Re: [PATCH] proc: Make inline name size calculation automatic

Re: [PATCH] proc: Make inline name size calculation automatic

[Alexey Dobriyan]

> Require a minimum inline name size of 33+1 to allow for names that look
> like two hex numbers with a dash between.

Hi, David.

Why 34? /proc will fallback to separate allocation for name anyway.

I sent nearly identical patch earlier.
https://marc.info/?l=linux-kernel&m=152667374404900&w=4

If you compare, the differences are:

* no BUILD_BUG_ON,
* 64 bytes is too litle even on 32-bit,
* 512 bytes is probably too much even on 64-bit.

> -	.name		= proc_root.inline_name,
> -	.inline_name	= "/proc",
> +	.name		= "/proc",

This will "uninline" 5 bytes wasting space.

Re: [PATCH] proc: Make inline name size calculation automatic

[David Howells]

Alexey Dobriyan <adobriyan@gmail.com> wrote:

> > Require a minimum inline name size of 33+1 to allow for names that look
> > like two hex numbers with a dash between.
> 
> Why 34? /proc will fallback to separate allocation for name anyway.

See above comment.  I ran find on /proc and there were a bunch of files whose
names were "<hex>-<hex>".  Allow for 16-char hex addresses and add a NUL char
to that and you get 34.

David

Re: [PATCH] proc: Make inline name size calculation automatic

[Alexey Dobriyan]

On Thu, Jun 14, 2018 at 09:30:42PM +0100, David Howells wrote:
> Alexey Dobriyan <adobriyan@gmail.com> wrote:
> 
> > > Require a minimum inline name size of 33+1 to allow for names that look
> > > like two hex numbers with a dash between.
> > 
> > Why 34? /proc will fallback to separate allocation for name anyway.
> 
> See above comment.  I ran find on /proc and there were a bunch of files whose
> names were "<hex>-<hex>".  Allow for 16-char hex addresses and add a NUL char
> to that and you get 34.

Those must be /proc/*/map_files symlinks. If that's the case,
they don't have PDEs allocated.

[PATCH] proc: fixup PDE allocation bloat

[Alexey Dobriyan]

commit 24074a35c5c975c94cd9691ae962855333aac47f
("proc: Make inline name size calculation automatic")
started to put PDE allocations into kmalloc-256 which is unnecessary as
~40 character names are very rare.

Put allocation back into kmalloc-192 cache for 64-bit non-debug builds.

Put BUILD_BUG_ON to know when PDE size is gotten out of control.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
---

 fs/proc/inode.c    |    6 ++++--
 fs/proc/internal.h |   17 +++++++----------
 2 files changed, 11 insertions(+), 12 deletions(-)

--- a/fs/proc/inode.c
+++ b/fs/proc/inode.c
@@ -105,8 +105,10 @@ void __init proc_init_kmemcache(void)
 		kmem_cache_create("pde_opener", sizeof(struct pde_opener), 0,
 				  SLAB_ACCOUNT|SLAB_PANIC, NULL);
 	proc_dir_entry_cache = kmem_cache_create_usercopy(
-		"proc_dir_entry", SIZEOF_PDE_SLOT, 0, SLAB_PANIC,
-		OFFSETOF_PDE_NAME, SIZEOF_PDE_INLINE_NAME, NULL);
+		"proc_dir_entry", SIZEOF_PDE, 0, SLAB_PANIC,
+		offsetof(struct proc_dir_entry, inline_name),
+		SIZEOF_PDE_INLINE_NAME, NULL);
+	BUILD_BUG_ON(sizeof(struct proc_dir_entry) >= SIZEOF_PDE);
 }
 
 static int proc_show_options(struct seq_file *seq, struct dentry *root)
--- a/fs/proc/internal.h
+++ b/fs/proc/internal.h
@@ -65,16 +65,13 @@ struct proc_dir_entry {
 	char inline_name[];
 } __randomize_layout;
 
-#define OFFSETOF_PDE_NAME offsetof(struct proc_dir_entry, inline_name)
-#define SIZEOF_PDE_SLOT					\
-	(OFFSETOF_PDE_NAME + 34 <= 64 ? 64 :		\
-	 OFFSETOF_PDE_NAME + 34 <= 128 ? 128 :		\
-	 OFFSETOF_PDE_NAME + 34 <= 192 ? 192 :		\
-	 OFFSETOF_PDE_NAME + 34 <= 256 ? 256 :		\
-	 OFFSETOF_PDE_NAME + 34 <= 512 ? 512 :		\
-	 0)
-
-#define SIZEOF_PDE_INLINE_NAME (SIZEOF_PDE_SLOT - OFFSETOF_PDE_NAME)
+#ifdef CONFIG_64BIT
+#define SIZEOF_PDE	(sizeof(spinlock_t) <= 4 ? 192 : 256)
+#else
+#define SIZEOF_PDE	(sizeof(spinlock_t) <= 4 ? 128 : 192)
+#endif
+
+#define SIZEOF_PDE_INLINE_NAME (SIZEOF_PDE - sizeof(struct proc_dir_entry))
 
 extern struct kmem_cache *proc_dir_entry_cache;
 void pde_free(struct proc_dir_entry *pde);

