20260821-linuxkm-WC_SVR_USE_NATIVE_REG_BUFS by douzzer · Pull Request #11240 · wolfSSL/wolfssl

douzzer · 2026-08-21T20:39:54Z

linuxkm/x86_vector_register_glue.c: add WC_SVR_USE_NATIVE_REG_BUFS: per-CPU
native vector register save buffers, serving hardirq and softirq contexts
where kernel_fpu_begin() is unusable.

previously these contexts were refused (WC_ACCEL_INHIBIT_E): hardirq
outright, softirq whenever !may_use_simd() (pre-6.15, a softirq
interrupting a foreign kernel_fpu section). an exact save/compute/restore
cycle is invisible to the interrupted context and to kernel FPU
bookkeeping, and run-to-completion freezes the blocking condition beneath
the section, so both classes are natively serviceable.

per-CPU {softirq, hardirq} state, each {depth, pin_preempt, save_area};
xsave64/xrstor64 with RFBM 0xe7 & XCR0 (fxsave64 fallback), plain XSAVE
only; feature detection by direct CPUID/XGETBV, independent of kernel
cpufeature machinery.
recursion is a per-(CPU, class) depth count -- save on the 0->1 edge,
restore on 1->0; softirq sections hold preempt_disable() (load-bearing
on PREEMPT_RT).
hardirq serves plain saves; softirq serves plain and _MAYBE_INHIBIT
(its legacy short-circuit is bypassed when native is ready); _INHIBIT,
hardirq _MAYBE, and NMI keep refusal semantics; task context is
untouched.
wc_svr_native_init() activates in place from
wc_linuxkm_allocate_svr_states() in any context but NMI -- the FIPS
pre-Init lazy allocation arrives inside the outermost save's own
bracket, and wolfCrypt_Init() has been observed with irqs disabled. a
self-bracketed register round-trip selftest gates wc_svr_native_ready;
on any failure the long-standing refusal semantics remain fully in
force.
wc_linuxkm_svr_native_is_ready(): runtime readiness accessor for policy
decisions elsewhere (e.g. the FIPS<v7 default-bank NO_VECTOR_OPS pin).
WOLFSSL_LINUXKM_VERBOSE_DEBUG: split the softirq WC_ACCEL_INHIBIT_E
counter into per-class counters (softirq, hardirq, NMI, other),
reported at module shutdown.

in non-VERBOSE builds with WC_SVR_DONT_USE_NATIVE_REG_BUFS (or in non-SVR
builds), generated object code is byte-identical to the previous revision.

also add linuxkm/simd_hammer/, a stress testing tool for kernel FPU/SIMD context handling.

tested with

wolfssl-multi-test.sh ...
check-source-text
'.*dev.*cust-kern.*'
'linuxkm-legacy-.*'
quantum-safe-wolfssl-all-crypto-only-intelasm-sp-asm-sp-fips-dev-linuxkm-next-insmod
quantum-safe-wolfssl-all-crypto-only-intelasm-sp-asm-fips-dev-linuxkm-next-clang-tidy

…er-CPU native vector register save buffers, serving hardirq and softirq contexts where kernel_fpu_begin() is unusable. previously these contexts were refused (WC_ACCEL_INHIBIT_E): hardirq outright, softirq whenever !may_use_simd() (pre-6.15, a softirq interrupting a foreign kernel_fpu section). an exact save/compute/restore cycle is invisible to the interrupted context and to kernel FPU bookkeeping, and run-to-completion freezes the blocking condition beneath the section, so both classes are natively serviceable. - per-CPU {softirq, hardirq} state, each {depth, pin_preempt, save_area}; xsave64/xrstor64 with RFBM 0xe7 & XCR0 (fxsave64 fallback), plain XSAVE only; feature detection by direct CPUID/XGETBV, independent of kernel cpufeature machinery. - recursion is a per-(CPU, class) depth count -- save on the 0->1 edge, restore on 1->0; softirq sections hold preempt_disable() (load-bearing on PREEMPT_RT). - hardirq serves plain saves; softirq serves plain and _MAYBE_INHIBIT (its legacy short-circuit is bypassed when native is ready); _INHIBIT, hardirq _MAYBE, and NMI keep refusal semantics; task context is untouched. - wc_svr_native_init() activates in place from wc_linuxkm_allocate_svr_states() in any context but NMI -- the FIPS pre-Init lazy allocation arrives inside the outermost save's own bracket, and wolfCrypt_Init() has been observed with irqs disabled. a self-bracketed register round-trip selftest gates wc_svr_native_ready; on any failure the long-standing refusal semantics remain fully in force. - wc_linuxkm_svr_native_is_ready(): runtime readiness accessor for policy decisions elsewhere (e.g. the FIPS<v7 default-bank NO_VECTOR_OPS pin). - WOLFSSL_LINUXKM_VERBOSE_DEBUG: split the softirq WC_ACCEL_INHIBIT_E counter into per-class counters (softirq, hardirq, NMI, other), reported at module shutdown. in non-VERBOSE builds with WC_SVR_DONT_USE_NATIVE_REG_BUFS (or in non-SVR builds), generated object code is byte-identical to the previous revision.

