From: Pavel Begunkov <asml.silence@gmail.com>
To: io-uring@vger.kernel.org
Cc: asml.silence@gmail.com, netdev@vger.kernel.org
Subject: [PATCH review-only 15/17] io_uring/zcrx: add dynamic area creation
Date: Sat, 11 Jul 2026 10:11:38 +0100 [thread overview]
Message-ID: <9dbb082763ba53ee6f2ceda51772ed9f6ca72383.1783616211.git.asml.silence@gmail.com> (raw)
In-Reply-To: <cover.1783616211.git.asml.silence@gmail.com>
It's not always possible for the user to predict during registration how
much memory zcrx will need to sustain the traffic. Allow to dynamically
add more areas with a new ctrl code ZCRX_CTRL_ADD_AREA.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
include/uapi/linux/io_uring/zcrx.h | 7 +++
io_uring/zcrx.c | 84 +++++++++++++++++++++++++-----
2 files changed, 79 insertions(+), 12 deletions(-)
diff --git a/include/uapi/linux/io_uring/zcrx.h b/include/uapi/linux/io_uring/zcrx.h
index 15c05c45ce36..08cdb173b04b 100644
--- a/include/uapi/linux/io_uring/zcrx.h
+++ b/include/uapi/linux/io_uring/zcrx.h
@@ -116,6 +116,7 @@ enum zcrx_ctrl_op {
ZCRX_CTRL_FLUSH_RQ,
ZCRX_CTRL_EXPORT,
ZCRX_CTRL_ARM_NOTIFICATION,
+ ZCRX_CTRL_ADD_AREA,
__ZCRX_CTRL_LAST,
};
@@ -134,6 +135,11 @@ struct zcrx_ctrl_arm_notif {
__u32 __resv[11];
};
+struct zcrx_ctrl_add_area {
+ __u64 area_ptr; /* pointer to struct io_uring_zcrx_area_reg */
+ __u64 __resv[5];
+};
+
struct zcrx_ctrl {
__u32 zcrx_id;
__u32 op; /* see enum zcrx_ctrl_op */
@@ -143,6 +149,7 @@ struct zcrx_ctrl {
struct zcrx_ctrl_export zc_export;
struct zcrx_ctrl_flush_rq zc_flush;
struct zcrx_ctrl_arm_notif zc_arm_notif;
+ struct zcrx_ctrl_add_area zc_area;
};
};
diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c
index 3f61f942c393..f7592a3c058d 100644
--- a/io_uring/zcrx.c
+++ b/io_uring/zcrx.c
@@ -36,6 +36,7 @@
#define ZCRX_REFILL_CAP MIN(64 * ZCRX_MAX_FRAGS_PER_PAGE, 1024)
#define IO_ZCRX_AREA_SUPPORTED_FLAGS (IORING_ZCRX_AREA_DMABUF)
+#define ZCRX_MAX_AREAS 1024
#define IO_DMA_ATTR (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING)
@@ -46,7 +47,7 @@ static inline u64 zcrx_area_id_to_token(u32 area_id)
static inline u32 zcrx_next_area_id(struct io_zcrx_ifq *zcrx)
{
- return zcrx->nr_areas;
+ return READ_ONCE(zcrx->nr_areas);
}
static inline struct io_zcrx_ifq *io_pp_to_ifq(struct page_pool *pp)
@@ -295,8 +296,6 @@ static int io_import_area(struct io_zcrx_ifq *ifq,
if (area_reg->flags & ~IO_ZCRX_AREA_SUPPORTED_FLAGS)
return -EINVAL;
- if (area_reg->rq_area_token)
- return -EINVAL;
if (area_reg->__resv2[0] || area_reg->__resv2[1])
return -EINVAL;
@@ -311,15 +310,11 @@ static int io_import_area(struct io_zcrx_ifq *ifq,
return io_import_umem(ifq, mem, area_reg);
}
-static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
- struct io_zcrx_area *area)
+static void __io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
+ struct io_zcrx_area *area)
{
int i;
- if (!area)
- return;
-
- guard(mutex)(&ifq->pp_lock);
if (!area->is_mapped)
return;
area->is_mapped = false;
@@ -337,6 +332,15 @@ static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
}
}
+static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
+ struct io_zcrx_area *area)
+{
+ if (!area)
+ return;
+ guard(mutex)(&ifq->pp_lock);
+ __io_zcrx_unmap_area(ifq, area);
+}
+
static void io_zcrx_unmap_areas(struct io_zcrx_ifq *ifq)
{
unsigned area_idx;
@@ -475,7 +479,9 @@ static int io_zcrx_append_area(struct io_zcrx_ifq *ifq,
