fanotify: add pre-content hooks

Previously we would only include optional information if you requested it via an FAN_ flag at fanotify_init time (FAN_REPORT_FID for example). However this isn't necessary as the event length is encoded in the metadata, and if the user doesn't want to consume the information they don't have to.

With the PRE_ACCESS events we will always generate range
information, so drop this check in order to allow this extra information to be exported without needing to have another flag.

Signed-off-by Josef Bacik <[email protected]>

---
 fs/notify/fanotify/fanotify_user.c | 13 ++++---------
 1 file changed, 4 insertions(+), 9 deletions(-)

fs/notify/fanotify/fanotify_user.c +4 -9

index 9ec313e9f6e1..2e2fba8a9d20 100644
@@ -160,9 +160,6 @@ static size_t fanotify_event_len(unsigned int info_mode,
 	int fh_len;
 	int dot_len = 0;
 
-	if (!info_mode)
-		return event_len;
-
 	if (fanotify_is_error_event(event->mask))
 		event_len += FANOTIFY_ERROR_INFO_LEN;
 

@@ -740,12 +737,10 @@ static ssize_t copy_event_to_user(struct fsnotify_group *group,
 	if (fanotify_is_perm_event(event->mask))
 		FANOTIFY_PERM(event)->fd = fd;
 
-	if (info_mode) {
-		ret = copy_info_records_to_user(event, info, info_mode, pidfd,
-						buf, count);
-		if (ret < 0)
-			goto out_close_fd;
-	}
+	ret = copy_info_records_to_user(event, info, info_mode, pidfd,
+					buf, count);
+	if (ret < 0)
+		goto out_close_fd;
 
 	if (f)
 		fd_install(fd, f);

Signature · 1 line

2.43.0

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

Quoted 12 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

Quoted 7 lines

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

The patch looks good. Just out of curiosity:

Quoted 8 lines

include/linux/fsnotify_backend.h +0 -0

index 8be029bc50b1..21e72b837ec5 100644
@@ -56,6 +56,9 @@
 #define FS_ACCESS_PERM               0x00020000      /* access event in a permissions hook */
 #define FS_OPEN_EXEC_PERM    0x00040000      /* open/exec event in a permission hook */
 

include/linux/fsnotify_backend.h line 60

Quoted 2 lines

include/linux/fsnotify_backend.h +2 -0

+#define FS_PRE_ACCESS                0x00100000      /* Pre-content access hook */
+#define FS_PRE_MODIFY                0x00200000      /* Pre-content modify hook */

Why is a hole left here in the flag space?

Quoted 32 lines · Jan Kara

Quoted 12 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

Quoted 7 lines

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

The patch looks good. Just out of curiosity:

Quoted 8 lines

include/linux/fsnotify_backend.h +0 -0

index 8be029bc50b1..21e72b837ec5 100644
@@ -56,6 +56,9 @@
 #define FS_ACCESS_PERM               0x00020000      /* access event in a permissions hook */
 #define FS_OPEN_EXEC_PERM    0x00040000      /* open/exec event in a permission hook */
 

Quoted 4 lines

include/linux/fsnotify_backend.h line 60

Quoted 2 lines

include/linux/fsnotify_backend.h +2 -0

+#define FS_PRE_ACCESS                0x00100000      /* Pre-content access hook */
+#define FS_PRE_MODIFY                0x00200000      /* Pre-content modify hook */

Why is a hole left here in the flag space?

Can't remember.

Currently we have a draft design for two more events FS_PATH_ACCESS, FS_PATH_MODIFY
https://github.com/amir73il/man-pages/commits/fan_pre_path

So might have been a desire to keep the pre-events group on the nibble.

Quoted 37 lines · Amir Goldstein

Quoted 31 lines · Jan Kara

Quoted 12 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

Quoted 7 lines

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

The patch looks good. Just out of curiosity:

Quoted 8 lines

include/linux/fsnotify_backend.h +0 -0

index 8be029bc50b1..21e72b837ec5 100644
@@ -56,6 +56,9 @@
 #define FS_ACCESS_PERM               0x00020000      /* access event in a permissions hook */
 #define FS_OPEN_EXEC_PERM    0x00040000      /* open/exec event in a permission hook */
 

Quoted 4 lines

include/linux/fsnotify_backend.h line 60

Quoted 2 lines

include/linux/fsnotify_backend.h +2 -0

+#define FS_PRE_ACCESS                0x00100000      /* Pre-content access hook */
+#define FS_PRE_MODIFY                0x00200000      /* Pre-content modify hook */

Why is a hole left here in the flag space?

Quoted 7 lines

Can't remember.

Currently we have a draft design for two more events FS_PATH_ACCESS, FS_PATH_MODIFY
https://github.com/amir73il/man-pages/commits/fan_pre_path

So might have been a desire to keep the pre-events group on the nibble.

Funny story.

I straced a program with latest FS_PRE_ACCESS (0x00080000) and see what I got:

fanotify_mark(3, FAN_MARK_ADD|FAN_MARK_MOUNT, FAN_CLOSE_WRITE|FAN_OPEN_PERM|FAN_ACCESS_PERM|FAN_DIR_MODIFY|FAN_ONDIR, AT_FDCWD, "/vdd") = 0

"FAN_DIR_MODIFY"! a blast from the past [1]

It would have been nice if we reserved 0x00080000 for FAN_PATH_MODIFY [2] to be a bit less confusing for users with old strace.

WDYT?

Quoted 57 lines · Amir Goldstein

Quoted 37 lines · Amir Goldstein

Quoted 31 lines · Jan Kara

Quoted 12 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

Quoted 7 lines

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

The patch looks good. Just out of curiosity:

Quoted 8 lines

include/linux/fsnotify_backend.h +0 -0

index 8be029bc50b1..21e72b837ec5 100644
@@ -56,6 +56,9 @@
 #define FS_ACCESS_PERM               0x00020000      /* access event in a permissions hook */
 #define FS_OPEN_EXEC_PERM    0x00040000      /* open/exec event in a permission hook */
 

Quoted 4 lines

include/linux/fsnotify_backend.h line 60

Quoted 2 lines

include/linux/fsnotify_backend.h +2 -0

+#define FS_PRE_ACCESS                0x00100000      /* Pre-content access hook */
+#define FS_PRE_MODIFY                0x00200000      /* Pre-content modify hook */

Why is a hole left here in the flag space?

Quoted 7 lines

Can't remember.

Currently we have a draft design for two more events FS_PATH_ACCESS, FS_PATH_MODIFY
https://github.com/amir73il/man-pages/commits/fan_pre_path

So might have been a desire to keep the pre-events group on the nibble.

Funny story.

I straced a program with latest FS_PRE_ACCESS (0x00080000) and see what I got:

fanotify_mark(3, FAN_MARK_ADD|FAN_MARK_MOUNT, FAN_CLOSE_WRITE|FAN_OPEN_PERM|FAN_ACCESS_PERM|FAN_DIR_MODIFY|FAN_ONDIR, AT_FDCWD, "/vdd") = 0

"FAN_DIR_MODIFY"! a blast from the past [1]

Quoted 4 lines

It would have been nice if we reserved 0x00080000 for FAN_PATH_MODIFY [2] to be a bit less confusing for users with old strace.

WDYT?

Yeah, reusing that bit for something semantically close would reduce some confusion. But realistically I don't think FAN_DIR_MODIFY go wide use when it was never supported in a released upstream kernel.

Quoted 62 lines · Jan Kara

Quoted 60 lines · Amir Goldstein

Quoted 37 lines · Amir Goldstein

Quoted 31 lines · Jan Kara

Quoted 12 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

Quoted 7 lines

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

The patch looks good. Just out of curiosity:

Quoted 8 lines

include/linux/fsnotify_backend.h +0 -0

index 8be029bc50b1..21e72b837ec5 100644
@@ -56,6 +56,9 @@
 #define FS_ACCESS_PERM               0x00020000      /* access event in a permissions hook */
 #define FS_OPEN_EXEC_PERM    0x00040000      /* open/exec event in a permission hook */
 

Quoted 4 lines

include/linux/fsnotify_backend.h line 60

Quoted 2 lines

include/linux/fsnotify_backend.h +2 -0

+#define FS_PRE_ACCESS                0x00100000      /* Pre-content access hook */
+#define FS_PRE_MODIFY                0x00200000      /* Pre-content modify hook */

Why is a hole left here in the flag space?

Quoted 7 lines

Can't remember.

Currently we have a draft design for two more events FS_PATH_ACCESS, FS_PATH_MODIFY
https://github.com/amir73il/man-pages/commits/fan_pre_path

So might have been a desire to keep the pre-events group on the nibble.

Funny story.

I straced a program with latest FS_PRE_ACCESS (0x00080000) and see what I got:

fanotify_mark(3, FAN_MARK_ADD|FAN_MARK_MOUNT, FAN_CLOSE_WRITE|FAN_OPEN_PERM|FAN_ACCESS_PERM|FAN_DIR_MODIFY|FAN_ONDIR, AT_FDCWD, "/vdd") = 0

"FAN_DIR_MODIFY"! a blast from the past [1]

It would have been nice if we reserved 0x00080000 for FAN_PATH_MODIFY [2] to be a bit less confusing for users with old strace.

Quoted 5 lines

Quoted 1 line

WDYT?

Yeah, reusing that bit for something semantically close would reduce some confusion. But realistically I don't think FAN_DIR_MODIFY go wide use when it was never supported in a released upstream kernel.

No, but its legacy lives in strace forever...

Quoted 64 lines · Amir Goldstein

Quoted 62 lines · Jan Kara

Quoted 60 lines · Amir Goldstein

Quoted 37 lines · Amir Goldstein

Quoted 31 lines · Jan Kara

Quoted 12 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

The new FS_PRE_ACCESS permission event is similar to FS_ACCESS_PERM, but it meant for a different use case of filling file content before access to a file range, so it has slightly different semantics.

Generate FS_PRE_ACCESS/FS_ACCESS_PERM as two seperate events, same as we did for FS_OPEN_PERM/FS_OPEN_EXEC_PERM.

FS_PRE_MODIFY is a new permission event, with similar semantics as FS_PRE_ACCESS, which is called before a file is modified.

