On Wed, Jan 22, 2020 at 09:06:51AM +0100, Michal Hocko wrote: >On Tue 21-01-20 11:01:30, Yang Shi wrote: >> >> >> On 1/21/20 12:40 AM, Michal Hocko wrote: >> > On Tue 21-01-20 09:44:16, Wei Yang wrote: >> > > On Mon, Jan 20, 2020 at 02:17:44PM +0100, Michal Hocko wrote: >> > > > On Mon 20-01-20 14:06:26, Michal Hocko wrote: >> > > > > On Sat 18-01-20 13:26:43, Yang Shi wrote: >> > > > > > The do_move_pages_to_node() might return > 0 value, the number of pages >> > > > > > that are not migrated, then the value will be returned to userspace >> > > > > > directly. But, move_pages() syscall would just return 0 or errno. So, >> > > > > > we need reset the return value to 0 for such case as what pre-v4.17 did. >> > > > > The patch is wrong. migrate_pages returns the number of pages it >> > > > > _hasn't_ migrated or -errno. Yeah that semantic sucks but... >> > > > > So err != 0 is always an error. Except err > 0 doesn't really provide >> > > > > any useful information to the userspace. I cannot really remember what >> > > > > was the actual behavior before my rework because there were some gotchas >> > > > > hidden there. >> > > > OK, so I've double checked. do_move_page_to_node_array would carry the >> > > > error code over to do_pages_move and it would store the status stored >> > > > in the pm array. It contains page_to_nid(page) so the resulting code >> > > > indeed behaves properly before my change and this is a regression. I >> > > Thanks, I see the change. >> > > >> > > > have a very vague recollection that this has been brought up already. >> > > > <...looks in notes...> >> > > > Found it! The report is >> > > > http://lkml.kernel.org/r/0329efa0984b9b0252ef166abb4498c0795fab36.1535113317.git.jstancek@xxxxxxxxxx >> > > > and my proposed workaround was http://lkml.kernel.org/r/20180829145537.GZ10223@xxxxxxxxxxxxxx >> > > Well, the above two links return 404. >> > You are right. They are not archived for some reason. Anyway, the patch >> > I was proposing back then is below: >> > >> > commit cfb88c266b645197135cde2905c2bfc82f6d82a9 >> > Author: Michal Hocko <mhocko@xxxxxxxx> >> > Date: Wed Nov 14 12:19:09 2018 +0100 >> > >> > mm: fix do_pages_move error reporting >> > a49bd4d71637 ("mm, numa: rework do_pages_move") has changed the way how >> > we report error to layers above. As the changelog mentioned the semantic >> > was quite unclear previously because the return 0 could mean both >> > success and failure. >> > The above mentioned commit didn't get all the way down to fix this >> > completely because it doesn't report pages that we even haven't >> > attempted to migrate and therefore we cannot simply say that the >> > semantic is: >> > - err < 0 - errno >> > - err >= 0 number of non-migrated pages. >> > Fixes: a49bd4d71637 ("mm, numa: rework do_pages_move") >> > Signed-off-by: Michal Hocko <mhocko@xxxxxxxx> >> >> Thanks, Michal. But, it looks this patch still could return > 0 value (the >> total number of non-migrated pages, including not even attempted pages) too, >> but the problem we are trying to fix is to make do_pages_move() return <= 0 >> value only since the man page of move_pages() doesn't allow return > 0 >> value. > >Yes this patch just lives with the changed semantic and tries to make it >sensible. So if some page cannot be migrated then we just stop and >return the number of non migrated pages at the tail of the given array. >This would make error handling slightly easier because you know that >count - ret pages of the array can be skipped if ret >= 0. > Got some different idea for this. Replied in the patch thread. >> And, by looking into the old code (v4.16), I spotted another problem. The >> migrate_pages() would store the migration failure error code into >> page_to_node->status. So, When do_move_page_to_node_array() returns > 0 >> value, the return value would be reset to 0 and the migration error codes >> for non-migrated pages would be stored into status to return to userspace. >> But, the rework removed this. >> >> I didn't dig into the intention of the rework, is it expected? > >I have tried to preserve the original semantic as possible. As explained >in the changelog there were quite some discrepancies even before. This >new one was not really intentional. We have effectively two options >here. Either somebody really depend on the former semantic and we have >to fix this or we can relax the semantic as the above patch attempts. > >I would be more inclined for the second option as nobody has complained >about the new semantic except for few ltp tests which do not represent >real workload. If you have a real usecase then speak up please. >-- >Michal Hocko >SUSE Labs -- Wei Yang Help you, Help me
