Commit 91d8f519 authored by Dennis Zhou's avatar Dennis Zhou Committed by Jens Axboe
Browse files

io_uring: add blkcg accounting to offloaded operations 



There are a few operations that are offloaded to the worker threads. In
this case, we lose process context and end up in kthread context. This
results in ios to be not accounted to the issuing cgroup and
consequently end up as issued by root. Just like others, adopt the
personality of the blkcg too when issuing via the workqueues.

For the SQPOLL thread, it will live and attach in the inited cgroup's
context.

Signed-off-by: default avatarDennis Zhou <dennis@kernel.org>
Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
parent de293938

fs/io-wq.c

+23 −0
Original line number Diff line number Diff line
@@ -17,6 +17,7 @@ 
#include <linux/rculist_nulls.h>

#include <linux/fs_struct.h>

#include <linux/task_work.h>

#include <linux/blk-cgroup.h>



#include "io-wq.h"
@@ -57,6 +58,9 @@ struct io_worker { 


	struct rcu_head rcu;

	struct mm_struct *mm;

#ifdef CONFIG_BLK_CGROUP

	struct cgroup_subsys_state *blkcg_css;

#endif

	const struct cred *cur_creds;

	const struct cred *saved_creds;

	struct files_struct *restore_files;
@@ -177,6 +181,13 @@ static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker) 
		worker->mm = NULL;

	}



#ifdef CONFIG_BLK_CGROUP

	if (worker->blkcg_css) {

		kthread_associate_blkcg(NULL);

		worker->blkcg_css = NULL;

	}

#endif



	return dropped_lock;

}
@@ -439,6 +450,17 @@ static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work) 
	work->flags |= IO_WQ_WORK_CANCEL;

}



static inline void io_wq_switch_blkcg(struct io_worker *worker,

				      struct io_wq_work *work)

{

#ifdef CONFIG_BLK_CGROUP

	if (work->blkcg_css != worker->blkcg_css) {

		kthread_associate_blkcg(work->blkcg_css);

		worker->blkcg_css = work->blkcg_css;

	}

#endif

}



static void io_wq_switch_creds(struct io_worker *worker,

			       struct io_wq_work *work)

{
@@ -467,6 +489,7 @@ static void io_impersonate_work(struct io_worker *worker, 
	if (worker->cur_creds != work->creds)

		io_wq_switch_creds(worker, work);

	current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->fsize;

	io_wq_switch_blkcg(worker, work);

}



static void io_assign_current_work(struct io_worker *worker,

fs/io-wq.h

+3 −0
Original line number Diff line number Diff line
@@ -87,6 +87,9 @@ struct io_wq_work { 
	struct io_wq_work_node list;

	struct files_struct *files;

	struct mm_struct *mm;

#ifdef CONFIG_BLK_CGROUP

	struct cgroup_subsys_state *blkcg_css;

#endif

	const struct cred *creds;

	struct nsproxy *nsproxy;

	struct fs_struct *fs;

fs/io_uring.c

+92 −0
Original line number Diff line number Diff line
@@ -80,6 +80,7 @@ 
#include <linux/task_work.h>

#include <linux/pagemap.h>

#include <linux/io_uring.h>

#include <linux/blk-cgroup.h>



#define CREATE_TRACE_POINTS

#include <trace/events/io_uring.h>
@@ -300,6 +301,10 @@ struct io_ring_ctx { 
	/* Only used for accounting purposes */

	struct mm_struct	*mm_account;



#ifdef CONFIG_BLK_CGROUP

	struct cgroup_subsys_state	*sqo_blkcg_css;

#endif



	struct io_sq_data	*sq_data;	/* if using sq thread polling */



	struct wait_queue_head	sqo_sq_wait;
@@ -748,6 +753,8 @@ struct io_op_def { 
	unsigned		needs_fsize : 1;

	/* must always have async data allocated */

	unsigned		needs_async_data : 1;

