2008/1/18, Linus Torvalds <[EMAIL PROTECTED]>: > > > On Fri, 18 Jan 2008, Anton Salikhmetov wrote: > > > > The current solution doesn't hit the performance at all when compared to > > the competitor POSIX-compliant systems. It is faster and does even more > > than the POSIX standard requires. > > Your current patches have two problems: > - they are simply unnecessarily invasive for a relatively simple issue > - all versions I've looked at closer are buggy too > > Example: > > + if (pte_dirty(*pte) && pte_write(*pte)) > + *pte = pte_wrprotect(*pte); > > Uhhuh. Looks simple enough. Except it does a non-atomic pte access while > other CPU's may be accessing it and updating it from their hw page table > walkers. What will happen? Who knows? I can see lost access bits at a > minimum. > > IOW, this isn't simple code. It's code that it is simple to screw up. In > this case, you really need to use ptep_set_wrprotect(), for example.
Before using pte_wrprotect() the vma_wrprotect() routine uses the pte_offset_map_lock() macro to get the PTE and to acquire the ptl spinlock. Why did you say that this code was not SMP-safe? It should be atomic, I think. > > So why not do it in many fewer lines with that simpler vma->dirty flag? Neither the dirty flag you suggest, nor the AS_MCTIME flag I've introduced in my previous solutions solve the following problem: - mmap() - a write reference - msync() with MS_ASYNC - a write reference - msync() with MS_ASYNC The POSIX standard requires the ctime and mtime stamps to be updated not later than at the second call to msync() with the MS_ASYNC flag. Some other POSIX-compliant operating system such as HP-UX and FreeBSD satisfy this POSIX requirement. Linux does not. > > Linus > -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
