Clear sensitive memory without getting optimized out (revival of #636)

[theStack]

This PR picks up #636 (which in turn picked up #448, so this is take number three) and is essentially a rebase on master.

Some changes to the original PR:

• the clearing function now has the secp256k1_ prefix again, since the related helper _memczero got it as well (see PR Don't use reserved identifiers memczero and benchmark_verify_t #835 / commit e89278f)
• the original commit b17a7df ("Make _set_fe_int( . , 0 ) set magnitude to 0") is not needed anymore, since it was already applied in PR VERIFY_CHECK precondition for secp256k1_fe_set_int. #943 (commit d49011f)
• clearing of stack memory with secp256k1_memclear is now also done on modules that have been newly introduced since then, i.e. schnorr and ellswift (of course, there is still no guarantee that all places where clearing is necessary are covered)

So far I haven't looked at any disassembly and possible performance implications yet (there were some concerns expressed in #636 (comment)), happy to go deeper there if this gets Concept ACKed.

The proposed method of using a memory barrier to prevent optimizating away the memset is still used in BoringSSL (where it was originally picked up from) and in the Linux Kernel, see e.g. https://github.com/google/boringssl/blob/5af122c3dfc163b5d1859f1f450756e8e320a142/crypto/mem.c#L335 and https://github.com/torvalds/linux/blob/d4560686726f7a357922f300fc81f5964be8df04/include/linux/string.h#L348 / https://github.com/torvalds/linux/blob/d4560686726f7a357922f300fc81f5964be8df04/include/linux/compiler.h#L102

Fixes #185.

[real-or-random]

Concept ACK (obviously)

Thanks for reviving this, I never had the time/motivation to come back to this PR, but it's important.

We should call SECP256K1_CHECKMEM_UNDEFINE (

secp256k1/src/checkmem.h

Line 27 in b307614

* - SECP256K1_CHECKMEM_UNDEFINE(p, len):

) in secp256k1_memclear after clearing the memory. Reading cleared memory would clearly be a bug.

[theStack]

We should call SECP256K1_CHECKMEM_UNDEFINE (

secp256k1/src/checkmem.h

Line 27 in b307614

* - SECP256K1_CHECKMEM_UNDEFINE(p, len):

) in secp256k1_memclear after clearing the memory. Reading cleared memory would clearly be a bug.

Thanks, added that, and rebased on master.

[real-or-random]

@theStack Can you rebase this on top of musig which has introduced a few more code locations that need clearing?

Personally, I'd love to have this in the next release.

[theStack]

@theStack Can you rebase this on top of musig which has introduced a few more code locations that need clearing?

Sure, done. With only five lines changed in the musig module, this needed less effort than expected, hope I didn't miss any instances (many of them are handled indirectly via the ..._{fe,scalar}_clear functions though).

Relevant excerpt of the range-diff (uncolored here)

$ git range-diff ac0e41...0b01d2
5:  6fcbae9 ! 15:  02ee811 Use secp256k1_memclear() to clear stack memory instead of memset()
    @@ src/modules/ellswift/main_impl.h: int secp256k1_ellswift_xdh(const secp256k1_con
          secp256k1_scalar_clear(&s);
      
