Hi Leonid, On Thu, Oct 09, 2014 at 05:07:11PM -0700, Leonid Yegoshin wrote: > On 10/09/2014 04:40 PM, James Hogan wrote: > > You could then avoid the whole stack and per-thread thing and just have > > a maximum of one emuframe dedicated to each thread or allocated on > > demand, and if there genuinely is a use case for nesting later on, worry > > about it then. > > As I understand, you propose to allocate some space in mmap. No, sorry if I wasn't very clear. I just mean that you can get away with a single kernel managed page per mm, with an emuframe allocated per-thread which that thread always uses, since they never nest, which I think simplifies the whole thing significantly. The allocation could be smarter than that of course in case you have thousands of threads and only a subset doing lots of FP branches, but a single thread should never need more than one at a time since the new signal behaviour effectively makes the delay slot emulation sort of atomic from the point of view of usermode, and the kernel knows for sure whether BD emulation is in progress from the PC. (If there is some other way than signals that I haven't taken into account that the emulation could be pre-empted then please let me know!) > > So long as the kernel handles a long sequence of sequential emulated > > branches gracefully (not necessarily correctly). > > > I don't understand a question. Each pair/single instruction is emulated > separately but there is some pipeline of that, even in FPU emulator, it > is just not this patch issue. I just mean an (illegal/undefined) sequence of FPU branch instructions in one anothers delay slots shouldn't be able to crash the kernel. Actually 2 of them would be enough to verify the kernel didn't get too confused. Maybe the second will be detected & ignored, or maybe it doesn't matter if the first emuframe gets overwritten by the second one from the kernels point of view. Cheers James
