In fact, I think you'll hit it even if you use the "for_each_sg()"
helper function. Exactly because it does the sg = sg_next(sg) at the end
of the for-loop, so it will do it for the last entry too, even though that
will access one past the end.
So it really *is* the case that every *single* use of that SG chain needs
to be switched over to only do the sg_next() when required, or sg_next()
needs to be fixed to look at the current-in-use entry (which we *can*
access) when it decides what to do about the next one (which we can *not*
access).
Moving the sg_is_chain() bit to the previous entry would work, but it
requires that nobody who could have a chained scatterlist must *ever*
access "sg->page" directly - you'd always need to use some helper function
that masks off the bit, eg
- rename "sg->page" into "sh->page_and_flag", and make it "unsigned long"
instead of a pointer.
- change every single "sg->page" access to use "sg_page(sg)" instead:
#define sg_pointer(sg) ((void *)((sg)->page_and_flag & ~1ul))
static inline struct page *sg_page(struct scatterist *sg)
{
return sg_pointer(sg);
}
- change "sg_next()" to do
static inline struct scatterlist *sg_next(struct scatterlist *sg)
{
if (sg->page_and_flag & 1)
sg = sg_pointer(sg+1)-1;
return ++sg;
}
where the magic is exactly the fact that now "sg_next()" will *not*
derefence the next SG entry unless the current one was marked
appropriately.
And then *creating* the chain obviously needs to properly mark the
next-to-last entry with the "next entry is a pointer" flag.
Big diff, it sounds like. But I don't see many alternatives. Jens?
Linus
-
