2016-09-21 18:43:57 +08:00
|
|
|
/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of version 2 of the GNU General Public
|
|
|
|
* License as published by the Free Software Foundation.
|
|
|
|
*/
|
|
|
|
#ifndef _LINUX_BPF_VERIFIER_H
|
|
|
|
#define _LINUX_BPF_VERIFIER_H 1
|
|
|
|
|
|
|
|
#include <linux/bpf.h> /* for enum bpf_reg_type */
|
|
|
|
#include <linux/filter.h> /* for MAX_BPF_STACK */
|
2017-08-07 22:26:19 +08:00
|
|
|
#include <linux/tnum.h>
|
2016-09-21 18:43:57 +08:00
|
|
|
|
2017-08-07 22:26:36 +08:00
|
|
|
/* Maximum variable offset umax_value permitted when resolving memory accesses.
|
|
|
|
* In practice this is far bigger than any realistic pointer offset; this limit
|
|
|
|
* ensures that umax_value + (int)off + (int)size cannot overflow a u64.
|
|
|
|
*/
|
2017-12-19 12:12:00 +08:00
|
|
|
#define BPF_MAX_VAR_OFF (1 << 29)
|
2017-08-07 22:26:36 +08:00
|
|
|
/* Maximum variable size permitted for ARG_CONST_SIZE[_OR_ZERO]. This ensures
|
|
|
|
* that converting umax_value to int cannot overflow.
|
|
|
|
*/
|
2017-12-19 12:12:00 +08:00
|
|
|
#define BPF_MAX_VAR_SIZ (1 << 29)
|
2016-09-28 22:54:32 +08:00
|
|
|
|
2017-08-23 22:11:21 +08:00
|
|
|
/* Liveness marks, used for registers and spilled-regs (in stack slots).
|
|
|
|
* Read marks propagate upwards until they find a write mark; they record that
|
|
|
|
* "one of this state's descendants read this reg" (and therefore the reg is
|
|
|
|
* relevant for states_equal() checks).
|
|
|
|
* Write marks collect downwards and do not propagate; they record that "the
|
|
|
|
* straight-line code that reached this state (from its parent) wrote this reg"
|
|
|
|
* (and therefore that reads propagated from this state or its descendants
|
|
|
|
* should not propagate to its parent).
|
|
|
|
* A state with a write mark can receive read marks; it just won't propagate
|
|
|
|
* them to its parent, since the write mark is a property, not of the state,
|
|
|
|
* but of the link between it and its parent. See mark_reg_read() and
|
|
|
|
* mark_stack_slot_read() in kernel/bpf/verifier.c.
|
|
|
|
*/
|
2017-08-16 03:34:35 +08:00
|
|
|
enum bpf_reg_liveness {
|
|
|
|
REG_LIVE_NONE = 0, /* reg hasn't been read or written this branch */
|
|
|
|
REG_LIVE_READ, /* reg was read, so we're sensitive to initial value */
|
|
|
|
REG_LIVE_WRITTEN, /* reg was written first, screening off later reads */
|
2018-12-14 03:42:34 +08:00
|
|
|
REG_LIVE_DONE = 4, /* liveness won't be updating this register anymore */
|
2017-08-16 03:34:35 +08:00
|
|
|
};
|
|
|
|
|
2016-09-21 18:43:57 +08:00
|
|
|
struct bpf_reg_state {
|
2018-08-23 03:02:19 +08:00
|
|
|
/* Ordering of fields matters. See states_equal() */
|
2016-09-21 18:43:57 +08:00
|
|
|
enum bpf_reg_type type;
|
|
|
|
union {
|
2017-08-07 22:26:19 +08:00
|
|
|
/* valid when type == PTR_TO_PACKET */
|
|
|
|
u16 range;
|
2016-09-21 18:43:57 +08:00
|
|
|
|
|
|
|
/* valid when type == CONST_PTR_TO_MAP | PTR_TO_MAP_VALUE |
|
|
|
|
* PTR_TO_MAP_VALUE_OR_NULL
|
|
|
|
*/
|
|
|
|
struct bpf_map *map_ptr;
|
2018-11-01 07:05:52 +08:00
|
|
|
|
|
|
|
/* Max size from any of the above. */
|
|
|
|
unsigned long raw;
|
2016-09-21 18:43:57 +08:00
|
|
|
};
|
2017-08-07 22:26:19 +08:00
|
|
|
/* Fixed part of pointer offset, pointer types only */
|
|
|
|
s32 off;
|
|
|
|
/* For PTR_TO_PACKET, used to find other pointers with the same variable
|
|
|
|
* offset, so they can share range knowledge.
|
|
|
|
* For PTR_TO_MAP_VALUE_OR_NULL this is used to share which map value we
|
|
|
|
* came from, when one is tested for != NULL.
|
2018-10-03 04:35:33 +08:00
|
|
|
* For PTR_TO_SOCKET this is used to share which pointers retain the
|
|
|
|
* same reference to the socket, to determine proper reference freeing.
|
2017-08-07 22:26:19 +08:00
|
|
|
*/
|
2016-12-08 02:57:59 +08:00
|
|
|
u32 id;
|
bpf: Fix bpf_tcp_sock and bpf_sk_fullsock issue related to bpf_sk_release
Lorenz Bauer [thanks!] reported that a ptr returned by bpf_tcp_sock(sk)
can still be accessed after bpf_sk_release(sk).
Both bpf_tcp_sock() and bpf_sk_fullsock() have the same issue.
This patch addresses them together.
A simple reproducer looks like this:
sk = bpf_sk_lookup_tcp();
/* if (!sk) ... */
tp = bpf_tcp_sock(sk);
/* if (!tp) ... */
bpf_sk_release(sk);
snd_cwnd = tp->snd_cwnd; /* oops! The verifier does not complain. */
The problem is the verifier did not scrub the register's states of
the tcp_sock ptr (tp) after bpf_sk_release(sk).
[ Note that when calling bpf_tcp_sock(sk), the sk is not always
refcount-acquired. e.g. bpf_tcp_sock(skb->sk). The verifier works
fine for this case. ]
Currently, the verifier does not track if a helper's return ptr (in REG_0)
is "carry"-ing one of its argument's refcount status. To carry this info,
the reg1->id needs to be stored in reg0.
One approach was tried, like "reg0->id = reg1->id", when calling
"bpf_tcp_sock()". The main idea was to avoid adding another "ref_obj_id"
for the same reg. However, overlapping the NULL marking and ref
tracking purpose in one "id" does not work well:
ref_sk = bpf_sk_lookup_tcp();
fullsock = bpf_sk_fullsock(ref_sk);
tp = bpf_tcp_sock(ref_sk);
if (!fullsock) {
bpf_sk_release(ref_sk);
return 0;
}
/* fullsock_reg->id is marked for NOT-NULL.
* Same for tp_reg->id because they have the same id.
*/
/* oops. verifier did not complain about the missing !tp check */
snd_cwnd = tp->snd_cwnd;
Hence, a new "ref_obj_id" is needed in "struct bpf_reg_state".
With a new ref_obj_id, when bpf_sk_release(sk) is called, the verifier can
scrub all reg states which has a ref_obj_id match. It is done with the
changes in release_reg_references() in this patch.
While fixing it, sk_to_full_sk() is removed from bpf_tcp_sock() and
bpf_sk_fullsock() to avoid these helpers from returning
another ptr. It will make bpf_sk_release(tp) possible:
sk = bpf_sk_lookup_tcp();
/* if (!sk) ... */
tp = bpf_tcp_sock(sk);
/* if (!tp) ... */
bpf_sk_release(tp);
A separate helper "bpf_get_listener_sock()" will be added in a later
patch to do sk_to_full_sk().
Misc change notes:
- To allow bpf_sk_release(tp), the arg of bpf_sk_release() is changed
from ARG_PTR_TO_SOCKET to ARG_PTR_TO_SOCK_COMMON. ARG_PTR_TO_SOCKET
is removed from bpf.h since no helper is using it.
- arg_type_is_refcounted() is renamed to arg_type_may_be_refcounted()
because ARG_PTR_TO_SOCK_COMMON is the only one and skb->sk is not
refcounted. All bpf_sk_release(), bpf_sk_fullsock() and bpf_tcp_sock()
take ARG_PTR_TO_SOCK_COMMON.
- check_refcount_ok() ensures is_acquire_function() cannot take
arg_type_may_be_refcounted() as its argument.
- The check_func_arg() can only allow one refcount-ed arg. It is
guaranteed by check_refcount_ok() which ensures at most one arg can be
refcounted. Hence, it is a verifier internal error if >1 refcount arg
found in check_func_arg().
- In release_reference(), release_reference_state() is called
first to ensure a match on "reg->ref_obj_id" can be found before
scrubbing the reg states with release_reg_references().
- reg_is_refcounted() is no longer needed.
1. In mark_ptr_or_null_regs(), its usage is replaced by
"ref_obj_id && ref_obj_id == id" because,
when is_null == true, release_reference_state() should only be
called on the ref_obj_id obtained by a acquire helper (i.e.
is_acquire_function() == true). Otherwise, the following
would happen:
sk = bpf_sk_lookup_tcp();
/* if (!sk) { ... } */
fullsock = bpf_sk_fullsock(sk);
if (!fullsock) {
/*
* release_reference_state(fullsock_reg->ref_obj_id)
* where fullsock_reg->ref_obj_id == sk_reg->ref_obj_id.
*
* Hence, the following bpf_sk_release(sk) will fail
* because the ref state has already been released in the
* earlier release_reference_state(fullsock_reg->ref_obj_id).
*/
bpf_sk_release(sk);
}
2. In release_reg_references(), the current reg_is_refcounted() call
is unnecessary because the id check is enough.
- The type_is_refcounted() and type_is_refcounted_or_null()
are no longer needed also because reg_is_refcounted() is removed.
