On Tue 21-11-17 17:48:31, John Hubbard wrote: [...] > Hi Michal, > > Yes, it really is useful for user space. I'll use CUDA as an example, but I > think anything that enforces a uniform virtual addressing scheme across CPUs > and devices, probably has to do something eerily similar. CUDA does this: > > a) Searches /proc/<pid>/maps for a "suitable" region of available VA space. > "Suitable" generally means it has to have a base address within a certain > limited range (a particular device model might have odd limitations, for > example), it has to be large enough, and alignment has to be large enough > (again, various devices may have constraints that lead us to do this). > > This is of course subject to races with other threads in the process. > > Let's say it finds a region starting at va. > > b) Next it does: > p = mmap(va, ...) > > *without* setting MAP_FIXED, of course (so va is just a hint), to attempt to > safely reserve that region. If p != va, then in most cases, this is a failure > (almost certainly due to another thread getting a mapping from that region > before we did), and so this layer now has to call munmap(), before returning > a "failure: retry" to upper layers. > > IMPROVEMENT: --> if instead, we could call this: > > p = mmap(va, ... MAP_FIXED_NO_CLOBBER ...) > > , then we could skip the munmap() call upon failure. This is a small thing, > but it is useful here. (Thanks to Piotr Jaroszynski and Mark Hairgrove > for helping me get that detail exactly right, btw.) > > c) After that, CUDA suballocates from p, via: > > q = mmap(sub_region_start, ... MAP_FIXED ...) > > Interestingly enough, "freeing" is also done via MAP_FIXED, and setting PROT_NONE > to the subregion. Anyway, I just included (c) for general interest. OK, I will add this to the changelog. This is basically the "Atomic address range probing in the multithreaded programs" I've had in the cover letter. > I expect that as we continue working on the open source compute software stack, > this new capability will be useful there, too. > > Oh, and on the naming, when I described how your implementation worked (without > naming it) to Piotr, he said, "oh, something like map-fixed-no-clobber?". So I > think my miniature sociology naming data point here can bolster the case ever so > slightly for calling it MAP_FIXED_NO_CLOBBER. haha. :) I will be probably stubborn and go with a shorter name I have currently. I am not very fond-of-very-long-names. -- Michal Hocko SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe linux-api" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
