Skip to contentSkip to conversation

Add support for memory-mapping BPF array maps

PATCH Superseded Andrii Nakryiko 8 participants 25 messages Nov 15, 2019 – Nov 18, 2019

Subsystems (11)

v4 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

v4 cover Add support for memory-mapping BPF array maps

This patch set adds ability to memory-map BPF array maps (single- and multi-element). The primary use case is memory-mapping BPF array maps, created to back global data variables, created by libbpf implicitly. This allows for much better usability, along with avoiding syscalls to read or update data completely.

Due to memory-mapping requirements, BPF array map that is supposed to be memory-mapped, has to be created with special BPF_F_MMAPABLE attribute, which triggers slightly different memory allocation strategy internally. See patch 1 for details.

Libbpf is extended to detect kernel support for this flag, and if supported, will specify it for all global data maps automatically.

Patch #1 refactors bpf_map_inc() and converts bpf_map's refcnt to atomic64_t to make refcounting never fail.

v3->v4:
- add mmap's open() callback to fix refcounting (Johannes);
- switch to remap_vmalloc_pages() instead of custom fault handler (Johannes);
- converted bpf_map's refcnt/usercnt into atomic64_t;
- provide default bpf_map_default_vmops handling open/close properly;

v2->v3:
- change allocation strategy to avoid extra pointer dereference (Jakub);

v1->v2:
- fix map lookup code generation for BPF_F_MMAPABLE case;
- prevent BPF_F_MMAPABLE flag for all but plain array map type;
- centralize ref-counting in generic bpf_map_mmap();
- don't use uref counting (Alexei);
- use vfree() directly;
- print flags with %x (Song);
- extend tests to verify bpf_map_{lookup,update}_elem() logic as well.

Andrii Nakryiko (4):
bpf: switch bpf_map ref counter to 64bit so bpf_map_inc never fails
bpf: add mmap() support for BPF_MAP_TYPE_ARRAY
libbpf: make global data internal arrays mmap()-able, if possible selftests/bpf: add BPF_TYPE_MAP_ARRAY mmap() tests

 .../net/ethernet/netronome/nfp/bpf/offload.c  |   4 +-
 include/linux/bpf.h                           |  21 +-
 include/linux/vmalloc.h                       |   1 +
 include/uapi/linux/bpf.h                      |   3 +
 kernel/bpf/arraymap.c                         |  59 ++++-
 kernel/bpf/inode.c                            |   2 +-
 kernel/bpf/map_in_map.c                       |   2 +-
 kernel/bpf/syscall.c                          | 150 +++++++++---
 kernel/bpf/verifier.c                         |   6 +-
 kernel/bpf/xskmap.c                           |   6 +-
 mm/vmalloc.c                                  |  20 ++
 net/core/bpf_sk_storage.c                     |   2 +-
 tools/include/uapi/linux/bpf.h                |   3 +
 tools/lib/bpf/libbpf.c                        |  32 ++-
 .../selftests/bpf/prog_tests/core_reloc.c     |  45 ++--
 tools/testing/selftests/bpf/prog_tests/mmap.c | 220 ++++++++++++++++++
 tools/testing/selftests/bpf/progs/test_mmap.c |  45 ++++
 17 files changed, 540 insertions(+), 81 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/mmap.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_mmap.c
Signature · 1 line
2.17.1
#1 lore raw

PATCH v4 1/4 bpf: switch bpf_map ref counter to 64bit so bpf_map_inc never fails

92117d8443bc ("bpf: fix refcnt overflow") turned refcounting of bpf_map into potentially failing operation, when refcount reaches BPF_MAX_REFCNT limit (32k). Due to using 32-bit counter, it's possible in practice to overflow refcounter and make it wrap around to 0, causing erroneous map free, while there are still references to it, causing use-after-free problems.

But having a failing refcounting operations are problematic in some cases. One example is mmap() interface. After establishing initial memory-mapping, user is allowed to arbitrarily map/remap/unmap parts of mapped memory, arbitrarily splitting it into multiple non-contiguous regions. All this happening without any control from the users of mmap subsystem. Rather mmap subsystem sends notifications to original creator of memory mapping through open/close callbacks, which are optionally specified during initial memory mapping creation. These callbacks are used to maintain accurate refcount for bpf_map (see next patch in this series). The problem is that open() callback is not supposed to fail, because memory-mapped resource is set up and properly referenced. This is posing a problem for using memory-mapping with BPF maps.

One solution to this is to maintain separate refcount for just memory-mappings and do single bpf_map_inc/bpf_map_put when it goes from/to zero, respectively. There are similar use cases in current work on tcp-bpf, necessitating extra counter as well. This seems like a rather unfortunate and ugly solution that doesn't scale well to various new use cases.

Another approach to solve this is to use non-failing refcount_t type, which uses 32-bit counter internally, but, once reaching overflow state at UINT_MAX, stays there. This utlimately causes memory leak, but prevents use after free.

But given refcounting is not the most performance-critical operation with BPF maps (it's not used from running BPF program code), we can also just switch to 64-bit counter that can't overflow in practice, potentially disadvantaging 32-bit platforms a tiny bit. This simplifies semantics and allows above described scenarios to not worry about failing refcount increment operation.

In terms of struct bpf_map size, we are still good and use the same amount of space:

BEFORE (3 cache lines, 8 bytes of padding at the end):
struct bpf_map {
const struct bpf_map_ops * ops __attribute__((__aligned__(64))); /* 0 8 */
struct bpf_map * inner_map_meta; /* 8 8 */
void * security; /* 16 8 */
enum bpf_map_type map_type; /* 24 4 */
u32 key_size; /* 28 4 */
u32 value_size; /* 32 4 */
u32 max_entries; /* 36 4 */
u32 map_flags; /* 40 4 */
int spin_lock_off; /* 44 4 */
u32 id; /* 48 4 */
int numa_node; /* 52 4 */
u32 btf_key_type_id; /* 56 4 */
u32 btf_value_type_id; /* 60 4 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct btf * btf; /* 64 8 */
struct bpf_map_memory memory; /* 72 16 */
bool unpriv_array; /* 88 1 */
bool frozen; /* 89 1 */

/* XXX 38 bytes hole, try to pack */

/* --- cacheline 2 boundary (128 bytes) --- */
atomic_t refcnt __attribute__((__aligned__(64))); /* 128 4 */
atomic_t usercnt; /* 132 4 */
struct work_struct work; /* 136 32 */
char name[16]; /* 168 16 */

/* size: 192, cachelines: 3, members: 21 */
/* sum members: 146, holes: 1, sum holes: 38 */
/* padding: 8 */
/* forced alignments: 2, forced holes: 1, sum forced holes: 38 */
} __attribute__((__aligned__(64)));

AFTER (same 3 cache lines, no extra padding now):
struct bpf_map {
const struct bpf_map_ops * ops __attribute__((__aligned__(64))); /* 0 8 */
struct bpf_map * inner_map_meta; /* 8 8 */
void * security; /* 16 8 */
enum bpf_map_type map_type; /* 24 4 */
u32 key_size; /* 28 4 */
u32 value_size; /* 32 4 */
u32 max_entries; /* 36 4 */
u32 map_flags; /* 40 4 */
int spin_lock_off; /* 44 4 */
u32 id; /* 48 4 */
int numa_node; /* 52 4 */
u32 btf_key_type_id; /* 56 4 */
u32 btf_value_type_id; /* 60 4 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct btf * btf; /* 64 8 */
struct bpf_map_memory memory; /* 72 16 */
bool unpriv_array; /* 88 1 */
bool frozen; /* 89 1 */

/* XXX 38 bytes hole, try to pack */

/* --- cacheline 2 boundary (128 bytes) --- */
atomic64_t refcnt __attribute__((__aligned__(64))); /* 128 8 */
atomic64_t usercnt; /* 136 8 */
struct work_struct work; /* 144 32 */
char name[16]; /* 176 16 */

/* size: 192, cachelines: 3, members: 21 */
/* sum members: 154, holes: 1, sum holes: 38 */
/* forced alignments: 2, forced holes: 1, sum forced holes: 38 */
} __attribute__((__aligned__(64)));

This patch, while modifying all users of bpf_map_inc, also cleans up its interface to match bpf_map_put with separate operations for bpf_map_inc and bpf_map_inc_with_uref (to match bpf_map_put and bpf_map_put_with_uref, respectively). Also, given there are no users of bpf_map_inc_not_zero specifying uref=true, remove uref flag and default to uref=false internally.

---
 .../net/ethernet/netronome/nfp/bpf/offload.c  |  4 +-
 include/linux/bpf.h                           | 10 ++--
 kernel/bpf/inode.c                            |  2 +-
 kernel/bpf/map_in_map.c                       |  2 +-
 kernel/bpf/syscall.c                          | 51 ++++++++-----------
 kernel/bpf/verifier.c                         |  6 +--
 kernel/bpf/xskmap.c                           |  6 +--
 net/core/bpf_sk_storage.c                     |  2 +-
 8 files changed, 34 insertions(+), 49 deletions(-)
drivers/net/ethernet/netronome/nfp/bpf/offload.c +1 -3
index 88fab6a82acf..06927ba5a3ae 100644
@@ -46,9 +46,7 @@ nfp_map_ptr_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog,
 	/* Grab a single ref to the map for our record.  The prog destroy ndo
 	 * happens after free_used_maps().
 	 */
-	map = bpf_map_inc(map, false);
-	if (IS_ERR(map))
-		return PTR_ERR(map);
+	bpf_map_inc(map);
 
 	record = kmalloc(sizeof(*record), GFP_KERNEL);
 	if (!record) {
include/linux/bpf.h +5 -5
index 7c7f518811a6..6fbe599fb977 100644
@@ -101,8 +101,8 @@ struct bpf_map {
 	/* The 3rd and 4th cacheline with misc members to avoid false sharing
 	 * particularly with refcounting.
 	 */
-	atomic_t refcnt ____cacheline_aligned;
-	atomic_t usercnt;
+	atomic64_t refcnt ____cacheline_aligned;
+	atomic64_t usercnt;
 	struct work_struct work;
 	char name[BPF_OBJ_NAME_LEN];
 };
@@ -677,9 +677,9 @@ void bpf_map_free_id(struct bpf_map *map, bool do_idr_lock);
 
