On 2014/8/11 19:54, Peter Zijlstra wrote:
On Mon, Aug 11, 2014 at 01:26:45AM +0000, Sha, Ruibin wrote:Hi Chintan, Thank you very much for your timely and kindly response and comments. Here is more detail about our Scenario: We have a big driver on Android product. The driver allocates lots of DDR pages. When applications mmap a file exported from the driver, driver would mmap the pages to the application space, usually with uncachable prot. On ia32/x86_64 arch, we have to avoid page cache alias issue. When driver allocates the pages, it would change page original mapping in page table with uncachable prot. Sometimes, the allocated page was used by kmap/kunmap. After kunmap, the page is still mapped in KMAP space. The entries in KMAP page table are not cleaned up until a kernel thread flushes the freed KMAP pages(usually it is woken up by kunmap). It means the driver need force to flush the KMAP page table entries before mapping pages to application space to be used. Otherwise, there is a race to create cache alias. To resolve this issue, we need export function kmap_flush_unused as the driver is compiled as module. Then, the driver calls kmap_flush_unused if the allocated pages are in HIGHMEM and being used by kmap.A: Because it messes up the order in which people normally read text. Q: Why is top-posting such a bad thing? A: Top-posting. Q: What is the most annoying thing in e-mail?
Sorry, Peter. Ruibin is a new guy in LKML community. He uses outlook to send emails. He would improve that.
That said, it sounds like you want set_memory_() to call kmap_flush_unused(). Because this race it not at all specific to your usage, it could happen to any set_memory_() site, right?
No. set_memory_() assumes the memory is not in HIGHMEM. This scenario is driver allocates HIGHMEM pages, which are kmapped before. Kernel uses a lazy method when kunmap a HIGHMEM page. The pages are not unmapped from KMAP page table entries immediately. When next kmap calling uses the same entry, kernel would change pte. Or when change_page_attr_set_clr is called. Our big driver doesn't call change_page_attr_set_clr when mmap the pages with UNCACHABLE prot. It need call kmap_flush_unused directly after allocating HIGHMEM pages. Thanks for the kind comments. Yanmin -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>