     
    + ## src/modules/musig/session_impl.h ##
    +@@ src/modules/musig/session_impl.h: static void secp256k1_nonce_function_musig(secp256k1_scalar *k, const unsigned c
    +         secp256k1_scalar_set_b32(&k[i], buf, NULL);
    + 
    +         /* Attempt to erase secret data */
    +-        memset(buf, 0, sizeof(buf));
    +-        memset(&sha_tmp, 0, sizeof(sha_tmp));
    ++        secp256k1_memclear(buf, sizeof(buf));
    ++        secp256k1_memclear(&sha_tmp, sizeof(sha_tmp));
    +     }
    +-    memset(rand, 0, sizeof(rand));
    +-    memset(&sha, 0, sizeof(sha));
    ++    secp256k1_memclear(rand, sizeof(rand));
    ++    secp256k1_memclear(&sha, sizeof(sha));
    + }
    + 
    + int secp256k1_musig_nonce_gen_internal(const secp256k1_context* ctx, secp256k1_musig_secnonce *secnonce, secp256k1_musig_pubnonce *pubnonce, const unsigned char *input_nonce, const unsigned char *seckey, const secp256k1_pubkey *pubkey, const unsigned char *msg32, const secp256k1_musig_keyagg_cache *keyagg_cache, const unsigned char *extra_input32) {
    +@@ src/modules/musig/session_impl.h: int secp256k1_musig_nonce_gen_counter(const secp256k1_context* ctx, secp256k1_mu
    +     if (!secp256k1_musig_nonce_gen_internal(ctx, secnonce, pubnonce, buf, seckey, &pubkey, msg32, keyagg_cache, extra_input32)) {
    +         return 0;
    +     }
    +-    memset(seckey, 0, sizeof(seckey));
    ++    secp256k1_memclear(seckey, sizeof(seckey));
    +     return 1;
    + }
    + 
    +

[real-or-random]

This one comes to my mind, too:

secp256k1/src/util.h

Lines 253 to 254 in a88aa93

	/* acc may contain secret values. Try to explicitly clear it. */
	acc = 0;

[theStack]

This one comes to my mind, too:

secp256k1/src/util.h

Lines 253 to 254 in a88aa93

	/* acc may contain secret values. Try to explicitly clear it. */
	acc = 0;

Ah good catch, missed that (only grepped for memset calls). Tackled now as well.

[real-or-random]

If you want motivated, you could look at git grep secp256k1_ge_set_gej.*(. @sipa's recent comment in the MuSig PR really caught my attention. When I worked on the previous PR, I really didn't consider the possibility that a gej could leak secret data.

---

I wonder if it makes sense to have the _finalize functions in the hash module clear the state automatically. And then have "unsafe" funtions like _finalize_keep if the callers knows that data is public or if the caller wants to reuse the midstate. It sounds like a neat idea, but I'm not really convinced because it special-cases the hash module somewhat: we'll need to manually clear everything else including scalars, etc... So we have to be careful with this anyway when writing and reviewing code, and perhaps having yet another safety mechanism that works only for the hash module. (This could perhaps be a nice follow-up if we add the same mechanism to more modules, e.g., ge_set_gej could clear the gejs, unless you use a _keep variant.)

[theStack]

If you want motivated, you could look at git grep secp256k1_ge_set_gej.*(. @sipa's recent comment in the MuSig PR really caught my attention. When I worked on the previous PR, I really didn't consider the possibility that a gej could leak secret data.

I found the following functions containing _gej instances resulting from point multiplication (secp256k1_ecmult_{gen,const}) with secret scalars:

• ECDH module: secp256k1_ecdh (res, resulting from the ECDH point multiplication; I think this should be cleared not only to avoid leaking the scalar, but even more so as it's a representation of the resulting shared secret)
• MuSig2 module: secp256k1_musig_nonce_gen_internal (public nonce nonce_ptj in the for loop)
• Schnorr module: secp256k1_schnorrsig_sign_internal (nonce commitment rj)
• main library: secp256k1_ec_pubkey_create_helper (public point pj)

Taking care of those could be probably go into an own PR, as its trivial to fix and review and hence has a significantly higher chance to land in the next release than this one? (and having a version where the compiler still might optimize it out seems still much better than not doing it at all)

Interestingly, the ECDSA signing function does clear out the nonce commitment for both the jacobi and affine points (though the latter wouldn't be needed according to #1479 (comment)).

