[PATCH bpf-next v2 0/5] File descriptor interface for BPF streams

[PATCH bpf-next v2 0/5] File descriptor interface for BPF streams

[Kumar Kartikeya Dwivedi]

BPF program stdout and stderr streams can currently be consumed only
through BPF_PROG_STREAM_READ_BY_FD. This requires repeated bpf() calls and
provides no way to block for output or integrate with poll-based event
loops.

Add BPF_PROG_STREAM_OPEN to return a read-only, close-on-exec descriptor
for a program stream. Reads block by default, with an option for
non-blocking mode. poll and epoll report readable data and report hangup
when the program is freed, while allowing buffered data to be drained
before EOF. Stream descriptors retain the stream storage without retaining
the program itself.

Expose the command through libbpf and switch bpftool tracelog to blocking
stream reads. bpftool falls back to BPF_PROG_STREAM_READ_BY_FD on older
kernels, where an unknown bpf() command returns EINVAL. The legacy command
remains supported.

The first patch fixes a pre-existing issue where empty stream writes
allocate elements that escape capacity accounting; the readiness tracking
added later relies on every element carrying data.

See commit logs for details.

Changelog:
----------
v1 -> v2
v1: https://lore.kernel.org/bpf/20260830093514.4105972-1-memxor@gmail.com

 * Fold the irq_work and readable-counter patches into the interface patch
   so no intermediate state wakes waiters directly or derives readiness
   from the capacity counter. (BPF CI)
 * Accept only BPF_F_STREAM_NONBLOCK; BPF_F_RDONLY is no longer accepted
   since the descriptor is always read-only.
 * Allocate streams only for programs loaded through BPF_PROG_LOAD, not for
   classic BPF filters, JIT subprograms or shim programs.
 * Add a patch that skips zero-length stream writes instead of allocating
   elements that bypass capacity accounting. (Emil, Sashiko)
 * Report EOF only when the stream was already dead before it was found
   empty, so data published right before program teardown is not lost.
   (Sashiko)
 * Document that POLLHUP follows program destruction, not the caller's own
   release, and that it may lag the final reference drop.
 * Drop the llseek operation so lseek fails with ESPIPE like other stream
   descriptors. (Emil)
 * Drop the unreachable length check in the file read path; the VFS caps
   read sizes below INT_MAX.
 * Let __bpf_prog_free() clean up after a failed stream allocation instead
   of freeing the streams twice, and drop redundant zeroing of the stream
   counters after kzalloc().
 * Explain why wakeups are always deferred through irq_work and drop the
   batching rationale. (Emil)
 * Skip irq_work_sync() at teardown for streams that never queued a
   notification; on PREEMPT_RT it waits for an RCU grace period that every
   program free, including cBPF filters, would otherwise pay. (Sashiko)
 * Qualify the POLLIN followed by EAGAIN claim to a single reader. (BPF CI)
 * Handle SIGINT, SIGHUP and SIGTERM in bpftool so a followed stream exits
   cleanly, and document the follow behavior. (Emil, BPF CI)
 * Drop bpftool's program reference once the stream is open so following
   ends with EOF when the program is unloaded, and restore the previous
   signal dispositions afterwards for batch mode. (Sashiko)
 * Report bpftool read errors on the fallback path as well.
 * Test lseek rejection and empty stream writes.
 * Retry the NMI test write on a later sample if the first attempt fails.
 * Drop Emil's Reviewed-by from the patches that changed.
 * Rebase on bpf-next.

Kumar Kartikeya Dwivedi (5):
  bpf: Skip zero-length stream writes
  bpf: Add file descriptor interface for program streams
  libbpf: Add bpf_prog_stream_open()
  bpftool: Read program streams through file descriptors
  selftests/bpf: Test program stream file descriptors

 include/linux/bpf.h                           |  14 +-
 include/uapi/linux/bpf.h                      |  39 ++
 kernel/bpf/core.c                             |   8 +-
 kernel/bpf/stream.c                           | 211 ++++++++-
 kernel/bpf/syscall.c                          |  29 ++
 .../bpftool/Documentation/bpftool-prog.rst    |   5 +
 tools/bpf/bpftool/prog.c                      |  62 ++-
 tools/include/uapi/linux/bpf.h                |  39 ++
 tools/lib/bpf/bpf.c                           |  19 +
 tools/lib/bpf/bpf.h                           |  24 +
 tools/lib/bpf/libbpf.map                      |   1 +
 .../testing/selftests/bpf/prog_tests/stream.c | 415 ++++++++++++++++++
 tools/testing/selftests/bpf/progs/stream.c    |  20 +
 13 files changed, 853 insertions(+), 33 deletions(-)


base-commit: 740eb74b8d2e3b7850e55b15257c8c0b0fc4a125
-- 
2.53.0

[PATCH bpf-next v2 1/5] bpf: Skip zero-length stream writes

[Kumar Kartikeya Dwivedi]

A stream write whose formatted output is empty still allocates a stream
element and links it into the stream log. Capacity accounting only counts
payload bytes, so such elements never count against the stream limit, and a
program that keeps producing empty output allocates elements without bound
until a reader drains them.

Return success without allocating anything when there is nothing to write,
for both bpf_stream_vprintk() and staged writes. Every element in a stream
log then carries at least one byte, which later changes rely on when they
derive read readiness from published byte counts.

Fixes: 5ab154f1463a ("bpf: Introduce BPF standard streams")
Suggested-by: Emil Tsalapatis <emil@etsalapatis.com>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 kernel/bpf/stream.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/kernel/bpf/stream.c b/kernel/bpf/stream.c
index 2b80a0599865..9829d1ebce70 100644
--- a/kernel/bpf/stream.c
+++ b/kernel/bpf/stream.c
@@ -75,8 +75,13 @@ static void bpf_stream_release_capacity(struct bpf_stream *stream, int len)
 
 static int bpf_stream_push_str(struct bpf_stream *stream, const char *str, int len)
 {
-	int ret = bpf_stream_consume_capacity(stream, len);
+	int ret;
+
+	/* Nothing to publish; do not allocate an element for it. */
+	if (!len)
+		return 0;
 
+	ret = bpf_stream_consume_capacity(stream, len);
 	if (ret)
 		return ret;
 
@@ -333,8 +338,8 @@ int bpf_stream_stage_printk(struct bpf_stream_stage *ss, const char *fmt, ...)
 	va_start(args, fmt);
 	len = vscnprintf(buf->buf, ARRAY_SIZE(buf->buf), fmt, args);
 	va_end(args);
-	/* Exclude NULL byte during push. */
-	ret = __bpf_stream_push_str(&ss->log, buf->buf, len);
+	/* Exclude NULL byte during push; skip empty output entirely. */
+	ret = len ? __bpf_stream_push_str(&ss->log, buf->buf, len) : 0;
 	if (!ret)
 		ss->len += len;
 	bpf_put_buffers();
-- 
2.53.0

[PATCH bpf-next v2 2/5] bpf: Add file descriptor interface for program streams

[Kumar Kartikeya Dwivedi]

The existing BPF_PROG_STREAM_READ_BY_FD command only supports polling a
program stream through repeated bpf() calls. It cannot block for new data
or integrate with poll-based event loops.

Add BPF_PROG_STREAM_OPEN to return a read-only, close-on-exec file
descriptor for a selected program stream. Reads block by default and
BPF_F_STREAM_NONBLOCK, the only accepted flag, provides non-blocking
behavior. poll reports readable data and reports hangup once the program
has been freed. Like pipes and sockets, the descriptor is not seekable and
lseek fails with ESPIPE.

A stream descriptor deliberately does not retain the program. Move each
stream into a separately refcounted allocation so program teardown can mark
it dead and wake descriptor users while outstanding descriptors drain
buffered data safely. Readers sample the dead flag before looking for data,
so EOF is reported only when the stream was already dead before it was
found empty; data published right before teardown is never skipped.

Only programs loaded through BPF_PROG_LOAD get streams. Classic BPF
filters, JIT subprograms and shim programs never write to one, and
kernel-side writers already resolve a subprogram to its main program, so
those programs no longer carry stream state.

