| Message ID | 48cbd548d153d1d2a1cf6c4f2127a6cef5d55deb.camel@redhat.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [GIT,PULL] Please pull rdma.git | expand |
On Mon, Apr 29, 2019 at 9:29 AM Doug Ledford <dledford@redhat.com> wrote: > > > drivers/infiniband/core/uverbs_main.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) This trivial one-liner is actually incorrect. It should use 'vmf->address', because the point of the ZERO_PAGE argument is to pick the page with the right virtual address alias for broken architectures that need those kinds. I'm actually surprised s390 wants it, usually it's just MIPS that has the horribly broken virtual address translation stuff. But it looks like for s390 it's at least only a performance issue (ie it causes some aliases in L1 that cause cacheline ping-pong rather than anything else). Linus
On Mon, 2019-04-29 at 09:48 -0700, Linus Torvalds wrote: > On Mon, Apr 29, 2019 at 9:29 AM Doug Ledford <dledford@redhat.com> wrote: > > > > drivers/infiniband/core/uverbs_main.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > This trivial one-liner is actually incorrect. > > It should use 'vmf->address', because the point of the ZERO_PAGE > argument is to pick the page with the right virtual address alias for > broken architectures that need those kinds. > > I'm actually surprised s390 wants it, usually it's just MIPS that has > the horribly broken virtual address translation stuff. But it looks > like for s390 it's at least only a performance issue (ie it causes > some aliases in L1 that cause cacheline ping-pong rather than anything > else). That's what I get for listening to Jason ;-) Well, since you have just essentially re-written the patch to be correct, you are now the developer of origin. Do you want to commit the fix directly or shall I respin it for you to pull?
On Mon, Apr 29, 2019 at 01:13:01PM -0400, Doug Ledford wrote: > On Mon, 2019-04-29 at 09:48 -0700, Linus Torvalds wrote: > > On Mon, Apr 29, 2019 at 9:29 AM Doug Ledford <dledford@redhat.com> wrote: > > > > > > drivers/infiniband/core/uverbs_main.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > This trivial one-liner is actually incorrect. > > > > It should use 'vmf->address', because the point of the ZERO_PAGE > > argument is to pick the page with the right virtual address alias for > > broken architectures that need those kinds. > > > > I'm actually surprised s390 wants it, usually it's just MIPS that has > > the horribly broken virtual address translation stuff. But it looks > > like for s390 it's at least only a performance issue (ie it causes > > some aliases in L1 that cause cacheline ping-pong rather than anything > > else). > > That's what I get for listening to Jason ;-) > > Well, since you have just essentially re-written the patch to be > correct, you are now the developer of origin. Do you want to commit the > fix directly or shall I respin it for you to pull? Linus already committed it. The whole breakage is mystery for me, the patch which introduced breakage, got successful build report from 0-build. Thanks > > -- > Doug Ledford <dledford@redhat.com> > GPG KeyID: B826A3330E572FDD > Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
On Mon, 2019-04-29 at 21:00 +0300, Leon Romanovsky wrote: > On Mon, Apr 29, 2019 at 01:13:01PM -0400, Doug Ledford wrote: > > On Mon, 2019-04-29 at 09:48 -0700, Linus Torvalds wrote: > > > On Mon, Apr 29, 2019 at 9:29 AM Doug Ledford <dledford@redhat.com> wrote: > > > > drivers/infiniband/core/uverbs_main.c | 2 +- > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > This trivial one-liner is actually incorrect. > > > > > > It should use 'vmf->address', because the point of the ZERO_PAGE > > > argument is to pick the page with the right virtual address alias for > > > broken architectures that need those kinds. > > > > > > I'm actually surprised s390 wants it, usually it's just MIPS that has > > > the horribly broken virtual address translation stuff. But it looks > > > like for s390 it's at least only a performance issue (ie it causes > > > some aliases in L1 that cause cacheline ping-pong rather than anything > > > else). > > > > That's what I get for listening to Jason ;-) > > > > Well, since you have just essentially re-written the patch to be > > correct, you are now the developer of origin. Do you want to commit the > > fix directly or shall I respin it for you to pull? > > Linus already committed it. > > The whole breakage is mystery for me, the patch which introduced > breakage, got successful build report from 0-build. Yeah, successful 0-day build is a good indicator, but not a guarantee. You just can't test all possible permutations and sometimes things slip by. But it's still way better than 20 years ago when half the patches weren't even compile tested before coming to Linus ;-)
Hi Linus, As per the thread on our last pull request, this is the two line build fix for s390 and mips. The following changes since commit 2557fabd6e29f349bfa0ac13f38ac98aa5eafc74: RDMA/hns: Bugfix for mapping user db (2019-04-25 10:40:04 -0300) are available in the Git repository at: git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git tags/for- linus for you to fetch changes up to d79a26b99f5f40db6863b1973750fd1d134d99b4: RDMA/uverbs: Fix compilation error on s390 and mips platforms (2019- 04-29 11:47:55 -0400) ---------------------------------------------------------------- Pull request for 5.1-rc - Build fix for the last pull request on s390 ---------------------------------------------------------------- Leon Romanovsky (1): RDMA/uverbs: Fix compilation error on s390 and mips platforms drivers/infiniband/core/uverbs_main.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)