ROX-30296: track POSIX ACL changes via inode_set_acl LSM hook

Author: Stringy

fact-ebpf/src/bpf/bound_path.h

@@ -38,6 +38,29 @@ __always_inline static struct bound_path_t* _path_read(struct path* path, bound_
   return bound_path;
 }
 
+/**
+ * Read a filesystem-relative path from a bare dentry.
+ *
+ * This is for hooks that only provide a struct dentry without a
+ * struct path (e.g. inode_set_acl). The resulting path can be used
+ * for LPM trie matching and inode monitoring checks.
+ */
+__always_inline static struct bound_path_t* dentry_read(struct dentry* dentry) {
+  struct bound_path_t* bound_path = get_bound_path(BOUND_PATH_MAIN);
+  if (bound_path == NULL) {
+    return NULL;
+  }
+
+  bound_path->len = __d_path_from_dentry(dentry, bound_path->path, PATH_MAX);
+  if (bound_path->len <= 0) {
+    return NULL;
+  }
+
+  bound_path->len = PATH_LEN_CLAMP(bound_path->len);
+
+  return bound_path;
+}
+
 __always_inline static struct bound_path_t* path_read_unchecked(struct path* path) {
   return _path_read(path, BOUND_PATH_MAIN, true);
 }

fact-ebpf/src/bpf/d_path.h

@@ -24,6 +24,14 @@
  */
 #define PATH_LEN_CLAMP(len) ((len) & PATH_MAX_MASK)
 
+// Context for __d_path_inner.
+//
+// Supports two modes:
+//   Full path mode: mnt and root must both be set. Crosses mount
+//     boundaries and terminates at the process root.
+//   Dentry-only mode: mnt and root must both be NULL. Walks the
+//     dentry chain to the filesystem root, producing a
+//     filesystem-relative path.
 struct d_path_ctx {
   struct helper_t* helper;
   struct path* root;
@@ -38,39 +46,49 @@ static long __d_path_inner(uint32_t index, void* _ctx) {
   struct d_path_ctx* ctx = (struct d_path_ctx*)_ctx;
   struct dentry* dentry = ctx->dentry;
   struct dentry* parent = BPF_CORE_READ(dentry, d_parent);
-  struct mount* mnt = ctx->mnt;
-  struct dentry* mnt_root = BPF_CORE_READ(mnt, mnt.mnt_root);
 
-  if (dentry == ctx->root->dentry && &mnt->mnt == ctx->root->mnt) {
-    // Found the root of the process, we are done
-    ctx->success = true;
-    return 1;
-  }
+  if (ctx->mnt != NULL) {
+    // Full path mode: we have mount context and can cross mount
+    // boundaries and detect the process root.
+    struct mount* mnt = ctx->mnt;
+    struct dentry* mnt_root = BPF_CORE_READ(mnt, mnt.mnt_root);
 
-  if (dentry == mnt_root) {
-    struct mount* m = BPF_CORE_READ(mnt, mnt_parent);
-    if (m != mnt) {
-      // Current dentry is a mount root different to the previous one we
-      // had (to prevent looping), switch over to that mount position
-      // and keep walking up the path.
-      ctx->dentry = BPF_CORE_READ(mnt, mnt_mountpoint);
-      ctx->mnt = m;
-      return 0;
+    if (dentry == ctx->root->dentry && &mnt->mnt == ctx->root->mnt) {
+      // Found the root of the process, we are done
+      ctx->success = true;
+      return 1;
     }
 
-    // Ended up in a global root, the path might need re-processing or
-    // the root is not attached yet, we are not getting a better path,
-    // so we assume we are correct and stop iterating.
-    ctx->success = true;
-    return 1;
+    if (dentry == mnt_root) {
+      struct mount* m = BPF_CORE_READ(mnt, mnt_parent);
+      if (m != mnt) {
+        // Current dentry is a mount root different to the previous one we
+        // had (to prevent looping), switch over to that mount position
+        // and keep walking up the path.
+        ctx->dentry = BPF_CORE_READ(mnt, mnt_mountpoint);
+        ctx->mnt = m;
+        return 0;
+      }
+
+      // Ended up in a global root, the path might need re-processing or
+      // the root is not attached yet, we are not getting a better path,
+      // so we assume we are correct and stop iterating.
+      ctx->success = true;
+      return 1;
+    }
   }
 