Readiness needs its own counter. Stream capacity is charged before
allocation and before an element is published to the stream log, so using
that reservation as the read and poll condition can report readable data
while no element exists: a blocking reader retries instead of sleeping and
a lone non-blocking reader can see POLLIN followed by EAGAIN. Publish bytes
with release ordering after adding elements to the lockless log, use
acquire loads before consuming them or reporting readiness, limit each read
to its readable snapshot and subtract only bytes actually copied. This
keeps the aggregate count correct even when concurrent publishers update it
out of publication order. With several readers on one stream, readiness
remains advisory, as it is for pipes. The capacity counter is kept solely
for enforcing the stream size limit.

Wakeups are always deferred through irq_work. Stream writers run in
whatever context the program runs in: NMI context for perf_event programs,
sections with interrupts disabled inside bpf_spin_lock or rqspinlock
critical sections since bpf_stream_vprintk() is KF_SPINLOCK_SAFE, and
tracing programs attached anywhere in the kernel, including inside the wait
queue and epoll code itself. Waking waiters directly from there can
deadlock, and no cheap context check covers every case: on PREEMPT_RT,
spinlock_t sections do not disable interrupts, so in_nmi() or
irqs_disabled() cannot tell such a program apart from a benign one. Queue
an irq_work item after publishing instead, as bpf_ringbuf does. Coalescing
of concurrent publications is a side effect that is compatible with poll
and epoll semantics: notifications prompt waiters to re-evaluate persistent
readable state, and EPOLLET consumers drain until EAGAIN.

Synchronize pending work before releasing the final stream reference so
the callback cannot outlive the stream, but only when the work was ever
queued: irq_work_sync() waits for an RCU grace period on PREEMPT_RT and on
architectures without an irq_work interrupt.

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 include/linux/bpf.h            |  14 ++-
 include/uapi/linux/bpf.h       |  39 +++++++
 kernel/bpf/core.c              |   8 +-
 kernel/bpf/stream.c            | 200 ++++++++++++++++++++++++++++++---
 kernel/bpf/syscall.c           |  29 +++++
 tools/include/uapi/linux/bpf.h |  39 +++++++
 6 files changed, 306 insertions(+), 23 deletions(-)

diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index 1e1ce2afe2ed..a38badb1c2f7 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -17,6 +17,7 @@
 #include <linux/numa.h>
 #include <linux/mm_types.h>
 #include <linux/wait.h>
+#include <linux/irq_work_types.h>
 #include <linux/refcount.h>
 #include <linux/mutex.h>
 #include <linux/module.h>
@@ -1770,12 +1771,18 @@ enum {
 };
 
 struct bpf_stream {
-	atomic_t capacity;
+	refcount_t refcnt;
+	atomic_t capacity;	/* bytes reserved against the stream limit */
+	atomic_t readable;	/* published bytes available to readers */
 	struct llist_head log;	/* list of in-flight stream elements in LIFO order */
 
 	struct mutex lock;  /* lock protecting backlog_{head,tail} */
 	struct llist_node *backlog_head; /* list of in-flight stream elements in FIFO order */
 	struct llist_node *backlog_tail; /* tail of the list above */
+	wait_queue_head_t waitq;
+	struct irq_work notify_work;
+	bool notify_used;	/* notify_work was queued at least once */
+	bool dead;
 };
 
 struct bpf_stream_stage {
@@ -1914,7 +1921,7 @@ struct bpf_prog_aux {
 		struct work_struct work;
 		struct rcu_head	rcu;
 	};
-	struct bpf_stream stream[2];
+	struct bpf_stream *stream[2];
 	struct mutex st_ops_assoc_mutex;
 	struct bpf_map __rcu *st_ops_assoc;
 };
@@ -4223,9 +4230,10 @@ void bpf_bprintf_cleanup(struct bpf_bprintf_data *data);
 int bpf_try_get_buffers(struct bpf_bprintf_buffers **bufs);
 void bpf_put_buffers(void);
 
-void bpf_prog_stream_init(struct bpf_prog *prog);
+int bpf_prog_stream_init(struct bpf_prog *prog, gfp_t gfp_extra_flags);
 void bpf_prog_stream_free(struct bpf_prog *prog);
 int bpf_prog_stream_read(struct bpf_prog *prog, enum bpf_stream_id stream_id, void __user *buf, u32 len);
+int bpf_prog_stream_new_fd(struct bpf_prog *prog, enum bpf_stream_id stream_id, u32 flags);
 void bpf_stream_stage_init(struct bpf_stream_stage *ss);
 void bpf_stream_stage_free(struct bpf_stream_stage *ss);
 __printf(2, 3)
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index 0aaa54359aeb..4687c3310996 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -936,6 +936,33 @@ union bpf_iter_link_info {
  * 		0 on success or -1 if an error occurred (in which case,
  * 		*errno* is set appropriately).
  *
+ * BPF_PROG_STREAM_OPEN
+ *	Description
+ *		Open a file descriptor for one of the BPF streams associated
+ *		with the program identified by *prog_fd*. The stream is selected
+ *		by *stream_id*.
+ *
+ *		The returned file descriptor supports **read**\ (2) and
+ *		**poll**\ (2). Reads block while the stream is empty unless
+ *		**BPF_F_STREAM_NONBLOCK** is specified in *flags*. A non-blocking
+ *		read of an empty stream fails with **EAGAIN**.
+ *
+ *		**poll**\ (2) reports **POLLIN** when data is available and
+ *		**POLLHUP** once the program has been freed, that is, after every
+ *		reference to it, including links and other file descriptors, has
+ *		been dropped. Hangup may lag the final release because program
+ *		teardown is deferred. Buffered data remains readable after
+ *		**POLLHUP** and a read returns zero after all such data has been
+ *		consumed.
+ *
+ *		The file descriptor is read-only and has the close-on-exec flag
+ *		set. It is not seekable and **lseek**\ (2) fails with **ESPIPE**.
+ *		*flags* may only contain **BPF_F_STREAM_NONBLOCK**.
+ *
+ *	Return
+ *		A new file descriptor (a nonnegative integer), or -1 if an
+ *		error occurred (in which case, *errno* is set appropriately).
+ *
  * NOTES
  *	eBPF objects (maps and programs) can be shared between processes.
  *
@@ -993,6 +1020,7 @@ enum bpf_cmd {
 	BPF_TOKEN_CREATE,
 	BPF_PROG_STREAM_READ_BY_FD,
 	BPF_PROG_ASSOC_STRUCT_OPS,
+	BPF_PROG_STREAM_OPEN,
 	__MAX_BPF_CMD,
 	BPF_COMMON_ATTRS = 1 << 16, /* Indicate carrying syscall common attrs. */
 };
@@ -1524,6 +1552,11 @@ enum {
 	BPF_STREAM_STDERR = 2,
 };
 
+/* flags for BPF_PROG_STREAM_OPEN command */
+enum {
+	BPF_F_STREAM_NONBLOCK = (1U << 0),
+};
+
 union bpf_attr {
 	struct { /* anonymous struct used by BPF_MAP_CREATE command */
 		__u32	map_type;	/* one of enum bpf_map_type */
@@ -1950,6 +1983,12 @@ union bpf_attr {
 		__u32		flags;
 	} prog_assoc_struct_ops;
 
+	struct {
+		__u32		prog_fd;
+		__u32		stream_id;
+		__u32		flags;
+	} prog_stream_open;
+
 } __attribute__((aligned(8)));
 
 /* The description below is an attempt at providing documentation to eBPF
diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index 273f74068068..134700f42531 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c
@@ -141,10 +141,6 @@ struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flag
 	mutex_init(&fp->aux->dst_mutex);
 	mutex_init(&fp->aux->st_ops_assoc_mutex);
 
