user/sven/linux.git/kernel/bpf, branch v4.20.16

bpf: Fix syscall's stackmap lookup potential deadlock

2019-03-13T21:04:14Z

[ Upstream commit 7c4cd051add3d00bbff008a133c936c515eaa8fe ] The map_lookup_elem used to not acquiring spinlock in order to optimize the reader. It was true until commit 557c0c6e7df8 ("bpf: convert stackmap to pre-allocation") The syscall's map_lookup_elem(stackmap) calls bpf_stackmap_copy(). bpf_stackmap_copy() may find the elem no longer needed after the copy is done. If that is the case, pcpu_freelist_push() saves this elem for reuse later. This push requires a spinlock. If a tracing bpf_prog got run in the middle of the syscall's map_lookup_elem(stackmap) and this tracing bpf_prog is calling bpf_get_stackid(stackmap) which also requires the same pcpu_freelist's spinlock, it may end up with a dead lock situation as reported by Eric Dumazet in https://patchwork.ozlabs.org/patch/1030266/ The situation is the same as the syscall's map_update_elem() which needs to acquire the pcpu_freelist's spinlock and could race with tracing bpf_prog. Hence, this patch fixes it by protecting bpf_stackmap_copy() with this_cpu_inc(bpf_prog_active) to prevent tracing bpf_prog from running. A later syscall's map_lookup_elem commit f1a2e44a3aec ("bpf: add queue and stack maps") also acquires a spinlock and races with tracing bpf_prog similarly. Hence, this patch is forward looking and protects the majority of the map lookups. bpf_map_offload_lookup_elem() is the exception since it is for network bpf_prog only (i.e. never called by tracing bpf_prog). Fixes: 557c0c6e7df8 ("bpf: convert stackmap to pre-allocation") Reported-by: Eric Dumazet Acked-by: Alexei Starovoitov Signed-off-by: Martin KaFai Lau Signed-off-by: Alexei Starovoitov Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: fix lockdep false positive in percpu_freelist

