Message-ID: <20180920173026.141333-18-namit@vmware.com>
Date:   Thu, 20 Sep 2018 10:30:23 -0700
From:   Nadav Amit <namit@...are.com>
To:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Arnd Bergmann <arnd@...db.de>
CC:     <linux-kernel@...r.kernel.org>,
        Xavier Deguillard <xdeguillard@...are.com>,
        Nadav Amit <namit@...are.com>
Subject: [PATCH v2 17/20] vmw_balloon: compaction support

Add support for compaction for VMware balloon. Since unlike the virtio
balloon, we also support huge-pages, which are not going through
compaction, we keep these pages in vmballoon and handle this list
separately. We use the same lock to protect both lists, as this lock is
not supposed to be contended.

Doing so also eliminates the need for the page_size lists. We update the
accounting as needed to reflect inflation, deflation and migration to be
reflected in vmstat.

Since VMware balloon now provides statistics for inflation, deflation
and migration in vmstat, select MEMORY_BALLOON in Kconfig.

Reviewed-by: Xavier Deguillard <xdeguillard@...are.com>
Signed-off-by: Nadav Amit <namit@...are.com>
---
 drivers/misc/Kconfig       |   1 +
 drivers/misc/vmw_balloon.c | 299 ++++++++++++++++++++++++++++++++-----
 include/uapi/linux/magic.h |   1 +
 3 files changed, 263 insertions(+), 38 deletions(-)

diff --git a/drivers/misc/Kconfig b/drivers/misc/Kconfig
index 3726eacdf65d..98e684961ece 100644
--- a/drivers/misc/Kconfig
+++ b/drivers/misc/Kconfig
@@ -409,6 +409,7 @@ config SPEAR13XX_PCIE_GADGET
 config VMWARE_BALLOON
 	tristate "VMware Balloon Driver"
 	depends on VMWARE_VMCI && X86 && HYPERVISOR_GUEST
+	select MEMORY_BALLOON
 	help
 	  This is VMware physical memory management driver which acts
 	  like a "balloon" that can be inflated to reclaim physical pages
diff --git a/drivers/misc/vmw_balloon.c b/drivers/misc/vmw_balloon.c
index 9b0b3fa4f836..cace823bc68c 100644
--- a/drivers/misc/vmw_balloon.c
+++ b/drivers/misc/vmw_balloon.c
@@ -28,6 +28,9 @@
 #include <linux/rwsem.h>
 #include <linux/slab.h>
 #include <linux/spinlock.h>
+#include <linux/mount.h>
+#include <linux/magic.h>
+#include <linux/balloon_compaction.h>
 #include <linux/vmw_vmci_defs.h>
 #include <linux/vmw_vmci_api.h>
 #include <asm/hypervisor.h>
@@ -39,23 +42,6 @@ MODULE_ALIAS("dmi:*:svnVMware*:*");
 MODULE_ALIAS("vmware_vmmemctl");
 MODULE_LICENSE("GPL");
 
-/*
- * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't allow wait
- * (__GFP_RECLAIM) for huge page allocations. Use __GFP_NOWARN, to suppress page
- * allocation failure warnings. Disallow access to emergency low-memory pools.
- */
-#define VMW_HUGE_PAGE_ALLOC_FLAGS	(__GFP_HIGHMEM|__GFP_NOWARN|	\
-					 __GFP_NOMEMALLOC)
-
-/*
- * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We allow lightweight
- * reclamation (__GFP_NORETRY). Use __GFP_NOWARN, to suppress page allocation
- * failure warnings. Disallow access to emergency low-memory pools.
- */
-#define VMW_PAGE_ALLOC_FLAGS		(__GFP_HIGHMEM|__GFP_NOWARN|	\
-					 __GFP_NOMEMALLOC|__GFP_NORETRY)
-
-/* Maximum number of refused pages we accumulate during inflation cycle */
 #define VMW_BALLOON_MAX_REFUSED		16
 
 /*
@@ -237,11 +223,6 @@ struct vmballoon_ctl {
 	enum vmballoon_op op;
 };
 