-#ifdef CONFIG_BPF_SYSCALL
-	bpf_prog_stream_init(fp);
-#endif
-
 	return fp;
 }
 
@@ -288,6 +284,9 @@ struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
 void __bpf_prog_free(struct bpf_prog *fp)
 {
 	if (fp->aux) {
+#ifdef CONFIG_BPF_SYSCALL
+		bpf_prog_stream_free(fp);
+#endif
 		mutex_destroy(&fp->aux->used_maps_mutex);
 		mutex_destroy(&fp->aux->dst_mutex);
 		mutex_destroy(&fp->aux->st_ops_assoc_mutex);
@@ -3073,7 +3072,6 @@ static void bpf_prog_free_deferred(struct work_struct *work)
 	aux = container_of(work, struct bpf_prog_aux, work);
 #ifdef CONFIG_BPF_SYSCALL
 	bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
-	bpf_prog_stream_free(aux->prog);
 #endif
 #ifdef CONFIG_CGROUP_BPF
 	if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
diff --git a/kernel/bpf/stream.c b/kernel/bpf/stream.c
index 9829d1ebce70..e0ce37fcd557 100644
--- a/kernel/bpf/stream.c
+++ b/kernel/bpf/stream.c
@@ -2,11 +2,15 @@
 /* Copyright (c) 2025 Meta Platforms, Inc. and affiliates. */
 
 #include <linux/bpf.h>
+#include <linux/anon_inodes.h>
 #include <linux/filter.h>
 #include <linux/bpf_mem_alloc.h>
 #include <linux/gfp.h>
+#include <linux/irq_work.h>
 #include <linux/memory.h>
 #include <linux/mutex.h>
+#include <linux/poll.h>
+#include <linux/refcount.h>
 
 static void bpf_stream_elem_init(struct bpf_stream_elem *elem, int len)
 {
@@ -73,6 +77,42 @@ static void bpf_stream_release_capacity(struct bpf_stream *stream, int len)
 	atomic_sub(len, &stream->capacity);
 }
 
+static void bpf_stream_notify(struct irq_work *work)
+{
+	struct bpf_stream *stream = container_of(work, struct bpf_stream, notify_work);
+
+	/*
+	 * Writers run in arbitrary program contexts, including NMI and regions
+	 * that already hold wait queue or epoll locks. Wake waiters from
+	 * irq_work instead, where taking those locks is safe.
+	 */
+	wake_up_interruptible_poll(&stream->waitq, EPOLLIN | EPOLLRDNORM);
+}
+
+static void bpf_stream_queue_notify(struct bpf_stream *stream)
+{
+	/*
+	 * Record that the work has been used so that teardown only pays for
+	 * irq_work_sync(), which may wait for an RCU grace period, when a
+	 * callback could actually be in flight.
+	 */
+	if (!READ_ONCE(stream->notify_used))
+		WRITE_ONCE(stream->notify_used, true);
+	irq_work_queue(&stream->notify_work);
+}
+
+static int bpf_stream_readable_bytes(struct bpf_stream *stream)
+{
+	return atomic_read_acquire(&stream->readable);
+}
+
+static void bpf_stream_publish(struct bpf_stream *stream, int len)
+{
+	/* Pairs with atomic_read_acquire() in bpf_stream_readable_bytes(). */
+	(void)atomic_add_return_release(len, &stream->readable);
+	bpf_stream_queue_notify(stream);
+}
+
 static int bpf_stream_push_str(struct bpf_stream *stream, const char *str, int len)
 {
 	int ret;
@@ -88,6 +128,8 @@ static int bpf_stream_push_str(struct bpf_stream *stream, const char *str, int l
 	ret = __bpf_stream_push_str(&stream->log, str, len);
 	if (ret)
 		bpf_stream_release_capacity(stream, len);
+	else
+		bpf_stream_publish(stream, len);
 
 	return ret;
 }
@@ -96,7 +138,7 @@ static struct bpf_stream *bpf_stream_get(enum bpf_stream_id stream_id, struct bp
 {
 	if (stream_id != BPF_STDOUT && stream_id != BPF_STDERR)
 		return NULL;
-	return &aux->stream[stream_id - 1];
+	return aux->stream[stream_id - 1];
 }
 
 static void bpf_stream_free_elem(struct bpf_stream_elem *elem)
@@ -164,14 +206,16 @@ static bool bpf_stream_consume_elem(struct bpf_stream_elem *elem, int *len)
 
 static int bpf_stream_read(struct bpf_stream *stream, void __user *buf, int len)
 {
-	int rem_len = len, cons_len, ret = 0;
+	int read_len, rem_len, cons_len, ret = 0;
 	struct bpf_stream_elem *elem = NULL;
 	struct llist_node *node;
 
 	mutex_lock(&stream->lock);
+	read_len = min(len, bpf_stream_readable_bytes(stream));
+	rem_len = read_len;
 
 	while (rem_len) {
-		int pos = len - rem_len;
+		int pos = read_len - rem_len;
 		int chunk, n;
 		bool cont;
 
@@ -193,7 +237,7 @@ static int bpf_stream_read(struct bpf_stream *stream, void __user *buf, int len)
 			/* Keep any successfully copied bytes; -EFAULT only if none. */
 			elem->consumed_len -= n;
 			rem_len += n;
-			ret = (len == rem_len) ? -EFAULT : 0;
+			ret = (read_len == rem_len) ? -EFAULT : 0;
 			break;
 		}
 
@@ -204,8 +248,9 @@ static int bpf_stream_read(struct bpf_stream *stream, void __user *buf, int len)
 		bpf_stream_free_elem(elem);
 	}
 
+	atomic_sub(read_len - rem_len, &stream->readable);
 	mutex_unlock(&stream->lock);
-	return ret ? ret : len - rem_len;
+	return ret ? ret : read_len - rem_len;
 }
 
 int bpf_prog_stream_read(struct bpf_prog *prog, enum bpf_stream_id stream_id, void __user *buf, u32 len)
@@ -220,6 +265,111 @@ int bpf_prog_stream_read(struct bpf_prog *prog, enum bpf_stream_id stream_id, vo
 	return bpf_stream_read(stream, buf, len);
 }
 
+static bool bpf_stream_has_data(struct bpf_stream *stream)
+{
+	return bpf_stream_readable_bytes(stream) > 0;
+}
+
+static void bpf_stream_put(struct bpf_stream *stream)
+{
+	if (refcount_dec_and_test(&stream->refcnt)) {
+		struct llist_node *list;
+
+		/* Only a stream that ever queued its work can have a callback in flight. */
+		if (READ_ONCE(stream->notify_used))
+			irq_work_sync(&stream->notify_work);
+		list = llist_del_all(&stream->log);
+		bpf_stream_free_list(list);
+		bpf_stream_free_list(stream->backlog_head);
+		mutex_destroy(&stream->lock);
+		kfree(stream);
+	}
+}
+
+static int bpf_stream_release(struct inode *inode, struct file *file)
+{
+	bpf_stream_put(file->private_data);
+	return 0;
+}
+
+static ssize_t bpf_stream_file_read(struct file *file, char __user *buf, size_t len,
+				    loff_t *ppos)
+{
+	struct bpf_stream *stream = file->private_data;
+	bool dead;
+	int ret;
+
+	if (!len)
+		return 0;
+
+	for (;;) {
+		/*
+		 * Sample teardown state before looking for data. Nothing is
+		 * published once the program is gone, so finding the stream
+		 * empty after observing dead means EOF. The opposite order could
+		 * report EOF while data published just before teardown is still
+		 * buffered.
+		 */
+		dead = smp_load_acquire(&stream->dead);
+		ret = bpf_stream_read(stream, buf, len);
+		if (ret)
+			return ret;
+		if (dead)
+			return 0;
+		if (file->f_flags & O_NONBLOCK)
+			return -EAGAIN;
+
+		ret = wait_event_interruptible(stream->waitq,
+					       bpf_stream_has_data(stream) ||
+					       READ_ONCE(stream->dead));
+		if (ret)
+			return ret;
+	}
+}
+
+static __poll_t bpf_stream_poll(struct file *file, struct poll_table_struct *pts)
+{
+	struct bpf_stream *stream = file->private_data;
+	__poll_t events = 0;
+
+	/*
+	 * poll_wait() only registers the wait queue callback. Register before
+	 * checking persistent state so a concurrent publication or teardown is
+	 * observed either by the callback or by the checks below.
+	 */
+	poll_wait(file, &stream->waitq, pts);
+	if (bpf_stream_has_data(stream))
+		events |= EPOLLIN | EPOLLRDNORM;
+	if (READ_ONCE(stream->dead))
+		events |= EPOLLHUP;
+	return events;
+}
+
+static const struct file_operations bpf_stream_fops = {
+	.release = bpf_stream_release,
+	.read = bpf_stream_file_read,
+	.poll = bpf_stream_poll,
+};
+
+int bpf_prog_stream_new_fd(struct bpf_prog *prog, enum bpf_stream_id stream_id, u32 flags)
+{
+	struct bpf_stream *stream;
+	int fd_flags = O_RDONLY | O_CLOEXEC;
+	int fd;
+
+	stream = bpf_stream_get(stream_id, prog->aux);
+	if (!stream)
+		return -ENOENT;
+	if (flags & BPF_F_STREAM_NONBLOCK)
+		fd_flags |= O_NONBLOCK;
+
+	refcount_inc(&stream->refcnt);
+	fd = anon_inode_getfd("bpf-stream", &bpf_stream_fops, stream, fd_flags);
+	if (fd < 0)
+		bpf_stream_put(stream);
+	return fd;
+}
+
 __bpf_kfunc_start_defs();
 