I wonder if it makes sense to have the _finalize functions in the hash module clear the state automatically. And then have "unsafe" funtions like _finalize_keep if the callers knows that data is public or if the caller wants to reuse the midstate. It sounds like a neat idea, but I'm not really convinced because it special-cases the hash module somewhat: we'll need to manually clear everything else including scalars, etc... So we have to be careful with this anyway when writing and reviewing code, and perhaps having yet another safety mechanism that works only for the hash module. (This could perhaps be a nice follow-up if we add the same mechanism to more modules, e.g., ge_set_gej could clear the gejs, unless you use a _keep variant.)

Sounds like a good idea to me for a follow-up.

[real-or-random]

Taking care of those could be probably go into an own PR, as its trivial to fix and review and hence has a significantly higher chance to land in the next release than this one? (and having a version where the compiler still might optimize it out seems still much better than not doing it at all)

My guess is that what we do currently is useless on any modern compiler. I admit that I haven't looked at the compiler output, but I'd rather spend the time on resolving the problem.

I don't think that fixing these additional cases here will make it much more difficult to review the PR. And to be honest, while this is a great for defense in depth, we haven't deeply cared about this so far. It's not the end of the world if we need to wait a few more months. So I think it's good to add these cases to this PR here.

[theStack]

My guess is that what we do currently is useless on any modern compiler. I admit that I haven't looked at the compiler output, but I'd rather spend the time on resolving the problem.

I don't think that fixing these additional cases here will make it much more difficult to review the PR. And to be honest, while this is a great for defense in depth, we haven't deeply cared about this so far. It's not the end of the world if we need to wait a few more months. So I think it's good to add these cases to this PR here.

Thanks for the feedback! Makes sense, added an extra commit at the end for the gej clearing after point multiplications, it's only four lines of code anyway. Also went over the necessary hash clearing places and made a few small changes (see #1579 (comment)).

[sipa]

utACK c921078. I have not reviewed this for exhaustiveness (as in, are there more places where clearing is useful/necessary), but the code changes look good.

[theStack]

Hm, it seems that in production code, we usually don't call any of the _CHECKMEM_{DEFINE,UNDEFINE} macros. There is one reachable code-path in secp256k1_declassify, but it would only hit if the context was created with the SECP256K1_CONTEXT_DECLASSIFY flag, where the API documentation explicitly says "Do not use". Should that _CHECKMEM_UNDEFINE call here be conditional by a preprocessor #if(def) or something alike, so it only applies to tests and can't slow down (and bloat) release builds?

I noticed that while looking at the disassembly of the .so file and wondering why there was so much extra code after the memory clearing, until I realized this must be valgrind's VALGRIND_MAKE_MEM_UNDEFINED. I'm building now explicitly with -DSECP256K1_VALGRIND=OFF -DSECP256K1_BUILD_CTIME_TESTS=OFF to avoid that.

[real-or-random]

Hm, great point! The reason why secp256k1_declassify uses a run-time flag (instead of compile-time flag) is that we want to run the constant-time tests on the real binary.

I don't know what the performance impact of these additional instructions is, but I doubt that having a compile-time flag is a concern in this case. That means that we could just wrap the SECP256K1_CHECKMEM_UNDEFINE in an #ifdef VERIFY. @sipa What do you think?

[sipa]

I think the reasoning is that we want the release binaries to be as close as possible to what we actually test in the ctime test, but disable the effect at runtime using SECP256K1_CONTEXT_DECLASSIFY to avoid a performance impact.

It wouldn't surprise me that that is overkill; the conditional to determine whether or not to declassify operations at runtime may have a higher cost than actually executing the nop instructions that declassification actually correspond to when not running instrumented by valgrind.

[real-or-random]

Oh, sorry, I think we're talking past each other. My suggestion, for which I'd like to have your opinion on, is to wrap SECP256K1_CHECKMEM_UNDEFINE in an #ifdef VERIFY block here in this particular case of clearing memory, i.e., inside secp256k1_memclear (and not in secp256k1_declassify. Do you think that's a reasonable thing to do?

