Message-ID: <aWqs1bRhkuKpyIUo@CMGLRV3>
Date: Fri, 16 Jan 2026 15:25:41 -0600
From: Frederick Lawler <fred@...udflare.com>
To: Mimi Zohar <zohar@...ux.ibm.com>,
	Roberto Sassu <roberto.sassu@...wei.com>,
	Dmitry Kasatkin <dmitry.kasatkin@...il.com>,
	Eric Snowberg <eric.snowberg@...cle.com>,
	Paul Moore <paul@...l-moore.com>, James Morris <jmorris@...ei.org>,
	"Serge E. Hallyn" <serge@...lyn.com>,
	"Darrick J. Wong" <djwong@...nel.org>,
	Christian Brauner <brauner@...nel.org>,
	Josef Bacik <josef@...icpanda.com>,
	Jeff Layton <jlayton@...nel.org>
Cc: linux-kernel@...r.kernel.org, linux-integrity@...r.kernel.org,
	linux-security-module@...r.kernel.org, kernel-team@...udflare.com
Subject: Re: [PATCH] ima: Detect changes to files via kstat changes rather
 than i_version

On Mon, Jan 12, 2026 at 04:32:23PM -0600, Frederick Lawler wrote:
> Commit 1cf7e834a6fb ("xfs: switch to multigrain timestamps")
> introduced a means to track change detection for an inode
> via ctime updates, opposed to setting kstat.change_cookie when
> calling into xfs_vn_getattr().
> 
> This introduced a regression because IMA caches kstat.change_cookie
> to compare against an inode's i_version directly in
> integrity_inode_attrs_changed(), and thus could be out of date
> depending on how file systems increment i_version.
> 
> To address this, require integrity_inode_attrs_changed() to query
> vfs_getattr_nosec() to compare the cached version against
> kstat.change_cookie directly. This ensures that when updates occur,
> we're accessing the same changed inode version on changes, and fallback
> to compare against an artificial version generated from kstat.ctime
> via integrity_ctime_guard() when there's no detected change
> to the kstat.change_cookie.
> 
> This ensures that in the absence of i_version support for file systems,
> and in the absence of a kstat.change_cookie update, we ultimately have a
> unique-enough version to compare against.
> 
> The exact implementation for integrity_ctime_guard() is to ensure that
> if tv_sec or tv_nsec are zero, there's some value to store back into
> struct integrity_inode_attributes.version. This also avoids the need to
> add additional storage and comparisons.
> 
> Lastly, because EVM still relies on querying and caching a backing inode's
> i_version, the integrity_inode_attrs_changed() falls back to the
> original inode.i_version != cached comparison. This maintains the
> invariant that a re-evaluation in unknown change detection circumstances
> is required.
> 
> Link: https://lore.kernel.org/all/aTspr4_h9IU4EyrR@CMGLRV3
> Suggested-by: Jeff Layton <jlayton@...nel.org>
> Signed-off-by: Frederick Lawler <fred@...udflare.com>
> ---
> We uncovered a case in kernels >= 6.13 where XFS is no longer updating
> struct kstat.change_cookie on i_op getattr() access calls. Instead, XFS is
> using multigrain ctime (as well as other file systems) for
> change detection in commit 1cf7e834a6fb ("xfs: switch to
> multigrain timestamps").
> 
> Because file systems may implement i_version as they see fit, IMA
> caching may be behind as well as file systems that don't support/export
> i_version. Thus we're proposing to compare against the kstat.change_cookie
> directly to the cached version, and fall back to a ctime guard when
> that's not updated.
> 
> EVM is largely left alone since there's no trivial way to query a file
> directly in the LSM call paths to obtain kstat.change_cookie &
> kstat.ctime to cache. Thus retains accessing i_version directly.
> 
> Regression tests will be added to the Linux Test Project instead of
> selftest to help catch future file system changes that may impact
> future evaluation of IMA.
> 
> I'd like this to be backported to at least 6.18 if possible.
> 
> Below is a simplified test that demonstrates the issue:
> 
> _fragment.config_
> CONFIG_XFS_FS=y
> CONFIG_OVERLAY_FS=y
> CONFIG_IMA=y
> CONFIG_IMA_WRITE_POLICY=y
> CONFIG_IMA_READ_POLICY=y
> 
> _./test.sh_
> 
> IMA_POLICY="/sys/kernel/security/ima/policy"
> TEST_BIN="/bin/date"
> MNT_BASE="/tmp/ima_test_root"
> 
> mkdir -p "$MNT_BASE"
> mount -t tmpfs tmpfs "$MNT_BASE"
> mkdir -p "$MNT_BASE"/{xfs_disk,upper,work,ovl}
> 
> dd if=/dev/zero of="$MNT_BASE/xfs.img" bs=1M count=300
> mkfs.xfs -q "$MNT_BASE/xfs.img"
> mount "$MNT_BASE/xfs.img" "$MNT_BASE/xfs_disk"
> cp "$TEST_BIN" "$MNT_BASE/xfs_disk/test_prog"
> 
> mount -t overlay overlay -o \
> "lowerdir=$MNT_BASE/xfs_disk,upperdir=$MNT_BASE/upper,workdir=$MNT_BASE/work" \
> "$MNT_BASE/ovl"
> 
> echo "audit func=BPRM_CHECK uid=$(id -u nobody)" > "$IMA_POLICY"
> 
> target_prog="$MNT_BASE/ovl/test_prog"
> setpriv --reuid nobody "$target_prog"
> setpriv --reuid nobody "$target_prog"
> setpriv --reuid nobody "$target_prog"
> 
> audit_count=$(dmesg | grep -c "file=\"$target_prog\"")
> 
> if [[ "$audit_count" -eq 1 ]]; then
>         echo "PASS: Found exactly 1 audit event."
> else
>         echo "FAIL: Expected 1 audit event, but found $audit_count."
>         exit 1
> fi
> ---
> Changes since RFC:
> - Remove calls to I_IS_VERSION()
> - Function documentation/comments
> - Abide IMA/EVM change detection fallback invariants
> - Combined ctime guard into version for attributes struct
> - Link to RFC: https://lore.kernel.org/r/20251229-xfs-ima-fixup-v1-1-6a717c939f7c@cloudflare.com
> ---
>  include/linux/integrity.h         | 42 +++++++++++++++++++++++++++++++++++----
>  security/integrity/evm/evm_main.c |  5 ++---
