D54572.id169499.diff
No OneTemporary
Actions

Size

14 KB

Referenced Files

None

Subscribers

None

D54572.id169499.diff
View Options

	diff --git a/share/man/man7/tuning.7 b/share/man/man7/tuning.7
	--- a/share/man/man7/tuning.7
	+++ b/share/man/man7/tuning.7
	@@ -222,6 +222,20 @@
	.Va vm.stats.vm.v_wire_count
	sysctls, respectively).
	.Pp
	+Due to the architecture of the
	+.Fx
	+virtual memory subsystem, the use of copy on write (CoW) anonymous
	+memory, e.g. on
	+.Xr fork 2 ,
	+causes swap reservation for all three regions (VM objects),
	+as in the original pre-fork mapping, and its copies in
	+the parent and child, instead of only two.
	+Eventually the subsystem tries to optimize the internal layout
	+of the tracking for CoW and often removes (collapses) no longer
	+needed backing objects, re-assigning its pages and swap
	+reservations to the copies.
	+This frees the swap reserve, but it is not guaranteed to happen.
	+.Pp
	The
	.Va kern.ipc.maxpipekva
	loader tunable is used to set a hard limit on the
	diff --git a/sys/dev/md/md.c b/sys/dev/md/md.c
	--- a/sys/dev/md/md.c
	+++ b/sys/dev/md/md.c
	@@ -1617,7 +1617,6 @@
	0, 0);
	swap_release_by_cred(IDX_TO_OFF(oldpages -
	newpages), sc->cred);
	- sc->s_swap.object->charge = IDX_TO_OFF(newpages);
	sc->s_swap.object->size = newpages;
	VM_OBJECT_WUNLOCK(sc->s_swap.object);
	} else if (newpages > oldpages) {
	@@ -1637,7 +1636,6 @@
	}
	}
	VM_OBJECT_WLOCK(sc->s_swap.object);
	- sc->s_swap.object->charge = IDX_TO_OFF(newpages);
	sc->s_swap.object->size = newpages;
	VM_OBJECT_WUNLOCK(sc->s_swap.object);
	}
	diff --git a/sys/kern/uipc_shm.c b/sys/kern/uipc_shm.c
	--- a/sys/kern/uipc_shm.c
	+++ b/sys/kern/uipc_shm.c
	@@ -759,7 +759,6 @@

	/* Free the swap accounted for shm */
	swap_release_by_cred(delta, object->cred);
	- object->charge -= delta;
	} else {
	if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
	return (EPERM);
	@@ -768,7 +767,6 @@
	delta = IDX_TO_OFF(nobjsize - object->size);
	if (!swap_reserve_by_cred(delta, object->cred))
	return (ENOMEM);
	- object->charge += delta;
	}
	shmfd->shm_size = length;
	mtx_lock(&shm_timestamp_lock);
	diff --git a/sys/vm/swap_pager.c b/sys/vm/swap_pager.c
	--- a/sys/vm/swap_pager.c
	+++ b/sys/vm/swap_pager.c
	@@ -239,7 +239,8 @@
	#ifdef INVARIANTS
	prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
	KASSERT(prev >= pdecr,
	- ("negative vmsize for uid %d\n", uip->ui_uid));
	+ ("negative vmsize for uid %d, prev %#jx decr %#jx\n",
	+ uip->ui_uid, (uintmax_t)prev, (uintmax_t)pdecr));
	#else
	atomic_subtract_long(&uip->ui_vmsize, pdecr);
	#endif
	@@ -329,7 +330,7 @@
	}

	void
	-swap_reserve_force(vm_ooffset_t incr)
	+swap_reserve_force_by_cred(vm_ooffset_t incr, struct ucred *cred)
	{
	u_long pincr;

	@@ -345,7 +346,13 @@
	#endif
	pincr = atop(incr);
	atomic_add_long(&swap_reserved, pincr);
	- swap_reserve_force_rlimit(pincr, curthread->td_ucred);
	+ swap_reserve_force_rlimit(pincr, cred);
	+}
	+
	+void
	+swap_reserve_force(vm_ooffset_t incr)
	+{
	+ swap_reserve_force_by_cred(incr, curthread->td_ucred);
	}

	void
	@@ -373,7 +380,8 @@
	pdecr = atop(decr);
	#ifdef INVARIANTS
	prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
	- KASSERT(prev >= pdecr, ("swap_reserved < decr"));
	+ KASSERT(prev >= pdecr, ("swap_reserved %#jx < decr %#jx",
	+ (uintmax_t)prev, (uintmax_t)pdecr));
	#else
	atomic_subtract_long(&swap_reserved, pdecr);
	#endif
	@@ -776,10 +784,7 @@

	object->un_pager.swp.writemappings = 0;
	object->handle = handle;
	- if (cred != NULL) {
	- object->cred = cred;
	- object->charge = size;
	- }
	+ object->cred = cred;
	return (true);
	}

