diff mbox series

[RFC,4/4] UNFINISHED mm, fs: use kmem_cache_charge() in path_openat()

Message ID 20240301-slab-memcg-v1-4-359328a46596@suse.cz
State New
Headers
Series memcg_kmem hooks refactoring and kmem_cache_charge() |

Commit Message

Vlastimil Babka March 1, 2024, 5:07 p.m. UTC 
  This is just an example of using the kmem_cache_charge() API.  I think
it's placed in a place that's applicable for Linus's example [1]
although he mentions do_dentry_open() - I have followed from strace()
showing openat(2) to path_openat() doing the alloc_empty_file().

The idea is that filp_cachep stops being SLAB_ACCOUNT. Allocations that
want to be accounted immediately can use GFP_KERNEL_ACCOUNT. I did that
in alloc_empty_file_noaccount() (despite the contradictory name but the
noaccount refers to something else, right?) as IIUC it's about
kernel-internal opens.

alloc_empty_file() is now not doing the accounting, so I added
kmem_account_file() that calls the new kmem_cache_charge() API.

Why is this unfinished:

- there are other callers of alloc_empty_file() which I didn't adjust so
  they simply became memcg-unaccounted. I haven't investigated for which
  ones it would make also sense to separate the allocation and accounting.
  Maybe alloc_empty_file() would need to get a parameter to control
  this.

- I don't know how to properly unwind the accounting failure case. It
  seems like a new case because when we succeed the open, there's no
  further error path at least in path_openat().

Basically it boils down I'm unfamiliar with VFS so this depends if this
approach is deemed useful enough to finish it.

[1] https://lore.kernel.org/all/CAHk-=whYOOdM7jWy5jdrAm8LxcgCMFyk2bt8fYYvZzM4U-zAQA@mail.gmail.com/

Not-signed-off-by: Vlastimil Babka <vbabka@suse.cz>
---
 fs/file_table.c | 9 +++++++--
 fs/internal.h   | 1 +
 fs/namei.c      | 4 +++-
 3 files changed, 11 insertions(+), 3 deletions(-)

Comments

Linus Torvalds March 1, 2024, 5:51 p.m. UTC | #1 
On Fri, 1 Mar 2024 at 09:07, Vlastimil Babka <vbabka@suse.cz> wrote:
>
> This is just an example of using the kmem_cache_charge() API.  I think
> it's placed in a place that's applicable for Linus's example [1]
> although he mentions do_dentry_open() - I have followed from strace()
> showing openat(2) to path_openat() doing the alloc_empty_file().

Thanks. This is not the right patch,  but yes, patches 1-3 look very nice to me.

> The idea is that filp_cachep stops being SLAB_ACCOUNT. Allocations that
> want to be accounted immediately can use GFP_KERNEL_ACCOUNT. I did that
> in alloc_empty_file_noaccount() (despite the contradictory name but the
> noaccount refers to something else, right?) as IIUC it's about
> kernel-internal opens.

Yeah, the "noaccount" function is about not accounting it towards nr_files.

That said, I don't think it necessarily needs to do the memory
accounting either - it's literally for cases where we're never going
to install the file descriptor in any user space.

Your change to use GFP_KERNEL_ACCOUNT isn't exactly wrong, but I don't
think it's really the right thing either, because

> Why is this unfinished:
>
> - there are other callers of alloc_empty_file() which I didn't adjust so
>   they simply became memcg-unaccounted. I haven't investigated for which
>   ones it would make also sense to separate the allocation and accounting.
>   Maybe alloc_empty_file() would need to get a parameter to control
>   this.

Right. I think the natural and logical way to deal with this is to
just say "we account when we add the file to the fdtable".

IOW, just have fd_install() do it. That's the really natural point,
and also makes it very logical why alloc_empty_file_noaccount()
wouldn't need to do the GFP_KERNEL_ACCOUNT.

> - I don't know how to properly unwind the accounting failure case. It
>   seems like a new case because when we succeed the open, there's no
>   further error path at least in path_openat().

Yeah, let me think about this part. Becasue fd_install() is the right
point, but that too does not really allow for error handling.

Yes, we could close things and fail it, but it really is much too late
at this point.

What I *think* I'd want for this case is

 (a) allow the accounting to go over by a bit

 (b) make sure there's a cheap way to ask (before) about "did we go
over the limit"

IOW, the accounting never needed to be byte-accurate to begin with,
and making it fail (cheaply and early) on the next file allocation is
fine.

Just make it really cheap. Can we do that?

