UserfaultFD: Extension for non cooperative uffd usage

PATCH In review Pavel Emelyanov 2 participants 22 messages Mar 18, 2015 – May 8, 2015 1 message not in the archive

Subsystems (7)

Memory management VFS MM core Scheduler +3 more subsystems

lore

Share & tools

lore

Pavel Emelyanov

author

2015-03-18 19:34 UTC #0 lore raw

cover UserfaultFD: Extension for non cooperative uffd usage

Hi,

On the recent LSF Andrea presented his userfault-fd patches and I had shown some issues that appear in usage scenarios when the monitor task and mm task do not cooperate to each other on VM changes (and fork()-s).

Here's the implementation of the extended uffd API that would help us to address those issues.

As proof of concept I've implemented only fix for fork() case, but I also plan to add the mremap() and exit() notifications, both are also required for such non-cooperative usage.

More details about the extension itself is in patch #2 and the fork() notification description is in patch #3.

Comments and suggestion are warmly welcome :)

Andrea, what's the best way to go on with the patches -- would you prefer to include them in your git tree or should I instead continue with them on my own, re-sending them when required? Either way would be OK for me.

Thanks,
Pavel

Pavel Emelyanov

author

2015-03-18 19:34 UTC #1 lore raw

PATCH 1/3 uffd: Tossing bits around

Reformat the existing code a bit to make it easier for further patching.

Signed-off-by Pavel Emelyanov <[email protected]>

---
 fs/userfaultfd.c | 37 ++++++++++++++++++++++++++++---------
 1 file changed, 28 insertions(+), 9 deletions(-)

fs/userfaultfd.c +28 -9

index b4c7f25..6c9a2d6 100644
@@ -61,6 +61,23 @@ struct userfaultfd_wake_range {
 	unsigned long len;
 };
 
+static const struct file_operations userfaultfd_fops;
+
+static struct userfaultfd_ctx *userfaultfd_ctx_alloc(void)
+{
+	struct userfaultfd_ctx *ctx;
+
+	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
+	if (ctx) {
+		atomic_set(&ctx->refcount, 1);
+		init_waitqueue_head(&ctx->fault_wqh);
+		init_waitqueue_head(&ctx->fd_wqh);
+		ctx->released = false;
+	}
+
+	return ctx;
+}
+
 static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode,
 				     int wake_flags, void *key)
 {

@@ -307,15 +324,15 @@ static int userfaultfd_release(struct inode *inode, struct file *file)
 	return 0;
 }
 
-static inline unsigned int find_userfault(struct userfaultfd_ctx *ctx,
+static inline unsigned int do_find_userfault(wait_queue_head_t *wqh,
 					  struct userfaultfd_wait_queue **uwq)
 {
 	wait_queue_t *wq;
 	struct userfaultfd_wait_queue *_uwq;
 	unsigned int ret = 0;
 
-	spin_lock(&ctx->fault_wqh.lock);
-	list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) {
+	spin_lock(&wqh->lock);
+	list_for_each_entry(wq, &wqh->task_list, task_list) {
 		_uwq = container_of(wq, struct userfaultfd_wait_queue, wq);
 		if (_uwq->pending) {
 			ret = POLLIN;

@@ -324,11 +341,17 @@ static inline unsigned int find_userfault(struct userfaultfd_ctx *ctx,
 			break;
 		}
 	}
-	spin_unlock(&ctx->fault_wqh.lock);
+	spin_unlock(&wqh->lock);
 
 	return ret;
 }
 
+static inline unsigned int find_userfault(struct userfaultfd_ctx *ctx,
+		struct userfaultfd_wait_queue **uwq)
+{
+	return do_find_userfault(&ctx->fault_wqh, uwq);
+}
+
 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 {
 	struct userfaultfd_ctx *ctx = file->private_data;

@@ -1080,16 +1103,12 @@ static struct file *userfaultfd_file_create(int flags)
 		goto out;
 
 	file = ERR_PTR(-ENOMEM);
-	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
+	ctx = userfaultfd_ctx_alloc();
 	if (!ctx)
 		goto out;
 
-	atomic_set(&ctx->refcount, 1);
-	init_waitqueue_head(&ctx->fault_wqh);
-	init_waitqueue_head(&ctx->fd_wqh);
 	ctx->flags = flags;
 	ctx->state = UFFD_STATE_WAIT_API;
-	ctx->released = false;
 	ctx->mm = current->mm;
 	/* prevent the mm struct to be freed */
 	atomic_inc(&ctx->mm->mm_count);

Signature · 1 line

1.8.4.2

Pavel Emelyanov

author

2015-03-18 19:35 UTC #2 lore raw

PATCH 2/3 uffd: Introduce the v2 API

In reply to Pavel Emelyanov

The new API will report more than just the page-faults. The reason for this is -- when the task whose mm we monitor with uffd and the monitor task itself cannot cooperate with each other, the former one can screw things up. Like this.

If task fork()-s the child process is detached from uffd and thus all not-yet-faulted-in memory gets mapped with zero-pages on touch.

Another example is mremap(). When the victim remaps the uffd-ed region and starts touching it the monitor would receive fault messages with addresses that were not register-ed with uffd ioctl before it. Thus monitor will have no idea how to handle those faults.

To address both we can send more events to the monitor. In particular, on fork() we can create another uffd context, register the same set of regions in it and "send" the descriptor to monitor.

For mremap() we can send the message describing what change has been performed.

So this patch prepares to ground for the described above feature by introducing the v2 API of uffd. With new API the kernel would respond with a message containing the event type (pagefault, fork or remap) and argument (fault address, new uffd descriptor or region change respectively).

Signed-off-by Pavel Emelyanov <[email protected]>

---
 fs/userfaultfd.c                 | 56 ++++++++++++++++++++++++++++++----------
 include/uapi/linux/userfaultfd.h | 21 ++++++++++++++-
 2 files changed, 62 insertions(+), 15 deletions(-)

fs/userfaultfd.c +42 ~~-14~~

index 6c9a2d6..bd629b4 100644
@@ -41,6 +41,8 @@ struct userfaultfd_ctx {
 	wait_queue_head_t fd_wqh;
 	/* userfaultfd syscall flags */
 	unsigned int flags;
+	/* features flags */
+	unsigned int features;
 	/* state machine */
 	enum userfaultfd_state state;
 	/* released */

@@ -49,6 +51,8 @@ struct userfaultfd_ctx {
 	struct mm_struct *mm;
 };
 
+#define UFFD_FEATURE_LONGMSG	0x1
+
 struct userfaultfd_wait_queue {
 	unsigned long address;
 	wait_queue_t wq;

@@ -369,7 +373,7 @@ static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 }
 
 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
-				    __u64 *addr)
+				    __u64 *mtype, __u64 *addr)
 {
 	ssize_t ret;
 	DECLARE_WAITQUEUE(wait, current);

@@ -383,6 +387,7 @@ static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 		if (find_userfault(ctx, &uwq)) {
 			uwq->pending = false;
 			/* careful to always initialize addr if ret == 0 */
+			*mtype = UFFD_PAGEFAULT;
 			*addr = uwq->address;
 			ret = 0;
 			break;

@@ -411,8 +416,6 @@ static ssize_t userfaultfd_read(struct file *file, char __user *buf,
 {
 	struct userfaultfd_ctx *ctx = file->private_data;
 	ssize_t _ret, ret = 0;
-	/* careful to always initialize addr if ret == 0 */
-	__u64 uninitialized_var(addr);
 	int no_wait = file->f_flags & O_NONBLOCK;
 
 	if (ctx->state == UFFD_STATE_WAIT_API)

@@ -420,16 +423,34 @@ static ssize_t userfaultfd_read(struct file *file, char __user *buf,
 	BUG_ON(ctx->state != UFFD_STATE_RUNNING);
 
 	for (;;) {
-		if (count < sizeof(addr))
-			return ret ? ret : -EINVAL;
-		_ret = userfaultfd_ctx_read(ctx, no_wait, &addr);
-		if (_ret < 0)
-			return ret ? ret : _ret;
-		if (put_user(addr, (__u64 __user *) buf))
-			return ret ? ret : -EFAULT;
-		ret += sizeof(addr);
-		buf += sizeof(addr);
-		count -= sizeof(addr);
+		if (!(ctx->features & UFFD_FEATURE_LONGMSG)) {
+			/* careful to always initialize addr if ret == 0 */
+			__u64 uninitialized_var(addr);
+			__u64 uninitialized_var(mtype);
+			if (count < sizeof(addr))
+				return ret ? ret : -EINVAL;
+			_ret = userfaultfd_ctx_read(ctx, no_wait, &mtype, &addr);
+			if (_ret < 0)
+				return ret ? ret : _ret;
+			BUG_ON(mtype != UFFD_PAGEFAULT);
+			if (put_user(addr, (__u64 __user *) buf))
+				return ret ? ret : -EFAULT;
+			_ret = sizeof(addr);
+		} else {
+			struct uffd_v2_msg msg;
+			if (count < sizeof(msg))
+				return ret ? ret : -EINVAL;
+			_ret = userfaultfd_ctx_read(ctx, no_wait, &msg.type, &msg.arg);
+			if (_ret < 0)
+				return ret ? ret : _ret;
+			if (copy_to_user(buf, &msg, sizeof(msg)))
+				return ret ? ret : -EINVAL;
+			_ret = sizeof(msg);
+		}
+
+		ret += _ret;
+		buf += _ret;
+		count -= _ret;
 		/*
 		 * Allow to read more than one fault at time but only
 		 * block if waiting for the very first one.