	@@ -892,8 +897,7 @@
	* Release the allocation charge.
	*/
	if (object->cred != NULL) {
	- swap_release_by_cred(object->charge, object->cred);
	- object->charge = 0;
	+ swap_release_by_cred(ptoa(object->size), object->cred);
	crfree(object->cred);
	object->cred = NULL;
	}
	diff --git a/sys/vm/vm.h b/sys/vm/vm.h
	--- a/sys/vm/vm.h
	+++ b/sys/vm/vm.h
	@@ -168,6 +168,7 @@
	bool swap_reserve(vm_ooffset_t incr);
	bool swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred);
	void swap_reserve_force(vm_ooffset_t incr);
	+void swap_reserve_force_by_cred(vm_ooffset_t incr, struct ucred *cred);
	void swap_release(vm_ooffset_t decr);
	void swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred);

	diff --git a/sys/vm/vm_fault.c b/sys/vm/vm_fault.c
	--- a/sys/vm/vm_fault.c
	+++ b/sys/vm/vm_fault.c
	@@ -2310,13 +2310,12 @@
	* directly.
	*/
	dst_object = vm_object_allocate_anon(atop(dst_entry->end -
	- dst_entry->start), NULL, NULL, 0);
	+ dst_entry->start), NULL, NULL);
	#if VM_NRESERVLEVEL > 0
	dst_object->flags \|= OBJ_COLORED;
	dst_object->pg_color = atop(dst_entry->start);
	#endif
	dst_object->domain = src_object->domain;
	- dst_object->charge = dst_entry->end - dst_entry->start;

	dst_entry->object.vm_object = dst_object;
	dst_entry->offset = 0;
	@@ -2329,7 +2328,7 @@
	("vm_fault_copy_entry: leaked swp charge"));
	dst_object->cred = curthread->td_ucred;
	crhold(dst_object->cred);
	- *fork_charge += dst_object->charge;
	+ *fork_charge += ptoa(dst_object->size);
	} else if ((dst_object->flags & OBJ_SWAP) != 0 &&
	dst_object->cred == NULL) {
	KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
	diff --git a/sys/vm/vm_map.c b/sys/vm/vm_map.c
	--- a/sys/vm/vm_map.c
	+++ b/sys/vm/vm_map.c
	@@ -2428,7 +2428,7 @@
	KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
	("map entry %p is a submap", entry));
	object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL,
	- entry->cred, entry->end - entry->start);
	+ entry->cred);
	entry->object.vm_object = object;
	entry->offset = 0;
	entry->cred = NULL;
	@@ -2443,21 +2443,26 @@
	static inline void
	vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry)
	{
	+ vm_object_t object;

	VM_MAP_ASSERT_LOCKED(map);
	KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
	("map entry %p is a submap", entry));
	- if (entry->object.vm_object == NULL && !vm_map_is_system(map) &&
	+ object = entry->object.vm_object;
	+ if (object == NULL && !vm_map_is_system(map) &&
	(entry->eflags & MAP_ENTRY_GUARD) == 0)
	vm_map_entry_back(entry);
	- else if (entry->object.vm_object != NULL &&
	+ else if (object != NULL &&
	((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) &&
	entry->cred != NULL) {
	- VM_OBJECT_WLOCK(entry->object.vm_object);
	- KASSERT(entry->object.vm_object->cred == NULL,
	+ VM_OBJECT_WLOCK(object);
	+ KASSERT(object->cred == NULL,
	("OVERCOMMIT: %s: both cred e %p", __func__, entry));
	- entry->object.vm_object->cred = entry->cred;
	- entry->object.vm_object->charge = entry->end - entry->start;
	+ object->cred = entry->cred;
	+ if (entry->end - entry->start < ptoa(object->size)) {
	+ swap_reserve_force_by_cred(ptoa(object->size) -
	+ entry->end + entry->start, object->cred);
	+ }
	VM_OBJECT_WUNLOCK(entry->object.vm_object);
	entry->cred = NULL;
	}
	@@ -2956,7 +2961,7 @@
	* we cannot distinguish between non-charged and
	* charged clipped mapping of the same object later.
	*/
	- KASSERT(obj->charge == 0,
	+ KASSERT(obj->cred == NULL,
	("vm_map_protect: object %p overcharged (entry %p)",
	obj, entry));
	if (!swap_reserve(ptoa(obj->size))) {
	@@ -2968,7 +2973,6 @@

	crhold(cred);
	obj->cred = cred;
	- obj->charge = ptoa(obj->size);
	VM_OBJECT_WUNLOCK(obj);
	}