 struct bpf_map *bpf_map_get_with_uref(u32 ufd);
 struct bpf_map *__bpf_map_get(struct fd f);
-struct bpf_map * __must_check bpf_map_inc(struct bpf_map *map, bool uref);
-struct bpf_map * __must_check bpf_map_inc_not_zero(struct bpf_map *map,
-						   bool uref);
+void bpf_map_inc(struct bpf_map *map);
+void bpf_map_inc_with_uref(struct bpf_map *map);
+struct bpf_map * __must_check bpf_map_inc_not_zero(struct bpf_map *map);
 void bpf_map_put_with_uref(struct bpf_map *map);
 void bpf_map_put(struct bpf_map *map);
 int bpf_map_charge_memlock(struct bpf_map *map, u32 pages);
kernel/bpf/inode.c +1 -1
index a70f7209cda3..2f17f24258dc 100644
@@ -34,7 +34,7 @@ static void *bpf_any_get(void *raw, enum bpf_type type)
 		raw = bpf_prog_inc(raw);
 		break;
 	case BPF_TYPE_MAP:
-		raw = bpf_map_inc(raw, true);
+		bpf_map_inc_with_uref(raw);
 		break;
 	default:
 		WARN_ON_ONCE(1);
kernel/bpf/map_in_map.c +1 -1
index fab4fb134547..4cbe987be35b 100644
@@ -98,7 +98,7 @@ void *bpf_map_fd_get_ptr(struct bpf_map *map,
 		return inner_map;
 
 	if (bpf_map_meta_equal(map->inner_map_meta, inner_map))
-		inner_map = bpf_map_inc(inner_map, false);
+		bpf_map_inc(inner_map);
 	else
 		inner_map = ERR_PTR(-EINVAL);
 
kernel/bpf/syscall.c +22 -29

109 lines not on this page · raw message

kernel/bpf/verifier.c +1 -5
index 2f2374967b36..38d34afe979c 100644
@@ -8169,11 +8169,7 @@ static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
 			 * will be used by the valid program until it's unloaded
 			 * and all maps are released in free_used_maps()
 			 */
-			map = bpf_map_inc(map, false);
-			if (IS_ERR(map)) {
-				fdput(f);
-				return PTR_ERR(map);
-			}
+			bpf_map_inc(map);
 
 			aux->map_index = env->used_map_cnt;
 			env->used_maps[env->used_map_cnt++] = map;
kernel/bpf/xskmap.c +2 -4
index da16c30868f3..90c4fce1c981 100644
@@ -11,10 +11,8 @@
 
 int xsk_map_inc(struct xsk_map *map)
 {
-	struct bpf_map *m = &map->map;
-
-	m = bpf_map_inc(m, false);
-	return PTR_ERR_OR_ZERO(m);
+	bpf_map_inc(&map->map);
+	return 0;
 }
 
 void xsk_map_put(struct xsk_map *map)
net/core/bpf_sk_storage.c +1 -1
index da5639a5bd3b..458be6b3eda9 100644
@@ -798,7 +798,7 @@ int bpf_sk_storage_clone(const struct sock *sk, struct sock *newsk)
 		 * Try to grab map refcnt to make sure that it's still
 		 * alive and prevent concurrent removal.
 		 */
-		map = bpf_map_inc_not_zero(&smap->map, false);
+		map = bpf_map_inc_not_zero(&smap->map);
 		if (IS_ERR(map))
 			continue;
 
Signature · 1 line
2.17.1
Acked-by on v4 1/4
#2 lore raw

Re: [PATCH v4 bpf-next 1/4] bpf: switch bpf_map ref counter to 64bit so bpf_map_inc never fails

Quoted 4 lines · Andrii Nakryiko

[...]

Quoted 8 lines

This patch, while modifying all users of bpf_map_inc, also cleans up its interface to match bpf_map_put with separate operations for bpf_map_inc and bpf_map_inc_with_uref (to match bpf_map_put and bpf_map_put_with_uref, respectively). Also, given there are no users of bpf_map_inc_not_zero specifying uref=true, remove uref flag and default to uref=false internally.

maintainer
#3 lore raw

Re: [PATCH v4 bpf-next 1/4] bpf: switch bpf_map ref counter to 64bit so bpf_map_inc never fails

In reply to Andrii Nakryiko
Quoted 5 lines · Andrii Nakryiko

92117d8443bc ("bpf: fix refcnt overflow") turned refcounting of bpf_map into potentially failing operation, when refcount reaches BPF_MAX_REFCNT limit (32k). Due to using 32-bit counter, it's possible in practice to overflow refcounter and make it wrap around to 0, causing erroneous map free, while there are still references to it, causing use-after-free problems.

You mention 'it's possible in practice' to overflow in relation to bpf map refcount, do we need any fixing for bpf tree here?

Quoted 118 lines

But having a failing refcounting operations are problematic in some cases. One example is mmap() interface. After establishing initial memory-mapping, user is allowed to arbitrarily map/remap/unmap parts of mapped memory, arbitrarily splitting it into multiple non-contiguous regions. All this happening without any control from the users of mmap subsystem. Rather mmap subsystem sends notifications to original creator of memory mapping through open/close callbacks, which are optionally specified during initial memory mapping creation. These callbacks are used to maintain accurate refcount for bpf_map (see next patch in this series). The problem is that open() callback is not supposed to fail, because memory-mapped resource is set up and properly referenced. This is posing a problem for using memory-mapping with BPF maps.

One solution to this is to maintain separate refcount for just memory-mappings and do single bpf_map_inc/bpf_map_put when it goes from/to zero, respectively. There are similar use cases in current work on tcp-bpf, necessitating extra counter as well. This seems like a rather unfortunate and ugly solution that doesn't scale well to various new use cases.

Another approach to solve this is to use non-failing refcount_t type, which uses 32-bit counter internally, but, once reaching overflow state at UINT_MAX, stays there. This utlimately causes memory leak, but prevents use after free.

But given refcounting is not the most performance-critical operation with BPF maps (it's not used from running BPF program code), we can also just switch to 64-bit counter that can't overflow in practice, potentially disadvantaging 32-bit platforms a tiny bit. This simplifies semantics and allows above described scenarios to not worry about failing refcount increment operation.

In terms of struct bpf_map size, we are still good and use the same amount of space:

BEFORE (3 cache lines, 8 bytes of padding at the end):
struct bpf_map {
const struct bpf_map_ops * ops __attribute__((__aligned__(64))); /* 0 8 */
struct bpf_map * inner_map_meta; /* 8 8 */
void * security; /* 16 8 */
enum bpf_map_type map_type; /* 24 4 */
u32 key_size; /* 28 4 */
u32 value_size; /* 32 4 */
u32 max_entries; /* 36 4 */
u32 map_flags; /* 40 4 */
int spin_lock_off; /* 44 4 */
u32 id; /* 48 4 */
int numa_node; /* 52 4 */
u32 btf_key_type_id; /* 56 4 */
u32 btf_value_type_id; /* 60 4 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct btf * btf; /* 64 8 */
struct bpf_map_memory memory; /* 72 16 */
bool unpriv_array; /* 88 1 */
bool frozen; /* 89 1 */

/* XXX 38 bytes hole, try to pack */

/* --- cacheline 2 boundary (128 bytes) --- */
atomic_t refcnt __attribute__((__aligned__(64))); /* 128 4 */
atomic_t usercnt; /* 132 4 */
struct work_struct work; /* 136 32 */
char name[16]; /* 168 16 */

/* size: 192, cachelines: 3, members: 21 */
/* sum members: 146, holes: 1, sum holes: 38 */
/* padding: 8 */
/* forced alignments: 2, forced holes: 1, sum forced holes: 38 */
} __attribute__((__aligned__(64)));

AFTER (same 3 cache lines, no extra padding now):
struct bpf_map {
const struct bpf_map_ops * ops __attribute__((__aligned__(64))); /* 0 8 */
struct bpf_map * inner_map_meta; /* 8 8 */
void * security; /* 16 8 */
enum bpf_map_type map_type; /* 24 4 */
u32 key_size; /* 28 4 */
u32 value_size; /* 32 4 */
u32 max_entries; /* 36 4 */
u32 map_flags; /* 40 4 */
int spin_lock_off; /* 44 4 */
u32 id; /* 48 4 */
int numa_node; /* 52 4 */
u32 btf_key_type_id; /* 56 4 */
u32 btf_value_type_id; /* 60 4 */
/* --- cacheline 1 boundary (64 bytes) --- */
struct btf * btf; /* 64 8 */
struct bpf_map_memory memory; /* 72 16 */
bool unpriv_array; /* 88 1 */
bool frozen; /* 89 1 */

/* XXX 38 bytes hole, try to pack */

/* --- cacheline 2 boundary (128 bytes) --- */
atomic64_t refcnt __attribute__((__aligned__(64))); /* 128 8 */
atomic64_t usercnt; /* 136 8 */
struct work_struct work; /* 144 32 */
char name[16]; /* 176 16 */

/* size: 192, cachelines: 3, members: 21 */
/* sum members: 154, holes: 1, sum holes: 38 */
/* forced alignments: 2, forced holes: 1, sum forced holes: 38 */
} __attribute__((__aligned__(64)));

This patch, while modifying all users of bpf_map_inc, also cleans up its interface to match bpf_map_put with separate operations for bpf_map_inc and bpf_map_inc_with_uref (to match bpf_map_put and bpf_map_put_with_uref, respectively). Also, given there are no users of bpf_map_inc_not_zero specifying uref=true, remove uref flag and default to uref=false internally.

---
  .../net/ethernet/netronome/nfp/bpf/offload.c  |  4 +-
  include/linux/bpf.h                           | 10 ++--
  kernel/bpf/inode.c                            |  2 +-
  kernel/bpf/map_in_map.c                       |  2 +-
  kernel/bpf/syscall.c                          | 51 ++++++++-----------
  kernel/bpf/verifier.c                         |  6 +--
  kernel/bpf/xskmap.c                           |  6 +--
  net/core/bpf_sk_storage.c                     |  2 +-
  8 files changed, 34 insertions(+), 49 deletions(-)

[...]