@@ -981,7 +1002,7 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 	ret = -EFAULT;
 	if (copy_from_user(&uffdio_api, buf, sizeof(__u64)))
 		goto out;
-	if (uffdio_api.api != UFFD_API) {
+	if (uffdio_api.api != UFFD_API && uffdio_api.api != UFFD_API_V2) {
 		/* careful not to leak info, we only read the first 8 bytes */
 		memset(&uffdio_api, 0, sizeof(uffdio_api));
 		if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))

@@ -992,6 +1013,12 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 	/* careful not to leak info, we only read the first 8 bytes */
 	uffdio_api.bits = UFFD_API_BITS;
 	uffdio_api.ioctls = UFFD_API_IOCTLS;
+
+	if (uffdio_api.api == UFFD_API_V2) {
+		ctx->features |= UFFD_FEATURE_LONGMSG;
+		uffdio_api.bits |= UFFD_API_V2_BITS;
+	}
+
 	ret = -EFAULT;
 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
 		goto out;

@@ -1109,6 +1136,7 @@ static struct file *userfaultfd_file_create(int flags)
 
 	ctx->flags = flags;
 	ctx->state = UFFD_STATE_WAIT_API;
+	ctx->features = 0;
 	ctx->mm = current->mm;
 	/* prevent the mm struct to be freed */
 	atomic_inc(&ctx->mm->mm_count);

include/uapi/linux/userfaultfd.h +20 -1

index db6e99a..4e169b8 100644
@@ -9,7 +9,9 @@
 #ifndef _LINUX_USERFAULTFD_H
 #define _LINUX_USERFAULTFD_H
 
-#define UFFD_API ((__u64)0xAA)
+#define UFFD_API 	((__u64)0xAA)
+#define UFFD_API_V2	((__u64)0xAB)
+
 /* FIXME: add "|UFFD_BIT_WP" to UFFD_API_BITS after implementing it */
 #define UFFD_API_BITS (UFFD_BIT_WRITE)
 #define UFFD_API_IOCTLS				\

@@ -147,4 +149,21 @@ struct uffdio_remap {
 	__s64 wake;
 };
 
+struct uffd_v2_msg {
+	__u64	type;
+	__u64	arg;
+};
+
+#define UFFD_PAGEFAULT	0x1
+
+#define UFFD_PAGEFAULT_BIT	(1 << (UFFD_PAGEFAULT - 1))
+#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT)
+
+/*
+ * Lower PAGE_SHIFT bits are used to report those supported
+ * by the pagefault message itself. Other bits are used to
+ * report the message types v2 API supports
+ */
+#define UFFD_API_V2_BITS	(__UFFD_API_V2_BITS << 12)
+
 #endif /* _LINUX_USERFAULTFD_H */

Signature · 1 line

1.8.4.2

Andrea Arcangeli

2015-04-21 12:18 UTC #3 lore raw

Re: [PATCH 2/3] uffd: Introduce the v2 API

fs/userfaultfd.c static ssize_t userfaultfd_read(struct file *file, char __user *buf, · line 426

Quoted 1 line · Pavel Emelyanov

fs/userfaultfd.c +1 -0

+		if (!(ctx->features & UFFD_FEATURE_LONGMSG)) {

If we are to use different protocols, it'd be nicer to have two different methods to assign to userfaultfd_fops.read that calls an __always_inline function, so that the above check can be optimized away at build time when the inline is expanded. So the branch is converted to calling a different pointer to function which is zero additional cost.

Quoted 12 lines

fs/userfaultfd.c +12 -0

+			/* careful to always initialize addr if ret == 0 */
+			__u64 uninitialized_var(addr);
+			__u64 uninitialized_var(mtype);
+			if (count < sizeof(addr))
+				return ret ? ret : -EINVAL;
+			_ret = userfaultfd_ctx_read(ctx, no_wait, &mtype, &addr);
+			if (_ret < 0)
+				return ret ? ret : _ret;
+			BUG_ON(mtype != UFFD_PAGEFAULT);
+			if (put_user(addr, (__u64 __user *) buf))
+				return ret ? ret : -EFAULT;
+			_ret = sizeof(addr);

fs/userfaultfd.c static ssize_t userfaultfd_read(struct file *file, char __user *buf, · line 448

Quoted 10 lines

fs/userfaultfd.c +10 -0

+		} else {
+			struct uffd_v2_msg msg;
+			if (count < sizeof(msg))
+				return ret ? ret : -EINVAL;
+			_ret = userfaultfd_ctx_read(ctx, no_wait, &msg.type, &msg.arg);
+			if (_ret < 0)
+				return ret ? ret : _ret;
+			if (copy_to_user(buf, &msg, sizeof(msg)))
+				return ret ? ret : -EINVAL;
+			_ret = sizeof(msg);

Reading 16bytes instead of 8bytes for each fault, probably wouldn't move the needle much in terms of userfaultfd_read performance. Perhaps we could consider using the uffd_v2_msg unconditionally and then have a single protocol differentiated by the feature bits.

The only reason to have two different protocols would be to be able to read 8 bytes per userfault, in the cooperative usage (i.e. qemu postcopy). But if we do that we want to use the __always_inline trick to avoid branches and additional runtime costs (otherwise we may as well forget all microoptimizations and read 16bytes always).

fs/userfaultfd.c static int userfaultfd_api(struct userfaultfd_ctx *ctx, · line 1024

Quoted 13 lines

fs/userfaultfd.c +6 -0

@@ -992,6 +1013,12 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 	/* careful not to leak info, we only read the first 8 bytes */
 	uffdio_api.bits = UFFD_API_BITS;
 	uffdio_api.ioctls = UFFD_API_IOCTLS;
+
+	if (uffdio_api.api == UFFD_API_V2) {
+		ctx->features |= UFFD_FEATURE_LONGMSG;
+		uffdio_api.bits |= UFFD_API_V2_BITS;
+	}
+
 	ret = -EFAULT;
 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
 		goto out;

The original meaning of the bits is:

If UFFD_BIT_WRITE was set in api.bits, it means the !!(address&UFFD_BIT_WRITE) tells if it was a write fault (missing or WP).

If UFFD_BIT_WP was set in api.bits, it means the !!(address&UFFD_BIT_WP) tells if it was a WP fault (if not set it means it was a missing fault).

Currently api.bits sets only UFFD_BIT_WRITE, and UFFD_BIT_WP will be set later, after the WP tracking mode will be implemented.

I'm uncertain how bits translated to features and if they should be unified or only have features.

Quoted 7 lines

include/uapi/linux/userfaultfd.h +7 -0

+struct uffd_v2_msg {
+	__u64	type;
+	__u64	arg;
+};
+
+#define UFFD_PAGEFAULT	0x1
+

include/uapi/linux/userfaultfd.h struct uffdio_remap { · line 168

Quoted 10 lines

include/uapi/linux/userfaultfd.h +10 -0

+#define UFFD_PAGEFAULT_BIT	(1 << (UFFD_PAGEFAULT - 1))
+#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT)
+
+/*
+ * Lower PAGE_SHIFT bits are used to report those supported
+ * by the pagefault message itself. Other bits are used to
+ * report the message types v2 API supports
+ */
+#define UFFD_API_V2_BITS	(__UFFD_API_V2_BITS << 12)
+

And why exactly is this 12 hardcoded? And which field should be masked with the bits? In the V1 protocol it was the "arg" (userfault address) not the "type". So this is a bit confusing and probably requires simplification.

Thanks,
Andrea

Pavel Emelyanov

author

2015-04-23 06:29 UTC #4 lore raw

Re: [PATCH 2/3] uffd: Introduce the v2 API

Quoted 9 lines · Andrea Arcangeli

fs/userfaultfd.c static ssize_t userfaultfd_read(struct file *file, char __user *buf, · line 426
Quoted 1 line · Pavel Emelyanov
fs/userfaultfd.c +1 -0
+		if (!(ctx->features & UFFD_FEATURE_LONGMSG)) {
If we are to use different protocols, it'd be nicer to have two different methods to assign to userfaultfd_fops.read that calls an __always_inline function, so that the above check can be optimized away at build time when the inline is expanded. So the branch is converted to calling a different pointer to function which is zero additional cost.

OK :)

Quoted 19 lines

fs/userfaultfd.c +19 -0

+			/* careful to always initialize addr if ret == 0 */
+			__u64 uninitialized_var(addr);
+			__u64 uninitialized_var(mtype);
+			if (count < sizeof(addr))
+				return ret ? ret : -EINVAL;
+			_ret = userfaultfd_ctx_read(ctx, no_wait, &mtype, &addr);
+			if (_ret < 0)
+				return ret ? ret : _ret;
+			BUG_ON(mtype != UFFD_PAGEFAULT);
+			if (put_user(addr, (__u64 __user *) buf))
+				return ret ? ret : -EFAULT;
+			_ret = sizeof(addr);
+		} else {
+			struct uffd_v2_msg msg;
+			if (count < sizeof(msg))
+				return ret ? ret : -EINVAL;
+			_ret = userfaultfd_ctx_read(ctx, no_wait, &msg.type, &msg.arg);
+			if (_ret < 0)
+				return ret ? ret : _ret;

Quoted 8 lines

fs/userfaultfd.c static ssize_t userfaultfd_read(struct file *file, char __user *buf, · line 448
Quoted 3 lines
fs/userfaultfd.c +3 -0
+			if (copy_to_user(buf, &msg, sizeof(msg)))
+				return ret ? ret : -EINVAL;
+			_ret = sizeof(msg);
Reading 16bytes instead of 8bytes for each fault, probably wouldn't move the needle much in terms of userfaultfd_read performance. Perhaps we could consider using the uffd_v2_msg unconditionally and then have a single protocol differentiated by the feature bits.

So your proposal is to always report 16 bytes per PF from read() and let userspace decide itself how to handle the result?

