Re: [PATCH] - support inheritance of mlocks across fork/exec V2

Sorry for the resend.  I had the 'linux-api' address wrong on my reponse
to Andrew.  Want to include the api list in this exchange.

On Sat, 2008-12-06 at 22:07 -0800, Andrew Morton wrote:

(cc linux-abi.  Again.)

On Wed, 03 Dec 2008 14:04:29 -0500 Lee Schermerhorn [off-list ref] wrote:

quoted

Against;  2.6.28-rc7-mmotm-081203

V02:	rework vetting of flags argument as suggested by
	Kosaki Motohiro.
	enhance description as requested by Andrew Morton.

Add support for mlockall(MCL_INHERIT|MCL_RECURSIVE):

	MCL_CURRENT[|MCL_FUTURE]|MCL_INHERIT - inherit memory locks
		[vmas' VM_LOCKED flags] across fork(), and inherit
		MCL_FUTURE behavior [mm's def_flags] across fork()
		and exec().  Behaves as if child and/or new task
		called mlockall(MCL_CURRENT|MCL_FUTURE) as first
		instruction.

	MCL_RECURSIVE - inherit MCL_CURRENT|MCL_FUTURE|MCL_INHERIT
		[vmas' VM_LOCKED flags for fork() and mm's def_flags
		and mcl_inherit across fork() and exec()] for all
		future generations of calling process's descendants.
		Behaves as if child and/or new task called
		mlockall(MCL_CURRENT|MCL_FUTURE|MCL_INHERIT|MCL_RECURSIVE)
		as the first instruction.

In support of a "lock prefix command"--e.g., mlock <cmd> <args> ...
Analogous to taskset(1) for cpu affinity or numactl(8) for numa memory
policy.

Together with patches to keep mlocked pages off the LRU, this will
allow users/admins to lock down applications without modifying them,
if their RLIMIT_MEMLOCK is sufficiently large, keeping their pages
off the LRU and out of consideration for reclaim.

Potentially useful, as well, in real-time environments to force
prefaulting and residency for applications that don't mlock themselves.

Jeff Sharkey at Montana State developed a similar patch for Linux
[link no longer accessible], but apparently he never submitted the patch.

I submitted an earlier version of this patch around a year ago.  I
resurrected it to test the unevictable lru/mlocked pages patches--
e.g., by "mlock -r make -j<N*nr_cpus> all".  This did shake out a few
races and vmstat accounting bugs, but NOT something I'd recommend as
general practice--for kernel builds, that is.

----

Define MCL_INHERIT, MCL_RECURSIVE in <asm-*/mman.h>.
+ x86 and  ia64 versions included.
+ other arch can/will be created, if this patch deemed merge-worthy.

Similarly, I'll provide kernel man page update if/when needed.

Example "lock prefix command" in Documentation/vm/mlock.c


...

--- linux-2.6.28-rc7-mmotm-081203.orig/mm/mlock.c	2008-12-03 10:33:11.000000000 -0500
+++ linux-2.6.28-rc7-mmotm-081203/mm/mlock.c	2008-12-03 10:33:29.000000000 -0500

@@ -573,15 +573,18 @@ asmlinkage long sys_munlock(unsigned lon
 static int do_mlockall(int flags)
 {
 	struct vm_area_struct * vma, * prev = NULL;
+	struct mm_struct *mm = current->mm;
 	unsigned int def_flags = 0;
 
 	if (flags & MCL_FUTURE)
-		def_flags = VM_LOCKED;
-	current->mm->def_flags = def_flags;
-	if (flags == MCL_FUTURE)
+		def_flags = VM_LOCKED;;

You get paid by the semicolon?

Nope.  Interestingly, ckeckpatch was fine with this.  [It did warn about

80 char line in x86 mman.h, but there was already a much longer line,

so I left it.  Fixed now so that patch is squeaky clean, as far as
checkpatch is concerned.]

quoted

+	mm->def_flags = def_flags;
+	if (flags & MCL_INHERIT)
+		mm->mcl_inherit = flags & (MCL_INHERIT | MCL_RECURSIVE);

aaaaaaarrrrrrggggggghhhhhhh!

So mm->mcl_inherit is _not_ a boolean meaning "this mm has MCL_INHERIT
set", as I naively believed when I saw it.  Is it in fact a composite
of both MCL_INHERIT and MCL_RECURSIVE, given a badly misleading name
and a waffly comment.