Fixes: 655a51e536c0 ("bpf: Add struct bpf_tcp_sock and BPF_FUNC_tcp_sock")
Reported-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-03-13 01:23:02 +08:00
|
|
|
/* PTR_TO_SOCKET and PTR_TO_TCP_SOCK could be a ptr returned
|
|
|
|
* from a pointer-cast helper, bpf_sk_fullsock() and
|
|
|
|
* bpf_tcp_sock().
|
|
|
|
*
|
|
|
|
* Consider the following where "sk" is a reference counted
|
|
|
|
* pointer returned from "sk = bpf_sk_lookup_tcp();":
|
|
|
|
*
|
|
|
|
* 1: sk = bpf_sk_lookup_tcp();
|
|
|
|
* 2: if (!sk) { return 0; }
|
|
|
|
* 3: fullsock = bpf_sk_fullsock(sk);
|
|
|
|
* 4: if (!fullsock) { bpf_sk_release(sk); return 0; }
|
|
|
|
* 5: tp = bpf_tcp_sock(fullsock);
|
|
|
|
* 6: if (!tp) { bpf_sk_release(sk); return 0; }
|
|
|
|
* 7: bpf_sk_release(sk);
|
|
|
|
* 8: snd_cwnd = tp->snd_cwnd; // verifier will complain
|
|
|
|
*
|
|
|
|
* After bpf_sk_release(sk) at line 7, both "fullsock" ptr and
|
|
|
|
* "tp" ptr should be invalidated also. In order to do that,
|
|
|
|
* the reg holding "fullsock" and "sk" need to remember
|
|
|
|
* the original refcounted ptr id (i.e. sk_reg->id) in ref_obj_id
|
|
|
|
* such that the verifier can reset all regs which have
|
|
|
|
* ref_obj_id matching the sk_reg->id.
|
|
|
|
*
|
|
|
|
* sk_reg->ref_obj_id is set to sk_reg->id at line 1.
|
|
|
|
* sk_reg->id will stay as NULL-marking purpose only.
|
|
|
|
* After NULL-marking is done, sk_reg->id can be reset to 0.
|
|
|
|
*
|
|
|
|
* After "fullsock = bpf_sk_fullsock(sk);" at line 3,
|
|
|
|
* fullsock_reg->ref_obj_id is set to sk_reg->ref_obj_id.
|
|
|
|
*
|
|
|
|
* After "tp = bpf_tcp_sock(fullsock);" at line 5,
|
|
|
|
* tp_reg->ref_obj_id is set to fullsock_reg->ref_obj_id
|
|
|
|
* which is the same as sk_reg->ref_obj_id.
|
|
|
|
*
|
|
|
|
* From the verifier perspective, if sk, fullsock and tp
|
|
|
|
* are not NULL, they are the same ptr with different
|
|
|
|
* reg->type. In particular, bpf_sk_release(tp) is also
|
|
|
|
* allowed and has the same effect as bpf_sk_release(sk).
|
|
|
|
*/
|
|
|
|
u32 ref_obj_id;
|
2017-08-07 22:26:19 +08:00
|
|
|
/* For scalar types (SCALAR_VALUE), this represents our knowledge of
|
|
|
|
* the actual value.
|
|
|
|
* For pointer types, this represents the variable part of the offset
|
|
|
|
* from the pointed-to object, and is shared with all bpf_reg_states
|
|
|
|
* with the same id as us.
|
|
|
|
*/
|
|
|
|
struct tnum var_off;
|
2016-12-08 02:57:59 +08:00
|
|
|
/* Used to determine if any memory access using this register will
|
2017-08-07 22:26:19 +08:00
|
|
|
* result in a bad access.
|
|
|
|
* These refer to the same value as var_off, not necessarily the actual
|
|
|
|
* contents of the register.
|
2016-12-08 02:57:59 +08:00
|
|
|
*/
|
2017-08-07 22:26:36 +08:00
|
|
|
s64 smin_value; /* minimum possible (s64)value */
|
|
|
|
s64 smax_value; /* maximum possible (s64)value */
|
|
|
|
u64 umin_value; /* minimum possible (u64)value */
|
|
|
|
u64 umax_value; /* maximum possible (u64)value */
|
2018-08-23 03:02:19 +08:00
|
|
|
/* parentage chain for liveness checking */
|
|
|
|
struct bpf_reg_state *parent;
|
2017-12-15 09:55:06 +08:00
|
|
|
/* Inside the callee two registers can be both PTR_TO_STACK like
|
|
|
|
* R1=fp-8 and R2=fp-8, but one of them points to this function stack
|
|
|
|
* while another to the caller's stack. To differentiate them 'frameno'
|
|
|
|
* is used which is an index in bpf_verifier_state->frame[] array
|
|
|
|
* pointing to bpf_func_state.
|
|
|
|
*/
|
|
|
|
u32 frameno;
|
2017-08-16 03:34:35 +08:00
|
|
|
enum bpf_reg_liveness live;
|
2016-09-21 18:43:57 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
enum bpf_stack_slot_type {
|
|
|
|
STACK_INVALID, /* nothing was stored in this stack slot */
|
|
|
|
STACK_SPILL, /* register spilled into stack */
|
2017-12-15 09:55:08 +08:00
|
|
|
STACK_MISC, /* BPF program wrote some data into this slot */
|
|
|
|
STACK_ZERO, /* BPF program wrote constant zero */
|
2016-09-21 18:43:57 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
#define BPF_REG_SIZE 8 /* size of eBPF register in bytes */
|
|
|
|
|
2017-11-01 09:16:05 +08:00
|
|
|
struct bpf_stack_state {
|
|
|
|
struct bpf_reg_state spilled_ptr;
|
|
|
|
u8 slot_type[BPF_REG_SIZE];
|
|
|
|
};
|
|
|
|
|
2018-10-03 04:35:35 +08:00
|
|
|
struct bpf_reference_state {
|
|
|
|
/* Track each reference created with a unique id, even if the same
|
|
|
|
* instruction creates the reference multiple times (eg, via CALL).
|
|
|
|
*/
|
|
|
|
int id;
|
|
|
|
/* Instruction where the allocation of this reference occurred. This
|
|
|
|
* is used purely to inform the user of a reference leak.
|
|
|
|
*/
|
|
|
|
int insn_idx;
|
|
|
|
};
|
|
|
|
|
2016-09-21 18:43:57 +08:00
|
|
|
/* state of the program:
|
|
|
|
* type of all registers and stack info
|
|
|
|
*/
|
2017-12-15 09:55:06 +08:00
|
|
|
struct bpf_func_state {
|
2016-09-21 18:43:57 +08:00
|
|
|
struct bpf_reg_state regs[MAX_BPF_REG];
|
2017-12-15 09:55:06 +08:00
|
|
|
/* index of call instruction that called into this func */
|
|
|
|
int callsite;
|
|
|
|
/* stack frame number of this function state from pov of
|
|
|
|
* enclosing bpf_verifier_state.
|
|
|
|
* 0 = main function, 1 = first callee.
|
|
|
|
*/
|
|
|
|
u32 frameno;
|
|
|
|
/* subprog number == index within subprog_stack_depth
|
|
|
|
* zero == main subprog
|
|
|
|
*/
|
|
|
|
u32 subprogno;
|
|
|
|
|
2018-10-03 04:35:35 +08:00
|
|
|
/* The following fields should be last. See copy_func_state() */
|
|
|
|
int acquired_refs;
|
|
|
|
struct bpf_reference_state *refs;
|
2017-11-01 09:16:05 +08:00
|
|
|
int allocated_stack;
|
|
|
|
struct bpf_stack_state *stack;
|
2016-09-21 18:43:57 +08:00
|
|
|
};
|
|
|
|
|
2017-12-15 09:55:06 +08:00
|
|
|
#define MAX_CALL_FRAMES 8
|
|
|
|
struct bpf_verifier_state {
|
|
|
|
/* call stack tracking */
|
|
|
|
struct bpf_func_state *frame[MAX_CALL_FRAMES];
|
|
|
|
u32 curframe;
|
2019-02-01 07:40:04 +08:00
|
|
|
u32 active_spin_lock;
|
bpf: prevent out of bounds speculation on pointer arithmetic
Jann reported that the original commit back in b2157399cc98
("bpf: prevent out-of-bounds speculation") was not sufficient
to stop CPU from speculating out of bounds memory access:
While b2157399cc98 only focussed on masking array map access
for unprivileged users for tail calls and data access such
that the user provided index gets sanitized from BPF program
and syscall side, there is still a more generic form affected
from BPF programs that applies to most maps that hold user
data in relation to dynamic map access when dealing with
unknown scalars or "slow" known scalars as access offset, for
example:
- Load a map value pointer into R6
- Load an index into R7
- Do a slow computation (e.g. with a memory dependency) that
loads a limit into R8 (e.g. load the limit from a map for
high latency, then mask it to make the verifier happy)
- Exit if R7 >= R8 (mispredicted branch)
- Load R0 = R6[R7]
- Load R0 = R6[R0]
For unknown scalars there are two options in the BPF verifier
where we could derive knowledge from in order to guarantee
safe access to the memory: i) While </>/<=/>= variants won't
allow to derive any lower or upper bounds from the unknown
scalar where it would be safe to add it to the map value
pointer, it is possible through ==/!= test however. ii) another
option is to transform the unknown scalar into a known scalar,
for example, through ALU ops combination such as R &= <imm>
followed by R |= <imm> or any similar combination where the
original information from the unknown scalar would be destroyed
entirely leaving R with a constant. The initial slow load still
precedes the latter ALU ops on that register, so the CPU
executes speculatively from that point. Once we have the known
scalar, any compare operation would work then. A third option
only involving registers with known scalars could be crafted
as described in [0] where a CPU port (e.g. Slow Int unit)
would be filled with many dependent computations such that
the subsequent condition depending on its outcome has to wait
for evaluation on its execution port and thereby executing
speculatively if the speculated code can be scheduled on a
different execution port, or any other form of mistraining
as described in [1], for example. Given this is not limited
to only unknown scalars, not only map but also stack access
is affected since both is accessible for unprivileged users
and could potentially be used for out of bounds access under
speculation.
In order to prevent any of these cases, the verifier is now
sanitizing pointer arithmetic on the offset such that any
out of bounds speculation would be masked in a way where the
pointer arithmetic result in the destination register will
stay unchanged, meaning offset masked into zero similar as
in array_index_nospec() case. With regards to implementation,
there are three options that were considered: i) new insn
for sanitation, ii) push/pop insn and sanitation as inlined
BPF, iii) reuse of ax register and sanitation as inlined BPF.
Option i) has the downside that we end up using from reserved
bits in the opcode space, but also that we would require
each JIT to emit masking as native arch opcodes meaning
mitigation would have slow adoption till everyone implements
it eventually which is counter-productive. Option ii) and iii)
have both in common that a temporary register is needed in
order to implement the sanitation as inlined BPF since we
are not allowed to modify the source register. While a push /
pop insn in ii) would be useful to have in any case, it
requires once again that every JIT needs to implement it
first. While possible, amount of changes needed would also
be unsuitable for a -stable patch. Therefore, the path which
has fewer changes, less BPF instructions for the mitigation
and does not require anything to be changed in the JITs is
option iii) which this work is pursuing. The ax register is
already mapped to a register in all JITs (modulo arm32 where
it's mapped to stack as various other BPF registers there)
and used in constant blinding for JITs-only so far. It can
be reused for verifier rewrites under certain constraints.
The interpreter's tmp "register" has therefore been remapped
into extending the register set with hidden ax register and
reusing that for a number of instructions that needed the
prior temporary variable internally (e.g. div, mod). This
allows for zero increase in stack space usage in the interpreter,
and enables (restricted) generic use in rewrites otherwise as
long as such a patchlet does not make use of these instructions.
The sanitation mask is dynamic and relative to the offset the
map value or stack pointer currently holds.