Quoted 47 lines

The only reason to have two different protocols would be to be able to read 8 bytes per userfault, in the cooperative usage (i.e. qemu postcopy). But if we do that we want to use the __always_inline trick to avoid branches and additional runtime costs (otherwise we may as well forget all microoptimizations and read 16bytes always).
fs/userfaultfd.c static int userfaultfd_api(struct userfaultfd_ctx *ctx, · line 1024
Quoted 13 lines
fs/userfaultfd.c +6 -0
@@ -992,6 +1013,12 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 	/* careful not to leak info, we only read the first 8 bytes */
 	uffdio_api.bits = UFFD_API_BITS;
 	uffdio_api.ioctls = UFFD_API_IOCTLS;
+
+	if (uffdio_api.api == UFFD_API_V2) {
+		ctx->features |= UFFD_FEATURE_LONGMSG;
+		uffdio_api.bits |= UFFD_API_V2_BITS;
+	}
+
 	ret = -EFAULT;
 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))
 		goto out;
The original meaning of the bits is:
If UFFD_BIT_WRITE was set in api.bits, it means the !!(address&UFFD_BIT_WRITE) tells if it was a write fault (missing or WP).
If UFFD_BIT_WP was set in api.bits, it means the !!(address&UFFD_BIT_WP) tells if it was a WP fault (if not set it means it was a missing fault).
Currently api.bits sets only UFFD_BIT_WRITE, and UFFD_BIT_WP will be set later, after the WP tracking mode will be implemented.
I'm uncertain how bits translated to features and if they should be unified or only have features.
Quoted 11 lines
include/uapi/linux/userfaultfd.h +11 -0
+struct uffd_v2_msg {
+	__u64	type;
+	__u64	arg;
+};
+
+#define UFFD_PAGEFAULT	0x1
+
+#define UFFD_PAGEFAULT_BIT	(1 << (UFFD_PAGEFAULT - 1))
+#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT)
+
+/*

Quoted 8 lines

include/uapi/linux/userfaultfd.h struct uffdio_remap { · line 168
Quoted 6 lines
+ * Lower PAGE_SHIFT bits are used to report those supported
+ * by the pagefault message itself. Other bits are used to
+ * report the message types v2 API supports
+ */
include/uapi/linux/userfaultfd.h +2 -0
+#define UFFD_API_V2_BITS	(__UFFD_API_V2_BITS << 12)
+
And why exactly is this 12 hardcoded?

Ah, it should have been the PAGE_SHIFT one, but I was unsure whether it would be OK to have different shifts in different arches.

But taking into account your comment that bits field id bad for these values, if we introduce the new .features one for api message, then this 12 will just go away.

Quoted 4 lines

And which field should be masked
with the bits? In the V1 protocol it was the "arg" (userfault address) not the "type". So this is a bit confusing and probably requires simplification.

I see. Actually I decided that since bits higher than 12th (for x86) is always 0 in api message (no bits allowed there, since pfn sits in this place), it would be OK to put non-PF bits there.

Should I better introduce another .features field in uffd API message?

-- Pavel

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Andrea Arcangeli

2015-04-27 21:12 UTC #5 lore raw

Re: [PATCH 2/3] uffd: Introduce the v2 API

Hello,

Quoted 2 lines · Pavel Emelyanov

So your proposal is to always report 16 bytes per PF from read() and let userspace decide itself how to handle the result?

Reading 16bytes for each userfault (instead of 8) and sharing the same read(2) protocol (UFFD_API) for both the cooperative and non-cooperative usages, is something I just suggested for consideration after reading your patchset.

The pros of using a single protocol for both is that it would reduce amount of code and there would be just one file operation for the .read method. The cons is that it will waste 8 bytes per userfault in terms of memory footprint. The other major cons is that it would force us to define the format of the non cooperative protocol now despite it's not fully finished yet.

I'm also ok with two protocols if nobody else objects, but if we use two protocols, we should at least use different file operation methods and use __always_inline with constants passed as parameter to optimize away the branches at build time. This way we get the reduced memory footprint in the read syscall without other runtime overhead associated with it.

Quoted 20 lines

Quoted 11 lines
Quoted 11 lines
include/uapi/linux/userfaultfd.h +11 -0
+struct uffd_v2_msg {
+	__u64	type;
+	__u64	arg;
+};
+
+#define UFFD_PAGEFAULT	0x1
+
+#define UFFD_PAGEFAULT_BIT	(1 << (UFFD_PAGEFAULT - 1))
+#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT)
+
+/*
Quoted 8 lines
include/uapi/linux/userfaultfd.h struct uffdio_remap { · line 168
Quoted 6 lines
+ * Lower PAGE_SHIFT bits are used to report those supported
+ * by the pagefault message itself. Other bits are used to
+ * report the message types v2 API supports
+ */
include/uapi/linux/userfaultfd.h +2 -0
+#define UFFD_API_V2_BITS	(__UFFD_API_V2_BITS << 12)
+
And why exactly is this 12 hardcoded?

Quoted 6 lines

Ah, it should have been the PAGE_SHIFT one, but I was unsure whether it would be OK to have different shifts in different arches.
But taking into account your comment that bits field id bad for these values, if we introduce the new .features one for api message, then this 12 will just go away.

Ok.

Quoted 8 lines

Quoted 4 lines
And which field should be masked
with the bits? In the V1 protocol it was the "arg" (userfault address) not the "type". So this is a bit confusing and probably requires simplification.
I see. Actually I decided that since bits higher than 12th (for x86) is always 0 in api message (no bits allowed there, since pfn sits in this place), it would be OK to put non-PF bits there.

That was ok yes.

Quoted 1 line

Should I better introduce another .features field in uffd API message?

What about renaming "uffdio_api.bits" to "uffdio_api.features"?

And then we set uffdio_api.features to UFFD_FEATURE_WRITE|UFFD_FEATURE_WP|UFFD_FEATURE_FORK as needed.

UFFD_FEATURE_WRITE would always be enabled, it's there only in case we want to disable it later (mostly if some arch has trouble with it, which is unlikely, but qemu doesn't need that bit of information at all for example so qemu would be fine if UFFD_FEATURE_WRITE disappears).

UFFD_FEATURE_WP would signal also that the wrprotection feature (not implemented yet) is available (then later the register ioctl would also show the new wrprotection ioctl numbers available to mangle the wrprotection). The UFFD_FEATURE_WP feature in the cooperative usage (qemu live snapshotting) can use the UFFD_API first protocol too.

UFFD_FEATURE_FORK would be returned if the UFFD_API_V2 was set in uffdio.api, and it would be part of the incremental non-cooperative patchset.

We could also not define "UFFD_FEATURE_FORK" at all and imply that fork/mremap/MADV_DONTNEED are all available if UFFD_API_V2 uffdio_api ioctl succeeds... That's only doable if we keep two different read protocols though. UFFD_FEATURE_FORK (or UFFD_FEATURE_NON_COOPERATIVE) are really strictly needed only if we share the same read(2) protocol for both the cooperative and non-cooperative usages.

The idea is that there's not huge benefit of only having the "fork" feature supported but missing "mremap" and "madv_dontneed".

In fact if a new syscall that works like mremap is added later (call it mremap2), we would need to fail the UFFDIO_API_V2 and require a UFFDIO_API_V3 for such kernel that can return a new mremap2 type of event. Userland couldn't just assume it is ok to use postcopy live migration for containers, because UFFD_FEATURE_FORK|MREMAP|MADV_DONTNEED are present in the uffdio.features when it asked for API_V2. There shall be something that tells userland "hey there's a new mremap2 that the software inside the container can run on top of this kernel, so you are going to get a new mremap2 type of userfault event too".

In any case, regardless of how we solve the above, "uffdio_api.features" sounds better than ".bits".

If we retain two different UFFD_API, we'll be able to freeze the current one and decide later if UFFD_FEATURE_FORK|UFFD_FEATURE_MREMAP|UFFD_FEATURE_MADV_DONTNEED shall be added to the .features, or if to rely on UFFD_API_V2 succeeding to let userland know that the non-cooperative usage is fully supported by the kernel.

Not having to freeze these details now is the main benefit of having two different UFFD_API after all...

Pavel Emelyanov

author

2015-04-30 09:50 UTC #6 lore raw

Re: [PATCH 2/3] uffd: Introduce the v2 API

Quoted 18 lines · Andrea Arcangeli

Hello,
Quoted 2 lines · Pavel Emelyanov
So your proposal is to always report 16 bytes per PF from read() and let userspace decide itself how to handle the result?
Reading 16bytes for each userfault (instead of 8) and sharing the same read(2) protocol (UFFD_API) for both the cooperative and non-cooperative usages, is something I just suggested for consideration after reading your patchset.
The pros of using a single protocol for both is that it would reduce amount of code and there would be just one file operation for the .read method. The cons is that it will waste 8 bytes per userfault in terms of memory footprint. The other major cons is that it would force us to define the format of the non cooperative protocol now despite it's not fully finished yet.

Quoted 6 lines

I'm also ok with two protocols if nobody else objects, but if we use two protocols, we should at least use different file operation methods and use __always_inline with constants passed as parameter to optimize away the branches at build time. This way we get the reduced memory footprint in the read syscall without other runtime overhead associated with it.

OK. I would go with two protocols then and will reshuffle the code to use two ops.

