xen: speed up grant-table reclaim

When a grant entry is still in use by the remote domain, Linux must put
it on a deferred list.  Normally, this list is very short, because
the PV network and block protocols expect the backend to unmap the grant
first.  However, Qubes OS's GUI protocol is subject to the constraints
of the X Window System, and as such winds up with the frontend unmapping
the window first.  As a result, the list can grow very large, resulting
in a massive memory leak and eventual VM freeze.

To partially solve this problem, make the number of entries that the VM
will attempt to free at each iteration tunable.  The default is still
10, but it can be overridden via a module parameter.

This is Cc: stable because (when combined with appropriate userspace
changes) it fixes a severe performance and stability problem for Qubes
OS users.

Cc: stable@vger.kernel.org
Signed-off-by: Demi Marie Obenour <demi@invisiblethingslab.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Link: https://lore.kernel.org/r/20230726165354.1252-1-demi@invisiblethingslab.com
Signed-off-by: Juergen Gross <jgross@suse.com>

2 files changed, 40 insertions, 11 deletions

diff --git a/Documentation/ABI/testing/sysfs-module b/Documentation/ABI/testing/sysfs-module
index 08886367d047..62addab47d0c 100644
--- a/Documentation/ABI/testing/sysfs-module
+++ b/Documentation/ABI/testing/sysfs-module
@@ -60,3 +60,14 @@ Description:	Module taint flags:
 			C   staging driver module
 			E   unsigned module
 			==  =====================
+
+What:		/sys/module/grant_table/parameters/free_per_iteration
+Date:		July 2023
+KernelVersion:	6.5 but backported to all supported stable branches
+Contact:	Xen developer discussion <xen-devel@lists.xenproject.org>
+Description:	Read and write number of grant entries to attempt to free per iteration.
+
+		Note: Future versions of Xen and Linux may provide a better
+		interface for controlling the rate of deferred grant reclaim
+		or may not need it at all.
+Users:		Qubes OS (https://www.qubes-os.org)
diff --git a/drivers/xen/grant-table.c b/drivers/xen/grant-table.c
index e1ec725c2819..f13c3b76ad1e 100644
--- a/drivers/xen/grant-table.c
+++ b/drivers/xen/grant-table.c
@@ -498,14 +498,21 @@ static LIST_HEAD(deferred_list);
 static void gnttab_handle_deferred(struct timer_list *);
 static DEFINE_TIMER(deferred_timer, gnttab_handle_deferred);
 
+static atomic64_t deferred_count;
+static atomic64_t leaked_count;
+static unsigned int free_per_iteration = 10;
+module_param(free_per_iteration, uint, 0600);
+
 static void gnttab_handle_deferred(struct timer_list *unused)
 {
-	unsigned int nr = 10;
+	unsigned int nr = READ_ONCE(free_per_iteration);
+	const bool ignore_limit = nr == 0;
 	struct deferred_entry *first = NULL;
 	unsigned long flags;
+	size_t freed = 0;
 
 	spin_lock_irqsave(&gnttab_list_lock, flags);
-	while (nr--) {
+	while ((ignore_limit || nr--) && !list_empty(&deferred_list)) {
 		struct deferred_entry *entry
 			= list_first_entry(&deferred_list,
 					   struct deferred_entry, list);
@@ -515,10 +522,14 @@ static void gnttab_handle_deferred(struct timer_list *unused)
 		list_del(&entry->list);
 		spin_unlock_irqrestore(&gnttab_list_lock, flags);
 		if (_gnttab_end_foreign_access_ref(entry->ref)) {
+			uint64_t ret = atomic64_dec_return(&deferred_count);
+
 			put_free_entry(entry->ref);
-			pr_debug("freeing g.e. %#x (pfn %#lx)\n",
-				 entry->ref, page_to_pfn(entry->page));
+			pr_debug("freeing g.e. %#x (pfn %#lx), %llu remaining\n",
+				 entry->ref, page_to_pfn(entry->page),
+				 (unsigned long long)ret);
 			put_page(entry->page);
+			freed++;
 			kfree(entry);
 			entry = NULL;
 		} else {
@@ -530,21 +541,22 @@ static void gnttab_handle_deferred(struct timer_list *unused)
 		spin_lock_irqsave(&gnttab_list_lock, flags);
 		if (entry)
 			list_add_tail(&entry->list, &deferred_list);
-		else if (list_empty(&deferred_list))
-			break;
 	}
-	if (!list_empty(&deferred_list) && !timer_pending(&deferred_timer)) {
+	if (list_empty(&deferred_list))
+		WARN_ON(atomic64_read(&deferred_count));
+	else if (!timer_pending(&deferred_timer)) {
 		deferred_timer.expires = jiffies + HZ;
 		add_timer(&deferred_timer);
 	}
 	spin_unlock_irqrestore(&gnttab_list_lock, flags);
+	pr_debug("Freed %zu references", freed);
 }
 
 static void gnttab_add_deferred(grant_ref_t ref, struct page *page)
 {
 	struct deferred_entry *entry;
 	gfp_t gfp = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
-	const char *what = KERN_WARNING "leaking";
+	uint64_t leaked, deferred;
 
 	entry = kmalloc(sizeof(*entry), gfp);
 	if (!page) {
@@ -567,10 +579,16 @@ static void gnttab_add_deferred(grant_ref_t ref, struct page *page)
 			add_timer(&deferred_timer);
 		}
 		spin_unlock_irqrestore(&gnttab_list_lock, flags);
-		what = KERN_DEBUG "deferring";
+		deferred = atomic64_inc_return(&deferred_count);
+		leaked = atomic64_read(&leaked_count);
+		pr_debug("deferring g.e. %#x (pfn %#lx) (total deferred %llu, total leaked %llu)\n",
+			 ref, page ? page_to_pfn(page) : -1, deferred, leaked);
+	} else {
+		deferred = atomic64_read(&deferred_count);
+		leaked = atomic64_inc_return(&leaked_count);
+		pr_warn("leaking g.e. %#x (pfn %#lx) (total deferred %llu, total leaked %llu)\n",
+			ref, page ? page_to_pfn(page) : -1, deferred, leaked);
 	}
-	printk("%s g.e. %#x (pfn %#lx)\n",
-	       what, ref, page ? page_to_pfn(page) : -1);
 }
 
 int gnttab_try_end_foreign_access(grant_ref_t ref)