There are various cases that need to be taken under consideration
for the masking, e.g. such operation could look as follows:
ptr += val or val += ptr or ptr -= val. Thus, the value to be
sanitized could reside either in source or in destination
register, and the limit is different depending on whether
the ALU op is addition or subtraction and depending on the
current known and bounded offset. The limit is derived as
follows: limit := max_value_size - (smin_value + off). For
subtraction: limit := umax_value + off. This holds because
we do not allow any pointer arithmetic that would
temporarily go out of bounds or would have an unknown
value with mixed signed bounds where it is unclear at
verification time whether the actual runtime value would
be either negative or positive. For example, we have a
derived map pointer value with constant offset and bounded
one, so limit based on smin_value works because the verifier
requires that statically analyzed arithmetic on the pointer
must be in bounds, and thus it checks if resulting
smin_value + off and umax_value + off is still within map
value bounds at time of arithmetic in addition to time of
access. Similarly, for the case of stack access we derive
the limit as follows: MAX_BPF_STACK + off for subtraction
and -off for the case of addition where off := ptr_reg->off +
ptr_reg->var_off.value. Subtraction is a special case for
the masking which can be in form of ptr += -val, ptr -= -val,
or ptr -= val. In the first two cases where we know that
the value is negative, we need to temporarily negate the
value in order to do the sanitation on a positive value
where we later swap the ALU op, and restore original source
register if the value was in source.
The sanitation of pointer arithmetic alone is still not fully
sufficient as is, since a scenario like the following could
happen ...
PTR += 0x1000 (e.g. K-based imm)
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
PTR += 0x1000
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
[...]
... which under speculation could end up as ...
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
[...]
... and therefore still access out of bounds. To prevent such
case, the verifier is also analyzing safety for potential out
of bounds access under speculative execution. Meaning, it is
also simulating pointer access under truncation. We therefore
"branch off" and push the current verification state after the
ALU operation with known 0 to the verification stack for later
analysis. Given the current path analysis succeeded it is
likely that the one under speculation can be pruned. In any
case, it is also subject to existing complexity limits and
therefore anything beyond this point will be rejected. In
terms of pruning, it needs to be ensured that the verification
state from speculative execution simulation must never prune
a non-speculative execution path, therefore, we mark verifier
state accordingly at the time of push_stack(). If verifier
detects out of bounds access under speculative execution from
one of the possible paths that includes a truncation, it will
reject such program.
Given we mask every reg-based pointer arithmetic for
unprivileged programs, we've been looking into how it could
affect real-world programs in terms of size increase. As the
majority of programs are targeted for privileged-only use
case, we've unconditionally enabled masking (with its alu
restrictions on top of it) for privileged programs for the
sake of testing in order to check i) whether they get rejected
in its current form, and ii) by how much the number of
instructions and size will increase. We've tested this by
using Katran, Cilium and test_l4lb from the kernel selftests.
For Katran we've evaluated balancer_kern.o, Cilium bpf_lxc.o
and an older test object bpf_lxc_opt_-DUNKNOWN.o and l4lb
we've used test_l4lb.o as well as test_l4lb_noinline.o. We
found that none of the programs got rejected by the verifier
with this change, and that impact is rather minimal to none.
balancer_kern.o had 13,904 bytes (1,738 insns) xlated and
7,797 bytes JITed before and after the change. Most complex
program in bpf_lxc.o had 30,544 bytes (3,817 insns) xlated
and 18,538 bytes JITed before and after and none of the other
tail call programs in bpf_lxc.o had any changes either. For
the older bpf_lxc_opt_-DUNKNOWN.o object we found a small
increase from 20,616 bytes (2,576 insns) and 12,536 bytes JITed
before to 20,664 bytes (2,582 insns) and 12,558 bytes JITed
after the change. Other programs from that object file had
similar small increase. Both test_l4lb.o had no change and
remained at 6,544 bytes (817 insns) xlated and 3,401 bytes
JITed and for test_l4lb_noinline.o constant at 5,080 bytes
(634 insns) xlated and 3,313 bytes JITed. This can be explained
in that LLVM typically optimizes stack based pointer arithmetic
by using K-based operations and that use of dynamic map access
is not overly frequent. However, in future we may decide to
optimize the algorithm further under known guarantees from
branch and value speculation. Latter seems also unclear in
terms of prediction heuristics that today's CPUs apply as well
as whether there could be collisions in e.g. the predictor's
Value History/Pattern Table for triggering out of bounds access,
thus masking is performed unconditionally at this point but could
be subject to relaxation later on. We were generally also
brainstorming various other approaches for mitigation, but the
blocker was always lack of available registers at runtime and/or
overhead for runtime tracking of limits belonging to a specific
pointer. Thus, we found this to be minimally intrusive under
given constraints.
With that in place, a simple example with sanitized access on
unprivileged load at post-verification time looks as follows:
# bpftool prog dump xlated id 282
[...]
28: (79) r1 = *(u64 *)(r7 +0)
29: (79) r2 = *(u64 *)(r7 +8)
30: (57) r1 &= 15
31: (79) r3 = *(u64 *)(r0 +4608)
32: (57) r3 &= 1
33: (47) r3 |= 1
34: (2d) if r2 > r3 goto pc+19
35: (b4) (u32) r11 = (u32) 20479 |
36: (1f) r11 -= r2 | Dynamic sanitation for pointer
37: (4f) r11 |= r2 | arithmetic with registers
38: (87) r11 = -r11 | containing bounded or known
39: (c7) r11 s>>= 63 | scalars in order to prevent
40: (5f) r11 &= r2 | out of bounds speculation.
41: (0f) r4 += r11 |
42: (71) r4 = *(u8 *)(r4 +0)
43: (6f) r4 <<= r1
[...]
For the case where the scalar sits in the destination register
as opposed to the source register, the following code is emitted
for the above example:
[...]
16: (b4) (u32) r11 = (u32) 20479
17: (1f) r11 -= r2
18: (4f) r11 |= r2
19: (87) r11 = -r11
20: (c7) r11 s>>= 63
21: (5f) r2 &= r11
22: (0f) r2 += r0
23: (61) r0 = *(u32 *)(r2 +0)
[...]
JIT blinding example with non-conflicting use of r10:
[...]
d5: je 0x0000000000000106 _
d7: mov 0x0(%rax),%edi |
da: mov $0xf153246,%r10d | Index load from map value and
e0: xor $0xf153259,%r10 | (const blinded) mask with 0x1f.
e7: and %r10,%rdi |_
ea: mov $0x2f,%r10d |
f0: sub %rdi,%r10 | Sanitized addition. Both use r10
f3: or %rdi,%r10 | but do not interfere with each
f6: neg %r10 | other. (Neither do these instructions
f9: sar $0x3f,%r10 | interfere with the use of ax as temp
fd: and %r10,%rdi | in interpreter.)
100: add %rax,%rdi |_
103: mov 0x0(%rdi),%eax
[...]
Tested that it fixes Jann's reproducer, and also checked that test_verifier
and test_progs suite with interpreter, JIT and JIT with hardening enabled
on x86-64 and arm64 runs successfully.
[0] Speculose: Analyzing the Security Implications of Speculative
Execution in CPUs, Giorgi Maisuradze and Christian Rossow,
https://arxiv.org/pdf/1801.04084.pdf
[1] A Systematic Evaluation of Transient Execution Attacks and
Defenses, Claudio Canella, Jo Van Bulck, Michael Schwarz,
Moritz Lipp, Benjamin von Berg, Philipp Ortner, Frank Piessens,
Dmitry Evtyushkin, Daniel Gruss,
https://arxiv.org/pdf/1811.05441.pdf
Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-01-03 07:58:34 +08:00
|
|
|
bool speculative;
|
2017-12-15 09:55:06 +08:00
|
|
|
};
|
|
|
|
|
2018-10-03 04:35:29 +08:00
|
|
|
#define bpf_get_spilled_reg(slot, frame) \
|
|
|
|
(((slot < frame->allocated_stack / BPF_REG_SIZE) && \
|
|
|
|
(frame->stack[slot].slot_type[0] == STACK_SPILL)) \
|
|
|
|
? &frame->stack[slot].spilled_ptr : NULL)
|
|
|
|
|
|
|
|
/* Iterate over 'frame', setting 'reg' to either NULL or a spilled register. */
|
|
|
|
#define bpf_for_each_spilled_reg(iter, frame, reg) \
|
|
|
|
for (iter = 0, reg = bpf_get_spilled_reg(iter, frame); \
|
|
|
|
iter < frame->allocated_stack / BPF_REG_SIZE; \
|
|
|
|
iter++, reg = bpf_get_spilled_reg(iter, frame))
|
|
|
|
|
2016-09-21 18:43:57 +08:00
|
|
|
/* linked list of verifier states used to prune search */
|
|
|
|
struct bpf_verifier_state_list {
|
|
|
|
struct bpf_verifier_state state;
|
|
|
|
struct bpf_verifier_state_list *next;
|
|
|
|
};
|
|
|
|
|
bpf: prevent out of bounds speculation on pointer arithmetic
Jann reported that the original commit back in b2157399cc98
("bpf: prevent out-of-bounds speculation") was not sufficient
to stop CPU from speculating out of bounds memory access:
While b2157399cc98 only focussed on masking array map access
for unprivileged users for tail calls and data access such
that the user provided index gets sanitized from BPF program
and syscall side, there is still a more generic form affected
from BPF programs that applies to most maps that hold user
data in relation to dynamic map access when dealing with
unknown scalars or "slow" known scalars as access offset, for
example:
- Load a map value pointer into R6
- Load an index into R7
- Do a slow computation (e.g. with a memory dependency) that
loads a limit into R8 (e.g. load the limit from a map for
high latency, then mask it to make the verifier happy)
- Exit if R7 >= R8 (mispredicted branch)
- Load R0 = R6[R7]
- Load R0 = R6[R0]
For unknown scalars there are two options in the BPF verifier
where we could derive knowledge from in order to guarantee
safe access to the memory: i) While </>/<=/>= variants won't
allow to derive any lower or upper bounds from the unknown
scalar where it would be safe to add it to the map value
pointer, it is possible through ==/!= test however. ii) another
option is to transform the unknown scalar into a known scalar,
for example, through ALU ops combination such as R &= <imm>
followed by R |= <imm> or any similar combination where the
original information from the unknown scalar would be destroyed
entirely leaving R with a constant. The initial slow load still
precedes the latter ALU ops on that register, so the CPU
executes speculatively from that point. Once we have the known
scalar, any compare operation would work then. A third option
only involving registers with known scalars could be crafted
as described in [0] where a CPU port (e.g. Slow Int unit)
would be filled with many dependent computations such that
the subsequent condition depending on its outcome has to wait
for evaluation on its execution port and thereby executing
speculatively if the speculated code can be scheduled on a
different execution port, or any other form of mistraining
as described in [1], for example. Given this is not limited
to only unknown scalars, not only map but also stack access
is affected since both is accessible for unprivileged users
and could potentially be used for out of bounds access under
speculation.
In order to prevent any of these cases, the verifier is now
sanitizing pointer arithmetic on the offset such that any
out of bounds speculation would be masked in a way where the
pointer arithmetic result in the destination register will
stay unchanged, meaning offset masked into zero similar as
in array_index_nospec() case. With regards to implementation,
there are three options that were considered: i) new insn
for sanitation, ii) push/pop insn and sanitation as inlined
BPF, iii) reuse of ax register and sanitation as inlined BPF.