   if (dentry == parent) {
-    // We escaped the mounts and ended up at (most likely) the root of
-    // the device, the path we formed will be wrong.
+    // Reached the root of the filesystem's dentry tree.
+    //
+    // In full path mode (mnt != NULL) this means we escaped the mounts
+    // and the path may be wrong due to a race condition.
     //
-    // This may happen in race conditions where some dentries go away
-    // while we are iterating.
+    // In dentry-only mode (mnt == NULL) this is the expected
+    // termination: we've reached the filesystem root and have a
+    // filesystem-relative path. This is correct for overlayfs
+    // (containers) and for files on the root filesystem.
+    ctx->success = (ctx->mnt == NULL);
     return 1;
   }
 
@@ -140,6 +158,46 @@ __always_inline static long __d_path(const struct path* path, char* buf, int buf
   return buflen - ctx.offset;
 }
 
+/**
+ * Resolve a filesystem-relative path from a bare dentry.
+ *
+ * This is used when no struct path is available (e.g. inode_set_acl).
+ * It walks the dentry chain up to the filesystem root, producing a
+ * path relative to the filesystem's root dentry. This is correct for
+ * overlayfs (containers) and for files on the root filesystem. It
+ * cannot cross mount boundaries, so paths on nested host mounts (e.g.
+ * a separate /var partition) will be relative to that mount's root.
+ */
+__always_inline static long __d_path_from_dentry(struct dentry* dentry, char* buf, int buflen) {
+  if (buflen <= 0) {
+    return -1;
+  }
+
+  int offset = PATH_LEN_CLAMP(buflen - 1);
+  struct d_path_ctx ctx = {
+      .buflen = buflen,
+      .helper = get_helper(),
+      .offset = offset,
+      .mnt = NULL,
+      .root = NULL,
+  };
+
+  if (ctx.helper == NULL) {
+    return -1;
+  }
+
+  ctx.helper->buf[offset] = '\0';
+  ctx.dentry = dentry;
+
+  long res = bpf_loop(PATH_MAX, __d_path_inner, &ctx, 0);
+  if (res <= 0 || !ctx.success) {
+    return -1;
+  }
+
+  bpf_probe_read_str(buf, buflen, &ctx.helper->buf[PATH_LEN_CLAMP(ctx.offset)]);
+  return buflen - ctx.offset;
+}
+
 __always_inline static long d_path(struct path* path, char* buf, int buflen, bool use_bpf_helper) {
   if (use_bpf_helper) {
     return bpf_d_path(path, buf, buflen);

fact-ebpf/src/bpf/events.h

@@ -144,3 +144,45 @@ __always_inline static void submit_rmdir_event(struct submit_event_args_t* args)
 
   __submit_event(args, path_hooks_support_bpf_d_path);
 }
+
+__always_inline static void submit_acl_event(struct submit_event_args_t* args,
+                                             const char* acl_name,
+                                             struct posix_acl* kacl) {
+  if (!reserve_event(args)) {
+    return;
+  }
+
+  args->event->type = FILE_ACTIVITY_ACL_SET;
+
+  // Determine ACL type from the xattr name.
+  // "system.posix_acl_access" vs "system.posix_acl_default"
+  char name_buf[32] = {0};
+  long name_len = bpf_probe_read_kernel_str(name_buf, sizeof(name_buf), acl_name);
+  if (name_len == 25 && __builtin_memcmp(name_buf, "system.posix_acl_default", 24) == 0) {
+    args->event->acl.acl_type = FACT_ACL_TYPE_DEFAULT;
+  } else {
+    args->event->acl.acl_type = FACT_ACL_TYPE_ACCESS;
+  }
+
+  if (kacl == NULL) {
+    args->event->acl.count = 0;
+  } else {
+    unsigned int count = 0;
+    bpf_probe_read_kernel(&count, sizeof(count), &kacl->a_count);
+    if (count > FACT_MAX_ACL_ENTRIES) {
+      count = FACT_MAX_ACL_ENTRIES;
+    }
+    args->event->acl.count = count;
+
+    for (unsigned int i = 0; i < FACT_MAX_ACL_ENTRIES && i < count; i++) {
+      struct posix_acl_entry entry = {0};
+      bpf_probe_read_kernel(&entry, sizeof(entry), &kacl->a_entries[i]);
+      args->event->acl.entries[i].e_tag = entry.e_tag;
+      args->event->acl.entries[i].e_perm = entry.e_perm;
+      args->event->acl.entries[i].e_id = entry.e_uid.val;
+    }
+  }
+
+  // inode_set_acl does not support bpf_d_path (no struct path available)
+  __submit_event(args, false);
+}

