2012-08-01 03:42:12 +04:00
|
|
|
/*
|
|
|
|
*
|
|
|
|
* Copyright IBM Corporation, 2012
|
|
|
|
* Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
|
|
|
|
*
|
2019-12-16 22:38:31 +03:00
|
|
|
* Cgroup v2
|
|
|
|
* Copyright (C) 2019 Red Hat, Inc.
|
|
|
|
* Author: Giuseppe Scrivano <gscrivan@redhat.com>
|
|
|
|
*
|
2012-08-01 03:42:12 +04:00
|
|
|
* This program is free software; you can redistribute it and/or modify it
|
|
|
|
* under the terms of version 2.1 of the GNU Lesser General Public License
|
|
|
|
* as published by the Free Software Foundation.
|
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it would be useful, but
|
|
|
|
* WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/cgroup.h>
|
2014-12-11 02:42:34 +03:00
|
|
|
#include <linux/page_counter.h>
|
2012-08-01 03:42:12 +04:00
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/hugetlb.h>
|
|
|
|
#include <linux/hugetlb_cgroup.h>
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
enum hugetlb_memory_event {
|
|
|
|
HUGETLB_MAX,
|
|
|
|
HUGETLB_NR_MEMORY_EVENTS,
|
|
|
|
};
|
|
|
|
|
2012-08-01 03:42:12 +04:00
|
|
|
struct hugetlb_cgroup {
|
|
|
|
struct cgroup_subsys_state css;
|
2019-12-16 22:38:31 +03:00
|
|
|
|
2012-08-01 03:42:12 +04:00
|
|
|
/*
|
|
|
|
* the counter to account for hugepages from hugetlb.
|
|
|
|
*/
|
2014-12-11 02:42:34 +03:00
|
|
|
struct page_counter hugepage[HUGE_MAX_HSTATE];
|
2019-12-16 22:38:31 +03:00
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
/*
|
|
|
|
* the counter to account for hugepage reservations from hugetlb.
|
|
|
|
*/
|
|
|
|
struct page_counter rsvd_hugepage[HUGE_MAX_HSTATE];
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
atomic_long_t events[HUGE_MAX_HSTATE][HUGETLB_NR_MEMORY_EVENTS];
|
|
|
|
atomic_long_t events_local[HUGE_MAX_HSTATE][HUGETLB_NR_MEMORY_EVENTS];
|
|
|
|
|
|
|
|
/* Handle for "hugetlb.events" */
|
|
|
|
struct cgroup_file events_file[HUGE_MAX_HSTATE];
|
|
|
|
|
|
|
|
/* Handle for "hugetlb.events.local" */
|
|
|
|
struct cgroup_file events_local_file[HUGE_MAX_HSTATE];
|
2012-08-01 03:42:12 +04:00
|
|
|
};
|
|
|
|
|
2012-08-01 03:42:24 +04:00
|
|
|
#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
|
|
|
|
#define MEMFILE_IDX(val) (((val) >> 16) & 0xffff)
|
|
|
|
#define MEMFILE_ATTR(val) ((val) & 0xffff)
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
#define hugetlb_cgroup_from_counter(counter, idx) \
|
|
|
|
container_of(counter, struct hugetlb_cgroup, hugepage[idx])
|
|
|
|
|
2012-08-01 03:42:12 +04:00
|
|
|
static struct hugetlb_cgroup *root_h_cgroup __read_mostly;
|
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
static inline struct page_counter *
|
|
|
|
hugetlb_cgroup_counter_from_cgroup(struct hugetlb_cgroup *h_cg, int idx,
|
|
|
|
bool rsvd)
|
|
|
|
{
|
|
|
|
if (rsvd)
|
|
|
|
return &h_cg->rsvd_hugepage[idx];
|
|
|
|
return &h_cg->hugepage[idx];
|
|
|
|
}
|
|
|
|
|
2012-08-01 03:42:12 +04:00
|
|
|
static inline
|
|
|
|
struct hugetlb_cgroup *hugetlb_cgroup_from_css(struct cgroup_subsys_state *s)
|
|
|
|
{
|
2013-08-09 04:11:23 +04:00
|
|
|
return s ? container_of(s, struct hugetlb_cgroup, css) : NULL;
|
2012-08-01 03:42:12 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline
|
|
|
|
struct hugetlb_cgroup *hugetlb_cgroup_from_task(struct task_struct *task)
|
|
|
|
{
|
2014-02-08 19:36:58 +04:00
|
|
|
return hugetlb_cgroup_from_css(task_css(task, hugetlb_cgrp_id));
|
2012-08-01 03:42:12 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline bool hugetlb_cgroup_is_root(struct hugetlb_cgroup *h_cg)
|
|
|
|
{
|
|
|
|
return (h_cg == root_h_cgroup);
|
|
|
|
}
|
|
|
|
|
2013-08-09 04:11:22 +04:00
|
|
|
static inline struct hugetlb_cgroup *
|
|
|
|
parent_hugetlb_cgroup(struct hugetlb_cgroup *h_cg)
|
2012-08-01 03:42:12 +04:00
|
|
|
{
|
2014-05-16 21:22:48 +04:00
|
|
|
return hugetlb_cgroup_from_css(h_cg->css.parent);
|
2012-08-01 03:42:12 +04:00
|
|
|
}
|
|
|
|
|
2013-08-09 04:11:22 +04:00
|
|
|
static inline bool hugetlb_cgroup_have_usage(struct hugetlb_cgroup *h_cg)
|
