io_uring: split out open/close operations
authorJens Axboe <axboe@kernel.dk>
Wed, 25 May 2022 03:54:43 +0000 (21:54 -0600)
committerJens Axboe <axboe@kernel.dk>
Mon, 25 Jul 2022 00:39:11 +0000 (18:39 -0600)
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring/Makefile
io_uring/io_uring.c
io_uring/io_uring.h
io_uring/openclose.c [new file with mode: 0644]
io_uring/openclose.h [new file with mode: 0644]

index 5efc4fe565a172a66389d5fdf1c8f6ebe10cc0df..e60def39ca2c6a7bf62adefd62b3f706af66d3c2 100644 (file)
@@ -3,5 +3,6 @@
 # Makefile for io_uring
 
 obj-$(CONFIG_IO_URING)         += io_uring.o xattr.o nop.o fs.o splice.o \
-                                       sync.o advise.o filetable.o
+                                       sync.o advise.o filetable.o \
+                                       openclose.o
 obj-$(CONFIG_IO_WQ)            += io-wq.o
index 4b4d6fd509d13f64d138f03d677acf87ee6548e0..a79186ba8c44ec6c729e3304a432c33babfb51ae 100644 (file)
@@ -98,6 +98,7 @@
 #include "splice.h"
 #include "sync.h"
 #include "advise.h"
+#include "openclose.h"
 
 #define IORING_MAX_ENTRIES     32768
 #define IORING_MAX_CQ_ENTRIES  (2 * IORING_MAX_ENTRIES)
@@ -283,12 +284,6 @@ struct io_poll_update {
        bool                            update_user_data;
 };
 
-struct io_close {
-       struct file                     *file;
-       int                             fd;
-       u32                             file_slot;
-};
-
 struct io_timeout_data {
        struct io_kiocb                 *req;
        struct hrtimer                  timer;
@@ -371,15 +366,6 @@ struct io_sr_msg {
        unsigned int                    flags;
 };
 
-struct io_open {
-       struct file                     *file;
-       int                             dfd;
-       u32                             file_slot;
-       struct filename                 *filename;
-       struct open_how                 how;
-       unsigned long                   nofile;
-};
-
 struct io_rsrc_update {
        struct file                     *file;
        u64                             arg;
@@ -555,9 +541,6 @@ static int io_req_prep_async(struct io_kiocb *req);
 
 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
                                 unsigned int issue_flags, u32 slot_index);
-static int __io_close_fixed(struct io_kiocb *req, unsigned int issue_flags,
-                           unsigned int offset);
-static inline int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
 
 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
 static void io_eventfd_signal(struct io_ring_ctx *ctx);
@@ -670,10 +653,15 @@ const char *io_uring_get_opcode(u8 opcode)
        return "INVALID";
 }
 
+bool io_is_uring_fops(struct file *file)
+{
+       return file->f_op == &io_uring_fops;
+}
+
 struct sock *io_uring_get_socket(struct file *file)
 {
 #if defined(CONFIG_UNIX)
-       if (file->f_op == &io_uring_fops) {
+       if (io_is_uring_fops(file)) {
                struct io_ring_ctx *ctx = file->private_data;
 
                return ctx->ring_sock->sk;
@@ -699,26 +687,6 @@ static inline bool io_file_need_scm(struct file *filp)
 }
 #endif
 
-static void io_ring_submit_unlock(struct io_ring_ctx *ctx, unsigned issue_flags)
-{
-       lockdep_assert_held(&ctx->uring_lock);
-       if (issue_flags & IO_URING_F_UNLOCKED)
-               mutex_unlock(&ctx->uring_lock);
-}
-
-static void io_ring_submit_lock(struct io_ring_ctx *ctx, unsigned issue_flags)
-{
-       /*
-        * "Normal" inline submissions always hold the uring_lock, since we
-        * grab it from the system call. Same is true for the SQPOLL offload.
-        * The only exception is when we've detached the request and issue it
-        * from an async worker thread, grab the lock for that case.
-        */
-       if (issue_flags & IO_URING_F_UNLOCKED)
-               mutex_lock(&ctx->uring_lock);
-       lockdep_assert_held(&ctx->uring_lock);
-}
-
 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
 {
        if (!*locked) {
@@ -3899,74 +3867,12 @@ done:
        return IOU_OK;
 }
 
