diff mbox series

[RFC,01/19] mm: Introduce vm_account

Message ID 748338ffe4c42d86669923159fe0426808ecb04d.1674538665.git-series.apopple@nvidia.com (mailing list archive)
State RFC
Headers show
Series [RFC,01/19] mm: Introduce vm_account | expand

Commit Message

Alistair Popple Jan. 24, 2023, 5:42 a.m. UTC
Kernel drivers that pin pages should account these pages against
either user->locked_vm or mm->pinned_vm and fail the pinning if
RLIMIT_MEMLOCK is exceeded and CAP_IPC_LOCK isn't held.

Currently drivers open-code this accounting and use various methods to
update the atomic variables and check against the limits leading to
various bugs and inconsistencies. To fix this introduce a standard
interface for charging pinned and locked memory. As this involves
taking references on kernel objects such as mm_struct or user_struct
we introduce a new vm_account struct to hold these references. Several
helper functions are then introduced to grab references and check
limits.

As the way these limits are charged and enforced is visible to
userspace we need to be careful not to break existing applications by
charging to different counters. As a result the vm_account functions
support accounting to different counters as required.

A future change will extend this to also account against a cgroup for
pinned pages.

Signed-off-by: Alistair Popple <apopple@nvidia.com>
Cc: linux-kernel@vger.kernel.org
Cc: linuxppc-dev@lists.ozlabs.org
Cc: linux-fpga@vger.kernel.org
Cc: linux-rdma@vger.kernel.org
Cc: virtualization@lists.linux-foundation.org
Cc: kvm@vger.kernel.org
Cc: netdev@vger.kernel.org
Cc: cgroups@vger.kernel.org
Cc: io-uring@vger.kernel.org
Cc: linux-mm@kvack.org
Cc: bpf@vger.kernel.org
Cc: rds-devel@oss.oracle.com
Cc: linux-kselftest@vger.kernel.org
---
 include/linux/mm_types.h | 87 ++++++++++++++++++++++++++++++++++++++++-
 mm/util.c                | 89 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 176 insertions(+)

Comments

Christoph Hellwig Jan. 24, 2023, 6:29 a.m. UTC | #1
> +/**
> + * vm_account_init - Initialise a new struct vm_account.
> + * @vm_account: pointer to uninitialised vm_account.
> + * @task: task to charge against.
> + * @user: user to charge against. Must be non-NULL for VM_ACCOUNT_USER.
> + * @flags: flags to use when charging to vm_account.
> + *
> + * Initialise a new uninitialiused struct vm_account. Takes references
> + * on the task/mm/user/cgroup as required although callers must ensure
> + * any references passed in remain valid for the duration of this
> + * call.
> + */
> +void vm_account_init(struct vm_account *vm_account, struct task_struct *task,
> +		struct user_struct *user, enum vm_account_flags flags);


kerneldoc comments are supposed to be next to the implementation, and
not the declaration in the header.
Jason Gunthorpe Jan. 24, 2023, 2:32 p.m. UTC | #2
On Tue, Jan 24, 2023 at 04:42:30PM +1100, Alistair Popple wrote:
> +/**
> + * enum vm_account_flags - Determine how pinned/locked memory is accounted.
> + * @VM_ACCOUNT_TASK: Account pinned memory to mm->pinned_vm.
> + * @VM_ACCOUNT_BYPASS: Don't enforce rlimit on any charges.
> + * @VM_ACCOUNT_USER: Accounnt locked memory to user->locked_vm.
> + *
> + * Determines which statistic pinned/locked memory is accounted
> + * against. All limits will be enforced against RLIMIT_MEMLOCK and the
> + * pins cgroup if CONFIG_CGROUP_PINS is enabled.
> + *
> + * New drivers should use VM_ACCOUNT_TASK. VM_ACCOUNT_USER is used by
> + * pre-existing drivers to maintain existing accounting against
> + * user->locked_mm rather than mm->pinned_mm.

I thought the guidance was the opposite of this, it is the newer
places in the kernel that are using VM_ACCOUNT_USER?

I haven't got to the rest of the patches yet, but isn't there also a
mm->pinned_vm vs mm->locked_vm variation in the current drivers as
well?

> +void vm_account_init_current(struct vm_account *vm_account)
> +{
> +	vm_account_init(vm_account, current, NULL, VM_ACCOUNT_TASK);
> +}
> +EXPORT_SYMBOL_GPL(vm_account_init_current);

This can probably just be a static inline

You might consider putting all this in some new vm_account.h - given
how rarely it is used? Compile times and all

Jason
Alistair Popple Jan. 30, 2023, 11:36 a.m. UTC | #3
Jason Gunthorpe <jgg@nvidia.com> writes:

> On Tue, Jan 24, 2023 at 04:42:30PM +1100, Alistair Popple wrote:
>> +/**
>> + * enum vm_account_flags - Determine how pinned/locked memory is accounted.
>> + * @VM_ACCOUNT_TASK: Account pinned memory to mm->pinned_vm.
>> + * @VM_ACCOUNT_BYPASS: Don't enforce rlimit on any charges.
>> + * @VM_ACCOUNT_USER: Accounnt locked memory to user->locked_vm.
>> + *
>> + * Determines which statistic pinned/locked memory is accounted
>> + * against. All limits will be enforced against RLIMIT_MEMLOCK and the
>> + * pins cgroup if CONFIG_CGROUP_PINS is enabled.
>> + *
>> + * New drivers should use VM_ACCOUNT_TASK. VM_ACCOUNT_USER is used by
>> + * pre-existing drivers to maintain existing accounting against
>> + * user->locked_mm rather than mm->pinned_mm.
>
> I thought the guidance was the opposite of this, it is the newer
> places in the kernel that are using VM_ACCOUNT_USER?