Option i) has the downside that we end up using from reserved
bits in the opcode space, but also that we would require
each JIT to emit masking as native arch opcodes meaning
mitigation would have slow adoption till everyone implements
it eventually which is counter-productive. Option ii) and iii)
have both in common that a temporary register is needed in
order to implement the sanitation as inlined BPF since we
are not allowed to modify the source register. While a push /
pop insn in ii) would be useful to have in any case, it
requires once again that every JIT needs to implement it
first. While possible, amount of changes needed would also
be unsuitable for a -stable patch. Therefore, the path which
has fewer changes, less BPF instructions for the mitigation
and does not require anything to be changed in the JITs is
option iii) which this work is pursuing. The ax register is
already mapped to a register in all JITs (modulo arm32 where
it's mapped to stack as various other BPF registers there)
and used in constant blinding for JITs-only so far. It can
be reused for verifier rewrites under certain constraints.
The interpreter's tmp "register" has therefore been remapped
into extending the register set with hidden ax register and
reusing that for a number of instructions that needed the
prior temporary variable internally (e.g. div, mod). This
allows for zero increase in stack space usage in the interpreter,
and enables (restricted) generic use in rewrites otherwise as
long as such a patchlet does not make use of these instructions.
The sanitation mask is dynamic and relative to the offset the
map value or stack pointer currently holds.
There are various cases that need to be taken under consideration
for the masking, e.g. such operation could look as follows:
ptr += val or val += ptr or ptr -= val. Thus, the value to be
sanitized could reside either in source or in destination
register, and the limit is different depending on whether
the ALU op is addition or subtraction and depending on the
current known and bounded offset. The limit is derived as
follows: limit := max_value_size - (smin_value + off). For
subtraction: limit := umax_value + off. This holds because
we do not allow any pointer arithmetic that would
temporarily go out of bounds or would have an unknown
value with mixed signed bounds where it is unclear at
verification time whether the actual runtime value would
be either negative or positive. For example, we have a
derived map pointer value with constant offset and bounded
one, so limit based on smin_value works because the verifier
requires that statically analyzed arithmetic on the pointer
must be in bounds, and thus it checks if resulting
smin_value + off and umax_value + off is still within map
value bounds at time of arithmetic in addition to time of
access. Similarly, for the case of stack access we derive
the limit as follows: MAX_BPF_STACK + off for subtraction
and -off for the case of addition where off := ptr_reg->off +
ptr_reg->var_off.value. Subtraction is a special case for
the masking which can be in form of ptr += -val, ptr -= -val,
or ptr -= val. In the first two cases where we know that
the value is negative, we need to temporarily negate the
value in order to do the sanitation on a positive value
where we later swap the ALU op, and restore original source
register if the value was in source.
The sanitation of pointer arithmetic alone is still not fully
sufficient as is, since a scenario like the following could
happen ...
PTR += 0x1000 (e.g. K-based imm)
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
PTR += 0x1000
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
[...]
... which under speculation could end up as ...
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
[...]
... and therefore still access out of bounds. To prevent such
case, the verifier is also analyzing safety for potential out
of bounds access under speculative execution. Meaning, it is
also simulating pointer access under truncation. We therefore
"branch off" and push the current verification state after the
ALU operation with known 0 to the verification stack for later
analysis. Given the current path analysis succeeded it is
likely that the one under speculation can be pruned. In any
case, it is also subject to existing complexity limits and
therefore anything beyond this point will be rejected. In
terms of pruning, it needs to be ensured that the verification
state from speculative execution simulation must never prune
a non-speculative execution path, therefore, we mark verifier
state accordingly at the time of push_stack(). If verifier
detects out of bounds access under speculative execution from
one of the possible paths that includes a truncation, it will
reject such program.
Given we mask every reg-based pointer arithmetic for
unprivileged programs, we've been looking into how it could
affect real-world programs in terms of size increase. As the
majority of programs are targeted for privileged-only use
case, we've unconditionally enabled masking (with its alu
restrictions on top of it) for privileged programs for the
sake of testing in order to check i) whether they get rejected
in its current form, and ii) by how much the number of
instructions and size will increase. We've tested this by
using Katran, Cilium and test_l4lb from the kernel selftests.
For Katran we've evaluated balancer_kern.o, Cilium bpf_lxc.o
and an older test object bpf_lxc_opt_-DUNKNOWN.o and l4lb
we've used test_l4lb.o as well as test_l4lb_noinline.o. We
found that none of the programs got rejected by the verifier
with this change, and that impact is rather minimal to none.
balancer_kern.o had 13,904 bytes (1,738 insns) xlated and
7,797 bytes JITed before and after the change. Most complex
program in bpf_lxc.o had 30,544 bytes (3,817 insns) xlated
and 18,538 bytes JITed before and after and none of the other
tail call programs in bpf_lxc.o had any changes either. For
the older bpf_lxc_opt_-DUNKNOWN.o object we found a small
increase from 20,616 bytes (2,576 insns) and 12,536 bytes JITed
before to 20,664 bytes (2,582 insns) and 12,558 bytes JITed
after the change. Other programs from that object file had
similar small increase. Both test_l4lb.o had no change and
remained at 6,544 bytes (817 insns) xlated and 3,401 bytes
JITed and for test_l4lb_noinline.o constant at 5,080 bytes
(634 insns) xlated and 3,313 bytes JITed. This can be explained
in that LLVM typically optimizes stack based pointer arithmetic
by using K-based operations and that use of dynamic map access
is not overly frequent. However, in future we may decide to
optimize the algorithm further under known guarantees from
branch and value speculation. Latter seems also unclear in
terms of prediction heuristics that today's CPUs apply as well
as whether there could be collisions in e.g. the predictor's
Value History/Pattern Table for triggering out of bounds access,
thus masking is performed unconditionally at this point but could
be subject to relaxation later on. We were generally also
brainstorming various other approaches for mitigation, but the
blocker was always lack of available registers at runtime and/or
overhead for runtime tracking of limits belonging to a specific
pointer. Thus, we found this to be minimally intrusive under
given constraints.
With that in place, a simple example with sanitized access on
unprivileged load at post-verification time looks as follows:
# bpftool prog dump xlated id 282
[...]
28: (79) r1 = *(u64 *)(r7 +0)
29: (79) r2 = *(u64 *)(r7 +8)
30: (57) r1 &= 15
31: (79) r3 = *(u64 *)(r0 +4608)
32: (57) r3 &= 1
33: (47) r3 |= 1
34: (2d) if r2 > r3 goto pc+19
35: (b4) (u32) r11 = (u32) 20479 |
36: (1f) r11 -= r2 | Dynamic sanitation for pointer
37: (4f) r11 |= r2 | arithmetic with registers
38: (87) r11 = -r11 | containing bounded or known
39: (c7) r11 s>>= 63 | scalars in order to prevent
40: (5f) r11 &= r2 | out of bounds speculation.
41: (0f) r4 += r11 |
42: (71) r4 = *(u8 *)(r4 +0)
43: (6f) r4 <<= r1
[...]
For the case where the scalar sits in the destination register
as opposed to the source register, the following code is emitted
for the above example:
[...]
16: (b4) (u32) r11 = (u32) 20479
17: (1f) r11 -= r2
18: (4f) r11 |= r2
19: (87) r11 = -r11
20: (c7) r11 s>>= 63
21: (5f) r2 &= r11
22: (0f) r2 += r0
23: (61) r0 = *(u32 *)(r2 +0)
[...]
JIT blinding example with non-conflicting use of r10:
[...]
d5: je 0x0000000000000106 _
d7: mov 0x0(%rax),%edi |
da: mov $0xf153246,%r10d | Index load from map value and
e0: xor $0xf153259,%r10 | (const blinded) mask with 0x1f.
e7: and %r10,%rdi |_
ea: mov $0x2f,%r10d |
f0: sub %rdi,%r10 | Sanitized addition. Both use r10
f3: or %rdi,%r10 | but do not interfere with each
f6: neg %r10 | other. (Neither do these instructions
f9: sar $0x3f,%r10 | interfere with the use of ax as temp
fd: and %r10,%rdi | in interpreter.)
100: add %rax,%rdi |_
103: mov 0x0(%rdi),%eax
[...]
Tested that it fixes Jann's reproducer, and also checked that test_verifier
and test_progs suite with interpreter, JIT and JIT with hardening enabled
on x86-64 and arm64 runs successfully.
[0] Speculose: Analyzing the Security Implications of Speculative
Execution in CPUs, Giorgi Maisuradze and Christian Rossow,
https://arxiv.org/pdf/1801.04084.pdf
[1] A Systematic Evaluation of Transient Execution Attacks and
Defenses, Claudio Canella, Jo Van Bulck, Michael Schwarz,
Moritz Lipp, Benjamin von Berg, Philipp Ortner, Frank Piessens,
Dmitry Evtyushkin, Daniel Gruss,
https://arxiv.org/pdf/1811.05441.pdf
Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-01-03 07:58:34 +08:00
|
|
|
/* Possible states for alu_state member. */
|
|
|
|
#define BPF_ALU_SANITIZE_SRC 1U
|
|
|
|
#define BPF_ALU_SANITIZE_DST 2U
|
|
|
|
#define BPF_ALU_NEG_VALUE (1U << 2)
|
2019-01-06 07:54:37 +08:00
|
|
|
#define BPF_ALU_NON_POINTER (1U << 3)
|
bpf: prevent out of bounds speculation on pointer arithmetic
Jann reported that the original commit back in b2157399cc98
("bpf: prevent out-of-bounds speculation") was not sufficient
to stop CPU from speculating out of bounds memory access:
While b2157399cc98 only focussed on masking array map access
for unprivileged users for tail calls and data access such
that the user provided index gets sanitized from BPF program
and syscall side, there is still a more generic form affected
from BPF programs that applies to most maps that hold user
data in relation to dynamic map access when dealing with
unknown scalars or "slow" known scalars as access offset, for
example:
- Load a map value pointer into R6
- Load an index into R7
- Do a slow computation (e.g. with a memory dependency) that
loads a limit into R8 (e.g. load the limit from a map for
high latency, then mask it to make the verifier happy)
- Exit if R7 >= R8 (mispredicted branch)
- Load R0 = R6[R7]
- Load R0 = R6[R0]
For unknown scalars there are two options in the BPF verifier
where we could derive knowledge from in order to guarantee
safe access to the memory: i) While </>/<=/>= variants won't
allow to derive any lower or upper bounds from the unknown
scalar where it would be safe to add it to the map value
pointer, it is possible through ==/!= test however. ii) another
option is to transform the unknown scalar into a known scalar,
for example, through ALU ops combination such as R &= <imm>
followed by R |= <imm> or any similar combination where the
original information from the unknown scalar would be destroyed
entirely leaving R with a constant. The initial slow load still
precedes the latter ALU ops on that register, so the CPU
executes speculatively from that point. Once we have the known
scalar, any compare operation would work then. A third option
only involving registers with known scalars could be crafted
as described in [0] where a CPU port (e.g. Slow Int unit)
would be filled with many dependent computations such that
the subsequent condition depending on its outcome has to wait
for evaluation on its execution port and thereby executing
speculatively if the speculated code can be scheduled on a
different execution port, or any other form of mistraining
as described in [1], for example. Given this is not limited
to only unknown scalars, not only map but also stack access
is affected since both is accessible for unprivileged users
and could potentially be used for out of bounds access under
speculation.