Quoted 21 lines
kernel/bpf/syscall.c +4 -4
+	refold = atomic64_fetch_add_unless(&map->refcnt, 1, 0);
  	if (!refold)
  		return ERR_PTR(-ENOENT);
-
  	if (uref)
-		atomic_inc(&map->usercnt);
+		atomic64_inc(&map->usercnt);
 
  	return map;
  }
 
-struct bpf_map *bpf_map_inc_not_zero(struct bpf_map *map, bool uref)
+struct bpf_map *bpf_map_inc_not_zero(struct bpf_map *map)
  {
  	spin_lock_bh(&map_idr_lock);
-	map = __bpf_map_inc_not_zero(map, uref);
+	map = __bpf_map_inc_not_zero(map, false);
  	spin_unlock_bh(&map_idr_lock);
 
  	return map;

kernel/bpf/syscall.c static struct bpf_prog *____bpf_prog_get(struct fd f) · line 1449

Quoted 6 lines
kernel/bpf/syscall.c +3 -0
@@ -1454,6 +1444,9 @@ static struct bpf_prog *____bpf_prog_get(struct fd f)
  	return f.file->private_data;
  }
 
+/* prog's refcnt limit */
+#define BPF_MAX_REFCNT 32768
+

Would it make sense in this context to also convert prog refcount over to atomic64 so we have both in one go in order to align them again? This could likely simplify even more wrt error paths.

Quoted 3 lines
kernel/bpf/syscall.c +0 -0
  struct bpf_prog *bpf_prog_add(struct bpf_prog *prog, int i)
  {
  	if (atomic_add_return(i, &prog->aux->refcnt) > BPF_MAX_REFCNT) {
#4 lore raw

Re: [PATCH v4 bpf-next 1/4] bpf: switch bpf_map ref counter to 64bit so bpf_map_inc never fails

Quoted 10 lines · Daniel Borkmann
Quoted 5 lines · Andrii Nakryiko

92117d8443bc ("bpf: fix refcnt overflow") turned refcounting of bpf_map into potentially failing operation, when refcount reaches BPF_MAX_REFCNT limit (32k). Due to using 32-bit counter, it's possible in practice to overflow refcounter and make it wrap around to 0, causing erroneous map free, while there are still references to it, causing use-after-free problems.

You mention 'it's possible in practice' to overflow in relation to bpf map refcount, do we need any fixing for bpf tree here?

I meant without existing 32k limit. With limit we currently have, no, we'll just fail bpf_map_inc instead. So no need to do anything about bpf tree.

Quoted 23 lines
Quoted 23 lines

But having a failing refcounting operations are problematic in some cases. One example is mmap() interface. After establishing initial memory-mapping, user is allowed to arbitrarily map/remap/unmap parts of mapped memory, arbitrarily splitting it into multiple non-contiguous regions. All this happening without any control from the users of mmap subsystem. Rather mmap subsystem sends notifications to original creator of memory mapping through open/close callbacks, which are optionally specified during initial memory mapping creation. These callbacks are used to maintain accurate refcount for bpf_map (see next patch in this series). The problem is that open() callback is not supposed to fail, because memory-mapped resource is set up and properly referenced. This is posing a problem for using memory-mapping with BPF maps.

One solution to this is to maintain separate refcount for just memory-mappings and do single bpf_map_inc/bpf_map_put when it goes from/to zero, respectively. There are similar use cases in current work on tcp-bpf, necessitating extra counter as well. This seems like a rather unfortunate and ugly solution that doesn't scale well to various new use cases.

Another approach to solve this is to use non-failing refcount_t type, which uses 32-bit counter internally, but, once reaching overflow state at UINT_MAX, stays there. This utlimately causes memory leak, but prevents use after free.

[...]

Quoted 9 lines

kernel/bpf/syscall.c static struct bpf_prog *____bpf_prog_get(struct fd f) · line 1449

Quoted 4 lines
kernel/bpf/syscall.c +3 -0
+/* prog's refcnt limit */
+#define BPF_MAX_REFCNT 32768
+

Would it make sense in this context to also convert prog refcount over to atomic64 so we have both in one go in order to align them again? This could likely simplify even more wrt error paths.

Yes, of course, I was just trying to minimize amount of changes, but I can go and do the same for prog refcnt.

Quoted 3 lines
Quoted 3 lines
  struct bpf_prog *bpf_prog_add(struct bpf_prog *prog, int i)
  {
      if (atomic_add_return(i, &prog->aux->refcnt) > BPF_MAX_REFCNT) {
#5 lore raw

PATCH v4 2/4 bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

In reply to Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

---
 include/linux/bpf.h            | 11 ++--
 include/linux/vmalloc.h        |  1 +
 include/uapi/linux/bpf.h       |  3 ++
 kernel/bpf/arraymap.c          | 59 +++++++++++++++++---
 kernel/bpf/syscall.c           | 99 ++++++++++++++++++++++++++++++++--
 mm/vmalloc.c                   | 20 +++++++
 tools/include/uapi/linux/bpf.h |  3 ++
 7 files changed, 184 insertions(+), 12 deletions(-)
include/linux/bpf.h +8 -3
index 6fbe599fb977..8021fce98868 100644
@@ -12,6 +12,7 @@
 #include <linux/err.h>
 #include <linux/rbtree_latch.h>
 #include <linux/numa.h>
+#include <linux/mm_types.h>
 #include <linux/wait.h>
 #include <linux/u64_stats_sync.h>
 
@@ -66,6 +67,7 @@ struct bpf_map_ops {
 				     u64 *imm, u32 off);
 	int (*map_direct_value_meta)(const struct bpf_map *map,
 				     u64 imm, u32 *off);
+	int (*map_mmap)(struct bpf_map *map, struct vm_area_struct *vma);
 };
 
 struct bpf_map_memory {
@@ -94,9 +96,10 @@ struct bpf_map {
 	u32 btf_value_type_id;
 	struct btf *btf;
 	struct bpf_map_memory memory;
+	char name[BPF_OBJ_NAME_LEN];
 	bool unpriv_array;
-	bool frozen; /* write-once */
-	/* 48 bytes hole */
+	bool frozen; /* write-once; write-protected by freeze_mutex */
+	/* 22 bytes hole */
 
 	/* The 3rd and 4th cacheline with misc members to avoid false sharing
 	 * particularly with refcounting.
@@ -104,7 +107,8 @@ struct bpf_map {
 	atomic64_t refcnt ____cacheline_aligned;
 	atomic64_t usercnt;
 	struct work_struct work;
-	char name[BPF_OBJ_NAME_LEN];
+	struct mutex freeze_mutex;
+	u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
 };
 
 static inline bool map_value_has_spin_lock(const struct bpf_map *map)
@@ -689,6 +693,7 @@ void bpf_map_charge_finish(struct bpf_map_memory *mem);
 void bpf_map_charge_move(struct bpf_map_memory *dst,
 			 struct bpf_map_memory *src);
 void *bpf_map_area_alloc(size_t size, int numa_node);
+void *bpf_map_area_mmapable_alloc(size_t size, int numa_node);
 void bpf_map_area_free(void *base);
 void bpf_map_init_from_attr(struct bpf_map *map, union bpf_attr *attr);
 
include/linux/vmalloc.h +1 -0
index 4e7809408073..b4c58a191eb1 100644
@@ -93,6 +93,7 @@ extern void *vzalloc(unsigned long size);
 extern void *vmalloc_user(unsigned long size);
 extern void *vmalloc_node(unsigned long size, int node);
 extern void *vzalloc_node(unsigned long size, int node);
+extern void *vmalloc_user_node_flags(unsigned long size, int node, gfp_t flags);
 extern void *vmalloc_exec(unsigned long size);
 extern void *vmalloc_32(unsigned long size);
 extern void *vmalloc_32_user(unsigned long size);
include/uapi/linux/bpf.h +3 -0
index df6809a76404..bb39b53622d9 100644
@@ -346,6 +346,9 @@ enum bpf_attach_type {
 /* Clone map from listener for newly accepted socket */
 #define BPF_F_CLONE		(1U << 9)
 
+/* Enable memory-mapping BPF map */
+#define BPF_F_MMAPABLE		(1U << 10)
+
 /* flags for BPF_PROG_QUERY */
 #define BPF_F_QUERY_EFFECTIVE	(1U << 0)
 
kernel/bpf/arraymap.c +53 -6

108 lines not on this page · raw message

kernel/bpf/syscall.c +96 -3

151 lines not on this page · raw message

mm/vmalloc.c +20 -0
index a3c70e275f4e..4a7d7459c4f9 100644
@@ -2671,6 +2671,26 @@ void *vzalloc_node(unsigned long size, int node)
 }
 EXPORT_SYMBOL(vzalloc_node);
 
+/**
+ * vmalloc_user_node_flags - allocate memory for userspace on a specific node
+ * @size: allocation size
+ * @node: numa node
+ * @flags: flags for the page level allocator
+ *
+ * The resulting memory area is zeroed so it can be mapped to userspace
+ * without leaking data.
+ *
+ * Return: pointer to the allocated memory or %NULL on error
+ */
+void *vmalloc_user_node_flags(unsigned long size, int node, gfp_t flags)
+{
+	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
+				    flags | __GFP_ZERO, PAGE_KERNEL,
+				    VM_USERMAP, node,
+				    __builtin_return_address(0));
+}
+EXPORT_SYMBOL(vmalloc_user_node_flags);
+
 /**
  * vmalloc_exec - allocate virtually contiguous, executable memory
  * @size:	  allocation size
tools/include/uapi/linux/bpf.h +3 -0
index df6809a76404..bb39b53622d9 100644
@@ -346,6 +346,9 @@ enum bpf_attach_type {
 /* Clone map from listener for newly accepted socket */
 #define BPF_F_CLONE		(1U << 9)
 
+/* Enable memory-mapping BPF map */
+#define BPF_F_MMAPABLE		(1U << 10)
+
 /* flags for BPF_PROG_QUERY */
 #define BPF_F_QUERY_EFFECTIVE	(1U << 0)
 