-static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
-{
-       struct io_open *open = io_kiocb_to_cmd(req);
-       const char __user *fname;
-       int ret;
-
-       if (unlikely(sqe->buf_index))
-               return -EINVAL;
-       if (unlikely(req->flags & REQ_F_FIXED_FILE))
-               return -EBADF;
-
-       /* open.how should be already initialised */
-       if (!(open->how.flags & O_PATH) && force_o_largefile())
-               open->how.flags |= O_LARGEFILE;
-
-       open->dfd = READ_ONCE(sqe->fd);
-       fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
-       open->filename = getname(fname);
-       if (IS_ERR(open->filename)) {
-               ret = PTR_ERR(open->filename);
-               open->filename = NULL;
-               return ret;
-       }
-
-       open->file_slot = READ_ONCE(sqe->file_index);
-       if (open->file_slot && (open->how.flags & O_CLOEXEC))
-               return -EINVAL;
-
-       open->nofile = rlimit(RLIMIT_NOFILE);
-       req->flags |= REQ_F_NEED_CLEANUP;
-       return 0;
-}
-
-static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
-{
-       struct io_open *open = io_kiocb_to_cmd(req);
-       u64 mode = READ_ONCE(sqe->len);
-       u64 flags = READ_ONCE(sqe->open_flags);
-
-       open->how = build_open_how(flags, mode);
-       return __io_openat_prep(req, sqe);
-}
-
-static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
-{
-       struct io_open *open = io_kiocb_to_cmd(req);
-       struct open_how __user *how;
-       size_t len;
-       int ret;
-
-       how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
-       len = READ_ONCE(sqe->len);
-       if (len < OPEN_HOW_SIZE_VER0)
-               return -EINVAL;
-
-       ret = copy_struct_from_user(&open->how, sizeof(open->how), how, len);
-       if (ret)
-               return ret;
-
-       return __io_openat_prep(req, sqe);
-}
-
 /*
  * Note when io_fixed_fd_install() returns error value, it will ensure
  * fput() is called correspondingly.
  */
-static int io_fixed_fd_install(struct io_kiocb *req, unsigned int issue_flags,
-                              struct file *file, unsigned int file_slot)
+int io_fixed_fd_install(struct io_kiocb *req, unsigned int issue_flags,
+                       struct file *file, unsigned int file_slot)
 {
        bool alloc_slot = file_slot == IORING_FILE_INDEX_ALLOC;
        struct io_ring_ctx *ctx = req->ctx;
@@ -3993,86 +3899,6 @@ err:
        return ret;
 }
 
-static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
-{
-       struct io_open *open = io_kiocb_to_cmd(req);
-       struct open_flags op;
-       struct file *file;
-       bool resolve_nonblock, nonblock_set;
-       bool fixed = !!open->file_slot;
-       int ret;
-
-       ret = build_open_flags(&open->how, &op);
-       if (ret)
-               goto err;
-       nonblock_set = op.open_flag & O_NONBLOCK;
-       resolve_nonblock = open->how.resolve & RESOLVE_CACHED;
-       if (issue_flags & IO_URING_F_NONBLOCK) {
-               /*
-                * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
-                * it'll always -EAGAIN
-                */
-               if (open->how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
-                       return -EAGAIN;
-               op.lookup_flags |= LOOKUP_CACHED;
-               op.open_flag |= O_NONBLOCK;
-       }
-
-       if (!fixed) {
-               ret = __get_unused_fd_flags(open->how.flags, open->nofile);
-               if (ret < 0)
-                       goto err;
-       }
-
-       file = do_filp_open(open->dfd, open->filename, &op);
-       if (IS_ERR(file)) {
-               /*
-                * We could hang on to this 'fd' on retrying, but seems like
-                * marginal gain for something that is now known to be a slower
-                * path. So just put it, and we'll get a new one when we retry.
-                */
-               if (!fixed)
-                       put_unused_fd(ret);
-
-               ret = PTR_ERR(file);
-               /* only retry if RESOLVE_CACHED wasn't already set by application */
-               if (ret == -EAGAIN &&
-                   (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
-                       return -EAGAIN;
-               goto err;
-       }
-
-       if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
-               file->f_flags &= ~O_NONBLOCK;
-       fsnotify_open(file);
-
-       if (!fixed)
-               fd_install(ret, file);
-       else
-               ret = io_fixed_fd_install(req, issue_flags, file,
-                                               open->file_slot);
-err:
-       putname(open->filename);
-       req->flags &= ~REQ_F_NEED_CLEANUP;
-       if (ret < 0)
-               req_set_fail(req);
-       io_req_set_res(req, ret, 0);
-       return IOU_OK;
-}
-
-static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
-{
-       return io_openat2(req, issue_flags);
-}
-
-static void io_open_cleanup(struct io_kiocb *req)
-{
-       struct io_open *open = io_kiocb_to_cmd(req);
-
-       if (open->filename)
-               putname(open->filename);
-}
-
 static int io_remove_buffers_prep(struct io_kiocb *req,
                                  const struct io_uring_sqe *sqe)
 {
@@ -4424,69 +4250,6 @@ static void io_statx_cleanup(struct io_kiocb *req)
                putname(sx->filename);
 }
 