Quoted 38 lines

Quoted 20 lines
Quoted 11 lines
Quoted 11 lines
include/uapi/linux/userfaultfd.h +11 -0
+struct uffd_v2_msg {
+	__u64	type;
+	__u64	arg;
+};
+
+#define UFFD_PAGEFAULT	0x1
+
+#define UFFD_PAGEFAULT_BIT	(1 << (UFFD_PAGEFAULT - 1))
+#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT)
+
+/*
Quoted 8 lines
include/uapi/linux/userfaultfd.h struct uffdio_remap { · line 168
Quoted 6 lines
+ * Lower PAGE_SHIFT bits are used to report those supported
+ * by the pagefault message itself. Other bits are used to
+ * report the message types v2 API supports
+ */
include/uapi/linux/userfaultfd.h +2 -0
+#define UFFD_API_V2_BITS	(__UFFD_API_V2_BITS << 12)
+
And why exactly is this 12 hardcoded?
Quoted 6 lines
Ah, it should have been the PAGE_SHIFT one, but I was unsure whether it would be OK to have different shifts in different arches.
But taking into account your comment that bits field id bad for these values, if we introduce the new .features one for api message, then this 12 will just go away.
Ok.
Quoted 8 lines
Quoted 4 lines
And which field should be masked
with the bits? In the V1 protocol it was the "arg" (userfault address) not the "type". So this is a bit confusing and probably requires simplification.
I see. Actually I decided that since bits higher than 12th (for x86) is always 0 in api message (no bits allowed there, since pfn sits in this place), it would be OK to put non-PF bits there.

Quoted 5 lines

That was ok yes.
Quoted 1 line
Should I better introduce another .features field in uffd API message?
What about renaming "uffdio_api.bits" to "uffdio_api.features"?

Yup, agreed, will do.

Quoted 29 lines

And then we set uffdio_api.features to UFFD_FEATURE_WRITE|UFFD_FEATURE_WP|UFFD_FEATURE_FORK as needed.
UFFD_FEATURE_WRITE would always be enabled, it's there only in case we want to disable it later (mostly if some arch has trouble with it, which is unlikely, but qemu doesn't need that bit of information at all for example so qemu would be fine if UFFD_FEATURE_WRITE disappears).
UFFD_FEATURE_WP would signal also that the wrprotection feature (not implemented yet) is available (then later the register ioctl would also show the new wrprotection ioctl numbers available to mangle the wrprotection). The UFFD_FEATURE_WP feature in the cooperative usage (qemu live snapshotting) can use the UFFD_API first protocol too.
UFFD_FEATURE_FORK would be returned if the UFFD_API_V2 was set in uffdio.api, and it would be part of the incremental non-cooperative patchset.
We could also not define "UFFD_FEATURE_FORK" at all and imply that fork/mremap/MADV_DONTNEED are all available if UFFD_API_V2 uffdio_api ioctl succeeds... That's only doable if we keep two different read protocols though. UFFD_FEATURE_FORK (or UFFD_FEATURE_NON_COOPERATIVE) are really strictly needed only if we share the same read(2) protocol for both the cooperative and non-cooperative usages.
The idea is that there's not huge benefit of only having the "fork" feature supported but missing "mremap" and "madv_dontneed".

Quoted 10 lines

In fact if a new syscall that works like mremap is added later (call it mremap2), we would need to fail the UFFDIO_API_V2 and require a UFFDIO_API_V3 for such kernel that can return a new mremap2 type of event. Userland couldn't just assume it is ok to use postcopy live migration for containers, because UFFD_FEATURE_FORK|MREMAP|MADV_DONTNEED are present in the uffdio.features when it asked for API_V2. There shall be something that tells userland "hey there's a new mremap2 that the software inside the container can run on top of this kernel, so you are going to get a new mremap2 type of userfault event too".

But that's why I assumed to use per-sycall bits -- UFFD_FEATURE_FORK, _MREMAP, _MWHATEVER so that userspace can read those bits and make sure it contains only bits it understands with other bits set to zero.

If we had only one UFFD_API_NON_COOPERATIVE userspace would have no idea what kind of messages it may receive.

Quoted 10 lines

In any case, regardless of how we solve the above, "uffdio_api.features" sounds better than ".bits".
If we retain two different UFFD_API, we'll be able to freeze the current one and decide later if UFFD_FEATURE_FORK|UFFD_FEATURE_MREMAP|UFFD_FEATURE_MADV_DONTNEED shall be added to the .features, or if to rely on UFFD_API_V2 succeeding to let userland know that the non-cooperative usage is fully supported by the kernel.

Quoted 4 lines

Not having to freeze these details now is the main benefit of having two different UFFD_API after all...
.

-- Pavel

Pavel Emelyanov

author

2015-03-18 19:35 UTC #7 lore raw

PATCH 3/3 uffd: Introduce fork() notification

In reply to Pavel Emelyanov

As described in previous e-mail, we need to get informed when the task, whose mm is monitored with uffd, calls fork().

The fork notification is the new uffd with the same regions and flags as those on parent. When read()-ing from uffd the monitor task would receive the new uffd's descriptor number and will be able to start reading events from the new task.

The fork() of mm with uffd attached doesn't finish until the monitor "acks" the message by reading the new uffd descriptor.

Signed-off-by Pavel Emelyanov <[email protected]>

---
 fs/userfaultfd.c                 | 175 ++++++++++++++++++++++++++++++++++++++-
 include/linux/userfaultfd_k.h    |  12 +++
 include/uapi/linux/userfaultfd.h |   4 +-
 kernel/fork.c                    |   9 +-
 4 files changed, 194 insertions(+), 6 deletions(-)

fs/userfaultfd.c +172 -3

index bd629b4..265f031 100644
@@ -12,6 +12,7 @@
  *  mm/ksm.c (mm hashing).
  */
 
+#include <linux/list.h>
 #include <linux/hashtable.h>
 #include <linux/sched.h>
 #include <linux/mm.h>

@@ -37,6 +38,8 @@ struct userfaultfd_ctx {
 	atomic_t refcount;
 	/* waitqueue head for the userfaultfd page faults */
 	wait_queue_head_t fault_wqh;
+	/* waitqueue head for fork-s */
+	wait_queue_head_t fork_wqh;
 	/* waitqueue head for the pseudo fd to wakeup poll/read */
 	wait_queue_head_t fd_wqh;
 	/* userfaultfd syscall flags */

@@ -51,10 +54,21 @@ struct userfaultfd_ctx {
 	struct mm_struct *mm;
 };
 
+struct userfaultfd_fork_ctx {
+	struct userfaultfd_ctx *orig;
+	struct userfaultfd_ctx *new;
+	struct list_head list;
+};
+
 #define UFFD_FEATURE_LONGMSG	0x1
+#define UFFD_FEATURE_FORK	0x2
 
 struct userfaultfd_wait_queue {
-	unsigned long address;
+	union {
+		unsigned long address;
+		struct userfaultfd_ctx *nctx;
+		int fd;
+	};
 	wait_queue_t wq;
 	bool pending;
 	struct userfaultfd_ctx *ctx;

@@ -75,6 +89,7 @@ static struct userfaultfd_ctx *userfaultfd_ctx_alloc(void)
 	if (ctx) {
 		atomic_set(&ctx->refcount, 1);
 		init_waitqueue_head(&ctx->fault_wqh);
+		init_waitqueue_head(&ctx->fork_wqh);
 		init_waitqueue_head(&ctx->fd_wqh);
 		ctx->released = false;
 	}

@@ -270,6 +285,111 @@ int handle_userfault(struct vm_area_struct *vma, unsigned long address,
 	return VM_FAULT_RETRY;
 }
 
+int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
+{
+	struct userfaultfd_ctx *ctx = NULL, *octx;
+	struct userfaultfd_fork_ctx *fctx;
+
+	octx = vma->vm_userfaultfd_ctx.ctx;
+	if (!octx || !(octx->features & UFFD_FEATURE_FORK)) {
+		vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
+		vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING);
+		return 0;
+	}
+
+	list_for_each_entry(fctx, fcs, list)
+		if (fctx->orig == octx) {
+			ctx = fctx->new;
+			break;
+		}
+
+	if (!ctx) {
+		fctx = kmalloc(sizeof(*fctx), GFP_KERNEL);
+		if (!fctx)
+			return -ENOMEM;
+
+		ctx = userfaultfd_ctx_alloc();
+		if (!ctx) {
+			kfree(fctx);
+			return -ENOMEM;
+		}
+
+		ctx->flags = octx->flags;
+		ctx->state = UFFD_STATE_RUNNING;
+		ctx->features = UFFD_FEATURE_FORK | UFFD_FEATURE_LONGMSG;
+		ctx->mm = vma->vm_mm;
+		atomic_inc(&ctx->mm->mm_count);
+
+		userfaultfd_ctx_get(octx);
+		fctx->orig = octx;
+		fctx->new = ctx;
+		list_add_tail(&fctx->list, fcs);
+	}
+
+	vma->vm_userfaultfd_ctx.ctx = ctx;
+	return 0;
+}
+
+static int dup_fctx(struct userfaultfd_fork_ctx *fctx)
+{
+	int ret = 0;
+	struct userfaultfd_ctx *ctx = fctx->orig;
+	struct userfaultfd_wait_queue uwq;
+
+	init_waitqueue_entry(&uwq.wq, current);
+	uwq.pending = true;
+	uwq.ctx = ctx;
+	uwq.nctx = fctx->new;
+
+	spin_lock(&ctx->fork_wqh.lock);
+	/*
+	 * After the __add_wait_queue the uwq is visible to userland
+	 * through poll/read().
+	 */
+	__add_wait_queue(&ctx->fork_wqh, &uwq.wq);
+	for (;;) {
+		set_current_state(TASK_KILLABLE);
+		if (!uwq.pending)
+			break;
+		if (ACCESS_ONCE(ctx->released) ||
+				fatal_signal_pending(current)) {
+			ret = -1;
+			break;
+		}
+
+		spin_unlock(&ctx->fork_wqh.lock);
+
+		wake_up_poll(&ctx->fd_wqh, POLLIN);
+		schedule();
+
+		spin_lock(&ctx->fork_wqh.lock);
+	}
+	__remove_wait_queue(&ctx->fork_wqh, &uwq.wq);
+	__set_current_state(TASK_RUNNING);
+	spin_unlock(&ctx->fork_wqh.lock);
+
+	/*
+	 * ctx may go away after this if the userfault pseudo fd is
+	 * already released.
+	 */
+	userfaultfd_ctx_put(ctx);
+
+	return ret;
+}
+
+void dup_userfaultfd_complete(struct list_head *fcs)
+{
+	int ret = 0;
+	struct userfaultfd_fork_ctx *fctx, *n;
+
+	list_for_each_entry_safe(fctx, n, fcs, list) {
+		if (!ret)
+			ret = dup_fctx(fctx);
+		list_del(&fctx->list);
+		kfree(fctx);
+	}
+}
+
 static int userfaultfd_release(struct inode *inode, struct file *file)
 {
 	struct userfaultfd_ctx *ctx = file->private_data;