Signature · 1 line
2.17.1
#6 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 82 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

---
 include/linux/bpf.h            | 11 ++--
 include/linux/vmalloc.h        |  1 +
 include/uapi/linux/bpf.h       |  3 ++
 kernel/bpf/arraymap.c          | 59 +++++++++++++++++---
 kernel/bpf/syscall.c           | 99 ++++++++++++++++++++++++++++++++--
 mm/vmalloc.c                   | 20 +++++++
 tools/include/uapi/linux/bpf.h |  3 ++
 7 files changed, 184 insertions(+), 12 deletions(-)
include/linux/bpf.h +5 -2
index 6fbe599fb977..8021fce98868 100644
@@ -12,6 +12,7 @@
 #include <linux/err.h>
 #include <linux/rbtree_latch.h>
 #include <linux/numa.h>
+#include <linux/mm_types.h>
 #include <linux/wait.h>
 #include <linux/u64_stats_sync.h>
 
@@ -66,6 +67,7 @@ struct bpf_map_ops {
 				     u64 *imm, u32 off);
 	int (*map_direct_value_meta)(const struct bpf_map *map,
 				     u64 imm, u32 *off);
+	int (*map_mmap)(struct bpf_map *map, struct vm_area_struct *vma);
 };
 
 struct bpf_map_memory {
@@ -94,9 +96,10 @@ struct bpf_map {
 	u32 btf_value_type_id;
 	struct btf *btf;
 	struct bpf_map_memory memory;
+	char name[BPF_OBJ_NAME_LEN];
 	bool unpriv_array;
-	bool frozen; /* write-once */
-	/* 48 bytes hole */
+	bool frozen; /* write-once; write-protected by freeze_mutex */
+	/* 22 bytes hole */
 
 	/* The 3rd and 4th cacheline with misc members to avoid false sharing
 	 * particularly with refcounting.

include/linux/bpf.h struct bpf_map { · line 112

Quoted 7 lines
include/linux/bpf.h +2 -1
@@ -104,7 +107,8 @@ struct bpf_map {
 	atomic64_t refcnt ____cacheline_aligned;
 	atomic64_t usercnt;
 	struct work_struct work;
-	char name[BPF_OBJ_NAME_LEN];
+	struct mutex freeze_mutex;
+	u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
 };

Can the mutex be moved into bpf_array instead of being in bpf_map that is shared across all map types?
If so then can you reuse the mutex that Daniel is adding in his patch 6/8 of tail_call series? Not sure what would the right name for such mutex. It will be used for your map_freeze logic and for Daniel's text_poke.

#7 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 81 lines · Alexei Starovoitov
Quoted 79 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

---
 include/linux/bpf.h            | 11 ++--
 include/linux/vmalloc.h        |  1 +
 include/uapi/linux/bpf.h       |  3 ++
 kernel/bpf/arraymap.c          | 59 +++++++++++++++++---
 kernel/bpf/syscall.c           | 99 ++++++++++++++++++++++++++++++++--
 mm/vmalloc.c                   | 20 +++++++
 tools/include/uapi/linux/bpf.h |  3 ++
 7 files changed, 184 insertions(+), 12 deletions(-)
include/linux/bpf.h +5 -2
index 6fbe599fb977..8021fce98868 100644
@@ -12,6 +12,7 @@
 #include <linux/err.h>
 #include <linux/rbtree_latch.h>
 #include <linux/numa.h>
+#include <linux/mm_types.h>
 #include <linux/wait.h>
 #include <linux/u64_stats_sync.h>
 
@@ -66,6 +67,7 @@ struct bpf_map_ops {
                                   u64 *imm, u32 off);
      int (*map_direct_value_meta)(const struct bpf_map *map,
                                   u64 imm, u32 *off);
+     int (*map_mmap)(struct bpf_map *map, struct vm_area_struct *vma);
 };
 
 struct bpf_map_memory {
@@ -94,9 +96,10 @@ struct bpf_map {
      u32 btf_value_type_id;
      struct btf *btf;
      struct bpf_map_memory memory;
+     char name[BPF_OBJ_NAME_LEN];
      bool unpriv_array;
-     bool frozen; /* write-once */
-     /* 48 bytes hole */
+     bool frozen; /* write-once; write-protected by freeze_mutex */
+     /* 22 bytes hole */
 
Quoted 13 lines

include/linux/bpf.h struct bpf_map { · line 112

Quoted 10 lines
include/linux/bpf.h +2 -1
      /* The 3rd and 4th cacheline with misc members to avoid false sharing
       * particularly with refcounting.
@@ -104,7 +107,8 @@ struct bpf_map {
      atomic64_t refcnt ____cacheline_aligned;
      atomic64_t usercnt;
      struct work_struct work;
-     char name[BPF_OBJ_NAME_LEN];
+     struct mutex freeze_mutex;
+     u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
 };

Can the mutex be moved into bpf_array instead of being in bpf_map that is shared across all map types?

No, freezing logic is common to all maps. Same for writecnt and mmap()-ing overall.

Quoted 4 lines

If so then can you reuse the mutex that Daniel is adding in his patch 6/8 of tail_call series? Not sure what would the right name for such mutex. It will be used for your map_freeze logic and for Daniel's text_poke.

#8 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 97 lines · Andrii Nakryiko
Quoted 81 lines · Alexei Starovoitov
Quoted 79 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

---
 include/linux/bpf.h            | 11 ++--
 include/linux/vmalloc.h        |  1 +
 include/uapi/linux/bpf.h       |  3 ++
 kernel/bpf/arraymap.c          | 59 +++++++++++++++++---
 kernel/bpf/syscall.c           | 99 ++++++++++++++++++++++++++++++++--
 mm/vmalloc.c                   | 20 +++++++
 tools/include/uapi/linux/bpf.h |  3 ++
 7 files changed, 184 insertions(+), 12 deletions(-)
include/linux/bpf.h +5 -2
index 6fbe599fb977..8021fce98868 100644
@@ -12,6 +12,7 @@
 #include <linux/err.h>
 #include <linux/rbtree_latch.h>
 #include <linux/numa.h>
+#include <linux/mm_types.h>
 #include <linux/wait.h>
 #include <linux/u64_stats_sync.h>
 
@@ -66,6 +67,7 @@ struct bpf_map_ops {
                                   u64 *imm, u32 off);
      int (*map_direct_value_meta)(const struct bpf_map *map,
                                   u64 imm, u32 *off);
+     int (*map_mmap)(struct bpf_map *map, struct vm_area_struct *vma);
 };
 
 struct bpf_map_memory {
@@ -94,9 +96,10 @@ struct bpf_map {
      u32 btf_value_type_id;
      struct btf *btf;
      struct bpf_map_memory memory;
+     char name[BPF_OBJ_NAME_LEN];
      bool unpriv_array;
-     bool frozen; /* write-once */
-     /* 48 bytes hole */
+     bool frozen; /* write-once; write-protected by freeze_mutex */
+     /* 22 bytes hole */
 
Quoted 13 lines

include/linux/bpf.h struct bpf_map { · line 112

Quoted 10 lines
include/linux/bpf.h +2 -1
      /* The 3rd and 4th cacheline with misc members to avoid false sharing
       * particularly with refcounting.
@@ -104,7 +107,8 @@ struct bpf_map {
      atomic64_t refcnt ____cacheline_aligned;
      atomic64_t usercnt;
      struct work_struct work;
-     char name[BPF_OBJ_NAME_LEN];
+     struct mutex freeze_mutex;
+     u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
 };

Can the mutex be moved into bpf_array instead of being in bpf_map that is shared across all map types?

Quoted 2 lines

No, freezing logic is common to all maps. Same for writecnt and mmap()-ing overall.

How mmap is going to work for hash map ? and for prog_array?

#9 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 99 lines · Alexei Starovoitov
Quoted 97 lines · Andrii Nakryiko
Quoted 94 lines · Alexei Starovoitov
Quoted 82 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

---
 include/linux/bpf.h            | 11 ++--
 include/linux/vmalloc.h        |  1 +
 include/uapi/linux/bpf.h       |  3 ++
 kernel/bpf/arraymap.c          | 59 +++++++++++++++++---
 kernel/bpf/syscall.c           | 99 ++++++++++++++++++++++++++++++++--
 mm/vmalloc.c                   | 20 +++++++
 tools/include/uapi/linux/bpf.h |  3 ++
 7 files changed, 184 insertions(+), 12 deletions(-)
include/linux/bpf.h +5 -2
index 6fbe599fb977..8021fce98868 100644
@@ -12,6 +12,7 @@
 #include <linux/err.h>
 #include <linux/rbtree_latch.h>
 #include <linux/numa.h>
+#include <linux/mm_types.h>
 #include <linux/wait.h>
 #include <linux/u64_stats_sync.h>
 
@@ -66,6 +67,7 @@ struct bpf_map_ops {
                                   u64 *imm, u32 off);
      int (*map_direct_value_meta)(const struct bpf_map *map,
                                   u64 imm, u32 *off);
+     int (*map_mmap)(struct bpf_map *map, struct vm_area_struct *vma);
 };
 
 struct bpf_map_memory {
@@ -94,9 +96,10 @@ struct bpf_map {
      u32 btf_value_type_id;
      struct btf *btf;
      struct bpf_map_memory memory;
+     char name[BPF_OBJ_NAME_LEN];
      bool unpriv_array;
-     bool frozen; /* write-once */
-     /* 48 bytes hole */
+     bool frozen; /* write-once; write-protected by freeze_mutex */
+     /* 22 bytes hole */
 
      /* The 3rd and 4th cacheline with misc members to avoid false sharing
       * particularly with refcounting.

include/linux/bpf.h struct bpf_map { · line 112

Quoted 7 lines
include/linux/bpf.h +2 -1
@@ -104,7 +107,8 @@ struct bpf_map {
      atomic64_t refcnt ____cacheline_aligned;
      atomic64_t usercnt;
      struct work_struct work;
-     char name[BPF_OBJ_NAME_LEN];
+     struct mutex freeze_mutex;
+     u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
 };

Can the mutex be moved into bpf_array instead of being in bpf_map that is shared across all map types?