 /*
@@ -287,28 +437,47 @@ __bpf_kfunc_end_defs();
 
 /* Added kfunc to common_btf_ids */
 
-void bpf_prog_stream_init(struct bpf_prog *prog)
+int bpf_prog_stream_init(struct bpf_prog *prog, gfp_t gfp_extra_flags)
 {
 	int i;
 
 	for (i = 0; i < ARRAY_SIZE(prog->aux->stream); i++) {
-		atomic_set(&prog->aux->stream[i].capacity, 0);
-		init_llist_head(&prog->aux->stream[i].log);
-		mutex_init(&prog->aux->stream[i].lock);
-		prog->aux->stream[i].backlog_head = NULL;
-		prog->aux->stream[i].backlog_tail = NULL;
+		struct bpf_stream *stream;
+
+		/* On failure, bpf_prog_stream_free() releases the streams allocated so far. */
+		stream = kzalloc_obj(*stream,
+				     bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
+		if (!stream)
+			return -ENOMEM;
+
+		refcount_set(&stream->refcnt, 1);
+		init_llist_head(&stream->log);
+		mutex_init(&stream->lock);
+		init_waitqueue_head(&stream->waitq);
+		init_irq_work(&stream->notify_work, bpf_stream_notify);
+		prog->aux->stream[i] = stream;
 	}
+	return 0;
 }
 
 void bpf_prog_stream_free(struct bpf_prog *prog)
 {
-	struct llist_node *list;
 	int i;
 
 	for (i = 0; i < ARRAY_SIZE(prog->aux->stream); i++) {
-		list = llist_del_all(&prog->aux->stream[i].log);
-		bpf_stream_free_list(list);
-		bpf_stream_free_list(prog->aux->stream[i].backlog_head);
+		struct bpf_stream *stream = prog->aux->stream[i];
+
+		if (!stream)
+			continue;
+		/*
+		 * Pairs with smp_load_acquire() in bpf_stream_file_read(): every
+		 * publication precedes the dead flag, so a reader that observes
+		 * it also observes all buffered data.
+		 */
+		smp_store_release(&stream->dead, true);
+		wake_up_interruptible_poll(&stream->waitq, EPOLLHUP);
+		bpf_stream_put(stream);
+		prog->aux->stream[i] = NULL;
 	}
 }
 
@@ -371,6 +540,7 @@ int bpf_stream_stage_commit(struct bpf_stream_stage *ss, struct bpf_prog *prog,
 		list = tail;
 	}
 	llist_add_batch(head, tail, &stream->log);
+	bpf_stream_publish(stream, ss->len);
 	return 0;
 }
 
diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
index 74496fd716d3..8580f38b41c1 100644
--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c
@@ -3080,6 +3080,10 @@ static int bpf_prog_load(union bpf_attr *attr, bpfptr_t uattr, struct bpf_log_at
 	prog->aux->user = get_current_user();
 	prog->len = attr->insn_cnt;
 
+	err = bpf_prog_stream_init(prog, GFP_USER);
+	if (err)
+		goto free_prog;
+
 	err = -EFAULT;
 	if (copy_from_bpfptr(prog->insns,
 			     make_bpfptr(attr->insns, uattr.is_kernel),
@@ -6313,6 +6317,28 @@ static int prog_assoc_struct_ops(union bpf_attr *attr)
 	return ret;
 }
 
+#define BPF_PROG_STREAM_OPEN_LAST_FIELD prog_stream_open.flags
+
+static int prog_stream_open(union bpf_attr *attr)
+{
+	struct bpf_prog *prog;
+	u32 flags = attr->prog_stream_open.flags;
+	int ret;
+
+	if (CHECK_ATTR(BPF_PROG_STREAM_OPEN))
+		return -EINVAL;
+	if (flags & ~BPF_F_STREAM_NONBLOCK)
+		return -EINVAL;
+
+	prog = bpf_prog_get(attr->prog_stream_open.prog_fd);
+	if (IS_ERR(prog))
+		return PTR_ERR(prog);
+
+	ret = bpf_prog_stream_new_fd(prog, attr->prog_stream_open.stream_id, flags);
+	bpf_prog_put(prog);
+	return ret;
+}
+
 static int __sys_bpf(enum bpf_cmd cmd, bpfptr_t uattr, unsigned int size,
 		     bpfptr_t uattr_common, unsigned int size_common)
 {
@@ -6485,6 +6511,9 @@ static int __sys_bpf(enum bpf_cmd cmd, bpfptr_t uattr, unsigned int size,
 	case BPF_PROG_ASSOC_STRUCT_OPS:
 		err = prog_assoc_struct_ops(&attr);
 		break;
+	case BPF_PROG_STREAM_OPEN:
+		err = prog_stream_open(&attr);
+		break;
 	default:
 		err = -EINVAL;
 		break;
diff --git a/tools/include/uapi/linux/bpf.h b/tools/include/uapi/linux/bpf.h
index 0aaa54359aeb..4687c3310996 100644
--- a/tools/include/uapi/linux/bpf.h
+++ b/tools/include/uapi/linux/bpf.h
@@ -936,6 +936,33 @@ union bpf_iter_link_info {
  * 		0 on success or -1 if an error occurred (in which case,
  * 		*errno* is set appropriately).
  *
+ * BPF_PROG_STREAM_OPEN
+ *	Description
+ *		Open a file descriptor for one of the BPF streams associated
+ *		with the program identified by *prog_fd*. The stream is selected
+ *		by *stream_id*.
+ *
+ *		The returned file descriptor supports **read**\ (2) and
+ *		**poll**\ (2). Reads block while the stream is empty unless
+ *		**BPF_F_STREAM_NONBLOCK** is specified in *flags*. A non-blocking
+ *		read of an empty stream fails with **EAGAIN**.
+ *
+ *		**poll**\ (2) reports **POLLIN** when data is available and
+ *		**POLLHUP** once the program has been freed, that is, after every
+ *		reference to it, including links and other file descriptors, has
+ *		been dropped. Hangup may lag the final release because program
+ *		teardown is deferred. Buffered data remains readable after
+ *		**POLLHUP** and a read returns zero after all such data has been
+ *		consumed.
+ *
+ *		The file descriptor is read-only and has the close-on-exec flag
+ *		set. It is not seekable and **lseek**\ (2) fails with **ESPIPE**.
+ *		*flags* may only contain **BPF_F_STREAM_NONBLOCK**.
+ *
+ *	Return
+ *		A new file descriptor (a nonnegative integer), or -1 if an
+ *		error occurred (in which case, *errno* is set appropriately).
+ *
  * NOTES
  *	eBPF objects (maps and programs) can be shared between processes.
  *
@@ -993,6 +1020,7 @@ enum bpf_cmd {
 	BPF_TOKEN_CREATE,
 	BPF_PROG_STREAM_READ_BY_FD,
 	BPF_PROG_ASSOC_STRUCT_OPS,
+	BPF_PROG_STREAM_OPEN,
 	__MAX_BPF_CMD,
 	BPF_COMMON_ATTRS = 1 << 16, /* Indicate carrying syscall common attrs. */
 };
@@ -1524,6 +1552,11 @@ enum {
 	BPF_STREAM_STDERR = 2,
 };
 
+/* flags for BPF_PROG_STREAM_OPEN command */
+enum {
+	BPF_F_STREAM_NONBLOCK = (1U << 0),
+};
+
 union bpf_attr {
 	struct { /* anonymous struct used by BPF_MAP_CREATE command */
 		__u32	map_type;	/* one of enum bpf_map_type */
@@ -1950,6 +1983,12 @@ union bpf_attr {
 		__u32		flags;
 	} prog_assoc_struct_ops;
 
+	struct {
+		__u32		prog_fd;
+		__u32		stream_id;
+		__u32		flags;
+	} prog_stream_open;
+
 } __attribute__((aligned(8)));
 
 /* The description below is an attempt at providing documentation to eBPF
-- 
2.53.0

[PATCH bpf-next v2 3/5] libbpf: Add bpf_prog_stream_open()

[Kumar Kartikeya Dwivedi]

Expose the BPF_PROG_STREAM_OPEN command through a low-level libbpf wrapper.
The options structure carries stream open flags, including non-blocking
mode, while leaving room for future extensions.

Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 tools/lib/bpf/bpf.c      | 19 +++++++++++++++++++
 tools/lib/bpf/bpf.h      | 24 ++++++++++++++++++++++++
 tools/lib/bpf/libbpf.map |  1 +
 3 files changed, 44 insertions(+)

diff --git a/tools/lib/bpf/bpf.c b/tools/lib/bpf/bpf.c
index a9de7f107cf7..b49822d212ae 100644
--- a/tools/lib/bpf/bpf.c
+++ b/tools/lib/bpf/bpf.c
@@ -1466,6 +1466,25 @@ int bpf_prog_stream_read(int prog_fd, __u32 stream_id, void *buf, __u32 buf_len,
 	return libbpf_err_errno(err);
 }
 
+int bpf_prog_stream_open(int prog_fd, __u32 stream_id,
+			 const struct bpf_prog_stream_open_opts *opts)
+{
+	const size_t attr_sz = offsetofend(union bpf_attr, prog_stream_open);
+	union bpf_attr attr;
+	int fd;
+
+	if (!OPTS_VALID(opts, bpf_prog_stream_open_opts))
+		return libbpf_err(-EINVAL);
+
+	memset(&attr, 0, attr_sz);
+	attr.prog_stream_open.prog_fd = prog_fd;
+	attr.prog_stream_open.stream_id = stream_id;
+	attr.prog_stream_open.flags = OPTS_GET(opts, flags, 0);
+
+	fd = sys_bpf_fd(BPF_PROG_STREAM_OPEN, &attr, attr_sz);
+	return libbpf_err_errno(fd);
+}
+
 int bpf_prog_assoc_struct_ops(int prog_fd, int map_fd,
 			      struct bpf_prog_assoc_struct_ops_opts *opts)
 {
diff --git a/tools/lib/bpf/bpf.h b/tools/lib/bpf/bpf.h
index 490e8cb4ba53..826d9cc9ab65 100644
--- a/tools/lib/bpf/bpf.h
+++ b/tools/lib/bpf/bpf.h
@@ -759,10 +759,34 @@ struct bpf_prog_stream_read_opts {
  *
  * @return The number of bytes read, on success; negative error code, otherwise
  * (errno is also set to the error code)
+ *
+ * For blocking reads and polling, prefer **bpf_prog_stream_open**.
  */
 LIBBPF_API int bpf_prog_stream_read(int prog_fd, __u32 stream_id, void *buf, __u32 buf_len,
 				    struct bpf_prog_stream_read_opts *opts);
 
+struct bpf_prog_stream_open_opts {
+	size_t sz;
+	__u32 flags;
+	size_t :0;
+};
+#define bpf_prog_stream_open_opts__last_field flags
+
+/**
+ * @brief **bpf_prog_stream_open** opens a file descriptor for a BPF stream of
+ * a given BPF program.
+ *
+ * @param prog_fd FD for the BPF program whose BPF stream is to be opened.
+ * @param stream_id ID of the BPF stream to be opened.
+ * @param opts optional options, can be NULL. BPF_F_STREAM_NONBLOCK requests a
+ * non-blocking descriptor.
+ *
+ * @return A new stream FD, on success; negative error code, otherwise (errno
+ * is also set to the error code)
+ */
+LIBBPF_API int bpf_prog_stream_open(int prog_fd, __u32 stream_id,
+				    const struct bpf_prog_stream_open_opts *opts);
+
 struct bpf_prog_assoc_struct_ops_opts {
 	size_t sz;
 	__u32 flags;
diff --git a/tools/lib/bpf/libbpf.map b/tools/lib/bpf/libbpf.map
index 18d27d20102e..53b7b591325f 100644
--- a/tools/lib/bpf/libbpf.map
+++ b/tools/lib/bpf/libbpf.map
@@ -459,6 +459,7 @@ LIBBPF_1.7.0 {
 LIBBPF_1.8.0 {
 	global:
 		bpf_map__attach_cgroup_opts;
+		bpf_prog_stream_open;
 		bpf_program__add_flags;
 		bpf_program__attach_tracing_multi;
 		bpf_program__clear_flags;
-- 
2.53.0

[PATCH bpf-next v2 4/5] bpftool: Read program streams through file descriptors

[Kumar Kartikeya Dwivedi]

Use bpf_prog_stream_open() in its default blocking mode when following a
program stream. This lets bpftool wait for later output instead of exiting
as soon as it drains the currently buffered bytes. Flush each chunk as it
arrives so redirected output is not held in stdio buffers while the next
read blocks.

A stream descriptor does not keep its program alive, so bpftool drops the
program descriptor once the stream is open. The command then ends with EOF
when the program is unloaded and otherwise runs until interrupted, as
bpftool prog tracelog already does for the trace pipe. Install SIGINT,
SIGHUP and SIGTERM handlers without SA_RESTART so an interrupted read ends
the loop, check the stop flag before each read so a signal that lands while
output is being printed ends it as well, and exit successfully in both
cases. Restore the previous dispositions afterwards so batch mode keeps its
usual signal behavior. Report errors from the fallback read path too.

Fall back to BPF_PROG_STREAM_READ_BY_FD when the open command returns
EINVAL. The bpf() syscall uses EINVAL for unknown commands, so this keeps a
new bpftool compatible with kernels that provide program streams but
predate their file descriptor interface. On such kernels the command still
dumps the buffered output and exits. Document both behaviors.

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 .../bpftool/Documentation/bpftool-prog.rst    |  5 ++
 tools/bpf/bpftool/prog.c                      | 62 ++++++++++++++++---
 2 files changed, 60 insertions(+), 7 deletions(-)

diff --git a/tools/bpf/bpftool/Documentation/bpftool-prog.rst b/tools/bpf/bpftool/Documentation/bpftool-prog.rst
index 90fe8c61bf42..ef7e00eb5e47 100644
--- a/tools/bpf/bpftool/Documentation/bpftool-prog.rst
+++ b/tools/bpf/bpftool/Documentation/bpftool-prog.rst
@@ -186,6 +186,11 @@ bpftool prog tracelog { stdout | stderr } *PROG*
     error messages to the standard error stream. This facility should be used
     only for debugging purposes.
 