	@@ -3942,7 +3946,7 @@
	vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry)
	{
	vm_object_t object;
	- vm_pindex_t offidxstart, offidxend, size1;
	+ vm_pindex_t offidxstart, offidxend, oldsize;
	vm_size_t size;

	vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE);
	@@ -3989,15 +3993,11 @@
	OBJPR_NOTMAPPED);
	if (offidxend >= object->size &&
	offidxstart < object->size) {
	- size1 = object->size;
	+ oldsize = object->size;
	object->size = offidxstart;
	if (object->cred != NULL) {
	- size1 -= object->size;
	- KASSERT(object->charge >= ptoa(size1),
	- ("object %p charge < 0", object));
	- swap_release_by_cred(ptoa(size1),
	- object->cred);
	- object->charge -= ptoa(size1);
	+ swap_release_by_cred(ptoa(oldsize -
	+ ptoa(object->size)), object->cred);
	}
	}
	}
	@@ -4198,7 +4198,7 @@
	("OVERCOMMIT: vm_map_copy_anon_entry: cred %p",
	src_object));
	src_object->cred = src_entry->cred;
	- src_object->charge = size;
	+ *fork_charge += ptoa(src_object->size) - size;
	}
	dst_entry->object.vm_object = src_object;
	if (charged) {
	@@ -4455,7 +4455,7 @@
	KASSERT(object->cred == NULL,
	("vmspace_fork both cred"));
	object->cred = old_entry->cred;
	- object->charge = old_entry->end -
	+ *fork_charge += old_entry->end -
	old_entry->start;
	old_entry->cred = NULL;
	}
	@@ -4957,6 +4957,13 @@
	if (newvmspace == NULL)
	return (ENOMEM);
	if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) {
	+ /*
	+ * The swap reservation failed. The accounting from
	+ * the entries of the copied newvmspace will be
	+ * subtracted in vmspace_free(), so force the
	+ * reservation there.
	+ */
	+ swap_reserve_force_by_cred(fork_charge, p->p_ucred);
	vmspace_free(newvmspace);
	return (ENOMEM);
	}
	@@ -5138,7 +5145,7 @@
	if (vm_map_lock_upgrade(map))
	goto RetryLookup;
	entry->object.vm_object = vm_object_allocate_anon(atop(size),
	- NULL, entry->cred, size);
	+ NULL, entry->cred);
	entry->offset = 0;
	entry->cred = NULL;
	vm_map_lock_downgrade(map);
	@@ -5396,9 +5403,8 @@
	(void *)entry->object.vm_object,
	(uintmax_t)entry->offset);
	if (entry->object.vm_object && entry->object.vm_object->cred)
	- db_printf(", obj ruid %d charge %jx",
	- entry->object.vm_object->cred->cr_ruid,
	- (uintmax_t)entry->object.vm_object->charge);
	+ db_printf(", obj ruid %d ",
	+ entry->object.vm_object->cred->cr_ruid);
	if (entry->eflags & MAP_ENTRY_COW)
	db_printf(", copy (%s)",
	(entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
	diff --git a/sys/vm/vm_object.h b/sys/vm/vm_object.h
	--- a/sys/vm/vm_object.h
	+++ b/sys/vm/vm_object.h
	@@ -175,7 +175,6 @@
	} phys;
	} un_pager;
	struct ucred *cred;
	- vm_ooffset_t charge;
	void *umtx_data;
	};

	@@ -356,8 +355,7 @@
	extern int umtx_shm_vnobj_persistent;

	vm_object_t vm_object_allocate (objtype_t, vm_pindex_t);
	-vm_object_t vm_object_allocate_anon(vm_pindex_t, vm_object_t, struct ucred *,
	- vm_size_t);
	+vm_object_t vm_object_allocate_anon(vm_pindex_t, vm_object_t, struct ucred *);
	vm_object_t vm_object_allocate_dyn(objtype_t, vm_pindex_t, u_short);
	boolean_t vm_object_coalesce(vm_object_t, vm_ooffset_t, vm_size_t, vm_size_t,
	int);
	diff --git a/sys/vm/vm_object.c b/sys/vm/vm_object.c
	--- a/sys/vm/vm_object.c
	+++ b/sys/vm/vm_object.c
	@@ -196,9 +196,9 @@
	KASSERT(object->type == OBJT_DEAD,
	("object %p has non-dead type %d",
	object, object->type));
	- KASSERT(object->charge == 0 && object->cred == NULL,
	- ("object %p has non-zero charge %ju (%p)",
	- object, (uintmax_t)object->charge, object->cred));
	+ KASSERT(object->cred == NULL,
	+ ("object %p has non-zero charge cred %p",
	+ object, object->cred));
	}
	#endif