>  security/integrity/ima/ima_api.c  | 11 +++++++---
>  security/integrity/ima/ima_main.c | 15 +++++---------
>  4 files changed, 53 insertions(+), 20 deletions(-)
> 
> diff --git a/include/linux/integrity.h b/include/linux/integrity.h
> index f5842372359be5341b6870a43b92e695e8fc78af..5eca8aa2769f9238c68bb40885ecc46910524f11 100644
> --- a/include/linux/integrity.h
> +++ b/include/linux/integrity.h
> @@ -9,6 +9,7 @@
>  
>  #include <linux/fs.h>
>  #include <linux/iversion.h>
> +#include <linux/kernel.h>
>  
>  enum integrity_status {
>  	INTEGRITY_PASS = 0,
> @@ -36,6 +37,14 @@ struct integrity_inode_attributes {
>  	dev_t dev;
>  };
>  
> +/*
> + * Wrapper to generate an artificial version for a file.
> + */
> +static inline u64 integrity_ctime_guard(struct kstat stat)
> +{
> +	return stat.ctime.tv_sec ^ stat.ctime.tv_nsec;
> +}
> +
>  /*
>   * On stacked filesystems the i_version alone is not enough to detect file data
>   * or metadata change. Additional metadata is required.
> @@ -51,14 +60,39 @@ integrity_inode_attrs_store(struct integrity_inode_attributes *attrs,
>  
>  /*
>   * On stacked filesystems detect whether the inode or its content has changed.
> + *
> + * Must be called in process context.
>   */
>  static inline bool
>  integrity_inode_attrs_changed(const struct integrity_inode_attributes *attrs,
> -			      const struct inode *inode)
> +			      struct file *file, struct inode *inode)
>  {
> -	return (inode->i_sb->s_dev != attrs->dev ||
> -		inode->i_ino != attrs->ino ||
> -		!inode_eq_iversion(inode, attrs->version));
> +	struct kstat stat;
> +
> +	might_sleep();
> +
> +	if (inode->i_sb->s_dev != attrs->dev || inode->i_ino != attrs->ino)
> +		return true;
> +
> +	/*
> +	 * EVM currently relies on backing inode i_version. While IS_I_VERSION
> +	 * is not a good indicator of i_version support, this still retains
> +	 * the logic such that a re-evaluation should still occur for EVM, and
> +	 * only for IMA if vfs_getattr_nosec() fails.
> +	 */
> +	if (!file || vfs_getattr_nosec(&file->f_path, &stat,
> +				       STATX_CHANGE_COOKIE | STATX_CTIME,
> +				       AT_STATX_SYNC_AS_STAT))
> +		return !IS_I_VERSION(inode) ||
> +			!inode_eq_iversion(inode, attrs->version);
> +
> +	if (stat.result_mask & STATX_CHANGE_COOKIE)
> +		return stat.change_cookie != attrs->version;
> +
> +	if (stat.result_mask & STATX_CTIME)
> +		return integrity_ctime_guard(stat) != attrs->version;
> +
> +	return true;
>  }
>  
>  
> diff --git a/security/integrity/evm/evm_main.c b/security/integrity/evm/evm_main.c
> index 73d500a375cb37a54f295b0e1e93fd6e5d9ecddc..6a4e0e246005246d5700b1db590c1759242b9cb6 100644
> --- a/security/integrity/evm/evm_main.c
> +++ b/security/integrity/evm/evm_main.c
> @@ -752,9 +752,8 @@ bool evm_metadata_changed(struct inode *inode, struct inode *metadata_inode)
>  	bool ret = false;
>  
>  	if (iint) {
> -		ret = (!IS_I_VERSION(metadata_inode) ||
> -		       integrity_inode_attrs_changed(&iint->metadata_inode,
> -						     metadata_inode));
> +		ret = integrity_inode_attrs_changed(&iint->metadata_inode,
> +						    NULL, metadata_inode);
>  		if (ret)
>  			iint->evm_status = INTEGRITY_UNKNOWN;
>  	}
> diff --git a/security/integrity/ima/ima_api.c b/security/integrity/ima/ima_api.c
> index c35ea613c9f8d404ba4886e3b736c3bab29d1668..8096986f3689781d3cdf6595f330033782f9cc45 100644
> --- a/security/integrity/ima/ima_api.c
> +++ b/security/integrity/ima/ima_api.c
> @@ -272,10 +272,15 @@ int ima_collect_measurement(struct ima_iint_cache *iint, struct file *file,
>  	 * to an initial measurement/appraisal/audit, but was modified to
>  	 * assume the file changed.
>  	 */
> -	result = vfs_getattr_nosec(&file->f_path, &stat, STATX_CHANGE_COOKIE,
> +	result = vfs_getattr_nosec(&file->f_path, &stat,
> +				   STATX_CHANGE_COOKIE | STATX_CTIME,
>  				   AT_STATX_SYNC_AS_STAT);
> -	if (!result && (stat.result_mask & STATX_CHANGE_COOKIE))
> -		i_version = stat.change_cookie;
> +	if (!result) {
> +		if (stat.result_mask & STATX_CHANGE_COOKIE)
> +			i_version = stat.change_cookie;
> +		else if (stat.result_mask & STATX_CTIME)
> +			i_version = integrity_ctime_guard(stat);
> +	}
>  	hash.hdr.algo = algo;
>  	hash.hdr.length = hash_digest_size[algo];
>  
> diff --git a/security/integrity/ima/ima_main.c b/security/integrity/ima/ima_main.c
> index 5770cf691912aa912fc65280c59f5baac35dd725..3a4c32e254f925bba85cb91b63744ac142b3b049 100644
> --- a/security/integrity/ima/ima_main.c
> +++ b/security/integrity/ima/ima_main.c
> @@ -22,6 +22,7 @@
>  #include <linux/mount.h>
>  #include <linux/mman.h>
>  #include <linux/slab.h>
> +#include <linux/stat.h>
>  #include <linux/xattr.h>
>  #include <linux/ima.h>
>  #include <linux/fs.h>
> @@ -191,18 +192,13 @@ static void ima_check_last_writer(struct ima_iint_cache *iint,
>  
>  	mutex_lock(&iint->mutex);
>  	if (atomic_read(&inode->i_writecount) == 1) {
> -		struct kstat stat;
> -
>  		clear_bit(IMA_EMITTED_OPENWRITERS, &iint->atomic_flags);
>  
>  		update = test_and_clear_bit(IMA_UPDATE_XATTR,
>  					    &iint->atomic_flags);
>  		if ((iint->flags & IMA_NEW_FILE) ||
> -		    vfs_getattr_nosec(&file->f_path, &stat,
> -				      STATX_CHANGE_COOKIE,
> -				      AT_STATX_SYNC_AS_STAT) ||
> -		    !(stat.result_mask & STATX_CHANGE_COOKIE) ||
> -		    stat.change_cookie != iint->real_inode.version) {
> +		    integrity_inode_attrs_changed(&iint->real_inode, file,
> +						  inode)) {