I'd just assumed mm->pinned_vm was preferred because that's what most
drivers use. user->locked_mm does seem more sensible though as at least
it's possible to meaningfully enforce some overall limit. Will switch
the flags/comment around to suggest new users use VM_ACCOUNT_USER.

> I haven't got to the rest of the patches yet, but isn't there also a
> mm->pinned_vm vs mm->locked_vm variation in the current drivers as
> well?
>
>> +void vm_account_init_current(struct vm_account *vm_account)
>> +{
>> +	vm_account_init(vm_account, current, NULL, VM_ACCOUNT_TASK);
>> +}
>> +EXPORT_SYMBOL_GPL(vm_account_init_current);
>
> This can probably just be a static inline
>
> You might consider putting all this in some new vm_account.h - given
> how rarely it is used? Compile times and all

Works for me.

> Jason
David Hildenbrand Jan. 31, 2023, 2 p.m. UTC | #4
On 24.01.23 06:42, Alistair Popple wrote:
> Kernel drivers that pin pages should account these pages against
> either user->locked_vm or mm->pinned_vm and fail the pinning if
> RLIMIT_MEMLOCK is exceeded and CAP_IPC_LOCK isn't held.
> 
> Currently drivers open-code this accounting and use various methods to
> update the atomic variables and check against the limits leading to
> various bugs and inconsistencies. To fix this introduce a standard
> interface for charging pinned and locked memory. As this involves
> taking references on kernel objects such as mm_struct or user_struct
> we introduce a new vm_account struct to hold these references. Several
> helper functions are then introduced to grab references and check
> limits.
> 
> As the way these limits are charged and enforced is visible to
> userspace we need to be careful not to break existing applications by
> charging to different counters. As a result the vm_account functions
> support accounting to different counters as required.
> 
> A future change will extend this to also account against a cgroup for
> pinned pages.

The term "vm_account" is misleading, no? VM_ACCOUNT is for accounting 
towards the commit limit ....
diff mbox series

Patch

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 9757067..7de2168 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -1085,4 +1085,91 @@  enum fault_flag {
 
 typedef unsigned int __bitwise zap_flags_t;
 
+/**
+ * enum vm_account_flags - Determine how pinned/locked memory is accounted.
+ * @VM_ACCOUNT_TASK: Account pinned memory to mm->pinned_vm.
+ * @VM_ACCOUNT_BYPASS: Don't enforce rlimit on any charges.
+ * @VM_ACCOUNT_USER: Accounnt locked memory to user->locked_vm.
+ *
+ * Determines which statistic pinned/locked memory is accounted
+ * against. All limits will be enforced against RLIMIT_MEMLOCK and the
+ * pins cgroup if CONFIG_CGROUP_PINS is enabled.
+ *
+ * New drivers should use VM_ACCOUNT_TASK. VM_ACCOUNT_USER is used by
+ * pre-existing drivers to maintain existing accounting against
+ * user->locked_mm rather than mm->pinned_mm.
+ *
+ * VM_ACCOUNT_BYPASS may also be specified to bypass rlimit
+ * checks. Typically this is used to cache CAP_IPC_LOCK from when a
+ * driver is first initialised. Note that this does not bypass cgroup
+ * limit checks.
+ */
+enum vm_account_flags {
+	VM_ACCOUNT_TASK = 0,
+	VM_ACCOUNT_BYPASS = 1,
+	VM_ACCOUNT_USER = 2,
+};
+
+struct vm_account {
+	struct task_struct *task;
+	union {
+		struct mm_struct *mm;
+		struct user_struct *user;
+	} a;
+	enum vm_account_flags flags;
+};
+
+/**
+ * vm_account_init - Initialise a new struct vm_account.
+ * @vm_account: pointer to uninitialised vm_account.
+ * @task: task to charge against.
+ * @user: user to charge against. Must be non-NULL for VM_ACCOUNT_USER.
+ * @flags: flags to use when charging to vm_account.
+ *
+ * Initialise a new uninitialiused struct vm_account. Takes references
+ * on the task/mm/user/cgroup as required although callers must ensure
+ * any references passed in remain valid for the duration of this
+ * call.
+ */
+void vm_account_init(struct vm_account *vm_account, struct task_struct *task,
+		struct user_struct *user, enum vm_account_flags flags);
+/**
+ * vm_account_init_current - Initialise a new struct vm_account.
+ * @vm_account: pointer to uninitialised vm_account.
+ *
+ * Helper to initialise a vm_account for the common case of charging
+ * with VM_ACCOUNT_TASK against current.
+ */
+void vm_account_init_current(struct vm_account *vm_account);
+
+/**
+ * vm_account_release - Initialise a new struct vm_account.
+ * @vm_account: pointer to initialised vm_account.
+ *
+ * Drop any object references obtained by vm_account_init(). The
+ * vm_account must not be used after calling this unless reinitialised
+ * with vm_account_init().
+ */
+void vm_account_release(struct vm_account *vm_account);
+
+/**
+ * vm_account_pinned - Charge pinned or locked memory to the vm_account.
+ * @vm_account: pointer to an initialised vm_account.
+ * @npages: number of pages to charge.
+ *
+ * Return: 0 on success, -ENOMEM if a limit would be exceeded.
+ *
+ * Note: All pages must be explicitly uncharged with
+ * vm_unaccount_pinned() prior to releasing the vm_account with
+ * vm_account_release().
+ */
+int vm_account_pinned(struct vm_account *vm_account, unsigned long npages);
+
+/**
+ * vm_unaccount_pinned - Uncharge pinned or locked memory to the vm_account.
+ * @vm_account: pointer to an initialised vm_account.
+ * @npages: number of pages to uncharge.
+ */
+void vm_unaccount_pinned(struct vm_account *vm_account, unsigned long npages);
+
 #endif /* _LINUX_MM_TYPES_H */
diff --git a/mm/util.c b/mm/util.c
index b56c92f..af40b1e 100644
--- a/mm/util.c
+++ b/mm/util.c
@@ -430,6 +430,95 @@  void arch_pick_mmap_layout(struct mm_struct *mm, struct rlimit *rlim_stack)
 }
 #endif
 