…ontext handling.

github-actions · 2026-08-21T23:07:22Z

MemBrowse Memory Report

linuxkm-pie

Data: __patchable_function_entries +56 B (+0.2%, 26,744 B)

linuxkm-standard

Data: __patchable_function_entries +56 B (+0.1%, 49,552 B)
No memory changes detected for:
gcc-arm-cortex-m0plus
gcc-arm-cortex-m3
gcc-arm-cortex-m4
gcc-arm-cortex-m4-baremetal
gcc-arm-cortex-m4-crypto-only
gcc-arm-cortex-m4-dtls13
gcc-arm-cortex-m4-min-ecc
gcc-arm-cortex-m4-openssl-compat
gcc-arm-cortex-m4-pkcs7
gcc-arm-cortex-m4-pq
gcc-arm-cortex-m4-rsa-only
gcc-arm-cortex-m4-sp-math
gcc-arm-cortex-m4-tls12
gcc-arm-cortex-m4-tls13
gcc-arm-cortex-m7
gcc-arm-cortex-m7-pq
gcc-arm-cortex-m7-tls13
stm32-sim-stm32h753

douzzer · 2026-08-23T03:24:32Z

retest this please
(no logs)

douzzer · 2026-08-25T15:46:16Z

retest this please
(no logs)
(clean check-source-text and no changes outside the kernel...)

kaleb-himes

Passed the battery of tests thrown at it. Sha 256 was a little slow under load but I think that will get cleaned up by the coming ASM fixes so not holding the PR up over it.

wolfSSL-Fenrir-bot

Fenrir Automated Review — PR #11240

Scan targets checked: linuxkm-bugs, linuxkm-src, wolfcrypt-bugs, wolfcrypt-rs-bugs, wolfcrypt-src

Findings: 6
6 finding(s) posted as inline comments (see file-level comments below)

This review was generated automatically by Fenrir. Reported findings require changes before merge.

wolfSSL-Fenrir-bot · 2026-08-26T03:32:29Z


-#if defined(HAVE_FIPS) && FIPS_VERSION3_LT(7,0,0)
+#if defined(WOLFSSL_USE_SAVE_VECTOR_REGISTERS) && \
+    !defined(WC_SVR_USE_NATIVE_REG_BUFS) && \


FIPSv7 default-bank NO_VECTOR_OPS pin gated on compile-time WC_SVR_USE_NATIVE_REG_BUFS instead of runtime readiness · Logic errors

The pin is compiled out whenever WC_SVR_USE_NATIVE_REG_BUFS is built in, but native service can fail at runtime (no XSAVE/FXSR, allocation failure, selftest failure). On such boots linuxkm_get_drbg() still sets WC_RNG_BANK_FLAG_NO_VECTOR_OPS only in non-blockable contexts, so a FIPS v5/v6 wc_Sha256 DRBG instance alternates between asm and C mid-object — the exact breakage the pin prevents. linuxkm_get_drbg() at line 2211 already uses the runtime accessor.

Related known finding #7893 (similar but distinct): Both involve Linux DRBG default-bank startup state, but the candidate faults in wc_linuxkm_rng_bank_init's vector-operation flag selection when native register buffers are unavailable at runtime; #7893 faults in wc_linuxkm_drbg_startup's missing rollback after later failures. The root causes and corrective patches are distinct.