Quoted 7 lines

FS_ACCESS_PERM is reported also on blockdev and pipes, but the new pre-content events are only reported for regular files and dirs.

The pre-content events are meant to be used by hierarchical storage managers that want to fill the content of files on first access.

Signed-off-by Amir Goldstein <[email protected]>

The patch looks good. Just out of curiosity:

Quoted 8 lines

include/linux/fsnotify_backend.h +0 -0

index 8be029bc50b1..21e72b837ec5 100644
@@ -56,6 +56,9 @@
 #define FS_ACCESS_PERM               0x00020000      /* access event in a permissions hook */
 #define FS_OPEN_EXEC_PERM    0x00040000      /* open/exec event in a permission hook */
 

Quoted 4 lines

include/linux/fsnotify_backend.h line 60

Quoted 2 lines

include/linux/fsnotify_backend.h +2 -0

+#define FS_PRE_ACCESS                0x00100000      /* Pre-content access hook */
+#define FS_PRE_MODIFY                0x00200000      /* Pre-content modify hook */

Why is a hole left here in the flag space?

Quoted 7 lines

Can't remember.

Currently we have a draft design for two more events FS_PATH_ACCESS, FS_PATH_MODIFY
https://github.com/amir73il/man-pages/commits/fan_pre_path

So might have been a desire to keep the pre-events group on the nibble.

Funny story.

I straced a program with latest FS_PRE_ACCESS (0x00080000) and see what I got:

fanotify_mark(3, FAN_MARK_ADD|FAN_MARK_MOUNT, FAN_CLOSE_WRITE|FAN_OPEN_PERM|FAN_ACCESS_PERM|FAN_DIR_MODIFY|FAN_ONDIR, AT_FDCWD, "/vdd") = 0

"FAN_DIR_MODIFY"! a blast from the past [1]

It would have been nice if we reserved 0x00080000 for FAN_PATH_MODIFY [2] to be a bit less confusing for users with old strace.

Quoted 8 lines

Quoted 5 lines

Quoted 1 line

WDYT?

Yeah, reusing that bit for something semantically close would reduce some confusion. But realistically I don't think FAN_DIR_MODIFY go wide use when it was never supported in a released upstream kernel.

No, but its legacy lives in strace forever...

Speaking of legacy events, you will notice that in the fan_pre_access branch I swapped the order of FS_PRE_ACCESS to be generated before FS_ACCESS_PERM.

It is a semantic difference that probably does not matter much in practice, but I justified it as "need to fill the content before content can be inspected" because FS_ACCESS_PERM is the legacy Anti-malware event.

This order is also aligned with the priority group associated with those events (PRE_CONTENT before CONTENT).

But from a wider POV, my feeling is that FS_ACCESS_PERM is not really used by anyone and it is baggage that we need to try to get rid of. It is not worth the bloat of the inlined fsnotify_file_area_perm() hook. It is not worth the wasted cycles in the __fsnotify_parent() call that will not be optimized when there is any high priority group listener on the sb.

I am tempted to try and combine the PRE/PERM access events into a single event and make sure that no fanotify group can subscribe to both of them at the same time, so a combined event can never be seen, but it is not very easy to rationalize this API.

For example, if we would have required FAN_REPORT_RANGE init flag for subscribing to FAN_PRE_ACCESS, then we could have denied the legacy FAN_ACCESS_PERM in this group, but I don't think that we want to do that (?).

From: Amir Goldstein <[email protected]>

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on open depending on file open mode. The pre-content event will be generated in addition to FS_OPEN_PERM, but without sb_writers held and after file was truncated in case file was opened with O_CREAT and/or O_TRUNC.

The event will have a range info of (0..0) to provide an opportunity to fill entire file content on open.

Signed-off-by Amir Goldstein <[email protected]>

Quoted 11 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on open depending on file open mode. The pre-content event will be generated in addition to FS_OPEN_PERM, but without sb_writers held and after file was truncated in case file was opened with O_CREAT and/or O_TRUNC.

The event will have a range info of (0..0) to provide an opportunity to fill entire file content on open.

Signed-off-by Amir Goldstein <[email protected]>

...

include/linux/fsnotify.h static inline int fsnotify_file_area_perm(struct file *file, int perm_mask, · line 184

Quoted 6 lines

include/linux/fsnotify.h +2 -0

@@ -176,10 +180,14 @@ static inline int fsnotify_file_area_perm(struct file *file, int perm_mask,
 
 /*
  * fsnotify_file_perm - permission hook before file access
+ *
+ * Called from read()/write() with perm_mas MAY_READ/MAY_WRITE.

                                        ^^^ perm_mask

include/linux/fsnotify.h static inline int fsnotify_file_area_perm(struct file *file, int perm_mask, · line 186

Quoted 2 lines

include/linux/fsnotify.h +2 -0

+ * Called from open() with MAY_OPEN in addition to fsnotify_open_perm(),
+ * but without sb_writers held and after the file was truncated.

This sentence is a bit confusing to me (although I think I understand what you want to say). How about just:

* Called from open() with MAY_OPEN without sb_writers held and after the
* file was truncated. Note that this is a different event from
* fsnotify_open_perm().

From: Amir Goldstein <[email protected]>

Similar to FAN_ACCESS_PERM permission event, but it is only allowed with class FAN_CLASS_PRE_CONTENT and only allowed on regular files are dirs.

Unlike FAN_ACCESS_PERM, it is safe to write to the file being accessed in the context of the event handler.

This pre-content event is meant to be used by hierarchical storage managers that want to fill the content of files on first read access.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify.c      |  3 ++-
 fs/notify/fanotify/fanotify_user.c | 17 ++++++++++++++---
 include/linux/fanotify.h           | 14 ++++++++++----
 include/uapi/linux/fanotify.h      |  2 ++
 4 files changed, 28 insertions(+), 8 deletions(-)

fs/notify/fanotify/fanotify.c +2 -1

index 224bccaab4cc..7dac8e4486df 100644
@@ -910,8 +910,9 @@ static int fanotify_handle_event(struct fsnotify_group *group, u32 mask,
 	BUILD_BUG_ON(FAN_OPEN_EXEC_PERM != FS_OPEN_EXEC_PERM);
 	BUILD_BUG_ON(FAN_FS_ERROR != FS_ERROR);
 	BUILD_BUG_ON(FAN_RENAME != FS_RENAME);
+	BUILD_BUG_ON(FAN_PRE_ACCESS != FS_PRE_ACCESS);
 
-	BUILD_BUG_ON(HWEIGHT32(ALL_FANOTIFY_EVENT_BITS) != 21);
+	BUILD_BUG_ON(HWEIGHT32(ALL_FANOTIFY_EVENT_BITS) != 22);
 
 	mask = fanotify_group_event_mask(group, iter_info, &match_mask,
 					 mask, data, data_type, dir);

fs/notify/fanotify/fanotify_user.c +14 -3

index 2e2fba8a9d20..c294849e474f 100644
@@ -1628,6 +1628,7 @@ static int fanotify_events_supported(struct fsnotify_group *group,
 				     unsigned int flags)
 {
 	unsigned int mark_type = flags & FANOTIFY_MARK_TYPE_BITS;
+	bool is_dir = d_is_dir(path->dentry);
 	/* Strict validation of events in non-dir inode mask with v5.17+ APIs */
 	bool strict_dir_events = FAN_GROUP_FLAG(group, FAN_REPORT_TARGET_FID) ||
 				 (mask & FAN_RENAME) ||

@@ -1665,9 +1666,15 @@ static int fanotify_events_supported(struct fsnotify_group *group,
 	 * but because we always allowed it, error only when using new APIs.
 	 */
 	if (strict_dir_events && mark_type == FAN_MARK_INODE &&
-	    !d_is_dir(path->dentry) && (mask & FANOTIFY_DIRONLY_EVENT_BITS))
+	    !is_dir && (mask & FANOTIFY_DIRONLY_EVENT_BITS))
 		return -ENOTDIR;
 
+	/* Pre-content events are only supported on regular files and dirs */
+	if (mask & FANOTIFY_PRE_CONTENT_EVENTS) {
+		if (!is_dir && !d_is_reg(path->dentry))
+			return -EINVAL;
+	}
+
 	return 0;
 }
 

@@ -1769,11 +1776,15 @@ static int do_fanotify_mark(int fanotify_fd, unsigned int flags, __u64 mask,
 		goto fput_and_out;
 
 	/*
-	 * Permission events require minimum priority FAN_CLASS_CONTENT.
+	 * Permission events are not allowed for FAN_CLASS_NOTIF.
+	 * Pre-content permission events are not allowed for FAN_CLASS_CONTENT.
 	 */
 	ret = -EINVAL;
 	if (mask & FANOTIFY_PERM_EVENTS &&
-	    group->priority < FSNOTIFY_PRIO_CONTENT)
+	    group->priority == FSNOTIFY_PRIO_NORMAL)
+		goto fput_and_out;
+	else if (mask & FANOTIFY_PRE_CONTENT_EVENTS &&
+		 group->priority == FSNOTIFY_PRIO_CONTENT)
 		goto fput_and_out;
 
 	if (mask & FAN_FS_ERROR &&

include/linux/fanotify.h +10 -4

index 4f1c4f603118..5c811baf44d2 100644
@@ -88,6 +88,16 @@
 #define FANOTIFY_DIRENT_EVENTS	(FAN_MOVE | FAN_CREATE | FAN_DELETE | \
 				 FAN_RENAME)
 
+/* Content events can be used to inspect file content */
+#define FANOTIFY_CONTENT_PERM_EVENTS (FAN_OPEN_PERM | FAN_OPEN_EXEC_PERM | \
+				      FAN_ACCESS_PERM)
+/* Pre-content events can be used to fill file content */
+#define FANOTIFY_PRE_CONTENT_EVENTS  (FAN_PRE_ACCESS)
+
+/* Events that require a permission response from user */
+#define FANOTIFY_PERM_EVENTS	(FANOTIFY_CONTENT_PERM_EVENTS | \
+				 FANOTIFY_PRE_CONTENT_EVENTS)
+
 /* Events that can be reported with event->fd */
 #define FANOTIFY_FD_EVENTS (FANOTIFY_PATH_EVENTS | FANOTIFY_PERM_EVENTS)
 

@@ -103,10 +113,6 @@
 				 FANOTIFY_INODE_EVENTS | \
 				 FANOTIFY_ERROR_EVENTS)
 
-/* Events that require a permission response from user */
-#define FANOTIFY_PERM_EVENTS	(FAN_OPEN_PERM | FAN_ACCESS_PERM | \
-				 FAN_OPEN_EXEC_PERM)
-
 /* Extra flags that may be reported with event or control handling of events */
 #define FANOTIFY_EVENT_FLAGS	(FAN_EVENT_ON_CHILD | FAN_ONDIR)
 

include/uapi/linux/fanotify.h +2 -0

index a37de58ca571..3ae43867d318 100644
@@ -26,6 +26,8 @@
 #define FAN_ACCESS_PERM		0x00020000	/* File accessed in perm check */
 #define FAN_OPEN_EXEC_PERM	0x00040000	/* File open/exec in perm check */
 
+#define FAN_PRE_ACCESS		0x00100000	/* Pre-content access hook */
+
 #define FAN_EVENT_ON_CHILD	0x08000000	/* Interested in child events */
 
 #define FAN_RENAME		0x10000000	/* File was renamed */

Signature · 1 line

2.43.0

From: Amir Goldstein <[email protected]>

Generate FAN_PRE_MODIFY permission event from fsnotify_file_perm() pre-write hook to notify fanotify listeners on an intent to make modification to a file.

