io_uring: Implement eBPF iterator for registered buffers

This change adds eBPF iterator for buffers registered in io_uring ctx.
It gives access to the ctx, the index of the registered buffer, and a
pointer to the io_uring_ubuf itself. This allows the iterator to save
info related to buffers added to an io_uring instance, that isn't easy
to export using the fdinfo interface (like exact struct page composing
the registered buffer).

The primary usecase this is enabling is checkpoint/restore support.

Note that we need to use mutex_trylock when the file is read from, in
seq_start functions, as the order of lock taken is opposite of what it
would be when io_uring operation reads the same file.  We take
seq_file->lock, then ctx->uring_lock, while io_uring would first take
ctx->uring_lock and then seq_file->lock for the same ctx.

This can lead to a deadlock scenario described below:

The sequence on CPU 0 is for normal read(2) on iterator.
For CPU 1, it is an io_uring instance trying to do same on iterator attached to
itself.

So CPU 0 does

sys_read
vfs_read
 bpf_seq_read
 mutex_lock(&seq_file->lock)    # A
  io_uring_buf_seq_start
  mutex_lock(&ctx->uring_lock)  # B

and CPU 1 does

io_uring_enter
mutex_lock(&ctx->uring_lock)    # B
 io_read
  bpf_seq_read
  mutex_lock(&seq_file->lock)   # A
  ...

Since the order of locks is opposite, it can deadlock. So we switch the
mutex_lock in io_uring_buf_seq_start to trylock, so it can return an
error for this case, then it will release seq_file->lock and CPU 1 will
make progress.

The trylock also protects the case where io_uring tries to read from
iterator attached to itself (same ctx), where the order of locks would
be:
 io_uring_enter
 mutex_lock(&ctx->uring_lock) <------------.
  io_read				    \
   seq_read				     \
    mutex_lock(&seq_file->lock)		     /
    mutex_lock(&ctx->uring_lock) # deadlock-`

In both these cases (recursive read and contended uring_lock), -EDEADLK
is returned to userspace.

In the future, this iterator will be extended to directly support
iteration of bvec Flexible Array Member, so that when there is no
corresponding VMA that maps to the registered buffer (e.g. if VMA is
destroyed after pinning pages), we are able to reconstruct the
registration on restore by dumping the page contents and then replaying
them into a temporary mapping used for registration later. All this is
out of scope for the current series however, but builds upon this
iterator.

Cc: Jens Axboe <axboe@kernel.dk>
Cc: Pavel Begunkov <asml.silence@gmail.com>
Cc: io-uring@vger.kernel.org
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>

fs/io_uring.c

@@ -81,6 +81,7 @@
 #include <linux/tracehook.h>
 #include <linux/audit.h>
 #include <linux/security.h>
+#include <linux/btf_ids.h>
 
