On 7/12/23 12:04, Yu Kuai wrote:

Subject:
Re: [PATCH v5 02/11] block: Block Device Filtering Mechanism
From:
Yu Kuai [off-list ref]
Date:
7/12/23, 12:04

To:
Yu Kuai [off-list ref], Sergei Shtepa [off-list ref], axboe@kernel.dk, hch@infradead.org, corbet@lwn.net, snitzer@kernel.org
CC:
viro@zeniv.linux.org.uk, brauner@kernel.org, dchinner@redhat.com, willy@infradead.org, dlemoal@kernel.org, linux@weissschuh.net, jack@suse.cz, ming.lei@redhat.com, linux-block@vger.kernel.org, linux-doc@vger.kernel.org, linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, Donald Buczek [off-list ref], "yukuai (C)" [off-list ref]


Hi,

在 2023/07/11 10:02, Yu Kuai 写道:

quoted

quoted

+static bool submit_bio_filter(struct bio *bio)
+{
+    if (bio_flagged(bio, BIO_FILTERED))
+        return false;
+
+    bio_set_flag(bio, BIO_FILTERED);
+    return bio->bi_bdev->bd_filter->ops->submit_bio(bio);
+}
+
  static void __submit_bio(struct bio *bio)
  {
+    /*
+     * If there is a filter driver attached, check if the BIO needs to go to
+     * the filter driver first, which can then pass on the bio or consume it.
+     */
+    if (bio->bi_bdev->bd_filter && submit_bio_filter(bio))
+        return;
+
      if (unlikely(!blk_crypto_bio_prep(&bio)))
          return;

...

quoted

quoted

+static void __blkfilter_detach(struct block_device *bdev)
+{
+    struct blkfilter *flt = bdev->bd_filter;
+    const struct blkfilter_operations *ops = flt->ops;
+
+    bdev->bd_filter = NULL;
+    ops->detach(flt);
+    module_put(ops->owner);
+}
+
+void blkfilter_detach(struct block_device *bdev)
+{
+    if (bdev->bd_filter) {
+        blk_mq_freeze_queue(bdev->bd_queue);

And this is not sate as well, for bio-based device, q_usage_counter is
not grabbed while submit_bio_filter() is called, hence there is a risk
of uaf from submit_bio_filter().

Thanks,
Kuai

Hi Kuai.

Indeed, the filter call is performed before calling bio_queue_enter().
I must admit, you are very attentive. 

I didn't keep track of the change in the position of the
bio_queue_enter() function deeper on the stack.

I think I need to add a check for q_usage_counter, for the debug build.
So that I don't miss it in the future.