Like FAN_PRE_ACCESS, it is only allowed with FAN_CLASS_PRE_CONTENT and unlike FAN_MODIFY, it is only allowed on regular files.

Like FAN_PRE_ACCESS, it is generated without sb_start_write() held, so it is safe for to perform filesystem modifications in the the

context of event handler.

Quoted 11 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

Generate FAN_PRE_MODIFY permission event from fsnotify_file_perm() pre-write hook to notify fanotify listeners on an intent to make modification to a file.

Like FAN_PRE_ACCESS, it is only allowed with FAN_CLASS_PRE_CONTENT and unlike FAN_MODIFY, it is only allowed on regular files.

Like FAN_PRE_ACCESS, it is generated without sb_start_write() held, so it is safe for to perform filesystem modifications in the the

^^^ seems superfluous ^^^ twice "the"

Quoted 1 line

context of event handler.

...

include/linux/fanotify.h line 96

Quoted 11 lines

include/linux/fanotify.h +2 -1

index 5c811baf44d2..ae6cb2688d52 100644
@@ -92,7 +92,8 @@
 #define FANOTIFY_CONTENT_PERM_EVENTS (FAN_OPEN_PERM | FAN_OPEN_EXEC_PERM | \
                                    FAN_ACCESS_PERM)
 /* Pre-content events can be used to fill file content */
-#define FANOTIFY_PRE_CONTENT_EVENTS  (FAN_PRE_ACCESS)
+#define FANOTIFY_PRE_CONTENT_EVENTS  (FAN_PRE_ACCESS | FAN_PRE_MODIFY)
+#define FANOTIFY_PRE_MODIFY_EVENTS   (FAN_PRE_MODIFY)

I didn't find FANOTIFY_PRE_MODIFY_EVENTS used anywhere?

Quoted 17 lines · Jan Kara

Quoted 11 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

Generate FAN_PRE_MODIFY permission event from fsnotify_file_perm() pre-write hook to notify fanotify listeners on an intent to make modification to a file.

Like FAN_PRE_ACCESS, it is only allowed with FAN_CLASS_PRE_CONTENT and unlike FAN_MODIFY, it is only allowed on regular files.

Like FAN_PRE_ACCESS, it is generated without sb_start_write() held, so it is safe for to perform filesystem modifications in the the

^^^ seems superfluous ^^^ twice "the"

Quoted 1 line

context of event handler.

...

Quoted 13 lines

include/linux/fanotify.h line 96

Quoted 11 lines

include/linux/fanotify.h +2 -1

index 5c811baf44d2..ae6cb2688d52 100644
@@ -92,7 +92,8 @@
 #define FANOTIFY_CONTENT_PERM_EVENTS (FAN_OPEN_PERM | FAN_OPEN_EXEC_PERM | \
                                    FAN_ACCESS_PERM)
 /* Pre-content events can be used to fill file content */
-#define FANOTIFY_PRE_CONTENT_EVENTS  (FAN_PRE_ACCESS)
+#define FANOTIFY_PRE_CONTENT_EVENTS  (FAN_PRE_ACCESS | FAN_PRE_MODIFY)
+#define FANOTIFY_PRE_MODIFY_EVENTS   (FAN_PRE_MODIFY)

I didn't find FANOTIFY_PRE_MODIFY_EVENTS used anywhere?

Right. It is used later in the sb_write_barrier patches. We can introduce it later if you prefer.

From: Amir Goldstein <[email protected]>

We would like to add file range information to pre-content events.

Pass a struct file_range with optional offset and length to event handler along with pre-content permission event.

Signed-off-by Amir Goldstein <[email protected]>

Quoted 8 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

We would like to add file range information to pre-content events.

Pass a struct file_range with optional offset and length to event handler along with pre-content permission event.

Signed-off-by Amir Goldstein <[email protected]>

...

fs/notify/fanotify/fanotify.c static struct fanotify_event *fanotify_alloc_perm_event(const struct path *path, · line 575

Quoted 8 lines

fs/notify/fanotify/fanotify.c +4 -0

@@ -565,6 +569,10 @@ static struct fanotify_event *fanotify_alloc_perm_event(const struct path *path,
      pevent->hdr.len = 0;
      pevent->state = FAN_EVENT_INIT;
      pevent->path = *path;
+     if (range) {
+             pevent->ppos = range->ppos;
+             pevent->count = range->count;
+     }

Shouldn't we initialze ppos & count in case range info isn't passed?

Quoted 13 lines · Jan Kara

Quoted 8 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

We would like to add file range information to pre-content events.

Pass a struct file_range with optional offset and length to event handler along with pre-content permission event.

Signed-off-by Amir Goldstein <[email protected]>

...

Quoted 10 lines

fs/notify/fanotify/fanotify.c static struct fanotify_event *fanotify_alloc_perm_event(const struct path *path, · line 575

Quoted 8 lines

fs/notify/fanotify/fanotify.c +4 -0

@@ -565,6 +569,10 @@ static struct fanotify_event *fanotify_alloc_perm_event(const struct path *path,
      pevent->hdr.len = 0;
      pevent->state = FAN_EVENT_INIT;
      pevent->path = *path;
+     if (range) {
+             pevent->ppos = range->ppos;
+             pevent->count = range->count;
+     }

Shouldn't we initialze ppos & count in case range info isn't passed?

Currently, range info is always passed in case of fanotify_event_has_access_range(), but for robustness I guess we should.

From: Amir Goldstein <[email protected]>

FANOTIFY_PIDFD_INFO_HDR_LEN is not the length of the header.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify_user.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

fs/notify/fanotify/fanotify_user.c +0 -0

index 3a7101544f30..5ece186d5c50 100644

Quoted 9 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

FANOTIFY_PIDFD_INFO_HDR_LEN is not the length of the header.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify_user.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

fs/notify/fanotify/fanotify_user.c struct kmem_cache *fanotify_perm_event_cachep __ro_after_init; · line 122

Quoted 10 lines

fs/notify/fanotify/fanotify_user.c +1 -1

index 3a7101544f30..5ece186d5c50 100644
@@ -119,7 +119,7 @@ struct kmem_cache *fanotify_perm_event_cachep __ro_after_init;
 #define FANOTIFY_EVENT_ALIGN 4
 #define FANOTIFY_FID_INFO_HDR_LEN \
 	(sizeof(struct fanotify_event_info_fid) + sizeof(struct file_handle))
-#define FANOTIFY_PIDFD_INFO_HDR_LEN \
+#define FANOTIFY_PIDFD_INFO_LEN \

OK, but then FANOTIFY_FID_INFO_HDR_LEN should be renamed as well? Or what's the difference?

From: Amir Goldstein <[email protected]>

With group class FAN_CLASS_PRE_CONTENT, report offset and length info along with FAN_PRE_ACCESS and FAN_PRE_MODIFY permission events.

This information is meant to be used by hierarchical storage managers that want to fill partial content of files on first access to range.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify.h      |  8 +++++++
 fs/notify/fanotify/fanotify_user.c | 38 ++++++++++++++++++++++++++++++
 include/uapi/linux/fanotify.h      |  7 ++++++
 3 files changed, 53 insertions(+)

fs/notify/fanotify/fanotify.h +0 -0

index 93598b7d5952..7f06355afa1f 100644

fs/notify/fanotify/fanotify.h +12 -0

+
+	info.hdr.info_type = FAN_EVENT_INFO_TYPE_RANGE;
+	info.hdr.len = info_len;
+	info.offset = *(pevent->ppos);
+	info.count = pevent->count;
+
+	if (copy_to_user(buf, &info, info_len))
+		return -EFAULT;
+
+	return info_len;
+}
+
 static int copy_info_records_to_user(struct fanotify_event *event,
 				     struct fanotify_info *info,
 				     unsigned int info_mode, int pidfd,

Quoted 23 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

With group class FAN_CLASS_PRE_CONTENT, report offset and length info along with FAN_PRE_ACCESS and FAN_PRE_MODIFY permission events.

This information is meant to be used by hierarchical storage managers that want to fill partial content of files on first access to range.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify.h      |  8 +++++++
 fs/notify/fanotify/fanotify_user.c | 38 ++++++++++++++++++++++++++++++
 include/uapi/linux/fanotify.h      |  7 ++++++
 3 files changed, 53 insertions(+)

fs/notify/fanotify/fanotify.h +0 -0

index 93598b7d5952..7f06355afa1f 100644
@@ -448,6 +448,14 @@ static inline bool fanotify_is_perm_event(u32 mask)
              mask & FANOTIFY_PERM_EVENTS;
 }
 

fs/notify/fanotify/fanotify.h static inline bool fanotify_is_perm_event(u32 mask) · line 457

Quoted 7 lines

fs/notify/fanotify/fanotify.h +7 -0

+static inline bool fanotify_event_has_access_range(struct fanotify_event *event)
+{
+     if (!(event->mask & FANOTIFY_PRE_CONTENT_EVENTS))
+             return false;
+
+     return FANOTIFY_PERM(event)->ppos;
+}

Now I'm a bit confused. Can we have legally NULL ppos for an event from FANOTIFY_PRE_CONTENT_EVENTS?

