[LU-1883] osd-ldiskfs fills file offsets into lnb->offset fields used as page offsets

• Type: Bug
• Resolution: Fixed
• Priority: Blocker
• Fix Version/s: Lustre 2.4.0
• Affects Version/s: Lustre 2.4.0
• Labels: USE_OFD
• Severity: 3
• Assignee: Li Wei (Inactive)
• Reporter: Li Wei (Inactive)

Description

This affects ofd + osd-ldiskfs but not obdfilter or current MDS/MGS stack.

static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
                                   struct niobuf_local *lnb)
{
        ENTRY;

        *nrpages = 0;  

        while (len > 0) {
                int poff = offset & (CFS_PAGE_SIZE - 1);
                int plen = CFS_PAGE_SIZE - poff;

                if (plen > len)
                        plen = len;
                lnb->offset = offset;
                /* lnb->lnb_page_offset = poff; */

"offset" contains the file offset, while "lnb->offset" is expected to be the page offset:

static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
                          struct niobuf_local *lnb, int npages)
{
[...]

                if (maxidx >= lnb[i].page->index) {
                        osd_iobuf_add_page(iobuf, lnb[i].page);
                } else {
                        long off;
                        char *p = kmap(lnb[i].page);

                        off = lnb[i].offset;
                        if (off)
                                memset(p, 0, off);

The memset() call might zero out quite a number of pages that may not belong to Lustre.

Issue Links

• is related to: LU-2406 Interop 2.3<->2.4 Failure: unable to handle kernel NULL pointer dereference at (null)
• is related to: LU-1844 sanityn, subtest test_16: list_del corruption when run ofd + ldiskfs

Comments

[Jodi Levi (Inactive)]

Please let me know if this is still an issue and needs to be reopened.

[Li Wei (Inactive)]

The master patch has landed.

[Li Wei (Inactive)]

This is actually a blocker for 2.4. I have to remove 2.3.0 from the "affects version/s" list, so that this does not block 2.3.

[Li Wei (Inactive)]

http://review.whamcloud.com/4051

This is for master.

[Li Wei (Inactive)]

Yes, the patch makes sure lnb_file_offset and lnb_page_offset both store what the names imply.

[Mikhail Pershin]

lnb has already lnb_file_offset field in master, so we need just use it as file offset and lnb->offset as page offset. Separate patch would rename other lnb fields to have lnb_ prefix.

[Li Wei (Inactive)]

http://review.whamcloud.com/3940

This is for b2_3.

[Alex Zhuravlev]

the correct fix would be to introduce lnb_page_offset and lnb_file_offset as in orion.