2019-03-13T21:04:13Z

[ Upstream commit a89fac57b5d080771efd4d71feaae19877cf68f0 ] Lockdep warns about false positive: [ 12.492084] 00000000e6b28347 (&head->lock){+...}, at: pcpu_freelist_push+0x2a/0x40 [ 12.492696] but this lock was taken by another, HARDIRQ-safe lock in the past: [ 12.493275] (&rq->lock){-.-.} [ 12.493276] [ 12.493276] [ 12.493276] and interrupts could create inverse lock ordering between them. [ 12.493276] [ 12.494435] [ 12.494435] other info that might help us debug this: [ 12.494979] Possible interrupt unsafe locking scenario: [ 12.494979] [ 12.495518] CPU0 CPU1 [ 12.495879] ---- ---- [ 12.496243] lock(&head->lock); [ 12.496502] local_irq_disable(); [ 12.496969] lock(&rq->lock); [ 12.497431] lock(&head->lock); [ 12.497890] [ 12.498104] lock(&rq->lock); [ 12.498368] [ 12.498368] *** DEADLOCK *** [ 12.498368] [ 12.498837] 1 lock held by dd/276: [ 12.499110] #0: 00000000c58cb2ee (rcu_read_lock){....}, at: trace_call_bpf+0x5e/0x240 [ 12.499747] [ 12.499747] the shortest dependencies between 2nd lock and 1st lock: [ 12.500389] -> (&rq->lock){-.-.} { [ 12.500669] IN-HARDIRQ-W at: [ 12.500934] _raw_spin_lock+0x2f/0x40 [ 12.501373] scheduler_tick+0x4c/0xf0 [ 12.501812] update_process_times+0x40/0x50 [ 12.502294] tick_periodic+0x27/0xb0 [ 12.502723] tick_handle_periodic+0x1f/0x60 [ 12.503203] timer_interrupt+0x11/0x20 [ 12.503651] __handle_irq_event_percpu+0x43/0x2c0 [ 12.504167] handle_irq_event_percpu+0x20/0x50 [ 12.504674] handle_irq_event+0x37/0x60 [ 12.505139] handle_level_irq+0xa7/0x120 [ 12.505601] handle_irq+0xa1/0x150 [ 12.506018] do_IRQ+0x77/0x140 [ 12.506411] ret_from_intr+0x0/0x1d [ 12.506834] _raw_spin_unlock_irqrestore+0x53/0x60 [ 12.507362] __setup_irq+0x481/0x730 [ 12.507789] setup_irq+0x49/0x80 [ 12.508195] hpet_time_init+0x21/0x32 [ 12.508644] x86_late_time_init+0xb/0x16 [ 12.509106] start_kernel+0x390/0x42a [ 12.509554] secondary_startup_64+0xa4/0xb0 [ 12.510034] IN-SOFTIRQ-W at: [ 12.510305] _raw_spin_lock+0x2f/0x40 [ 12.510772] try_to_wake_up+0x1c7/0x4e0 [ 12.511220] swake_up_locked+0x20/0x40 [ 12.511657] swake_up_one+0x1a/0x30 [ 12.512070] rcu_process_callbacks+0xc5/0x650 [ 12.512553] __do_softirq+0xe6/0x47b [ 12.512978] irq_exit+0xc3/0xd0 [ 12.513372] smp_apic_timer_interrupt+0xa9/0x250 [ 12.513876] apic_timer_interrupt+0xf/0x20 [ 12.514343] default_idle+0x1c/0x170 [ 12.514765] do_idle+0x199/0x240 [ 12.515159] cpu_startup_entry+0x19/0x20 [ 12.515614] start_kernel+0x422/0x42a [ 12.516045] secondary_startup_64+0xa4/0xb0 [ 12.516521] INITIAL USE at: [ 12.516774] _raw_spin_lock_irqsave+0x38/0x50 [ 12.517258] rq_attach_root+0x16/0xd0 [ 12.517685] sched_init+0x2f2/0x3eb [ 12.518096] start_kernel+0x1fb/0x42a [ 12.518525] secondary_startup_64+0xa4/0xb0 [ 12.518986] } [ 12.519132] ... key at: [] __key.71384+0x0/0x8 [ 12.519649] ... acquired at: [ 12.519892] pcpu_freelist_pop+0x7b/0xd0 [ 12.520221] bpf_get_stackid+0x1d2/0x4d0 [ 12.520563] ___bpf_prog_run+0x8b4/0x11a0 [ 12.520887] [ 12.521008] -> (&head->lock){+...