For example, maybe don't bother with the whole "bytes and pages"
stuff. Just a simple "are we more than one page over?" kind of
question. Without the 'stock_lock' mess for sub-page bytes etc

How would that look? Would it result in something that can be done
cheaply without locking and atomics and without excessive pointer
indirection through many levels of memcg data structures?

             Linus
Roman Gushchin March 1, 2024, 6:53 p.m. UTC | #2 
On Fri, Mar 01, 2024 at 09:51:18AM -0800, Linus Torvalds wrote:
> On Fri, 1 Mar 2024 at 09:07, Vlastimil Babka <vbabka@suse.cz> wrote:
> >
> > This is just an example of using the kmem_cache_charge() API.  I think
> > it's placed in a place that's applicable for Linus's example [1]
> > although he mentions do_dentry_open() - I have followed from strace()
> > showing openat(2) to path_openat() doing the alloc_empty_file().
> 
> Thanks. This is not the right patch,  but yes, patches 1-3 look very nice to me.
> 
> > The idea is that filp_cachep stops being SLAB_ACCOUNT. Allocations that
> > want to be accounted immediately can use GFP_KERNEL_ACCOUNT. I did that
> > in alloc_empty_file_noaccount() (despite the contradictory name but the
> > noaccount refers to something else, right?) as IIUC it's about
> > kernel-internal opens.
> 
> Yeah, the "noaccount" function is about not accounting it towards nr_files.
> 
> That said, I don't think it necessarily needs to do the memory
> accounting either - it's literally for cases where we're never going
> to install the file descriptor in any user space.
> 
> Your change to use GFP_KERNEL_ACCOUNT isn't exactly wrong, but I don't
> think it's really the right thing either, because
> 
> > Why is this unfinished:
> >
> > - there are other callers of alloc_empty_file() which I didn't adjust so
> >   they simply became memcg-unaccounted. I haven't investigated for which
> >   ones it would make also sense to separate the allocation and accounting.
> >   Maybe alloc_empty_file() would need to get a parameter to control
> >   this.
> 
> Right. I think the natural and logical way to deal with this is to
> just say "we account when we add the file to the fdtable".
> 
> IOW, just have fd_install() do it. That's the really natural point,
> and also makes it very logical why alloc_empty_file_noaccount()
> wouldn't need to do the GFP_KERNEL_ACCOUNT.
> 
> > - I don't know how to properly unwind the accounting failure case. It
> >   seems like a new case because when we succeed the open, there's no
> >   further error path at least in path_openat().
> 
> Yeah, let me think about this part. Becasue fd_install() is the right
> point, but that too does not really allow for error handling.
> 
> Yes, we could close things and fail it, but it really is much too late
> at this point.
> 
> What I *think* I'd want for this case is
> 
>  (a) allow the accounting to go over by a bit
> 
>  (b) make sure there's a cheap way to ask (before) about "did we go
> over the limit"
> 
> IOW, the accounting never needed to be byte-accurate to begin with,
> and making it fail (cheaply and early) on the next file allocation is
> fine.
> 
> Just make it really cheap. Can we do that?
> 
> For example, maybe don't bother with the whole "bytes and pages"
> stuff. Just a simple "are we more than one page over?" kind of
> question. Without the 'stock_lock' mess for sub-page bytes etc
> 
> How would that look? Would it result in something that can be done
> cheaply without locking and atomics and without excessive pointer
> indirection through many levels of memcg data structures?

I think it's possible and I'm currently looking into batching charge,
objcg refcnt management and vmstats using per-task caching. It should
speed up things for the majority of allocations.
For allocations from an irq context and targeted allocations
(where the target memcg != memcg of the current task) we'd probably need to
keep the old scheme. I hope to post some patches relatively soon.

I tried to optimize the current implementation but failed to get any
significant gains. It seems that the overhead is very evenly spread across
objcg pointer access, charge management, objcg refcnt management and vmstats.

Thanks!
Christian Brauner March 4, 2024, 12:47 p.m. UTC | #3 
On Fri, Mar 01, 2024 at 09:51:18AM -0800, Linus Torvalds wrote:
> On Fri, 1 Mar 2024 at 09:07, Vlastimil Babka <vbabka@suse.cz> wrote:
> >
> > This is just an example of using the kmem_cache_charge() API.  I think
> > it's placed in a place that's applicable for Linus's example [1]
> > although he mentions do_dentry_open() - I have followed from strace()
> > showing openat(2) to path_openat() doing the alloc_empty_file().
> 
> Thanks. This is not the right patch,  but yes, patches 1-3 look very nice to me.
> 
> > The idea is that filp_cachep stops being SLAB_ACCOUNT. Allocations that
> > want to be accounted immediately can use GFP_KERNEL_ACCOUNT. I did that
> > in alloc_empty_file_noaccount() (despite the contradictory name but the
> > noaccount refers to something else, right?) as IIUC it's about
> > kernel-internal opens.
> 
> Yeah, the "noaccount" function is about not accounting it towards nr_files.
> That said, I don't think it necessarily needs to do the memory
> accounting either - it's literally for cases where we're never going
> to install the file descriptor in any user space.

