marko, please review 10 top patches at https://github.com/MariaDB/server/commits/bb-10.4-MDEV-17441

Please also pay special attention to buf_tmp_buffer_t::reserved patch.

Looks generally OK. I posted a few comments to some of the commits.

I still see a few references to my_atomic in InnoDB after the commits that I reviewed. Could all of them be replaced?

I’d also like to have a few member functions in rw_lock_t, to reduce some typing. Ideally, the memory ordering would not be repeated on each operation, but it should be part of the object declaration. Could we somehow the memory ordering a template parameter, or otherwise make relaxed memory ordering the default, to reduce typing?

Original head with 17 patches just for reference: https://github.com/MariaDB/server/commits/bf64d634b831052e8f603fb6cc43296051ae24a3
Second review round head with 21 patches: https://github.com/MariaDB/server/commits/b9f945a186255741319f621c94aeadf43d04c6f7

We should eventually eliminate them all. Unfortunately trying to eliminate them at once may lead to project failure, since it is not that trivial.

I don't like idea of making relaxed default memory ordering for two reasons:
1. unfortunately implicit memory order in general makes people make unexpected assumptions of how memory ordering works, which leads to horribly broken code
2. sequentially consistent is c++11 defined default memory ordering; even though I agree relaxed is more appropriate, it'd be very confusing to have atomics implementation with different default

But I agree we could reduce some typing by having proper methods in rw_lock_t or alike.

marko, I implemented most of your suggestions and pushed to https://github.com/MariaDB/server/commits/bb-10.4-MDEV-17441

Would you like to have another look?

Looks OK. I sent a few comments, requesting minor improvement for code clarity.

I see that os0once.h and srv0mon.h still have not been converted to use C++11 atomics. Shall we leave those for a later version?

I have WiP patches for os0once.h, which I'm planning to complete soon hopefully.
srv0mon.h requires larger refactoring and we should probably postpone it.

The srv0mon.h would be refactored or removed by MDEV-15706 in a later release.

kevg, could you complete this effort? Specifically get rid of os0once. I pushed 2 WiP patches to bb-10.4-svoj-MDEV-17441.

Looks good. I only had a minor request to replace a comment with a debug assertion.