powerpc/pseries/cmm: Implement balloon compaction
authorDavid Hildenbrand <david@redhat.com>
Thu, 31 Oct 2019 14:29:29 +0000 (15:29 +0100)
committerMichael Ellerman <mpe@ellerman.id.au>
Wed, 13 Nov 2019 05:58:01 +0000 (16:58 +1100)
We can now get rid of the cmm_lock and completely rely on the balloon
compaction internals, which now also manage the page list and the
lock.

Inflated/"loaned" pages are now movable. Memory blocks that contain
such pages can get offlined. Also, all such pages will be marked
PageOffline() and can therefore be excluded in memory dumps using
recent versions of makedumpfile.

Don't switch to balloon_page_alloc() yet (due to the GFP_NOIO). Will
do that separately to discuss this change in detail.

Signed-off-by: David Hildenbrand <david@redhat.com>
[mpe: Add isolated_pages-- in cmm_migratepage() as suggested by David]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20191031142933.10779-9-david@redhat.com
arch/powerpc/platforms/pseries/Kconfig
arch/powerpc/platforms/pseries/cmm.c
include/uapi/linux/magic.h

index 9e35cddddf7307df8c39782dd9cd008ec3aa1638..595e9f8a65398d6653cf9bcb5c7606d7ab920c47 100644 (file)
@@ -108,6 +108,7 @@ config PPC_SMLPAR
 config CMM
        tristate "Collaborative memory management"
        depends on PPC_SMLPAR
+       select MEMORY_BALLOON
        default y
        help
          Select this option, if you want to enable the kernel interface
index 3a55dd1fdd392105df129ee36a4c6957d72c774c..0663ddad964af95b248934ec5a2ed6e41ff1fb05 100644 (file)
 #include <linux/stringify.h>
 #include <linux/swap.h>
 #include <linux/device.h>
+#include <linux/mount.h>
+#include <linux/pseudo_fs.h>
+#include <linux/magic.h>
+#include <linux/balloon_compaction.h>
 #include <asm/firmware.h>
 #include <asm/hvcall.h>
 #include <asm/mmu.h>
@@ -77,13 +81,11 @@ static atomic_long_t loaned_pages;
 static unsigned long loaned_pages_target;
 static unsigned long oom_freed_pages;
 
-static LIST_HEAD(cmm_page_list);
-static DEFINE_SPINLOCK(cmm_lock);
-
 static DEFINE_MUTEX(hotplug_mutex);
 static int hotplug_occurred; /* protected by the hotplug mutex */
 
 static struct task_struct *cmm_thread_ptr;
+static struct balloon_dev_info b_dev_info;
 
 static long plpar_page_set_loaned(struct page *page)
 {
@@ -149,19 +151,16 @@ static long cmm_alloc_pages(long nr)
                                  __GFP_NOMEMALLOC);
                if (!page)
                        break;
-               spin_lock(&cmm_lock);
                rc = plpar_page_set_loaned(page);
                if (rc) {
                        pr_err("%s: Can not set page to loaned. rc=%ld\n", __func__, rc);
-                       spin_unlock(&cmm_lock);
                        __free_page(page);
                        break;
                }
 
-               list_add(&page->lru, &cmm_page_list);
+               balloon_page_enqueue(&b_dev_info, page);
                atomic_long_inc(&loaned_pages);
                adjust_managed_page_count(page, -1);
-               spin_unlock(&cmm_lock);
                nr--;
        }
 
@@ -178,21 +177,19 @@ static long cmm_alloc_pages(long nr)
  **/
 static long cmm_free_pages(long nr)
 {
-       struct page *page, *tmp;
+       struct page *page;
 
        cmm_dbg("Begin free of %ld pages.\n", nr);
-       spin_lock(&cmm_lock);
-       list_for_each_entry_safe(page, tmp, &cmm_page_list, lru) {
-               if (!nr)
+       while (nr) {
+               page = balloon_page_dequeue(&b_dev_info);
+               if (!page)
                        break;
                plpar_page_set_active(page);
-               list_del(&page->lru);
                adjust_managed_page_count(page, 1);
                __free_page(page);
                atomic_long_dec(&loaned_pages);
                nr--;
        }
-       spin_unlock(&cmm_lock);
        cmm_dbg("End request with %ld pages unfulfilled\n", nr);
        return nr;
 }
@@ -484,6 +481,106 @@ static struct notifier_block cmm_mem_nb = {
        .priority = CMM_MEM_HOTPLUG_PRI
 };
 
