Fix kmem read on 32-bit archs
Quoted 3 lines · Linus Torvalds
Yes, the thing needs to be opened with
O_LARGEFILEand you need to use "llseek()" to seek into it, but once you do that, everything should be fine.
GCC warns about using llseek and suggests lseek64 instead. That works
for me, but up till 2.6.11 plain lseek worked too. I guess it really shouldn't have?
__
Chuck
Quoted 4 lines · Chuck Ebbert
GCC warns about using llseek and suggests lseek64 instead. That works
for me, but up till 2.6.11 plain lseek worked too. I guess it really shouldn't have?
Well, we have had various special-cases for /dev/[k]mem to try to make it work even with 32-bit lseek in the past (or with 64-bit lseek on archs that need the high bit set). So I wouldn't say "shouldn't have" - I'd love it if it worked, but my ove for it working is somewhat tempered by the need for it not screwing up everything else in the kernel ;)
Now, it's entirely possible that we shoul djust make "sys_lseek()" work
differently: right now it _always_ sign-extends its offset from "off_t" to "loff_t", but the thing is, for a SEEK_SET, it probably makes more sense to zero-extend it (since a negative SEEK_SET just doesn't make any sense).
So a hacky alternative might be something like the appended, but I have to admit that it's pretty damn hacky even if it makes sense at some level..
Oh, btw, this will cause it to still return EOVERFLOW, you'd need to also change that error return logic.
Linus
-----
fs/read_write.c +5 -1
@@ -137,7 +137,11 @@ asmlinkage off_t sys_lseek(unsigned int retval = -EINVAL ;if (origin <=2 ) { -loff_t res =vfs_llseek (file, offset, origin); +/* SEEK_SET zero-extends, others sign-extend */ +loff_t res = offset; +if (!origin) + res = (unsigned long )offset; + res =vfs_llseek (file, res, origin); retval = res;if (res != (loff_t )retval) retval = -EOVERFLOW ;/* LFS: should only happen on 32 bit platforms */