	/* needs blkcg context, issues async io potentially */

	unsigned		needs_blkcg : 1;

	/* size of async data needed, if any */

	unsigned short		async_size;

};
@@ -761,6 +768,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.pollin			= 1,

		.buffer_select		= 1,

		.needs_async_data	= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_rw),

	},

	[IORING_OP_WRITEV] = {
@@ -771,15 +779,18 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.pollout		= 1,

		.needs_fsize		= 1,

		.needs_async_data	= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_rw),

	},

	[IORING_OP_FSYNC] = {

		.needs_file		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_READ_FIXED] = {

		.needs_file		= 1,

		.unbound_nonreg_file	= 1,

		.pollin			= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_rw),

	},

	[IORING_OP_WRITE_FIXED] = {
@@ -788,6 +799,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.unbound_nonreg_file	= 1,

		.pollout		= 1,

		.needs_fsize		= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_rw),

	},

	[IORING_OP_POLL_ADD] = {
@@ -797,6 +809,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
	[IORING_OP_POLL_REMOVE] = {},

	[IORING_OP_SYNC_FILE_RANGE] = {

		.needs_file		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_SENDMSG] = {

		.needs_mm		= 1,
@@ -805,6 +818,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.needs_fs		= 1,

		.pollout		= 1,

		.needs_async_data	= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_msghdr),

	},

	[IORING_OP_RECVMSG] = {
@@ -815,6 +829,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.pollin			= 1,

		.buffer_select		= 1,

		.needs_async_data	= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_msghdr),

	},

	[IORING_OP_TIMEOUT] = {
@@ -847,15 +862,18 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
	[IORING_OP_FALLOCATE] = {

		.needs_file		= 1,

		.needs_fsize		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_OPENAT] = {

		.file_table		= 1,

		.needs_fs		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_CLOSE] = {

		.needs_file		= 1,

		.needs_file_no_error	= 1,

		.file_table		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_FILES_UPDATE] = {

		.needs_mm		= 1,
@@ -865,6 +883,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.needs_mm		= 1,

		.needs_fs		= 1,

		.file_table		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_READ] = {

		.needs_mm		= 1,
@@ -872,6 +891,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.unbound_nonreg_file	= 1,

		.pollin			= 1,

		.buffer_select		= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_rw),

	},

	[IORING_OP_WRITE] = {
@@ -880,19 +900,23 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.unbound_nonreg_file	= 1,

		.pollout		= 1,

		.needs_fsize		= 1,

		.needs_blkcg		= 1,

		.async_size		= sizeof(struct io_async_rw),

	},

	[IORING_OP_FADVISE] = {

		.needs_file		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_MADVISE] = {

		.needs_mm		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_SEND] = {

		.needs_mm		= 1,

		.needs_file		= 1,

		.unbound_nonreg_file	= 1,

		.pollout		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_RECV] = {

		.needs_mm		= 1,
@@ -900,10 +924,12 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.unbound_nonreg_file	= 1,

		.pollin			= 1,

		.buffer_select		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_OPENAT2] = {

		.file_table		= 1,

		.needs_fs		= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_EPOLL_CTL] = {

		.unbound_nonreg_file	= 1,
@@ -913,6 +939,7 @@ static const struct io_op_def io_op_defs[] __read_mostly = { 
		.needs_file		= 1,

		.hash_reg_file		= 1,

		.unbound_nonreg_file	= 1,

		.needs_blkcg		= 1,

	},

	[IORING_OP_PROVIDE_BUFFERS] = {},

	[IORING_OP_REMOVE_BUFFERS] = {},
@@ -1011,6 +1038,26 @@ static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx, 
	return __io_sq_thread_acquire_mm(ctx);

}



static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,

					 struct cgroup_subsys_state **cur_css)



{

#ifdef CONFIG_BLK_CGROUP

	/* puts the old one when swapping */

	if (*cur_css != ctx->sqo_blkcg_css) {

		kthread_associate_blkcg(ctx->sqo_blkcg_css);