 #define CREATE_TRACE_POINTS
 #include <trace/events/io_uring.h>
@@ -11125,3 +11126,205 @@ static int __init io_uring_init(void)
 	return 0;
 };
 __initcall(io_uring_init);
+
+#ifdef CONFIG_BPF_SYSCALL
+
+BTF_ID_LIST(btf_io_uring_ids)
+BTF_ID(struct, io_ring_ctx)
+BTF_ID(struct, io_mapped_ubuf)
+
+struct bpf_io_uring_seq_info {
+	struct io_ring_ctx *ctx;
+	u64 index;
+};
+
+static int bpf_io_uring_init_seq(void *priv_data, struct bpf_iter_aux_info *aux)
+{
+	struct bpf_io_uring_seq_info *info = priv_data;
+	struct io_ring_ctx *ctx = aux->io_uring.ctx;
+
+	info->ctx = ctx;
+	return 0;
+}
+
+static int bpf_io_uring_iter_attach(struct bpf_prog *prog,
+				    union bpf_iter_link_info *linfo,
+				    struct bpf_iter_aux_info *aux)
+{
+	struct io_ring_ctx *ctx;
+	struct fd f;
+	int ret;
+
+	f = fdget(linfo->io_uring.io_uring_fd);
+	if (unlikely(!f.file))
+		return -EBADF;
+
+	ret = -EOPNOTSUPP;
+	if (unlikely(f.file->f_op != &io_uring_fops))
+		goto out_fput;
+
+	ret = -ENXIO;
+	ctx = f.file->private_data;
+	if (unlikely(!percpu_ref_tryget(&ctx->refs)))
+		goto out_fput;
+
+	ret = 0;
+	aux->io_uring.ctx = ctx;
+	/* each io_uring file's inode is unique, since it uses
+	 * anon_inode_getfile_secure, which can be used to search
+	 * through files and map link fd back to the io_uring.
+	 */
+	aux->io_uring.inode = f.file->f_inode->i_ino;
+
+out_fput:
+	fdput(f);
+	return ret;
+}
+
+static void bpf_io_uring_iter_detach(struct bpf_iter_aux_info *aux)
+{
+	percpu_ref_put(&aux->io_uring.ctx->refs);
+}
+
+#ifdef CONFIG_PROC_FS
+void bpf_io_uring_iter_show_fdinfo(const struct bpf_iter_aux_info *aux,
+				   struct seq_file *seq)
+{
+	seq_printf(seq, "io_uring_inode:\t%lu\n", aux->io_uring.inode);
+}
+#endif
+
+int bpf_io_uring_iter_fill_link_info(const struct bpf_iter_aux_info *aux,
+				     struct bpf_link_info *info)
+{
+	info->iter.io_uring.inode = aux->io_uring.inode;
+	return 0;
+}
+
+/* io_uring iterator for registered buffers */
+
+struct bpf_iter__io_uring_buf {
+	__bpf_md_ptr(struct bpf_iter_meta *, meta);
+	__bpf_md_ptr(struct io_ring_ctx *, ctx);
+	__bpf_md_ptr(struct io_mapped_ubuf *, ubuf);
+	u64 index;
+};
+
+static void *__bpf_io_uring_buf_seq_get_next(struct bpf_io_uring_seq_info *info)
+{
+	if (info->index < info->ctx->nr_user_bufs)
+		return info->ctx->user_bufs[info->index++];
+	return NULL;
+}
+
+static void *bpf_io_uring_buf_seq_start(struct seq_file *seq, loff_t *pos)
+{
+	struct bpf_io_uring_seq_info *info = seq->private;
+	struct io_mapped_ubuf *ubuf;
+
+	/* Indicate to userspace that the uring lock is contended */
+	if (!mutex_trylock(&info->ctx->uring_lock))
+		return ERR_PTR(-EDEADLK);
+
+	ubuf = __bpf_io_uring_buf_seq_get_next(info);
+	if (!ubuf)
+		return NULL;
+
+	if (*pos == 0)
+		++*pos;
+	return ubuf;
+}
+
+static void *bpf_io_uring_buf_seq_next(struct seq_file *seq, void *v, loff_t *pos)
+{
+	struct bpf_io_uring_seq_info *info = seq->private;
+
+	++*pos;
+	return __bpf_io_uring_buf_seq_get_next(info);
+}
+
+DEFINE_BPF_ITER_FUNC(io_uring_buf, struct bpf_iter_meta *meta,
+		     struct io_ring_ctx *ctx, struct io_mapped_ubuf *ubuf,
+		     u64 index)
+
+static int __bpf_io_uring_buf_seq_show(struct seq_file *seq, void *v, bool in_stop)
+{
+	struct bpf_io_uring_seq_info *info = seq->private;
+	struct bpf_iter__io_uring_buf ctx;
+	struct bpf_iter_meta meta;
+	struct bpf_prog *prog;
+
+	meta.seq = seq;
+	prog = bpf_iter_get_info(&meta, in_stop);
+	if (!prog)
+		return 0;
+
+	ctx.meta = &meta;
+	ctx.ctx = info->ctx;
+	ctx.ubuf = v;
+	ctx.index = info->index ? info->index - !in_stop : 0;
+
+	return bpf_iter_run_prog(prog, &ctx);
+}
+
+static int bpf_io_uring_buf_seq_show(struct seq_file *seq, void *v)
+{
+	return __bpf_io_uring_buf_seq_show(seq, v, false);
+}
+
+static void bpf_io_uring_buf_seq_stop(struct seq_file *seq, void *v)
+{
+	struct bpf_io_uring_seq_info *info = seq->private;
+
+	/* If IS_ERR(v) is true, then ctx->uring_lock wasn't taken */
+	if (IS_ERR(v))
+		return;
+	if (!v)
+		__bpf_io_uring_buf_seq_show(seq, v, true);
+	else if (info->index) /* restart from index */
+		info->index--;
+	mutex_unlock(&info->ctx->uring_lock);
+}
+
+static const struct seq_operations bpf_io_uring_buf_seq_ops = {
+	.start = bpf_io_uring_buf_seq_start,
+	.next  = bpf_io_uring_buf_seq_next,
+	.stop  = bpf_io_uring_buf_seq_stop,
+	.show  = bpf_io_uring_buf_seq_show,
+};
+
+static const struct bpf_iter_seq_info bpf_io_uring_buf_seq_info = {
+	.seq_ops          = &bpf_io_uring_buf_seq_ops,
+	.init_seq_private = bpf_io_uring_init_seq,
+	.fini_seq_private = NULL,
+	.seq_priv_size    = sizeof(struct bpf_io_uring_seq_info),
+};
+
+static struct bpf_iter_reg io_uring_buf_reg_info = {
+	.target            = "io_uring_buf",
+	.feature	   = BPF_ITER_RESCHED,
+	.attach_target     = bpf_io_uring_iter_attach,
+	.detach_target     = bpf_io_uring_iter_detach,
+#ifdef CONFIG_PROC_FS
+	.show_fdinfo	   = bpf_io_uring_iter_show_fdinfo,
+#endif
+	.fill_link_info    = bpf_io_uring_iter_fill_link_info,
+	.ctx_arg_info_size = 2,
+	.ctx_arg_info = {
+		{ offsetof(struct bpf_iter__io_uring_buf, ctx),
+		  PTR_TO_BTF_ID },
+		{ offsetof(struct bpf_iter__io_uring_buf, ubuf),
+		  PTR_TO_BTF_ID_OR_NULL },
+	},
+	.seq_info	   = &bpf_io_uring_buf_seq_info,
+};
+
+static int __init io_uring_iter_init(void)
+{
+	io_uring_buf_reg_info.ctx_arg_info[0].btf_id = btf_io_uring_ids[0];
+	io_uring_buf_reg_info.ctx_arg_info[1].btf_id = btf_io_uring_ids[1];
+	return bpf_iter_reg_target(&io_uring_buf_reg_info);
+}
+late_initcall(io_uring_iter_init);
+
+#endif