2012-08-01 03:42:12 +04:00
|
|
|
{
|
|
|
|
int idx;
|
|
|
|
|
|
|
|
for (idx = 0; idx < hugetlb_max_hstate; idx++) {
|
2014-12-11 02:42:34 +03:00
|
|
|
if (page_counter_read(&h_cg->hugepage[idx]))
|
2012-08-01 03:42:12 +04:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2016-05-21 02:57:50 +03:00
|
|
|
static void hugetlb_cgroup_init(struct hugetlb_cgroup *h_cgroup,
|
|
|
|
struct hugetlb_cgroup *parent_h_cgroup)
|
|
|
|
{
|
|
|
|
int idx;
|
|
|
|
|
|
|
|
for (idx = 0; idx < HUGE_MAX_HSTATE; idx++) {
|
|
|
|
struct page_counter *counter = &h_cgroup->hugepage[idx];
|
|
|
|
struct page_counter *parent = NULL;
|
|
|
|
unsigned long limit;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
if (parent_h_cgroup)
|
|
|
|
parent = &parent_h_cgroup->hugepage[idx];
|
|
|
|
page_counter_init(counter, parent);
|
|
|
|
|
|
|
|
limit = round_down(PAGE_COUNTER_MAX,
|
|
|
|
1 << huge_page_order(&hstates[idx]));
|
2018-06-08 03:06:18 +03:00
|
|
|
ret = page_counter_set_max(counter, limit);
|
2016-05-21 02:57:50 +03:00
|
|
|
VM_BUG_ON(ret);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-08-09 04:11:23 +04:00
|
|
|
static struct cgroup_subsys_state *
|
|
|
|
hugetlb_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
|
2012-08-01 03:42:12 +04:00
|
|
|
{
|
2013-08-09 04:11:23 +04:00
|
|
|
struct hugetlb_cgroup *parent_h_cgroup = hugetlb_cgroup_from_css(parent_css);
|
|
|
|
struct hugetlb_cgroup *h_cgroup;
|
2012-08-01 03:42:12 +04:00
|
|
|
|
|
|
|
h_cgroup = kzalloc(sizeof(*h_cgroup), GFP_KERNEL);
|
|
|
|
if (!h_cgroup)
|
|
|
|
return ERR_PTR(-ENOMEM);
|
|
|
|
|
2016-05-21 02:57:50 +03:00
|
|
|
if (!parent_h_cgroup)
|
2012-08-01 03:42:12 +04:00
|
|
|
root_h_cgroup = h_cgroup;
|
2016-05-21 02:57:50 +03:00
|
|
|
|
|
|
|
hugetlb_cgroup_init(h_cgroup, parent_h_cgroup);
|
2012-08-01 03:42:12 +04:00
|
|
|
return &h_cgroup->css;
|
|
|
|
}
|
|
|
|
|
2013-08-09 04:11:23 +04:00
|
|
|
static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css)
|
2012-08-01 03:42:12 +04:00
|
|
|
{
|
|
|
|
struct hugetlb_cgroup *h_cgroup;
|
|
|
|
|
2013-08-09 04:11:23 +04:00
|
|
|
h_cgroup = hugetlb_cgroup_from_css(css);
|
2012-08-01 03:42:12 +04:00
|
|
|
kfree(h_cgroup);
|
|
|
|
}
|
|
|
|
|
2012-08-01 03:42:21 +04:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Should be called with hugetlb_lock held.
|
|
|
|
* Since we are holding hugetlb_lock, pages cannot get moved from
|
|
|
|
* active list or uncharged from the cgroup, So no need to get
|
|
|
|
* page reference and test for page active here. This function
|
|
|
|
* cannot fail.
|
|
|
|
*/
|
2013-08-09 04:11:22 +04:00
|
|
|
static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg,
|
2012-08-01 03:42:21 +04:00
|
|
|
struct page *page)
|
|
|
|
{
|
2014-12-11 02:42:34 +03:00
|
|
|
unsigned int nr_pages;
|
|
|
|
struct page_counter *counter;
|
2012-08-01 03:42:21 +04:00
|
|
|
struct hugetlb_cgroup *page_hcg;
|
2013-08-09 04:11:22 +04:00
|
|
|
struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(h_cg);
|
2012-08-01 03:42:21 +04:00
|
|
|
|
|
|
|
page_hcg = hugetlb_cgroup_from_page(page);
|
|
|
|
/*
|
|
|
|
* We can have pages in active list without any cgroup
|
|
|
|
* ie, hugepage with less than 3 pages. We can safely
|
|
|
|
* ignore those pages.
|
|
|
|
*/
|
|
|
|
if (!page_hcg || page_hcg != h_cg)
|
|
|
|
goto out;
|
|
|
|
|
2019-09-24 01:34:30 +03:00
|
|
|
nr_pages = compound_nr(page);
|
2012-08-01 03:42:21 +04:00
|
|
|
if (!parent) {
|
|
|
|
parent = root_h_cgroup;
|
|
|
|
/* root has no limit */
|
2014-12-11 02:42:34 +03:00
|
|
|
page_counter_charge(&parent->hugepage[idx], nr_pages);
|
2012-08-01 03:42:21 +04:00
|
|
|
}
|
|
|
|
counter = &h_cg->hugepage[idx];
|
2014-12-11 02:42:34 +03:00
|
|
|
/* Take the pages off the local counter */
|
|
|
|
page_counter_cancel(counter, nr_pages);
|
2012-08-01 03:42:21 +04:00
|
|
|
|
|
|
|
set_hugetlb_cgroup(page, parent);
|
|
|
|
out:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Force the hugetlb cgroup to empty the hugetlb resources by moving them to
|
|
|
|
* the parent cgroup.