In order to prevent any of these cases, the verifier is now
sanitizing pointer arithmetic on the offset such that any
out of bounds speculation would be masked in a way where the
pointer arithmetic result in the destination register will
stay unchanged, meaning offset masked into zero similar as
in array_index_nospec() case. With regards to implementation,
there are three options that were considered: i) new insn
for sanitation, ii) push/pop insn and sanitation as inlined
BPF, iii) reuse of ax register and sanitation as inlined BPF.
Option i) has the downside that we end up using from reserved
bits in the opcode space, but also that we would require
each JIT to emit masking as native arch opcodes meaning
mitigation would have slow adoption till everyone implements
it eventually which is counter-productive. Option ii) and iii)
have both in common that a temporary register is needed in
order to implement the sanitation as inlined BPF since we
are not allowed to modify the source register. While a push /
pop insn in ii) would be useful to have in any case, it
requires once again that every JIT needs to implement it
first. While possible, amount of changes needed would also
be unsuitable for a -stable patch. Therefore, the path which
has fewer changes, less BPF instructions for the mitigation
and does not require anything to be changed in the JITs is
option iii) which this work is pursuing. The ax register is
already mapped to a register in all JITs (modulo arm32 where
it's mapped to stack as various other BPF registers there)
and used in constant blinding for JITs-only so far. It can
be reused for verifier rewrites under certain constraints.
The interpreter's tmp "register" has therefore been remapped
into extending the register set with hidden ax register and
reusing that for a number of instructions that needed the
prior temporary variable internally (e.g. div, mod). This
allows for zero increase in stack space usage in the interpreter,
and enables (restricted) generic use in rewrites otherwise as
long as such a patchlet does not make use of these instructions.
The sanitation mask is dynamic and relative to the offset the
map value or stack pointer currently holds.
There are various cases that need to be taken under consideration
for the masking, e.g. such operation could look as follows:
ptr += val or val += ptr or ptr -= val. Thus, the value to be
sanitized could reside either in source or in destination
register, and the limit is different depending on whether
the ALU op is addition or subtraction and depending on the
current known and bounded offset. The limit is derived as
follows: limit := max_value_size - (smin_value + off). For
subtraction: limit := umax_value + off. This holds because
we do not allow any pointer arithmetic that would
temporarily go out of bounds or would have an unknown
value with mixed signed bounds where it is unclear at
verification time whether the actual runtime value would
be either negative or positive. For example, we have a
derived map pointer value with constant offset and bounded
one, so limit based on smin_value works because the verifier
requires that statically analyzed arithmetic on the pointer
must be in bounds, and thus it checks if resulting
smin_value + off and umax_value + off is still within map
value bounds at time of arithmetic in addition to time of
access. Similarly, for the case of stack access we derive
the limit as follows: MAX_BPF_STACK + off for subtraction
and -off for the case of addition where off := ptr_reg->off +
ptr_reg->var_off.value. Subtraction is a special case for
the masking which can be in form of ptr += -val, ptr -= -val,
or ptr -= val. In the first two cases where we know that
the value is negative, we need to temporarily negate the
value in order to do the sanitation on a positive value
where we later swap the ALU op, and restore original source
register if the value was in source.
The sanitation of pointer arithmetic alone is still not fully
sufficient as is, since a scenario like the following could
happen ...
PTR += 0x1000 (e.g. K-based imm)
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
PTR += 0x1000
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
[...]
... which under speculation could end up as ...
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
[...]
... and therefore still access out of bounds. To prevent such
case, the verifier is also analyzing safety for potential out
of bounds access under speculative execution. Meaning, it is
also simulating pointer access under truncation. We therefore
"branch off" and push the current verification state after the
ALU operation with known 0 to the verification stack for later
analysis. Given the current path analysis succeeded it is
likely that the one under speculation can be pruned. In any
case, it is also subject to existing complexity limits and
therefore anything beyond this point will be rejected. In
terms of pruning, it needs to be ensured that the verification
state from speculative execution simulation must never prune
a non-speculative execution path, therefore, we mark verifier
state accordingly at the time of push_stack(). If verifier
detects out of bounds access under speculative execution from
one of the possible paths that includes a truncation, it will
reject such program.
Given we mask every reg-based pointer arithmetic for
unprivileged programs, we've been looking into how it could
affect real-world programs in terms of size increase. As the
majority of programs are targeted for privileged-only use
case, we've unconditionally enabled masking (with its alu
restrictions on top of it) for privileged programs for the
sake of testing in order to check i) whether they get rejected
in its current form, and ii) by how much the number of
instructions and size will increase. We've tested this by
using Katran, Cilium and test_l4lb from the kernel selftests.
For Katran we've evaluated balancer_kern.o, Cilium bpf_lxc.o
and an older test object bpf_lxc_opt_-DUNKNOWN.o and l4lb
we've used test_l4lb.o as well as test_l4lb_noinline.o. We
found that none of the programs got rejected by the verifier
with this change, and that impact is rather minimal to none.
balancer_kern.o had 13,904 bytes (1,738 insns) xlated and
7,797 bytes JITed before and after the change. Most complex
program in bpf_lxc.o had 30,544 bytes (3,817 insns) xlated
and 18,538 bytes JITed before and after and none of the other
tail call programs in bpf_lxc.o had any changes either. For
the older bpf_lxc_opt_-DUNKNOWN.o object we found a small
increase from 20,616 bytes (2,576 insns) and 12,536 bytes JITed
before to 20,664 bytes (2,582 insns) and 12,558 bytes JITed
after the change. Other programs from that object file had
similar small increase. Both test_l4lb.o had no change and
remained at 6,544 bytes (817 insns) xlated and 3,401 bytes
JITed and for test_l4lb_noinline.o constant at 5,080 bytes
(634 insns) xlated and 3,313 bytes JITed. This can be explained
in that LLVM typically optimizes stack based pointer arithmetic
by using K-based operations and that use of dynamic map access
is not overly frequent. However, in future we may decide to
optimize the algorithm further under known guarantees from
branch and value speculation. Latter seems also unclear in
terms of prediction heuristics that today's CPUs apply as well
as whether there could be collisions in e.g. the predictor's
Value History/Pattern Table for triggering out of bounds access,
thus masking is performed unconditionally at this point but could
be subject to relaxation later on. We were generally also
brainstorming various other approaches for mitigation, but the
blocker was always lack of available registers at runtime and/or
overhead for runtime tracking of limits belonging to a specific
pointer. Thus, we found this to be minimally intrusive under
given constraints.
With that in place, a simple example with sanitized access on
unprivileged load at post-verification time looks as follows:
# bpftool prog dump xlated id 282
[...]
28: (79) r1 = *(u64 *)(r7 +0)
29: (79) r2 = *(u64 *)(r7 +8)
30: (57) r1 &= 15
31: (79) r3 = *(u64 *)(r0 +4608)
32: (57) r3 &= 1
33: (47) r3 |= 1
34: (2d) if r2 > r3 goto pc+19
35: (b4) (u32) r11 = (u32) 20479 |
36: (1f) r11 -= r2 | Dynamic sanitation for pointer
37: (4f) r11 |= r2 | arithmetic with registers
38: (87) r11 = -r11 | containing bounded or known
39: (c7) r11 s>>= 63 | scalars in order to prevent
40: (5f) r11 &= r2 | out of bounds speculation.
41: (0f) r4 += r11 |
42: (71) r4 = *(u8 *)(r4 +0)
43: (6f) r4 <<= r1
[...]
For the case where the scalar sits in the destination register
as opposed to the source register, the following code is emitted
for the above example:
[...]
16: (b4) (u32) r11 = (u32) 20479
17: (1f) r11 -= r2
18: (4f) r11 |= r2
19: (87) r11 = -r11
20: (c7) r11 s>>= 63
21: (5f) r2 &= r11
22: (0f) r2 += r0
23: (61) r0 = *(u32 *)(r2 +0)
[...]
JIT blinding example with non-conflicting use of r10:
[...]
d5: je 0x0000000000000106 _
d7: mov 0x0(%rax),%edi |
da: mov $0xf153246,%r10d | Index load from map value and
e0: xor $0xf153259,%r10 | (const blinded) mask with 0x1f.
e7: and %r10,%rdi |_
ea: mov $0x2f,%r10d |
f0: sub %rdi,%r10 | Sanitized addition. Both use r10
f3: or %rdi,%r10 | but do not interfere with each
f6: neg %r10 | other. (Neither do these instructions
f9: sar $0x3f,%r10 | interfere with the use of ax as temp
fd: and %r10,%rdi | in interpreter.)
100: add %rax,%rdi |_
103: mov 0x0(%rdi),%eax
[...]
Tested that it fixes Jann's reproducer, and also checked that test_verifier
and test_progs suite with interpreter, JIT and JIT with hardening enabled
on x86-64 and arm64 runs successfully.
[0] Speculose: Analyzing the Security Implications of Speculative
Execution in CPUs, Giorgi Maisuradze and Christian Rossow,
https://arxiv.org/pdf/1801.04084.pdf
[1] A Systematic Evaluation of Transient Execution Attacks and
Defenses, Claudio Canella, Jo Van Bulck, Michael Schwarz,
Moritz Lipp, Benjamin von Berg, Philipp Ortner, Frank Piessens,
Dmitry Evtyushkin, Daniel Gruss,
https://arxiv.org/pdf/1811.05441.pdf
Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-01-03 07:58:34 +08:00
|
|
|
#define BPF_ALU_SANITIZE (BPF_ALU_SANITIZE_SRC | \
|
|
|
|
BPF_ALU_SANITIZE_DST)
|
|
|
|
|
2016-09-21 18:43:57 +08:00
|
|
|
struct bpf_insn_aux_data {
|
2017-03-16 09:26:42 +08:00
|
|
|
union {
|
|
|
|
enum bpf_reg_type ptr_type; /* pointer type for load/store insns */
|
bpf: properly enforce index mask to prevent out-of-bounds speculation
While reviewing the verifier code, I recently noticed that the
following two program variants in relation to tail calls can be
loaded.