Quoted 5 lines
Quoted 2 lines

No, freezing logic is common to all maps. Same for writecnt and mmap()-ing overall.

How mmap is going to work for hash map ? and for prog_array?

It probably won't. But one day we might have hash map using open adressing, which will be more prone to memory-mapping. Or, say, some sort of non-per-CPU ring buffer would be a good candidate as well. It doesn't seem like a good idea to restrict mmap()-ability to just array for no good reason.

But speaking about freeze_mutex specifically. It's to coordinate writable memory-mapping and frozen flag. Even if we make memory-mapping bpf_array specific, map->frozen is generic flag and handled in syscall.c generically, so I just can't protect it only from bpf_array side.

#10 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 114 lines · Andrii Nakryiko
Quoted 99 lines · Alexei Starovoitov
Quoted 97 lines · Andrii Nakryiko
Quoted 94 lines · Alexei Starovoitov
Quoted 82 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

---
 include/linux/bpf.h            | 11 ++--
 include/linux/vmalloc.h        |  1 +
 include/uapi/linux/bpf.h       |  3 ++
 kernel/bpf/arraymap.c          | 59 +++++++++++++++++---
 kernel/bpf/syscall.c           | 99 ++++++++++++++++++++++++++++++++--
 mm/vmalloc.c                   | 20 +++++++
 tools/include/uapi/linux/bpf.h |  3 ++
 7 files changed, 184 insertions(+), 12 deletions(-)
include/linux/bpf.h +5 -2
index 6fbe599fb977..8021fce98868 100644
@@ -12,6 +12,7 @@
 #include <linux/err.h>
 #include <linux/rbtree_latch.h>
 #include <linux/numa.h>
+#include <linux/mm_types.h>
 #include <linux/wait.h>
 #include <linux/u64_stats_sync.h>
 
@@ -66,6 +67,7 @@ struct bpf_map_ops {
                                   u64 *imm, u32 off);
      int (*map_direct_value_meta)(const struct bpf_map *map,
                                   u64 imm, u32 *off);
+     int (*map_mmap)(struct bpf_map *map, struct vm_area_struct *vma);
 };
 
 struct bpf_map_memory {
@@ -94,9 +96,10 @@ struct bpf_map {
      u32 btf_value_type_id;
      struct btf *btf;
      struct bpf_map_memory memory;
+     char name[BPF_OBJ_NAME_LEN];
      bool unpriv_array;
-     bool frozen; /* write-once */
-     /* 48 bytes hole */
+     bool frozen; /* write-once; write-protected by freeze_mutex */
+     /* 22 bytes hole */
 
      /* The 3rd and 4th cacheline with misc members to avoid false sharing
       * particularly with refcounting.

include/linux/bpf.h struct bpf_map { · line 112

Quoted 7 lines
include/linux/bpf.h +2 -1
@@ -104,7 +107,8 @@ struct bpf_map {
      atomic64_t refcnt ____cacheline_aligned;
      atomic64_t usercnt;
      struct work_struct work;
-     char name[BPF_OBJ_NAME_LEN];
+     struct mutex freeze_mutex;
+     u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
 };

Can the mutex be moved into bpf_array instead of being in bpf_map that is shared across all map types?

Quoted 5 lines
Quoted 2 lines

No, freezing logic is common to all maps. Same for writecnt and mmap()-ing overall.

How mmap is going to work for hash map ? and for prog_array?

It probably won't. But one day we might have hash map using open adressing, which will be more prone to memory-mapping. Or, say, some sort of non-per-CPU ring buffer would be a good candidate as well. It doesn't seem like a good idea to restrict mmap()-ability to just array for no good reason.

Quoted 5 lines

But speaking about freeze_mutex specifically. It's to coordinate writable memory-mapping and frozen flag. Even if we make memory-mapping bpf_array specific, map->frozen is generic flag and handled in syscall.c generically, so I just can't protect it only from bpf_array side.

Alternatively we can use spinlock. I don't think it's too ugly, tbh. See below.

From 0da495b911adad495857f1c0fc3596f1d06a705f Mon Sep 17 00:00:00 2001
From: Andrii Nakryiko <[email protected]>
Date: Fri, 15 Nov 2019 08:32:43 -0800
Subject: [PATCH bpf-next] bpf: switch freeze locking to use spin_lock and save
space

Switch to spin_lock in favor of mutex. Due to mmap-ing itself happening not under spinlock, there needs to be an extra "correction" step for writecnt, if mapping fails.

---
 include/linux/bpf.h  |  6 ++---
 kernel/bpf/syscall.c | 52 ++++++++++++++++++++++++--------------------
 2 files changed, 31 insertions(+), 27 deletions(-)
include/linux/bpf.h +3 -3
index 8021fce98868..ceaefd69c6bb 100644
@@ -98,7 +98,7 @@ struct bpf_map {
     struct bpf_map_memory memory;
     char name[BPF_OBJ_NAME_LEN];
     bool unpriv_array;
-    bool frozen; /* write-once; write-protected by freeze_mutex */
+    bool frozen; /* write-once; write-protected by freeze_lock */
     /* 22 bytes hole */
 
     /* The 3rd and 4th cacheline with misc members to avoid false sharing
@@ -107,8 +107,8 @@ struct bpf_map {
     atomic64_t refcnt ____cacheline_aligned;
     atomic64_t usercnt;
     struct work_struct work;
-    struct mutex freeze_mutex;
-    u64 writecnt; /* writable mmap cnt; protected by freeze_mutex */
+    u64 writecnt; /* writable mmap cnt; protected by freeze_lock */
+    spinlock_t freeze_lock;
 };
 
 static inline bool map_value_has_spin_lock(const struct bpf_map *map)
kernel/bpf/syscall.c +4 -4
index 810e53990641..fc782ab06d57 100644
@@ -448,9 +448,9 @@ static void bpf_map_mmap_open(struct vm_area_struct *vma)
     bpf_map_inc(map);
 
     if (vma->vm_flags & VM_WRITE) {
-        mutex_lock(&map->freeze_mutex);
+        spin_lock(&map->freeze_lock);
         map->writecnt++;
-        mutex_unlock(&map->freeze_mutex);
+        spin_unlock(&map->freeze_lock);
     }
 }
 
@@ -460,9 +460,9 @@ static void bpf_map_mmap_close(struct vm_area_struct *vma)
     struct bpf_map *map = vma->vm_file->private_data;
 
     if (vma->vm_flags & VM_WRITE) {
-        mutex_lock(&map->freeze_mutex);
+        spin_lock(&map->freeze_lock);
         map->writecnt--;
-        mutex_unlock(&map->freeze_mutex);
+        spin_unlock(&map->freeze_lock);
     }
 
     bpf_map_put(map);
@@ -484,28 +484,34 @@ static int bpf_map_mmap(struct file *filp,

struct vm_area_struct *vma)

     if (!(vma->vm_flags & VM_SHARED))
         return -EINVAL;

- mutex_lock(&map->freeze_mutex);
-

+    spin_lock(&map->freeze_lock);
     if ((vma->vm_flags & VM_WRITE) && map->frozen) {
-        err = -EPERM;
-        goto out;
+        spin_unlock(&map->freeze_lock);
+        return -EPERM;
     }

-
+ /* speculatively increase writecnt assuming mmap will succeed */
+ if (vma->vm_flags & VM_WRITE)
+ map->writecnt++;
+ spin_unlock(&map->freeze_lock);
+

     /* set default open/close callbacks */
     vma->vm_ops = &bpf_map_default_vmops;
     vma->vm_private_data = map;

     err = map->ops->map_mmap(map, vma);

- if (err)
- goto out;
+ if (err) {
+ /* mmap failed, undo speculative writecnt increment */
+ if (vma->vm_flags & VM_WRITE) {

+            spin_lock(&map->freeze_lock);
+            map->writecnt--;
+            spin_unlock(&map->freeze_lock);

+ }
+ return err;
+ }

     bpf_map_inc(map);

- if (vma->vm_flags & VM_WRITE)
- map->writecnt++;
-out:

-    mutex_unlock(&map->freeze_mutex);
-    return err;
+    return 0;

}

const struct file_operations bpf_map_fops = {
@@ -661,7 +667,7 @@ static int map_create(union bpf_attr *attr)

     atomic64_set(&map->refcnt, 1);
     atomic64_set(&map->usercnt, 1);

- mutex_init(&map->freeze_mutex);
+ spin_lock_init(&map->freeze_lock);

     if (attr->btf_key_type_id || attr->btf_value_type_id) {
         struct btf *btf;
@@ -1244,12 +1250,15 @@ static int map_freeze(const union bpf_attr *attr)
     if (CHECK_ATTR(BPF_MAP_FREEZE))
         return -EINVAL;

+ if (!capable(CAP_SYS_ADMIN))
+ return -EPERM;
+

     f = fdget(ufd);
     map = __bpf_map_get(f);
     if (IS_ERR(map))
         return PTR_ERR(map);

- mutex_lock(&map->freeze_mutex);
+ spin_lock(&map->freeze_lock);

     if (map->writecnt) {
         err = -EBUSY;
@@ -1259,14 +1268,9 @@ static int map_freeze(const union bpf_attr *attr)
         err = -EBUSY;
         goto err_put;
     }

- if (!capable(CAP_SYS_ADMIN)) {
- err = -EPERM;
- goto err_put;
- }
-

     WRITE_ONCE(map->frozen, true);
 err_put:
-    mutex_unlock(&map->freeze_mutex);
+    spin_unlock(&map->freeze_lock);
     fdput(f);
     return err;

}

Signature · 1 line
2.17.1
#11 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 11 lines · Andrii Nakryiko

Alternatively we can use spinlock. I don't think it's too ugly, tbh. See below.

From 0da495b911adad495857f1c0fc3596f1d06a705f Mon Sep 17 00:00:00 2001
From: Andrii Nakryiko <[email protected]>
Date: Fri, 15 Nov 2019 08:32:43 -0800
Subject: [PATCH bpf-next] bpf: switch freeze locking to use spin_lock and save
space

Switch to spin_lock in favor of mutex. Due to mmap-ing itself happening not under spinlock, there needs to be an extra "correction" step for writecnt, if mapping fails.