-struct vmballoon_page_size {
-	/* list of reserved physical pages */
-	struct list_head pages;
-};
-
 /**
  * struct vmballoon_batch_entry - a batch entry for lock or unlock.
  *
@@ -256,8 +237,6 @@ struct vmballoon_batch_entry {
 } __packed;
 
 struct vmballoon {
-	struct vmballoon_page_size page_sizes[VMW_BALLOON_NUM_PAGE_SIZES];
-
 	/**
 	 * @max_page_size: maximum supported page size for ballooning.
 	 *
@@ -338,8 +317,20 @@ struct vmballoon {
 	struct dentry *dbg_entry;
 #endif
 
+	/**
+	 * @b_dev_info: balloon device information descriptor.
+	 */
+	struct balloon_dev_info b_dev_info;
+
 	struct delayed_work dwork;
 
+	/**
+	 * @huge_pages - list of the inflated 2MB pages.
+	 *
+	 * Protected by @b_dev_info.pages_lock .
+	 */
+	struct list_head huge_pages;
+
 	/**
 	 * @vmci_doorbell.
 	 *
@@ -602,10 +593,10 @@ static int vmballoon_alloc_page_list(struct vmballoon *b,
 
 	for (i = 0; i < req_n_pages; i++) {
 		if (ctl->page_size == VMW_BALLOON_2M_PAGE)
-			page = alloc_pages(VMW_HUGE_PAGE_ALLOC_FLAGS,
-					   VMW_BALLOON_2M_ORDER);
+			page = alloc_pages(__GFP_HIGHMEM|__GFP_NOWARN|
+					__GFP_NOMEMALLOC, VMW_BALLOON_2M_ORDER);
 		else
-			page = alloc_page(VMW_PAGE_ALLOC_FLAGS);
+			page = balloon_page_alloc();
 
 		/* Update statistics */
 		vmballoon_stats_page_inc(b, VMW_BALLOON_PAGE_STAT_ALLOC,
@@ -918,9 +909,22 @@ static void vmballoon_enqueue_page_list(struct vmballoon *b,
 					unsigned int *n_pages,
 					enum vmballoon_page_size_type page_size)
 {
-	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
+	unsigned long flags;
+
+	if (page_size == VMW_BALLOON_4K_PAGE) {
+		balloon_page_list_enqueue(&b->b_dev_info, pages);
+	} else {
+		/*
+		 * Keep the huge pages in a local list which is not available
+		 * for the balloon compaction mechanism.
+		 */
+		spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
+		list_splice_init(pages, &b->huge_pages);
+		__count_vm_events(BALLOON_INFLATE, *n_pages *
+				  vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
+		spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
+	}
 
-	list_splice_init(pages, &page_size_info->pages);
 	*n_pages = 0;
 }
 
@@ -943,15 +947,28 @@ static void vmballoon_dequeue_page_list(struct vmballoon *b,
 					enum vmballoon_page_size_type page_size,
 					unsigned int n_req_pages)
 {
-	struct vmballoon_page_size *page_size_info = &b->page_sizes[page_size];
 	struct page *page, *tmp;
 	unsigned int i = 0;
+	unsigned long flags;
 
-	list_for_each_entry_safe(page, tmp, &page_size_info->pages, lru) {
+	/* In the case of 4k pages, use the compaction infrastructure */
+	if (page_size == VMW_BALLOON_4K_PAGE) {
+		*n_pages = balloon_page_list_dequeue(&b->b_dev_info, pages,
+						     n_req_pages);
+		return;
+	}
+
+	/* 2MB pages */
+	spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
+	list_for_each_entry_safe(page, tmp, &b->huge_pages, lru) {
 		list_move(&page->lru, pages);
 		if (++i == n_req_pages)
 			break;
 	}
+
+	__count_vm_events(BALLOON_DEFLATE,
+			  i * vmballoon_page_in_frames(VMW_BALLOON_2M_PAGE));
+	spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
 	*n_pages = i;
 }
 
@@ -1519,9 +1536,204 @@ static inline void vmballoon_debugfs_exit(struct vmballoon *b)
 
