Skip to contentSkip to conversation

gfs2: gfs2_read_sb: put gfs2_assert inside the loop

PATCH In review Fox Chen 3 participants 4 messages Oct 3, 2020 – Oct 5, 2020 GFS2 file system

lore
Share & tools
lore

Fox Chen

author
#0 lore raw

PATCH 1/1 gfs2: gfs2_read_sb: put gfs2_assert inside the loop

for (x = 2;; x++) {
        ...
        gfs2_assert(sdp, x <= GFS2_MAX_META_HEIGHT);  <--- after
        ...
        if (d != sdp->sd_heightsize[x - 1] || m)
                break;
        sdp->sd_heightsize[x] = space;
}

sdp->sd_max_height = x
gfs2_assert(sdp, sdp->sd_max_height <= GFS2_MAX_META_HEIGHT) <--- before

Before this patch, gfs2_assert is put outside of the loop of sdp->sd_heightsize[x] calculation. When something goes wrong, x exceeds the size of GFS2_MAX_META_HEIGHT, it may already crash inside the loop when

sdp->sd_heightsize[x] = space

tries to reach the out-of-bound
location, gfs2_assert won't help here.

This patch fixes this by moving gfs2_assert into the loop.
We will check x value each time to see if it exceeds GFS2_MAX_META_HEIGHT.

---
 fs/gfs2/ops_fstype.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/gfs2/ops_fstype.c +1 -1
index 6d18d2c91add..6cc32e3010f2 100644
@@ -333,6 +333,7 @@ static int gfs2_read_sb(struct gfs2_sbd *sdp, int silent)
 		u64 space, d;
 		u32 m;
 
+		gfs2_assert(sdp, x <= GFS2_MAX_META_HEIGHT);
 		space = sdp->sd_heightsize[x - 1] * sdp->sd_inptrs;
 		d = space;
 		m = do_div(d, sdp->sd_inptrs);
@@ -343,7 +344,6 @@ static int gfs2_read_sb(struct gfs2_sbd *sdp, int silent)
 	}
 	sdp->sd_max_height = x;
 	sdp->sd_heightsize[x] = ~0;
-	gfs2_assert(sdp, sdp->sd_max_height <= GFS2_MAX_META_HEIGHT);
 
 	sdp->sd_max_dents_per_leaf = (sdp->sd_sb.sb_bsize -
 				      sizeof(struct gfs2_leaf)) /
Signature · 1 line
2.25.1
#1 lore raw

Hi Fox Chen,

Quoted 9 lines · Fox Chen

Before this patch, gfs2_assert is put outside of the loop of sdp->sd_heightsize[x] calculation. When something goes wrong, x exceeds the size of GFS2_MAX_META_HEIGHT, it may already crash inside the loop when

sdp->sd_heightsize[x] = space

tries to reach the out-of-bound
location, gfs2_assert won't help here.

that's true, but the smallest possible block size is 1024 bytes, and with that, the height cannot grow bigger than 10. So the assert is basically there only for documentation purposes.

Thanks,
Andreas

#2 lore raw
In reply to Fox Chen
Quoted 9 lines · Fox Chen
for (x = 2;; x++) {
         ...
         gfs2_assert(sdp, x <= GFS2_MAX_META_HEIGHT);  <--- after
         ...
         if (d != sdp->sd_heightsize[x - 1] || m)
                 break;
         sdp->sd_heightsize[x] = space;
}
Quoted 5 lines

sdp->sd_max_height = x
gfs2_assert(sdp, sdp->sd_max_height <= GFS2_MAX_META_HEIGHT) <--- before

Before this patch, gfs2_assert is put outside of the loop of sdp->sd_heightsize[x] calculation. When something goes wrong,

So this looks related to one of the recent syzbot reports, where the "something goes wrong" is the block size in the on-disk superblock was zeroed and that leads eventually to this out-of-bounds write. The correct fix in that case would be to add a validity check for the block size in gfs2_check_sb().

Andy

Quoted 37 lines

x exceeds the size of GFS2_MAX_META_HEIGHT, it may already crash inside the loop when

sdp->sd_heightsize[x] = space

tries to reach the out-of-bound
location, gfs2_assert won't help here.

This patch fixes this by moving gfs2_assert into the loop.
We will check x value each time to see if it exceeds GFS2_MAX_META_HEIGHT.

---
  fs/gfs2/ops_fstype.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
fs/gfs2/ops_fstype.c +1 -1
index 6d18d2c91add..6cc32e3010f2 100644
@@ -333,6 +333,7 @@ static int gfs2_read_sb(struct gfs2_sbd *sdp, int silent)
  		u64 space, d;
  		u32 m;
 
+		gfs2_assert(sdp, x <= GFS2_MAX_META_HEIGHT);
  		space = sdp->sd_heightsize[x - 1] * sdp->sd_inptrs;
  		d = space;
  		m = do_div(d, sdp->sd_inptrs);
@@ -343,7 +344,6 @@ static int gfs2_read_sb(struct gfs2_sbd *sdp, int silent)
  	}
  	sdp->sd_max_height = x;
  	sdp->sd_heightsize[x] = ~0;
-	gfs2_assert(sdp, sdp->sd_max_height <= GFS2_MAX_META_HEIGHT);
 
  	sdp->sd_max_dents_per_leaf = (sdp->sd_sb.sb_bsize -
  				      sizeof(struct gfs2_leaf)) /
author
#3 lore raw
Quoted 14 lines · Andrew Price
Quoted 12 lines · Fox Chen
for (x = 2;; x++) {
         ...
         gfs2_assert(sdp, x <= GFS2_MAX_META_HEIGHT);  <--- after
         ...
         if (d != sdp->sd_heightsize[x - 1] || m)
                 break;
         sdp->sd_heightsize[x] = space;
}

sdp->sd_max_height = x
gfs2_assert(sdp, sdp->sd_max_height <= GFS2_MAX_META_HEIGHT) <--- before

Quoted 9 lines
Quoted 2 lines

Before this patch, gfs2_assert is put outside of the loop of sdp->sd_heightsize[x] calculation. When something goes wrong,

So this looks related to one of the recent syzbot reports, where the "something goes wrong" is the block size in the on-disk superblock was zeroed and that leads eventually to this out-of-bounds write. The correct fix in that case would be to add a validity check for the block size in gfs2_check_sb().

Yes, I saw this bug from the syzbot report and I though instead of KASAN gfs2_assert should be able to catch it so I proposed this patch. :)

thank you both for your comments.

fox