Exactly.

> Your change to use GFP_KERNEL_ACCOUNT isn't exactly wrong, but I don't
> think it's really the right thing either, because
> 
> > Why is this unfinished:
> >
> > - there are other callers of alloc_empty_file() which I didn't adjust so
> >   they simply became memcg-unaccounted. I haven't investigated for which
> >   ones it would make also sense to separate the allocation and accounting.
> >   Maybe alloc_empty_file() would need to get a parameter to control
> >   this.
> 
> Right. I think the natural and logical way to deal with this is to
> just say "we account when we add the file to the fdtable".
> IOW, just have fd_install() do it. That's the really natural point,
> and also makes it very logical why alloc_empty_file_noaccount()
> wouldn't need to do the GFP_KERNEL_ACCOUNT.
> 
> > - I don't know how to properly unwind the accounting failure case. It
> >   seems like a new case because when we succeed the open, there's no
> >   further error path at least in path_openat().
> 
> Yeah, let me think about this part. Becasue fd_install() is the right
> point, but that too does not really allow for error handling.
> 
> Yes, we could close things and fail it, but it really is much too late
> at this point.

It would also mean massaging 100+ callsites. And having a non-subsystems
specific failure step between file allocation, fd reservation and
fd_install() would be awkward and an invitation for bugs.

> What I *think* I'd want for this case is
> 
>  (a) allow the accounting to go over by a bit
> 
>  (b) make sure there's a cheap way to ask (before) about "did we go
> over the limit"
> 
> IOW, the accounting never needed to be byte-accurate to begin with,
> and making it fail (cheaply and early) on the next file allocation is
> fine.

I think that's a good idea.
diff mbox series

Patch

diff --git a/fs/file_table.c b/fs/file_table.c
index b991f90571b4..6401b6f175ae 100644
--- a/fs/file_table.c
+++ b/fs/file_table.c
@@ -223,6 +223,11 @@  struct file *alloc_empty_file(int flags, const struct cred *cred)
 	return ERR_PTR(-ENFILE);
 }
 
+int kmem_account_file(struct file *f)
+{
+	return kmem_cache_charge(filp_cachep, GFP_KERNEL_ACCOUNT, f);
+}
+
 /*
  * Variant of alloc_empty_file() that doesn't check and modify nr_files.
  *
@@ -234,7 +239,7 @@  struct file *alloc_empty_file_noaccount(int flags, const struct cred *cred)
 	struct file *f;
 	int error;
 
-	f = kmem_cache_zalloc(filp_cachep, GFP_KERNEL);
+	f = kmem_cache_zalloc(filp_cachep, GFP_KERNEL_ACCOUNT);
 	if (unlikely(!f))
 		return ERR_PTR(-ENOMEM);
 
@@ -468,7 +473,7 @@  void __init files_init(void)
 {
 	filp_cachep = kmem_cache_create("filp", sizeof(struct file), 0,
 				SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN |
-				SLAB_PANIC | SLAB_ACCOUNT, NULL);
+				SLAB_PANIC, NULL);
 	percpu_counter_init(&nr_files, 0, GFP_KERNEL);
 }
 
diff --git a/fs/internal.h b/fs/internal.h
index b67406435fc0..06ada11b71d0 100644
--- a/fs/internal.h
+++ b/fs/internal.h
@@ -96,6 +96,7 @@  extern void chroot_fs_refs(const struct path *, const struct path *);
 struct file *alloc_empty_file(int flags, const struct cred *cred);
 struct file *alloc_empty_file_noaccount(int flags, const struct cred *cred);
 struct file *alloc_empty_backing_file(int flags, const struct cred *cred);
+int kmem_account_file(struct file *file);
 
 static inline void file_put_write_access(struct file *file)
 {
diff --git a/fs/namei.c b/fs/namei.c
index 4e0de939fea1..fcf3f3fcd059 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -3799,8 +3799,10 @@  static struct file *path_openat(struct nameidata *nd,
 		terminate_walk(nd);
 	}
 	if (likely(!error)) {
-		if (likely(file->f_mode & FMODE_OPENED))
+		if (likely(file->f_mode & FMODE_OPENED)) {
+			kmem_account_file(file);
 			return file;
+		}
 		WARN_ON(1);
 		error = -EINVAL;
 	}