Quoted 24 lines · Jan Kara

Quoted 23 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

With group class FAN_CLASS_PRE_CONTENT, report offset and length info along with FAN_PRE_ACCESS and FAN_PRE_MODIFY permission events.

This information is meant to be used by hierarchical storage managers that want to fill partial content of files on first access to range.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify.h      |  8 +++++++
 fs/notify/fanotify/fanotify_user.c | 38 ++++++++++++++++++++++++++++++
 include/uapi/linux/fanotify.h      |  7 ++++++
 3 files changed, 53 insertions(+)

fs/notify/fanotify/fanotify.h +0 -0

index 93598b7d5952..7f06355afa1f 100644
@@ -448,6 +448,14 @@ static inline bool fanotify_is_perm_event(u32 mask)
              mask & FANOTIFY_PERM_EVENTS;
 }
 

Quoted 11 lines

fs/notify/fanotify/fanotify.h static inline bool fanotify_is_perm_event(u32 mask) · line 457

Quoted 7 lines

fs/notify/fanotify/fanotify.h +7 -0

+static inline bool fanotify_event_has_access_range(struct fanotify_event *event)
+{
+     if (!(event->mask & FANOTIFY_PRE_CONTENT_EVENTS))
+             return false;
+
+     return FANOTIFY_PERM(event)->ppos;
+}

Now I'm a bit confused. Can we have legally NULL ppos for an event from FANOTIFY_PRE_CONTENT_EVENTS?

Sorry for the very late reply...

The short answer is that NULL FANOTIFY_PERM(event)->ppos simply means that fanotify_alloc_perm_event() was called with NULL range, which is the very common case of legacy permission events.

The long answer is a bit convoluted, so bare with me.
The long answer is to the question whether fsnotify_file_range() can be called with a NULL ppos.

This shouldn't be possible AFAIK for regular files and directories, unless some fs that is marked with FS_ALLOW_HSM opens a regular file with FMODE_STREAM, which should not be happening IMO, but then the assertion belongs inside fsnotify_file_range().

However, there was another way to get NULL ppos before I added the patch "fsnotify: generate pre-content permission event on open"

Which made this "half intentional" change:
static inline int fsnotify_file_perm(struct file *file, int perm_mask) {

(quoted diff) +1 -1

-       return fsnotify_file_area_perm(file, perm_mask, NULL, 0);
+       return fsnotify_file_area_perm(file, perm_mask, &file->f_pos, 0);
 }

In order to implement:
"The event will have a range info of (0..0) to provide an opportunity
to fill the entire file content on open."

The problem is that do_open() was not the only caller of fsnotify_file_perm(). There is another call from iterate_dir() and the change above causes FS_PRE_ACCESS events on readdir to report the directory f_pos - Do we want that? I think we do, but HSM should be able to tell the difference between opendir() and readdir(), because my HSM only wants to fill dir content on the latter.

Quoted 66 lines · Amir Goldstein

Quoted 24 lines · Jan Kara

Quoted 23 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

With group class FAN_CLASS_PRE_CONTENT, report offset and length info along with FAN_PRE_ACCESS and FAN_PRE_MODIFY permission events.

This information is meant to be used by hierarchical storage managers that want to fill partial content of files on first access to range.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify.h      |  8 +++++++
 fs/notify/fanotify/fanotify_user.c | 38 ++++++++++++++++++++++++++++++
 include/uapi/linux/fanotify.h      |  7 ++++++
 3 files changed, 53 insertions(+)

fs/notify/fanotify/fanotify.h +0 -0

index 93598b7d5952..7f06355afa1f 100644
@@ -448,6 +448,14 @@ static inline bool fanotify_is_perm_event(u32 mask)
              mask & FANOTIFY_PERM_EVENTS;
 }
 

Quoted 11 lines

fs/notify/fanotify/fanotify.h static inline bool fanotify_is_perm_event(u32 mask) · line 457

Quoted 7 lines

fs/notify/fanotify/fanotify.h +7 -0

+static inline bool fanotify_event_has_access_range(struct fanotify_event *event)
+{
+     if (!(event->mask & FANOTIFY_PRE_CONTENT_EVENTS))
+             return false;
+
+     return FANOTIFY_PERM(event)->ppos;
+}

Now I'm a bit confused. Can we have legally NULL ppos for an event from FANOTIFY_PRE_CONTENT_EVENTS?

Sorry for the very late reply...

The short answer is that NULL FANOTIFY_PERM(event)->ppos simply means that fanotify_alloc_perm_event() was called with NULL range, which is the very common case of legacy permission events.

The long answer is a bit convoluted, so bare with me.
The long answer is to the question whether fsnotify_file_range() can be called with a NULL ppos.

This shouldn't be possible AFAIK for regular files and directories, unless some fs that is marked with FS_ALLOW_HSM opens a regular file with FMODE_STREAM, which should not be happening IMO, but then the assertion belongs inside fsnotify_file_range().

However, there was another way to get NULL ppos before I added the patch "fsnotify: generate pre-content permission event on open"

Which made this "half intentional" change:
static inline int fsnotify_file_perm(struct file *file, int perm_mask) {

(quoted diff) +1 -1

-       return fsnotify_file_area_perm(file, perm_mask, NULL, 0);
+       return fsnotify_file_area_perm(file, perm_mask, &file->f_pos, 0);
 }

In order to implement:
"The event will have a range info of (0..0) to provide an opportunity
to fill the entire file content on open."

Quoted 6 lines

The problem is that do_open() was not the only caller of fsnotify_file_perm(). There is another call from iterate_dir() and the change above causes FS_PRE_ACCESS events on readdir to report the directory f_pos - Do we want that? I think we do, but HSM should be able to tell the difference between opendir() and readdir(), because my HSM only wants to fill dir content on the latter.

Well, I'm not so sure we want to report fpos on opendir / readdir(). For directories fpos is an opaque cookie that is filesystem dependent and you are not even allowed to carry it from open to open. It is valid only within that one open-close session if I remember right. So userspace HSM cannot do much with it and in my opinion reporting it to userspace is a recipe for abuse...

I'm undecided whether we want to allow pre-access events without range or enforce 0-0 range. I don't think there's a big practical difference.

Quoted 82 lines · Jan Kara

Quoted 66 lines · Amir Goldstein

Quoted 24 lines · Jan Kara

Quoted 23 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

With group class FAN_CLASS_PRE_CONTENT, report offset and length info along with FAN_PRE_ACCESS and FAN_PRE_MODIFY permission events.

This information is meant to be used by hierarchical storage managers that want to fill partial content of files on first access to range.

Signed-off-by Amir Goldstein <[email protected]>

---
 fs/notify/fanotify/fanotify.h      |  8 +++++++
 fs/notify/fanotify/fanotify_user.c | 38 ++++++++++++++++++++++++++++++
 include/uapi/linux/fanotify.h      |  7 ++++++
 3 files changed, 53 insertions(+)

fs/notify/fanotify/fanotify.h +0 -0

index 93598b7d5952..7f06355afa1f 100644
@@ -448,6 +448,14 @@ static inline bool fanotify_is_perm_event(u32 mask)
              mask & FANOTIFY_PERM_EVENTS;
 }
 

Quoted 11 lines

fs/notify/fanotify/fanotify.h static inline bool fanotify_is_perm_event(u32 mask) · line 457

Quoted 7 lines

fs/notify/fanotify/fanotify.h +7 -0

+static inline bool fanotify_event_has_access_range(struct fanotify_event *event)
+{
+     if (!(event->mask & FANOTIFY_PRE_CONTENT_EVENTS))
+             return false;
+
+     return FANOTIFY_PERM(event)->ppos;
+}

Now I'm a bit confused. Can we have legally NULL ppos for an event from FANOTIFY_PRE_CONTENT_EVENTS?

Sorry for the very late reply...

The short answer is that NULL FANOTIFY_PERM(event)->ppos simply means that fanotify_alloc_perm_event() was called with NULL range, which is the very common case of legacy permission events.

The long answer is a bit convoluted, so bare with me.
The long answer is to the question whether fsnotify_file_range() can be called with a NULL ppos.

This shouldn't be possible AFAIK for regular files and directories, unless some fs that is marked with FS_ALLOW_HSM opens a regular file with FMODE_STREAM, which should not be happening IMO, but then the assertion belongs inside fsnotify_file_range().

However, there was another way to get NULL ppos before I added the patch "fsnotify: generate pre-content permission event on open"

Which made this "half intentional" change:
static inline int fsnotify_file_perm(struct file *file, int perm_mask) {

(quoted diff) +1 -1

-       return fsnotify_file_area_perm(file, perm_mask, NULL, 0);
+       return fsnotify_file_area_perm(file, perm_mask, &file->f_pos, 0);
 }

In order to implement:
"The event will have a range info of (0..0) to provide an opportunity
to fill the entire file content on open."

Quoted 6 lines

The problem is that do_open() was not the only caller of fsnotify_file_perm(). There is another call from iterate_dir() and the change above causes FS_PRE_ACCESS events on readdir to report the directory f_pos - Do we want that? I think we do, but HSM should be able to tell the difference between opendir() and readdir(), because my HSM only wants to fill dir content on the latter.

Well, I'm not so sure we want to report fpos on opendir / readdir(). For directories fpos is an opaque cookie that is filesystem dependent and you are not even allowed to carry it from open to open. It is valid only within that one open-close session if I remember right. So userspace HSM cannot do much with it and in my opinion reporting it to userspace is a recipe for abuse...

Quoted 3 lines

I'm undecided whether we want to allow pre-access events without range or enforce 0-0 range. I don't think there's a big practical difference.

So there is a practical difference.
My HSM wants to fill dir content on readdir() and not on opendir(). Other HSMs may want to fill dir content on opendir().
It could do that if opendir() (as does open()) reports range [0..0] and readdir() reports no range.