FWIW I was pondering that too, and thought your initial design was nicer, the transient errors sometimes become a major PITA.

Acked-by on v4 2/4
#12 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

In reply to Andrii Nakryiko
Quoted 33 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

    map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

Quoted 5 lines
maintainer
#13 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

In reply to Andrii Nakryiko
Quoted 38 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful for working with BPF global data from userspace programs. It allows to avoid typical bpf_map_{lookup,update}_elem operations, improving both performance and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),

     map freezing will keep failing with -EBUSY;

- once number of writable memory mappings drops to zero, map freezing can be performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc() to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that struct bpf_array is at the end of a single page of memory with array->value being aligned with the start of the second page. On deallocation we need to accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions. Memory-mapping subsystem provides few optional callbacks, among them open() and close(). close() is called for each memory region that is unmapped, so that users can decrease their reference counters and free up resources, if necessary. open() is *almost* symmetrical: it's called for each memory region that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

[...]

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 448

Quoted 6 lines
kernel/bpf/syscall.c +6 -0
+/* called for any extra memory-mapped regions (except initial) */
+static void bpf_map_mmap_open(struct vm_area_struct *vma)
+{
+	struct bpf_map *map = vma->vm_file->private_data;
+
+	bpf_map_inc(map);

This would also need to inc uref counter since it's technically a reference of this map into user space as otherwise if map->ops->map_release_uref would be used for maps supporting mmap, then the callback would trigger even if user space still has a reference to it.

Quoted 9 lines
kernel/bpf/syscall.c +9 -0
+	if (vma->vm_flags & VM_WRITE) {
+		mutex_lock(&map->freeze_mutex);
+		map->writecnt++;
+		mutex_unlock(&map->freeze_mutex);
+	}
+}
+
+/* called for all unmapped memory region (including initial) */
+static void bpf_map_mmap_close(struct vm_area_struct *vma)

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 468

Quoted 10 lines
kernel/bpf/syscall.c +10 -0
+{
+	struct bpf_map *map = vma->vm_file->private_data;
+
+	if (vma->vm_flags & VM_WRITE) {
+		mutex_lock(&map->freeze_mutex);
+		map->writecnt--;
+		mutex_unlock(&map->freeze_mutex);
+	}
+
+	bpf_map_put(map);

Ditto.

Quoted 24 lines
kernel/bpf/syscall.c +24 -0
+}
+
+static const struct vm_operations_struct bpf_map_default_vmops = {
+	.open		= bpf_map_mmap_open,
+	.close		= bpf_map_mmap_close,
+};
+
+static int bpf_map_mmap(struct file *filp, struct vm_area_struct *vma)
+{
+	struct bpf_map *map = filp->private_data;
+	int err;
+
+	if (!map->ops->map_mmap || map_value_has_spin_lock(map))
+		return -ENOTSUPP;
+
+	if (!(vma->vm_flags & VM_SHARED))
+		return -EINVAL;
+
+	mutex_lock(&map->freeze_mutex);
+
+	if ((vma->vm_flags & VM_WRITE) && map->frozen) {
+		err = -EPERM;
+		goto out;
+	}

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 502

Quoted 10 lines
kernel/bpf/syscall.c +10 -0
+
+	/* set default open/close callbacks */
+	vma->vm_ops = &bpf_map_default_vmops;
+	vma->vm_private_data = map;
+
+	err = map->ops->map_mmap(map, vma);
+	if (err)
+		goto out;
+
+	bpf_map_inc(map);

Ditto.

Quoted 7 lines
kernel/bpf/syscall.c +7 -0
+	if (vma->vm_flags & VM_WRITE)
+		map->writecnt++;
+out:
+	mutex_unlock(&map->freeze_mutex);
+	return err;
+}
+
#14 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 57 lines · Daniel Borkmann
Quoted 54 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful
for working with BPF global data from userspace programs. It allows to avoid
typical bpf_map_{lookup,update}_elem operations, improving both performance
and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and
mmap-ing is happening under mutex now:
   - if map is already frozen, no writable mapping is allowed;
   - if map has writable memory mappings active (accounted in map->writecnt),
     map freezing will keep failing with -EBUSY;
   - once number of writable memory mappings drops to zero, map freezing can be
     performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks
can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc()
to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that
struct bpf_array is at the end of a single page of memory with array->value
being aligned with the start of the second page. On deallocation we need to
accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions.
Memory-mapping subsystem provides few optional callbacks, among them open()
and close().  close() is called for each memory region that is unmapped, so
that users can decrease their reference counters and free up resources, if
necessary. open() is *almost* symmetrical: it's called for each memory region
that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

[...]

Quoted 13 lines

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 448

Quoted 6 lines
kernel/bpf/syscall.c +6 -0
+/* called for any extra memory-mapped regions (except initial) */
+static void bpf_map_mmap_open(struct vm_area_struct *vma)
+{
+    struct bpf_map *map = vma->vm_file->private_data;
+
+    bpf_map_inc(map);

This would also need to inc uref counter since it's technically a reference of this map into user space as otherwise if map->ops->map_release_uref would
be used for maps supporting mmap, then the callback would trigger even if user
space still has a reference to it.

I thought we use uref only for array that can hold FDs ? That's why I suggested Andrii earlier to drop uref++.

maintainer
#15 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 72 lines · Alexei Starovoitov
Quoted 57 lines · Daniel Borkmann
Quoted 54 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful
for working with BPF global data from userspace programs. It allows to avoid
typical bpf_map_{lookup,update}_elem operations, improving both performance
and usability.

There had to be special considerations for map freezing, to avoid having writable memory view into a frozen map. To solve this issue, map freezing and
mmap-ing is happening under mutex now:
   - if map is already frozen, no writable mapping is allowed;
   - if map has writable memory mappings active (accounted in
map->writecnt),
     map freezing will keep failing with -EBUSY;
   - once number of writable memory mappings drops to zero, map
freezing can be
     performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks
can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc()
to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that
struct bpf_array is at the end of a single page of memory with array->value
being aligned with the start of the second page. On deallocation we need to
accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions.
Memory-mapping subsystem provides few optional callbacks, among them open()
and close().  close() is called for each memory region that is unmapped, so
that users can decrease their reference counters and free up resources, if
necessary. open() is *almost* symmetrical: it's called for each memory region
that is being mapped, **except** the very first one. So bpf_map_mmap does initial refcnt bump, while open() will do any extra ones after that. Thus number of close() calls is equal to number of open() calls plus one more.

[...]

Quoted 13 lines

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 448

Quoted 6 lines
kernel/bpf/syscall.c +6 -0
+/* called for any extra memory-mapped regions (except initial) */
+static void bpf_map_mmap_open(struct vm_area_struct *vma)
+{
+    struct bpf_map *map = vma->vm_file->private_data;
+
+    bpf_map_inc(map);

This would also need to inc uref counter since it's technically a reference of this map into user space as otherwise if map->ops->map_release_uref would
be used for maps supporting mmap, then the callback would trigger even if user
space still has a reference to it.

Quoted 2 lines

I thought we use uref only for array that can hold FDs ? That's why I suggested Andrii earlier to drop uref++.

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref callback in future for something else, will we remember that we earlier missed to add it here? :/

#16 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 79 lines · Daniel Borkmann
Quoted 78 lines · Alexei Starovoitov
Quoted 76 lines · Daniel Borkmann
Quoted 59 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful
for working with BPF global data from userspace programs. It allows to avoid
typical bpf_map_{lookup,update}_elem operations, improving both performance
and usability.

There had to be special considerations for map freezing, to avoid having
writable memory view into a frozen map. To solve this issue, map freezing and
mmap-ing is happening under mutex now:
    - if map is already frozen, no writable mapping is allowed;
    - if map has writable memory mappings active (accounted in map->writecnt),
      map freezing will keep failing with -EBUSY;
    - once number of writable memory mappings drops to zero, map freezing can be
      performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks
can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc()
to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that
struct bpf_array is at the end of a single page of memory with array->value
being aligned with the start of the second page. On deallocation we need to
accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions.
Memory-mapping subsystem provides few optional callbacks, among them open()
and close().  close() is called for each memory region that is unmapped, so
that users can decrease their reference counters and free up resources, if
necessary. open() is *almost* symmetrical: it's called for each memory region
that is being mapped, **except** the very first one. So bpf_map_mmap does
initial refcnt bump, while open() will do any extra ones after that. Thus
number of close() calls is equal to number of open() calls plus one more.

[...]

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 448

Quoted 6 lines
kernel/bpf/syscall.c +6 -0
+/* called for any extra memory-mapped regions (except initial) */
+static void bpf_map_mmap_open(struct vm_area_struct *vma)
+{
+    struct bpf_map *map = vma->vm_file->private_data;
+
+    bpf_map_inc(map);

This would also need to inc uref counter since it's technically a reference
of this map into user space as otherwise if map->ops->map_release_uref would
be used for maps supporting mmap, then the callback would trigger even if user
space still has a reference to it.

Quoted 8 lines
Quoted 2 lines

I thought we use uref only for array that can hold FDs ? That's why I suggested Andrii earlier to drop uref++.

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?
This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

maintainer
#17 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 83 lines · Alexei Starovoitov
Quoted 82 lines · Daniel Borkmann
Quoted 80 lines · Alexei Starovoitov
Quoted 62 lines · Daniel Borkmann
Quoted 59 lines · Andrii Nakryiko

Add ability to memory-map contents of BPF array map. This is extremely useful
for working with BPF global data from userspace programs. It allows to avoid
typical bpf_map_{lookup,update}_elem operations, improving both performance
and usability.

There had to be special considerations for map freezing, to avoid having
writable memory view into a frozen map. To solve this issue, map freezing and
mmap-ing is happening under mutex now:
    - if map is already frozen, no writable mapping is allowed;
    - if map has writable memory mappings active (accounted in
map->writecnt),
      map freezing will keep failing with -EBUSY;
    - once number of writable memory mappings drops to zero, map
freezing can be
      performed again.

