Skip to contentSkip to conversation

memcg: charge io_uring related objects

PATCH In review Yutian Yang 3 participants 3 messages Jul 17, 2021

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 memcg: charge io_uring related objects

This patch adds accounting flags to allocations of io_uring related objects. Allocations of the objects are all triggerable by io_uring_setup() syscall from userspace.

We have written a PoC to show that the missing-charging objects lead to breaking memcg limits. The PoC program takes around 835MB unaccounted memory, while it is charged for only 23MB memory usage. We evaluate the PoC on QEMU x86_64 v5.2.90 + Linux kernel v5.10.19 + Debian buster. All the limitations including ulimits and sysctl variables are set as default. Specifically, the MEMLOCK in prlimit is set as 65536KB and hard limit of NOFILE is set as 1,048,576.

The PoC is written as a testcase under Linux LTP. To compile it, put the source to ltp_dir/testcases/kernel/syscalls/io_uring/ and make.

/*------------------------- POC code ----------------------------*/

#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <string.h>
#include <fcntl.h>
#include "config.h"
#include "tst_test.h"
#include "lapi/io_uring.h"
#include "lapi/namespaces_constants.h"

#define TEST_FILE "test_file"

#define QUEUE_DEPTH 1
#define BLOCK_SZ 1024

#define errExit(msg)    do { perror(msg); exit(EXIT_FAILURE);  \
                        } while (0)

#define STACK_SIZE (16 * 1024)

static char thread_stack[512][STACK_SIZE];

int thread_fn(void* arg)
{
struct io_uring_params p;
memset(&p, 0 ,sizeof(p));

  for (int i = 0; i< 10000 ; ++i) {
    int ringfd = io_uring_setup(QUEUE_DEPTH, &p);
    if (ringfd == -1) {
      errExit("io_uring_setup");
    }

}
while(1);
return 0;
}

static void run(unsigned int n) {
int thread_pid;

  for (int i = 0; i < 1; ++i) {
    thread_pid = ltp_clone(SIGCHLD, thread_fn, NULL, STACK_SIZE, \
      thread_stack[i]);

}
while(1);
}

static struct tst_test test = {
.test = run,
.tcnt = 1,
.timeout = -1,
};

/*-------------------------- end --------------------------------*/

---
 fs/io-wq.c    |  6 +++---
 fs/io_uring.c | 10 +++++-----
 2 files changed, 8 insertions(+), 8 deletions(-)
fs/io-wq.c +3 -3
index f72d53848..ab31d01cc 100644
@@ -1086,11 +1086,11 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
 		return ERR_PTR(-EINVAL);
 
-	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
+	wq = kzalloc(sizeof(*wq), GFP_KERNEL_ACCOUNT);
 	if (!wq)
 		return ERR_PTR(-ENOMEM);
 
-	wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
+	wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL_ACCOUNT);
 	if (!wq->wqes)
 		goto err_wq;
 
@@ -1111,7 +1111,7 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 
 		if (!node_online(alloc_node))
 			alloc_node = NUMA_NO_NODE;
-		wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
+		wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL_ACCOUNT, alloc_node);
 		if (!wqe)
 			goto err;
 		wq->wqes[node] = wqe;
fs/io_uring.c +5 -5
index d0b7332ca..175fd5b0e 100644
@@ -1177,7 +1177,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	struct io_ring_ctx *ctx;
 	int hash_bits;
 
-	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
+	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL_ACCOUNT);
 	if (!ctx)
 		return NULL;
 
@@ -1195,7 +1195,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 		hash_bits = 1;
 	ctx->cancel_hash_bits = hash_bits;
 	ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
-					GFP_KERNEL);
+					GFP_KERNEL_ACCOUNT);
 	if (!ctx->cancel_hash)
 		goto err;
 	__hash_init(ctx->cancel_hash, 1U << hash_bits);
@@ -7850,7 +7850,7 @@ static int io_uring_alloc_task_context(struct task_struct *task)
 	struct io_uring_task *tctx;
 	int ret;
 
-	tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
+	tctx = kmalloc(sizeof(*tctx), GFP_KERNEL_ACCOUNT);
 	if (unlikely(!tctx))
 		return -ENOMEM;
 
@@ -8038,7 +8038,7 @@ static void io_mem_free(void *ptr)
 
 static void *io_mem_alloc(size_t size)
 {
-	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
+	gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
 				__GFP_NORETRY;
 
 	return (void *) __get_free_pages(gfp_flags, get_order(size));
@@ -9874,7 +9874,7 @@ static int __init io_uring_init(void)
 
 	BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
 	BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
-	req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
+	req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
 	return 0;
 };
 __initcall(io_uring_init);
Signature · 1 line
2.25.1
#1 lore raw

+ Jens and Andrew

Quoted 128 lines · Yutian Yang

This patch adds accounting flags to allocations of io_uring related objects. Allocations of the objects are all triggerable by io_uring_setup() syscall from userspace.