+    On kernels that support opening a stream as a file descriptor, bpftool
+    keeps printing new output as the program produces it, until the program is
+    unloaded or <Ctrl+C> is hit. Older kernels only allow dumping the output
+    buffered so far, after which bpftool exits.
+
 bpftool prog run *PROG* data_in *FILE* [data_out *FILE* [data_size_out *L*]] [ctx_in *FILE* [ctx_out *FILE* [ctx_size_out *M*]]] [repeat *N*]
     Run BPF program *PROG* in the kernel testing infrastructure for BPF,
     meaning that the program works on the data and context provided by the
diff --git a/tools/bpf/bpftool/prog.c b/tools/bpf/bpftool/prog.c
index 24e40dfab469..f0241dada548 100644
--- a/tools/bpf/bpftool/prog.c
+++ b/tools/bpf/bpftool/prog.c
@@ -1119,21 +1119,69 @@ enum prog_tracelog_mode {
 	TRACE_STDERR,
 };
 
+static volatile sig_atomic_t stream_stop;
+
+static void stop_stream(int signo)
+{
+	stream_stop = 1;
+}
+
+/* Consumes prog_fd. */
 static int
 prog_tracelog_stream(int prog_fd, enum prog_tracelog_mode mode)
 {
+	/* No SA_RESTART: an interrupted read() must return EINTR to end the loop. */
+	const struct sigaction act = { .sa_handler = stop_stream };
+	const int signals[] = { SIGHUP, SIGINT, SIGTERM };
+	struct sigaction old[ARRAY_SIZE(signals)];
 	FILE *file = mode == TRACE_STDOUT ? stdout : stderr;
 	int stream_id = mode == TRACE_STDOUT ? 1 : 2;
 	char buf[512];
-	int ret;
+	unsigned int i;
+	int fd, ret;
+
+	fd = bpf_prog_stream_open(prog_fd, stream_id, NULL);
+	if (fd == -EINVAL) {
+		/* Kernel predates BPF_PROG_STREAM_OPEN: dump buffered output and exit. */
+		do {
+			ret = bpf_prog_stream_read(prog_fd, stream_id, buf, sizeof(buf), NULL);
+			if (ret > 0)
+				fwrite(buf, sizeof(buf[0]), ret, file);
+		} while (ret > 0);
+		if (ret < 0)
+			p_err("failed to read stream: %s", strerror(-ret));
+		close(prog_fd);
+		goto out;
+	}
+	/*
+	 * The stream descriptor does not keep the program alive. Drop the
+	 * program reference so that reads return EOF once the program is gone.
+	 */
+	close(prog_fd);
+	if (fd < 0) {
+		p_err("failed to open stream: %s", strerror(-fd));
+		return -1;
+	}
 
+	stream_stop = 0;
+	for (i = 0; i < ARRAY_SIZE(signals); i++)
+		sigaction(signals[i], &act, &old[i]);
 	ret = 0;
-	do {
-		ret = bpf_prog_stream_read(prog_fd, stream_id, buf, sizeof(buf), NULL);
-		if (ret > 0)
-			fwrite(buf, sizeof(buf[0]), ret, file);
-	} while (ret > 0);
-
+	while (!stream_stop) {
+		ret = read(fd, buf, sizeof(buf));
+		if (ret <= 0)
+			break;
+		fwrite(buf, sizeof(buf[0]), ret, file);
+		fflush(file);
+	}
+	if (ret < 0 && !(stream_stop && errno == EINTR))
+		p_err("failed to read stream: %s", strerror(errno));
+	else
+		ret = 0;
+	for (i = 0; i < ARRAY_SIZE(signals); i++)
+		sigaction(signals[i], &old[i], NULL);
+	close(fd);
+out:
 	fflush(file);
 	return ret ? -1 : 0;
 }
-- 
2.53.0

[PATCH bpf-next v2 5/5] selftests/bpf: Test program stream file descriptors

[Kumar Kartikeya Dwivedi]

Exercise validation and normal file semantics of BPF_PROG_STREAM_OPEN.
Cover read-only and close-on-exec state, write and lseek rejection,
non-blocking reads, partial reads, poll readiness, and a blocking reader
woken by new stream data. Also check that an empty stream write produces
neither readable data nor poll readiness.

Register the descriptor with EPOLLET, consume the initial data event
without draining buffered output, then release the program. Require a
second edge carrying POLLIN and POLLHUP, drain the buffered bytes, and
verify that poll continues to report HUP before read returns EOF. This
ensures program teardown notifies edge-triggered loops even when
readability never transitions.

Also block a reader on an empty live stream, release the program, and
verify that it wakes and returns EOF. This directly covers the teardown
wait condition and its no-lost-wakeup ordering.

Finally, attach a stream-writing program to a hardware perf event and
busy-poll its epoll registration until output arrives. This both drives the
per-thread event into NMI context and directly verifies that the deferred
irq_work notification reaches epoll, without a separate BSS completion
handshake.

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
---
 .../testing/selftests/bpf/prog_tests/stream.c | 415 ++++++++++++++++++
 tools/testing/selftests/bpf/progs/stream.c    |  20 +
 2 files changed, 435 insertions(+)

diff --git a/tools/testing/selftests/bpf/prog_tests/stream.c b/tools/testing/selftests/bpf/prog_tests/stream.c
index 74bd15c4bfba..2c6b44155286 100644
--- a/tools/testing/selftests/bpf/prog_tests/stream.c
+++ b/tools/testing/selftests/bpf/prog_tests/stream.c
@@ -1,11 +1,17 @@
 // SPDX-License-Identifier: GPL-2.0
 /* Copyright (c) 2025 Meta Platforms, Inc. and affiliates. */
 #include <test_progs.h>
+#include <linux/perf_event.h>
+#include <poll.h>
+#include <sys/epoll.h>
 #include <sys/mman.h>
+#include <sys/syscall.h>
 
 #include "stream.skel.h"
 #include "stream_fail.skel.h"
 
+#define NMI_TIMEOUT_NS (5ULL * 1000 * 1000 * 1000)
+
 void test_stream_failure(void)
 {
 	RUN_TESTS(stream_fail);
@@ -62,6 +68,415 @@ void test_stream_syscall(void)
 	stream__destroy(skel);
 }
 
