question about potential integer truncation in default_erasesize
hi all,
drivers/mtd/chips/map_rom.c:default_erasesize can truncate map_info.size from unsigned long to unsigned int on 64 bit archs and i'm wondering if this is intentional or should/could map_info.size be turned into an unsigned int field? FTR, this issue was detected with the upcoming version of the size overflow plugin we have in PaX/grsecurity and there're a handful of similar cases in the tree where potentially unwanted or unnecessary integer truncations occur, this being one of these. any opinion/help is welcome!
cheers,
PaX Team
Quoted 9 lines · PaX Team
hi all,
drivers/mtd/chips/map_rom.c:default_erasesize can truncate map_info.size from unsigned long to unsigned int on 64 bit archs and i'm wondering if this is intentional or should/could map_info.size be turned into an unsigned int field? FTR, this issue was detected with the upcoming version of the size overflow plugin we have in PaX/grsecurity and there're a handful of similar cases in the tree where potentially unwanted or unnecessary integer truncations occur, this being one of these. any opinion/help is welcome!
This is being assigned to the erasesize, which is 32-bit already, and all of MTD expects a 32-bit erasesize, so it'd be a pretty big job to "fix" the truncation. But really, we can't handle (and shouldn't; there's really no need) >32-bit eraseblocks. That would be an un-manageable flash geometry.
Regards,
Brian
Quoted 11 lines · Brian Norris
Quoted 9 lines · PaX Team
hi all,
drivers/mtd/chips/map_rom.c:default_erasesize can truncate map_info.size from unsigned long to unsigned int on 64 bit archs and i'm wondering if this is intentional or should/could map_info.size be turned into an unsigned int field? FTR, this issue was detected with the upcoming version of the size overflow plugin we have in PaX/grsecurity and there're a handful of similar cases in the tree where potentially unwanted or unnecessary integer truncations occur, this being one of these. any opinion/help is welcome!
Quoted 3 lines
This is being assigned to the erasesize, which is 32-bit already, and all of MTD expects a 32-bit erasesize, so it'd be a pretty big job to "fix" the truncation.
to make sure i got this right, map_info.size in other uses can hold a value larger than 4GB (so it has to stay 64 bit, at least on 64 bit archs) but for erasesize it should never have such a big value? that'd actually be fine for us since it means when the overflow plugin instruments this code any runtime trigger means a real problem, not a false positive.
cheers,
PaX Team
Quoted 13 lines · PaX Team
Quoted 11 lines · Brian Norris
Quoted 9 lines · PaX Team
hi all,
drivers/mtd/chips/map_rom.c:default_erasesize can truncate map_info.size from unsigned long to unsigned int on 64 bit archs and i'm wondering if this is intentional or should/could map_info.size be turned into an unsigned int field? FTR, this issue was detected with the upcoming version of the size overflow plugin we have in PaX/grsecurity and there're a handful of similar cases in the tree where potentially unwanted or unnecessary integer truncations occur, this being one of these. any opinion/help is welcome!
Quoted 7 lines
Quoted 3 lines
This is being assigned to the erasesize, which is 32-bit already, and all of MTD expects a 32-bit erasesize, so it'd be a pretty big job to "fix" the truncation.
to make sure i got this right, map_info.size in other uses can hold a value larger than 4GB (so it has to stay 64 bit, at least on 64 bit archs) but for erasesize it should never have such a big value?
Correct.
Quoted 3 lines
that'd
actually be fine for us since it means when the overflow plugin instruments this code any runtime trigger means a real problem, not a false positive.
OK, good. As long as you aren't going to start complaining about theoretical concerns we're OK, but a dynamic check is cool.
Brian
Quoted 6 lines · Brian Norris
Quoted 3 lines
that'd
actually be fine for us since it means when the overflow plugin instruments this code any runtime trigger means a real problem, not a false positive.OK, good. As long as you aren't going to start complaining about theoretical concerns we're OK, but a dynamic check is cool.
we don't care about theoretical bugs but real ones too. the reason for asking about this code is that while investigating a false positive report in some other code (that turned out to be due to a limitation in how gcc represents certain constructs at its gimple layer) we decided to check the underlying cause in the whole tree so that we could decide how to fix the FP. the choice is usually either changing the source code (so that gcc produces a different representation) or disabling instrumentation for either the particular expression or all similar constructs. the decision comes down to how many potentially useful checks we'd omit vs. how many more false positives we'd risk having. in this case the code construct was copying between structure fields of different sizes and linux has fortunately a few of these only that can be manually verified, save for a few cases i could not decide myself, hence this email and few others elsewhere (some turned out to be real bugs or just inefficient code in fact).
in short, if you ever hear from us again it'll be because there was a real integer truncation/overflow/etc event (and we'll have precise source location for it ;).