Only non-per-CPU plain arrays are supported right now. Maps with spinlocks
can't be memory mapped either.

For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc()
to be mmap()'able. We also need to make sure that array data memory is page-sized and page-aligned, so we over-allocate memory in such a way that
struct bpf_array is at the end of a single page of memory with array->value
being aligned with the start of the second page. On deallocation we need to
accomodate this memory arrangement to free vmalloc()'ed memory correctly.

One important consideration regarding how memory-mapping subsystem functions.
Memory-mapping subsystem provides few optional callbacks, among them open()
and close().  close() is called for each memory region that is unmapped, so
that users can decrease their reference counters and free up resources, if
necessary. open() is *almost* symmetrical: it's called for each memory region
that is being mapped, **except** the very first one. So bpf_map_mmap does
initial refcnt bump, while open() will do any extra ones after that. Thus
number of close() calls is equal to number of open() calls plus one more.

[...]

Quoted 14 lines

kernel/bpf/syscall.c static ssize_t bpf_dummy_write(struct file *filp, const char __user *buf, · line 448

Quoted 6 lines
kernel/bpf/syscall.c +6 -0
+/* called for any extra memory-mapped regions (except initial) */
+static void bpf_map_mmap_open(struct vm_area_struct *vma)
+{
+    struct bpf_map *map = vma->vm_file->private_data;
+
+    bpf_map_inc(map);

This would also need to inc uref counter since it's technically a reference
of this map into user space as otherwise if map->ops->map_release_uref would
be used for maps supporting mmap, then the callback would trigger even if user
space still has a reference to it.

I thought we use uref only for array that can hold FDs ? That's why I suggested Andrii earlier to drop uref++.

Quoted 7 lines
Quoted 5 lines

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?

Was saying missed to add the inc/put for the uref counter.

Quoted 2 lines

This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

Right, I meant if in future we ever have another use case outside of it for some reason (unrelated to those maps you mention above). Can we guarantee this is never going to happen? Seemed less fragile at least to maintain proper count here.

Thanks,
Daniel

#18 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 10 lines · Daniel Borkmann
Quoted 7 lines
Quoted 5 lines

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?

Was saying missed to add the inc/put for the uref counter.

Quoted 7 lines
Quoted 2 lines

This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

Right, I meant if in future we ever have another use case outside of it for some reason (unrelated to those maps you mention above). Can we guarantee this is never going to happen? Seemed less fragile at least to maintain proper count here.

I'm struggling to understand the concern.
map-in-map, xskmap, socket local storage are doing bpf_map_inc(, false) when they need to hold the map. Why this case is any different?

#19 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 12 lines · Alexei Starovoitov
Quoted 10 lines · Daniel Borkmann
Quoted 7 lines
Quoted 5 lines

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?

Was saying missed to add the inc/put for the uref counter.

Quoted 7 lines
Quoted 7 lines
Quoted 2 lines

This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

Right, I meant if in future we ever have another use case outside of it for some reason (unrelated to those maps you mention above). Can we guarantee this is never going to happen? Seemed less fragile at least to maintain proper count here.

I don't think we'll ever going to allow mmaping anything that contains not just pure data. E.g., we disallow mmaping array that contains spin lock for that reason. So I think it's safe to assume that this is not going to happen even for future maps. At least not without some serious considerations before that. So I'm going to keep it as just plain bpf_map_inc for now.

I'm going to convert bpf_prog_add/bpf_prog_inc, though, and will do it as a separate patch, on top of bpf_map_inc refactor. It touches quite a lot drivers, so would benefit from having being separate.

Quoted 4 lines

I'm struggling to understand the concern.
map-in-map, xskmap, socket local storage are doing bpf_map_inc(, false) when they need to hold the map. Why this case is any different?

maintainer
#20 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 20 lines · Andrii Nakryiko
Quoted 11 lines · Alexei Starovoitov
Quoted 10 lines · Daniel Borkmann
Quoted 7 lines
Quoted 5 lines

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?

Was saying missed to add the inc/put for the uref counter.

Quoted 7 lines
Quoted 7 lines
Quoted 2 lines

This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

Right, I meant if in future we ever have another use case outside of it for some reason (unrelated to those maps you mention above). Can we guarantee this is never going to happen? Seemed less fragile at least to maintain proper count here.

Quoted 6 lines

I don't think we'll ever going to allow mmaping anything that contains not just pure data. E.g., we disallow mmaping array that contains spin lock for that reason. So I think it's safe to assume that this is not going to happen even for future maps. At least not without some serious considerations before that. So I'm going to keep it as just plain bpf_map_inc for now.

Fair enough, then keep it as it is. The purpose of the uref counter is to track whatever map holds a reference either in form of fd or inode in bpf fs which are the only two things till now where user space can refer to the map, and once it hits 0, we perform the map's map_release_uref() callback.

The fact that some maps make use of it and some others not is an implementation detail in my opinion, but the concept itself is generic and can be used by whatever map implementation would need it in future. From my perspective not breaking with this semantic would allow to worry about one less issue once this callback gets reused for whatever reason.

As I understand, from your PoV, you think that this uref counter is and will be exactly only tied to the fd based maps that currently use it and given they understandably won't ever need a mmap interface we don't need to inc/dec it there.

Fair enough, but could we add some assertion then which adds a check if a map ever uses both that we bail out so we don't forget about this detail in a few weeks from now? Given complexity we have in our BPF codebase these days, I'm mainly worried about the latter if we can catch such details with a trivial check easily, for example, it would be trivial enough to add a test for the existence of map_release_uref callback inside bpf_map_mmap() and bail out in order to guarantee this, similar as you do with the spinlock.

Quoted 3 lines

I'm going to convert bpf_prog_add/bpf_prog_inc, though, and will do it as a separate patch, on top of bpf_map_inc refactor. It touches quite a lot drivers, so would benefit from having being separate.

Yeah, sounds good to me. Thanks for converting!

Quoted 3 lines
Quoted 3 lines

I'm struggling to understand the concern.
map-in-map, xskmap, socket local storage are doing bpf_map_inc(, false) when they need to hold the map. Why this case is any different?

(See above.)

#21 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 29 lines · Daniel Borkmann
Quoted 20 lines · Andrii Nakryiko
Quoted 11 lines · Alexei Starovoitov
Quoted 10 lines · Daniel Borkmann
Quoted 7 lines
Quoted 5 lines

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?

Was saying missed to add the inc/put for the uref counter.

Quoted 7 lines
Quoted 7 lines
Quoted 2 lines

This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

Right, I meant if in future we ever have another use case outside of it for some reason (unrelated to those maps you mention above). Can we guarantee this is never going to happen? Seemed less fragile at least to maintain proper count here.

Quoted 6 lines

I don't think we'll ever going to allow mmaping anything that contains not just pure data. E.g., we disallow mmaping array that contains spin lock for that reason. So I think it's safe to assume that this is not going to happen even for future maps. At least not without some serious considerations before that. So I'm going to keep it as just plain bpf_map_inc for now.

Quoted 4 lines

Fair enough, then keep it as it is. The purpose of the uref counter is to track whatever map holds a reference either in form of fd or inode in bpf fs which are the only two things till now where user space can refer to the map, and once it hits 0, we perform the map's map_release_uref() callback.

To be honest, I don't exactly understand why we need both refcnt and usercnt. Does it have anything to do with some circular dependencies for those maps containing other FDs? And once userspace doesn't have any more referenced, we release FDs, which might decrement refcnt, thus breaking circular refcnt between map FD and special FDs inside a map? Or that's not the case and there is a different reason?

Either way, I looked at map creation and bpf_map_release() implementation again. map_create() does set usercnt to 1, and bpf_map_release(), which I assume is called when map FD is closed, does decrement usercnt, so yeah, we do with_uref() manipulations for cases when userspace maintains some sort of handle to map. In that regard, mmap() does fall into the same category, so I'm going to convert everything mmap-related back to bpf_map_inc_with_uref()/bpf_map_put_with_uref(), to stay consistent.

Quoted 31 lines

The fact that some maps make use of it and some others not is an implementation detail in my opinion, but the concept itself is generic and can be used by whatever map implementation would need it in future. From my perspective not breaking with this semantic would allow to worry about one less issue once this callback gets reused for whatever reason.

As I understand, from your PoV, you think that this uref counter is and will be exactly only tied to the fd based maps that currently use it and given they understandably won't ever need a mmap interface we don't need to inc/dec it there.

Fair enough, but could we add some assertion then which adds a check if a map ever uses both that we bail out so we don't forget about this detail in a few weeks from now? Given complexity we have in our BPF codebase these days, I'm mainly worried about the latter if we can catch such details with a trivial check easily, for example, it would be trivial enough to add a test for the existence of map_release_uref callback inside bpf_map_mmap() and bail out in order to guarantee this, similar as you do with the spinlock.

Quoted 3 lines

I'm going to convert bpf_prog_add/bpf_prog_inc, though, and will do it as a separate patch, on top of bpf_map_inc refactor. It touches quite a lot drivers, so would benefit from having being separate.

Yeah, sounds good to me. Thanks for converting!

Quoted 3 lines
Quoted 3 lines

I'm struggling to understand the concern.
map-in-map, xskmap, socket local storage are doing bpf_map_inc(, false) when they need to hold the map. Why this case is any different?

(See above.)

maintainer
#22 lore raw

Re: [PATCH v4 bpf-next 2/4] bpf: add mmap() support for BPF_MAP_TYPE_ARRAY

Quoted 34 lines · Andrii Nakryiko
Quoted 28 lines · Daniel Borkmann
Quoted 20 lines · Andrii Nakryiko
Quoted 11 lines · Alexei Starovoitov
Quoted 10 lines · Daniel Borkmann
Quoted 7 lines
Quoted 5 lines

Yeah, only for fd array currently. Question is, if we ever reuse that map_release_uref
callback in future for something else, will we remember that we earlier missed to add
it here? :/

What do you mean 'missed to add' ?

Was saying missed to add the inc/put for the uref counter.