From: Amir Goldstein <[email protected]>

With FAN_DENY response, user trying to perform the filesystem operation gets an error with errno set to EPERM.

It is useful for hierarchical storage management (HSM) service to be able to deny access for reasons more diverse than EPERM, for example EAGAIN, if HSM could retry the operation later.

Allow fanotify groups with priority FAN_CLASSS_PRE_CONTENT to responsd to permission events with the response value FAN_DENY_ERRNO(errno), instead of FAN_DENY to return a custom error.

Limit custom error to values to some errors expected on read(2)/write(2)

and open(2) of regular files. This list could be extended in the future. Userspace can test for legitimate values of FAN_DENY_ERRNO(errno) by writing a response to an fanotify group fd with a value of FAN_NOFD in the fd field of the response.

The change in fanotify_response is backward compatible, because errno is written in the high 8 bits of the 32bit response field and old kernels reject respose value with high bits set.

Signed-off-by Amir Goldstein <[email protected]>

fs/notify/fanotify/fanotify.c +7 -1

@@ -258,18 +258,25 @@ static int fanotify_get_response(struct fsnotify_group *group,
 	}
 
 	/* userspace responded, convert to something usable */
-	switch (event->response & FANOTIFY_RESPONSE_ACCESS) {
+	switch (FAN_RESPONSE_ACCESS(event->response)) {
 	case FAN_ALLOW:
 		ret = 0;
 		break;
 	case FAN_DENY:
+		/* Check custom errno from pre-content events */
+		errno = FAN_RESPONSE_ERRNO(event->response);
+		if (errno) {
+			ret = -errno;
+			break;
+		}

fs/notify/fanotify/fanotify_user.c +28 -3

index c3c8b2ea80b6..b4d810168521 100644
@@ -337,11 +337,12 @@ static int process_access_response(struct fsnotify_group *group,
 	struct fanotify_perm_event *event;
 	int fd = response_struct->fd;
 	u32 response = response_struct->response;
+	int errno = FAN_RESPONSE_ERRNO(response);
 	int ret = info_len;
 	struct fanotify_response_info_audit_rule friar;
 
-	pr_debug("%s: group=%p fd=%d response=%u buf=%p size=%zu\n", __func__,
-		 group, fd, response, info, info_len);
+	pr_debug("%s: group=%p fd=%d response=%x errno=%d buf=%p size=%zu\n",
+		 __func__, group, fd, response, errno, info, info_len);
 	/*
 	 * make sure the response is valid, if invalid we do nothing and either
 	 * userspace can send a valid response or we will clean it up after the

@@ -350,9 +351,33 @@ static int process_access_response(struct fsnotify_group *group,
 	if (response & ~FANOTIFY_RESPONSE_VALID_MASK)
 		return -EINVAL;
 
-	switch (response & FANOTIFY_RESPONSE_ACCESS) {
+	switch (FAN_RESPONSE_ACCESS(response)) {
 	case FAN_ALLOW:
+		if (errno)
+			return -EINVAL;
+		break;
 	case FAN_DENY:
+		/* Custom errno is supported only for pre-content groups */
+		if (errno && group->priority != FSNOTIFY_PRIO_PRE_CONTENT)
+			return -EINVAL;
+
+		/*
+		 * Limit errno to values expected on open(2)/read(2)/write(2)
+		 * of regular files.
+		 */
+		switch (errno) {
+		case 0:
+		case EIO:
+		case EPERM:
+		case EBUSY:
+		case ETXTBSY:
+		case EAGAIN:
+		case ENOSPC:
+		case EDQUOT:
+			break;
+		default:
+			return -EINVAL;
+		}
 		break;
 	default:
 		return -EINVAL;

include/linux/fanotify.h +0 -0

index ae6cb2688d52..76d818a7d654 100644
@@ -132,7 +132,14 @@
 /* These masks check for invalid bits in permission responses. */
 #define FANOTIFY_RESPONSE_ACCESS (FAN_ALLOW | FAN_DENY)

Quoted 9 lines · Josef Bacik

From: Amir Goldstein <[email protected]>

With FAN_DENY response, user trying to perform the filesystem operation gets an error with errno set to EPERM.

It is useful for hierarchical storage management (HSM) service to be able to deny access for reasons more diverse than EPERM, for example EAGAIN, if HSM could retry the operation later.

Quoted 5 lines

Allow fanotify groups with priority FAN_CLASSS_PRE_CONTENT to responsd to permission events with the response value FAN_DENY_ERRNO(errno), instead of FAN_DENY to return a custom error.

Limit custom error to values to some errors expected on read(2)/write(2)

^^^ parse error. Perhaps: "Limit custom error values to errors
expected on read..."

Quoted 10 lines

and open(2) of regular files. This list could be extended in the future. Userspace can test for legitimate values of FAN_DENY_ERRNO(errno) by writing a response to an fanotify group fd with a value of FAN_NOFD in the fd field of the response.

The change in fanotify_response is backward compatible, because errno is written in the high 8 bits of the 32bit response field and old kernels reject respose value with high bits set.

Signed-off-by Amir Goldstein <[email protected]>

...

Quoted 16 lines

fs/notify/fanotify/fanotify.c +7 -1

@@ -258,18 +258,25 @@ static int fanotify_get_response(struct fsnotify_group *group,
      }
 
      /* userspace responded, convert to something usable */
-     switch (event->response & FANOTIFY_RESPONSE_ACCESS) {
+     switch (FAN_RESPONSE_ACCESS(event->response)) {
      case FAN_ALLOW:
              ret = 0;
              break;
      case FAN_DENY:
+             /* Check custom errno from pre-content events */
+             errno = FAN_RESPONSE_ERRNO(event->response);
+             if (errno) {
+                     ret = -errno;
+                     break;
+             }

fs/notify/fanotify/fanotify.c static int fanotify_get_response(struct fsnotify_group *group, · line 280

Quoted 10 lines

fs/notify/fanotify/fanotify.c +2 -1

+             fallthrough;
      default:
              ret = -EPERM;
      }
 
      /* Check if the response should be audited */
      if (event->response & FAN_AUDIT)
-             audit_fanotify(event->response & ~FAN_AUDIT,
+             audit_fanotify(FAN_RESPONSE_ACCESS(event->response),
                             &event->audit_rule);

I think you need to also keep FAN_INFO in the flags not to break some userspace possibly parsing audit requests.

Quoted 61 lines

fs/notify/fanotify/fanotify_user.c +28 -3

index c3c8b2ea80b6..b4d810168521 100644
@@ -337,11 +337,12 @@ static int process_access_response(struct fsnotify_group *group,
      struct fanotify_perm_event *event;
      int fd = response_struct->fd;
      u32 response = response_struct->response;
+     int errno = FAN_RESPONSE_ERRNO(response);
      int ret = info_len;
      struct fanotify_response_info_audit_rule friar;
 
-     pr_debug("%s: group=%p fd=%d response=%u buf=%p size=%zu\n", __func__,
-              group, fd, response, info, info_len);
+     pr_debug("%s: group=%p fd=%d response=%x errno=%d buf=%p size=%zu\n",
+              __func__, group, fd, response, errno, info, info_len);
      /*
       * make sure the response is valid, if invalid we do nothing and either
       * userspace can send a valid response or we will clean it up after the

@@ -350,9 +351,33 @@ static int process_access_response(struct fsnotify_group *group,
      if (response & ~FANOTIFY_RESPONSE_VALID_MASK)
              return -EINVAL;
 
-     switch (response & FANOTIFY_RESPONSE_ACCESS) {
+     switch (FAN_RESPONSE_ACCESS(response)) {
      case FAN_ALLOW:
+             if (errno)
+                     return -EINVAL;
+             break;
      case FAN_DENY:
+             /* Custom errno is supported only for pre-content groups */
+             if (errno && group->priority != FSNOTIFY_PRIO_PRE_CONTENT)
+                     return -EINVAL;
+
+             /*
+              * Limit errno to values expected on open(2)/read(2)/write(2)
+              * of regular files.
+              */
+             switch (errno) {
+             case 0:
+             case EIO:
+             case EPERM:
+             case EBUSY:
+             case ETXTBSY:
+             case EAGAIN:
+             case ENOSPC:
+             case EDQUOT:
+                     break;
+             default:
+                     return -EINVAL;
+             }
              break;
      default:
              return -EINVAL;

include/linux/fanotify.h +0 -0

index ae6cb2688d52..76d818a7d654 100644
@@ -132,7 +132,14 @@
 /* These masks check for invalid bits in permission responses. */
 #define FANOTIFY_RESPONSE_ACCESS (FAN_ALLOW | FAN_DENY)

include/linux/fanotify.h line 142

Quoted 10 lines

include/linux/fanotify.h +8 -1

 #define FANOTIFY_RESPONSE_FLAGS (FAN_AUDIT | FAN_INFO)
-#define FANOTIFY_RESPONSE_VALID_MASK (FANOTIFY_RESPONSE_ACCESS | FANOTIFY_RESPONSE_FLAGS)
+#define FANOTIFY_RESPONSE_ERRNO      (FAN_ERRNO_MASK << FAN_ERRNO_SHIFT)
+#define FANOTIFY_RESPONSE_VALID_MASK \
+     (FANOTIFY_RESPONSE_ACCESS | FANOTIFY_RESPONSE_FLAGS | \
+      FANOTIFY_RESPONSE_ERRNO)
+
+/* errno other than EPERM can specified in upper byte of deny response */
+#define FAN_RESPONSE_ACCESS(res)     ((res) & FANOTIFY_RESPONSE_ACCESS)
+#define FAN_RESPONSE_ERRNO(res)              ((int)((res) >> FAN_ERRNO_SHIFT))

I have to say I find the names FANOTIFY_RESPONSE_ERRNO and FAN_RESPONSE_ERRNO() (and similarly with FAN_RESPONSE_ACCESS) very similar and thus confusing. I was staring at it for 5 minutes wondering how comes it compiles before I realized one prefix is shorter than the other one so the indentifiers are indeed different. Maybe we'd make these inline functions instead of macros and name them like:

fanotify_get_response_decision()
fanotify_get_response_errno()

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on page fault depending on the faulting method.