@@ -356,6 +476,12 @@ static inline unsigned int find_userfault(struct userfaultfd_ctx *ctx,
 	return do_find_userfault(&ctx->fault_wqh, uwq);
 }
 
+static inline unsigned int find_userfault_fork(struct userfaultfd_ctx *ctx,
+		struct userfaultfd_wait_queue **uwq)
+{
+	return do_find_userfault(&ctx->fork_wqh, uwq);
+}
+
 static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 {
 	struct userfaultfd_ctx *ctx = file->private_data;

@@ -366,12 +492,40 @@ static unsigned int userfaultfd_poll(struct file *file, poll_table *wait)
 	case UFFD_STATE_WAIT_API:
 		return POLLERR;
 	case UFFD_STATE_RUNNING:
-		return find_userfault(ctx, NULL);
+		return find_userfault(ctx, NULL) || find_userfault_fork(ctx, NULL);
 	default:
 		BUG();
 	}
 }
 
+static ssize_t resolve_userfault_fork(struct userfaultfd_ctx *ctx,
+		struct userfaultfd_wait_queue **uwq)
+{
+	struct userfaultfd_ctx *new;
+	int fd;
+	struct file *file;
+
+	if (!find_userfault_fork(ctx, uwq))
+		return 0;
+
+	new = (*uwq)->nctx;
+	fd = get_unused_fd_flags(new->flags & UFFD_SHARED_FCNTL_FLAGS);
+	if (fd < 0)
+		return fd;
+
+	file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new,
+				  O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS));
+	if (IS_ERR(file)) {
+		put_unused_fd(fd);
+		return PTR_ERR(file);
+	}
+
+	fd_install(fd, file);
+	(*uwq)->fd = fd;
+
+	return 1;
+}
+
 static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 				    __u64 *mtype, __u64 *addr)
 {

@@ -392,6 +546,21 @@ static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait,
 			ret = 0;
 			break;
 		}
+
+		ret = resolve_userfault_fork(ctx, &uwq);
+		if (ret < 0)
+			break;
+		if (ret > 0) {
+			*mtype = UFFD_FORK;
+			*addr = uwq->fd;
+
+			uwq->pending = false;
+			wake_up(&ctx->fork_wqh);
+
+			ret = 0;
+			break;
+		}
+
 		if (signal_pending(current)) {
 			ret = -ERESTARTSYS;
 			break;

@@ -1015,7 +1184,7 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 	uffdio_api.ioctls = UFFD_API_IOCTLS;
 
 	if (uffdio_api.api == UFFD_API_V2) {
-		ctx->features |= UFFD_FEATURE_LONGMSG;
+		ctx->features |= UFFD_FEATURE_FORK | UFFD_FEATURE_LONGMSG;
 		uffdio_api.bits |= UFFD_API_V2_BITS;
 	}

include/linux/userfaultfd_k.h +12 -0

index 81f0d11..44827f7 100644
@@ -75,6 +75,9 @@ static inline bool userfaultfd_armed(struct vm_area_struct *vma)
 	return vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP);
 }
 
+extern int dup_userfaultfd(struct vm_area_struct *, struct list_head *);
+extern void dup_userfaultfd_complete(struct list_head *);
+
 #else /* CONFIG_USERFAULTFD */
 
 /* mm helpers */

@@ -107,6 +110,15 @@ static inline bool userfaultfd_armed(struct vm_area_struct *vma)
 	return false;
 }
 
+static inline int dup_userfaultfd(struct vm_area_struct *, struct list_head *)
+{
+	return 0;
+}
+
+static inline void dup_userfaultfd_complete(struct list_head *)
+{
+}
+
 #endif /* CONFIG_USERFAULTFD */
 
 #endif /* _LINUX_USERFAULTFD_K_H */

include/uapi/linux/userfaultfd.h +3 -1

index 4e169b8..f6cfea3 100644
@@ -155,9 +155,11 @@ struct uffd_v2_msg {
 };
 
 #define UFFD_PAGEFAULT	0x1
+#define UFFD_FORK	0x2
 
 #define UFFD_PAGEFAULT_BIT	(1 << (UFFD_PAGEFAULT - 1))
-#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT)
+#define UFFD_FORK_BIT		(1 << (UFFD_FORK - 1))
+#define __UFFD_API_V2_BITS	(UFFD_PAGEFAULT_BIT | UFFD_FORK_BIT)
 
 /*
  * Lower PAGE_SHIFT bits are used to report those supported

kernel/fork.c +7 -2

index cfab6e9..532882d 100644
@@ -55,6 +55,7 @@
 #include <linux/rmap.h>
 #include <linux/ksm.h>
 #include <linux/acct.h>
+#include <linux/userfaultfd_k.h>
 #include <linux/tsacct_kern.h>
 #include <linux/cn_proc.h>
 #include <linux/freezer.h>

@@ -370,6 +371,7 @@ static int dup_mmap(struct mm_struct *mm, struct mm_struct *oldmm)
 	struct rb_node **rb_link, *rb_parent;
 	int retval;
 	unsigned long charge;
+	LIST_HEAD(uf);
 
 	uprobe_start_dup_mmap();
 	down_write(&oldmm->mmap_sem);

@@ -421,11 +423,13 @@ static int dup_mmap(struct mm_struct *mm, struct mm_struct *oldmm)
 		if (retval)
 			goto fail_nomem_policy;
 		tmp->vm_mm = mm;
+		retval = dup_userfaultfd(tmp, &uf);
+		if (retval)
+			goto fail_nomem_anon_vma_fork;
 		if (anon_vma_fork(tmp, mpnt))
 			goto fail_nomem_anon_vma_fork;
-		tmp->vm_flags &= ~(VM_LOCKED|VM_UFFD_MISSING|VM_UFFD_WP);
+		tmp->vm_flags &= ~(VM_LOCKED);
 		tmp->vm_next = tmp->vm_prev = NULL;
-		tmp->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
 		file = tmp->vm_file;
 		if (file) {
 			struct inode *inode = file_inode(file);

@@ -481,6 +485,7 @@ out:
 	up_write(&mm->mmap_sem);
 	flush_tlb_mm(oldmm);
 	up_write(&oldmm->mmap_sem);
+	dup_userfaultfd_complete(&uf);
 	uprobe_end_dup_mmap();
 	return retval;
 fail_nomem_anon_vma_fork:

Signature · 1 line

1.8.4.2

Andrea Arcangeli

2015-04-21 12:02 UTC #8 lore raw

In reply to Pavel Emelyanov

Hi Pavel,

Quoted 10 lines · Pavel Emelyanov

Hi,
On the recent LSF Andrea presented his userfault-fd patches and I had shown some issues that appear in usage scenarios when the monitor task and mm task do not cooperate to each other on VM changes (and fork()-s).
Here's the implementation of the extended uffd API that would help us to address those issues.

Quoted 8 lines

As proof of concept I've implemented only fix for fork() case, but I also plan to add the mremap() and exit() notifications, both are also required for such non-cooperative usage.
More details about the extension itself is in patch #2 and the fork() notification description is in patch #3.
Comments and suggestion are warmly welcome :)

This looks feasible.

Quoted 4 lines

Andrea, what's the best way to go on with the patches -- would you prefer to include them in your git tree or should I instead continue with them on my own, re-sending them when required? Either way would be OK for me.

Ok so various improvements happened in userfaultfd patchset over the last month so your incremental patchset likely requires a rebase sorry. When you posted it I was in the middle of the updates. Now things are working stable and I have no pending updates, so it would be a good time for a rebase.

I can merge it if you like, it's your call if you prefer to maintain it incrementally or if I should merge it, but I wouldn't push it to Andrew for upstream integration in the first batch, as this complicates things further and it's not fully complete yet (at least the version posted only handled fork). I think it can be merged incrementally in a second stage.

The major updates of the userfaultfd patchset over the last month were:

1) Various mixed fixes thanks to the feedback from Dave Hansen and David Gilbert.

The most notable one is the use of mm_users instead of mm_count to pin the mm to avoid crashes that assumed the vma still existed (in the userfaultfd_release method and in the various ioctl). exit_mmap doesn't even set mm->mmap to NULL, so unless I introduce a userfaultfd_exit to call in mmput, I have to pin the mm_users to be safe. This is mainly an issue for the non-cooperative usage you're implementing. Can you catch the exit somehow so you can close the fd? The memory won't be released until you do. Is this ok with you? I suppose you had to close the fd somehow anyway.

