On 17/12/2021 09:28, Dan Carpenter wrote: > On Fri, Dec 17, 2021 at 09:16:19AM +0000, Steven Price wrote: >> On 17/12/2021 09:10, Dan Carpenter wrote: >>> On Fri, Dec 17, 2021 at 08:55:50AM +0000, Steven Price wrote: >>>> However this one is harder to fix without setting an arbitrary cap on >>>> the number of BOs during a sumbit. I'm not sure how other drivers handle >>>> this - the ones I've looked at so far all have the same issue. There's >>>> obviously the list that Dan already sent, but e.g. msm has the same bug >>>> in msm_gem_submit.c:submit_create() with an amusing bug where the check >>>> for (sz > SIZE_MAX) will never hit, although the call is to kzalloc() so >>>> large allocations are going to fail anyway. >>> >>> sz is u64 and SIZE_MAX is ULONG_MAX so the (sz > SIZE_MAX) condition >>> does work to prevent an integer overflow on 32bit systems. But it's not >>> beautiful. >> >> sz is the result of struct_size() which returns a size_t, and SIZE_MAX >> in case of an overflow. > > Correct. > >> However the check is *greater than* SIZE_MAX >> which will never occur even on 32 bit systems. > > You've missed a part. We add ((u64)nr_cmds * sizeof(submit->cmd[0])) > to SIZE_MAX. If nr_cmds is zero then, whatever, the kzmalloc() will > fail. No big deal. But if it's non-zero then "sz" is larger than > SIZE_MAX and we allocate a smaller buffer than expected leading to > memory corruption.
Ah, my bracket matching is obviously off today - somehow I hadn't noticed that the second line wasn't part of the call to struct_size(). > Btw, it turns out that I had a hand in writing that check so hooray for > me. :) #HoorayForMe Indeed hooray for you! ;) Although it still all seems like a very round-a-bout way of enforcing an arbitrary maximum on the size! Thanks, Steve