|
|
|
|
*/
|
2013-08-09 04:11:23 +04:00
|
|
|
static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
|
2012-08-01 03:42:12 +04:00
|
|
|
{
|
2013-08-09 04:11:23 +04:00
|
|
|
struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
|
2012-08-01 03:42:21 +04:00
|
|
|
struct hstate *h;
|
|
|
|
struct page *page;
|
2012-10-26 15:37:33 +04:00
|
|
|
int idx = 0;
|
2012-08-01 03:42:21 +04:00
|
|
|
|
|
|
|
do {
|
|
|
|
for_each_hstate(h) {
|
|
|
|
spin_lock(&hugetlb_lock);
|
|
|
|
list_for_each_entry(page, &h->hugepage_activelist, lru)
|
2013-08-09 04:11:22 +04:00
|
|
|
hugetlb_cgroup_move_parent(idx, h_cg, page);
|
2012-08-01 03:42:21 +04:00
|
|
|
|
|
|
|
spin_unlock(&hugetlb_lock);
|
|
|
|
idx++;
|
|
|
|
}
|
|
|
|
cond_resched();
|
2013-08-09 04:11:22 +04:00
|
|
|
} while (hugetlb_cgroup_have_usage(h_cg));
|
2012-08-01 03:42:12 +04:00
|
|
|
}
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
static inline void hugetlb_event(struct hugetlb_cgroup *hugetlb, int idx,
|
|
|
|
enum hugetlb_memory_event event)
|
|
|
|
{
|
|
|
|
atomic_long_inc(&hugetlb->events_local[idx][event]);
|
|
|
|
cgroup_file_notify(&hugetlb->events_local_file[idx]);
|
|
|
|
|
|
|
|
do {
|
|
|
|
atomic_long_inc(&hugetlb->events[idx][event]);
|
|
|
|
cgroup_file_notify(&hugetlb->events_file[idx]);
|
|
|
|
} while ((hugetlb = parent_hugetlb_cgroup(hugetlb)) &&
|
|
|
|
!hugetlb_cgroup_is_root(hugetlb));
|
|
|
|
}
|
|
|
|
|
2012-08-01 03:42:18 +04:00
|
|
|
int hugetlb_cgroup_charge_cgroup(int idx, unsigned long nr_pages,
|
|
|
|
struct hugetlb_cgroup **ptr)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
2014-12-11 02:42:34 +03:00
|
|
|
struct page_counter *counter;
|
2012-08-01 03:42:18 +04:00
|
|
|
struct hugetlb_cgroup *h_cg = NULL;
|
|
|
|
|
|
|
|
if (hugetlb_cgroup_disabled())
|
|
|
|
goto done;
|
|
|
|
/*
|
|
|
|
* We don't charge any cgroup if the compound page have less
|
|
|
|
* than 3 pages.
|
|
|
|
*/
|
|
|
|
if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
|
|
|
|
goto done;
|
|
|
|
again:
|
|
|
|
rcu_read_lock();
|
|
|
|
h_cg = hugetlb_cgroup_from_task(current);
|
2019-11-16 04:34:46 +03:00
|
|
|
if (!css_tryget(&h_cg->css)) {
|
2012-08-01 03:42:18 +04:00
|
|
|
rcu_read_unlock();
|
|
|
|
goto again;
|
|
|
|
}
|
|
|
|
rcu_read_unlock();
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
if (!page_counter_try_charge(&h_cg->hugepage[idx], nr_pages,
|
|
|
|
&counter)) {
|
2015-11-06 05:50:26 +03:00
|
|
|
ret = -ENOMEM;
|
2020-03-29 05:17:22 +03:00
|
|
|
hugetlb_event(h_cg, idx, HUGETLB_MAX);
|
2019-12-16 22:38:31 +03:00
|
|
|
}
|
2012-08-01 03:42:18 +04:00
|
|
|
css_put(&h_cg->css);
|
|
|
|
done:
|
|
|
|
*ptr = h_cg;
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2012-08-01 03:42:35 +04:00
|
|
|
/* Should be called with hugetlb_lock held */
|
2012-08-01 03:42:18 +04:00
|
|
|
void hugetlb_cgroup_commit_charge(int idx, unsigned long nr_pages,
|
|
|
|
struct hugetlb_cgroup *h_cg,
|
|
|
|
struct page *page)
|
|
|
|
{
|
|
|
|
if (hugetlb_cgroup_disabled() || !h_cg)
|
|
|
|
return;
|
|
|
|
|
|
|
|
set_hugetlb_cgroup(page, h_cg);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Should be called with hugetlb_lock held
|
|
|
|
*/
|
|
|
|
void hugetlb_cgroup_uncharge_page(int idx, unsigned long nr_pages,
|
|
|
|
struct page *page)
|
|
|
|
{
|
|
|
|
struct hugetlb_cgroup *h_cg;
|
|
|
|
|
|
|
|
if (hugetlb_cgroup_disabled())
|
|
|
|
return;
|
2014-08-30 02:18:42 +04:00
|
|
|
lockdep_assert_held(&hugetlb_lock);
|
2012-08-01 03:42:18 +04:00
|
|
|
h_cg = hugetlb_cgroup_from_page(page);
|
|
|
|
if (unlikely(!h_cg))
|
|
|
|
return;
|
|
|
|
set_hugetlb_cgroup(page, NULL);
|
2014-12-11 02:42:34 +03:00
|
|
|
page_counter_uncharge(&h_cg->hugepage[idx], nr_pages);
|
2012-08-01 03:42:18 +04:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
void hugetlb_cgroup_uncharge_cgroup(int idx, unsigned long nr_pages,
|
|
|
|
struct hugetlb_cgroup *h_cg)
|
|
|
|
{
|
|
|
|
if (hugetlb_cgroup_disabled() || !h_cg)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (huge_page_order(&hstates[idx]) < HUGETLB_CGROUP_MIN_ORDER)
|
|
|
|
return;
|
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
page_counter_uncharge(&h_cg->hugepage[idx], nr_pages);
|
2012-08-01 03:42:18 +04:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
enum {
|
|
|
|
RES_USAGE,