} { [ 12.521292] HARDIRQ-ON-W at: [ 12.521539] _raw_spin_lock+0x2f/0x40 [ 12.521950] pcpu_freelist_push+0x2a/0x40 [ 12.522396] bpf_get_stackid+0x494/0x4d0 [ 12.522828] ___bpf_prog_run+0x8b4/0x11a0 [ 12.523296] INITIAL USE at: [ 12.523537] _raw_spin_lock+0x2f/0x40 [ 12.523944] pcpu_freelist_populate+0xc0/0x120 [ 12.524417] htab_map_alloc+0x405/0x500 [ 12.524835] __do_sys_bpf+0x1a3/0x1a90 [ 12.525253] do_syscall_64+0x4a/0x180 [ 12.525659] entry_SYSCALL_64_after_hwframe+0x49/0xbe [ 12.526167] } [ 12.526311] ... key at: [] __key.13130+0x0/0x8 [ 12.526812] ... acquired at: [ 12.527047] __lock_acquire+0x521/0x1350 [ 12.527371] lock_acquire+0x98/0x190 [ 12.527680] _raw_spin_lock+0x2f/0x40 [ 12.527994] pcpu_freelist_push+0x2a/0x40 [ 12.528325] bpf_get_stackid+0x494/0x4d0 [ 12.528645] ___bpf_prog_run+0x8b4/0x11a0 [ 12.528970] [ 12.529092] [ 12.529092] stack backtrace: [ 12.529444] CPU: 0 PID: 276 Comm: dd Not tainted 5.0.0-rc3-00018-g2fa53f892422 #475 [ 12.530043] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-2.el7 04/01/2014 [ 12.530750] Call Trace: [ 12.530948] dump_stack+0x5f/0x8b [ 12.531248] check_usage_backwards+0x10c/0x120 [ 12.531598] ? ___bpf_prog_run+0x8b4/0x11a0 [ 12.531935] ? mark_lock+0x382/0x560 [ 12.532229] mark_lock+0x382/0x560 [ 12.532496] ? print_shortest_lock_dependencies+0x180/0x180 [ 12.532928] __lock_acquire+0x521/0x1350 [ 12.533271] ? find_get_entry+0x17f/0x2e0 [ 12.533586] ? find_get_entry+0x19c/0x2e0 [ 12.533902] ? lock_acquire+0x98/0x190 [ 12.534196] lock_acquire+0x98/0x190 [ 12.534482] ? pcpu_freelist_push+0x2a/0x40 [ 12.534810] _raw_spin_lock+0x2f/0x40 [ 12.535099] ? pcpu_freelist_push+0x2a/0x40 [ 12.535432] pcpu_freelist_push+0x2a/0x40 [ 12.535750] bpf_get_stackid+0x494/0x4d0 [ 12.536062] ___bpf_prog_run+0x8b4/0x11a0 It has been explained that is a false positive here: https://lkml.org/lkml/2018/7/25/756 Recap: - stackmap uses pcpu_freelist - The lock in pcpu_freelist is a percpu lock - stackmap is only used by tracing bpf_prog - A tracing bpf_prog cannot be run if another bpf_prog has already been running (ensured by the percpu bpf_prog_active counter). Eric pointed out that this lockdep splats stops other legit lockdep splats in selftests/bpf/test_progs.c. Fix this by calling local_irq_save/restore for stackmap. Another false positive had also been worked around by calling local_irq_save in commit 89ad2fa3f043 ("bpf: fix lockdep splat"). That commit added unnecessary irq_save/restore to fast path of bpf hash map. irqs are already disabled at that point, since htab is holding per bucket spin_lock with irqsave. Let's reduce overhead for htab by introducing __pcpu_freelist_push/pop function w/o irqsave and convert pcpu_freelist_push/pop to irqsave to be used elsewhere (right now only in stackmap). It stops lockdep false positive in stackmap with a bit of acceptable overhead. Fixes: 557c0c6e7df8 ("bpf: convert stackmap to pre-allocation") Reported-by: Naresh Kamboju Reported-by: Eric Dumazet Acked-by: Martin KaFai Lau Signed-off-by: Alexei Starovoitov Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: run bpf programs with preemption disabled