This pre-content event is meant to be used by hierarchical storage managers that want to fill in the file content on first read access.

Signed-off-by Josef Bacik <[email protected]>

---
 fs/notify/fsnotify.c             | 13 +++++++++
 include/linux/fsnotify_backend.h | 14 +++++++++
 mm/filemap.c                     | 50 ++++++++++++++++++++++++++++----
 3 files changed, 71 insertions(+), 6 deletions(-)

fs/notify/fsnotify.c +0 -0

index 1ca4a8da7f29..435232d46b4f 100644

fs/notify/fsnotify.c +2 -0

+}
+
 /**
  * fsnotify_unmount_inodes - an sb is unmounting.  handle any watched inodes.
  * @sb: superblock being unmounted.

include/linux/fsnotify_backend.h +14 -0

index 36c3d18cc40a..6983fbf096b8 100644
@@ -900,6 +900,15 @@ static inline void fsnotify_init_event(struct fsnotify_event *event)
        INIT_LIST_HEAD(&event->list);
 }
 
+#ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
+bool fsnotify_file_has_content_watches(struct file *file);
+#else
+static inline bool fsnotify_file_has_content_watches(struct file *file)
+{
+       return false;
+}
+#endif /* CONFIG_FANOTIFY_ACCESS_PERMISSIONS */
+
 #else
 
 static inline int fsnotify(__u32 mask, const void *data, int data_type,

@@ -938,6 +947,11 @@ static inline u32 fsnotify_get_cookie(void)
 static inline void fsnotify_unmount_inodes(struct super_block *sb)
 {}
 
+static inline bool fsnotify_file_has_content_watches(struct file *file)
+{
+       return false;
+}
+
 #endif /* CONFIG_FSNOTIFY */
 
 #endif /* __KERNEL __ */

mm/filemap.c +3 -2

index ca8c8d889eef..cc9d7885bbe3 100644
@@ -46,6 +46,7 @@
 #include <linux/pipe_fs_i.h>
 #include <linux/splice.h>
 #include <linux/rcupdate_wait.h>
+#include <linux/fsnotify.h>
 #include <asm/pgalloc.h>
 #include <asm/tlbflush.h>
 #include "internal.h"

@@ -3112,13 +3113,13 @@ static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
  * that.  If we didn't pin a file then we return NULL.  The file that is
  * returned needs to be fput()'ed when we're done with it.
  */
-static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
+static struct file *do_sync_mmap_readahead(struct vm_fault *vmf,
+                                          struct file *fpin)
 {
        struct file *file = vmf->vma->vm_file;
        struct file_ra_state *ra = &file->f_ra;
        struct address_space *mapping = file->f_mapping;
        DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
-       struct file *fpin = NULL;
        unsigned long vm_flags = vmf->vma->vm_flags;
        unsigned int mmap_miss;
 

mm/filemap.c +12 -1

        struct file *file = vmf->vma->vm_file;
        struct file_ra_state *ra = &file->f_ra;
        DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff);
-       struct file *fpin = NULL;
        unsigned int mmap_miss;
 
        /* If we don't want any read-ahead, don't bother */

@@ -3287,6 +3288,35 @@ vm_fault_t filemap_fault(struct vm_fault *vmf)
        if (unlikely(index >= max_idx))
                return VM_FAULT_SIGBUS;
 
+       /*
+        * If we have pre-content watchers then we need to generate events on
+        * page fault so that we can populate any data before the fault.
+        *
+        * We only do this on the first pass through, otherwise the populating
+        * application could potentially deadlock on the mmap lock if it tries
+        * to populate it with mmap.
+        */
+       if (fault_flag_allow_retry_first(vmf->flags) &&
+           fsnotify_file_has_content_watches(file)) {
+               int mask = (vmf->flags & FAULT_FLAG_WRITE) ? MAY_WRITE : MAY_READ;
+               loff_t pos = vmf->pgoff << PAGE_SHIFT;

Quoted 19 lines · Josef Bacik

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on page fault depending on the faulting method.

This pre-content event is meant to be used by hierarchical storage managers that want to fill in the file content on first read access.

Signed-off-by Josef Bacik <[email protected]>

---
 fs/notify/fsnotify.c             | 13 +++++++++
 include/linux/fsnotify_backend.h | 14 +++++++++
 mm/filemap.c                     | 50 ++++++++++++++++++++++++++++----
 3 files changed, 71 insertions(+), 6 deletions(-)

fs/notify/fsnotify.c +0 -0

index 1ca4a8da7f29..435232d46b4f 100644

fs/notify/fsnotify.c void __fsnotify_vfsmount_delete(struct vfsmount *mnt) · line 31

Quoted 4 lines

fs/notify/fsnotify.c +1 -0

@@ -28,6 +28,19 @@ void __fsnotify_vfsmount_delete(struct vfsmount *mnt)
        fsnotify_clear_marks_by_mount(mnt);
 }
 
+bool fsnotify_file_has_content_watches(struct file *file)

nit: has_pre_content_watches...

fs/notify/fsnotify.c void __fsnotify_vfsmount_delete(struct vfsmount *mnt) · line 41

Quoted 10 lines

fs/notify/fsnotify.c +10 -0

+{
+       struct inode *inode = file_inode(file);
+       struct super_block *sb = inode->i_sb;
+       struct mount *mnt = real_mount(file->f_path.mnt);
+       u32 mask = inode->i_fsnotify_mask;
+
+       mask |= mnt->mnt_fsnotify_mask;
+       mask |= sb->s_fsnotify_mask;
+
+       return !!(mask & FSNOTIFY_PRE_CONTENT_EVENTS);

This can use the fsnotify_object_watched() helper, and it will need the READ_ONCE() that are just being added to avoid data races.

Quoted 66 lines

fs/notify/fsnotify.c +2 -0

+}
+
 /**
  * fsnotify_unmount_inodes - an sb is unmounting.  handle any watched inodes.
  * @sb: superblock being unmounted.

include/linux/fsnotify_backend.h +14 -0

index 36c3d18cc40a..6983fbf096b8 100644
@@ -900,6 +900,15 @@ static inline void fsnotify_init_event(struct fsnotify_event *event)
        INIT_LIST_HEAD(&event->list);
 }
 
+#ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
+bool fsnotify_file_has_content_watches(struct file *file);
+#else
+static inline bool fsnotify_file_has_content_watches(struct file *file)
+{
+       return false;
+}
+#endif /* CONFIG_FANOTIFY_ACCESS_PERMISSIONS */
+
 #else
 
 static inline int fsnotify(__u32 mask, const void *data, int data_type,

@@ -938,6 +947,11 @@ static inline u32 fsnotify_get_cookie(void)
 static inline void fsnotify_unmount_inodes(struct super_block *sb)
 {}
 
+static inline bool fsnotify_file_has_content_watches(struct file *file)
+{
+       return false;
+}
+
 #endif /* CONFIG_FSNOTIFY */
 
 #endif /* __KERNEL __ */

mm/filemap.c +3 -2

index ca8c8d889eef..cc9d7885bbe3 100644
@@ -46,6 +46,7 @@
 #include <linux/pipe_fs_i.h>
 #include <linux/splice.h>
 #include <linux/rcupdate_wait.h>
+#include <linux/fsnotify.h>
 #include <asm/pgalloc.h>
 #include <asm/tlbflush.h>
 #include "internal.h"

@@ -3112,13 +3113,13 @@ static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
  * that.  If we didn't pin a file then we return NULL.  The file that is
  * returned needs to be fput()'ed when we're done with it.
  */
-static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
+static struct file *do_sync_mmap_readahead(struct vm_fault *vmf,
+                                          struct file *fpin)
 {
        struct file *file = vmf->vma->vm_file;
        struct file_ra_state *ra = &file->f_ra;
        struct address_space *mapping = file->f_mapping;
        DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
-       struct file *fpin = NULL;
        unsigned long vm_flags = vmf->vma->vm_flags;
        unsigned int mmap_miss;
 

mm/filemap.c static struct file *do_sync_mmap_readahead(struct vm_fault *vmf) · line 3188

Quoted 7 lines

mm/filemap.c +2 -1

@@ -3182,12 +3183,12 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
  * was pinned if we have to drop the mmap_lock in order to do IO.
  */
 static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
-                                           struct folio *folio)
+                                           struct folio *folio,
+                                           struct file *fpin)
 {

If I am reading correctly, iomap (i.e. xfs) write shared memory fault does not reach this code?

Do we care about writable shared memory faults use case for HSM? It does not sound very relevant to HSM, but we cannot just ignore it..

Quoted 118 lines · Amir Goldstein

Quoted 19 lines · Josef Bacik

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on page fault depending on the faulting method.

This pre-content event is meant to be used by hierarchical storage managers that want to fill in the file content on first read access.

Signed-off-by Josef Bacik <[email protected]>

---
 fs/notify/fsnotify.c             | 13 +++++++++
 include/linux/fsnotify_backend.h | 14 +++++++++
 mm/filemap.c                     | 50 ++++++++++++++++++++++++++++----
 3 files changed, 71 insertions(+), 6 deletions(-)

fs/notify/fsnotify.c +0 -0

index 1ca4a8da7f29..435232d46b4f 100644

fs/notify/fsnotify.c void __fsnotify_vfsmount_delete(struct vfsmount *mnt) · line 31

Quoted 4 lines

fs/notify/fsnotify.c +1 -0

@@ -28,6 +28,19 @@ void __fsnotify_vfsmount_delete(struct vfsmount *mnt)
        fsnotify_clear_marks_by_mount(mnt);
 }
 
+bool fsnotify_file_has_content_watches(struct file *file)

nit: has_pre_content_watches...

fs/notify/fsnotify.c void __fsnotify_vfsmount_delete(struct vfsmount *mnt) · line 41

Quoted 10 lines

fs/notify/fsnotify.c +10 -0

+{
+       struct inode *inode = file_inode(file);
+       struct super_block *sb = inode->i_sb;
+       struct mount *mnt = real_mount(file->f_path.mnt);
+       u32 mask = inode->i_fsnotify_mask;
+
+       mask |= mnt->mnt_fsnotify_mask;
+       mask |= sb->s_fsnotify_mask;
+
+       return !!(mask & FSNOTIFY_PRE_CONTENT_EVENTS);

This can use the fsnotify_object_watched() helper, and it will need the READ_ONCE() that are just being added to avoid data races.

