virtio-mem: prepare page onlining code for granularity smaller than MAX_ORDER - 1
authorDavid Hildenbrand <david@redhat.com>
Fri, 26 Nov 2021 13:42:08 +0000 (14:42 +0100)
committerMichael S. Tsirkin <mst@redhat.com>
Fri, 14 Jan 2022 23:50:52 +0000 (18:50 -0500)
Let's prepare our page onlining code for subblock size smaller than
MAX_ORDER - 1: we'll get called for a MAX_ORDER - 1 page but might have
some subblocks in the range plugged and some unplugged. In that case,
fallback to subblock granularity to properly only expose the plugged
parts to the buddy.

Signed-off-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/20211126134209.17332-2-david@redhat.com
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Eric Ren <renzhengeek@gmail.com>
drivers/virtio/virtio_mem.c

index 033fb93..434aba8 100644 (file)
@@ -20,6 +20,7 @@
 #include <linux/mutex.h>
 #include <linux/bitmap.h>
 #include <linux/lockdep.h>
+#include <linux/log2.h>
 
 #include <acpi/acpi_numa.h>
 
@@ -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.
         */