Can we please make this clearer?

How about "mcl_inherit_mlocks_possibly_recursively" ?

Seriously, it made/makes sense to me.  If mcl_inherit is non-zero, it
WILL contain MCL_INHERIT, meaning that existing mlocks and any VM_LOCKED
in def_flags will be inherited across the next fork or exec [def_flags
only].  It may also contain MCL_RECURSIVE--tested separately--meaning
that the contents of mcl_inherit, itself, will also be inherited across
the next and future forks/execs, until/unless a descendant invokes
mlockall() w/o MCL_INHERIT or MCL_RECURSIVE to clear out mcl_inherit.

This seemed a bit much to try to explain on line where mcl_inherit is
declared in mm_types.  More appropriate to a man page, I would have
thought.   How about a Documentation/vm doc?

Meanwhile, perhaps:  "mcl_inherit" -> "mcl_inherit_flags" or "mcl_flags"
or "mlock_flags", or such?

quoted

...

--- linux-2.6.28-rc7-mmotm-081203.orig/kernel/fork.c	2008-12-03 10:18:15.000000000 -0500
+++ linux-2.6.28-rc7-mmotm-081203/kernel/fork.c	2008-12-03 10:33:29.000000000 -0500

@@ -278,7 +278,8 @@ static int dup_mmap(struct mm_struct *mm
 	 */
 	down_write_nested(&mm->mmap_sem, SINGLE_DEPTH_NESTING);
 
-	mm->locked_vm = 0;
+	if (!mm->mcl_inherit)
+		mm->locked_vm = 0;
 	mm->mmap = NULL;
 	mm->mmap_cache = NULL;
 	mm->free_area_cache = oldmm->mmap_base;

@@ -316,7 +317,8 @@ static int dup_mmap(struct mm_struct *mm
 		if (IS_ERR(pol))
 			goto fail_nomem_policy;
 		vma_set_policy(tmp, pol);
-		tmp->vm_flags &= ~VM_LOCKED;
+		if (!mm->mcl_inherit)
+			tmp->vm_flags &= ~VM_LOCKED;
 		tmp->vm_mm = mm;
 		tmp->vm_next = NULL;
 		anon_vma_link(tmp);

@@ -406,6 +408,8 @@ __cacheline_aligned_in_smp DEFINE_SPINLO
 
 static struct mm_struct * mm_init(struct mm_struct * mm, struct task_struct *p)
 {
+	unsigned long def_flags = 0;
+
 	atomic_set(&mm->mm_users, 1);
 	atomic_set(&mm->mm_count, 1);
 	init_rwsem(&mm->mmap_sem);

@@ -422,9 +426,14 @@ static struct mm_struct * mm_init(struct
 	mm->free_area_cache = TASK_UNMAPPED_BASE;
 	mm->cached_hole_size = ~0UL;
 	mm_init_owner(mm, p);
+	if (current->mm && current->mm->mcl_inherit) {
+		def_flags = current->mm->def_flags & VM_LOCKED;
+		if (mm->mcl_inherit & MCL_RECURSIVE)
+			mm->mcl_inherit  = current->mm->mcl_inherit;
+	}

hm.  That code looks familiar.  Not worth a helper function?

If we could remove the copy in load_elf_binary(), point would be moot.

More below.

quoted

 	if (likely(!mm_alloc_pgd(mm))) {
-		mm->def_flags = 0;
+		mm->def_flags = def_flags;
 		mmu_notifier_mm_init(mm);
 		return mm;
 	}
Index: linux-2.6.28-rc7-mmotm-081203/fs/binfmt_elf.c
===================================================================

--- linux-2.6.28-rc7-mmotm-081203.orig/fs/binfmt_elf.c	2008-12-03 10:19:21.000000000 -0500
+++ linux-2.6.28-rc7-mmotm-081203/fs/binfmt_elf.c	2008-12-03 10:33:29.000000000 -0500

@@ -585,6 +585,7 @@ static int load_elf_binary(struct linux_
 	unsigned long reloc_func_desc = 0;
 	int executable_stack = EXSTACK_DEFAULT;
 	unsigned long def_flags = 0;
+	int mcl_inherit = 0;
 	struct {
 		struct elfhdr elf_ex;
 		struct elfhdr interp_elf_ex;

@@ -749,6 +750,13 @@ static int load_elf_binary(struct linux_
 		SET_PERSONALITY(loc->elf_ex);
 	}
 
+	/* Optionally inherit MCL_FUTURE state before destroying old mm */
+	if (current->mm && current->mm->mcl_inherit) {

umm, OK, it looks like current->mm can be null here.