Variant 1:
# bpftool p d x i 15
0: (15) if r1 == 0x0 goto pc+3
1: (18) r2 = map[id:5]
3: (05) goto pc+2
4: (18) r2 = map[id:6]
6: (b7) r3 = 7
7: (35) if r3 >= 0xa0 goto pc+2
8: (54) (u32) r3 &= (u32) 255
9: (85) call bpf_tail_call#12
10: (b7) r0 = 1
11: (95) exit
# bpftool m s i 5
5: prog_array flags 0x0
key 4B value 4B max_entries 4 memlock 4096B
# bpftool m s i 6
6: prog_array flags 0x0
key 4B value 4B max_entries 160 memlock 4096B
Variant 2:
# bpftool p d x i 20
0: (15) if r1 == 0x0 goto pc+3
1: (18) r2 = map[id:8]
3: (05) goto pc+2
4: (18) r2 = map[id:7]
6: (b7) r3 = 7
7: (35) if r3 >= 0x4 goto pc+2
8: (54) (u32) r3 &= (u32) 3
9: (85) call bpf_tail_call#12
10: (b7) r0 = 1
11: (95) exit
# bpftool m s i 8
8: prog_array flags 0x0
key 4B value 4B max_entries 160 memlock 4096B
# bpftool m s i 7
7: prog_array flags 0x0
key 4B value 4B max_entries 4 memlock 4096B
In both cases the index masking inserted by the verifier in order
to control out of bounds speculation from a CPU via b2157399cc98
("bpf: prevent out-of-bounds speculation") seems to be incorrect
in what it is enforcing. In the 1st variant, the mask is applied
from the map with the significantly larger number of entries where
we would allow to a certain degree out of bounds speculation for
the smaller map, and in the 2nd variant where the mask is applied
from the map with the smaller number of entries, we get buggy
behavior since we truncate the index of the larger map.
The original intent from commit b2157399cc98 is to reject such
occasions where two or more different tail call maps are used
in the same tail call helper invocation. However, the check on
the BPF_MAP_PTR_POISON is never hit since we never poisoned the
saved pointer in the first place! We do this explicitly for map
lookups but in case of tail calls we basically used the tail
call map in insn_aux_data that was processed in the most recent
path which the verifier walked. Thus any prior path that stored
a pointer in insn_aux_data at the helper location was always
overridden.
Fix it by moving the map pointer poison logic into a small helper
that covers both BPF helpers with the same logic. After that in
fixup_bpf_calls() the poison check is then hit for tail calls
and the program rejected. Latter only happens in unprivileged
case since this is the *only* occasion where a rewrite needs to
happen, and where such rewrite is specific to the map (max_entries,
index_mask). In the privileged case the rewrite is generic for
the insn->imm / insn->code update so multiple maps from different
paths can be handled just fine since all the remaining logic
happens in the instruction processing itself. This is similar
to the case of map lookups: in case there is a collision of
maps in fixup_bpf_calls() we must skip the inlined rewrite since
this will turn the generic instruction sequence into a non-
generic one. Thus the patch_call_imm will simply update the
insn->imm location where the bpf_map_lookup_elem() will later
take care of the dispatch. Given we need this 'poison' state
as a check, the information of whether a map is an unpriv_array
gets lost, so enforcing it prior to that needs an additional
state. In general this check is needed since there are some
complex and tail call intensive BPF programs out there where
LLVM tends to generate such code occasionally. We therefore
convert the map_ptr rather into map_state to store all this
w/o extra memory overhead, and the bit whether one of the maps
involved in the collision was from an unpriv_array thus needs
to be retained as well there.
Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2018-05-24 08:32:53 +08:00
|
|
|
unsigned long map_state; /* pointer/poison value for maps */
|
2017-12-15 09:55:15 +08:00
|
|
|
s32 call_imm; /* saved imm field of call insn */
|
bpf: prevent out of bounds speculation on pointer arithmetic
Jann reported that the original commit back in b2157399cc98
("bpf: prevent out-of-bounds speculation") was not sufficient
to stop CPU from speculating out of bounds memory access:
While b2157399cc98 only focussed on masking array map access
for unprivileged users for tail calls and data access such
that the user provided index gets sanitized from BPF program
and syscall side, there is still a more generic form affected
from BPF programs that applies to most maps that hold user
data in relation to dynamic map access when dealing with
unknown scalars or "slow" known scalars as access offset, for
example:
- Load a map value pointer into R6
- Load an index into R7
- Do a slow computation (e.g. with a memory dependency) that
loads a limit into R8 (e.g. load the limit from a map for
high latency, then mask it to make the verifier happy)
- Exit if R7 >= R8 (mispredicted branch)
- Load R0 = R6[R7]
- Load R0 = R6[R0]
For unknown scalars there are two options in the BPF verifier
where we could derive knowledge from in order to guarantee
safe access to the memory: i) While </>/<=/>= variants won't
allow to derive any lower or upper bounds from the unknown
scalar where it would be safe to add it to the map value
pointer, it is possible through ==/!= test however. ii) another
option is to transform the unknown scalar into a known scalar,
for example, through ALU ops combination such as R &= <imm>
followed by R |= <imm> or any similar combination where the
original information from the unknown scalar would be destroyed
entirely leaving R with a constant. The initial slow load still
precedes the latter ALU ops on that register, so the CPU
executes speculatively from that point. Once we have the known
scalar, any compare operation would work then. A third option
only involving registers with known scalars could be crafted
as described in [0] where a CPU port (e.g. Slow Int unit)
would be filled with many dependent computations such that
the subsequent condition depending on its outcome has to wait
for evaluation on its execution port and thereby executing
speculatively if the speculated code can be scheduled on a
different execution port, or any other form of mistraining
as described in [1], for example. Given this is not limited
to only unknown scalars, not only map but also stack access
is affected since both is accessible for unprivileged users
and could potentially be used for out of bounds access under
speculation.
In order to prevent any of these cases, the verifier is now
sanitizing pointer arithmetic on the offset such that any
out of bounds speculation would be masked in a way where the
pointer arithmetic result in the destination register will
stay unchanged, meaning offset masked into zero similar as
in array_index_nospec() case. With regards to implementation,
there are three options that were considered: i) new insn
for sanitation, ii) push/pop insn and sanitation as inlined
BPF, iii) reuse of ax register and sanitation as inlined BPF.
Option i) has the downside that we end up using from reserved
bits in the opcode space, but also that we would require
each JIT to emit masking as native arch opcodes meaning
mitigation would have slow adoption till everyone implements
it eventually which is counter-productive. Option ii) and iii)
have both in common that a temporary register is needed in
order to implement the sanitation as inlined BPF since we
are not allowed to modify the source register. While a push /
pop insn in ii) would be useful to have in any case, it
requires once again that every JIT needs to implement it
first. While possible, amount of changes needed would also
be unsuitable for a -stable patch. Therefore, the path which
has fewer changes, less BPF instructions for the mitigation
and does not require anything to be changed in the JITs is
option iii) which this work is pursuing. The ax register is
already mapped to a register in all JITs (modulo arm32 where
it's mapped to stack as various other BPF registers there)
and used in constant blinding for JITs-only so far. It can
be reused for verifier rewrites under certain constraints.
The interpreter's tmp "register" has therefore been remapped
into extending the register set with hidden ax register and
reusing that for a number of instructions that needed the
prior temporary variable internally (e.g. div, mod). This
allows for zero increase in stack space usage in the interpreter,
and enables (restricted) generic use in rewrites otherwise as
long as such a patchlet does not make use of these instructions.
The sanitation mask is dynamic and relative to the offset the
map value or stack pointer currently holds.
There are various cases that need to be taken under consideration
for the masking, e.g. such operation could look as follows:
ptr += val or val += ptr or ptr -= val. Thus, the value to be
sanitized could reside either in source or in destination
register, and the limit is different depending on whether
the ALU op is addition or subtraction and depending on the
current known and bounded offset. The limit is derived as
follows: limit := max_value_size - (smin_value + off). For
subtraction: limit := umax_value + off. This holds because
we do not allow any pointer arithmetic that would
temporarily go out of bounds or would have an unknown
value with mixed signed bounds where it is unclear at
verification time whether the actual runtime value would
be either negative or positive. For example, we have a
derived map pointer value with constant offset and bounded
one, so limit based on smin_value works because the verifier
requires that statically analyzed arithmetic on the pointer
must be in bounds, and thus it checks if resulting
smin_value + off and umax_value + off is still within map
value bounds at time of arithmetic in addition to time of
access. Similarly, for the case of stack access we derive
the limit as follows: MAX_BPF_STACK + off for subtraction
and -off for the case of addition where off := ptr_reg->off +
ptr_reg->var_off.value. Subtraction is a special case for
the masking which can be in form of ptr += -val, ptr -= -val,
or ptr -= val. In the first two cases where we know that
the value is negative, we need to temporarily negate the
value in order to do the sanitation on a positive value
where we later swap the ALU op, and restore original source
register if the value was in source.
The sanitation of pointer arithmetic alone is still not fully
sufficient as is, since a scenario like the following could
happen ...
PTR += 0x1000 (e.g. K-based imm)
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
PTR += 0x1000
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
[...]
... which under speculation could end up as ...
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
[...]
... and therefore still access out of bounds. To prevent such
case, the verifier is also analyzing safety for potential out
of bounds access under speculative execution. Meaning, it is
also simulating pointer access under truncation. We therefore
"branch off" and push the current verification state after the
ALU operation with known 0 to the verification stack for later
analysis. Given the current path analysis succeeded it is
likely that the one under speculation can be pruned. In any
case, it is also subject to existing complexity limits and
therefore anything beyond this point will be rejected. In
terms of pruning, it needs to be ensured that the verification
state from speculative execution simulation must never prune
a non-speculative execution path, therefore, we mark verifier
state accordingly at the time of push_stack(). If verifier
detects out of bounds access under speculative execution from
one of the possible paths that includes a truncation, it will
reject such program.
Given we mask every reg-based pointer arithmetic for
unprivileged programs, we've been looking into how it could
affect real-world programs in terms of size increase. As the
majority of programs are targeted for privileged-only use
case, we've unconditionally enabled masking (with its alu
restrictions on top of it) for privileged programs for the
sake of testing in order to check i) whether they get rejected
in its current form, and ii) by how much the number of
instructions and size will increase. We've tested this by
using Katran, Cilium and test_l4lb from the kernel selftests.
For Katran we've evaluated balancer_kern.o, Cilium bpf_lxc.o
and an older test object bpf_lxc_opt_-DUNKNOWN.o and l4lb
we've used test_l4lb.o as well as test_l4lb_noinline.o. We
found that none of the programs got rejected by the verifier
with this change, and that impact is rather minimal to none.
balancer_kern.o had 13,904 bytes (1,738 insns) xlated and
7,797 bytes JITed before and after the change. Most complex
program in bpf_lxc.o had 30,544 bytes (3,817 insns) xlated
and 18,538 bytes JITed before and after and none of the other
tail call programs in bpf_lxc.o had any changes either. For
the older bpf_lxc_opt_-DUNKNOWN.o object we found a small
increase from 20,616 bytes (2,576 insns) and 12,536 bytes JITed
before to 20,664 bytes (2,582 insns) and 12,558 bytes JITed
after the change. Other programs from that object file had
similar small increase. Both test_l4lb.o had no change and
remained at 6,544 bytes (817 insns) xlated and 3,401 bytes
JITed and for test_l4lb_noinline.o constant at 5,080 bytes
(634 insns) xlated and 3,313 bytes JITed. This can be explained
in that LLVM typically optimizes stack based pointer arithmetic
by using K-based operations and that use of dynamic map access
is not overly frequent. However, in future we may decide to
optimize the algorithm further under known guarantees from
branch and value speculation. Latter seems also unclear in
terms of prediction heuristics that today's CPUs apply as well
as whether there could be collisions in e.g. the predictor's
Value History/Pattern Table for triggering out of bounds access,
thus masking is performed unconditionally at this point but could
be subject to relaxation later on. We were generally also
brainstorming various other approaches for mitigation, but the
blocker was always lack of available registers at runtime and/or
overhead for runtime tracking of limits belonging to a specific
pointer. Thus, we found this to be minimally intrusive under
given constraints.