-static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
-{
-       struct io_close *close = io_kiocb_to_cmd(req);
-
-       if (sqe->off || sqe->addr || sqe->len || sqe->rw_flags || sqe->buf_index)
-               return -EINVAL;
-       if (req->flags & REQ_F_FIXED_FILE)
-               return -EBADF;
-
-       close->fd = READ_ONCE(sqe->fd);
-       close->file_slot = READ_ONCE(sqe->file_index);
-       if (close->file_slot && close->fd)
-               return -EINVAL;
-
-       return 0;
-}
-
-static int io_close(struct io_kiocb *req, unsigned int issue_flags)
-{
-       struct files_struct *files = current->files;
-       struct io_close *close = io_kiocb_to_cmd(req);
-       struct fdtable *fdt;
-       struct file *file;
-       int ret = -EBADF;
-
-       if (close->file_slot) {
-               ret = io_close_fixed(req, issue_flags);
-               goto err;
-       }
-
-       spin_lock(&files->file_lock);
-       fdt = files_fdtable(files);
-       if (close->fd >= fdt->max_fds) {
-               spin_unlock(&files->file_lock);
-               goto err;
-       }
-       file = rcu_dereference_protected(fdt->fd[close->fd],
-                       lockdep_is_held(&files->file_lock));
-       if (!file || file->f_op == &io_uring_fops) {
-               spin_unlock(&files->file_lock);
-               goto err;
-       }
-
-       /* if the file has a flush method, be safe and punt to async */
-       if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
-               spin_unlock(&files->file_lock);
-               return -EAGAIN;
-       }
-
-       file = __close_fd_get_file(close->fd);
-       spin_unlock(&files->file_lock);
-       if (!file)
-               goto err;
-
-       /* No ->flush() or already async, safely close from here */
-       ret = filp_close(file, current->files);
-err:
-       if (ret < 0)
-               req_set_fail(req);
-       io_req_set_res(req, ret, 0);
-       return IOU_OK;
-}
-
 #if defined(CONFIG_NET)
 static int io_shutdown_prep(struct io_kiocb *req,
                            const struct io_uring_sqe *sqe)
@@ -7744,8 +7507,8 @@ static struct io_rsrc_node *io_rsrc_node_alloc(void)
        return ref_node;
 }
 
-static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
-                               struct io_rsrc_data *data_to_kill)
+void io_rsrc_node_switch(struct io_ring_ctx *ctx,
+                        struct io_rsrc_data *data_to_kill)
        __must_hold(&ctx->uring_lock)
 {
        WARN_ON_ONCE(!ctx->rsrc_backup_node);
@@ -7772,7 +7535,7 @@ static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
        }
 }
 
-static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
+int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
 {
        if (ctx->rsrc_backup_node)
                return 0;
@@ -8319,8 +8082,8 @@ fail:
        return ret;
 }
 
-static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
-                                struct io_rsrc_node *node, void *rsrc)
+int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
+                         struct io_rsrc_node *node, void *rsrc)
 {
        u64 *tag_slot = io_get_tag_slot(data, idx);
        struct io_rsrc_put *prsrc;
@@ -8386,52 +8149,6 @@ err:
        return ret;
 }
 