2) userfaults are waken immediately even if they're not been "read" yet, this can lead to POLLIN false positives (so I only allow poll if the fd is open in nonblocking mode to be sure it won't hang). Is it too paranoid to return POLLERR if the fd is not open in nonblocking mode?

	http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=f222d9de0a5302dc8ac62d6fab53a84251098751

3) optimize read to return entries in O(1) and poll which was already O(1) becomes lockless. This required to split the waitqueue in two, one for pending faults and one for non pending faults, and the faults are refiled across the two waitqueues when they're read. Both waitqueues are protected by a single lock to be simpler and faster at runtime (the fault_pending_wqh one).

	http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=9aa033ed43a1134c2223dac8c5d9e02e0100fca1

4) Allocate the ctx with kmem_cache_alloc. This is going to collide a bit with your cleanup patch sorry.

	http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=f5a8db16d2876eed8906a4d36f1d0e06ca5490f6

5) Originally qemu had two bitflags for each page and kept 3 states (of the 4 possible with two bits) for each page in order to deal with the races that can happen if one thread is reading the userfaults and another thread is calling the UFFDIO_COPY ioctl in the background. This patch solves all races in the kernel so the two bits per page can be dropped from qemu codebase. I started documenting the races that can materialize by using 2 threads (instead of running the workload single threaded with a single poll event loop), and how userland had to solve them until I decided it was simpler to fix the race in the kernel by running an ad-hoc pagetable walk inside the wait_event()-kind-of-section. This simplified qemu significantly and it doesn't make the kernel much more complicated.

I tried this before in much older versions but I use gup_fast for it and it didn't work well with gup_fast for various reasons.

http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=41efeae4e93f0296436f2a9fc6b28b6b0158512a

After this patch the only reason to call UFFDIO_WAKE is to handle the userfaults in batches in combination with the DONT_WAKE flag of UFFDIO_COPY.

6) I removed the read recursion from mcopy_atomic. This avoids to depend on the write-starvation behavior of rwsem to be safe. After this change the rwsem is free to stop any further down_read if there's a down_write waiting on the lock.

http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=b1e3a08acc9e3f6c2614e89fc3b8e338daa58e18

About other troubles for the non cooperative usage: MADV_DONTNEED likely needs to be trapped too or how do you know that you shall map a zero page instead of the old data at the faulting address?

Thanks,
Andrea

Pavel Emelyanov

author

2015-04-23 06:34 UTC #9 lore raw

Quoted 29 lines · Andrea Arcangeli

Hi Pavel,
Quoted 10 lines · Pavel Emelyanov
Hi,
On the recent LSF Andrea presented his userfault-fd patches and I had shown some issues that appear in usage scenarios when the monitor task and mm task do not cooperate to each other on VM changes (and fork()-s).
Here's the implementation of the extended uffd API that would help us to address those issues.
Quoted 8 lines
As proof of concept I've implemented only fix for fork() case, but I also plan to add the mremap() and exit() notifications, both are also required for such non-cooperative usage.
More details about the extension itself is in patch #2 and the fork() notification description is in patch #3.
Comments and suggestion are warmly welcome :)
This looks feasible.
Quoted 4 lines
Andrea, what's the best way to go on with the patches -- would you prefer to include them in your git tree or should I instead continue with them on my own, re-sending them when required? Either way would be OK for me.

Quoted 5 lines

Ok so various improvements happened in userfaultfd patchset over the last month so your incremental patchset likely requires a rebase sorry. When you posted it I was in the middle of the updates. Now things are working stable and I have no pending updates, so it would be a good time for a rebase.

OK, thanks for the heads up! I will rebase my patches.

Quoted 6 lines

I can merge it if you like, it's your call if you prefer to maintain it incrementally or if I should merge it, but I wouldn't push it to Andrew for upstream integration in the first batch, as this complicates things further and it's not fully complete yet (at least the version posted only handled fork). I think it can be merged incrementally in a second stage.

Sure!

Quoted 75 lines

The major updates of the userfaultfd patchset over the last month were:
1) Various mixed fixes thanks to the feedback from Dave Hansen and David Gilbert.
The most notable one is the use of mm_users instead of mm_count to pin the mm to avoid crashes that assumed the vma still existed (in the userfaultfd_release method and in the various ioctl). exit_mmap doesn't even set mm->mmap to NULL, so unless I introduce a userfaultfd_exit to call in mmput, I have to pin the mm_users to be safe. This is mainly an issue for the non-cooperative usage you're implementing. Can you catch the exit somehow so you can close the fd? The memory won't be released until you do. Is this ok with you? I suppose you had to close the fd somehow anyway.
2) userfaults are waken immediately even if they're not been "read" yet, this can lead to POLLIN false positives (so I only allow poll if the fd is open in nonblocking mode to be sure it won't hang). Is it too paranoid to return POLLERR if the fd is not open in nonblocking mode?
	http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=f222d9de0a5302dc8ac62d6fab53a84251098751
3) optimize read to return entries in O(1) and poll which was already O(1) becomes lockless. This required to split the waitqueue in two, one for pending faults and one for non pending faults, and the faults are refiled across the two waitqueues when they're read. Both waitqueues are protected by a single lock to be simpler and faster at runtime (the fault_pending_wqh one).
	http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=9aa033ed43a1134c2223dac8c5d9e02e0100fca1
4) Allocate the ctx with kmem_cache_alloc. This is going to collide a bit with your cleanup patch sorry.
	http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=f5a8db16d2876eed8906a4d36f1d0e06ca5490f6
5) Originally qemu had two bitflags for each page and kept 3 states (of the 4 possible with two bits) for each page in order to deal with the races that can happen if one thread is reading the userfaults and another thread is calling the UFFDIO_COPY ioctl in the background. This patch solves all races in the kernel so the two bits per page can be dropped from qemu codebase. I started documenting the races that can materialize by using 2 threads (instead of running the workload single threaded with a single poll event loop), and how userland had to solve them until I decided it was simpler to fix the race in the kernel by running an ad-hoc pagetable walk inside the wait_event()-kind-of-section. This simplified qemu significantly and it doesn't make the kernel much more complicated.
I tried this before in much older versions but I use gup_fast for it and it didn't work well with gup_fast for various reasons.
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=41efeae4e93f0296436f2a9fc6b28b6b0158512a
After this patch the only reason to call UFFDIO_WAKE is to handle the userfaults in batches in combination with the DONT_WAKE flag of UFFDIO_COPY.
6) I removed the read recursion from mcopy_atomic. This avoids to depend on the write-starvation behavior of rwsem to be safe. After this change the rwsem is free to stop any further down_read if there's a down_write waiting on the lock.
   	  http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=b1e3a08acc9e3f6c2614e89fc3b8e338daa58e18
About other troubles for the non cooperative usage: MADV_DONTNEED likely needs to be trapped too or how do you know that you shall map a zero page instead of the old data at the faulting address?
Thanks,
Andrea
.

A message that is not in the archive (sent off-list or not yet received): [email protected]

Pavel Emelyanov

author

2015-04-30 16:38 UTC #11 lore raw

PATCH 1/1 UserfaultFD: Rename uffd_api.bits into .features

Hi,

This is (seem to be) the minimal thing that is required to unblock standard uffd usage from the non-cooperative one. Now more bits can be added to the features field indicating e.g. UFFD_FEATURE_FORK and others needed for the latter use-case.

Signed-off-by Pavel Emelyanov <[email protected]>

---

fs/userfaultfd.c +2 -2

index d2db3e1..c4d0216 100644
@@ -1090,7 +1090,7 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
 		goto out;
 	}
 	/* careful not to leak info, we only read the first 8 bytes */
-	uffdio_api.bits = UFFD_API_BITS;
+	uffdio_api.features = UFFD_API_FEATURES;
 	uffdio_api.ioctls = UFFD_API_IOCTLS;
 	ret = -EFAULT;
 	if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api)))

@@ -1159,7 +1159,7 @@ static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f)
 	 *	protocols: aa:... bb:...
 	 */
 	seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n",
-		   pending, total, UFFD_API, UFFD_API_BITS,
+		   pending, total, UFFD_API, UFFD_API_FEATURES,
 		   UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS);
 }
 #endif

include/uapi/linux/userfaultfd.h +8 -2

index aecd64b..3dc0cba 100644
@@ -11,7 +11,7 @@
 
 #define UFFD_API ((__u64)0xAA)
 /* FIXME: add "|UFFD_BIT_WP" to UFFD_API_BITS after implementing it */
-#define UFFD_API_BITS (UFFD_BIT_WRITE)
+#define UFFD_API_FEATURES (UFFD_FEATURE_WRITE_BIT)
 #define UFFD_API_IOCTLS				\
 	((__u64)1 << _UFFDIO_REGISTER |		\
 	 (__u64)1 << _UFFDIO_UNREGISTER |	\

@@ -63,12 +63,18 @@
 #define UFFD_BIT_WP	(1<<1)	/* handle_userfault() reason VM_UFFD_WP */
 #define UFFD_BITS	2	/* two above bits used for UFFD_BIT_* mask */
 
+/*
+ * Features reported in uffdio_api.features field
+ */
+#define UFFD_FEATURE_WRITE_BIT	(1<<0) /* Corresponds to UFFD_BIT_WRITE */
+#define UFFD_FEATURE_WP_BIT	(1<<1) /* Corresponds to UFFD_BIT_WP */
+
 struct uffdio_api {
 	/* userland asks for an API number */
 	__u64 api;
 
 	/* kernel answers below with the available features for the API */
-	__u64 bits;
+	__u64 features;
 	__u64 ioctls;
 };

Andrea Arcangeli

Applied

2015-05-07 13:42 UTC #12 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Hi Pavel,

Quoted 8 lines · Pavel Emelyanov

Hi,
This is (seem to be) the minimal thing that is required to unblock standard uffd usage from the non-cooperative one. Now more bits can be added to the features field indicating e.g. UFFD_FEATURE_FORK and others needed for the latter use-case.
Signed-off-by Pavel Emelyanov <[email protected]>

Applied.

http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=c2dee3384770a953cbad27b46854aa6fd13656c6
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=d0df59f21f2cde4c49879c00586ce3cb1e3860fe

I was also asked if we could return the full address of the fault including the page offset. In the end I also implemented this incremental to your change:

http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=c308fc81b0a9c53c11b33331ad00d8e5b9763e60

Let me know if you're ok with it. The commit header explains more why I think the bits below PAGE_SHIFT of the fault address aren't interesting but why I did this change anyway.

