lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <1434134610-2831-7-git-send-email-moltmann@vmware.com>
Date:	Fri, 12 Jun 2015 11:43:27 -0700
From:	"Philip P. Moltmann" <moltmann@...are.com>
To:	dmitry.torokhov@...il.com
Cc:	"Philip P. Moltmann" <moltmann@...are.com>,
	gregkh@...uxfoundation.org, linux-kernel@...r.kernel.org,
	xdeguillard@...are.com, akpm@...ux-foundation.org,
	pv-drivers@...are.com
Subject: [PATCH v3 6/9] VMware balloon: Do not limit the amount of frees and allocations in non-sleep mode.

When VMware's hypervisor requests a VM to reclaim memory this is preferrably done
via ballooning. If the balloon driver does not return memory fast enough, more
drastic methods, such as hypervisor-level swapping are needed. These other methods
cause performance issues, e.g. hypervisor-level swapping requires the hypervisor to
swap in a page syncronously while the virtual CPU is blocked.

Hence it is in the interest of the VM to balloon memory as fast as possible. The
problem with doing this is that the VM might end up doing nothing else than
ballooning and the user might notice that the VM is stalled, esp. when the VM has
only a single virtual CPU.

This is less of a problem if the VM and the hypervisor perform balloon operations
faster. Also the balloon driver yields regularly, hence on a single virtual CPU
the Linux scheduler should be able to properly time-slice between ballooning and
other tasks.

Testing Done: quickly ballooned a lot of pages while wathing if there are any
perceived hickups (periods of non-responsiveness) in the execution of the
linux VM. No such hickups were seen.

Signed-off-by: Xavier Deguillard <xdeguillard@...are.com>
---
 drivers/misc/vmw_balloon.c | 68 +++++++++++-----------------------------------
 1 file changed, 16 insertions(+), 52 deletions(-)

diff --git a/drivers/misc/vmw_balloon.c b/drivers/misc/vmw_balloon.c
index 6eaf7f7..904b6a6 100644
--- a/drivers/misc/vmw_balloon.c
+++ b/drivers/misc/vmw_balloon.c
@@ -46,7 +46,7 @@
 
 MODULE_AUTHOR("VMware, Inc.");
 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
-MODULE_VERSION("1.3.3.0-k");
+MODULE_VERSION("1.3.4.0-k");
 MODULE_ALIAS("dmi:*:svnVMware*:*");
 MODULE_ALIAS("vmware_vmmemctl");
 MODULE_LICENSE("GPL");
@@ -57,12 +57,6 @@ MODULE_LICENSE("GPL");
  */
 
 /*
- * Rate of allocating memory when there is no memory pressure
- * (driver performs non-sleeping allocations).
- */
-#define VMW_BALLOON_NOSLEEP_ALLOC_MAX	16384U
-
-/*
  * Rates of memory allocaton when guest experiences memory pressure
  * (driver performs sleeping allocations).
  */
@@ -71,13 +65,6 @@ MODULE_LICENSE("GPL");
 #define VMW_BALLOON_RATE_ALLOC_INC	16U
 
 /*
- * Rates for releasing pages while deflating balloon.
- */
-#define VMW_BALLOON_RATE_FREE_MIN	512U
-#define VMW_BALLOON_RATE_FREE_MAX	16384U
-#define VMW_BALLOON_RATE_FREE_INC	16U
-
-/*
  * When guest is under memory pressure, use a reduced page allocation
  * rate for next several cycles.
  */
@@ -99,9 +86,6 @@ MODULE_LICENSE("GPL");
  */
 #define VMW_PAGE_ALLOC_CANSLEEP		(GFP_HIGHUSER)
 
-/* Maximum number of page allocations without yielding processor */
-#define VMW_BALLOON_YIELD_THRESHOLD	1024
-
 /* Maximum number of refused pages we accumulate during inflation cycle */
 #define VMW_BALLOON_MAX_REFUSED		16
 
@@ -278,7 +262,6 @@ struct vmballoon {
 
 	/* adjustment rates (pages per second) */
 	unsigned int rate_alloc;
-	unsigned int rate_free;
 
