On Thu, Mar 20, 2025 at 04:39:11PM +0000, David Howells wrote:
Jarkko Sakkinen jarkko@kernel.org wrote:
if (test_bit(KEY_FLAG_FINAL_PUT, &key->flags)) {smp_mb(); /* Clobber key->user after FINAL_PUT seen. */test_bit() is already atomic.
Atomiticity doesn't apply to test_bit() - it only matters when it does two (or more) accesses that must be perceptually indivisible (e.g. set_bit doing RMW).
But atomiticity isn't the issue here, hence the barrier. You need to be looking at memory-barriers.txt, not atomic_bitops.txt.
We have two things to correctly order and set_bit() does not imply sufficient barriering; test_and_set_bit() does, but not set_bit(), hence Linus's comment about really wanting a set_bit_release().
Oops, I was hallucinating here. And yeah, test_and_set_bit() does imply full mb as you said.
I was somehow remembering what I did in SGX driver incorrectly and that led me into misconclusions, sorry.
if (test_and_set_bit(SGX_ENCL_IOCTL, &encl->flags)) return -EBUSY;
smp_mb(); /* key->user before FINAL_PUT set. */set_bit(KEY_FLAG_FINAL_PUT, &key->flags);Ditto.
Ditto. ;-)
Duh, no need poke with the stick further (or deeper) ;-)
Nit: I'm just thinking should the name imply more like that "now key_put() is actually done". E.g., even something like KEY_FLAG_PUT_DONE would be more self-descriptive.
KEY_FLAG_PUT_DONE isn't right. There can be lots of puts on a single key - only the one that reduces it to 0 matters for this. You could call it KEY_FLAG_CAN_NOW_GC or KEY_FLAG_GC_ABLE.
Well all alternatives are fine but my thinking was that one that finally zeros the refcount, "finalizes put" (pick whatever you want anyway).
David
BR, Jarkko