Re: [PATCH] proc: fixup PDE allocation bloat

[Shakeel Butt]

On Sun, Jun 17, 2018 at 2:57 PM Alexey Dobriyan <adobriyan@gmail.com> wrote:
>
> commit 24074a35c5c975c94cd9691ae962855333aac47f
> ("proc: Make inline name size calculation automatic")
> started to put PDE allocations into kmalloc-256 which is unnecessary as
> ~40 character names are very rare.
>
> Put allocation back into kmalloc-192 cache for 64-bit non-debug builds.
>
> Put BUILD_BUG_ON to know when PDE size is gotten out of control.
>
> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> ---
>
>  fs/proc/inode.c    |    6 ++++--
>  fs/proc/internal.h |   17 +++++++----------
>  2 files changed, 11 insertions(+), 12 deletions(-)
>
> --- a/fs/proc/inode.c
> +++ b/fs/proc/inode.c
> @@ -105,8 +105,10 @@ void __init proc_init_kmemcache(void)
>                 kmem_cache_create("pde_opener", sizeof(struct pde_opener), 0,
>                                   SLAB_ACCOUNT|SLAB_PANIC, NULL);
>         proc_dir_entry_cache = kmem_cache_create_usercopy(
> -               "proc_dir_entry", SIZEOF_PDE_SLOT, 0, SLAB_PANIC,
> -               OFFSETOF_PDE_NAME, SIZEOF_PDE_INLINE_NAME, NULL);
> +               "proc_dir_entry", SIZEOF_PDE, 0, SLAB_PANIC,

Hi Alexey, can you comment if proc_dir_entry_cache should or shouldn't
have SLAB_ACCOUNT flag?

Shakeel

Re: [PATCH] proc: fixup PDE allocation bloat

[Alexey Dobriyan]

On Thu, Jul 19, 2018 at 05:06:55PM -0700, Shakeel Butt wrote:
> On Sun, Jun 17, 2018 at 2:57 PM Alexey Dobriyan <adobriyan@gmail.com> wrote:
> >
> > commit 24074a35c5c975c94cd9691ae962855333aac47f
> > ("proc: Make inline name size calculation automatic")
> > started to put PDE allocations into kmalloc-256 which is unnecessary as
> > ~40 character names are very rare.
> >
> > Put allocation back into kmalloc-192 cache for 64-bit non-debug builds.
> >
> > Put BUILD_BUG_ON to know when PDE size is gotten out of control.
> >
> > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > ---
> >
> >  fs/proc/inode.c    |    6 ++++--
> >  fs/proc/internal.h |   17 +++++++----------
> >  2 files changed, 11 insertions(+), 12 deletions(-)
> >
> > --- a/fs/proc/inode.c
> > +++ b/fs/proc/inode.c
> > @@ -105,8 +105,10 @@ void __init proc_init_kmemcache(void)
> >                 kmem_cache_create("pde_opener", sizeof(struct pde_opener), 0,
> >                                   SLAB_ACCOUNT|SLAB_PANIC, NULL);
> >         proc_dir_entry_cache = kmem_cache_create_usercopy(
> > -               "proc_dir_entry", SIZEOF_PDE_SLOT, 0, SLAB_PANIC,
> > -               OFFSETOF_PDE_NAME, SIZEOF_PDE_INLINE_NAME, NULL);
> > +               "proc_dir_entry", SIZEOF_PDE, 0, SLAB_PANIC,
> 
> Hi Alexey, can you comment if proc_dir_entry_cache should or shouldn't
> have SLAB_ACCOUNT flag?

It should not (but see below):

SLAB_ACCOUNT is for allocations which can be done by userspace directly:
open(2) directly allocates "struct file".

But /proc entries aren't like that: say, /proc/cpuinfo is created by kernel
and userspace can't do anything about it.

Some subsystems create /proc entries based on userspace actions and
those aren't related to hardware (example: xt_hashlimit.c) but those are
few so kernel doesn't bother accounting those.

Or in other words: user can't mkdir(1) and touch(1) and ln(1) inside /proc
at will and therefore PDEs aren't accounted.