to non-atomic update.If I understand correctly, only the manipulations of ->first pointer and ->next pointer need to take care of nulls pointer.
hmm. I feel we're still talking past each other.
See if (get_nulls_value()== ...) inlookup_nulls_elem_raw(). It's there because of reuse. The lookup can start in one bucket and finish in another.Quoted 39 lines
Quoted 9 lines
And for the future, please please focus on "why" part in the cover letter and commit logs instead of "what".
Since the only thing I got from the log was:
"Currently, the update is not atomic
because the overwrite of existing element happens in a two-steps way, but the support of atomic update is feasible".Quoted 6 lines
"is feasible" doesn't explain "why".
Link to xdp-newbie question is nice for additional context, but reviewers should not need to go and read some thread somewhere to understand "why" part.
All of it should be in the commit log.OK. My original thought is that is a reported problem, so an extra link will be enough. Will try to add more context next time.
Quoted 9 lines
Quoted 6 lines
map may still be incorrect (as shown long time ago [1]), so I think maybe for other types of map, the atomic update doesn't matter too much.
A thread from 3 years ago ?! Sorry, it's not helpful to ask people to page-in such an old context with lots of follow ups that may or may not be relevant today.
Let me reuse part of the diagram above to explain how does the lookup procedure may return a incorrect value:
lookup procedure A
A: find the old element (instead of the new element)update procedure B
B: delete the old element
update procedure C on the same CPU:Quoted 6 lines
A: the key is matched and return the value in the element
C: reuse the old element (overwrite its key and value)
A: read the value (it is incorrect, because it has been reused and overwritten)
... and it's fine. It's by design. It's an element reuse behavior.
Quoted 7 lines
Long ago hashmap had two modes: prealloc (default) and
NO_PREALLOC(call_rcu+ kfree)The
call_rcupart was there to make things safe. The memory cannot be kfree-ed to the kernel until RCU GP. Withbpf_mem_allochashmap elements are freed back tobpf_maright away. Hashmap is doingbpf_mem_cache_free()
We (emnify.com) missed this change and kept writing code with an assumption that NO_PREALLOC implies rcu.
Is there something we can do as of today to reduce the likelihood of an item getting reused immediately? We are concerned with lookups yielding bogus results when racing with updates. Worse, a program could corrupt an unrelated item when writing via a pointer obtained from lookup.
You wrote that "lookup on one cpu and deletes the same element on the other cpu is asking for trouble.” It puzzles me since user space updating a map while (say) TC program is consulting the map to make a routing decision look like a supported and widespread use case.
For us, implications vary in severity, e.g.:
- 1-in-1e? packet mis-delivered (bogus lookup: LOW)
- a tenant getting billed for a packet of another tenant delivered via satellite and costing USD 0.10 (writing into unrelated item: LOW)
- a network flow state corrupted (writing into unrelated item: MEDIUM)
We need to audit our code to ensure that e.g. a flow state getting corrupted self-corrects and the damage doesn’t spread.
It would be nice if we as eBPF users could decide whether we wish to live dangerously or prefer to trade speed for safety, on a case-by-case basis.
Quoted 11 lines
(instead of
bpf_mem_cache_free_rcu()) because users need speed. So since 2022 both prealloc andno_preallocreuse elements. We can consider a new flag for the hash map likeF_REUSE_AFTER_RCU_GPthat will use_rcu()flavor of freeing intobpf_ma,
but it has to have a strong reason.
And soon as we add it the default with prealloc would need to usecall_rcu()too, right?
and that becomes nightmare, since bpf prog can easily DoS the system. Even if we usebpf_mem_cache_free_rcu()only, the DoS is a concern. Unlike new things like bpf_obj_new/obj_drop the hashmap
is unpriv, so concerns are drastically different.
Would it be an option to gate F_REUSE_AFTER_RCU under CAP_BPF?
It looks like sleepable programs would need to bpf_rcu_read_lock explicitly to reap the benefits, but why not.