Fix: Keep the block compiled under WC_SVR_USE_NATIVE_REG_BUFS and gate the pin at runtime with !wc_linuxkm_svr_native_is_ready(), mirroring linuxkm_get_drbg().

wolfSSL-Fenrir-bot · 2026-08-26T03:32:29Z

+    #if defined(WC_SVR_DONT_USE_NATIVE_REG_BUFS)
+        #undef WC_SVR_USE_NATIVE_REG_BUFS
+    #elif defined(WOLFSSL_USE_SAVE_VECTOR_REGISTERS) && !defined(WC_SVR_USE_NATIVE_REG_BUFS)
+        #define WC_SVR_USE_NATIVE_REG_BUFS


WC_SVR_USE_NATIVE_REG_BUFS is auto-defined on non-x86, but its implementation and declaration are CONFIG_X86-only · LKCAPI interface correctness

The auto-define keys only on WOLFSSL_USE_SAVE_VECTOR_REGISTERS, which is also set for the ARM path at line 258. wc_linuxkm_svr_native_is_ready() is declared only inside the defined(CONFIG_X86) block at line 746 and defined only in x86_vector_register_glue.c, so the #ifdef WC_SVR_USE_NATIVE_REG_BUFS call at lkcapi_sha_glue.c:2211 has no declaration or symbol on ARM linuxkm builds.

Fix: Add && defined(CONFIG_X86) to the #elif condition that auto-defines WC_SVR_USE_NATIVE_REG_BUFS.

wolfSSL-Fenrir-bot · 2026-08-26T03:32:29Z


-#if defined(HAVE_FIPS) && FIPS_VERSION3_LT(7,0,0)
+#if defined(WOLFSSL_USE_SAVE_VECTOR_REGISTERS) && \
+    !defined(WC_SVR_USE_NATIVE_REG_BUFS) && \


FIPS7 default-bank NO_VECTOR_OPS pin is dropped on a compile-time macro rather than runtime native readiness · Cryptographic correctness

WC_SVR_USE_NATIVE_REG_BUFS only means the native save path is compiled in; wc_svr_native_ready stays 0 when the CPU lacks XSAVE/FXSR, the GFP_ATOMIC save-buffer allocation fails, or wc_svr_native_selftest() fails. In those builds the bank-level pin is gone while native never serves, so bankref checkouts via wc_local_rng_bank_checkout_for_bankref() (which sets no per-call NO_VECTOR_OPS) run unpinned. The sibling site linuxkm_get_drbg() at line 2211 correctly uses the runtime accessor.

Related known finding #7929 (similar but distinct): Both affect LinuxKM DRBG/default-bank initialization in lkcapi_sha_glue.c and can leave default-RNG behavior unsafe or inconsistent. However, #7929 faults after DRBG registration by omitting rollback on later startup failures, while this finding selects vector-operation pinning from compile-time availability instead of runtime native readiness; the operations, root causes, and patches differ.

Fix: Keep the FIPS<7 branch compiled in and gate the pin at runtime on !wc_linuxkm_svr_native_is_ready() instead of the build-time macro.

wolfSSL-Fenrir-bot · 2026-08-26T03:32:29Z


-#if defined(HAVE_FIPS) && FIPS_VERSION3_LT(7,0,0)
+#if defined(WOLFSSL_USE_SAVE_VECTOR_REGISTERS) && \
+    !defined(WC_SVR_USE_NATIVE_REG_BUFS) && \


FIPSv7 NO_VECTOR_OPS pin is gated at compile time instead of on runtime native readiness · Logic errors

The WC_RNG_BANK_FLAG_NO_VECTOR_OPS pin is now suppressed whenever WC_SVR_USE_NATIVE_REG_BUFS is merely compiled in. Native service can fail to come up at runtime (no XSAVE/FXSR, allocation failure, selftest failure in wc_svr_native_init()), and the pin is then dropped anyway, reintroducing mid-object asm/C switching that FIPS v5/v6 SHA-2 cannot survive. linuxkm_get_drbg() in the same file uses the runtime accessor.

Fix: Keep the #if on WOLFSSL_USE_SAVE_VECTOR_REGISTERS && HAVE_FIPS && FIPS_VERSION3_LT(7,0,0) and add a runtime !wc_linuxkm_svr_native_is_ready() condition to the if.

wolfSSL-Fenrir-bot · 2026-08-26T03:32:29Z

+                return 1;
+        }
+#endif /* WC_SVR_USE_NATIVE_REG_BUFS */
+#ifdef DEBUG_VECTOR_REGISTER_ACCESS_HARDIRQ_INFO


New DEBUG_VECTOR_REGISTER_ACCESS_HARDIRQ_INFO macro not registered in .wolfssl_known_macro_extras · Incorrect macro expansion

DEBUG_VECTOR_REGISTER_ACCESS_HARDIRQ_INFO is introduced at lines 505 and 607 but is defined nowhere in the tree. Its siblings DEBUG_VECTOR_REGISTER_ACCESS_ALWAYS_ON/OFF are listed in .wolfssl_known_macro_extras:255-256; this PR adds CONFIG_X86_64 and WC_SVR_DONT_USE_NATIVE_REG_BUFS there but omits this one, so the unknown-macro check flags it.