+void vm_account_init(struct vm_account *vm_account, struct task_struct *task,
+		struct user_struct *user, enum vm_account_flags flags)
+{
+	vm_account->task = get_task_struct(task);
+
+	if (flags & VM_ACCOUNT_USER) {
+		vm_account->a.user = get_uid(user);
+	} else {
+		mmgrab(task->mm);
+		vm_account->a.mm = task->mm;
+	}
+
+	vm_account->flags = flags;
+}
+EXPORT_SYMBOL_GPL(vm_account_init);
+
+void vm_account_init_current(struct vm_account *vm_account)
+{
+	vm_account_init(vm_account, current, NULL, VM_ACCOUNT_TASK);
+}
+EXPORT_SYMBOL_GPL(vm_account_init_current);
+
+void vm_account_release(struct vm_account *vm_account)
+{
+	put_task_struct(vm_account->task);
+	if (vm_account->flags & VM_ACCOUNT_USER)
+		free_uid(vm_account->a.user);
+	else
+		mmdrop(vm_account->a.mm);
+}
+EXPORT_SYMBOL_GPL(vm_account_release);
+
+/*
+ * Charge pages with an atomic compare and swap. Returns -ENOMEM on
+ * failure, 1 on success and 0 for retry.
+ */
+static int vm_account_cmpxchg(struct vm_account *vm_account,
+				unsigned long npages, unsigned long lock_limit)
+{
+	u64 cur_pages, new_pages;
+
+	if (vm_account->flags & VM_ACCOUNT_USER)
+		cur_pages = atomic_long_read(&vm_account->a.user->locked_vm);
+	else
+		cur_pages = atomic64_read(&vm_account->a.mm->pinned_vm);
+
+	new_pages = cur_pages + npages;
+	if (lock_limit != RLIM_INFINITY && new_pages > lock_limit)
+		return -ENOMEM;
+
+	if (vm_account->flags & VM_ACCOUNT_USER) {
+		return atomic_long_cmpxchg(&vm_account->a.user->locked_vm,
+					   cur_pages, new_pages) == cur_pages;
+	} else {
+		return atomic64_cmpxchg(&vm_account->a.mm->pinned_vm,
+					   cur_pages, new_pages) == cur_pages;
+	}
+}
+
+int vm_account_pinned(struct vm_account *vm_account, unsigned long npages)
+{
+	unsigned long lock_limit = RLIM_INFINITY;
+	int ret;
+
+	if (!(vm_account->flags & VM_ACCOUNT_BYPASS) && !capable(CAP_IPC_LOCK))
+		lock_limit = task_rlimit(vm_account->task,
+					RLIMIT_MEMLOCK) >> PAGE_SHIFT;
+
+	while (true) {
+		ret = vm_account_cmpxchg(vm_account, npages, lock_limit);
+		if (ret > 0)
+			break;
+		else if (ret < 0)
+			return ret;
+	}
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(vm_account_pinned);
+
+void vm_unaccount_pinned(struct vm_account *vm_account, unsigned long npages)
+{
+	if (vm_account->flags & VM_ACCOUNT_USER)
+		atomic_long_sub(npages, &vm_account->a.user->locked_vm);
+	else
+		atomic64_sub(npages, &vm_account->a.mm->pinned_vm);
+}
+EXPORT_SYMBOL_GPL(vm_unaccount_pinned);
+
 /**
  * __account_locked_vm - account locked pages to an mm's locked_vm
  * @mm:          mm to account against