Eric Blake, le Tue 08 May 2012 12:54:08 -0600, a écrit : > On 05/08/2012 12:39 PM, Samuel Thibault wrote: > > The misc/nice test currently fails on hurd-i386 because the nice support > > there is not precise enough: the Mach kernel, which handles scheduling > > priorities, has only 32 priority levels, and not 40, so conversion is > > used, but that becomes not precise enough for the misc/nice test. Apart > > from that, the nice tool works fine. Could you disable that test on > > hurd-i386? > > What is the mapping between the kernel's 32 levels and the > POSIX-mandated minimum of 40 niceness levels (that is, which niceness > levels are rounded to a neighbor)? Can we rewrite the test to use just > niceness levels provided by Hurd?
It's a simple factor 2: #define MACH_PRIORITY_TO_NICE(prio) (2 * ((prio) - 12)) #define NICE_TO_MACH_PRIORITY(nice) (12 + ((nice) / 2)) > In fact, > if you could track 64 niceness levels, and just have every pair of > niceness levels map to the kernel priorities, then you would be > POSIX-compliant (POSIX doesn't require that niceness level 0 and 1 have > any difference in scheduling, only that you can have niceness levels > between 0 and 2*NZERO-1). Right, but that'd means that libc has to remember all priorities, to be able to return them, including other processes (it's not only about nice, but also get/setpriority). OTOH, we could as well extend mach's priorities to 40 levels and change the mappping. The only issue with that is that when using a patched libc (using >32 priority values) with a non-patched kernel (refusing them), some nice values will be refused by kernel. Maybe we can consider it's not a problem that people have to upgrade their kernel to fix it. Samuel