-static int __io_close_fixed(struct io_kiocb *req, unsigned int issue_flags,
-                           unsigned int offset)
-{
-       struct io_ring_ctx *ctx = req->ctx;
-       struct io_fixed_file *file_slot;
-       struct file *file;
-       int ret;
-
-       io_ring_submit_lock(ctx, issue_flags);
-       ret = -ENXIO;
-       if (unlikely(!ctx->file_data))
-               goto out;
-       ret = -EINVAL;
-       if (offset >= ctx->nr_user_files)
-               goto out;
-       ret = io_rsrc_node_switch_start(ctx);
-       if (ret)
-               goto out;
-
-       offset = array_index_nospec(offset, ctx->nr_user_files);
-       file_slot = io_fixed_file_slot(&ctx->file_table, offset);
-       ret = -EBADF;
-       if (!file_slot->file_ptr)
-               goto out;
-
-       file = (struct file *)(file_slot->file_ptr & FFS_MASK);
-       ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
-       if (ret)
-               goto out;
-
-       file_slot->file_ptr = 0;
-       io_file_bitmap_clear(&ctx->file_table, offset);
-       io_rsrc_node_switch(ctx, ctx->file_data);
-       ret = 0;
-out:
-       io_ring_submit_unlock(ctx, issue_flags);
-       return ret;
-}
-
-static inline int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
-{
-       struct io_close *close = io_kiocb_to_cmd(req);
-
-       return __io_close_fixed(req, issue_flags, close->file_slot - 1);
-}
-
 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
                                 struct io_uring_rsrc_update2 *up,
                                 unsigned nr_args)
index 02c00122b97a31512c246225b0ab5a29c5c5847f..ebb225e85012513e29add4c6ddd4dcf37be9cacf 100644 (file)
@@ -2,6 +2,7 @@
 #define IOU_CORE_H
 
 #include <linux/errno.h>
+#include <linux/lockdep.h>
 #include "io_uring_types.h"
 
 enum {
@@ -30,8 +31,39 @@ static inline void io_put_file(struct file *file)
                fput(file);
 }
 
+static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
+                                        unsigned issue_flags)
+{
+       lockdep_assert_held(&ctx->uring_lock);
+       if (issue_flags & IO_URING_F_UNLOCKED)
+               mutex_unlock(&ctx->uring_lock);
+}
+
+static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
+                                      unsigned issue_flags)
+{
+       /*
+        * "Normal" inline submissions always hold the uring_lock, since we
+        * grab it from the system call. Same is true for the SQPOLL offload.
+        * The only exception is when we've detached the request and issue it
+        * from an async worker thread, grab the lock for that case.
+        */
+       if (issue_flags & IO_URING_F_UNLOCKED)
+               mutex_lock(&ctx->uring_lock);
+       lockdep_assert_held(&ctx->uring_lock);
+}
+
 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
                               unsigned issue_flags);
+int io_fixed_fd_install(struct io_kiocb *req, unsigned int issue_flags,
+                       struct file *file, unsigned int file_slot);
+
+int io_rsrc_node_switch_start(struct io_ring_ctx *ctx);
+int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
+                         struct io_rsrc_node *node, void *rsrc);
+void io_rsrc_node_switch(struct io_ring_ctx *ctx,
+                        struct io_rsrc_data *data_to_kill);
+bool io_is_uring_fops(struct file *file);
 