Probably not here.  IIRC, it can be NULL [for init] up in mm_init() and
I probably just cut and pasted, instead of a helper function...

quoted

+		def_flags = current->mm->def_flags & VM_LOCKED;

ooh, we finally used local variable def_flags for something.  That's
been wasting space since 2.2.26 (or earlier).

def_flags was being used to clear out the current->mm->def_flags.  Don't
know why, tho'.  See below...

quoted

+		if (current->mm->mcl_inherit & MCL_RECURSIVE)
+			mcl_inherit  = current->mm->mcl_inherit;
+	}
+
 	/* Flush all traces of the currently running executable */
 	retval = flush_old_exec(bprm);
 	if (retval)

@@ -757,6 +765,7 @@ static int load_elf_binary(struct linux_
 	/* OK, This is the point of no return */
 	current->flags &= ~PF_FORKNOEXEC;
 	current->mm->def_flags = def_flags;
+	current->mm->mcl_inherit = mcl_inherit;
 
 	/* Do this immediately, since STACK_TOP as used in setup_arg_pages
 	   may depend on the personality.  */

Why is this done in binfmt_elf.c?  Doesn't this preclude the operation
of these new flags for other binary formats?

Because binfmt_elf.c is the only binary format loader [that I can see]
that takes it upon itself to reinitialize mm->def_flags, which has
already been set up in mm_init().  I wondered, myself, why this is the
case, but I didn't dare just remove it, lest it reintroduce some obscure
bug.  So, I duplicated my version of def_flags [and mcl_inherit]
initialization here. 

Now that you've raised the question:  why, indeed, does the elf format
loader reinit def_flags?  Can we just remove this?

quoted

Index: linux-2.6.28-rc7-mmotm-081203/arch/x86/include/asm/mman.h
===================================================================

--- linux-2.6.28-rc7-mmotm-081203.orig/arch/x86/include/asm/mman.h	2008-12-03 10:16:26.000000000 -0500
+++ linux-2.6.28-rc7-mmotm-081203/arch/x86/include/asm/mman.h	2008-12-03 10:33:29.000000000 -0500

@@ -16,5 +16,8 @@
 
 #define MCL_CURRENT	1		/* lock all current mappings */
 #define MCL_FUTURE	2		/* lock all future mappings */
+#define MCL_INHERIT	4		/* inherit mlocks across fork */
+					/* inherit '_FUTURE flag across fork/exec */
+#define MCL_RECURSIVE	8		/* inherit mlocks recursively */
 
 #endif /* _ASM_X86_MMAN_H */

Index: linux-2.6.28-rc7-mmotm-081203/include/linux/mm_types.h
===================================================================

--- linux-2.6.28-rc7-mmotm-081203.orig/include/linux/mm_types.h	2008-12-03 10:18:01.000000000 -0500
+++ linux-2.6.28-rc7-mmotm-081203/include/linux/mm_types.h	2008-12-03 10:33:29.000000000 -0500

@@ -235,6 +235,8 @@ struct mm_struct {
 	unsigned int token_priority;
 	unsigned int last_interval;
 
+	int mcl_inherit;		/* inherit current/future locks */
+
 	unsigned long flags; /* Must use atomic bitops to access the bits */

I was thinking that the boolean mcl_inherit could become a bit in
mm.flags.  That was before I unmisled myself about it.

Perhaps we could use two bits.  That might get a bit nasty.

I did look for someplace to add these flags, before deciding on a new
member.  The "Must use atomic bitops", and the fact that mm->flags seems
used soley for controlling core dumping [and defined in sched.h, no
less!], put me off from going that route.  Similarly, because def_flags
is an initializer for the vma's vm_flags, I didn't want to pollute that.
I figured the 'int' would fill a padding slot on 64-bit archs, but does
add to mm size for 32-bit archs.

Suggestions?
I'll hold off on a re-spin, pending resolution of some of these issues.

Thanks for the thorough review.

Lee


--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html