| Message ID | 1507028295-9353-3-git-send-email-mguzik@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Oct 3, 2017 at 3:58 AM, Mateusz Guzik <mguzik@redhat.com> wrote: > Explicit locking in the fallback case provides a safe state of the > table. Getting rid of blocking semantics makes __fd_install usable > again in non-sleepable contexts, which easies backporting efforts. > > There is a side effect of slightly nicer assembly for the common case > as might_sleep can now be removed. > > Signed-off-by: Mateusz Guzik <mguzik@redhat.com> > --- > Documentation/filesystems/porting | 4 ---- > fs/file.c | 11 +++++++---- > 2 files changed, 7 insertions(+), 8 deletions(-) Nice change ! Reviewed-by: Eric Dumazet <edumazet@google.com>
On Tue, Oct 03, 2017 at 07:41:11AM -0700, Eric Dumazet wrote: > On Tue, Oct 3, 2017 at 3:58 AM, Mateusz Guzik <mguzik@redhat.com> wrote: > > Explicit locking in the fallback case provides a safe state of the > > table. Getting rid of blocking semantics makes __fd_install usable > > again in non-sleepable contexts, which easies backporting efforts. > > > > There is a side effect of slightly nicer assembly for the common case > > as might_sleep can now be removed. > > > > Signed-off-by: Mateusz Guzik <mguzik@redhat.com> > > --- > > Documentation/filesystems/porting | 4 ---- > > fs/file.c | 11 +++++++---- > > 2 files changed, 7 insertions(+), 8 deletions(-) > > Nice change ! > > Reviewed-by: Eric Dumazet <edumazet@google.com> Hey Eric, Any chance you could review the patches from Sandhya that make this entire codepath obsolete? https://lkml.org/lkml/2017/4/29/20
On Wed, Oct 4, 2017 at 6:55 AM, Matthew Wilcox <willy@infradead.org> wrote: > On Tue, Oct 03, 2017 at 07:41:11AM -0700, Eric Dumazet wrote: >> On Tue, Oct 3, 2017 at 3:58 AM, Mateusz Guzik <mguzik@redhat.com> wrote: >> > Explicit locking in the fallback case provides a safe state of the >> > table. Getting rid of blocking semantics makes __fd_install usable >> > again in non-sleepable contexts, which easies backporting efforts. >> > >> > There is a side effect of slightly nicer assembly for the common case >> > as might_sleep can now be removed. >> > >> > Signed-off-by: Mateusz Guzik <mguzik@redhat.com> >> > --- >> > Documentation/filesystems/porting | 4 ---- >> > fs/file.c | 11 +++++++---- >> > 2 files changed, 7 insertions(+), 8 deletions(-) >> >> Nice change ! >> >> Reviewed-by: Eric Dumazet <edumazet@google.com> > > Hey Eric, > > Any chance you could review the patches from Sandhya that make this entire > codepath obsolete? > > https://lkml.org/lkml/2017/4/29/20 > Hmm... 18 files changed, 578 insertions(+), 585 deletions(-) Frankly I need to be convinced with solid performance numbers before I am taking a look at this series. I do not believe an IDR will be faster than current implementation, so I am not quite convinced at this moment. Thanks.
On Wed, Oct 04, 2017 at 07:00:40AM -0700, Eric Dumazet wrote: > On Wed, Oct 4, 2017 at 6:55 AM, Matthew Wilcox <willy@infradead.org> wrote: > > Any chance you could review the patches from Sandhya that make this entire > > codepath obsolete? > > > > https://lkml.org/lkml/2017/4/29/20 > > > > Hmm... > > 18 files changed, 578 insertions(+), 585 deletions(-) > > Frankly I need to be convinced with solid performance numbers before I > am taking a look at this series. I was hoping you'd help us get some solid performance numbers ... you seem to have workloads available to you that help find weaknesses in implementations. The number of lines inserted is a bit of a red herring. Over 100 are in the test suite (you surely aren't going to review those) and another ~300 are adding enhancements to the IDR & radix tree that should be useful for other users (eg I think I have a way to speed up writing out dirty pages by using get_tag_batch()). > I do not believe an IDR will be faster than current implementation, so > I am not quite convinced at this moment. I don't think it should be significantly different in performance. Let's look at the layout of data for a typical bash shell (fds 0-2 and 255 open). Current implementation: files_struct -> fdt -> fd -> struct file IDR: files_struct -> radix node (shift 6) -> radix node (shift 0) -> struct file In either case, it's the same number of dependent loads. It'll start to look worse for the radix tree above 4096 open fds in a given process.
