diff --git a/drivers/virtio/virtio_mem.c b/drivers/virtio/virtio_mem.c index 033fb93ed52847a5d31c81bacae2e28e9b169f37..434aba84524587f8f46d388795203dc5084ef879 100644 --- a/drivers/virtio/virtio_mem.c +++ b/drivers/virtio/virtio_mem.c @@ -20,6 +20,7 @@ #include #include #include +#include #include @@ -1228,28 +1229,46 @@ static void virtio_mem_fake_offline_cancel_offline(unsigned long pfn, page_ref_inc(pfn_to_page(pfn + i)); } -static void virtio_mem_online_page_cb(struct page *page, unsigned int order) +static void virtio_mem_online_page(struct virtio_mem *vm, + struct page *page, unsigned int order) { - const unsigned long addr = page_to_phys(page); - unsigned long id, sb_id; - struct virtio_mem *vm; + const unsigned long start = page_to_phys(page); + const unsigned long end = start + PFN_PHYS(1 << order); + unsigned long addr, next, id, sb_id, count; bool do_online; - rcu_read_lock(); - list_for_each_entry_rcu(vm, &virtio_mem_devices, next) { - if (!virtio_mem_contains_range(vm, addr, PFN_PHYS(1 << order))) - continue; + /* + * We can get called with any order up to MAX_ORDER - 1. If our + * subblock size is smaller than that and we have a mixture of plugged + * and unplugged subblocks within such a page, we have to process in + * smaller granularity. In that case we'll adjust the order exactly once + * within the loop. + */ + for (addr = start; addr < end; ) { + next = addr + PFN_PHYS(1 << order); if (vm->in_sbm) { - /* - * We exploit here that subblocks have at least - * MAX_ORDER_NR_PAGES size/alignment - so we cannot - * cross subblocks within one call. - */ id = virtio_mem_phys_to_mb_id(addr); sb_id = virtio_mem_phys_to_sb_id(vm, addr); - do_online = virtio_mem_sbm_test_sb_plugged(vm, id, - sb_id, 1); + count = virtio_mem_phys_to_sb_id(vm, next - 1) - sb_id + 1; + + if (virtio_mem_sbm_test_sb_plugged(vm, id, sb_id, count)) { + /* Fully plugged. */ + do_online = true; + } else if (count == 1 || + virtio_mem_sbm_test_sb_unplugged(vm, id, sb_id, count)) { + /* Fully unplugged. */ + do_online = false; + } else { + /* + * Mixture, process sub-blocks instead. This + * will be at least the size of a pageblock. + * We'll run into this case exactly once. + */ + order = ilog2(vm->sbm.sb_size) - PAGE_SHIFT; + do_online = virtio_mem_sbm_test_sb_plugged(vm, id, sb_id, 1); + continue; + } } else { /* * If the whole block is marked fake offline, keep @@ -1260,18 +1279,38 @@ static void virtio_mem_online_page_cb(struct page *page, unsigned int order) VIRTIO_MEM_BBM_BB_FAKE_OFFLINE; } + if (do_online) + generic_online_page(pfn_to_page(PFN_DOWN(addr)), order); + else + virtio_mem_set_fake_offline(PFN_DOWN(addr), 1 << order, + false); + addr = next; + } +} + +static void virtio_mem_online_page_cb(struct page *page, unsigned int order) +{ + const unsigned long addr = page_to_phys(page); + struct virtio_mem *vm; + + rcu_read_lock(); + list_for_each_entry_rcu(vm, &virtio_mem_devices, next) { /* - * virtio_mem_set_fake_offline() might sleep, we don't need - * the device anymore. See virtio_mem_remove() how races + * Pages we're onlining will never cross memory blocks and, + * therefore, not virtio-mem devices. + */ + if (!virtio_mem_contains_range(vm, addr, PFN_PHYS(1 << order))) + continue; + + /* + * virtio_mem_set_fake_offline() might sleep. We can safely + * drop the RCU lock at this point because the device + * cannot go away. See virtio_mem_remove() how races * between memory onlining and device removal are handled. */ rcu_read_unlock(); - if (do_online) - generic_online_page(page, order); - else - virtio_mem_set_fake_offline(PFN_DOWN(addr), 1 << order, - false); + virtio_mem_online_page(vm, page, order); return; } rcu_read_unlock(); @@ -2438,8 +2477,7 @@ static int virtio_mem_init_hotplug(struct virtio_mem *vm) /* * We want subblocks to span at least MAX_ORDER_NR_PAGES and * pageblock_nr_pages pages. This: - * - Simplifies our page onlining code (virtio_mem_online_page_cb) - * and fake page onlining code (virtio_mem_fake_online). + * - Simplifies our fake page onlining code (virtio_mem_fake_online). * - Is required for now for alloc_contig_range() to work reliably - * it doesn't properly handle smaller granularity on ZONE_NORMAL. */