> On Thu, 30 Jul 2009, erik quanstrom wrote: > >> On Thu Jul 30 00:05:45 EDT 2009, el...@andrew.cmu.edu wrote: >>> My familiarity with the kernel source code is superficial to say the >>> least, but it seems to me that this code (from /sys/src/9/pc/trap.c) >>> contains a race condition: >>> >>> 702 if(sp<(USTKTOP-BY2PG) || sp>(USTKTOP-sizeof(Sargs)-BY2WD)) >>> 703 validaddr(sp, sizeof(Sargs)+BY2WD, 0); >>> 704 >>> 705 up->s = *((Sargs*)(sp+BY2WD)); >>> >>> We verify that the address is correct; is there any reason another thread >>> in the same address space cannot start running after line 703 completes >>> and unmap that memory, causing us to access unmapped memory at line 705? >>> The system call entry is itself an interrupt gate, but line 689 is >>> spllo(), and we appear to hold no locks at this point. >> >> plan 9 threads are cooperatively scheduled. so >> the correct term is proc. but you are correct, >> another proc sharing memory with this one >> could be running. however, that proc would >> not have access to this proc's stack. (rfork >> doesn't allow shared stack.) and even if it >> did, plan 9 stacks don't shrink. > > What if sp points inside a segment which is not the actual stack segment? > Then could someone else come along and segdetach() it in between the two > mentioned lines? > >> let's suppose that the address is invalid later. >> the kernel always moves data to/from user >> buffers outside of any locks because even >> valid targets may be paged out. if the address >> is truely invalid, waserror() will be true and >> the else branch starting at 714 will be executed >> >> - erik > > -- Elly
I think you may be right, Elly. Multithreaded programs indeed have their stack running outside the stack segment, so this could happen there. splhi won't even do on a multiprocessor. One should probably lock down the segment. We've never seen this happen, of course — or rather, we haven't noticed this as the cause of a crash. Sape