The lazy-initialisation of BN_MONT_CTX was serialising all threads, as noted by Daniel Sands and co at Sandia. This was to handle the case that 2 or more threads race to lazy-init the same context, but stunted all scalability in the case where 2 or more threads are doing unrelated things! We favour the latter case by punishing the former. The init work gets done by each thread that finds the context to be uninitialised, and we then lock the "set" logic after that work is done - the winning thread's work gets used, the losing threads throw away what they've done.
Signed-off-by: Geoff Thorpe <ge...@openssl.org> --- Hi Daniel, This patch is relative to the current 'master' branch, but it applies cleanly ("git am") to the 1.0.2 and 1.0.1 branches too, so should presumably be workable if you're using something similar (like a patched source tree from some distro). "make test" in openssl still passes after this, so I'm reasonably sure it doesn't logically break anything, but as you have some scalability use-case to hand I'd be grateful if you could confirm for me that it improves the performance issue too. Thanks, Geoff crypto/bn/bn_mont.c | 46 ++++++++++++++++++++++++++-------------------- 1 file changed, 26 insertions(+), 20 deletions(-) diff --git a/crypto/bn/bn_mont.c b/crypto/bn/bn_mont.c index e6f6e3f..799763e 100644 --- a/crypto/bn/bn_mont.c +++ b/crypto/bn/bn_mont.c @@ -480,32 +480,38 @@ BN_MONT_CTX *BN_MONT_CTX_copy(BN_MONT_CTX *to, BN_MONT_CTX *from) BN_MONT_CTX *BN_MONT_CTX_set_locked(BN_MONT_CTX **pmont, int lock, const BIGNUM *mod, BN_CTX *ctx) { - int got_write_lock = 0; BN_MONT_CTX *ret; CRYPTO_r_lock(lock); - if (!*pmont) + ret = *pmont; + CRYPTO_r_unlock(lock); + if (ret) + return ret; + + /* We don't want to serialise globally while doing our lazy-init math in + * BN_MONT_CTX_set. That punishes threads that are doing independent + * things. Instead, punish the case where more than one thread tries to + * lazy-init the same 'pmont', by having each do the lazy-init math work + * independently and only use the one from the thread that wins the race + * (the losers throw away the work they've done). */ + ret = BN_MONT_CTX_new(); + if (!ret) + return NULL; + if (!BN_MONT_CTX_set(ret, mod, ctx)) { - CRYPTO_r_unlock(lock); - CRYPTO_w_lock(lock); - got_write_lock = 1; + BN_MONT_CTX_free(ret); + return NULL; + } - if (!*pmont) - { - ret = BN_MONT_CTX_new(); - if (ret && !BN_MONT_CTX_set(ret, mod, ctx)) - BN_MONT_CTX_free(ret); - else - *pmont = ret; - } + /* The locked compare-and-set, after the local work is done. */ + CRYPTO_w_lock(lock); + if (*pmont) + { + BN_MONT_CTX_free(ret); + ret = *pmont; } - - ret = *pmont; - - if (got_write_lock) - CRYPTO_w_unlock(lock); else - CRYPTO_r_unlock(lock); - + *pmont = ret; + CRYPTO_w_unlock(lock); return ret; } -- 1.9.1 ______________________________________________________________________ OpenSSL Project http://www.openssl.org Development Mailing List openssl-dev@openssl.org Automated List Manager majord...@openssl.org