After reviewing this last change I think it's time to make a proper submit and it's polished enough for merging in -mm after proper review of the full patchset.

Thanks,
Andrea

Pavel Emelyanov

author

2015-05-07 14:28 UTC #13 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 10 lines · Andrea Arcangeli

Hi Pavel,
Quoted 7 lines · Pavel Emelyanov
Hi,
This is (seem to be) the minimal thing that is required to unblock standard uffd usage from the non-cooperative one. Now more bits can be added to the features field indicating e.g. UFFD_FEATURE_FORK and others needed for the latter use-case.

Quoted 6 lines

Quoted 1 line
Signed-off-by Pavel Emelyanov <[email protected]>
Applied.
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=c2dee3384770a953cbad27b46854aa6fd13656c6
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=d0df59f21f2cde4c49879c00586ce3cb1e3860fe

Great!

Quoted 7 lines

I was also asked if we could return the full address of the fault including the page offset. In the end I also implemented this incremental to your change:
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=c308fc81b0a9c53c11b33331ad00d8e5b9763e60
Let me know if you're ok with it.

Yup, this is very close to what I did in my set -- introduced a message to report back to the user-space on read. But my message is more than 8+2*1 bytes, so we'll have one message for 0xAA API and another one for 0xAB (new) one :)

Quoted 7 lines

The commit header explains more why
I think the bits below PAGE_SHIFT of the fault address aren't interesting but why I did this change anyway.
After reviewing this last change I think it's time to make a proper submit and it's polished enough for merging in -mm after proper review of the full patchset.

Yup, fully agree :) And I will soon send the re-based non-cooperative patchset with new API version, longer messages, fork and remap events reporting.

-- Pavel

Andrea Arcangeli

2015-05-07 14:33 UTC #14 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 3 lines · Pavel Emelyanov

Yup, this is very close to what I did in my set -- introduced a message to report back to the user-space on read. But my message is more than 8+2*1 bytes, so we'll have one message for 0xAA API and another one for 0xAB (new) one :)

I slightly altered it to fix an issue with packet alignments so it'd be 16bytes.

How big is your msg currently? Could we get to use the same API?

UFFDIO_REGISTER_MODE_FORK

UFFDIO_REGISTER_MODE_NON_COOPERATIVE would differentiate if you want to register for fork/mremap/dontneed events as well or only the default (UFFD_EVENT_PAGEFAULT).

Pavel Emelyanov

author

2015-05-07 14:42 UTC #15 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 9 lines · Andrea Arcangeli

Quoted 3 lines · Pavel Emelyanov
Yup, this is very close to what I did in my set -- introduced a message to report back to the user-space on read. But my message is more than 8+2*1 bytes, so we'll have one message for 0xAA API and another one for 0xAB (new) one :)
I slightly altered it to fix an issue with packet alignments so it'd be 16bytes.
How big is your msg currently? Could we get to use the same API?

Right now it's like this

struct uffd_event {
        __u64 type;
        union {
                struct {
                        __u64 addr;
                } pagefault;

                struct {
                        __u32 ufd;
                } fork;

                struct {
                        __u64 from;
                        __u64 to;
                        __u64 len;
                } remap;
        } arg;
};

where .type is your uffd_msg.event and the rest is event-specific.

Quoted 7 lines

UFFDIO_REGISTER_MODE_FORK
or
UFFDIO_REGISTER_MODE_NON_COOPERATIVE would differentiate if you want to register for fork/mremap/dontneed events as well or only the default (UFFD_EVENT_PAGEFAULT).

I planned to use this in UFFDIO_API call -- the uffdio_api.features will be in-out argument denoting the bits user needs and reporting what kernel can.

-- Pavel

Andrea Arcangeli

2015-05-07 15:11 UTC #16 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 32 lines · Pavel Emelyanov

Quoted 9 lines · Andrea Arcangeli
Quoted 3 lines · Pavel Emelyanov
Yup, this is very close to what I did in my set -- introduced a message to report back to the user-space on read. But my message is more than 8+2*1 bytes, so we'll have one message for 0xAA API and another one for 0xAB (new) one :)
I slightly altered it to fix an issue with packet alignments so it'd be 16bytes.
How big is your msg currently? Could we get to use the same API?
Right now it's like this
struct uffd_event {
        __u64 type;
        union {
                struct {
                        __u64 addr;
                } pagefault;

                struct {
                        __u32 ufd;
                } fork;

                struct {
                        __u64 from;
                        __u64 to;
                        __u64 len;
                } remap;
        } arg;
};

Quoted 1 line

where .type is your uffd_msg.event and the rest is event-specific.

So you have two more __u64.

In theory if msg.event == UFFD_EVENT_MREMAP you could have the "from" encoded in the msg.arg and then userland could read 16 more bytes knowing it'll get "to len" and we won't have to alter the UFFD_API for adding new EVENT that requires bigger msg size. But it's probably not worth it as an enter/exit kernel is way more costly than reading 16 more bytes, if we already know we need 32bytes in the end.

MADV_DONTNEED shouldn't need more bytes than mremap either.

I think it's ok if I enlarge it to 32bytes.

Quoted 5 lines

Quoted 5 lines
UFFDIO_REGISTER_MODE_FORK
or

Quoted 7 lines

Quoted 3 lines
UFFDIO_REGISTER_MODE_NON_COOPERATIVE would differentiate if you want to register for fork/mremap/dontneed events as well or only the default (UFFD_EVENT_PAGEFAULT).
I planned to use this in UFFDIO_API call -- the uffdio_api.features will be in-out argument denoting the bits user needs and reporting what kernel can.

Ok I guess in-out and checking api.features is easier than checking the vma during the fault. That's ok for me, plus if needed the registration flag can still be added later in addition of the in-out api.features.

So I also need to error out the UFFDIO_API call if api.features is not zero, ok?

After those two changes you should be ok with the same API?

We may still need a new API later of course, it's hard to predict the future and all possible future usages of the userfaultfd... but perhaps this will be enough...

Pavel Emelyanov

author

2015-05-07 15:20 UTC #17 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 37 lines · Andrea Arcangeli

Quoted 32 lines · Pavel Emelyanov
Quoted 9 lines · Andrea Arcangeli
Quoted 3 lines · Pavel Emelyanov
Yup, this is very close to what I did in my set -- introduced a message to report back to the user-space on read. But my message is more than 8+2*1 bytes, so we'll have one message for 0xAA API and another one for 0xAB (new) one :)
I slightly altered it to fix an issue with packet alignments so it'd be 16bytes.
How big is your msg currently? Could we get to use the same API?
Right now it's like this
struct uffd_event {
        __u64 type;
        union {
                struct {
                        __u64 addr;
                } pagefault;

                struct {
                        __u32 ufd;
                } fork;

                struct {
                        __u64 from;
                        __u64 to;
                        __u64 len;
                } remap;
        } arg;
};
Quoted 1 line
where .type is your uffd_msg.event and the rest is event-specific.
So you have two more __u64.

Quoted 8 lines

In theory if msg.event == UFFD_EVENT_MREMAP you could have the "from" encoded in the msg.arg and then userland could read 16 more bytes knowing it'll get "to len" and we won't have to alter the UFFD_API for adding new EVENT that requires bigger msg size. But it's probably not worth it as an enter/exit kernel is way more costly than reading 16 more bytes, if we already know we need 32bytes in the end.
MADV_DONTNEED shouldn't need more bytes than mremap either.

Yes, this one only needs an address and length.

Quoted 1 line

I think it's ok if I enlarge it to 32bytes.

Cool, then we don't need the 2nd API for that :) At least for now.

Quoted 13 lines

Quoted 5 lines
Quoted 5 lines
UFFDIO_REGISTER_MODE_FORK
or
Quoted 7 lines
Quoted 3 lines
UFFDIO_REGISTER_MODE_NON_COOPERATIVE would differentiate if you want to register for fork/mremap/dontneed events as well or only the default (UFFD_EVENT_PAGEFAULT).
I planned to use this in UFFDIO_API call -- the uffdio_api.features will be in-out argument denoting the bits user needs and reporting what kernel can.

Quoted 7 lines

Ok I guess in-out and checking api.features is easier than checking the vma during the fault. That's ok for me, plus if needed the registration flag can still be added later in addition of the in-out api.features.
So I also need to error out the UFFDIO_API call if api.features is not zero, ok?

Exactly!

Quoted 1 line

After those two changes you should be ok with the same API?

Yes. Longer message (type + 3 u64-s) and the ability to request for extra events is all I need. If you're OK with this being in the 0xAA API, then let's do it. I'll rebase my patches again once this appears in your repo :)

Quoted 3 lines

We may still need a new API later of course, it's hard to predict the future and all possible future usages of the userfaultfd... but perhaps this will be enough...

-- Pavel

Andrea Arcangeli

2015-05-07 17:08 UTC #18 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 2 lines · Pavel Emelyanov

Yes. Longer message (type + 3 u64-s) and the ability to request for extra events is all I need. If you're OK with this being in the 0xAA API, then

This started from the request to get the full address (even if personally I'm not convinced that the bits below PAGE_SHIFT can be meaningful to userland) but I thought we could achieve both things and hopefully this change is for the best.