fact-ebpf/src/bpf/main.c

@@ -389,6 +389,38 @@ int BPF_PROG(trace_d_instantiate, struct dentry* dentry, struct inode* inode) {
   return 0;
 }
 
+SEC("lsm/inode_set_acl")
+int BPF_PROG(trace_inode_set_acl, struct mnt_idmap* idmap, struct dentry* dentry,
+             const char* acl_name, struct posix_acl* kacl) {
+  struct metrics_t* m = get_metrics();
+  if (m == NULL) {
+    return 0;
+  }
+  struct submit_event_args_t args = {.metrics = &m->inode_set_acl};
+
+  args.metrics->total++;
+
+  struct bound_path_t* bound_path = dentry_read(dentry);
+  if (bound_path == NULL) {
+    bpf_printk("Failed to read path from dentry");
+    args.metrics->error++;
+    return 0;
+  }
+  args.filename = bound_path->path;
+
+  struct inode* inode_ptr = BPF_CORE_READ(dentry, d_inode);
+  args.inode = inode_to_key(inode_ptr);
+  args.monitored = is_monitored(&args.inode, bound_path, NULL);
+
+  if (args.monitored == NOT_MONITORED) {
+    args.metrics->ignored++;
+    return 0;
+  }
+
+  submit_acl_event(&args, acl_name, kacl);
+  return 0;
+}
+
 SEC("lsm/path_rmdir")
 int BPF_PROG(trace_path_rmdir, struct path* dir, struct dentry* dentry) {
   struct metrics_t* m = get_metrics();

fact-ebpf/src/bpf/types.h

@@ -54,6 +54,18 @@ typedef enum monitored_t {
 // For the time being we just keep a char.
 typedef char inode_value_t;
 
+#define FACT_MAX_ACL_ENTRIES 32
+
+// ACL type constants matching the xattr names
+#define FACT_ACL_TYPE_ACCESS 0
+#define FACT_ACL_TYPE_DEFAULT 1
+
+struct acl_entry_t {
+  short e_tag;
+  unsigned short e_perm;
+  unsigned int e_id;
+};
+
 typedef enum file_activity_type_t {
   FILE_ACTIVITY_INIT = -1,
   FILE_ACTIVITY_OPEN = 0,
@@ -64,6 +76,7 @@ typedef enum file_activity_type_t {
   FILE_ACTIVITY_RENAME,
   DIR_ACTIVITY_CREATION,
   DIR_ACTIVITY_UNLINK,
+  FILE_ACTIVITY_ACL_SET,
 } file_activity_type_t;
 
 struct event_t {
@@ -90,6 +103,11 @@ struct event_t {
       inode_key_t inode;
       monitored_t monitored;
     } rename;
+    struct {
+      unsigned int count;
+      unsigned int acl_type;
+      struct acl_entry_t entries[FACT_MAX_ACL_ENTRIES];
+    } acl;
   };
 };
 
@@ -132,4 +150,5 @@ struct metrics_t {
   struct metrics_by_hook_t path_mkdir;
   struct metrics_by_hook_t d_instantiate;
   struct metrics_by_hook_t path_rmdir;
+  struct metrics_by_hook_t inode_set_acl;
 };

fact-ebpf/src/lib.rs

@@ -145,6 +145,7 @@ impl metrics_t {
         self.path_mkdir = self.path_mkdir.accumulate(&other.path_mkdir);
         self.path_rmdir = self.path_rmdir.accumulate(&other.path_rmdir);
         self.d_instantiate = self.d_instantiate.accumulate(&other.d_instantiate);
+        self.inode_set_acl = self.inode_set_acl.accumulate(&other.inode_set_acl);
         self
     }
 }