---

But regarding what you said:

I think the reasoning is that we want the release binaries to be as close as possible to what we actually test in the ctime test, but disable the effect at runtime using SECP256K1_CONTEXT_DECLASSIFY to avoid a performance impact.

Yes, I agree. Perhaps an additional reason is to avoid any shenanigans that the valgrind syscalls may have. They should be noops, but it's certainly a bit safer not to run them at all.

It wouldn't surprise me that that is overkill; the conditional to determine whether or not to declassify operations at runtime may have a higher cost than actually executing the nop instructions that declassification actually correspond to when not running instrumented by valgrind.

I can imagine that the overhead of the noops is negligible, but at least checking the conditional should be negligible as well because we use EXCEPT to help the compiler predict the branch:

secp256k1/src/secp256k1.c

Lines 236 to 238 in 68b5520

	static SECP256K1_INLINE void secp256k1_declassify(const secp256k1_context* ctx, const void *p, size_t len) {
	if (EXPECT(ctx->declassify, 0)) SECP256K1_CHECKMEM_DEFINE(p, len);
	}

[theStack]

What bothered me in general (independent on how lightweight the code generated by the valgrind macros might be) is that when users build with the default CMake settings, they could currently end up with different release binaries, depending on whether valgrind is installed or not. Unrelated to this PR, but maybe the SECP256K1_VALGRIND build setting should only be auto-detected in the "dev-mode" preset, and OFF by default? (I only looked at the CMake build, don't know if the valgrind setting is also auto-detected on autotools).

[real-or-random]

I see. That's a valid point, which has been brought up before, see #813 (comment). (This thread is also very helpful to understand why we have run-time flag etc.) The conclusion was that the benefits outweigh the drawback that the build outputs depends on the availability of valgrind, but yeah, there's no perfect solution here. And yes, it's also auto-detected in autotools.

[sipa]

I see, so the idea is to have

...
#ifndef VERIFY
     SECP256K1_CHECKMEM_UNDEFINE(ptr, len);
#endif
}

in secp256k1_memclear()?

That would mean tests_noverify wouldn't detect violations of use-after-clear, nor would applications linking against the production library when running in valgrind. The upside would be that the production library definitely does not have its behavior affected by being compiled with valgrind present.

[real-or-random]

Indeed.

That would mean tests_noverify wouldn't detect violations of use-after-clear

Fwiw, I had this downside in mind, and I don't think it's a big deal because we still have the normal tests.

nor would applications linking against the production library when running in valgrind.

Okay, I had not considered this one... I still think that's acceptable, but yeah, it's a good point...

[sipa]

ACK. I guess it makes sense that if the primary goal is making sure no runtime impact for production code, this should be guarded by a compile-time check, as the runtime check isn't available here (being just for ctime testing).

[theStack]

For a first check on whether this PR fulfills its promise, I looked at the generated assembler code of secp256k1_ec_seckey_verify, as this function is particularly short and should clear out the secret scalar at the end. The following commands were used to generate the disassembled output:

$ cmake --build build && objdump --no-addresses --no-show-raw-insn --disassemble=secp256k1_ec_seckey_verify ./build/lib/libsecp256k1.so > seckey_verify_disasm.txt

compiler output diff of `secp256k1_ec_seckey_verify` (master vs. PR branch)

running on Debian 12 stable (Bookworm) and gcc 12.2.0 on a x86-64 machine

--- master.txt  2024-10-22 16:09:38.985554751 +0200
+++ pr1579.txt  2024-10-22 16:12:37.324112394 +0200
@@ -11,31 +11,38 @@
 Disassembly of section .text:
 