 #endif	/* CONFIG_DEBUG_FS */
 
+
+#ifdef CONFIG_BALLOON_COMPACTION
+
+static struct dentry *vmballoon_mount(struct file_system_type *fs_type,
+				      int flags, const char *dev_name,
+				      void *data)
+{
+	static const struct dentry_operations ops = {
+		.d_dname = simple_dname,
+	};
+
+	return mount_pseudo(fs_type, "balloon-vmware:", NULL, &ops,
+			    BALLOON_VMW_MAGIC);
+}
+
+static struct file_system_type vmballoon_fs = {
+	.name           = "balloon-vmware",
+	.mount          = vmballoon_mount,
+	.kill_sb        = kill_anon_super,
+};
+
+static struct vfsmount *vmballoon_mnt;
+
+/**
+ * vmballoon_migratepage() - migrates a balloon page.
+ * @b_dev_info: balloon device information descriptor.
+ * @newpage: the page to which @page should be migrated.
+ * @page: a ballooned page that should be migrated.
+ * @mode: migration mode, ignored.
+ *
+ * This function is really open-coded, but that is according to the interface
+ * that balloon_compaction provides.
+ *
+ * Return: zero on success, -EAGAIN when migration cannot be performed
+ *	   momentarily, and -EBUSY if migration failed and should be retried
+ *	   with that specific page.
+ */
+static int vmballoon_migratepage(struct balloon_dev_info *b_dev_info,
+				 struct page *newpage, struct page *page,
+				 enum migrate_mode mode)
+{
+	unsigned long status, flags;
+	struct vmballoon *b;
+	int ret;
+
+	b = container_of(b_dev_info, struct vmballoon, b_dev_info);
+
+	/*
+	 * If the semaphore is taken, there is ongoing configuration change
+	 * (i.e., balloon reset), so try again.
+	 */
+	if (!down_read_trylock(&b->conf_sem))
+		return -EAGAIN;
+
+	spin_lock(&b->comm_lock);
+	/*
+	 * We must start by deflating and not inflating, as otherwise the
+	 * hypervisor may tell us that it has enough memory and the new page is
+	 * not needed. Since the old page is isolated, we cannot use the list
+	 * interface to unlock it, as the LRU field is used for isolation.
+	 * Instead, we use the native interface directly.
+	 */
+	vmballoon_add_page(b, 0, page);
+	status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
+				   VMW_BALLOON_DEFLATE);
+
+	if (status == VMW_BALLOON_SUCCESS)
+		status = vmballoon_status_page(b, 0, &page);
+
+	/*
+	 * If a failure happened, let the migration mechanism know that it
+	 * should not retry.
+	 */
+	if (status != VMW_BALLOON_SUCCESS) {
+		spin_unlock(&b->comm_lock);
+		ret = -EBUSY;
+		goto out_unlock;
+	}
+
+	/*
+	 * The page is isolated, so it is safe to delete it without holding
+	 * @pages_lock . We keep holding @comm_lock since we will need it in a
+	 * second.
+	 */
+	balloon_page_delete(page);
+
+	put_page(page);
+
+	/* Inflate */
+	vmballoon_add_page(b, 0, newpage);
+	status = vmballoon_lock_op(b, 1, VMW_BALLOON_4K_PAGE,
+				   VMW_BALLOON_INFLATE);
+
+	if (status == VMW_BALLOON_SUCCESS)
+		status = vmballoon_status_page(b, 0, &newpage);
+
+	spin_unlock(&b->comm_lock);
+
+	if (status != VMW_BALLOON_SUCCESS) {
+		/*
+		 * A failure happened. While we can deflate the page we just
+		 * inflated, this deflation can also encounter an error. Instead
+		 * we will decrease the size of the balloon to reflect the
+		 * change and report failure.
+		 */
+		atomic64_dec(&b->size);
+		ret = -EBUSY;
+	} else {
+		/*
+		 * Success. Take a reference for the page, and we will add it to
+		 * the list after acquiring the lock.
+		 */
+		get_page(newpage);
+		ret = MIGRATEPAGE_SUCCESS;
+	}
+
+	/* Update the balloon list under the @pages_lock */
+	spin_lock_irqsave(&b->b_dev_info.pages_lock, flags);
+
+	/*
+	 * On inflation success, we already took a reference for the @newpage.
+	 * If we succeed just insert it to the list and update the statistics
+	 * under the lock.
+	 */
+	if (ret == MIGRATEPAGE_SUCCESS) {
+		balloon_page_insert(&b->b_dev_info, newpage);
+		__count_vm_event(BALLOON_MIGRATE);
+	}
+
+	/*
+	 * We deflated successfully, so regardless to the inflation success, we
+	 * need to reduce the number of isolated_pages.
+	 */
+	b->b_dev_info.isolated_pages--;
+	spin_unlock_irqrestore(&b->b_dev_info.pages_lock, flags);
+
+out_unlock:
+	up_read(&b->conf_sem);
+	return ret;
+}
+
+/**
+ * vmballoon_compaction_deinit() - removes compaction related data.
+ *
+ * @b: pointer to the balloon.
+ */
+static void vmballoon_compaction_deinit(struct vmballoon *b)
+{
+	if (!IS_ERR(b->b_dev_info.inode))
+		iput(b->b_dev_info.inode);
+
+	b->b_dev_info.inode = NULL;
+	kern_unmount(vmballoon_mnt);
+	vmballoon_mnt = NULL;
+}
+
+/**
+ * vmballoon_compaction_init() - initialized compaction for the balloon.
+ *
+ * @b: pointer to the balloon.
+ *
+ * If during the initialization a failure occurred, this function does not
+ * perform cleanup. The caller must call vmballoon_compaction_deinit() in this
+ * case.
+ *
+ * Return: zero on success or error code on failure.
+ */
+static __init int vmballoon_compaction_init(struct vmballoon *b)
+{
+	vmballoon_mnt = kern_mount(&vmballoon_fs);
+	if (IS_ERR(vmballoon_mnt))
+		return PTR_ERR(vmballoon_mnt);
+
+	b->b_dev_info.migratepage = vmballoon_migratepage;
+	b->b_dev_info.inode = alloc_anon_inode(vmballoon_mnt->mnt_sb);
+
+	if (IS_ERR(b->b_dev_info.inode))
+		return PTR_ERR(b->b_dev_info.inode);
+
+	b->b_dev_info.inode->i_mapping->a_ops = &balloon_aops;
+	return 0;
+}
+
+#else /* CONFIG_BALLOON_COMPACTION */
+
+static void vmballoon_compaction_deinit(struct vmballoon *b)
+{
+}
+
+static int vmballoon_compaction_init(struct vmballoon *b)
+{
+	return 0;
+}
+
+#endif /* CONFIG_BALLOON_COMPACTION */
+
 static int __init vmballoon_init(void)
 {
-	enum vmballoon_page_size_type page_size;
 	int error;
 
