Memory Allocation

For allocating large areas of virtually contiguous memory that do not have to be physically contiguous for interfacing with hardware, the new vmalloc() function (with the same calling convention as conventional malloc()) will cause less stress on the memory subsystem. It allocates possibly non-contiguous blocks of free memory, and maps them into one contiguous space in high memory. It is less likely to fail than kmalloc in many situations. It does not take a priority like kamlloc does. It cannot be called from within an interrupt, and it may implicitly cause pre-emption to occur.

Memory allocated with vmalloc is not DMA-able, even on systems without DMA restrictions, because DMA under Linux assumes a 1-1 logical-physical page mapping. This simplifies memory management in several ways, and is not a severe restriction, because kmalloc provides a way to get DMA-capable memory.

Just because it is addressed virtually does not mean that this memory is subject to paging to disk, despite rumors to the contrary. The “virtual” in vmalloc refers only to the addressing, which is not a 1-1 mapping from virtual to physical address space, unlike the rest of the kernel. Swapping may be initiated to provide the memory during a call to vmalloc(), but the vmalloced memory will not then be swapped out.

Memory allocated with vmalloc() is freed with vfree().

Now we learn what the GFP above stands for: get_free_page (well, perhaps __get_free_pages) and simply specifies how exactly this function goes about attempting to find free pages of memory. As you may guess, the same GFP_* values are used for these functions as well.

This is the way to request an amount of memory that is easy for the memory subsystem to allocate. This is the lowest-level—and therefore the lowest-overhead—way of dynamicly allocating memory. If you need a chunk of memory larger than half a page but no larger than a page (when deciding this, be aware that page size varies from architecture to architecture; it is 4Kb on the i86 and 8Kb on the DEC Alpha, for instance), especially if you only need it for the duration of the current procedure, this can be the right way to go. Also, if you are working on subsystem-specific memory management, you almost certainly want to allocate your memory this way.

If you want only one page, call get_free_page(priority), where priority is one of the GFP_* values. Of course, the same rules about which GFP_* value is correct apply as for kmalloc. If you only want one page and don't care if it has been cleared (set to all zero values), use __get_free_page(priority) instead, since most of the overhead of allocating a page with get_free_page goes to clearing the page.

If you need to allocate more than one consecutive page, you can do so, although this is more likely to fail than allocating a single page, and the more pages you wish to allocate, the less likely you are to succeed. You can only allocate a number of pages which is a power of two. __get_free_pages(priority, order) is called with the same priority argument; the order argument gives the size according to the following formula: PAGE_SIZE<\!s>*<\!s>2<+>order<+> so an order of 0 gives one page, of 1 gives two pages, of 2 gives four pages, and so on (in current kernels, at least) up to 5, which gives 32 pages, which on the i86 architecture is 128Kb. PAGE_SIZE, as you may have guessed, is the standard macro for the number of bytes in a page.

The __get_dma_pages() function works exactly like __get_free_pages(), except that it allocates pages which are capable of being used for DMA, and it puts more stress on the memory allocation system.

Pages allocated with get_free_page() or __get_free_page() are freed with free_page(), and pages allocated with __get_free_pages() and __get_dma_pages() are freed with free_pages().

Now we approach the deprecated strategies. They may be useful in some circumstances, mostly in situations where they are the “easy way” to get a driver working. In those cases, it is usually best to eventually find another way to do the same thing, as neither strategy is very flexible. Neither strategy is applicable to loadable modules.

When a device which is compiled into the kernel is initialized, it is passed a pointer to available memory. It is then required to return a pointer. If the pointer it returns is higher than the pointer it got, the memory in between the two pointers is reserved for the device. That memory will be in the first few megabytes of memory. The exact location will depend on how the kernel is booted. This is (perhaps unfortunately...) documented in the Linux Kernel Hackers' Guide.

Once allocated, that memory cannot be freed.