+static bool stream_fd_trigger(struct bpf_program *prog)
+{
+	LIBBPF_OPTS(bpf_test_run_opts, opts);
+	int ret;
+
+	ret = bpf_prog_test_run_opts(bpf_program__fd(prog), &opts);
+	return ASSERT_OK(ret, "test_run") && ASSERT_OK(opts.retval, "retval");
+}
+
+static void test_stream_fd_open(void)
+{
+	LIBBPF_OPTS(bpf_prog_stream_open_opts, opts);
+	struct stream *skel;
+	int fd, prog_fd;
+
+	skel = stream__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "stream__open_and_load"))
+		return;
+
+	prog_fd = bpf_program__fd(skel->progs.stream_syscall);
+	fd = bpf_prog_stream_open(0, BPF_STREAM_STDOUT, NULL);
+	ASSERT_EQ(fd, -EINVAL, "bad_prog_fd");
+
+	fd = bpf_prog_stream_open(prog_fd, 0, NULL);
+	ASSERT_EQ(fd, -ENOENT, "bad_stream_id");
+
+	opts.flags = BPF_F_RDONLY;
+	fd = bpf_prog_stream_open(prog_fd, BPF_STREAM_STDOUT, &opts);
+	ASSERT_EQ(fd, -EINVAL, "access_flag");
+
+	opts.flags = 1U << 31;
+	fd = bpf_prog_stream_open(prog_fd, BPF_STREAM_STDOUT, &opts);
+	ASSERT_EQ(fd, -EINVAL, "unknown_flag");
+
+	fd = bpf_prog_stream_open(prog_fd, BPF_STREAM_STDERR, NULL);
+	if (ASSERT_OK_FD(fd, "stderr"))
+		close(fd);
+
+	stream__destroy(skel);
+}
+
+static void test_stream_fd_nonblock(void)
+{
+	LIBBPF_OPTS(bpf_prog_stream_open_opts, opts,
+		.flags = BPF_F_STREAM_NONBLOCK,
+	);
+	struct pollfd pfd = { .events = POLLIN | POLLHUP };
+	struct stream *skel;
+	char buf[4] = {};
+	int fd, flags, ret;
+
+	skel = stream__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "stream__open_and_load"))
+		return;
+
+	fd = bpf_prog_stream_open(bpf_program__fd(skel->progs.stream_syscall),
+				  BPF_STREAM_STDOUT, &opts);
+	if (!ASSERT_OK_FD(fd, "stream_open"))
+		goto out_destroy;
+	pfd.fd = fd;
+
+	flags = fcntl(fd, F_GETFD);
+	ASSERT_GE(flags, 0, "getfd");
+	ASSERT_NEQ(flags & FD_CLOEXEC, 0, "cloexec");
+	flags = fcntl(fd, F_GETFL);
+	ASSERT_GE(flags, 0, "getfl");
+	ASSERT_EQ(flags & O_ACCMODE, O_RDONLY, "readonly");
+	ASSERT_NEQ(flags & O_NONBLOCK, 0, "nonblock");
+
+	ret = write(fd, "x", 1);
+	ASSERT_EQ(ret, -1, "write");
+	ASSERT_EQ(errno, EBADF, "write_errno");
+	ret = lseek(fd, 0, SEEK_SET);
+	ASSERT_EQ(ret, -1, "lseek");
+	ASSERT_EQ(errno, ESPIPE, "lseek_errno");
+
+	ret = poll(&pfd, 1, 0);
+	ASSERT_EQ(ret, 0, "poll_empty");
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, -1, "read_empty");
+	ASSERT_EQ(errno, EAGAIN, "read_empty_errno");
+
+	if (!stream_fd_trigger(skel->progs.stream_syscall))
+		goto out_close;
+	ret = poll(&pfd, 1, 0);
+	ASSERT_EQ(ret, 1, "poll_data");
+	ASSERT_NEQ(pfd.revents & POLLIN, 0, "pollin");
+	ASSERT_EQ(pfd.revents & POLLHUP, 0, "no_pollhup");
+
+	ret = read(fd, buf, 2);
+	ASSERT_EQ(ret, 2, "read_first");
+	ASSERT_OK(memcmp(buf, "fo", 2), "read_first_data");
+	pfd.revents = 0;
+	ret = poll(&pfd, 1, 0);
+	ASSERT_EQ(ret, 1, "poll_partial");
+	ASSERT_NEQ(pfd.revents & POLLIN, 0, "pollin_partial");
+
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, 1, "read_rest");
+	ASSERT_EQ(buf[0], 'o', "read_rest_data");
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, -1, "read_drained");
+	ASSERT_EQ(errno, EAGAIN, "read_drained_errno");
+
+out_close:
+	close(fd);
+out_destroy:
+	stream__destroy(skel);
+}
+
+static void test_stream_fd_empty(void)
+{
+	LIBBPF_OPTS(bpf_prog_stream_open_opts, opts,
+		.flags = BPF_F_STREAM_NONBLOCK,
+	);
+	struct pollfd pfd = { .events = POLLIN | POLLHUP };
+	struct stream *skel;
+	char buf[4];
+	int fd, ret;
+
+	skel = stream__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "stream__open_and_load"))
+		return;
+
+	fd = bpf_prog_stream_open(bpf_program__fd(skel->progs.stream_empty),
+				  BPF_STREAM_STDOUT, &opts);
+	if (!ASSERT_OK_FD(fd, "stream_open"))
+		goto out_destroy;
+	pfd.fd = fd;
+
+	/* Empty output produces neither data nor readiness. */
+	if (!stream_fd_trigger(skel->progs.stream_empty))
+		goto out_close;
+	ret = poll(&pfd, 1, 0);
+	ASSERT_EQ(ret, 0, "poll_empty_write");
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, -1, "read_empty_write");
+	ASSERT_EQ(errno, EAGAIN, "read_empty_write_errno");
+
+out_close:
+	close(fd);
+out_destroy:
+	stream__destroy(skel);
+}
+
+struct stream_fd_read_ctx {
+	int fd;
+	ssize_t ret;
+	char buf[4];
+};
+
+static void *stream_fd_read_thread(void *arg)
+{
+	struct stream_fd_read_ctx *ctx = arg;
+
+	ctx->ret = read(ctx->fd, ctx->buf, sizeof(ctx->buf));
+	return NULL;
+}
+
+static int stream_fd_timed_join(pthread_t thread)
+{
+	struct timespec timeout;
+
+	clock_gettime(CLOCK_REALTIME, &timeout);
+	timeout.tv_sec += 5;
+	return pthread_timedjoin_np(thread, NULL, &timeout);
+}
+
+static void test_stream_fd_blocking(void)
+{
+	struct stream_fd_read_ctx ctx = {};
+	struct stream *skel;
+	pthread_t thread;
+	bool thread_live = false;
+	int fd, flags, ret;
+
+	skel = stream__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "stream__open_and_load"))
+		return;
+
+	fd = bpf_prog_stream_open(bpf_program__fd(skel->progs.stream_syscall),
+				  BPF_STREAM_STDOUT, NULL);
+	if (!ASSERT_OK_FD(fd, "stream_open"))
+		goto out_destroy;
+	ctx.fd = fd;
+
+	flags = fcntl(fd, F_GETFL);
+	ASSERT_GE(flags, 0, "getfl");
+	ASSERT_EQ(flags & O_NONBLOCK, 0, "blocking");
+
+	ret = pthread_create(&thread, NULL, stream_fd_read_thread, &ctx);
+	if (!ASSERT_OK(ret, "pthread_create"))
+		goto out_close;
+	thread_live = true;
+
+	usleep(50000);
+	ret = pthread_tryjoin_np(thread, NULL);
+	if (!ASSERT_EQ(ret, EBUSY, "read_blocks")) {
+		thread_live = ret != 0;
+		goto out_thread;
+	}
+	if (!stream_fd_trigger(skel->progs.stream_syscall))
+		goto out_thread;
+
+	ret = stream_fd_timed_join(thread);
+	if (!ASSERT_OK(ret, "pthread_join"))
+		goto out_thread;
+	thread_live = false;
+	ASSERT_EQ(ctx.ret, 3, "read_len");
+	ASSERT_OK(memcmp(ctx.buf, "foo", 3), "read_data");
+
+	memset(&ctx, 0, sizeof(ctx));
+	ctx.fd = fd;
+	ret = pthread_create(&thread, NULL, stream_fd_read_thread, &ctx);
+	if (!ASSERT_OK(ret, "pthread_create_eof"))
+		goto out_close;
+	thread_live = true;
+
+	usleep(50000);
+	ret = pthread_tryjoin_np(thread, NULL);
+	if (!ASSERT_EQ(ret, EBUSY, "read_eof_blocks")) {
+		thread_live = ret != 0;
+		goto out_thread;
+	}
+
+	stream__destroy(skel);
+	skel = NULL;
+	ret = stream_fd_timed_join(thread);
+	if (!ASSERT_OK(ret, "pthread_join_eof"))
+		goto out_thread;
+	thread_live = false;
+	ASSERT_EQ(ctx.ret, 0, "read_eof");
+
+out_thread:
+	if (thread_live) {
+		stream__destroy(skel);
+		skel = NULL;
+		ret = stream_fd_timed_join(thread);
+		if (ret) {
+			pthread_cancel(thread);
+			pthread_join(thread, NULL);
+		}
+	}
+out_close:
+	close(fd);
+out_destroy:
+	stream__destroy(skel);
+}
+
+static void test_stream_fd_hup(void)
+{
+	LIBBPF_OPTS(bpf_prog_stream_open_opts, opts,
+		.flags = BPF_F_STREAM_NONBLOCK,
+	);
+	struct epoll_event event = {
+		.events = EPOLLIN | EPOLLET,
+	};
+	struct pollfd pfd = { .events = POLLIN | POLLHUP };
+	struct stream *skel;
+	char buf[4] = {};
+	int epfd = -1, fd, ret;
+
+	skel = stream__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "stream__open_and_load"))
+		return;
+
+	fd = bpf_prog_stream_open(bpf_program__fd(skel->progs.stream_syscall),
+				  BPF_STREAM_STDOUT, &opts);
+	if (!ASSERT_OK_FD(fd, "stream_open"))
+		goto out_destroy;
+	pfd.fd = fd;
+	epfd = epoll_create1(EPOLL_CLOEXEC);
+	if (!ASSERT_OK_FD(epfd, "epoll_create"))
+		goto out_close;
+	event.data.fd = fd;
+	ret = epoll_ctl(epfd, EPOLL_CTL_ADD, fd, &event);
+	if (!ASSERT_OK(ret, "epoll_ctl"))
+		goto out_close;
+
+	if (!stream_fd_trigger(skel->progs.stream_syscall))
+		goto out_close;
+	ret = epoll_wait(epfd, &event, 1, 5000);
+	if (!ASSERT_EQ(ret, 1, "epoll_wait_data"))
+		goto out_close;
+	ASSERT_NEQ(event.events & EPOLLIN, 0, "epollin");
+	ASSERT_EQ(event.events & EPOLLHUP, 0, "no_epollhup");
+
+	stream__destroy(skel);
+	skel = NULL;
+	event.events = 0;
+	ret = epoll_wait(epfd, &event, 1, 5000);
+	if (!ASSERT_EQ(ret, 1, "epoll_wait_hup"))
+		goto out_close;
+	ASSERT_NEQ(event.events & EPOLLIN, 0, "epollin_with_hup");
+	ASSERT_NEQ(event.events & EPOLLHUP, 0, "epollhup");
+
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, 3, "read_buffered");
+	ASSERT_OK(memcmp(buf, "foo", 3), "read_buffered_data");
+	pfd.revents = 0;
+	ret = poll(&pfd, 1, 0);
+	ASSERT_EQ(ret, 1, "poll_drained_hup");
+	ASSERT_EQ(pfd.revents & POLLIN, 0, "no_pollin_after_drain");
+	ASSERT_NEQ(pfd.revents & POLLHUP, 0, "pollhup_after_drain");
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, 0, "read_eof");
+
+out_close:
+	if (epfd >= 0)
+		close(epfd);
+	close(fd);
+out_destroy:
+	stream__destroy(skel);
+}
+
+static void test_stream_fd_nmi_epoll(void)
+{
+	LIBBPF_OPTS(bpf_prog_stream_open_opts, opts,
+		.flags = BPF_F_STREAM_NONBLOCK,
+	);
+	struct perf_event_attr attr = {
+		.size = sizeof(attr),
+		.type = PERF_TYPE_HARDWARE,
+		.config = PERF_COUNT_HW_CPU_CYCLES,
+		.freq = 1,
+		.sample_freq = 10,
+	};
+	struct epoll_event event = {
+		.events = EPOLLIN,
+	};
+	struct bpf_link *link = NULL;
+	struct stream *skel;
+	__u64 deadline;
+	char buf[4] = {};
+	int epfd = -1, fd = -1, pmu_fd = -1;
+	int ret;
+
+	skel = stream__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "stream__open_and_load"))
+		return;
+	fd = bpf_prog_stream_open(bpf_program__fd(skel->progs.stream_nmi),
+				  BPF_STREAM_STDOUT, &opts);
+	if (!ASSERT_OK_FD(fd, "stream_open"))
+		goto out;
+
+	epfd = epoll_create1(EPOLL_CLOEXEC);
+	if (!ASSERT_OK_FD(epfd, "epoll_create"))
+		goto out;
+	event.data.fd = fd;
+	ret = epoll_ctl(epfd, EPOLL_CTL_ADD, fd, &event);
+	if (!ASSERT_OK(ret, "epoll_ctl"))
+		goto out;
+
+	pmu_fd = syscall(__NR_perf_event_open, &attr, 0, -1, -1,
+			 PERF_FLAG_FD_CLOEXEC);
+	if (pmu_fd < 0 && (errno == ENOENT || errno == EOPNOTSUPP)) {
+		printf("%s:SKIP:no PERF_COUNT_HW_CPU_CYCLES\n", __func__);
+		test__skip();
+		goto out;
+	}
+	if (!ASSERT_GE(pmu_fd, 0, "perf_event_open"))
+		goto out;
+
+	link = bpf_program__attach_perf_event(skel->progs.stream_nmi, pmu_fd);
+	if (!ASSERT_OK_PTR(link, "attach_perf_event")) {
+		link = NULL;
+		goto out;
+	}
+	pmu_fd = -1;
+
+	deadline = get_time_ns() + NMI_TIMEOUT_NS;
+	do {
+		ret = epoll_wait(epfd, &event, 1, 0);
+	} while (ret == 0 && get_time_ns() < deadline);
+	if (!ASSERT_EQ(ret, 1, "epoll_wait"))
+		goto out;
+	ASSERT_NEQ(event.events & EPOLLIN, 0, "epollin");
+
+	ret = read(fd, buf, sizeof(buf));
+	ASSERT_EQ(ret, 3, "read_len");
+	ASSERT_OK(memcmp(buf, "nmi", 3), "read_data");
+
+out:
+	bpf_link__destroy(link);
+	if (pmu_fd >= 0)
+		close(pmu_fd);
+	if (epfd >= 0)
+		close(epfd);
+	if (fd >= 0)
+		close(fd);
+	stream__destroy(skel);
+}
+
+void test_stream_fd(void)
+{
+	if (test__start_subtest("open"))
+		test_stream_fd_open();
+	if (test__start_subtest("nonblock"))
+		test_stream_fd_nonblock();
+	if (test__start_subtest("empty"))
+		test_stream_fd_empty();
+	if (test__start_subtest("blocking"))
+		test_stream_fd_blocking();
+	if (test__start_subtest("hup"))
+		test_stream_fd_hup();
+	if (test__start_subtest("nmi_epoll"))
+		test_stream_fd_nmi_epoll();
+}
+
 void test_stream_oversize(void)
 {
 	LIBBPF_OPTS(bpf_test_run_opts, opts);
diff --git a/tools/testing/selftests/bpf/progs/stream.c b/tools/testing/selftests/bpf/progs/stream.c
index 12fc29e45487..1ae015c4e6b4 100644
--- a/tools/testing/selftests/bpf/progs/stream.c
+++ b/tools/testing/selftests/bpf/progs/stream.c
@@ -44,6 +44,7 @@ struct {
 		  _X64 _X64 _X64 _X64 _X64 _X64 _X64 _X64)
 