On Wed, Oct 4, 2017 at 7:58 AM, Matthew Wilcox <willy@infradead.org> wrote: > On Wed, Oct 04, 2017 at 07:00:40AM -0700, Eric Dumazet wrote: >> On Wed, Oct 4, 2017 at 6:55 AM, Matthew Wilcox <willy@infradead.org> wrote: >> > Any chance you could review the patches from Sandhya that make this entire >> > codepath obsolete? >> > >> > https://lkml.org/lkml/2017/4/29/20 >> > >> >> Hmm... >> >> 18 files changed, 578 insertions(+), 585 deletions(-) >> >> Frankly I need to be convinced with solid performance numbers before I >> am taking a look at this series. > > I was hoping you'd help us get some solid performance numbers ... you > seem to have workloads available to you that help find weaknesses in > implementations. > > The number of lines inserted is a bit of a red herring. Over 100 are in > the test suite (you surely aren't going to review those) and another ~300 > are adding enhancements to the IDR & radix tree that should be useful > for other users (eg I think I have a way to speed up writing out dirty > pages by using get_tag_batch()). > >> I do not believe an IDR will be faster than current implementation, so >> I am not quite convinced at this moment. > > I don't think it should be significantly different in performance. Let's > look at the layout of data for a typical bash shell (fds 0-2 and 255 open). > > Current implementation: > > files_struct -> fdt -> fd -> struct file > > IDR: > > files_struct -> radix node (shift 6) -> radix node (shift 0) -> struct file > > In either case, it's the same number of dependent loads. It'll start > to look worse for the radix tree above 4096 open fds in a given process. I am interested in performance for process with 10,000,000 fds, and ~100 threads constantly adding/deleting/using fds. Thanks
diff --git a/Documentation/filesystems/porting b/Documentation/filesystems/porting index 93e0a24..17bb4dc 100644 --- a/Documentation/filesystems/porting +++ b/Documentation/filesystems/porting @@ -502,10 +502,6 @@ in your dentry operations instead. store it as cookie. -- [mandatory] - __fd_install() & fd_install() can now sleep. Callers should not - hold a spinlock or other resources that do not allow a schedule. --- -[mandatory] any symlink that might use page_follow_link_light/page_put_link() must have inode_nohighmem(inode) called before anything might start playing with its pagecache. No highmem pages should end up in the pagecache of such diff --git a/fs/file.c b/fs/file.c index 9d047bd..4115503 100644 --- a/fs/file.c +++ b/fs/file.c @@ -592,13 +592,16 @@ void __fd_install(struct files_struct *files, unsigned int fd, { struct fdtable *fdt; - might_sleep(); rcu_read_lock_sched(); - while (unlikely(files->resize_in_progress)) { + if (unlikely(files->resize_in_progress)) { rcu_read_unlock_sched(); - wait_event(files->resize_wait, !files->resize_in_progress); - rcu_read_lock_sched(); + spin_lock(&files->file_lock); + fdt = files_fdtable(files); + BUG_ON(fdt->fd[fd] != NULL); + rcu_assign_pointer(fdt->fd[fd], file); + spin_unlock(&files->file_lock); + return; } /* coupled with smp_wmb() in expand_fdtable() */ smp_rmb();
Explicit locking in the fallback case provides a safe state of the table. Getting rid of blocking semantics makes __fd_install usable again in non-sleepable contexts, which easies backporting efforts. There is a side effect of slightly nicer assembly for the common case as might_sleep can now be removed. Signed-off-by: Mateusz Guzik <mguzik@redhat.com> --- Documentation/filesystems/porting | 4 ---- fs/file.c | 11 +++++++---- 2 files changed, 7 insertions(+), 8 deletions(-)