+#ifdef CONFIG_BALLOON_COMPACTION
+static struct vfsmount *balloon_mnt;
+
+static int cmm_init_fs_context(struct fs_context *fc)
+{
+       return init_pseudo(fc, PPC_CMM_MAGIC) ? 0 : -ENOMEM;
+}
+
+static struct file_system_type balloon_fs = {
+       .name = "ppc-cmm",
+       .init_fs_context = cmm_init_fs_context,
+       .kill_sb = kill_anon_super,
+};
+
+static int cmm_migratepage(struct balloon_dev_info *b_dev_info,
+                          struct page *newpage, struct page *page,
+                          enum migrate_mode mode)
+{
+       unsigned long flags;
+
+       /*
+        * loan/"inflate" the newpage first.
+        *
+        * We might race against the cmm_thread who might discover after our
+        * loan request that another page is to be unloaned. However, once
+        * the cmm_thread runs again later, this error will automatically
+        * be corrected.
+        */
+       if (plpar_page_set_loaned(newpage)) {
+               /* Unlikely, but possible. Tell the caller not to retry now. */
+               pr_err_ratelimited("%s: Cannot set page to loaned.", __func__);
+               return -EBUSY;
+       }
+
+       /* balloon page list reference */
+       get_page(newpage);
+
+       spin_lock_irqsave(&b_dev_info->pages_lock, flags);
+       balloon_page_insert(b_dev_info, newpage);
+       balloon_page_delete(page);
+       b_dev_info->isolated_pages--;
+       spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
+
+       /*
+        * activate/"deflate" the old page. We ignore any errors just like the
+        * other callers.
+        */
+       plpar_page_set_active(page);
+
+       /* balloon page list reference */
+       put_page(page);
+
+       return MIGRATEPAGE_SUCCESS;
+}
+
+static int cmm_balloon_compaction_init(void)
+{
+       int rc;
+
+       balloon_devinfo_init(&b_dev_info);
+       b_dev_info.migratepage = cmm_migratepage;
+
+       balloon_mnt = kern_mount(&balloon_fs);
+       if (IS_ERR(balloon_mnt)) {
+               rc = PTR_ERR(balloon_mnt);
+               balloon_mnt = NULL;
+               return rc;
+       }
+
+       b_dev_info.inode = alloc_anon_inode(balloon_mnt->mnt_sb);
+       if (IS_ERR(b_dev_info.inode)) {
+               rc = PTR_ERR(b_dev_info.inode);
+               b_dev_info.inode = NULL;
+               kern_unmount(balloon_mnt);
+               balloon_mnt = NULL;
+               return rc;
+       }
+
+       b_dev_info.inode->i_mapping->a_ops = &balloon_aops;
+       return 0;
+}
+static void cmm_balloon_compaction_deinit(void)
+{
+       if (b_dev_info.inode)
+               iput(b_dev_info.inode);
+       b_dev_info.inode = NULL;
+       kern_unmount(balloon_mnt);
+       balloon_mnt = NULL;
+}
+#else /* CONFIG_BALLOON_COMPACTION */
+static int cmm_balloon_compaction_init(void)
+{
+       return 0;
+}
+
+static void cmm_balloon_compaction_deinit(void)
+{
+}
+#endif /* CONFIG_BALLOON_COMPACTION */
+
 /**
  * cmm_init - Module initialization
  *
@@ -497,9 +594,14 @@ static int cmm_init(void)
        if (!firmware_has_feature(FW_FEATURE_CMO))
                return -EOPNOTSUPP;
 
-       if ((rc = register_oom_notifier(&cmm_oom_nb)) < 0)
+       rc = cmm_balloon_compaction_init();
+       if (rc)
                return rc;
 
+       rc = register_oom_notifier(&cmm_oom_nb);
+       if (rc < 0)
+               goto out_balloon_compaction;
+
        if ((rc = register_reboot_notifier(&cmm_reboot_nb)))
                goto out_oom_notifier;
 
@@ -527,6 +629,8 @@ out_reboot_notifier:
        unregister_reboot_notifier(&cmm_reboot_nb);
 out_oom_notifier:
        unregister_oom_notifier(&cmm_oom_nb);
+out_balloon_compaction:
+       cmm_balloon_compaction_deinit();
        return rc;
 }
 
@@ -545,6 +649,7 @@ static void cmm_exit(void)
        unregister_memory_notifier(&cmm_mem_nb);
        cmm_free_pages(atomic_long_read(&loaned_pages));
        cmm_unregister_sysfs(&cmm_dev);
+       cmm_balloon_compaction_deinit();
 }
 
 /**
index 903cc2d2750b24f4bc48cbdce4a71c5543e41137..3ac436376d79b6313fc8be536b4bde09cb32ea92 100644 (file)
@@ -94,5 +94,6 @@
 #define ZSMALLOC_MAGIC         0x58295829
 #define DMA_BUF_MAGIC          0x444d4142      /* "DMAB" */
 #define Z3FOLD_MAGIC           0x33
+#define PPC_CMM_MAGIC          0xc7571590
 
 #endif /* __LINUX_MAGIC_H__ */