2019-03-13T21:04:13Z

[ Upstream commit 6cab5e90ab2bd323c9f3811b6c70a4687df51e27 ] Disabled preemption is necessary for proper access to per-cpu maps from BPF programs. But the sender side of socket filters didn't have preemption disabled: unix_dgram_sendmsg->sk_filter->sk_filter_trim_cap->bpf_prog_run_save_cb->BPF_PROG_RUN and a combination of af_packet with tun device didn't disable either: tpacket_snd->packet_direct_xmit->packet_pick_tx_queue->ndo_select_queue-> tun_select_queue->tun_ebpf_select_queue->bpf_prog_run_clear_cb->BPF_PROG_RUN Disable preemption before executing BPF programs (both classic and extended). Reported-by: Jann Horn Signed-off-by: Alexei Starovoitov Acked-by: Song Liu Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: fix sanitation rewrite in case of non-pointers

2019-03-10T06:10:16Z

commit 3612af783cf52c74a031a2f11b82247b2599d3cd upstream. Marek reported that he saw an issue with the below snippet in that timing measurements where off when loaded as unpriv while results were reasonable when loaded as privileged: [...] uint64_t a = bpf_ktime_get_ns(); uint64_t b = bpf_ktime_get_ns(); uint64_t delta = b - a; if ((int64_t)delta > 0) { [...] Turns out there is a bug where a corner case is missing in the fix d3bd7413e0ca ("bpf: fix sanitation of alu op with pointer / scalar type from different paths"), namely fixup_bpf_calls() only checks whether aux has a non-zero alu_state, but it also needs to test for the case of BPF_ALU_NON_POINTER since in both occasions we need to skip the masking rewrite (as there is nothing to mask). Fixes: d3bd7413e0ca ("bpf: fix sanitation of alu op with pointer / scalar type from different paths") Reported-by: Marek Majkowski Reported-by: Arthur Fabre Signed-off-by: Daniel Borkmann Link: https://lore.kernel.org/netdev/CAJPywTJqP34cK20iLM5YmUMz9KXQOdu1-+BZrGMAGgLuBWz7fg@mail.gmail.com/T/ Acked-by: Song Liu Signed-off-by: Alexei Starovoitov Signed-off-by: Greg Kroah-Hartman

bpf: zero out build_id for BPF_STACK_BUILD_ID_IP

2019-02-27T09:09:50Z

[ Upstream commit 4af396ae4836c4ecab61e975b8e61270c551894d ] When returning BPF_STACK_BUILD_ID_IP from stack_map_get_build_id_offset, make sure that build_id field is empty. Since we are using percpu free list, there is a possibility that we might reuse some previous bpf_stack_build_id with non-zero build_id. Fixes: 615755a77b24 ("bpf: extend stackmap to save binary_build_id+offset instead of address") Acked-by: Song Liu Signed-off-by: Stanislav Fomichev Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: don't assume build-id length is always 20 bytes

2019-02-27T09:09:50Z

[ Upstream commit 0b698005a9d11c0e91141ec11a2c4918a129f703 ] Build-id length is not fixed to 20, it can be (`man ld` /--build-id): * 128-bit (uuid) * 160-bit (sha1) * any length specified in ld --build-id=0xhexstring To fix the issue of missing BPF_STACK_BUILD_ID_VALID for shorter build-ids, assume that build-id is somewhere in the range of 1 .. 20. Set the remaining bytes to zero. v2: * don't introduce new "len = min(BPF_BUILD_ID_SIZE, nhdr->n_descsz)", we already know that nhdr->n_descsz <= BPF_BUILD_ID_SIZE if we enter this 'if' condition Fixes: 615755a77b24 ("bpf: extend stackmap to save binary_build_id+offset instead of address") Acked-by: Song Liu Signed-off-by: Stanislav Fomichev Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: fix panic in stack_map_get_build_id() on i386 and arm32

2019-02-27T09:09:46Z

[ Upstream commit beaf3d1901f4ea46fbd5c9d857227d99751de469 ] As Naresh reported, test_stacktrace_build_id() causes panic on i386 and arm32 systems. This is caused by page_address() returns NULL in certain cases. This patch fixes this error by using kmap_atomic/kunmap_atomic instead of page_address. Fixes: 615755a77b24 (" bpf: extend stackmap to save binary_build_id+offset instead of address") Reported-by: Naresh Kamboju Signed-off-by: Song Liu Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: fix inner map masking to prevent oob under speculation