include/linux/bpf.h

@@ -1509,8 +1509,20 @@ int bpf_obj_get_user(const char __user *pathname, int flags);
 	extern int bpf_iter_ ## target(args);			\
 	int __init bpf_iter_ ## target(args) { return 0; }
 
+struct io_ring_ctx;
+
 struct bpf_iter_aux_info {
+	/* Map member must not alias any other members, due to the check in
+	 * bpf_trace.c:__get_seq_info, since in case of map the seq_ops for
+	 * iterator is different from others. The seq_ops is not from main
+	 * iter registration but from map_ops. Nullability of 'map' allows
+	 * to skip this check for non-map iterator cheaply.
+	 */
 	struct bpf_map *map;
+	struct {
+		struct io_ring_ctx *ctx;
+		ino_t inode;
+	} io_uring;
 };
 
 typedef int (*bpf_iter_attach_target_t)(struct bpf_prog *prog,

include/uapi/linux/bpf.h

@@ -91,6 +91,9 @@ union bpf_iter_link_info {
 	struct {
 		__u32	map_fd;
 	} map;
+	struct {
+		__u32   io_uring_fd;
+	} io_uring;
 };
 
 /* BPF syscall commands, see bpf(2) man-page for more details. */
@@ -5720,6 +5723,9 @@ struct bpf_link_info {
 				struct {
 					__u32 map_id;
 				} map;
+				struct {
+					__u64 inode;
+				} io_uring;
 			};
 		} iter;
 		struct  {

tools/include/uapi/linux/bpf.h

@@ -91,6 +91,9 @@ union bpf_iter_link_info {
 	struct {
 		__u32	map_fd;
 	} map;
+	struct {
+		__u32   io_uring_fd;
+	} io_uring;
 };
 
 /* BPF syscall commands, see bpf(2) man-page for more details. */
@@ -5720,6 +5723,9 @@ struct bpf_link_info {
 				struct {
 					__u32 map_id;
 				} map;
+				struct {
+					__u64 inode;
+				} io_uring;
 			};
 		} iter;
 		struct  {