		*cur_css = ctx->sqo_blkcg_css;

	}

#endif

}



static void io_sq_thread_unassociate_blkcg(void)

{

#ifdef CONFIG_BLK_CGROUP

	kthread_associate_blkcg(NULL);

#endif

}



static inline void req_set_fail_links(struct io_kiocb *req)

{

	if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
@@ -1148,6 +1195,10 @@ static bool io_req_clean_work(struct io_kiocb *req) 
		mmdrop(req->work.mm);

		req->work.mm = NULL;

	}

#ifdef CONFIG_BLK_CGROUP

	if (req->work.blkcg_css)

		css_put(req->work.blkcg_css);

#endif

	if (req->work.creds) {

		put_cred(req->work.creds);

		req->work.creds = NULL;
@@ -1187,6 +1238,19 @@ static void io_prep_async_work(struct io_kiocb *req) 
		mmgrab(current->mm);

		req->work.mm = current->mm;

	}

#ifdef CONFIG_BLK_CGROUP

	if (!req->work.blkcg_css && def->needs_blkcg) {

		rcu_read_lock();

		req->work.blkcg_css = blkcg_css();

		/*

		 * This should be rare, either the cgroup is dying or the task

		 * is moving cgroups. Just punt to root for the handful of ios.

		 */

		if (!css_tryget_online(req->work.blkcg_css))

			req->work.blkcg_css = NULL;

		rcu_read_unlock();

	}

#endif

	if (!req->work.creds)

		req->work.creds = get_current_cred();

	if (!req->work.fs && def->needs_fs) {
@@ -6789,6 +6853,7 @@ static void io_sqd_init_new(struct io_sq_data *sqd) 


static int io_sq_thread(void *data)

{

	struct cgroup_subsys_state *cur_css = NULL;

	const struct cred *old_cred = NULL;

	struct io_sq_data *sqd = data;

	struct io_ring_ctx *ctx;
@@ -6818,6 +6883,7 @@ static int io_sq_thread(void *data) 
					revert_creds(old_cred);

				old_cred = override_creds(ctx->creds);

			}

			io_sq_thread_associate_blkcg(ctx, &cur_css);



			ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
@@ -6841,6 +6907,8 @@ static int io_sq_thread(void *data) 


	io_run_task_work();



	if (cur_css)

		io_sq_thread_unassociate_blkcg();

	if (old_cred)

		revert_creds(old_cred);
@@ -8304,6 +8372,11 @@ static void io_ring_ctx_free(struct io_ring_ctx *ctx) 
		ctx->mm_account = NULL;

	}



#ifdef CONFIG_BLK_CGROUP

	if (ctx->sqo_blkcg_css)

		css_put(ctx->sqo_blkcg_css);

#endif



	io_sqe_files_unregister(ctx);

	io_eventfd_unregister(ctx);

	io_destroy_buffers(ctx);
@@ -9288,6 +9361,25 @@ static int io_uring_create(unsigned entries, struct io_uring_params *p, 
	mmgrab(current->mm);

	ctx->mm_account = current->mm;



#ifdef CONFIG_BLK_CGROUP

	/*

	 * The sq thread will belong to the original cgroup it was inited in.

	 * If the cgroup goes offline (e.g. disabling the io controller), then

	 * issued bios will be associated with the closest cgroup later in the

	 * block layer.

	 */

	rcu_read_lock();

	ctx->sqo_blkcg_css = blkcg_css();

	ret = css_tryget_online(ctx->sqo_blkcg_css);

	rcu_read_unlock();

	if (!ret) {

		/* don't init against a dying cgroup, have the user try again */

		ctx->sqo_blkcg_css = NULL;

		ret = -ENODEV;

		goto err;

	}

#endif



	/*

	 * Account memory _before_ installing the file descriptor. Once

	 * the descriptor is installed, it can get closed at any time. Also

CRA Git | Maintained and supported by SUSTech CRA and CCSE