 #endif
diff --git a/io_uring/openclose.c b/io_uring/openclose.c
new file mode 100644 (file)
index 0000000..fa35bd5
--- /dev/null
@@ -0,0 +1,283 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <linux/kernel.h>
+#include <linux/errno.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/fdtable.h>
+#include <linux/fsnotify.h>
+#include <linux/namei.h>
+#include <linux/io_uring.h>
+
+#include <uapi/linux/io_uring.h>
+
+#include "../fs/internal.h"
+
+#include "io_uring_types.h"
+#include "io_uring.h"
+#include "openclose.h"
+
+struct io_open {
+       struct file                     *file;
+       int                             dfd;
+       u32                             file_slot;
+       struct filename                 *filename;
+       struct open_how                 how;
+       unsigned long                   nofile;
+};
+
+struct io_close {
+       struct file                     *file;
+       int                             fd;
+       u32                             file_slot;
+};
+
+static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+       struct io_open *open = io_kiocb_to_cmd(req);
+       const char __user *fname;
+       int ret;
+
+       if (unlikely(sqe->buf_index))
+               return -EINVAL;
+       if (unlikely(req->flags & REQ_F_FIXED_FILE))
+               return -EBADF;
+
+       /* open.how should be already initialised */
+       if (!(open->how.flags & O_PATH) && force_o_largefile())
+               open->how.flags |= O_LARGEFILE;
+
+       open->dfd = READ_ONCE(sqe->fd);
+       fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
+       open->filename = getname(fname);
+       if (IS_ERR(open->filename)) {
+               ret = PTR_ERR(open->filename);
+               open->filename = NULL;
+               return ret;
+       }
+
+       open->file_slot = READ_ONCE(sqe->file_index);
+       if (open->file_slot && (open->how.flags & O_CLOEXEC))
+               return -EINVAL;
+
+       open->nofile = rlimit(RLIMIT_NOFILE);
+       req->flags |= REQ_F_NEED_CLEANUP;
+       return 0;
+}
+
+int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+       struct io_open *open = io_kiocb_to_cmd(req);
+       u64 mode = READ_ONCE(sqe->len);
+       u64 flags = READ_ONCE(sqe->open_flags);
+
+       open->how = build_open_how(flags, mode);
+       return __io_openat_prep(req, sqe);
+}
+
+int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+       struct io_open *open = io_kiocb_to_cmd(req);
+       struct open_how __user *how;
+       size_t len;
+       int ret;
+
+       how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
+       len = READ_ONCE(sqe->len);
+       if (len < OPEN_HOW_SIZE_VER0)
+               return -EINVAL;
+
+       ret = copy_struct_from_user(&open->how, sizeof(open->how), how, len);
+       if (ret)
+               return ret;
+
+       return __io_openat_prep(req, sqe);
+}
+
+int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
+{
+       struct io_open *open = io_kiocb_to_cmd(req);
+       struct open_flags op;
+       struct file *file;
+       bool resolve_nonblock, nonblock_set;
+       bool fixed = !!open->file_slot;
+       int ret;
+
+       ret = build_open_flags(&open->how, &op);
+       if (ret)
+               goto err;
+       nonblock_set = op.open_flag & O_NONBLOCK;
+       resolve_nonblock = open->how.resolve & RESOLVE_CACHED;
+       if (issue_flags & IO_URING_F_NONBLOCK) {
+               /*
+                * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
+                * it'll always -EAGAIN
+                */
+               if (open->how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
+                       return -EAGAIN;
+               op.lookup_flags |= LOOKUP_CACHED;
+               op.open_flag |= O_NONBLOCK;
+       }
+
+       if (!fixed) {
+               ret = __get_unused_fd_flags(open->how.flags, open->nofile);
+               if (ret < 0)
+                       goto err;
+       }
+
+       file = do_filp_open(open->dfd, open->filename, &op);
+       if (IS_ERR(file)) {
+               /*
+                * We could hang on to this 'fd' on retrying, but seems like
+                * marginal gain for something that is now known to be a slower
+                * path. So just put it, and we'll get a new one when we retry.
+                */
+               if (!fixed)
+                       put_unused_fd(ret);
+
+               ret = PTR_ERR(file);
+               /* only retry if RESOLVE_CACHED wasn't already set by application */
+               if (ret == -EAGAIN &&
+                   (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
+                       return -EAGAIN;
+               goto err;
+       }
+
+       if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
