On Tue, May 11, 2021 at 06:04:26PM +0000, Leah Rumancik wrote:

In the ocfs2 changes, I noticed you are using "false", instead of 0,
in the second argument to jbd2_journal_flush.

When I looked more closely, the function signature of
jbd2_journal_flush is also using an unsigned long long for flags,
which struck me as strange:

+extern int	 jbd2_journal_flush(journal_t *journal, unsigned long long flags);

I then noticed that later in the patch series, the ioctl argument is
taking an unsigned long long and we're passing that straight through
to jbd2_journal_flush().

First of all, unsigned long long is not very efficient on many
platforms (especially 32-bit platforms), but also on platforms where
int is 32 bits.  If we don't expect us to need more than 32 flag bits,
I'd suggest explicit ly using __u32 in ioctl interface.  (__u32 is
fine; it's the use of the base int type which can get us into trouble,
since int can be either 32 or 64 bits depending on the architecture).

Secondly, I'd suggest using a different set of flags for
jbd2_journal_flush(), which is an internal kernel interface, and the
EXT4_IOC_CHECKPOINT interface.  We might in the future want to add
some internal flags to jbd2_journal_flush that we do *not* want to
expose via EXT4_IOC_CHECKPOINT, and so it's best that we keep those
two interfaces separate.

I'd suggest defining these in include/linux/jbd2.h, and giving them
names like: JBD2_JOURNAL_FLUSH_DISCARD and JBD2_JOURNAL_FLUSH_ERASE...
(and making the flags parameter an unsigned int).

+	/* flags must be set to either discard or zeroout */
+	if ((flags & JBD2_ERASE_FLAG_DISCARD & JBD2_ERASE_FLAG_ZEROOUT) || !flags)
+		return -EINVAL;

The expression (flags & JBD2_ERASE_FLAG_DISCARD & JBD2_ERASE_FLAG_ZEROOUT)
is always going to evaluate to zero, since (1 & 2) is 0.

What you probably want is something like:

#define JBD2_JOURNAL_FLUSH_DISCARD	0x0001
#define JBD2_JOURNAL_FLUSH_ZEROOUT	0x0002
#define JBD2_JOURNAL_FLUSH_VALID	0x0003

     if ((flags & ~JBD2_JOURNAL_FLUSH_VALID) ||
         ((flags & JBD2_JOURNAL_FLUSH_DISCARD) &&
	  (flags & JBD2_JOURNAL_FLUSH_ZEROOUT)))
	return -EINVAL;

+
+	err = jbd2_journal_bmap(journal, log_offset, &block_start);
+	if (err) {
+		printk(KERN_ERR "JBD2: bad block at offset %lu", log_offset);
+		return err;
+	}

We could get rid of this, and instead make sure block_start is initialized
to ~((unsigned long long) 0).  Then in the loop we can do...

+
+	/*
+	 * use block_start - 1 to meet check for contiguous with previous region:
+	 * phys_block == block_stop + 1
+	 */
+	block_stop = block_start - 1;
+
+	for (block = log_offset; block < journal->j_total_len; block++) {
+		err = jbd2_journal_bmap(journal, block, &phys_block);
+		if (err) {
+			printk(KERN_ERR "JBD2: bad block at offset %lu", block);
+			return err;
+		}

		if (block_start == ~((unsigned long long) 0)) {
			block_start = phys_block;
			block_Stop = block_start - 1;
		}

+
+		if (block == journal->j_total_len - 1) {
+			block_stop = phys_block;
+		} else if (phys_block == block_stop + 1) {
+			block_stop++;
+			continue;
+		}
+
+		/*
+		 * not contiguous with prior physical block or this is last
+		 * block of journal, take care of the region
+		 */
+		byte_start = block_start * journal->j_blocksize;
+		byte_stop = block_stop * journal->j_blocksize;
+		byte_count = (block_stop - block_start + 1) *
+				journal->j_blocksize;
+
+		truncate_inode_pages_range(journal->j_dev->bd_inode->i_mapping,
+				byte_start, byte_stop);
+
+		if (flags & JBD2_ERASE_FLAG_DISCARD) {
+			err = blkdev_issue_discard(journal->j_dev,
+					byte_start >> SECTOR_SHIFT,
+					byte_count >> SECTOR_SHIFT,
+					GFP_NOFS, 0);
+		} else if (flags & JBD2_ERASE_FLAG_ZEROOUT) {
+			err = blkdev_issue_zeroout(journal->j_dev,
+					byte_start >> SECTOR_SHIFT,
+					byte_count >> SECTOR_SHIFT,
+					GFP_NOFS, 0);
+		}
+
+		if (unlikely(err != 0)) {
+			printk(KERN_ERR "JBD2: (error %d) unable to wipe journal at physical blocks %llu - %llu",
+					err, block_start, block_stop);
+			return err;
+		}
+
+		block_start = phys_block;
+		block_stop = phys_block;

Is this right?  When we initialized the loop, above, block_stop was
set to block_start-1 (where block_start == phys_block).  So I think it
might be more correct to replace the above two lines with:

		block_start = ~((unsigned long long) 0);

... and then let block_start and block_stop be initialized in a single
place.  Do you agree?  Does this make sense to you?

	       	       	    	      	    - Ted