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
RES_RSVD_USAGE,
|
2014-12-11 02:42:34 +03:00
|
|
|
RES_LIMIT,
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
RES_RSVD_LIMIT,
|
2014-12-11 02:42:34 +03:00
|
|
|
RES_MAX_USAGE,
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
RES_RSVD_MAX_USAGE,
|
2014-12-11 02:42:34 +03:00
|
|
|
RES_FAILCNT,
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
RES_RSVD_FAILCNT,
|
2014-12-11 02:42:34 +03:00
|
|
|
};
|
|
|
|
|
2013-12-05 21:28:03 +04:00
|
|
|
static u64 hugetlb_cgroup_read_u64(struct cgroup_subsys_state *css,
|
|
|
|
struct cftype *cft)
|
2012-08-01 03:42:24 +04:00
|
|
|
{
|
2014-12-11 02:42:34 +03:00
|
|
|
struct page_counter *counter;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
struct page_counter *rsvd_counter;
|
2013-08-09 04:11:24 +04:00
|
|
|
struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
|
2012-08-01 03:42:24 +04:00
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
counter = &h_cg->hugepage[MEMFILE_IDX(cft->private)];
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(cft->private)];
|
2012-08-01 03:42:24 +04:00
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
switch (MEMFILE_ATTR(cft->private)) {
|
|
|
|
case RES_USAGE:
|
|
|
|
return (u64)page_counter_read(counter) * PAGE_SIZE;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_USAGE:
|
|
|
|
return (u64)page_counter_read(rsvd_counter) * PAGE_SIZE;
|
2014-12-11 02:42:34 +03:00
|
|
|
case RES_LIMIT:
|
2018-06-08 03:06:18 +03:00
|
|
|
return (u64)counter->max * PAGE_SIZE;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_LIMIT:
|
|
|
|
return (u64)rsvd_counter->max * PAGE_SIZE;
|
2014-12-11 02:42:34 +03:00
|
|
|
case RES_MAX_USAGE:
|
|
|
|
return (u64)counter->watermark * PAGE_SIZE;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_MAX_USAGE:
|
|
|
|
return (u64)rsvd_counter->watermark * PAGE_SIZE;
|
2014-12-11 02:42:34 +03:00
|
|
|
case RES_FAILCNT:
|
|
|
|
return counter->failcnt;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_FAILCNT:
|
|
|
|
return rsvd_counter->failcnt;
|
2014-12-11 02:42:34 +03:00
|
|
|
default:
|
|
|
|
BUG();
|
|
|
|
}
|
2012-08-01 03:42:24 +04:00
|
|
|
}
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
static int hugetlb_cgroup_read_u64_max(struct seq_file *seq, void *v)
|
|
|
|
{
|
|
|
|
int idx;
|
|
|
|
u64 val;
|
|
|
|
struct cftype *cft = seq_cft(seq);
|
|
|
|
unsigned long limit;
|
|
|
|
struct page_counter *counter;
|
|
|
|
struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
|
|
|
|
|
|
|
|
idx = MEMFILE_IDX(cft->private);
|
|
|
|
counter = &h_cg->hugepage[idx];
|
|
|
|
|
|
|
|
limit = round_down(PAGE_COUNTER_MAX,
|
|
|
|
1 << huge_page_order(&hstates[idx]));
|
|
|
|
|
|
|
|
switch (MEMFILE_ATTR(cft->private)) {
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_USAGE:
|
|
|
|
counter = &h_cg->rsvd_hugepage[idx];
|
|
|
|
/* Fall through. */
|
2019-12-16 22:38:31 +03:00
|
|
|
case RES_USAGE:
|
|
|
|
val = (u64)page_counter_read(counter);
|
|
|
|
seq_printf(seq, "%llu\n", val * PAGE_SIZE);
|
|
|
|
break;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_LIMIT:
|
|
|
|
counter = &h_cg->rsvd_hugepage[idx];
|
|
|
|
/* Fall through. */
|
2019-12-16 22:38:31 +03:00
|
|
|
case RES_LIMIT:
|
|
|
|
val = (u64)counter->max;
|
|
|
|
if (val == limit)
|
|
|
|
seq_puts(seq, "max\n");
|
|
|
|
else
|
|
|
|
seq_printf(seq, "%llu\n", val * PAGE_SIZE);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
BUG();
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
static DEFINE_MUTEX(hugetlb_limit_mutex);
|
|
|
|
|
2014-05-13 20:16:21 +04:00
|
|
|
static ssize_t hugetlb_cgroup_write(struct kernfs_open_file *of,
|
2019-12-16 22:38:31 +03:00
|
|
|
char *buf, size_t nbytes, loff_t off,
|
|
|
|
const char *max)
|
2012-08-01 03:42:24 +04:00
|
|
|
{
|
2014-12-11 02:42:34 +03:00
|
|
|
int ret, idx;
|
|
|
|
unsigned long nr_pages;
|
2014-05-13 20:16:21 +04:00
|
|
|
struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
bool rsvd = false;
|
2012-08-01 03:42:24 +04:00
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
if (hugetlb_cgroup_is_root(h_cg)) /* Can't set limit on root */
|
|
|
|
return -EINVAL;
|
|
|
|
|
2014-05-13 20:16:21 +04:00
|
|
|
buf = strstrip(buf);
|
2019-12-16 22:38:31 +03:00
|
|
|
ret = page_counter_memparse(buf, max, &nr_pages);
|