+               file->f_flags &= ~O_NONBLOCK;
+       fsnotify_open(file);
+
+       if (!fixed)
+               fd_install(ret, file);
+       else
+               ret = io_fixed_fd_install(req, issue_flags, file,
+                                               open->file_slot);
+err:
+       putname(open->filename);
+       req->flags &= ~REQ_F_NEED_CLEANUP;
+       if (ret < 0)
+               req_set_fail(req);
+       io_req_set_res(req, ret, 0);
+       return IOU_OK;
+}
+
+int io_openat(struct io_kiocb *req, unsigned int issue_flags)
+{
+       return io_openat2(req, issue_flags);
+}
+
+void io_open_cleanup(struct io_kiocb *req)
+{
+       struct io_open *open = io_kiocb_to_cmd(req);
+
+       if (open->filename)
+               putname(open->filename);
+}
+
+int __io_close_fixed(struct io_kiocb *req, unsigned int issue_flags,
+                    unsigned int offset)
+{
+       struct io_ring_ctx *ctx = req->ctx;
+       struct io_fixed_file *file_slot;
+       struct file *file;
+       int ret;
+
+       io_ring_submit_lock(ctx, issue_flags);
+       ret = -ENXIO;
+       if (unlikely(!ctx->file_data))
+               goto out;
+       ret = -EINVAL;
+       if (offset >= ctx->nr_user_files)
+               goto out;
+       ret = io_rsrc_node_switch_start(ctx);
+       if (ret)
+               goto out;
+
+       offset = array_index_nospec(offset, ctx->nr_user_files);
+       file_slot = io_fixed_file_slot(&ctx->file_table, offset);
+       ret = -EBADF;
+       if (!file_slot->file_ptr)
+               goto out;
+
+       file = (struct file *)(file_slot->file_ptr & FFS_MASK);
+       ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
+       if (ret)
+               goto out;
+
+       file_slot->file_ptr = 0;
+       io_file_bitmap_clear(&ctx->file_table, offset);
+       io_rsrc_node_switch(ctx, ctx->file_data);
+       ret = 0;
+out:
+       io_ring_submit_unlock(ctx, issue_flags);
+       return ret;
+}
+
+static inline int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
+{
+       struct io_close *close = io_kiocb_to_cmd(req);
+
+       return __io_close_fixed(req, issue_flags, close->file_slot - 1);
+}
+
+int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+       struct io_close *close = io_kiocb_to_cmd(req);
+
+       if (sqe->off || sqe->addr || sqe->len || sqe->rw_flags || sqe->buf_index)
+               return -EINVAL;
+       if (req->flags & REQ_F_FIXED_FILE)
+               return -EBADF;
+
+       close->fd = READ_ONCE(sqe->fd);
+       close->file_slot = READ_ONCE(sqe->file_index);
+       if (close->file_slot && close->fd)
+               return -EINVAL;
+
+       return 0;
+}
+
+int io_close(struct io_kiocb *req, unsigned int issue_flags)
+{
+       struct files_struct *files = current->files;
+       struct io_close *close = io_kiocb_to_cmd(req);
+       struct fdtable *fdt;
+       struct file *file;
+       int ret = -EBADF;
+
+       if (close->file_slot) {
+               ret = io_close_fixed(req, issue_flags);
+               goto err;
+       }
+
+       spin_lock(&files->file_lock);
+       fdt = files_fdtable(files);
+       if (close->fd >= fdt->max_fds) {
+               spin_unlock(&files->file_lock);
+               goto err;
+       }
+       file = rcu_dereference_protected(fdt->fd[close->fd],
+                       lockdep_is_held(&files->file_lock));
+       if (!file || io_is_uring_fops(file)) {
+               spin_unlock(&files->file_lock);
+               goto err;
+       }
+
+       /* if the file has a flush method, be safe and punt to async */
+       if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
+               spin_unlock(&files->file_lock);
+               return -EAGAIN;
+       }
+
+       file = __close_fd_get_file(close->fd);
+       spin_unlock(&files->file_lock);
+       if (!file)
+               goto err;
+
+       /* No ->flush() or already async, safely close from here */
+       ret = filp_close(file, current->files);
+err:
+       if (ret < 0)
+               req_set_fail(req);
+       io_req_set_res(req, ret, 0);
+       return IOU_OK;
+}
diff --git a/io_uring/openclose.h b/io_uring/openclose.h
new file mode 100644 (file)
index 0000000..9f578f3
--- /dev/null
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: GPL-2.0
+
+int __io_close_fixed(struct io_kiocb *req, unsigned int issue_flags,
+                    unsigned int offset);
+
+int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
+int io_openat(struct io_kiocb *req, unsigned int issue_flags);
+void io_open_cleanup(struct io_kiocb *req);
+
+int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
+int io_openat2(struct io_kiocb *req, unsigned int issue_flags);
+
+int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
+int io_close(struct io_kiocb *req, unsigned int issue_flags);