Re: [PATCH 1/4] midx repack: avoid integer overflow on 32 bit systems
- From
- Phillip Wood <phillip.wood123@gmail.com>
- Date
- May 21, 2025, 15:19 UTC
- Message-ID
- <a87e6f8f-e6b3-4d91-8b0a-312962819eb4@gmail.com>
- In-Reply-To
- <aCzBvvZDS2OFJ30h@nand.local>
Hi Taylor
On 20/05/2025 18:54, Taylor Blau wrote:
Show 42 quoted lines
> On Tue, May 20, 2025 at 04:04:24PM +0100, Phillip Wood wrote:
>> diff --git a/midx-write.c b/midx-write.c
>> index dd3b3070e55..c7cb2315431 100644
>> --- a/midx-write.c
>> +++ b/midx-write.c
>> @@ -1699,19 +1699,23 @@ static void fill_included_packs_batch(struct repository *r,
>> for (i = 0; total_size < batch_size && i < m->num_packs; i++) {
>> int pack_int_id = pack_info[i].pack_int_id;
>> struct packed_git *p = m->packs[pack_int_id];
>> - size_t expected_size;
>> + uint64_t expected_size;
>>
>> if (!want_included_pack(r, m, pack_kept_objects, pack_int_id))
>> continue;
>>
>> - expected_size = st_mult(p->pack_size,
>> - pack_info[i].referenced_objects);
>> + expected_size = uint64_mult(p->pack_size,
>> + pack_info[i].referenced_objects);
>
> Makes sense.
>
>> expected_size /= p->num_objects;
>>
>> if (expected_size >= batch_size)
>> continue;
>>
>> - total_size += expected_size;
>> + if (unsigned_add_overflows (total_size, (size_t)expected_size))
>> + total_size = SIZE_MAX;
>> + else
>> + total_size += expected_size;
>> +
>
> But this part I am not totally following. Here we have 'total_size'
> declared as a size_t, and 'expected_size' as a uint64_t, and (on 32-bit
> systems) down-cast to a 32-bit unsigned value.
>
> So if 'expected_size' is larger than SIZE_MAX, we should set
> 'total_size' to SIZE_MAX. But that may not happen, say if
> 'expected_size' is (2^32-1<<32). Should total_size also be declared as a
> uint64_t here?By this point we know that expected_size < SIZE_MAX due to the test in the context lines above this change. batch_size is declared as size_t and to get here expected_size < batch_size. I'll add a sentence to the commit message to make that clearer.
> I wondered if it might be easier to count down from the given batch_size > instead of adding up to it (requiring the second > unsigned_add_overflows() check). I tried it out and got this instead:
I think you're right that we if we counted down we'd need one less comparison but I'm not sure if it is worth the churn. In the diff below
factor = pack_info[i].referenced_objects / p->num_objects;
can only ever be zero or one as factor is declared as uint64_t so I don't think it works as-is. If you're happy with the shifted-integer approach in the next patch I'd rather just stick with that.
Thanks
Phillip
Show 53 quoted lines
> --- 8< ---
> diff --git a/midx-write.c b/midx-write.c
> index 48a4dc5e94..f81dd9ff6d 100644
> --- a/midx-write.c
> +++ b/midx-write.c
> @@ -1671,7 +1671,7 @@ static void fill_included_packs_batch(struct repository *r,
> size_t batch_size)
> {
> uint32_t i;
> - size_t total_size;
> + uint64_t remaining = batch_size;
> struct repack_info *pack_info;
> int pack_kept_objects = 0;
>
> @@ -1695,23 +1695,23 @@ static void fill_included_packs_batch(struct repository *r,
>
> QSORT(pack_info, m->num_packs, compare_by_mtime);
>
> - total_size = 0;
> - for (i = 0; total_size < batch_size && i < m->num_packs; i++) {
> + for (i = 0; i < m->num_packs; i++) {
> int pack_int_id = pack_info[i].pack_int_id;
> struct packed_git *p = m->packs[pack_int_id];
> - size_t expected_size;
> + uint64_t expected_size, factor;
>
> if (!want_included_pack(r, m, pack_kept_objects, pack_int_id))
> continue;
>
> - expected_size = st_mult(p->pack_size,
> - pack_info[i].referenced_objects);
> - expected_size /= p->num_objects;
> + factor = pack_info[i].referenced_objects / p->num_objects;
> + if (p->pack_size > UINT64_MAX / factor)
> + die(...);
>
> - if (expected_size >= batch_size)
> - continue;
> + expected_size = p->pack_size * factor;
> + if (expected_size > remaining)
> + break;
>
> - total_size += expected_size;
> + remaining -= expected_size;
> include_pack[pack_int_id] = 1;
> }
> --- >8 ---
>
> That reduces the two overflow checks down to one, and avoids the need to
> introduce a uint64_t-specific variant of the st_add() function.
>
> Thanks,
> Taylor