Skip to contentSkip to conversation

cifs: fix skipping to incorrect offset in emit_cached_diretns

PATCH In review Ronnie Sahlberg 4 participants 5 messages Oct 11, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Steve,

V3 of the patch. We have to handle that there may be indicies that are skipped in the ->pos sequence so add support to handle this. As we now can handle this case we no longer need to modify cifs_filldir to return -EEXIST for these cases.

This fixes generic/006/100/111/257

#1 lore raw

PATCH 1/1 cifs: fix skipping to incorrect offset in emit_cached_dirents

When application has done lseek() to a different offset on a directory fd we skipped one entry too many before we start emitting directory entries from the cache.

We need to also make sure that when we are starting to emit directory entries from the cache, the ->pos sequence might have holes and skip some indices.

---
 fs/cifs/readdir.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)
fs/cifs/readdir.c +10 -5
index 8e060c00c969..7170614434a1 100644
@@ -847,14 +847,19 @@ static bool emit_cached_dirents(struct cached_dirents *cde,
 	int rc;
 
 	list_for_each_entry(dirent, &cde->entries, entry) {
-		if (ctx->pos >= dirent->pos)
+		if (ctx->pos > dirent->pos)
 			continue;
+		/*
+		 * There may be holes in the ->pos sequence
+		 * so always force ctx->pos to the current position.
+		 */
 		ctx->pos = dirent->pos;
 		rc = dir_emit(ctx, dirent->name, dirent->namelen,
 			      dirent->fattr.cf_uniqueid,
 			      dirent->fattr.cf_dtype);
 		if (!rc)
 			return rc;
+		ctx->pos++;
 	}
 	return true;
 }
@@ -1202,10 +1207,10 @@ int cifs_readdir(struct file *file, struct dir_context *ctx)
 				 ctx->pos, tmp_buf);
 			cifs_save_resume_key(current_entry, cifsFile);
 			break;
-		} else
-			current_entry =
-				nxt_dir_entry(current_entry, end_of_smb,
-					cifsFile->srch_inf.info_level);
+		}
+		current_entry =
+			nxt_dir_entry(current_entry, end_of_smb,
+				      cifsFile->srch_inf.info_level);
 	}
 	kfree(tmp_buf);
 
Signature · 1 line
2.35.3
reviewer
#2 lore raw

Re: [PATCH] cifs: fix skipping to incorrect offset in emit_cached_dirents

Quoted 18 lines · Ronnie Sahlberg

When application has done lseek() to a different offset on a directory fd we skipped one entry too many before we start emitting directory entries from the cache.

We need to also make sure that when we are starting to emit directory entries from the cache, the ->pos sequence might have holes and skip some indices.

---
  fs/cifs/readdir.c | 15 ++++++++++-----
  1 file changed, 10 insertions(+), 5 deletions(-)
fs/cifs/readdir.c +0 -0
index 8e060c00c969..7170614434a1 100644

fs/cifs/readdir.c static bool emit_cached_dirents(struct cached_dirents *cde, · line 855

Quoted 10 lines
fs/cifs/readdir.c +5 -1
@@ -847,14 +847,19 @@ static bool emit_cached_dirents(struct cached_dirents *cde,
  	int rc;
 
  	list_for_each_entry(dirent, &cde->entries, entry) {
-		if (ctx->pos >= dirent->pos)
+		if (ctx->pos > dirent->pos)
  			continue;
+		/*
+		 * There may be holes in the ->pos sequence
+		 * so always force ctx->pos to the current position.
+		 */

The comment is a bit vague by referring to "->pos", because it's the same name in both ctx and dirent.

But I have a second question, does squeezing out the holes affect a later possible lseek on the dirhandle? I'm having trouble tracking down where that might happen.

fs/cifs/readdir.c static bool emit_cached_dirents(struct cached_dirents *cde, · line 861

Quoted 6 lines
fs/cifs/readdir.c +0 -0
  		ctx->pos = dirent->pos;
  		rc = dir_emit(ctx, dirent->name, dirent->namelen,
  			      dirent->fattr.cf_uniqueid,
  			      dirent->fattr.cf_dtype);
  		if (!rc)
  			return rc;

It's weird that "rc" is an integer, but dir_emit() returns a bool. It's also confusing that this isn't coded as

		if (!rc)
			return false;

Other than those questions, it looks like a clever fix.

Tom.

Quoted 19 lines
fs/cifs/readdir.c +5 -4
+		ctx->pos++;
  	}
  	return true;
  }
@@ -1202,10 +1207,10 @@ int cifs_readdir(struct file *file, struct dir_context *ctx)
  				 ctx->pos, tmp_buf);
  			cifs_save_resume_key(current_entry, cifsFile);
  			break;
-		} else
-			current_entry =
-				nxt_dir_entry(current_entry, end_of_smb,
-					cifsFile->srch_inf.info_level);
+		}
+		current_entry =
+			nxt_dir_entry(current_entry, end_of_smb,
+				      cifsFile->srch_inf.info_level);
  	}
  	kfree(tmp_buf);
 
#3 lore raw

Re: [PATCH] cifs: fix skipping to incorrect offset in emit_cached_dirents

I ran the buildbot on this and it looks like this version of the patch fixed the tests which failed with the earlier version of the patch

Quoted 73 lines · Tom Talpey
Quoted 18 lines · Ronnie Sahlberg

When application has done lseek() to a different offset on a directory fd we skipped one entry too many before we start emitting directory entries from the cache.

We need to also make sure that when we are starting to emit directory entries from the cache, the ->pos sequence might have holes and skip some indices.