We have written a PoC to show that the missing-charging objects lead to breaking memcg limits. The PoC program takes around 835MB unaccounted memory, while it is charged for only 23MB memory usage. We evaluate the PoC on QEMU x86_64 v5.2.90 + Linux kernel v5.10.19 + Debian buster. All the limitations including ulimits and sysctl variables are set as default. Specifically, the MEMLOCK in prlimit is set as 65536KB and hard limit of NOFILE is set as 1,048,576.

The PoC is written as a testcase under Linux LTP. To compile it, put the source to ltp_dir/testcases/kernel/syscalls/io_uring/ and make.

/*------------------------- POC code ----------------------------*/

#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <string.h>
#include <fcntl.h>
#include "config.h"
#include "tst_test.h"
#include "lapi/io_uring.h"
#include "lapi/namespaces_constants.h"

#define TEST_FILE "test_file"

#define QUEUE_DEPTH 1
#define BLOCK_SZ 1024

#define errExit(msg)    do { perror(msg); exit(EXIT_FAILURE);  \
                        } while (0)

#define STACK_SIZE (16 * 1024)

static char thread_stack[512][STACK_SIZE];

int thread_fn(void* arg)
{
struct io_uring_params p;
memset(&p, 0 ,sizeof(p));

  for (int i = 0; i< 10000 ; ++i) {
    int ringfd = io_uring_setup(QUEUE_DEPTH, &p);
    if (ringfd == -1) {
      errExit("io_uring_setup");
    }

}
while(1);
return 0;
}

static void run(unsigned int n) {
int thread_pid;

  for (int i = 0; i < 1; ++i) {
    thread_pid = ltp_clone(SIGCHLD, thread_fn, NULL, STACK_SIZE, \
      thread_stack[i]);

}
while(1);
}

static struct tst_test test = {
.test = run,
.tcnt = 1,
.timeout = -1,
};

/*-------------------------- end --------------------------------*/

---
 fs/io-wq.c    |  6 +++---
 fs/io_uring.c | 10 +++++-----
 2 files changed, 8 insertions(+), 8 deletions(-)
fs/io-wq.c +3 -3
index f72d53848..ab31d01cc 100644
@@ -1086,11 +1086,11 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
        if (WARN_ON_ONCE(!data->free_work || !data->do_work))
                return ERR_PTR(-EINVAL);
 
-       wq = kzalloc(sizeof(*wq), GFP_KERNEL);
+       wq = kzalloc(sizeof(*wq), GFP_KERNEL_ACCOUNT);
        if (!wq)
                return ERR_PTR(-ENOMEM);
 
-       wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
+       wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL_ACCOUNT);
        if (!wq->wqes)
                goto err_wq;
 
@@ -1111,7 +1111,7 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 
                if (!node_online(alloc_node))
                        alloc_node = NUMA_NO_NODE;
-               wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
+               wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL_ACCOUNT, alloc_node);
                if (!wqe)
                        goto err;
                wq->wqes[node] = wqe;
fs/io_uring.c +2 -2
index d0b7332ca..175fd5b0e 100644
@@ -1177,7 +1177,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
        struct io_ring_ctx *ctx;
        int hash_bits;
 
-       ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
+       ctx = kzalloc(sizeof(*ctx), GFP_KERNEL_ACCOUNT);
        if (!ctx)
                return NULL;
 
@@ -1195,7 +1195,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
                hash_bits = 1;
        ctx->cancel_hash_bits = hash_bits;
        ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
-                                       GFP_KERNEL);
+                                       GFP_KERNEL_ACCOUNT);
        if (!ctx->cancel_hash)
                goto err;
        __hash_init(ctx->cancel_hash, 1U << hash_bits);

fs/io_uring.c static int io_uring_alloc_task_context(struct task_struct *task) · line 7855

Quoted 7 lines
fs/io_uring.c +1 -1
@@ -7850,7 +7850,7 @@ static int io_uring_alloc_task_context(struct task_struct *task)
        struct io_uring_task *tctx;
        int ret;
 
-       tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
+       tctx = kmalloc(sizeof(*tctx), GFP_KERNEL_ACCOUNT);
        if (unlikely(!tctx))
                return -ENOMEM;

What about percpu_counter_init() in this function and io_wq_hash in io_init_wq_offload()?

Quoted 22 lines
fs/io_uring.c +2 -3
@@ -8038,7 +8038,7 @@ static void io_mem_free(void *ptr)
 
 static void *io_mem_alloc(size_t size)
 {
-       gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
+       gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
                                __GFP_NORETRY;
 
        return (void *) __get_free_pages(gfp_flags, get_order(size));
@@ -9874,7 +9874,7 @@ static int __init io_uring_init(void)
 
        BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
        BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
-       req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
+       req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
        return 0;
 };
 __initcall(io_uring_init);
--

2.25.1

#2 lore raw
Quoted 1 line · Shakeel Butt

+ Jens and Andrew

A lot of these are probably (mostly) pointless to account, I suspect you'd get 99% of the way there by just doing io_mem_alloc(). But as far as the patch is concerned, looks fine to me. If we're doing all of them, then the patch is incomplete. If you send a v2 (that's for the original author) I'll be happy to get it queude for 5.14.

Signature · 1 line
Jens Axboe