[LU-6174] do_div() silently truncates divisor to uint32_t

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.16.0
• Severity: 3
• Assignee: James A Simmons
• Reporter: Isaac Huang (Inactive)

Description

The Linux do_div() silently truncates divisor to uint32_t, even on x86_64:

 
25 # define do_div(n,base) ({                                      \
26         uint32_t __base = (base);

Any code that uses 64-bit divisor will end up with incorrect result whenever the higher 32 bits aren't all zero. And when the lower 32 bits are all zero, it will end up with a division by zero error.

It happens on both x86_64 and i686. I've verified it on a x86_64 VM.

So any code that looks like the following is BAD:

 
uint64_t foo, bar;
......
if (bar != 0)
    do_div(foo, bar);

There seemed to be plenty of such code in Lustre, e.g.:

 
void fld_cache_fini(struct fld_cache *cache)
        __u64 pct;
......
        if (cache->fci_stat.fst_count > 0) {
                pct = cache->fci_stat.fst_cache * 100;
                do_div(pct, cache->fci_stat.fst_count);

I'd suggest to go through and verify all callers of do_div(). In addition, it'd be good to add a warning on do_div() in our patch checker script.

Issue Links

• is related to: LU-19059 overflow in lov_stripe_size()
• is related to: LU-9887 sanity-lfsck test_9a: FAIL: (4) Got speed 952, expected less than 144
• is related to: LU-6163 64-bit divides should use do_div()
• is related to: LU-8137 fix llverdev to be able to write and verify large files in the filesystem