Mac allow system ucontext #90
Conversation
|
This still needs some cleanup, but the intent here is to enable initial support for m1 macs using the system ucontext implementation while we work on support for assembly-based fast context switching for arm64 (primarily being developed on our ThunderX2 Linux arm systems for now). The initial support given here requires the changes in this pull request, along with the following build recipe: Additionally, you'll see warnings about using deprecated functions, which you can suppress via compiler flags if they bother you. @ronawho was wondering about why we had disallowed the use of the system ucontext on mac systems in the first place. I don't have an answer to that, except that I have found that our preferred default small stack sizes are not viable for it: signal.h specifies a minimum of 32k and a maximum of 128k. All tests pass for me using 64k. Not sure if @m3m0ryh0l3 had a particular reason in mind. |
|
@olivier-snl I don't have access to an m1 mac to test on, but I'm wondering if we can automate this detection instead of forcing users to reconfigure. We have a lot of users coming from a pure python world and many of them are not comfortable or familiar with reconfiguring. |
|
Tested on x86 Mac that ./configure && make check works for both setting --disable-fastcontext and not setting --disable-fastcontext. M1 Mac requires the use of --disable-fastcontext. With setting this option on M1 Mac both, configure and make check succeed. We are working on the fast context implementation on M1 Mac hardware. |
|
We could add a autoconf check for M1 MAC and set --disable-fastcontext in that case. |
Default to fifo tasking layer on arm based macs [reviewed by @bradcray and @gbtitus] Qthreads doesn't currently support arm based macs. Configure fails with `error: What kind of a Mac is this?` while trying to figure out what native assembly to use for context switching. We started looking at using system `ucontext` in sandialabs/qthreads#90, but given how close our release is we want to default to fifo since we know some users have had success with that. Resolves Cray/chapel-private#2424 Part of #17911
| *-darwin*) | ||
| qthread_cv_swapcontext_works=no | ||
| ;; | ||
| # *-darwin*) |
There was a problem hiding this comment.
Since current x86 and M1 Macs seemingly support system ucontext, I think it is ok to drop this case statement.
There was a problem hiding this comment.
Consider adding something like
AS_IF([test "$host" in *-darvin*], [qt_host_based_enable_fastcontext=no], [qt_host_based_enable_fastcontext=yes] ;;
to L37 to disable fast context automatically on M1 Macs while the support is not available yet.
There was a problem hiding this comment.
darvin -> darwin. But really, it's the combo of darwin and aarch64 (however that parses). We don't want to disable fast context on earlier macs.
There was a problem hiding this comment.
Yes, that combo should come together in L37 if I see this correctly when adding that additional check for Darwin.
There was a problem hiding this comment.
There was a problem hiding this comment.
https://github.com/janciesko/qthreads/tree/mac-allow-system-ucontext-2
This is not correct. m1 macs are arm v8 not arm v7. The arch to AC will be something like arm64 or aarch64.
There was a problem hiding this comment.
Yes, this was more an idea / place holder. We'd need to add the right case depending on what config gives us.
There was a problem hiding this comment.
We can match on aarch64-*-darwin* | arm64*-*-darwin*
There was a problem hiding this comment.
Actually, since it shouldn't really match any of the architectures in the case, it should fall through without any change. I just tried from the qthreads repo branch (without the m4 change and without the disable fast context switch on configure) and it worked just fine.
|
Superseded by #92 |
No description provided.