Mimi, On Wed, Aug 18, 2021 at 05:07:20PM -0400, Mimi Zohar wrote: > On Fri, 2021-08-13 at 17:31 -0400, Mimi Zohar wrote: > > On Fri, 2021-08-13 at 00:46 +0300, Vitaly Chikunov wrote: > > > On Fri, Aug 13, 2021 at 12:21:43AM +0300, Vitaly Chikunov wrote: > > > > After CRYPTO_secure_malloc_init OpenSSL will store private keys in > > > > secure heap. This facility is only available since OpenSSL_1_1_0-pre1 > > > > and enabled for 'ima_sign', 'sign', 'sign_hash', and 'hmac'. > > > > From the manpage: > > CRYPTO_secure_malloc_init() returns 0 on failure, 1 if successful, and > > 2 if successful but the heap could not be protected by memory mapping. > > > > Not sure what we would do on failure ( 0, 2), but we should at least > > check the return code. > > > > > > > setvbuf(3) _IONBF is used to hopefully avoid private key and password > > > > being stored inside of stdio buffers. > > > > > > I should note that usefulness of this method (of avoiding buffering) is > > > not proven. I don't find other implementations doing it. So, I'm open to > > > suggestion of removing it. > > > > > Probably would be better to split the patch. > > According to the man page "OPENSSL_secure_malloc() allocates num bytes > from the heap. If CRYPTO_secure_malloc_init() is not called, this is > equivalent to calling OPENSSL_malloc()". OPENSSL_malloc() is > supported in the older openssl versions. > > Does it make sense to replace allocating memory for the password via > malloc() with OPENSSL_secure_malloc()? For older openssl versions, > define OPENSSL_secure_malloc() and OPENSSL_secure_free() as > OPENSSL_malloc() and OPENSSL_free(). It makes sense to use OPENSSL_secure_malloc for passwords, but it will look strange to define OPENSSL_secure_malloc as OPENSSL_malloc for older versions. Maybe we could add a #warning in that case. Thanks, > > This doesn't solve the memory handling for private keys and passwords > for older openssl versions, but it is a path forward. > > thanks, > > Mimi