Quoted 66 lines

fs/notify/fsnotify.c +2 -0

+}
+
 /**
  * fsnotify_unmount_inodes - an sb is unmounting.  handle any watched inodes.
  * @sb: superblock being unmounted.

include/linux/fsnotify_backend.h +14 -0

index 36c3d18cc40a..6983fbf096b8 100644
@@ -900,6 +900,15 @@ static inline void fsnotify_init_event(struct fsnotify_event *event)
        INIT_LIST_HEAD(&event->list);
 }
 
+#ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
+bool fsnotify_file_has_content_watches(struct file *file);
+#else
+static inline bool fsnotify_file_has_content_watches(struct file *file)
+{
+       return false;
+}
+#endif /* CONFIG_FANOTIFY_ACCESS_PERMISSIONS */
+
 #else
 
 static inline int fsnotify(__u32 mask, const void *data, int data_type,

@@ -938,6 +947,11 @@ static inline u32 fsnotify_get_cookie(void)
 static inline void fsnotify_unmount_inodes(struct super_block *sb)
 {}
 
+static inline bool fsnotify_file_has_content_watches(struct file *file)
+{
+       return false;
+}
+
 #endif /* CONFIG_FSNOTIFY */
 
 #endif /* __KERNEL __ */

mm/filemap.c +3 -2

index ca8c8d889eef..cc9d7885bbe3 100644
@@ -46,6 +46,7 @@
 #include <linux/pipe_fs_i.h>
 #include <linux/splice.h>
 #include <linux/rcupdate_wait.h>
+#include <linux/fsnotify.h>
 #include <asm/pgalloc.h>
 #include <asm/tlbflush.h>
 #include "internal.h"

@@ -3112,13 +3113,13 @@ static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio,
  * that.  If we didn't pin a file then we return NULL.  The file that is
  * returned needs to be fput()'ed when we're done with it.
  */
-static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
+static struct file *do_sync_mmap_readahead(struct vm_fault *vmf,
+                                          struct file *fpin)
 {
        struct file *file = vmf->vma->vm_file;
        struct file_ra_state *ra = &file->f_ra;
        struct address_space *mapping = file->f_mapping;
        DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff);
-       struct file *fpin = NULL;
        unsigned long vm_flags = vmf->vma->vm_flags;
        unsigned int mmap_miss;
 

mm/filemap.c static struct file *do_sync_mmap_readahead(struct vm_fault *vmf) · line 3188

Quoted 7 lines

mm/filemap.c +2 -1

@@ -3182,12 +3183,12 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
  * was pinned if we have to drop the mmap_lock in order to do IO.
  */
 static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
-                                           struct folio *folio)
+                                           struct folio *folio,
+                                           struct file *fpin)
 {

If I am reading correctly, iomap (i.e. xfs) write shared memory fault does not reach this code?

Quoted 3 lines

Do we care about writable shared memory faults use case for HSM? It does not sound very relevant to HSM, but we cannot just ignore it..

Sorry I realized I went off to try and solve this problem and never responded to you. I'm addressing the other comments, but this one is a little tricky.

We're kind of stuck between a rock and a hard place with this. I had originally put this before the ->fault() callback, but purposefully moved it into filemap_fault() because I want to be able to drop the mmap lock while we're waiting for a response from the HSM.

The reason to do this is because there are things that take the mmap lock for simple things outside of the process, like /proc/$PID/smaps and other related things, and this can cause high priority tasks to block behind possibly low priority IO, creating a priority inversion.

Now, I'm not sure how widespread of a problem this is anymore, I know there's been work done to the kernel and tools to avoid this style of problem. I'm ok with a "try it and see" approach, but I don't love that.

However I think putting fsnotify hooks into XFS itself for this particular path is a good choice either.

What do you think? Just move it to before ->fault(), leave the mmap lock in place, and be done with it?

Quoted 15 lines · Josef Bacik

Quoted 6 lines

If I am reading correctly, iomap (i.e. xfs) write shared memory fault does not reach this code?

Do we care about writable shared memory faults use case for HSM? It does not sound very relevant to HSM, but we cannot just ignore it..

Sorry I realized I went off to try and solve this problem and never responded to you. I'm addressing the other comments, but this one is a little tricky.

We're kind of stuck between a rock and a hard place with this. I had originally put this before the ->fault() callback, but purposefully moved it into filemap_fault() because I want to be able to drop the mmap lock while we're waiting for a response from the HSM.

Quoted 9 lines

The reason to do this is because there are things that take the mmap lock for simple things outside of the process, like /proc/$PID/smaps and other related things, and this can cause high priority tasks to block behind possibly low priority IO, creating a priority inversion.

Now, I'm not sure how widespread of a problem this is anymore, I know there's been work done to the kernel and tools to avoid this style of problem. I'm ok with a "try it and see" approach, but I don't love that.

I defer this question to Jan.

Quoted 2 lines

However I think putting fsnotify hooks into XFS itself for this particular path is a good choice either.

I think you meant "not a good choice" and I agree -
it is not only xfs, but could be any fs that will be converted to iomap Other fs have ->fault != filemap_fault, even if they do end up calling filemap_fault, IOW, there is no API guarantee that they will.

Quoted 2 lines

What do you think? Just move it to before ->fault(), leave the mmap lock in place, and be done with it?

If Jan blesses the hook called with mmap lock, then yeh, putting the hook in the most generic "vfs" code would be the best choice for maintenance.

Quoted 36 lines · Amir Goldstein

Quoted 15 lines · Josef Bacik

Quoted 6 lines

If I am reading correctly, iomap (i.e. xfs) write shared memory fault does not reach this code?

Do we care about writable shared memory faults use case for HSM? It does not sound very relevant to HSM, but we cannot just ignore it..

Sorry I realized I went off to try and solve this problem and never responded to you. I'm addressing the other comments, but this one is a little tricky.

We're kind of stuck between a rock and a hard place with this. I had originally put this before the ->fault() callback, but purposefully moved it into filemap_fault() because I want to be able to drop the mmap lock while we're waiting for a response from the HSM.

Quoted 9 lines

The reason to do this is because there are things that take the mmap lock for simple things outside of the process, like /proc/$PID/smaps and other related things, and this can cause high priority tasks to block behind possibly low priority IO, creating a priority inversion.

Now, I'm not sure how widespread of a problem this is anymore, I know there's been work done to the kernel and tools to avoid this style of problem. I'm ok with a "try it and see" approach, but I don't love that.

I defer this question to Jan.

Quoted 2 lines

However I think putting fsnotify hooks into XFS itself for this particular path is a good choice either.

I think you meant "not a good choice" and I agree -
it is not only xfs, but could be any fs that will be converted to iomap Other fs have ->fault != filemap_fault, even if they do end up calling filemap_fault, IOW, there is no API guarantee that they will.

Quoted 6 lines

Quoted 2 lines

What do you think? Just move it to before ->fault(), leave the mmap lock in place, and be done with it?

If Jan blesses the hook called with mmap lock, then yeh, putting the hook in the most generic "vfs" code would be the best choice for maintenance.

Well, I agree with Josef's comment about a rock and a hard place. For once, regardless whether the hook will happen from before ->fault or from inside the ->fault handler there will be fault callers where we cannot drop mmap_lock (not all archs support dropping mmap_lock inside a fault AFAIR - but a quick grep seems to show that these days maybe they do, also some callers - most notably through GUP - don't allow dropping of mmap_lock inside fault). So we have to have a way to handle a fault without FAULT_FLAG_ALLOW_RETRY flag.

Now of course waiting for userspace reply to fanotify event with mmap_lock held is ... dangerous. For example consider application with two threads:

T1 T2
page fault on inode I write to inode I
lock mm->mmap_lock inode_lock(I)
send fanotify event ...
fault_in_iov_iter_readable()
lock mm->mmap_lock -> blocks
behind T1

now the HSM handler needs to fill in contents of inode I requested by the page fault:

inode_lock(I) -> deadlock

So conceptually I think the flow could look like (in __do_fault):

	if (!(vmf->flags & FAULT_FLAG_TRIED) &&
	    fsnotify_may_send_pre_content_event()) {
		if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
			return VM_FAULT_RETRY;
		fpin = maybe_unlock_mmap_for_io(vmf, NULL);
		if (!fpin)
			return ???VM_FAULT_SIGSEGV???;
		err = fsnotify_fault(...);
		if (err)
			return VM_FAULT_SIGBUS | VM_FAULT_RETRY;
		/*
		 * We are fine with proceeding with the fault. Retry the fault
		 * to let the filesystem handle it.
		 */
		return VM_FAULT_RETRY;
	}

The downside is that if we enter the page fault without ability to drop mmap_lock on a file needing HSM handling, we get SIGSEGV. I'm not sure it matters in practice because these are not that common paths e.g. stuff like putting a breakpoint / uprobe on a file but maybe there are some surprises.

Quoted 88 lines · Jan Kara

Quoted 36 lines · Amir Goldstein

Quoted 15 lines · Josef Bacik

Quoted 6 lines

If I am reading correctly, iomap (i.e. xfs) write shared memory fault does not reach this code?

Do we care about writable shared memory faults use case for HSM? It does not sound very relevant to HSM, but we cannot just ignore it..

Sorry I realized I went off to try and solve this problem and never responded to you. I'm addressing the other comments, but this one is a little tricky.

We're kind of stuck between a rock and a hard place with this. I had originally put this before the ->fault() callback, but purposefully moved it into filemap_fault() because I want to be able to drop the mmap lock while we're waiting for a response from the HSM.

Quoted 9 lines

The reason to do this is because there are things that take the mmap lock for simple things outside of the process, like /proc/$PID/smaps and other related things, and this can cause high priority tasks to block behind possibly low priority IO, creating a priority inversion.

Now, I'm not sure how widespread of a problem this is anymore, I know there's been work done to the kernel and tools to avoid this style of problem. I'm ok with a "try it and see" approach, but I don't love that.

I defer this question to Jan.

Quoted 2 lines

However I think putting fsnotify hooks into XFS itself for this particular path is a good choice either.