With that in place, a simple example with sanitized access on
unprivileged load at post-verification time looks as follows:
# bpftool prog dump xlated id 282
[...]
28: (79) r1 = *(u64 *)(r7 +0)
29: (79) r2 = *(u64 *)(r7 +8)
30: (57) r1 &= 15
31: (79) r3 = *(u64 *)(r0 +4608)
32: (57) r3 &= 1
33: (47) r3 |= 1
34: (2d) if r2 > r3 goto pc+19
35: (b4) (u32) r11 = (u32) 20479 |
36: (1f) r11 -= r2 | Dynamic sanitation for pointer
37: (4f) r11 |= r2 | arithmetic with registers
38: (87) r11 = -r11 | containing bounded or known
39: (c7) r11 s>>= 63 | scalars in order to prevent
40: (5f) r11 &= r2 | out of bounds speculation.
41: (0f) r4 += r11 |
42: (71) r4 = *(u8 *)(r4 +0)
43: (6f) r4 <<= r1
[...]
For the case where the scalar sits in the destination register
as opposed to the source register, the following code is emitted
for the above example:
[...]
16: (b4) (u32) r11 = (u32) 20479
17: (1f) r11 -= r2
18: (4f) r11 |= r2
19: (87) r11 = -r11
20: (c7) r11 s>>= 63
21: (5f) r2 &= r11
22: (0f) r2 += r0
23: (61) r0 = *(u32 *)(r2 +0)
[...]
JIT blinding example with non-conflicting use of r10:
[...]
d5: je 0x0000000000000106 _
d7: mov 0x0(%rax),%edi |
da: mov $0xf153246,%r10d | Index load from map value and
e0: xor $0xf153259,%r10 | (const blinded) mask with 0x1f.
e7: and %r10,%rdi |_
ea: mov $0x2f,%r10d |
f0: sub %rdi,%r10 | Sanitized addition. Both use r10
f3: or %rdi,%r10 | but do not interfere with each
f6: neg %r10 | other. (Neither do these instructions
f9: sar $0x3f,%r10 | interfere with the use of ax as temp
fd: and %r10,%rdi | in interpreter.)
100: add %rax,%rdi |_
103: mov 0x0(%rdi),%eax
[...]
Tested that it fixes Jann's reproducer, and also checked that test_verifier
and test_progs suite with interpreter, JIT and JIT with hardening enabled
on x86-64 and arm64 runs successfully.
[0] Speculose: Analyzing the Security Implications of Speculative
Execution in CPUs, Giorgi Maisuradze and Christian Rossow,
https://arxiv.org/pdf/1801.04084.pdf
[1] A Systematic Evaluation of Transient Execution Attacks and
Defenses, Claudio Canella, Jo Van Bulck, Michael Schwarz,
Moritz Lipp, Benjamin von Berg, Philipp Ortner, Frank Piessens,
Dmitry Evtyushkin, Daniel Gruss,
https://arxiv.org/pdf/1811.05441.pdf
Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-01-03 07:58:34 +08:00
|
|
|
u32 alu_limit; /* limit for add/sub register with pointer */
|
2017-03-16 09:26:42 +08:00
|
|
|
};
|
2017-06-23 06:07:39 +08:00
|
|
|
int ctx_field_size; /* the ctx field size for load insn, maybe 0 */
|
2018-05-16 00:27:05 +08:00
|
|
|
int sanitize_stack_off; /* stack slot to be cleared */
|
2017-11-23 08:42:05 +08:00
|
|
|
bool seen; /* this insn was processed by the verifier */
|
bpf: prevent out of bounds speculation on pointer arithmetic
Jann reported that the original commit back in b2157399cc98
("bpf: prevent out-of-bounds speculation") was not sufficient
to stop CPU from speculating out of bounds memory access:
While b2157399cc98 only focussed on masking array map access
for unprivileged users for tail calls and data access such
that the user provided index gets sanitized from BPF program
and syscall side, there is still a more generic form affected
from BPF programs that applies to most maps that hold user
data in relation to dynamic map access when dealing with
unknown scalars or "slow" known scalars as access offset, for
example:
- Load a map value pointer into R6
- Load an index into R7
- Do a slow computation (e.g. with a memory dependency) that
loads a limit into R8 (e.g. load the limit from a map for
high latency, then mask it to make the verifier happy)
- Exit if R7 >= R8 (mispredicted branch)
- Load R0 = R6[R7]
- Load R0 = R6[R0]
For unknown scalars there are two options in the BPF verifier
where we could derive knowledge from in order to guarantee
safe access to the memory: i) While </>/<=/>= variants won't
allow to derive any lower or upper bounds from the unknown
scalar where it would be safe to add it to the map value
pointer, it is possible through ==/!= test however. ii) another
option is to transform the unknown scalar into a known scalar,
for example, through ALU ops combination such as R &= <imm>
followed by R |= <imm> or any similar combination where the
original information from the unknown scalar would be destroyed
entirely leaving R with a constant. The initial slow load still
precedes the latter ALU ops on that register, so the CPU
executes speculatively from that point. Once we have the known
scalar, any compare operation would work then. A third option
only involving registers with known scalars could be crafted
as described in [0] where a CPU port (e.g. Slow Int unit)
would be filled with many dependent computations such that
the subsequent condition depending on its outcome has to wait
for evaluation on its execution port and thereby executing
speculatively if the speculated code can be scheduled on a
different execution port, or any other form of mistraining
as described in [1], for example. Given this is not limited
to only unknown scalars, not only map but also stack access
is affected since both is accessible for unprivileged users
and could potentially be used for out of bounds access under
speculation.
In order to prevent any of these cases, the verifier is now
sanitizing pointer arithmetic on the offset such that any
out of bounds speculation would be masked in a way where the
pointer arithmetic result in the destination register will
stay unchanged, meaning offset masked into zero similar as
in array_index_nospec() case. With regards to implementation,
there are three options that were considered: i) new insn
for sanitation, ii) push/pop insn and sanitation as inlined
BPF, iii) reuse of ax register and sanitation as inlined BPF.
Option i) has the downside that we end up using from reserved
bits in the opcode space, but also that we would require
each JIT to emit masking as native arch opcodes meaning
mitigation would have slow adoption till everyone implements
it eventually which is counter-productive. Option ii) and iii)
have both in common that a temporary register is needed in
order to implement the sanitation as inlined BPF since we
are not allowed to modify the source register. While a push /
pop insn in ii) would be useful to have in any case, it
requires once again that every JIT needs to implement it
first. While possible, amount of changes needed would also
be unsuitable for a -stable patch. Therefore, the path which
has fewer changes, less BPF instructions for the mitigation
and does not require anything to be changed in the JITs is
option iii) which this work is pursuing. The ax register is
already mapped to a register in all JITs (modulo arm32 where
it's mapped to stack as various other BPF registers there)
and used in constant blinding for JITs-only so far. It can
be reused for verifier rewrites under certain constraints.
The interpreter's tmp "register" has therefore been remapped
into extending the register set with hidden ax register and
reusing that for a number of instructions that needed the
prior temporary variable internally (e.g. div, mod). This
allows for zero increase in stack space usage in the interpreter,
and enables (restricted) generic use in rewrites otherwise as
long as such a patchlet does not make use of these instructions.
The sanitation mask is dynamic and relative to the offset the
map value or stack pointer currently holds.
There are various cases that need to be taken under consideration
for the masking, e.g. such operation could look as follows:
ptr += val or val += ptr or ptr -= val. Thus, the value to be
sanitized could reside either in source or in destination
register, and the limit is different depending on whether
the ALU op is addition or subtraction and depending on the
current known and bounded offset. The limit is derived as
follows: limit := max_value_size - (smin_value + off). For
subtraction: limit := umax_value + off. This holds because
we do not allow any pointer arithmetic that would
temporarily go out of bounds or would have an unknown
value with mixed signed bounds where it is unclear at
verification time whether the actual runtime value would
be either negative or positive. For example, we have a
derived map pointer value with constant offset and bounded
one, so limit based on smin_value works because the verifier
requires that statically analyzed arithmetic on the pointer
must be in bounds, and thus it checks if resulting
smin_value + off and umax_value + off is still within map
value bounds at time of arithmetic in addition to time of
access. Similarly, for the case of stack access we derive
the limit as follows: MAX_BPF_STACK + off for subtraction
and -off for the case of addition where off := ptr_reg->off +
ptr_reg->var_off.value. Subtraction is a special case for
the masking which can be in form of ptr += -val, ptr -= -val,
or ptr -= val. In the first two cases where we know that
the value is negative, we need to temporarily negate the
value in order to do the sanitation on a positive value
where we later swap the ALU op, and restore original source
register if the value was in source.
The sanitation of pointer arithmetic alone is still not fully
sufficient as is, since a scenario like the following could
happen ...
PTR += 0x1000 (e.g. K-based imm)
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
PTR += 0x1000
PTR -= BIG_NUMBER_WITH_SLOW_COMPARISON
[...]
... which under speculation could end up as ...
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
PTR += 0x1000
PTR -= 0 [ truncated by mitigation ]
[...]
... and therefore still access out of bounds. To prevent such
case, the verifier is also analyzing safety for potential out
of bounds access under speculative execution. Meaning, it is
also simulating pointer access under truncation. We therefore
"branch off" and push the current verification state after the
ALU operation with known 0 to the verification stack for later
analysis. Given the current path analysis succeeded it is
likely that the one under speculation can be pruned. In any
case, it is also subject to existing complexity limits and
therefore anything beyond this point will be rejected. In
terms of pruning, it needs to be ensured that the verification
state from speculative execution simulation must never prune
a non-speculative execution path, therefore, we mark verifier
state accordingly at the time of push_stack(). If verifier
detects out of bounds access under speculative execution from
one of the possible paths that includes a truncation, it will
reject such program.
Given we mask every reg-based pointer arithmetic for
unprivileged programs, we've been looking into how it could
affect real-world programs in terms of size increase. As the
majority of programs are targeted for privileged-only use
case, we've unconditionally enabled masking (with its alu
restrictions on top of it) for privileged programs for the
sake of testing in order to check i) whether they get rejected
in its current form, and ii) by how much the number of
instructions and size will increase. We've tested this by
using Katran, Cilium and test_l4lb from the kernel selftests.