Quoted 7 lines
Quoted 7 lines
Quoted 2 lines

This is mmap path. Anything that needs releasing (like FDs for prog_array or progs for sockmap) cannot be mmap-able.

Right, I meant if in future we ever have another use case outside of it for some reason (unrelated to those maps you mention above). Can we guarantee this is never going to happen? Seemed less fragile at least to maintain proper count here.

Quoted 6 lines

I don't think we'll ever going to allow mmaping anything that contains not just pure data. E.g., we disallow mmaping array that contains spin lock for that reason. So I think it's safe to assume that this is not going to happen even for future maps. At least not without some serious considerations before that. So I'm going to keep it as just plain bpf_map_inc for now.

Quoted 4 lines

Fair enough, then keep it as it is. The purpose of the uref counter is to track whatever map holds a reference either in form of fd or inode in bpf fs which are the only two things till now where user space can refer to the map, and once it hits 0, we perform the map's map_release_uref() callback.

Quoted 6 lines

To be honest, I don't exactly understand why we need both refcnt and usercnt. Does it have anything to do with some circular dependencies for those maps containing other FDs? And once userspace doesn't have any more referenced, we release FDs, which might decrement refcnt, thus breaking circular refcnt between map FD and special FDs inside a map? Or that's not the case and there is a different reason?

Yes, back then we added it to break up circular dependencies in relation to tail calls which is why these are pinned before the loader process finishes (e.g. as in Cilium's case).

Quoted 8 lines

Either way, I looked at map creation and bpf_map_release() implementation again. map_create() does set usercnt to 1, and bpf_map_release(), which I assume is called when map FD is closed, does decrement usercnt, so yeah, we do with_uref() manipulations for cases when userspace maintains some sort of handle to map. In that regard, mmap() does fall into the same category, so I'm going to convert everything mmap-related back to bpf_map_inc_with_uref()/bpf_map_put_with_uref(), to stay consistent.

Ok, fair enough, either way would have been fine.

Thanks a lot,
Daniel

#23 lore raw

PATCH v4 3/4 libbpf: make global data internal arrays mmap()-able, if possible

In reply to Andrii Nakryiko

Add detection of BPF_F_MMAPABLE flag support for arrays and add it as an extra flag to internal global data maps, if supported by kernel. This allows users to memory-map global data and use it without BPF map operations, greatly simplifying user experience.

---
 tools/lib/bpf/libbpf.c | 32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)
tools/lib/bpf/libbpf.c +30 -2
index 96ef18cfeffb..7e2be1288fa1 100644
@@ -142,6 +142,8 @@ struct bpf_capabilities {
 	__u32 btf_func:1;
 	/* BTF_KIND_VAR and BTF_KIND_DATASEC support */
 	__u32 btf_datasec:1;
+	/* BPF_F_MMAPABLE is supported for arrays */
+	__u32 array_mmap:1;
 };
 
 /*
@@ -856,8 +858,6 @@ bpf_object__init_internal_map(struct bpf_object *obj, enum libbpf_map_type type,
 		pr_warn("failed to alloc map name\n");
 		return -ENOMEM;
 	}
-	pr_debug("map '%s' (global data): at sec_idx %d, offset %zu.\n",
-		 map_name, map->sec_idx, map->sec_offset);
 
 	def = &map->def;
 	def->type = BPF_MAP_TYPE_ARRAY;
@@ -865,6 +865,12 @@ bpf_object__init_internal_map(struct bpf_object *obj, enum libbpf_map_type type,
 	def->value_size = data->d_size;
 	def->max_entries = 1;
 	def->map_flags = type == LIBBPF_MAP_RODATA ? BPF_F_RDONLY_PROG : 0;
+	if (obj->caps.array_mmap)
+		def->map_flags |= BPF_F_MMAPABLE;
+
+	pr_debug("map '%s' (global data): at sec_idx %d, offset %zu, flags %x.\n",
+		 map_name, map->sec_idx, map->sec_offset, def->map_flags);
+
 	if (data_buff) {
 		*data_buff = malloc(data->d_size);
 		if (!*data_buff) {
@@ -2160,6 +2166,27 @@ static int bpf_object__probe_btf_datasec(struct bpf_object *obj)
 	return 0;
 }
 
+static int bpf_object__probe_array_mmap(struct bpf_object *obj)
+{
+	struct bpf_create_map_attr attr = {
+		.map_type = BPF_MAP_TYPE_ARRAY,
+		.map_flags = BPF_F_MMAPABLE,
+		.key_size = sizeof(int),
+		.value_size = sizeof(int),
+		.max_entries = 1,
+	};
+	int fd;
+
+	fd = bpf_create_map_xattr(&attr);
+	if (fd >= 0) {
+		obj->caps.array_mmap = 1;
+		close(fd);
+		return 1;
+	}
+
+	return 0;
+}
+
 static int
 bpf_object__probe_caps(struct bpf_object *obj)
 {
@@ -2168,6 +2195,7 @@ bpf_object__probe_caps(struct bpf_object *obj)
 		bpf_object__probe_global_data,
 		bpf_object__probe_btf_func,
 		bpf_object__probe_btf_datasec,
+		bpf_object__probe_array_mmap,
 	};
 	int i, ret;
 
Signature · 1 line
2.17.1
#24 lore raw

PATCH v4 4/4 selftests/bpf: add BPF_TYPE_MAP_ARRAY mmap() tests

In reply to Andrii Nakryiko

Add selftests validating mmap()-ing BPF array maps: both single-element and multi-element ones. Check that plain bpf_map_update_elem() and bpf_map_lookup_elem() work correctly with memory-mapped array. Also convert CO-RE relocation tests to use memory-mapped views of global data.

---
 .../selftests/bpf/prog_tests/core_reloc.c     |  45 ++--
 tools/testing/selftests/bpf/prog_tests/mmap.c | 220 ++++++++++++++++++
 tools/testing/selftests/bpf/progs/test_mmap.c |  45 ++++
 3 files changed, 292 insertions(+), 18 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/mmap.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_mmap.c
tools/testing/selftests/bpf/prog_tests/core_reloc.c +27 -18
index f94bd071536b..ec9e2fdd6b89 100644
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 #include <test_progs.h>
 #include "progs/core_reloc_types.h"
+#include <sys/mman.h>
 
 #define STRUCT_TO_CHAR_PTR(struct_name) (const char *)&(struct struct_name)
 
@@ -453,8 +454,15 @@ struct data {
 	char out[256];
 };
 
+static size_t roundup_page(size_t sz)
+{
+	long page_size = sysconf(_SC_PAGE_SIZE);
+	return (sz + page_size - 1) / page_size * page_size;
+}
+
 void test_core_reloc(void)
 {
+	const size_t mmap_sz = roundup_page(sizeof(struct data));
 	struct bpf_object_load_attr load_attr = {};
 	struct core_reloc_test_case *test_case;
 	const char *tp_name, *probe_name;
@@ -463,8 +471,8 @@ void test_core_reloc(void)
 	struct bpf_map *data_map;
 	struct bpf_program *prog;
 	struct bpf_object *obj;
-	const int zero = 0;
-	struct data data;
+	struct data *data;
+	void *mmap_data = NULL;
 
 	for (i = 0; i < ARRAY_SIZE(test_cases); i++) {
 		test_case = &test_cases[i];
@@ -476,8 +484,7 @@ void test_core_reloc(void)
 		);
 
 		obj = bpf_object__open_file(test_case->bpf_obj_file, &opts);
-		if (CHECK(IS_ERR_OR_NULL(obj), "obj_open",
-			  "failed to open '%s': %ld\n",
+		if (CHECK(IS_ERR(obj), "obj_open", "failed to open '%s': %ld\n",
 			  test_case->bpf_obj_file, PTR_ERR(obj)))
 			continue;
 
@@ -519,24 +526,22 @@ void test_core_reloc(void)
 		if (CHECK(!data_map, "find_data_map", "data map not found\n"))
 			goto cleanup;
 
-		memset(&data, 0, sizeof(data));
-		memcpy(data.in, test_case->input, test_case->input_len);
-
-		err = bpf_map_update_elem(bpf_map__fd(data_map),
-					  &zero, &data, 0);
-		if (CHECK(err, "update_data_map",
-			  "failed to update .data map: %d\n", err))
+		mmap_data = mmap(NULL, mmap_sz, PROT_READ | PROT_WRITE,
+				 MAP_SHARED, bpf_map__fd(data_map), 0);
+		if (CHECK(mmap_data == MAP_FAILED, "mmap",
+			  ".bss mmap failed: %d", errno)) {
+			mmap_data = NULL;
 			goto cleanup;
+		}
+		data = mmap_data;
+
+		memset(mmap_data, 0, sizeof(*data));
+		memcpy(data->in, test_case->input, test_case->input_len);
 
 		/* trigger test run */
 		usleep(1);
 
-		err = bpf_map_lookup_elem(bpf_map__fd(data_map), &zero, &data);
-		if (CHECK(err, "get_result",
-			  "failed to get output data: %d\n", err))
-			goto cleanup;
-
-		equal = memcmp(data.out, test_case->output,
+		equal = memcmp(data->out, test_case->output,
 			       test_case->output_len) == 0;
 		if (CHECK(!equal, "check_result",
 			  "input/output data don't match\n")) {
@@ -548,12 +553,16 @@ void test_core_reloc(void)
 			}
 			for (j = 0; j < test_case->output_len; j++) {
 				printf("output byte #%d: EXP 0x%02hhx GOT 0x%02hhx\n",
-				       j, test_case->output[j], data.out[j]);
+				       j, test_case->output[j], data->out[j]);
 			}
 			goto cleanup;
 		}
 
 cleanup:
+		if (mmap_data) {
+			CHECK_FAIL(munmap(mmap_data, mmap_sz));
+			mmap_data = NULL;
+		}
 		if (!IS_ERR_OR_NULL(link)) {
 			bpf_link__destroy(link);
 			link = NULL;
tools/testing/selftests/bpf/prog_tests/mmap.c new file +220 -0

220 lines not on this page · raw message

tools/testing/selftests/bpf/progs/test_mmap.c new file +45 -0

45 lines not on this page · raw message

Signature · 1 line
2.17.1