I think you meant "not a good choice" and I agree -
it is not only xfs, but could be any fs that will be converted to iomap Other fs have ->fault != filemap_fault, even if they do end up calling filemap_fault, IOW, there is no API guarantee that they will.

Quoted 6 lines

Quoted 2 lines

What do you think? Just move it to before ->fault(), leave the mmap lock in place, and be done with it?

If Jan blesses the hook called with mmap lock, then yeh, putting the hook in the most generic "vfs" code would be the best choice for maintenance.

Well, I agree with Josef's comment about a rock and a hard place. For once, regardless whether the hook will happen from before ->fault or from inside the ->fault handler there will be fault callers where we cannot drop mmap_lock (not all archs support dropping mmap_lock inside a fault AFAIR - but a quick grep seems to show that these days maybe they do, also some callers - most notably through GUP - don't allow dropping of mmap_lock inside fault). So we have to have a way to handle a fault without FAULT_FLAG_ALLOW_RETRY flag.

Now of course waiting for userspace reply to fanotify event with mmap_lock held is ... dangerous. For example consider application with two threads:

T1 T2
page fault on inode I write to inode I
lock mm->mmap_lock inode_lock(I)
send fanotify event ...
fault_in_iov_iter_readable()
lock mm->mmap_lock -> blocks
behind T1

now the HSM handler needs to fill in contents of inode I requested by the page fault:

inode_lock(I) -> deadlock

So conceptually I think the flow could look like (in __do_fault):

	if (!(vmf->flags & FAULT_FLAG_TRIED) &&
	    fsnotify_may_send_pre_content_event()) {
		if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
			return VM_FAULT_RETRY;
		fpin = maybe_unlock_mmap_for_io(vmf, NULL);
		if (!fpin)
			return ???VM_FAULT_SIGSEGV???;
		err = fsnotify_fault(...);
		if (err)
			return VM_FAULT_SIGBUS | VM_FAULT_RETRY;
		/*
		 * We are fine with proceeding with the fault. Retry the fault
		 * to let the filesystem handle it.
		 */
		return VM_FAULT_RETRY;
	}

Quoted 5 lines

The downside is that if we enter the page fault without ability to drop mmap_lock on a file needing HSM handling, we get SIGSEGV. I'm not sure it matters in practice because these are not that common paths e.g. stuff like putting a breakpoint / uprobe on a file but maybe there are some surprises.

The only thing I don't like about this is that now the fault handler loses the ability to drop the mmap sem. I think in practice this doesn't really matter, because we're trying to avoid doing IO while under the mmap sem, and presumably this will have instantiated the pages to avoid the IO.

However if you use reflink this wouldn't be the case, and now we've re-introduced the priority inversion possiblity.

I'm leaning more towards just putting the fsnotify hook in the xfs code for the write case, and anybody else who implements their own ->fault without calling the generic one will just have to do the same.

That being said I'm not going to die on any particular hill here, if you still want to do the above before the ->fault() handler then I'll update my code to do that and we can move on. Thanks,

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on page fault depending on the faulting method.

This pre-content event is meant to be used by hierarchical storage managers that want to fill in the file content on first read access.

Signed-off-by Josef Bacik <[email protected]>

mm/filemap.c +0 -0

@@ -3287,6 +3288,35 @@ vm_fault_t filemap_fault(struct vm_fault *vmf)
 	if (unlikely(index >= max_idx))
 		return VM_FAULT_SIGBUS;
 

mm/filemap.c +19 -0

+		int mask = (vmf->flags & FAULT_FLAG_WRITE) ? MAY_WRITE : MAY_READ;
+		loff_t pos = vmf->pgoff << PAGE_SHIFT;
+
+		fpin = maybe_unlock_mmap_for_io(vmf, fpin);
+
+		/*
+		 * We can only emit the event if we did actually release the
+		 * mmap lock.
+		 */
+		if (fpin) {
+			error = fsnotify_file_area_perm(fpin, mask, &pos,
+							PAGE_SIZE);
+			if (error) {
+				fput(fpin);
+				return VM_FAULT_ERROR;
+			}
+		}
+	}
+
 	/*
 	 * Do we have something in the page cache already?
 	 */

Quoted 7 lines · Josef Bacik

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on page fault depending on the faulting method.

This pre-content event is meant to be used by hierarchical storage managers that want to fill in the file content on first read access.

Signed-off-by Josef Bacik <[email protected]>

...

Quoted 4 lines

mm/filemap.c +0 -0

@@ -3287,6 +3288,35 @@ vm_fault_t filemap_fault(struct vm_fault *vmf)
 	if (unlikely(index >= max_idx))
 		return VM_FAULT_SIGBUS;
 

mm/filemap.c vm_fault_t filemap_fault(struct vm_fault *vmf) · line 3300

Quoted 10 lines

mm/filemap.c +10 -0

+	/*
+	 * If we have pre-content watchers then we need to generate events on
+	 * page fault so that we can populate any data before the fault.
+	 *
+	 * We only do this on the first pass through, otherwise the populating
+	 * application could potentially deadlock on the mmap lock if it tries
+	 * to populate it with mmap.
+	 */
+	if (fault_flag_allow_retry_first(vmf->flags) &&
+	    fsnotify_file_has_content_watches(file)) {

I'm somewhat nervous that if ALLOW_RETRY isn't set, we'd silently jump into readpage code without ever sending pre-content event and thus we'd possibly expose invalid content to userspace? I think we should fail the fault if fsnotify_file_has_content_watches(file) && !(vmf->flags & FAULT_FLAG_ALLOW_RETRY).

Quoted 23 lines

mm/filemap.c +19 -0

+		int mask = (vmf->flags & FAULT_FLAG_WRITE) ? MAY_WRITE : MAY_READ;
+		loff_t pos = vmf->pgoff << PAGE_SHIFT;
+
+		fpin = maybe_unlock_mmap_for_io(vmf, fpin);
+
+		/*
+		 * We can only emit the event if we did actually release the
+		 * mmap lock.
+		 */
+		if (fpin) {
+			error = fsnotify_file_area_perm(fpin, mask, &pos,
+							PAGE_SIZE);
+			if (error) {
+				fput(fpin);
+				return VM_FAULT_ERROR;
+			}
+		}
+	}
+
 	/*
 	 * Do we have something in the page cache already?
 	 */

...

Quoted 4 lines

mm/filemap.c +0 -0

@@ -3612,6 +3643,13 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf,
 	unsigned long rss = 0;
 	unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type;
 

mm/filemap.c vm_fault_t filemap_map_pages(struct vm_fault *vmf, · line 3652

Quoted 7 lines

mm/filemap.c +7 -0

+	/*
+	 * We are under RCU, we can't emit events here, we need to force a
+	 * normal fault to make sure the events get sent.
+	 */
+	if (fsnotify_file_has_content_watches(file))
+		return ret;
+

I don't think we need to do anything for filemap_map_pages(). The call just inserts page cache content into page tables and whatever is in the page cache and has folio_uptodate() set should be already valid file content, shouldn't it?

Quoted 13 lines · Jan Kara

Quoted 7 lines · Josef Bacik

FS_PRE_ACCESS or FS_PRE_MODIFY will be generated on page fault depending on the faulting method.

This pre-content event is meant to be used by hierarchical storage managers that want to fill in the file content on first read access.

Signed-off-by Josef Bacik <[email protected]>

...

Quoted 4 lines

mm/filemap.c +0 -0

@@ -3287,6 +3288,35 @@ vm_fault_t filemap_fault(struct vm_fault *vmf)
 	if (unlikely(index >= max_idx))
 		return VM_FAULT_SIGBUS;
 

Quoted 16 lines

mm/filemap.c vm_fault_t filemap_fault(struct vm_fault *vmf) · line 3300

Quoted 10 lines

mm/filemap.c +10 -0

+	/*
+	 * If we have pre-content watchers then we need to generate events on
+	 * page fault so that we can populate any data before the fault.
+	 *
+	 * We only do this on the first pass through, otherwise the populating
+	 * application could potentially deadlock on the mmap lock if it tries
+	 * to populate it with mmap.
+	 */
+	if (fault_flag_allow_retry_first(vmf->flags) &&
+	    fsnotify_file_has_content_watches(file)) {

I'm somewhat nervous that if ALLOW_RETRY isn't set, we'd silently jump into readpage code without ever sending pre-content event and thus we'd possibly expose invalid content to userspace? I think we should fail the fault if fsnotify_file_has_content_watches(file) && !(vmf->flags & FAULT_FLAG_ALLOW_RETRY).

I was worried about this too but it seems to not be the case that we'll not ever have ALLOW_RETRY. That being said I'm fine turning this into a sigbus.

Quoted 27 lines

Quoted 22 lines

mm/filemap.c +19 -0

+		int mask = (vmf->flags & FAULT_FLAG_WRITE) ? MAY_WRITE : MAY_READ;
+		loff_t pos = vmf->pgoff << PAGE_SHIFT;
+
+		fpin = maybe_unlock_mmap_for_io(vmf, fpin);
+
+		/*
+		 * We can only emit the event if we did actually release the
+		 * mmap lock.
+		 */
+		if (fpin) {
+			error = fsnotify_file_area_perm(fpin, mask, &pos,
+							PAGE_SIZE);
+			if (error) {
+				fput(fpin);
+				return VM_FAULT_ERROR;
+			}
+		}
+	}
+
 	/*
 	 * Do we have something in the page cache already?
 	 */

...

Quoted 4 lines

mm/filemap.c +0 -0

@@ -3612,6 +3643,13 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf,
 	unsigned long rss = 0;
 	unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type;
 

Quoted 12 lines

mm/filemap.c vm_fault_t filemap_map_pages(struct vm_fault *vmf, · line 3652

Quoted 7 lines

mm/filemap.c +7 -0

+	/*
+	 * We are under RCU, we can't emit events here, we need to force a
+	 * normal fault to make sure the events get sent.
+	 */
+	if (fsnotify_file_has_content_watches(file))
+		return ret;
+

I don't think we need to do anything for filemap_map_pages(). The call just inserts page cache content into page tables and whatever is in the page cache and has folio_uptodate() set should be already valid file content, shouldn't it?