	@@ -254,7 +254,6 @@
	refcount_init(&object->ref_count, 1);
	object->memattr = VM_MEMATTR_DEFAULT;
	object->cred = NULL;
	- object->charge = 0;
	object->handle = handle;
	object->backing_object = NULL;
	object->backing_object_offset = (vm_ooffset_t) 0;
	@@ -452,7 +451,7 @@
	*/
	vm_object_t
	vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
	- struct ucred *cred, vm_size_t charge)
	+ struct ucred *cred)
	{
	vm_object_t handle, object;

	@@ -466,7 +465,6 @@
	_vm_object_allocate(OBJT_SWAP, size,
	OBJ_ANON \| OBJ_ONEMAPPING \| OBJ_SWAP, object, handle);
	object->cred = cred;
	- object->charge = cred != NULL ? charge : 0;
	return (object);
	}

	@@ -1448,7 +1446,7 @@
	/*
	* Allocate a new object with the given length.
	*/
	- result = vm_object_allocate_anon(atop(length), source, cred, length);
	+ result = vm_object_allocate_anon(atop(length), source, cred);

	/*
	* Store the offset into the source object, and fix up the offset into
	@@ -1511,6 +1509,7 @@
	struct pctrie_iter pages;
	vm_page_t m;
	vm_object_t orig_object, new_object, backing_object;
	+ struct ucred *cred;
	vm_pindex_t offidxstart;
	vm_size_t size;

	@@ -1525,9 +1524,26 @@

	offidxstart = OFF_TO_IDX(entry->offset);
	size = atop(entry->end - entry->start);
	+ if (orig_object->cred != NULL) {
	+ /*
	+ * vm_object_split() is currently called from
	+ * vmspace_fork(), and it might be tempting to add the
	+ * charge for the split object to fork_charge. But
	+ * fork_charge is discharged on error when the copied
	+ * vmspace is destroyed. Since the split object is
	+ * inserted into the shadow hierarchy serving the
	+ * source vm_map, it is kept even after the
	+ * unsuccessful fork, meaning that we have to force
	+ * its swap usage.
	+ */
	+ cred = curthread->td_ucred;
	+ crhold(cred);
	+ swap_reserve_force_by_cred(ptoa(size), cred);
	+ } else {
	+ cred = NULL;
	+ }

	- new_object = vm_object_allocate_anon(size, orig_object,
	- orig_object->cred, ptoa(size));
	+ new_object = vm_object_allocate_anon(size, orig_object, cred);

	/*
	* We must wait for the orig_object to complete any in-progress
	@@ -1550,12 +1566,6 @@
	new_object->backing_object_offset =
	orig_object->backing_object_offset + entry->offset;
	}
	- if (orig_object->cred != NULL) {
	- crhold(orig_object->cred);
	- KASSERT(orig_object->charge >= ptoa(size),
	- ("orig_object->charge < 0"));
	- orig_object->charge -= ptoa(size);
	- }

	/*
	* Mark the split operation so that swap_pager_getpages() knows
	@@ -2233,7 +2243,6 @@
	swap_release_by_cred(ptoa(prev_object->size -
	next_pindex), prev_object->cred);
	}
	- prev_object->charge += charge;
	} else if ((cflags & OBJCO_CHARGED) != 0) {
	/*
	* The caller charged, but the object has
	@@ -2786,9 +2795,8 @@
	db_iprintf("Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x",
	object, (int)object->type, (uintmax_t)object->size,
	object->resident_page_count, object->ref_count, object->flags);
	- db_iprintf(" ruid %d charge %jx\n",
	- object->cred ? object->cred->cr_ruid : -1,
	- (uintmax_t)object->charge);
	+ db_iprintf(" ruid %d\n",
	+ object->cred ? object->cred->cr_ruid : -1);
	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
	atomic_load_int(&object->shadow_count),
	object->backing_object ? object->backing_object->ref_count : 0,

File Metadata

Mime Type: text/plain
Expires: Sun, Jan 18, 1:19 PM (8 h, 45 m)
Storage Engine: blob
Storage Format: Raw Data
Storage Handle: 27702847
Default Alt Text: D54572.id169499.diff (14 KB)

D54572.id169499.diffNo OneTemporaryActions

D54572.id169499.diffView Options

File Metadata

Event Timeline

D54572.id169499.diff
No OneTemporary
Actions

D54572.id169499.diff
View Options