2014-12-11 02:42:34 +03:00
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
2014-05-13 20:16:21 +04:00
|
|
|
idx = MEMFILE_IDX(of_cft(of)->private);
|
2016-05-21 02:57:50 +03:00
|
|
|
nr_pages = round_down(nr_pages, 1 << huge_page_order(&hstates[idx]));
|
2012-08-01 03:42:24 +04:00
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
switch (MEMFILE_ATTR(of_cft(of)->private)) {
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_LIMIT:
|
|
|
|
rsvd = true;
|
|
|
|
/* Fall through. */
|
2012-08-01 03:42:24 +04:00
|
|
|
case RES_LIMIT:
|
2014-12-11 02:42:34 +03:00
|
|
|
mutex_lock(&hugetlb_limit_mutex);
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
ret = page_counter_set_max(
|
|
|
|
hugetlb_cgroup_counter_from_cgroup(h_cg, idx, rsvd),
|
|
|
|
nr_pages);
|
2014-12-11 02:42:34 +03:00
|
|
|
mutex_unlock(&hugetlb_limit_mutex);
|
2012-08-01 03:42:24 +04:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
ret = -EINVAL;
|
|
|
|
break;
|
|
|
|
}
|
2014-05-13 20:16:21 +04:00
|
|
|
return ret ?: nbytes;
|
2012-08-01 03:42:24 +04:00
|
|
|
}
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
static ssize_t hugetlb_cgroup_write_legacy(struct kernfs_open_file *of,
|
|
|
|
char *buf, size_t nbytes, loff_t off)
|
|
|
|
{
|
|
|
|
return hugetlb_cgroup_write(of, buf, nbytes, off, "-1");
|
|
|
|
}
|
|
|
|
|
|
|
|
static ssize_t hugetlb_cgroup_write_dfl(struct kernfs_open_file *of,
|
|
|
|
char *buf, size_t nbytes, loff_t off)
|
|
|
|
{
|
|
|
|
return hugetlb_cgroup_write(of, buf, nbytes, off, "max");
|
|
|
|
}
|
|
|
|
|
2014-05-13 20:16:21 +04:00
|
|
|
static ssize_t hugetlb_cgroup_reset(struct kernfs_open_file *of,
|
|
|
|
char *buf, size_t nbytes, loff_t off)
|
2012-08-01 03:42:24 +04:00
|
|
|
{
|
2014-12-11 02:42:34 +03:00
|
|
|
int ret = 0;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
struct page_counter *counter, *rsvd_counter;
|
2014-05-13 20:16:21 +04:00
|
|
|
struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(of_css(of));
|
2012-08-01 03:42:24 +04:00
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
counter = &h_cg->hugepage[MEMFILE_IDX(of_cft(of)->private)];
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
rsvd_counter = &h_cg->rsvd_hugepage[MEMFILE_IDX(of_cft(of)->private)];
|
2012-08-01 03:42:24 +04:00
|
|
|
|
2014-12-11 02:42:34 +03:00
|
|
|
switch (MEMFILE_ATTR(of_cft(of)->private)) {
|
2012-08-01 03:42:24 +04:00
|
|
|
case RES_MAX_USAGE:
|
2014-12-11 02:42:34 +03:00
|
|
|
page_counter_reset_watermark(counter);
|
2012-08-01 03:42:24 +04:00
|
|
|
break;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_MAX_USAGE:
|
|
|
|
page_counter_reset_watermark(rsvd_counter);
|
|
|
|
break;
|
2012-08-01 03:42:24 +04:00
|
|
|
case RES_FAILCNT:
|
2014-12-11 02:42:34 +03:00
|
|
|
counter->failcnt = 0;
|
2012-08-01 03:42:24 +04:00
|
|
|
break;
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
case RES_RSVD_FAILCNT:
|
|
|
|
rsvd_counter->failcnt = 0;
|
|
|
|
break;
|
2012-08-01 03:42:24 +04:00
|
|
|
default:
|
|
|
|
ret = -EINVAL;
|
|
|
|
break;
|
|
|
|
}
|
2014-05-13 20:16:21 +04:00
|
|
|
return ret ?: nbytes;
|
2012-08-01 03:42:24 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
static char *mem_fmt(char *buf, int size, unsigned long hsize)
|
|
|
|
{
|
|
|
|
if (hsize >= (1UL << 30))
|
|
|
|
snprintf(buf, size, "%luGB", hsize >> 30);
|
|
|
|
else if (hsize >= (1UL << 20))
|
|
|
|
snprintf(buf, size, "%luMB", hsize >> 20);
|
|
|
|
else
|
|
|
|
snprintf(buf, size, "%luKB", hsize >> 10);
|
|
|
|
return buf;
|
|
|
|
}
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
static int __hugetlb_events_show(struct seq_file *seq, bool local)
|
|
|
|
{
|
|
|
|
int idx;
|
|
|
|
long max;
|
|
|
|
struct cftype *cft = seq_cft(seq);
|
|
|
|
struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(seq_css(seq));
|
|
|
|
|
|
|
|
idx = MEMFILE_IDX(cft->private);
|
|
|
|
|
|
|
|
if (local)
|
|
|
|
max = atomic_long_read(&h_cg->events_local[idx][HUGETLB_MAX]);
|
|
|
|
else
|
|
|
|
max = atomic_long_read(&h_cg->events[idx][HUGETLB_MAX]);
|
|
|
|
|
|
|
|
seq_printf(seq, "max %lu\n", max);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int hugetlb_events_show(struct seq_file *seq, void *v)
|
|
|
|
{
|
|
|
|
return __hugetlb_events_show(seq, false);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int hugetlb_events_local_show(struct seq_file *seq, void *v)