Can you have a look at this and let me know if it looks ok?

http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=d7ba2f23bb978820aa04f1e338789669eff33a7d
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/tree/fs/userfaultfd.c?h=userfault&id=d7ba2f23bb978820aa04f1e338789669eff33a7d
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/tree/include/uapi/linux/userfaultfd.h?h=userfault&id=d7ba2f23bb978820aa04f1e338789669eff33a7d

Thanks!
Andrea

Pavel Emelyanov

author

2015-05-07 18:35 UTC #19 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

Quoted 8 lines · Andrea Arcangeli

Quoted 2 lines · Pavel Emelyanov
Yes. Longer message (type + 3 u64-s) and the ability to request for extra events is all I need. If you're OK with this being in the 0xAA API, then
This started from the request to get the full address (even if personally I'm not convinced that the bits below PAGE_SHIFT can be meaningful to userland) but I thought we could achieve both things and hopefully this change is for the best.

Quoted 5 lines

Can you have a look at this and let me know if it looks ok?
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/commit/?h=userfault&id=d7ba2f23bb978820aa04f1e338789669eff33a7d
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/tree/fs/userfaultfd.c?h=userfault&id=d7ba2f23bb978820aa04f1e338789669eff33a7d
http://git.kernel.org/cgit/linux/kernel/git/andrea/aa.git/tree/include/uapi/linux/userfaultfd.h?h=userfault&id=d7ba2f23bb978820aa04f1e338789669eff33a7d

Yup, this works for me, Now I can re-base my patches on it and check for the features to contain FORK, REMAP and MADVDONTNEED and provide more info in the place now occupied with the uffd_msg.reserved fields. And all this w/o introducing new API :)

Thanks!

-- Pavel

Pavel Emelyanov

author

2015-05-08 13:40 UTC #20 lore raw

Re: [PATCH] UserfaultFD: Rename uffd_api.bits into .features

In reply to Andrea Arcangeli

Andrea,

Quoted 10 lines · Andrea Arcangeli

Quoted 2 lines · Pavel Emelyanov
Yes. Longer message (type + 3 u64-s) and the ability to request for extra events is all I need. If you're OK with this being in the 0xAA API, then
This started from the request to get the full address (even if personally I'm not convinced that the bits below PAGE_SHIFT can be meaningful to userland) but I thought we could achieve both things and hopefully this change is for the best.
Can you have a look at this and let me know if it looks ok?

On the recent userfaultfd branch (dee0a1d0) even w/o my patches I see the stack corruption panic. It only appeared after the uffd_msg introduction, but maybe I just was lucky before it :)

[   48.302949] Kernel panic - not syncing: stack-protector: Kernel stack is corrupted in: ffffffff81251ad9
[   48.302949] 
[   48.303032] CPU: 0 PID: 603 Comm: a.out Not tainted 4.1.0-rc2-uffd-criu+ #27
[   48.303032] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[   48.303032]  0000000000000000 00000000e6804e19 ffff88003da13c28 ffffffff816c584a
[   48.303032]  0000000000000000 ffffffff819fc4b8 ffff88003da13ca8 ffffffff816c136e
[   48.303032]  ffff880000000010 ffff88003da13cb8 ffff88003da13c58 00000000e6804e19
[   48.303032] Call Trace:
[   48.303032]  [<ffffffff816c584a>] dump_stack+0x45/0x57
[   48.303032]  [<ffffffff816c136e>] panic+0xd0/0x204
[   48.303032]  [<ffffffff81251ad9>] ? userfault_msg+0x69/0x70
[   48.303032]  [<ffffffff8108c76b>] __stack_chk_fail+0x1b/0x20
[   48.303032]  [<ffffffff81251ad9>] userfault_msg+0x69/0x70
[   48.303032]  [<ffffffff81251ce0>] ? userfaultfd_poll+0x80/0x80
[   48.303032]  [<ffffffff810cf7b0>] ? abort_exclusive_wait+0xb0/0xb0
[   48.303032]  [<ffffffff811be622>] ? handle_mm_fault+0x1692/0x1900
[   48.303032]  [<ffffffff810be358>] ? __enqueue_entity+0x78/0x80
[   48.303032]  [<ffffffff810c2369>] ? pick_next_entity+0xa9/0x190
[   48.303032]  [<ffffffff810c9450>] ? pick_next_task_fair+0x640/0x900
[   48.303032]  [<ffffffff81054e61>] ? __do_page_fault+0x181/0x420
[   48.303032]  [<ffffffff810551c7>] ? trace_do_page_fault+0x47/0x110
[   48.303032]  [<ffffffff8105028e>] ? do_async_page_fault+0x1e/0xe0
[   48.303032]  [<ffffffff816ceb98>] ? async_page_fault+0x28/0x30
[   48.303032] Kernel Offset: disabled
[   48.303032] ---[ end Kernel panic - not syncing: stack-protector: Kernel stack is corrupted in: ffffffff81251ad9

What I do is a trivial test -- fork a kid, make uffd, create and register a mapping, send one to child and access the memory. Child reads the uffd_msg and copies the data back. Crash happens when parent reads the memory and parents sits in read(). The address in question is in the userfaultfd_msg():

ffffffff81251a70 <userfault_msg>:

ffffffff81251a70:       55                      push   %rbp
ffffffff81251a71:       48 89 f8                mov    %rdi,%rax
ffffffff81251a74:       48 89 e5                mov    %rsp,%rbp
ffffffff81251a77:       48 83 ec 10             sub    $0x10,%rsp
ffffffff81251a7b:       c6 00 12                movb   $0x12,(%rax)
ffffffff81251a7e:       65 48 8b 3c 25 28 00    mov    %gs:0x28,%rdi
ffffffff81251a85:       00 00 
ffffffff81251a87:       48 89 7d f8             mov    %rdi,-0x8(%rbp)
ffffffff81251a8b:       31 ff                   xor    %edi,%edi
ffffffff81251a8d:       83 e2 01                and    $0x1,%edx
ffffffff81251a90:       48 c7 45 f0 00 00 00    movq   $0x0,-0x10(%rbp)
ffffffff81251a97:       00 
ffffffff81251a98:       48 c7 45 f8 00 00 00    movq   $0x0,-0x8(%rbp)
ffffffff81251a9f:       00 
ffffffff81251aa0:       48 c7 45 00 00 00 00    movq   $0x0,0x0(%rbp)
ffffffff81251aa7:       00 
ffffffff81251aa8:       48 c7 45 08 00 00 00    movq   $0x0,0x8(%rbp)
ffffffff81251aaf:       00 
ffffffff81251ab0:       48 89 70 10             mov    %rsi,0x10(%rax)
ffffffff81251ab4:       74 04                   je     ffffffff81251aba <userfault_msg+0x4a>
ffffffff81251ab6:       83 48 08 01             orl    $0x1,0x8(%rax)
ffffffff81251aba:       80 e5 10                and    $0x10,%ch
ffffffff81251abd:       74 04                   je     ffffffff81251ac3 <userfault_msg+0x53>
ffffffff81251abf:       83 48 08 02             orl    $0x2,0x8(%rax)
ffffffff81251ac3:       48 8b 7d f8             mov    -0x8(%rbp),%rdi
ffffffff81251ac7:       65 48 33 3c 25 28 00    xor    %gs:0x28,%rdi
ffffffff81251ace:       00 00 
ffffffff81251ad0:       75 02                   jne    ffffffff81251ad4 <userfault_msg+0x64>
ffffffff81251ad2:       c9                      leaveq 
ffffffff81251ad3:       c3                      retq   
ffffffff81251ad4:       e8 77 ac e3 ff          callq  ffffffff8108c750 <__stack_chk_fail>
ffffffff81251ad9:       0f 1f 80 00 00 00 00    nopl   0x0(%rax)

-- Pavel

Pavel Emelyanov

author

2015-05-08 14:14 UTC #21 lore raw

PATCH 1/1 UserfaultFD: Fix stack corruption when zeroing uffd_msg

Signed-off-by Pavel Emelyanov <[email protected]>

---

fs/userfaultfd.c +1 -1

index 026ef99..c89e96f 100644
@@ -134,7 +134,7 @@ static inline void msg_init(struct uffd_msg *msg)
 	 * Must use memset to zero out the paddings or kernel data is
 	 * leaked to userland.
 	 */
-	memset(&msg, 0, sizeof(struct uffd_msg));
+	memset(msg, 0, sizeof(struct uffd_msg));
 }
 
 static inline struct uffd_msg userfault_msg(unsigned long address,

Andrea Arcangeli

2015-05-08 17:54 UTC #22 lore raw

Re: [PATCH] UserfaultFD: Fix stack corruption when zeroing uffd_msg

Quoted 10 lines · Pavel Emelyanov

Signed-off-by Pavel Emelyanov <[email protected]>
---
fs/userfaultfd.c +0 -0
index 026ef99..c89e96f 100644

fs/userfaultfd.c static inline void msg_init(struct uffd_msg *msg) · line 137

Quoted 5 lines

fs/userfaultfd.c +1 -1

@@ -134,7 +134,7 @@ static inline void msg_init(struct uffd_msg *msg)
 	 * Must use memset to zero out the paddings or kernel data is
 	 * leaked to userland.
 	 */
-	memset(&msg, 0, sizeof(struct uffd_msg));
+	memset(msg, 0, sizeof(struct uffd_msg));

Applied (folded). The memset was originally inline in userfault_msg, when I introduced msg_init the compiler didn't warn of the & that had to be removed as it's now called as:

	msg_init(&msg);

All right thanks!