 <secp256k1_ec_seckey_verify>:
-       sub    $0x38,%rsp
+       push   %rbx
+       sub    $0x30,%rsp
        test   %rsi,%rsi
-       je     <secp256k1_ec_seckey_verify+0x48>
+       je     <secp256k1_ec_seckey_verify+0x58>
+       lea    0x10(%rsp),%rbx
        lea    0xc(%rsp),%rdx
-       lea    0x10(%rsp),%rdi
+       pxor   %xmm0,%xmm0
+       mov    %rbx,%rdi
        call   <secp256k1_scalar_set_b32>
        mov    0xc(%rsp),%ecx
        mov    0x10(%rsp),%rax
        or     0x18(%rsp),%rax
        or     0x20(%rsp),%rax
+       movaps %xmm0,0x10(%rsp)
        or     0x28(%rsp),%rax
+       movaps %xmm0,0x20(%rsp)
        setne  %dl
        xor    %eax,%eax
        test   %ecx,%ecx
        sete   %al
-       add    $0x38,%rsp
        and    %edx,%eax
+       add    $0x30,%rsp
+       pop    %rbx
        ret
        nopl   0x0(%rax)
        mov    %rdi,%rax
        mov    0xb0(%rdi),%rsi
-       lea    0x63b9(%rip),%rdi        # <_fini+0x1196>
+       lea    0x6199(%rip),%rdi        # <_fini+0xf46>
        call   *0xa8(%rax)
+       add    $0x30,%rsp
        xor    %eax,%eax
-       add    $0x38,%rsp
+       pop    %rbx
        ret

There is a little noise in the diff due to different register allocations and jump offsets, but in the middle one can see quite clearly that the memory clearing happens using the 16-byte SSE register xmm0: it is earlier zeroed with a `pxor` instruction and then written with `movaps` instruction twice to memory locations `0x10(%rsp)` and `0x20(%rsp)`, which correspond to the 32-byte `secp256k1_scalar` instance (interestingly, the clearing is interleaved with the `mov`/`or` instructions of the `_scalar_is_zero` check). So this seems to work as expected. Might be interesting to also try it with a variety of other compiler versions (especially newer ones) and different architectures, like arm64.

[real-or-random]

ACK mod my review comments

[theStack]

Fixed #1579 (comment) and rebased on master for easier comparison of branches (since #1553, the output location of the .so file changed from ./build/src to ./build/lib; not being aware of that, I was unintentionally comparing the library file generated from master with itself for quite some time, wondering why the PR doesn't change anything 🤦‍♂️ ).

[theStack]

Regarding the open question of whether or not to include the _CHECKMEM_UNDEFINE macro in secp256k1_memclear in non-verify binaries or not, I compared the valgrind-builds on the pr branch with the macro commented out vs. with the macro included. This is what the assembler diff looks like (created for the same function and using the same methodology as in #1579 (comment)):

disassemble of secp256k1_ec_seckey_verify, without and with the SECP256K1_CHECKMEM_UNDEFINE in a valgrind build

--- pr1579_without_checkmem_undefine.txt	2024-10-25 17:34:50.639344958 +0200
+++ pr1579_with_checkmem_undefine.txt	2024-10-25 17:33:35.632626088 +0200
@@ -11,38 +11,59 @@
 Disassembly of section .text:
 
