[LU-6489] osd-ldiskfs checks s_maxbytes limits incorrectly

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.8.0
• Affects Version/s: Lustre 2.8.0
• Labels: easy osd-ldiskfs
• Severity: 3
• Assignee: Yang Sheng
• Reporter: John Hammond

Description

See the !HAVE_LDISKFS_MAP_BLOCKS version of osd_ldiskfs_map_ext_inode_pages(). The extent_max_page_index comparison is not applied to the first page in the IO and IIUC the comparison should be against fp->index + clen rather than fp->index + i.

static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
                                           struct page **page,
                                           int pages, unsigned long *blocks,
                                           int create)
{
        int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
        int rc = 0, i = 0;
        struct page *fp = NULL;
        int clen = 0;
        pgoff_t extent_max_page_index;

        extent_max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;

        CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
                inode->i_ino, pages, (*page)->index);

        /* pages are sorted already. so, we just have to find
         * contig. space and process them properly */
        while (i < pages) {
                if (fp == NULL) {
                        /* start new extent */
                        fp = *page++;
                        clen = 1;
                        i++;
                        continue;
                } else if (fp->index + clen == (*page)->index) {
                        /* continue the extent */
                        page++;
                        clen++;
                        i++;
                        continue;
                }

                if (fp->index + i >= extent_max_page_index)
                        GOTO(cleanup, rc = -EFBIG);

The other version of osd_ldiskfs_map_inode_pages() looks better but also uses i instead of clen.

In practice this is OK because clients respect llli_maxbytes/lsm_maxbytes/ocd_maxbytes. If client checks are disabled then it's easy to submit IOs that exceed the 16T - 4K limits and cause LBUGs in osd_do_bio(). However dynamic layouts and HSM restore introduce TOCTTOU issues with lli_maxbytes based checks.

Issue Links

• is duplicated by: LU-3732 osd_io.c:320:osd_do_bio()) ASSERTION( iobuf->dr_rw == 0 ) failed: page_idx 4, block_idx 4, i 0