 int size;
+int nmi_stream_prints;
 u64 fault_addr;
 void *arena_ptr;
 
@@ -124,6 +125,25 @@ int stream_syscall(void *ctx)
 	return 0;
 }
 
+SEC("syscall")
+__success __retval(0)
+int stream_empty(void *ctx)
+{
+	return bpf_stream_printk(BPF_STDOUT, "");
+}
+
+SEC("perf_event")
+int stream_nmi(void *ctx)
+{
+	if (nmi_stream_prints)
+		return 0;
+	/* Retry on a later sample if the write failed, e.g. with -EBUSY. */
+	if (bpf_stream_printk(BPF_STDOUT, "nmi"))
+		return 0;
+	nmi_stream_prints = 1;
+	return 0;
+}
+
 SEC("syscall")
 __success __retval(0)
 int stream_oversize(void *ctx)
-- 
2.53.0

Re: [PATCH bpf-next v2 2/5] bpf: Add file descriptor interface for program streams

[Alexei Starovoitov]

On Thu, Sep 24, 2026 at 06:26 PM Kumar Kartikeya Dwivedi <memxor@gmail.com> wrote:
> +static void bpf_stream_publish(struct bpf_stream *stream, int len)
> +{
> +	/* Pairs with atomic_read_acquire() in bpf_stream_readable_bytes(). */
> +	(void)atomic_add_return_release(len, &stream->readable);
> +	bpf_stream_queue_notify(stream);
> +}

This queues irq_work on every bpf_stream_vprintk() and every stage
commit, whether anyone opened the stream or not.
That's an irq_work interrupt per print for every existing user.
ringbuf queues it only when the consumer caught up.

The result of atomic_add_return_release() is thrown away.
can we queue the work only when it returns len, i.e. when the stream
was empty ?

pw-bot: cr

Re: [PATCH bpf-next v2 4/5] bpftool: Read program streams through file descriptors

[Alexei Starovoitov]

On Thu, Sep 24, 2026 at 06:26 PM Kumar Kartikeya Dwivedi <memxor@gmail.com> wrote:
> Use bpf_prog_stream_open() in its default blocking mode when following a
> program stream. This lets bpftool wait for later output instead of exiting
> as soon as it drains the currently buffered bytes. Flush each chunk as it
> arrives so redirected output is not held in stdio buffers while the next
> read blocks.

So the same 'bpftool prog tracelog stderr PROG' prints what is buffered
and exits on old kernels and doesn't return on new ones while the prog
is loaded. Scripts that dump the stream will hang.
Probably should be opt-in?

[...]

> +	while (!stream_stop) {
> +		ret = read(fd, buf, sizeof(buf));

A signal that arrives after the stream_stop check and before read()
blocks only sets the flag. read() then sleeps until the next print or
the next signal.
exit_tracelog() calls exit(0) from the handler.
can we do the same here ?
Then stream_stop and save/restore of old[] are not needed.