Fix: Add DEBUG_VECTOR_REGISTER_ACCESS_HARDIRQ_INFO to .wolfssl_known_macro_extras.

wolfSSL-Fenrir-bot · 2026-08-26T03:32:29Z


-#if defined(HAVE_FIPS) && FIPS_VERSION3_LT(7,0,0)
+#if defined(WOLFSSL_USE_SAVE_VECTOR_REGISTERS) && \
+    !defined(WC_SVR_USE_NATIVE_REG_BUFS) && \


FIPS7 default-bank NO_VECTOR_OPS pin is dropped at compile time instead of on runtime native readiness · Preprocessor-conditional security bypass

The pin is gated on !defined(WC_SVR_USE_NATIVE_REG_BUFS), which the header defines for every SVR build. When wc_svr_native_init() bails out (no XSAVE/FXSR, allocation failure, or selftest mismatch at x86_vector_register_glue.c:1579), wc_svr_native_ready stays 0 and the old refusal semantics remain in force, yet the pin is already gone. linuxkm_get_drbg() at line 2211 makes the identical policy decision via the runtime accessor.

Related known finding #7929 (similar but distinct): Both affect LinuxKM DRBG/default-bank startup in the same file, but the candidate faults in wc_linuxkm_rng_bank_init's NO_VECTOR_OPS policy gate when native vector readiness fails; #7929 faults in wc_linuxkm_drbg_startup rollback after later registration failures. Their root causes and corrective patches are different.

Fix: Gate the pin on wc_linuxkm_svr_native_is_ready() at runtime, as linuxkm_get_drbg() does, rather than on the compile-time macro.

douzzer added 3 commits August 21, 2026 15:23

add linuxkm/simd_hammer/, a stress testing tool for kernel FPU/SIMD c…

a999679

…ontext handling.

update .wolfssl_known_macro_extras

2fc57e0

douzzer self-assigned this Aug 21, 2026

kaleb-himes approved these changes Aug 25, 2026

View reviewed changes

douzzer assigned wolfSSL-Bot and unassigned douzzer Aug 26, 2026

philljj requested review from philljj and wolfSSL-Fenrir-bot August 26, 2026 02:34

wolfSSL-Fenrir-bot requested changes Aug 26, 2026

View reviewed changes

Navigation Menu

Uh oh!

20260821-linuxkm-WC_SVR_USE_NATIVE_REG_BUFS - #11240

20260821-linuxkm-WC_SVR_USE_NATIVE_REG_BUFS#11240
douzzer wants to merge 3 commits into
wolfSSL:masterfrom
douzzer:20260821-linuxkm-WC_SVR_USE_NATIVE_REG_BUFS

douzzer commented Aug 21, 2026

Uh oh!

github-actions Bot commented Aug 21, 2026 •

edited

Loading

Uh oh!

douzzer commented Aug 23, 2026

Uh oh!

douzzer commented Aug 25, 2026

Uh oh!

kaleb-himes left a comment

Uh oh!

wolfSSL-Fenrir-bot left a comment

Uh oh!

wolfSSL-Fenrir-bot Aug 26, 2026

Uh oh!

wolfSSL-Fenrir-bot Aug 26, 2026

Uh oh!

wolfSSL-Fenrir-bot Aug 26, 2026

Uh oh!

wolfSSL-Fenrir-bot Aug 26, 2026

Uh oh!

wolfSSL-Fenrir-bot Aug 26, 2026

Uh oh!

wolfSSL-Fenrir-bot Aug 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

douzzer commented Aug 21, 2026

Uh oh!

github-actions Bot commented Aug 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

MemBrowse Memory Report

Uh oh!

douzzer commented Aug 23, 2026

Uh oh!

douzzer commented Aug 25, 2026

Uh oh!

kaleb-himes left a comment

Choose a reason for hiding this comment

Uh oh!

wolfSSL-Fenrir-bot left a comment

Choose a reason for hiding this comment

Fenrir Automated Review — PR #11240

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions Bot commented Aug 21, 2026 •

edited

Loading