---
  fs/cifs/readdir.c | 15 ++++++++++-----
  1 file changed, 10 insertions(+), 5 deletions(-)
fs/cifs/readdir.c +0 -0
index 8e060c00c969..7170614434a1 100644

fs/cifs/readdir.c static bool emit_cached_dirents(struct cached_dirents *cde, · line 855

Quoted 10 lines
fs/cifs/readdir.c +5 -1
@@ -847,14 +847,19 @@ static bool emit_cached_dirents(struct cached_dirents *cde,
      int rc;
 
      list_for_each_entry(dirent, &cde->entries, entry) {
-             if (ctx->pos >= dirent->pos)
+             if (ctx->pos > dirent->pos)
                      continue;
+             /*
+              * There may be holes in the ->pos sequence
+              * so always force ctx->pos to the current position.
+              */

The comment is a bit vague by referring to "->pos", because it's the same name in both ctx and dirent.

But I have a second question, does squeezing out the holes affect a later possible lseek on the dirhandle? I'm having trouble tracking down where that might happen.

fs/cifs/readdir.c static bool emit_cached_dirents(struct cached_dirents *cde, · line 861

Quoted 6 lines
fs/cifs/readdir.c +0 -0
              ctx->pos = dirent->pos;
              rc = dir_emit(ctx, dirent->name, dirent->namelen,
                            dirent->fattr.cf_uniqueid,
                            dirent->fattr.cf_dtype);
              if (!rc)
                      return rc;

It's weird that "rc" is an integer, but dir_emit() returns a bool. It's also confusing that this isn't coded as

                if (!rc)
                        return false;

Other than those questions, it looks like a clever fix.

Tom.

Quoted 19 lines
fs/cifs/readdir.c +5 -4
+             ctx->pos++;
      }
      return true;
  }
@@ -1202,10 +1207,10 @@ int cifs_readdir(struct file *file, struct dir_context *ctx)
                               ctx->pos, tmp_buf);
                      cifs_save_resume_key(current_entry, cifsFile);
                      break;
-             } else
-                     current_entry =
-                             nxt_dir_entry(current_entry, end_of_smb,
-                                     cifsFile->srch_inf.info_level);
+             }
+             current_entry =
+                     nxt_dir_entry(current_entry, end_of_smb,
+                                   cifsFile->srch_inf.info_level);
      }
      kfree(tmp_buf);
 
Signature · 3 lines
Thanks,

Steve
#4 lore raw

Re: [PATCH] cifs: fix skipping to incorrect offset in emit_cached_dirents

In reply to Tom Talpey
Quoted 34 lines · Tom Talpey
Quoted 18 lines · Ronnie Sahlberg

When application has done lseek() to a different offset on a directory fd we skipped one entry too many before we start emitting directory entries from the cache.

We need to also make sure that when we are starting to emit directory entries from the cache, the ->pos sequence might have holes and skip some indices.

---
  fs/cifs/readdir.c | 15 ++++++++++-----
  1 file changed, 10 insertions(+), 5 deletions(-)
fs/cifs/readdir.c +0 -0
index 8e060c00c969..7170614434a1 100644

fs/cifs/readdir.c static bool emit_cached_dirents(struct cached_dirents *cde, · line 855

Quoted 10 lines
fs/cifs/readdir.c +5 -1
@@ -847,14 +847,19 @@ static bool emit_cached_dirents(struct cached_dirents *cde,
      int rc;
 
      list_for_each_entry(dirent, &cde->entries, entry) {
-             if (ctx->pos >= dirent->pos)
+             if (ctx->pos > dirent->pos)
                      continue;
+             /*
+              * There may be holes in the ->pos sequence
+              * so always force ctx->pos to the current position.
+              */

The comment is a bit vague by referring to "->pos", because it's the same name in both ctx and dirent.

Quoted 3 lines

But I have a second question, does squeezing out the holes affect a later possible lseek on the dirhandle? I'm having trouble tracking down where that might happen.

I think if an application does an lseek() to a position we did not previously tell
the application about, then what happens is undefined. What happens in this patch is that we just "skip" to the next entry with a "valid" pos
and start returning data from there.
I.e. we just skip over the "hole" but that should be fine since the application should never
have lseek()ed to somewhere inside the hole to begin with.

Quoted 36 lines

fs/cifs/readdir.c static bool emit_cached_dirents(struct cached_dirents *cde, · line 861

Quoted 7 lines
fs/cifs/readdir.c +0 -0
              ctx->pos = dirent->pos;
              rc = dir_emit(ctx, dirent->name, dirent->namelen,
                            dirent->fattr.cf_uniqueid,
                            dirent->fattr.cf_dtype);
              if (!rc)
                      return rc;

It's weird that "rc" is an integer, but dir_emit() returns a bool. It's also confusing that this isn't coded as

                if (!rc)
                        return false;

Other than those questions, it looks like a clever fix.

Tom.

Quoted 19 lines
fs/cifs/readdir.c +5 -4
+             ctx->pos++;
      }
      return true;
  }
@@ -1202,10 +1207,10 @@ int cifs_readdir(struct file *file, struct dir_context *ctx)
                               ctx->pos, tmp_buf);
                      cifs_save_resume_key(current_entry, cifsFile);
                      break;
-             } else
-                     current_entry =
-                             nxt_dir_entry(current_entry, end_of_smb,
-                                     cifsFile->srch_inf.info_level);
+             }
+             current_entry =
+                     nxt_dir_entry(current_entry, end_of_smb,
+                                   cifsFile->srch_inf.info_level);
      }
      kfree(tmp_buf);