"Kirk True" <ktrue@movaris.com> writes: > Hi all, > > 1. Regarding non-contiguous memory allocation, what is the need to have > *virtually* contiguous but not *physically* contiguous pages? This is the essential purpose of virtual memory management. Say a user-space program wants to allocate a 1MB array for some reason. In userspace, we want that memory to appear virtually contiguous (so that int x = *(bigArray+50000) works as expected), but we really couldn't care less where the pages are physically located. The same is often true in the kernel - if you need to allocate a 32K buffer for some reason, and there's no particular reason (such as DMA requirements) to make the buffer physically contiguous, then it's usually a lot easier to find 8 singleton free pages than it is to find 8 contiguous free pages. > 2. UtLVMM says that vmalloc is only used in the kernel for storing swap > information - yet it's used by a bunch of drivers which are > considered part of the kernel; is it just semantics? Perhaps it really meant, the VM subsystem only uses vmalloc in this fashion? I don't know, the statement just sounds wrong to me. > 3. Is vmalloc called from user-mode ever? No. Nothing in the kernel is ever called from user-mode, ever. The only way for userspace code to invoke kernel code is by making a user->kernel context switch (system call). > 4. What is the difference between kmalloc, malloc, and vmalloc? kmalloc and vmalloc are kernel facilities that are totally unrelated to malloc, the C standard library function. kmalloc (IIRC) allocates contiguous regions of physical pages in kernelspace. (It is a wrapper around get_free_pages().) vmalloc allocates arbitrary collections of physical pages, but maps them virtually-contiguously into kernelspace. malloc is called from userspace when a user program wants to allocate anonymous virtual memory. It makes a system call and requests the appropriate allocation from the kernel, which merely does the VM-area bookkeeping necessary to *logically* associate the requested VM with the process. The memory is not physically allocated unless the userspace code attempts to access it (in which case the pagefault handler serves the fault by consulting the proc's VM-areas and physically allocating a page to the process). > 5. Anonymous memory is memory that is *not* backed by a file, such as > the stack or heap space, right? And mmap is called when mapping > files into memory, right? The why does mmap deal with anonymous > memory (sorry, I'm totally confused here)? Anonymous memory *is* (sometimes) backed by a file: the swap file. However, I've never really looked closely at the mmap() imeplementation, so I don't know if this fact addresses your question or not. It may also have to do with shared memory management (eg the shmat() etc. system calls). I believe these are implemented in terms of mmap(), though I may well be wrogn (as I've never bothered to look). HTH, -- Joe Knapka -- (let ((antichrist 'me) (anarchist 'me))) -- the sexp-pistols. -- Kernelnewbies: Help each other learn about the Linux kernel. Archive: http://mail.nl.linux.org/kernelnewbies/ FAQ: http://kernelnewbies.org/faq/