|
|
|
|
{
|
|
|
|
return __hugetlb_events_show(seq, true);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __init __hugetlb_cgroup_file_dfl_init(int idx)
|
2012-08-01 03:42:24 +04:00
|
|
|
{
|
|
|
|
char buf[32];
|
|
|
|
struct cftype *cft;
|
|
|
|
struct hstate *h = &hstates[idx];
|
|
|
|
|
|
|
|
/* format the size */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
mem_fmt(buf, sizeof(buf), huge_page_size(h));
|
2012-08-01 03:42:24 +04:00
|
|
|
|
|
|
|
/* Add the limit file */
|
2019-12-16 22:38:31 +03:00
|
|
|
cft = &h->cgroup_files_dfl[0];
|
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
|
|
|
|
cft->seq_show = hugetlb_cgroup_read_u64_max;
|
|
|
|
cft->write = hugetlb_cgroup_write_dfl;
|
|
|
|
cft->flags = CFTYPE_NOT_ON_ROOT;
|
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
/* Add the reservation limit file */
|
2019-12-16 22:38:31 +03:00
|
|
|
cft = &h->cgroup_files_dfl[1];
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
|
|
|
|
cft->seq_show = hugetlb_cgroup_read_u64_max;
|
|
|
|
cft->write = hugetlb_cgroup_write_dfl;
|
|
|
|
cft->flags = CFTYPE_NOT_ON_ROOT;
|
|
|
|
|
|
|
|
/* Add the current usage file */
|
|
|
|
cft = &h->cgroup_files_dfl[2];
|
2019-12-16 22:38:31 +03:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.current", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
|
|
|
|
cft->seq_show = hugetlb_cgroup_read_u64_max;
|
|
|
|
cft->flags = CFTYPE_NOT_ON_ROOT;
|
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
/* Add the current reservation usage file */
|
|
|
|
cft = &h->cgroup_files_dfl[3];
|
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.current", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
|
|
|
|
cft->seq_show = hugetlb_cgroup_read_u64_max;
|
|
|
|
cft->flags = CFTYPE_NOT_ON_ROOT;
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
/* Add the events file */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft = &h->cgroup_files_dfl[4];
|
2019-12-16 22:38:31 +03:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, 0);
|
|
|
|
cft->seq_show = hugetlb_events_show;
|
|
|
|
cft->file_offset = offsetof(struct hugetlb_cgroup, events_file[idx]),
|
|
|
|
cft->flags = CFTYPE_NOT_ON_ROOT;
|
|
|
|
|
|
|
|
/* Add the events.local file */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft = &h->cgroup_files_dfl[5];
|
2019-12-16 22:38:31 +03:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.events.local", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, 0);
|
|
|
|
cft->seq_show = hugetlb_events_local_show;
|
|
|
|
cft->file_offset = offsetof(struct hugetlb_cgroup,
|
|
|
|
events_local_file[idx]),
|
|
|
|
cft->flags = CFTYPE_NOT_ON_ROOT;
|
|
|
|
|
|
|
|
/* NULL terminate the last cft */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft = &h->cgroup_files_dfl[6];
|
2019-12-16 22:38:31 +03:00
|
|
|
memset(cft, 0, sizeof(*cft));
|
|
|
|
|
|
|
|
WARN_ON(cgroup_add_dfl_cftypes(&hugetlb_cgrp_subsys,
|
|
|
|
h->cgroup_files_dfl));
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __init __hugetlb_cgroup_file_legacy_init(int idx)
|
|
|
|
{
|
|
|
|
char buf[32];
|
|
|
|
struct cftype *cft;
|
|
|
|
struct hstate *h = &hstates[idx];
|
|
|
|
|
|
|
|
/* format the size */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
mem_fmt(buf, sizeof(buf), huge_page_size(h));
|
2019-12-16 22:38:31 +03:00
|
|
|
|
|
|
|
/* Add the limit file */
|
|
|
|
cft = &h->cgroup_files_legacy[0];
|
2012-08-01 03:42:24 +04:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.limit_in_bytes", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_LIMIT);
|
2013-12-05 21:28:03 +04:00
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
2019-12-16 22:38:31 +03:00
|
|
|
cft->write = hugetlb_cgroup_write_legacy;
|
2012-08-01 03:42:24 +04:00
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
/* Add the reservation limit file */
|
2019-12-16 22:38:31 +03:00
|
|
|
cft = &h->cgroup_files_legacy[1];
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.limit_in_bytes", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_LIMIT);
|
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
|
|
|
cft->write = hugetlb_cgroup_write_legacy;
|
|
|
|
|
|
|
|
/* Add the usage file */
|
|
|
|
cft = &h->cgroup_files_legacy[2];
|