2019-01-31T07:15:45Z

[ commit 9d5564ddcf2a0f5ba3fa1c3a1f8a1b59ad309553 upstream ] During review I noticed that inner meta map setup for map in map is buggy in that it does not propagate all needed data from the reference map which the verifier is later accessing. In particular one such case is index masking to prevent out of bounds access under speculative execution due to missing the map's unpriv_array/index_mask field propagation. Fix this such that the verifier is generating the correct code for inlined lookups in case of unpriviledged use. Before patch (test_verifier's 'map in map access' dump): # bpftool prog dump xla id 3 0: (62) *(u32 *)(r10 -4) = 0 1: (bf) r2 = r10 2: (07) r2 += -4 3: (18) r1 = map[id:4] 5: (07) r1 += 272 | 6: (61) r0 = *(u32 *)(r2 +0) | 7: (35) if r0 >= 0x1 goto pc+6 | Inlined map in map lookup 8: (54) (u32) r0 &= (u32) 0 | with index masking for 9: (67) r0 <<= 3 | map->unpriv_array. 10: (0f) r0 += r1 | 11: (79) r0 = *(u64 *)(r0 +0) | 12: (15) if r0 == 0x0 goto pc+1 | 13: (05) goto pc+1 | 14: (b7) r0 = 0 | 15: (15) if r0 == 0x0 goto pc+11 16: (62) *(u32 *)(r10 -4) = 0 17: (bf) r2 = r10 18: (07) r2 += -4 19: (bf) r1 = r0 20: (07) r1 += 272 | 21: (61) r0 = *(u32 *)(r2 +0) | Index masking missing (!) 22: (35) if r0 >= 0x1 goto pc+3 | for inner map despite 23: (67) r0 <<= 3 | map->unpriv_array set. 24: (0f) r0 += r1 | 25: (05) goto pc+1 | 26: (b7) r0 = 0 | 27: (b7) r0 = 0 28: (95) exit After patch: # bpftool prog dump xla id 1 0: (62) *(u32 *)(r10 -4) = 0 1: (bf) r2 = r10 2: (07) r2 += -4 3: (18) r1 = map[id:2] 5: (07) r1 += 272 | 6: (61) r0 = *(u32 *)(r2 +0) | 7: (35) if r0 >= 0x1 goto pc+6 | Same inlined map in map lookup 8: (54) (u32) r0 &= (u32) 0 | with index masking due to 9: (67) r0 <<= 3 | map->unpriv_array. 10: (0f) r0 += r1 | 11: (79) r0 = *(u64 *)(r0 +0) | 12: (15) if r0 == 0x0 goto pc+1 | 13: (05) goto pc+1 | 14: (b7) r0 = 0 | 15: (15) if r0 == 0x0 goto pc+12 16: (62) *(u32 *)(r10 -4) = 0 17: (bf) r2 = r10 18: (07) r2 += -4 19: (bf) r1 = r0 20: (07) r1 += 272 | 21: (61) r0 = *(u32 *)(r2 +0) | 22: (35) if r0 >= 0x1 goto pc+4 | Now fixed inlined inner map 23: (54) (u32) r0 &= (u32) 0 | lookup with proper index masking 24: (67) r0 <<= 3 | for map->unpriv_array. 25: (0f) r0 += r1 | 26: (05) goto pc+1 | 27: (b7) r0 = 0 | 28: (b7) r0 = 0 29: (95) exit Fixes: b2157399cc98 ("bpf: prevent out-of-bounds speculation") Signed-off-by: Daniel Borkmann Acked-by: Martin KaFai Lau Signed-off-by: Alexei Starovoitov Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: fix sanitation of alu op with pointer / scalar type from different paths

2019-01-31T07:15:45Z

[ commit d3bd7413e0ca40b60cf60d4003246d067cafdeda upstream ] While 979d63d50c0c ("bpf: prevent out of bounds speculation on pointer arithmetic") took care of rejecting alu op on pointer when e.g. pointer came from two different map values with different map properties such as value size, Jann reported that a case was not covered yet when a given alu op is used in both "ptr_reg += reg" and "numeric_reg += reg" from different branches where we would incorrectly try to sanitize based on the pointer's limit. Catch this corner case and reject the program instead. Fixes: 979d63d50c0c ("bpf: prevent out of bounds speculation on pointer arithmetic") Reported-by: Jann Horn Signed-off-by: Daniel Borkmann Acked-by: Alexei Starovoitov Signed-off-by: Alexei Starovoitov Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin

bpf: prevent out of bounds speculation on pointer arithmetic

2019-01-31T07:15:45Z

[ commit 979d63d50c0c0f7bc537bf821e056cc9fe5abd38 upstream ] 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 &= followed by R |= 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 Signed-off-by: Daniel Borkmann Acked-by: Alexei Starovoitov Signed-off-by: Alexei Starovoitov Signed-off-by: Daniel Borkmann Signed-off-by: Sasha Levin