Nick Piggin <npiggin@xxxxxxxxx> writes: > On 2 May 2012 00:08, Jeff Moyer <jmoyer@xxxxxxxxxx> wrote: >> Nick Piggin <npiggin@xxxxxxxxx> writes: >> >>> Not a bad fix. But it's kind of sad to have i_size checking logic also in >>> block_read_full_page, that does not cope with this. >>> >>> I have found there are parts of the kernel (readahead) that try to read >>> beyond EOF and seem to get angry if we return an error (by not >>> marking uptodate in readpage) in that case though :( >>> >>> But, either way, I think it's very reasonable to not mark buffers beyond >>> end of device as mapped. So I think your patch is fine. >>> >>> I guess for ext[234], it does not read metadata close to the end of the >>> device or you were using 4K sized blocks? >> >> Well, the test case just reads directly from the loop device, bypassing >> the file system, and I did use 1KB blocks when making the file system, so >> it is quite puzzling. > > It's because buffer_head creation does not go through the same paths > for bdev file access versus getblk APIs. > > blkdev_get_block does the right thing there > > In fact, it's probably good to unify the checks here, i.e., use max_blocks() You really think it's worth it? I mean, it's just an i_size_read and a shift, and there is precedent for it inside fs/buffer.c. I'd prefer to keep the patch as-is, but will change it if you feel that strongly about it. Cheers, Jeff -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html