2012-08-01 03:42:24 +04:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.usage_in_bytes", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_USAGE);
|
2013-12-05 21:28:03 +04:00
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
2012-08-01 03:42:24 +04:00
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
/* Add the reservation usage file */
|
|
|
|
cft = &h->cgroup_files_legacy[3];
|
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.usage_in_bytes", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_USAGE);
|
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
|
|
|
|
2012-08-01 03:42:24 +04:00
|
|
|
/* Add the MAX usage file */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft = &h->cgroup_files_legacy[4];
|
2012-08-01 03:42:24 +04:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.max_usage_in_bytes", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_MAX_USAGE);
|
2014-05-13 20:16:21 +04:00
|
|
|
cft->write = hugetlb_cgroup_reset;
|
2013-12-05 21:28:03 +04:00
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
2012-08-01 03:42:24 +04:00
|
|
|
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
/* Add the MAX reservation usage file */
|
|
|
|
cft = &h->cgroup_files_legacy[5];
|
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.max_usage_in_bytes", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_MAX_USAGE);
|
|
|
|
cft->write = hugetlb_cgroup_reset;
|
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
|
|
|
|
2012-08-01 03:42:24 +04:00
|
|
|
/* Add the failcntfile */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft = &h->cgroup_files_legacy[6];
|
2012-08-01 03:42:24 +04:00
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.failcnt", buf);
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_FAILCNT);
|
|
|
|
cft->write = hugetlb_cgroup_reset;
|
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
|
|
|
|
|
|
|
/* Add the reservation failcntfile */
|
|
|
|
cft = &h->cgroup_files_legacy[7];
|
|
|
|
snprintf(cft->name, MAX_CFTYPE_NAME, "%s.rsvd.failcnt", buf);
|
|
|
|
cft->private = MEMFILE_PRIVATE(idx, RES_RSVD_FAILCNT);
|
2014-05-13 20:16:21 +04:00
|
|
|
cft->write = hugetlb_cgroup_reset;
|
2013-12-05 21:28:03 +04:00
|
|
|
cft->read_u64 = hugetlb_cgroup_read_u64;
|
2012-08-01 03:42:24 +04:00
|
|
|
|
|
|
|
/* NULL terminate the last cft */
|
hugetlb_cgroup: add hugetlb_cgroup reservation counter
These counters will track hugetlb reservations rather than hugetlb memory
faulted in. This patch only adds the counter, following patches add the
charging and uncharging of the counter.
This is patch 1 of an 9 patch series.
Problem:
Currently tasks attempting to reserve more hugetlb memory than is
available get a failure at mmap/shmget time. This is thanks to Hugetlbfs
Reservations [1]. However, if a task attempts to reserve more hugetlb
memory than its hugetlb_cgroup limit allows, the kernel will allow the
mmap/shmget call, but will SIGBUS the task when it attempts to fault in
the excess memory.
We have users hitting their hugetlb_cgroup limits and thus we've been
looking at this failure mode. We'd like to improve this behavior such
that users violating the hugetlb_cgroup limits get an error on mmap/shmget
time, rather than getting SIGBUS'd when they try to fault the excess
memory in. This gives the user an opportunity to fallback more gracefully
to non-hugetlbfs memory for example.
The underlying problem is that today's hugetlb_cgroup accounting happens
at hugetlb memory *fault* time, rather than at *reservation* time. Thus,
enforcing the hugetlb_cgroup limit only happens at fault time, and the
offending task gets SIGBUS'd.
Proposed Solution:
A new page counter named
'hugetlb.xMB.rsvd.[limit|usage|max_usage]_in_bytes'. This counter has
slightly different semantics than
'hugetlb.xMB.[limit|usage|max_usage]_in_bytes':
- While usage_in_bytes tracks all *faulted* hugetlb memory,
rsvd.usage_in_bytes tracks all *reserved* hugetlb memory and hugetlb
memory faulted in without a prior reservation.
- If a task attempts to reserve more memory than limit_in_bytes allows,
the kernel will allow it to do so. But if a task attempts to reserve
more memory than rsvd.limit_in_bytes, the kernel will fail this
reservation.
This proposal is implemented in this patch series, with tests to verify
functionality and show the usage.