 	/* slowdown page allocations for next few cycles */
 	unsigned int slow_allocation_cycles;
@@ -502,18 +485,13 @@ static bool vmballoon_send_batched_unlock(struct vmballoon *b,
 static void vmballoon_pop(struct vmballoon *b)
 {
 	struct page *page, *next;
-	unsigned int count = 0;
 
 	list_for_each_entry_safe(page, next, &b->pages, lru) {
 		list_del(&page->lru);
 		__free_page(page);
 		STATS_INC(b->stats.free);
 		b->size--;
-
-		if (++count >= b->rate_free) {
-			count = 0;
-			cond_resched();
-		}
+		cond_resched();
 	}
 
 	if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
@@ -715,7 +693,7 @@ static void vmballoon_add_batched_page(struct vmballoon *b, int idx,
  */
 static void vmballoon_inflate(struct vmballoon *b)
 {
-	unsigned int rate;
+	unsigned rate;
 	unsigned int allocations = 0;
 	unsigned int num_pages = 0;
 	int error = 0;
@@ -742,13 +720,13 @@ static void vmballoon_inflate(struct vmballoon *b)
 	 * Start with no sleep allocation rate which may be higher
 	 * than sleeping allocation rate.
 	 */
-	rate = b->slow_allocation_cycles ?
-			b->rate_alloc : VMW_BALLOON_NOSLEEP_ALLOC_MAX;
+	rate = b->slow_allocation_cycles ? b->rate_alloc : UINT_MAX;
 
-	pr_debug("%s - goal: %d, no-sleep rate: %d, sleep rate: %d\n",
+	pr_debug("%s - goal: %d, no-sleep rate: %u, sleep rate: %d\n",
 		 __func__, b->target - b->size, rate, b->rate_alloc);
 
-	while (b->size < b->target && num_pages < b->target - b->size) {
+	while (!b->reset_required &&
+		b->size < b->target && num_pages < b->target - b->size) {
 		struct page *page;
 
 		if (flags == VMW_PAGE_ALLOC_NOSLEEP)
@@ -798,10 +776,7 @@ static void vmballoon_inflate(struct vmballoon *b)
 				break;
 		}
 
-		if (++allocations > VMW_BALLOON_YIELD_THRESHOLD) {
-			cond_resched();
-			allocations = 0;
-		}
+		cond_resched();
 
 		if (allocations >= rate) {
 			/* We allocated enough pages, let's take a break. */
@@ -837,36 +812,29 @@ static void vmballoon_deflate(struct vmballoon *b)
 	unsigned int num_pages = 0;
 	int error;
 
-	pr_debug("%s - size: %d, target %d, rate: %d\n", __func__, b->size,
-						b->target, b->rate_free);
+	pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
 
 	/* free pages to reach target */
 	list_for_each_entry_safe(page, next, &b->pages, lru) {
 		list_del(&page->lru);
 		b->ops->add_page(b, num_pages++, page);
 
+
 		if (num_pages == b->batch_max_pages) {
 			error = b->ops->unlock(b, num_pages, &b->target);
 			num_pages = 0;
-			if (error) {
-				/* quickly decrease rate in case of error */
-				b->rate_free = max(b->rate_free / 2,
-						VMW_BALLOON_RATE_FREE_MIN);
+			if (error)
 				return;
-			}
 		}
 
-		if (++i >= b->size - b->target)
+		if (b->reset_required || ++i >= b->size - b->target)
 			break;
+
+		cond_resched();
 	}
 
 	if (num_pages > 0)
 		b->ops->unlock(b, num_pages, &b->target);
-
-	/* slowly increase rate if there were no errors */
-	if (error == 0)
-		b->rate_free = min(b->rate_free + VMW_BALLOON_RATE_FREE_INC,
-				   VMW_BALLOON_RATE_FREE_MAX);
 }
 
 static const struct vmballoon_ops vmballoon_basic_ops = {
@@ -992,11 +960,8 @@ static int vmballoon_debug_show(struct seq_file *f, void *offset)
 
 	/* format rate info */
 	seq_printf(f,
-		   "rateNoSleepAlloc:   %8d pages/sec\n"
-		   "rateSleepAlloc:     %8d pages/sec\n"
-		   "rateFree:           %8d pages/sec\n",
-		   VMW_BALLOON_NOSLEEP_ALLOC_MAX,
-		   b->rate_alloc, b->rate_free);
+		   "rateSleepAlloc:     %8d pages/sec\n",
+		   b->rate_alloc);
 
 	seq_printf(f,
 		   "\n"
@@ -1087,7 +1052,6 @@ static int __init vmballoon_init(void)
 
 	/* initialize rates */
 	balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX;
-	balloon.rate_free = VMW_BALLOON_RATE_FREE_MAX;
 
 	INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
 
-- 
2.4.3

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