For Katran we've evaluated balancer_kern.o, Cilium bpf_lxc.o
and an older test object bpf_lxc_opt_-DUNKNOWN.o and l4lb
we've used test_l4lb.o as well as test_l4lb_noinline.o. We
found that none of the programs got rejected by the verifier
with this change, and that impact is rather minimal to none.
balancer_kern.o had 13,904 bytes (1,738 insns) xlated and
7,797 bytes JITed before and after the change. Most complex
program in bpf_lxc.o had 30,544 bytes (3,817 insns) xlated
and 18,538 bytes JITed before and after and none of the other
tail call programs in bpf_lxc.o had any changes either. For
the older bpf_lxc_opt_-DUNKNOWN.o object we found a small
increase from 20,616 bytes (2,576 insns) and 12,536 bytes JITed
before to 20,664 bytes (2,582 insns) and 12,558 bytes JITed
after the change. Other programs from that object file had
similar small increase. Both test_l4lb.o had no change and
remained at 6,544 bytes (817 insns) xlated and 3,401 bytes
JITed and for test_l4lb_noinline.o constant at 5,080 bytes
(634 insns) xlated and 3,313 bytes JITed. This can be explained
in that LLVM typically optimizes stack based pointer arithmetic
by using K-based operations and that use of dynamic map access
is not overly frequent. However, in future we may decide to
optimize the algorithm further under known guarantees from
branch and value speculation. Latter seems also unclear in
terms of prediction heuristics that today's CPUs apply as well
as whether there could be collisions in e.g. the predictor's
Value History/Pattern Table for triggering out of bounds access,
thus masking is performed unconditionally at this point but could
be subject to relaxation later on. We were generally also
brainstorming various other approaches for mitigation, but the
blocker was always lack of available registers at runtime and/or
overhead for runtime tracking of limits belonging to a specific
pointer. Thus, we found this to be minimally intrusive under
given constraints.
With that in place, a simple example with sanitized access on
unprivileged load at post-verification time looks as follows:
# bpftool prog dump xlated id 282
[...]
28: (79) r1 = *(u64 *)(r7 +0)
29: (79) r2 = *(u64 *)(r7 +8)
30: (57) r1 &= 15
31: (79) r3 = *(u64 *)(r0 +4608)
32: (57) r3 &= 1
33: (47) r3 |= 1
34: (2d) if r2 > r3 goto pc+19
35: (b4) (u32) r11 = (u32) 20479 |
36: (1f) r11 -= r2 | Dynamic sanitation for pointer
37: (4f) r11 |= r2 | arithmetic with registers
38: (87) r11 = -r11 | containing bounded or known
39: (c7) r11 s>>= 63 | scalars in order to prevent
40: (5f) r11 &= r2 | out of bounds speculation.
41: (0f) r4 += r11 |
42: (71) r4 = *(u8 *)(r4 +0)
43: (6f) r4 <<= r1
[...]
For the case where the scalar sits in the destination register
as opposed to the source register, the following code is emitted
for the above example:
[...]
16: (b4) (u32) r11 = (u32) 20479
17: (1f) r11 -= r2
18: (4f) r11 |= r2
19: (87) r11 = -r11
20: (c7) r11 s>>= 63
21: (5f) r2 &= r11
22: (0f) r2 += r0
23: (61) r0 = *(u32 *)(r2 +0)
[...]
JIT blinding example with non-conflicting use of r10:
[...]
d5: je 0x0000000000000106 _
d7: mov 0x0(%rax),%edi |
da: mov $0xf153246,%r10d | Index load from map value and
e0: xor $0xf153259,%r10 | (const blinded) mask with 0x1f.
e7: and %r10,%rdi |_
ea: mov $0x2f,%r10d |
f0: sub %rdi,%r10 | Sanitized addition. Both use r10
f3: or %rdi,%r10 | but do not interfere with each
f6: neg %r10 | other. (Neither do these instructions
f9: sar $0x3f,%r10 | interfere with the use of ax as temp
fd: and %r10,%rdi | in interpreter.)
100: add %rax,%rdi |_
103: mov 0x0(%rdi),%eax
[...]
Tested that it fixes Jann's reproducer, and also checked that test_verifier
and test_progs suite with interpreter, JIT and JIT with hardening enabled
on x86-64 and arm64 runs successfully.
[0] Speculose: Analyzing the Security Implications of Speculative
Execution in CPUs, Giorgi Maisuradze and Christian Rossow,
https://arxiv.org/pdf/1801.04084.pdf
[1] A Systematic Evaluation of Transient Execution Attacks and
Defenses, Claudio Canella, Jo Van Bulck, Michael Schwarz,
Moritz Lipp, Benjamin von Berg, Philipp Ortner, Frank Piessens,
Dmitry Evtyushkin, Daniel Gruss,
https://arxiv.org/pdf/1811.05441.pdf
Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2019-01-03 07:58:34 +08:00
|
|
|
u8 alu_state; /* used in combination with alu_limit */
|
2019-01-23 14:45:23 +08:00
|
|
|
unsigned int orig_idx; /* original instruction index */
|
2016-09-21 18:43:57 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
#define MAX_USED_MAPS 64 /* max number of maps accessed by one eBPF program */
|
|
|
|
|
2017-10-10 01:30:15 +08:00
|
|
|
#define BPF_VERIFIER_TMP_LOG_SIZE 1024
|
|
|
|
|
2018-03-25 02:44:22 +08:00
|
|
|
struct bpf_verifier_log {
|
2017-10-10 01:30:10 +08:00
|
|
|
u32 level;
|
2017-10-10 01:30:15 +08:00
|
|
|
char kbuf[BPF_VERIFIER_TMP_LOG_SIZE];
|
2017-10-10 01:30:10 +08:00
|
|
|
char __user *ubuf;
|
|
|
|
u32 len_used;
|
|
|
|
u32 len_total;
|
|
|
|
};
|
|
|
|
|
2018-03-25 02:44:22 +08:00
|
|
|
static inline bool bpf_verifier_log_full(const struct bpf_verifier_log *log)
|
2017-10-10 01:30:10 +08:00
|
|
|
{
|
|
|
|
return log->len_used >= log->len_total - 1;
|
|
|
|
}
|
|
|
|
|
2018-03-25 02:44:23 +08:00
|
|
|
static inline bool bpf_verifier_log_needed(const struct bpf_verifier_log *log)
|
|
|
|
{
|
|
|
|
return log->level && log->ubuf && !bpf_verifier_log_full(log);
|
|
|
|
}
|
|
|
|
|
2017-12-15 09:55:05 +08:00
|
|
|
#define BPF_MAX_SUBPROGS 256
|
|
|
|
|
2018-05-03 04:17:18 +08:00
|
|
|
struct bpf_subprog_info {
|
|
|
|
u32 start; /* insn idx of function entry point */
|
2018-12-08 08:42:25 +08:00
|
|
|
u32 linfo_idx; /* The idx to the main_prog->aux->linfo */
|
2018-05-03 04:17:18 +08:00
|
|
|
u16 stack_depth; /* max. stack depth used by this function */
|
|
|
|
};
|
|
|
|
|
2016-09-21 18:43:57 +08:00
|
|
|
/* single container for all structs
|
|
|
|
* one verifier_env per bpf_check() call
|
|
|
|
*/
|
|
|
|
struct bpf_verifier_env {
|
2019-01-03 07:58:27 +08:00
|
|
|
u32 insn_idx;
|
|
|
|
u32 prev_insn_idx;
|
2016-09-21 18:43:57 +08:00
|
|
|
struct bpf_prog *prog; /* eBPF program being verified */
|
2017-10-17 07:40:54 +08:00
|
|
|
const struct bpf_verifier_ops *ops;
|
2016-09-21 18:43:57 +08:00
|
|
|
struct bpf_verifier_stack_elem *head; /* stack of verifier states to be processed */
|
|
|
|
int stack_size; /* number of states to be processed */
|
2017-05-11 02:38:07 +08:00
|
|
|
bool strict_alignment; /* perform strict pointer alignment checks */
|
2017-11-01 09:16:05 +08:00
|
|
|
struct bpf_verifier_state *cur_state; /* current verifier state */
|
2016-09-21 18:43:57 +08:00
|
|
|
struct bpf_verifier_state_list **explored_states; /* search pruning optimization */
|
|
|
|
struct bpf_map *used_maps[MAX_USED_MAPS]; /* array of map's used by eBPF program */
|
|
|
|
u32 used_map_cnt; /* number of used maps */
|
|
|
|
u32 id_gen; /* used to generate unique reg IDs */
|
|
|
|
bool allow_ptr_leaks;
|
|
|
|
bool seen_direct_write;
|
|
|
|
struct bpf_insn_aux_data *insn_aux_data; /* array of per-insn state */
|
2018-12-14 02:41:48 +08:00
|
|
|
const struct bpf_line_info *prev_linfo;
|
2018-03-25 02:44:22 +08:00
|
|
|
struct bpf_verifier_log log;
|
2018-05-03 04:17:18 +08:00
|
|
|
struct bpf_subprog_info subprog_info[BPF_MAX_SUBPROGS + 1];
|
2017-12-15 09:55:05 +08:00
|
|
|
u32 subprog_cnt;
|
2016-09-21 18:43:57 +08:00
|
|
|
};
|
|
|
|
|
2018-05-17 04:27:41 +08:00
|
|
|
__printf(2, 0) void bpf_verifier_vlog(struct bpf_verifier_log *log,
|
|
|
|
const char *fmt, va_list args);
|
2018-01-10 20:26:06 +08:00
|
|
|
__printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
|
|
|
|
const char *fmt, ...);
|
|
|
|
|
2018-10-03 04:35:35 +08:00
|
|
|
static inline struct bpf_func_state *cur_func(struct bpf_verifier_env *env)
|
2017-11-01 09:16:05 +08:00
|
|
|
{
|
2017-12-15 09:55:06 +08:00
|
|
|
struct bpf_verifier_state *cur = env->cur_state;
|
|
|
|
|
2018-10-03 04:35:35 +08:00
|
|
|
return cur->frame[cur->curframe];
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct bpf_reg_state *cur_regs(struct bpf_verifier_env *env)
|
|
|
|
{
|
|
|
|
return cur_func(env)->regs;
|
2017-11-01 09:16:05 +08:00
|
|
|
}
|
|
|
|
|
2018-11-09 21:03:31 +08:00
|
|
|
int bpf_prog_offload_verifier_prep(struct bpf_prog *prog);
|
2017-12-28 10:39:05 +08:00
|
|
|
int bpf_prog_offload_verify_insn(struct bpf_verifier_env *env,
|
|
|
|
int insn_idx, int prev_insn_idx);
|
2018-10-07 19:56:47 +08:00
|
|
|
int bpf_prog_offload_finalize(struct bpf_verifier_env *env);
|
2019-01-23 14:45:24 +08:00
|
|
|
void
|
|
|
|
bpf_prog_offload_replace_insn(struct bpf_verifier_env *env, u32 off,
|
|
|
|
struct bpf_insn *insn);
|
|
|
|
void
|
|
|
|
bpf_prog_offload_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt);
|
2017-11-04 04:56:17 +08:00
|
|
|
|
2016-09-21 18:43:57 +08:00
|
|
|
#endif /* _LINUX_BPF_VERIFIER_H */
|