Hi Mauricio... > > I think, even the addresses seem equal, physically they are > > different > > > > pages. IMHO this is done so that the kernel doesn't need to > > "tangle" with SP (stack pointer). > > So you are saying that each thread belongs to an application contains > its own stack in terms of physical pages and not virtual memory, > right? Yes, that's what I meant. The address is still the same, but they use different physical pages... > If a CLONE_VM is used as a flag for clone() function, the stack area > pointed by memory descriptor "mm" just means the threads share > virtually the same stack. IMHO CLONE_VM means sharing address space. Thus, we see same set of virtual address reference between threads. Or more technically, meaning they share a same page directory. Then yes, it means "virtually sharing the same stack".... > According to your explanation, each thread uses a VMA as a stack > area, but the maps entry under proc filesystem shows an unique VMA > considered as stack for all threads, since their virtual addresses > are equal. This happens with Firefox web browser, for instance the > threads with PID 4274 and 4275 have the same VMA according to maps > output: > > bfd47000-bfd5b000 rwxp bfd47000 00:00 0 [stack]. > > Any comments? IMHO, that means each Firefox threads don't ask for creation of unique stack area (both virtually and physically). Thus, each threads refer to the same virtual address. > Or there is just one common VMA considered as a stack and all threads > point in different location in this virtually shared stack along the > time? I think, the stack area is COW-ed between threads. About Stack Pointers pointing to different virtual address, I am not so sure. I think that is a bit risky to do, since in Linux, stack segment does cover the whole user address space (flat model) so one cannot restrict how much the stack area can grow. > If COW happens to stack, it means that CLONE_VM was not set during > the creation of child process, so the parent and the child will not > share the same address space. My question is related when VM_CLONE is > set, so the parent and child will use the same memory space, but how > stack is used among many threads in this situation? Like I wrote above, IMHO CLONE_VM means sharing the same address space, thus COW is perfectly ok to be done even CLONE_VM flag is used. The kernel just need to update the related PTEs so that both threads still use the same virtual address set. > I agree with you that each thread should contain its own stack, but > again the maps output just show one VMA related to stack for all > threads. I hope with all I explained above, you can pick your own conclusion for your above question :) BTW, I am sorry if what I wrote isn't perfectly clear. Any feedbacks are greatly appreciated. regards, Mulyadi -- Kernelnewbies: Help each other learn about the Linux kernel. Archive: http://mail.nl.linux.org/kernelnewbies/ FAQ: http://kernelnewbies.org/faq/