 	/*
@@ -1531,17 +1743,22 @@ static int __init vmballoon_init(void)
 	if (x86_hyper_type != X86_HYPER_VMWARE)
 		return -ENODEV;
 
-	for (page_size = VMW_BALLOON_4K_PAGE;
-	     page_size <= VMW_BALLOON_LAST_SIZE; page_size++)
-		INIT_LIST_HEAD(&balloon.page_sizes[page_size].pages);
-
-
 	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
 
 	error = vmballoon_debugfs_init(&balloon);
 	if (error)
-		return error;
+		goto fail;
 
+	/*
+	 * Initialization of compaction must be done after the call to
+	 * balloon_devinfo_init() .
+	 */
+	balloon_devinfo_init(&balloon.b_dev_info);
+	error = vmballoon_compaction_init(&balloon);
+	if (error)
+		goto fail;
+
+	INIT_LIST_HEAD(&balloon.huge_pages);
 	spin_lock_init(&balloon.comm_lock);
 	init_rwsem(&balloon.conf_sem);
 	balloon.vmci_doorbell = VMCI_INVALID_HANDLE;
@@ -1552,6 +1769,9 @@ static int __init vmballoon_init(void)
 	queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
 
 	return 0;
+fail:
+	vmballoon_compaction_deinit(&balloon);
+	return error;
 }
 
 /*
@@ -1576,5 +1796,8 @@ static void __exit vmballoon_exit(void)
 	 */
 	vmballoon_send_start(&balloon, 0);
 	vmballoon_pop(&balloon);
+
+	/* Only once we popped the balloon, compaction can be deinit */
+	vmballoon_compaction_deinit(&balloon);
 }
 module_exit(vmballoon_exit);
diff --git a/include/uapi/linux/magic.h b/include/uapi/linux/magic.h
index 1a6fee974116..d4e432295b3f 100644
--- a/include/uapi/linux/magic.h
+++ b/include/uapi/linux/magic.h
@@ -89,5 +89,6 @@
 #define UDF_SUPER_MAGIC		0x15013346
 #define BALLOON_KVM_MAGIC	0x13661366
 #define ZSMALLOC_MAGIC		0x58295829
+#define BALLOON_VMW_MAGIC	0xba11007
 
 #endif /* __LINUX_MAGIC_H__ */
-- 
2.17.1

Powered by blists - more mailing lists