 <secp256k1_ec_seckey_verify>:
+	push   %rbp
 	push   %rbx
-	sub    $0x30,%rsp
+	sub    $0x68,%rsp
 	test   %rsi,%rsi
-	je     <secp256k1_ec_seckey_verify+0x58>
+	je     <secp256k1_ec_seckey_verify+0xb8>
+	lea    0x30(%rsp),%rbp
 	lea    0x10(%rsp),%rbx
-	lea    0xc(%rsp),%rdx
 	pxor   %xmm0,%xmm0
+	mov    %rbp,%rdx
 	mov    %rbx,%rdi
 	call   <secp256k1_scalar_set_b32>
-	mov    0xc(%rsp),%ecx
-	mov    0x10(%rsp),%rax
-	or     0x18(%rsp),%rax
-	or     0x20(%rsp),%rax
+	mov    0x10(%rsp),%rdx
+	or     0x18(%rsp),%rdx
 	movaps %xmm0,0x10(%rsp)
-	or     0x28(%rsp),%rax
+	or     0x20(%rsp),%rdx
+	or     0x28(%rsp),%rdx
 	movaps %xmm0,0x20(%rsp)
-	setne  %dl
-	xor    %eax,%eax
-	test   %ecx,%ecx
-	sete   %al
-	and    %edx,%eax
-	add    $0x30,%rsp
+	mov    0x30(%rsp),%edx
+	setne  %al
+	xor    %ecx,%ecx
+	test   %edx,%edx
+	sete   %cl
+	and    %eax,%ecx
+	movq   $0x4d430001,0x30(%rsp)
+	xor    %edx,%edx
+	mov    %rbp,%rax
+	mov    %rbx,0x38(%rsp)
+	movq   $0x20,0x40(%rsp)
+	movq   $0x0,0x48(%rsp)
+	movq   $0x0,0x50(%rsp)
+	movq   $0x0,0x58(%rsp)
+	rol    $0x3,%rdi
+	rol    $0xd,%rdi
+	rol    $0x3d,%rdi
+	rol    $0x33,%rdi
+	xchg   %rbx,%rbx
+	mov    %rdx,0x8(%rsp)
+	mov    0x8(%rsp),%rax
+	add    $0x68,%rsp
+	mov    %ecx,%eax
 	pop    %rbx
+	pop    %rbp
 	ret
-	nopl   0x0(%rax)
+	nopl   0x0(%rax,%rax,1)
 	mov    %rdi,%rax
 	mov    0xb0(%rdi),%rsi
-	lea    0x5d09(%rip),%rdi        # <_fini+0x766>
+	lea    0x7b99(%rip),%rdi        # <_fini+0x14e6>
 	call   *0xa8(%rax)
-	add    $0x30,%rsp
-	xor    %eax,%eax
+	add    $0x68,%rsp
+	xor    %ecx,%ecx
+	mov    %ecx,%eax
 	pop    %rbx
+	pop    %rbp
 	ret
 
 Disassembly of section .fini:

Seems like at least in this function, it causes at least 15 extra (no-op like) instructions to be emitted. Considering that previously, these valgrind placeholders were not reachable (unless explicitly specified in the context object) and these extra instructions are added at multiple places spread all over the code, it might be better to not include them in the release binaries? OTOH, "ctime-tests should run with the same binary as in the release" then doesn't hold anymore. So really not sure what's better 🤔

[real-or-random]

OTOH, "ctime-tests should run with the same binary as in the release" then doesn't hold anymore.

I don't think so. If we wrap the new _CHECKMEM_UNDEFINE within _memclear in #ifdef VERIFY, I don't see how this affects the ctimetests at all. We'd still include the _CHECKMEM_UNDEFINE in declassify in the release builds. And the ctimetests executable will still link against the (release) binary of the library. (Note that the ctimetests test executable uses the public API, it's really just an application that consumes the library by linking against it, as opposed to the other test binaries, which essentially #include all the library code.)

[theStack]

OTOH, "ctime-tests should run with the same binary as in the release" then doesn't hold anymore.

I don't think so. If we wrap the new _CHECKMEM_UNDEFINE within _memclear in #ifdef VERIFY, I don't see how this affects the ctimetests at all. We'd still include the _CHECKMEM_UNDEFINE in declassify in the release builds. And the ctimetests executable will still link against the (release) binary of the library. (Note that the ctimetests test executable uses the public API, it's really just an application that consumes the library by linking against it, as opposed to the other test binaries, which essentially #include all the library code.)

Oh right, for some reason I assumed that the ctimetests binary is also compiled with VERIFY defined, which is of course nonsensical. Added the "#ifdef VERIFY" block around _CHECKMEM_UNDEFINE now, as it looks we have consensus on doing that (#1579 (comment)).

[real-or-random]

ACK 765ef53

[sipa]

reACK 765ef53

[Vin757]

W