I'm working through my tests, and I don't think I can get away with this
change. The check for the inode->i_ino != attr->ino may result in a
re-evaluation because we're not updating the attr->ino while collecting
measurement on non-stacked file systems checks. Same for attr->dev not
updating.

I'll put this back in the next patch version, and still check ctime here
similar to the RFC version.

>  			iint->flags &= ~(IMA_DONE_MASK | IMA_NEW_FILE);
>  			iint->measured_pcrs = 0;
>  			if (update)
> @@ -328,9 +324,8 @@ static int process_measurement(struct file *file, const struct cred *cred,
>  	real_inode = d_real_inode(file_dentry(file));
>  	if (real_inode != inode &&
>  	    (action & IMA_DO_MASK) && (iint->flags & IMA_DONE_MASK)) {
> -		if (!IS_I_VERSION(real_inode) ||
> -		    integrity_inode_attrs_changed(&iint->real_inode,
> -						  real_inode)) {
> +		if (integrity_inode_attrs_changed(&iint->real_inode,
> +						  file, real_inode)) {
>  			iint->flags &= ~IMA_DONE_MASK;
>  			iint->measured_pcrs = 0;
>  		}
> 
> ---
> base-commit: 8f0b4cce4481fb22653697cced8d0d04027cb1e8
> change-id: 20251212-xfs-ima-fixup-931780a62c2c
> 
> Best regards,
> -- 
> Frederick Lawler <fred@...udflare.com>
> 

Powered by blists - more mailing lists