Re: Bug in commit aa511ff8218b ("badblocks: switch to the improved badblock handling

[Ira Weiny <ira.weiny@intel.com>]

Coly Li wrote:

[snip]

> 
> Hi Ira,
> 
> The above information is accurate and very helpful, thank you!
> 
> From __badblocks_check(), the problematic code block is,
> 1303 re_check:
> 1304         bad.start = s;
> 1305         bad.len = sectors;
> 1306
> 1307         if (badblocks_empty(bb)) {
> 1308                 len = sectors;
> 1309                 goto update_sectors;
> 1310         }
> 1311
> 1312         prev = prev_badblocks(bb, &bad, hint);
> 1313
> 1314         /* start after all badblocks */
> 1315         if ((prev + 1) >= bb->count && !overlap_front(bb, prev, &bad)) {
> 1316                 len = sectors;
> 1317                 goto update_sectors;
> 1318         }
> 1319
> 1320         if (overlap_front(bb, prev, &bad)) {
> 1321                 if (BB_ACK(p[prev]))
> 1322                         acked_badblocks++;
> 1323                 else
> 1324                         unacked_badblocks++;
> 1325
> 1326                 if (BB_END(p[prev]) >= (s + sectors))
> 1327                         len = sectors;
> 1328                 else
> 1329                         len = BB_END(p[prev]) - s;
> 1330
> 1331                 if (set == 0) {
> 1332                         *first_bad = BB_OFFSET(p[prev]);
> 1333                         *bad_sectors = BB_LEN(p[prev]);
> 1334                         set = 1;
> 1335                 }
> 1336                 goto update_sectors;
> 1337         }
> 1338
> 1339         /* Not front overlap, but behind overlap */
> 1340         if ((prev + 1) < bb->count && overlap_behind(bb, &bad, prev + 1)) {
> 1341                 len = BB_OFFSET(p[prev + 1]) - bad.start;
> 1342                 hint = prev + 1;
> 1343                 goto update_sectors;
> 1344         }
> 1345
> 1346         /* not cover any badblocks range in the table */
> 1347         len = sectors;
> 1348
> 1349 update_sectors:
> 
> If the checking range is before all badblocks records in the badblocks table,
> value -1 is returned from prev_badblock(). Code blocks between line 1314 and
> line 1337 doesn't hanle the implicit '-1' value properly. Then counter
> unacked_badblocks is increased at line 1324 mistakenly.
> 
> So the value prev should be checked and make sure '>= 0' before comparing
> the checking range with a badblock record returned by prev_badblocks(). Other
> wise it dones't make sense.
> 
> For badblocks_set() and badblocks_clear(), 'prev < 0' is explicitly checked,
> value '-1' doesn't go though into following code.
> 
> Could you please apply and try the attached patch? Hope it may help a bit.
> 
> And now it is weekend time, you may be out of office and not able to access
> the testing hardware. I will do more testing from myside and update more info
> if necessary.
> 
> Thanks for the report and debug!
> 
> Coly Li
> 
> [debug patch snipped]

This debug patch does fix our tests.  Thanks!

But Nan has submitted a series to fix this as well.[1]

I'm going to test his series as well.

Thanks!
Ira

[1] https://lore.kernel.org/linux-block/20231223063728.3229446-1-linan666@huaweicloud.com/