Overflow in libstd/io/cursor.rs "seek" returns incorrect error with large positions

[amosonn]

Seems to be a reroll #20678. I have some idea how to fix it, will try to draft a pull request later, but it basically requires checking the sign of the i64, negging if required and casting to u64, and using checked_add or checked_sub.

I will divide into cases to demonstrate that this is required, to save repeating phrases, I will use the following notation:
S means pos < i64::max_value(), L means pos > i64::max_value();
P means n >= 0, N means n < 0;
O+ means we will notice bads by checking OF on add (u64::checked_add), O- means we will notice them by checking OF on sub (u64::checked_sub after i64::wrapping_neg), Z+ means we will notice by adding as i64 and making sure result is non-negative, Z- is making sure it is negative, D+ mean we will notice bads by asserting OF is set on add (u64::overflowing_add), D- is likewise for sub (u64::overflowing_sub). Note the we don't need to consider n == i64::min_value(), because after i64::wrapping_neg and as u64 we still get the right, large, positive number.
So, we have the following cases and appropriate checks that spot trouble:

• SP -> O+Z+D- (there's never a problem, but the others false-positive)
• LP -> O+Z-D- (but D- requires also that the result is non-zero).
• SN -> O-Z+D+
• LN -> O-D+ (there's never a problem, but the others false-positive)

Overall, we see that there is no one test which works for all cases, and a minimum of two cases (and tests) yields the solution described above.

[alexcrichton]

Thanks for the report! I'd definitely believe that the exact implementation here is slightly wrong and could use some tweaking.

[amosonn]

I wrote a test that should fail and an implementation that should fix, just waiting for build & tests to finish on my machine before PR. Edit: PR #39874