On Tue, 3 Jun 2025 at 23:04, Luis Gerhorst [off-list ref] wrote:

This improves the expressiveness of unprivileged BPF by inserting
speculation barriers instead of rejecting the programs.

The approach was previously presented at LPC'24 [1] and RAID'24 [2].

To mitigate the Spectre v1 (PHT) vulnerability, the kernel rejects
potentially-dangerous unprivileged BPF programs as of
commit 9183671af6db ("bpf: Fix leakage under speculation on mispredicted
branches"). In [2], we have analyzed 364 object files from open source
projects (Linux Samples and Selftests, BCC, Loxilb, Cilium, libbpf
Examples, Parca, and Prevail) and found that this affects 31% to 54% of
programs.

To resolve this in the majority of cases this patchset adds a fall-back
for mitigating Spectre v1 using speculation barriers. The kernel still
optimistically attempts to verify all speculative paths but uses
speculation barriers against v1 when unsafe behavior is detected. This
allows for more programs to be accepted without disabling the BPF
Spectre mitigations (e.g., by setting cpu_mitigations_off()).

For this, it relies on the fact that speculation barriers generally
prevent all later instructions from executing if the speculation was not
correct (not only loads). See patch 7 ("bpf: Fall back to nospec for
Spectre v1") for a detailed description and references to the relevant
vendor documentation (AMD and Intel x86-64, ARM64, and PowerPC).

In [1] we have measured the overhead of this approach relative to having
mitigations off and including the upstream Spectre v4 mitigations. For
event tracing and stack-sampling profilers, we found that mitigations
increase BPF program execution time by 0% to 62%. For the Loxilb network
load balancer, we have measured a 14% slowdown in SCTP performance but
no significant slowdown for TCP. This overhead only applies to programs
that were previously rejected.

I reran the expressiveness-evaluation with v6.14 and made sure the main
results still match those from [1] and [2] (which used v6.5).

Main design decisions are:

* Do not use separate bytecode insns for v1 and v4 barriers (inspired by
  Daniel Borkmann's question at LPC). This simplifies the verifier
  significantly and has the only downside that performance on PowerPC is
  not as high as it could be.

* Allow archs to still disable v1/v4 mitigations separately by setting
  bpf_jit_bypass_spec_v1/v4(). This has the benefit that archs can
  benefit from improved BPF expressiveness / performance if they are not
  vulnerable (e.g., ARM64 for v4 in the kernel).

* Do not remove the empty BPF_NOSPEC implementation for backends for
  which it is unknown whether they are vulnerable to Spectre v1.

[1] https://lpc.events/event/18/contributions/1954/ ("Mitigating
    Spectre-PHT using Speculation Barriers in Linux eBPF")
[2] https://arxiv.org/pdf/2405.00078 ("VeriFence: Lightweight and
    Precise Spectre Defenses for Untrusted Linux Kernel Extensions")

Changes:

* v3 -> v4:
  - Remove insn parameter from do_check_insn() and extract
    process_bpf_exit_full as a function as requested by Eduard
  - Investigate apparent sanitize_check_bounds() bug reported by
    Kartikeya (does appear to not be a bug but only confusing code),
    sent separate patch to document it and add an assert
  - Remove already-merged commit 1 ("selftests/bpf: Fix caps for
    __xlated/jited_unpriv")
  - Drop former commit 10 ("bpf: Allow nospec-protected var-offset stack
    access") as it did not include a test and there are other places
    where var-off is rejected. Also, none of the tested real-world
    programs used var-off in the paper. Therefore keep the old behavior
    for now and potentially prepare a patch that converts all cases
    later if required.
  - Add link to AMD lfence and PowerPC speculation barrier (ori 31,31,0)
    documentation
  - Move detailed barrier documentation to commit 7 ("bpf: Fall back to
    nospec for Spectre v1")
  - Link to v3: https://lore.kernel.org/all/20250501073603.1402960-1-luis.gerhorst@fau.de/ (local)

LGTM. For the set,

Acked-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>

[...]