Alternatives considered:
1. A new cgroup, instead of only a new page_counter attached to the
existing hugetlb_cgroup. Adding a new cgroup seemed like a lot of code
duplication with hugetlb_cgroup. Keeping hugetlb related page counters
under hugetlb_cgroup seemed cleaner as well.
2. Instead of adding a new counter, we considered adding a sysctl that
modifies the behavior of hugetlb.xMB.[limit|usage]_in_bytes, to do
accounting at reservation time rather than fault time. Adding a new
page_counter seems better as userspace could, if it wants, choose to
enforce different cgroups differently: one via limit_in_bytes, and
another via rsvd.limit_in_bytes. This could be very useful if you're
transitioning how hugetlb memory is partitioned on your system one
cgroup at a time, for example. Also, someone may find usage for both
limit_in_bytes and rsvd.limit_in_bytes concurrently, and this approach
gives them the option to do so.
Testing:
- Added tests passing.
- Used libhugetlbfs for regression testing.
[1]: https://www.kernel.org/doc/html/latest/vm/hugetlbfs_reserv.html
Signed-off-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Sandipan Das <sandipan@linux.ibm.com>
Link: http://lkml.kernel.org/r/20200211213128.73302-1-almasrymina@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:11:11 +03:00
|
|
|
cft = &h->cgroup_files_legacy[8];
|
2012-08-01 03:42:24 +04:00
|
|
|
memset(cft, 0, sizeof(*cft));
|
|
|
|
|
2014-07-15 19:05:09 +04:00
|
|
|
WARN_ON(cgroup_add_legacy_cftypes(&hugetlb_cgrp_subsys,
|
2019-12-16 22:38:31 +03:00
|
|
|
h->cgroup_files_legacy));
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __init __hugetlb_cgroup_file_init(int idx)
|
|
|
|
{
|
|
|
|
__hugetlb_cgroup_file_dfl_init(idx);
|
|
|
|
__hugetlb_cgroup_file_legacy_init(idx);
|
2012-12-19 02:23:19 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
void __init hugetlb_cgroup_file_init(void)
|
|
|
|
{
|
|
|
|
struct hstate *h;
|
|
|
|
|
|
|
|
for_each_hstate(h) {
|
|
|
|
/*
|
|
|
|
* Add cgroup control files only if the huge page consists
|
|
|
|
* of more than two normal pages. This is because we use
|
2015-11-07 03:29:54 +03:00
|
|
|
* page[2].private for storing cgroup details.
|
2012-12-19 02:23:19 +04:00
|
|
|
*/
|
|
|
|
if (huge_page_order(h) >= HUGETLB_CGROUP_MIN_ORDER)
|
|
|
|
__hugetlb_cgroup_file_init(hstate_index(h));
|
|
|
|
}
|
2012-08-01 03:42:24 +04:00
|
|
|
}
|
|
|
|
|
2012-08-01 03:42:36 +04:00
|
|
|
/*
|
|
|
|
* hugetlb_lock will make sure a parallel cgroup rmdir won't happen
|
|
|
|
* when we migrate hugepages
|
|
|
|
*/
|
2012-08-01 03:42:27 +04:00
|
|
|
void hugetlb_cgroup_migrate(struct page *oldhpage, struct page *newhpage)
|
|
|
|
{
|
|
|
|
struct hugetlb_cgroup *h_cg;
|
2012-08-01 03:42:35 +04:00
|
|
|
struct hstate *h = page_hstate(oldhpage);
|
2012-08-01 03:42:27 +04:00
|
|
|
|
|
|
|
if (hugetlb_cgroup_disabled())
|
|
|
|
return;
|
|
|
|
|
2014-01-24 03:52:54 +04:00
|
|
|
VM_BUG_ON_PAGE(!PageHuge(oldhpage), oldhpage);
|
2012-08-01 03:42:27 +04:00
|
|
|
spin_lock(&hugetlb_lock);
|
|
|
|
h_cg = hugetlb_cgroup_from_page(oldhpage);
|
|
|
|
set_hugetlb_cgroup(oldhpage, NULL);
|
|
|
|
|
|
|
|
/* move the h_cg details to new cgroup */
|
|
|
|
set_hugetlb_cgroup(newhpage, h_cg);
|
2012-08-01 03:42:35 +04:00
|
|
|
list_move(&newhpage->lru, &h->hugepage_activelist);
|
2012-08-01 03:42:27 +04:00
|
|
|
spin_unlock(&hugetlb_lock);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2019-12-16 22:38:31 +03:00
|
|
|
static struct cftype hugetlb_files[] = {
|
|
|
|
{} /* terminate */
|
|
|
|
};
|
|
|
|
|
2014-02-08 19:36:58 +04:00
|
|
|
struct cgroup_subsys hugetlb_cgrp_subsys = {
|
2012-11-19 20:13:38 +04:00
|
|
|
.css_alloc = hugetlb_cgroup_css_alloc,
|
|
|
|
.css_offline = hugetlb_cgroup_css_offline,
|
|
|
|
.css_free = hugetlb_cgroup_css_free,
|
2019-12-16 22:38:31 +03:00
|
|
|
.dfl_cftypes = hugetlb_files,
|
|
|
|
.legacy_cftypes = hugetlb_files,
|
2012-08-01 03:42:12 +04:00
|
|
|
};
|