struct io_zcrx_area **areas, **old_areas;
unsigned old_nr;
- if (WARN_ON_ONCE(ifq->kern_readable != kern_readable))
+ if (ifq->kern_readable != kern_readable)
+ return -EINVAL;
+ if (ifq->nr_areas + 1 > ZCRX_MAX_AREAS)
return -EINVAL;
if (WARN_ON_ONCE(area->area_id != zcrx_next_area_id(ifq)))
return -EINVAL;
@@ -516,7 +522,7 @@ static int __zcrx_create_area(struct io_zcrx_ifq *ifq,
return -EINVAL;
buf_size_shift = ilog2(rx_buf_len);
}
- if (WARN_ON_ONCE(ifq->niov_shift))
+ if (ifq->niov_shift && ifq->niov_shift != buf_size_shift)
return -EINVAL;
if (!ifq->dev && buf_size_shift != PAGE_SHIFT)
return -EOPNOTSUPP;
@@ -578,7 +584,7 @@ static int __zcrx_create_area(struct io_zcrx_ifq *ifq,
return 0;
err:
if (area) {
- io_zcrx_unmap_area(ifq, area);
+ __io_zcrx_unmap_area(ifq, area);
io_zcrx_free_area(ifq, area);
}
return ret;
@@ -1012,6 +1018,8 @@ int io_register_zcrx(struct io_ring_ctx *ctx,
if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area)))
return -EFAULT;
+ if (area.rq_area_token)
+ return -EINVAL;
memset(¬if, 0, sizeof(notif));
if (reg.notif_desc && copy_from_user(¬if, u64_to_user_ptr(reg.notif_desc),
@@ -1074,6 +1082,8 @@ int io_register_zcrx(struct io_ring_ctx *ctx,
goto err;
}
+ WARN_ON_ONCE(!ifq->niov_shift);
+
reg.zcrx_id = id;
scoped_guard(mutex, &ctx->mmap_lock) {
@@ -1559,6 +1569,54 @@ static int zcrx_arm_notif(struct io_ring_ctx *ctx, struct io_zcrx_ifq *zcrx,
return 0;
}
+static int zcrx_ctrl_add_area(struct io_ring_ctx *ctx, struct io_zcrx_ifq *ifq,
+ struct zcrx_ctrl *ctrl)
+{
+ struct zcrx_ctrl_add_area *ctrl_add = &ctrl->zc_area;
+ struct io_uring_zcrx_area_reg __user *area_uptr;
+ struct io_uring_zcrx_area_reg area_reg;
+ struct io_zcrx_area *area = NULL;
+ int ret;
+
+ area_uptr = u64_to_user_ptr(ctrl_add->area_ptr);
+ if (copy_from_user(&area_reg, area_uptr, sizeof(area_reg)))
+ return -EFAULT;
+ if (!mem_is_zero(&ctrl_add->__resv, sizeof(ctrl_add->__resv)))
+ return -EINVAL;
+ if (area_reg.rq_area_token)
+ return -EINVAL;
+
+ while (true) {
+ u32 area_id = zcrx_next_area_id(ifq);
+
+ /*
+ * It's hard to roll back append and page faults under
+ * ->pp_lock is a bad idea. Grab and post an unstable area id
+ * first, and then check-retry under the lock.
+ */
+ area_reg.rq_area_token = zcrx_area_id_to_token(area_id);
+ if (copy_to_user(area_uptr, &area_reg, sizeof(area_reg)))
+ return -EFAULT;
+
+ guard(mutex)(&ifq->pp_lock);
+ if (area_id != zcrx_next_area_id(ifq))
+ continue;
+
+ ret = __zcrx_create_area(ifq, &area_reg, &area,
+ 1U << ifq->niov_shift, area_id);
+ if (ret)
+ break;
+
+ ret = io_zcrx_append_area(ifq, area);
+ if (ret)
+ __io_zcrx_unmap_area(ifq, area);
+ break;
+ }
+ if (ret && area)
+ io_zcrx_free_area(ifq, area);
+ return ret;
+}
+
int io_zcrx_ctrl(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
{
struct zcrx_ctrl ctrl;
@@ -1585,6 +1643,8 @@ int io_zcrx_ctrl(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
return zcrx_export(ctx, zcrx, &ctrl, arg);
case ZCRX_CTRL_ARM_NOTIFICATION:
return zcrx_arm_notif(ctx, zcrx, &ctrl);
+ case ZCRX_CTRL_ADD_AREA:
+ return zcrx_ctrl_add_area(ctx, zcrx, &ctrl);
}
return -EOPNOTSUPP;
--
2.54.0
WARNING: multiple messages have this Message-ID (diff)
From: Pavel Begunkov <asml.silence@gmail.com>
To: io-uring@vger.kernel.org
Cc: asml.silence@gmail.com, netdev@vger.kernel.org
Subject: [PATCH review-only 15/17] io_uring/zcrx: add dynamic area creation
Date: Sat, 11 Jul 2026 11:40:08 +0100 [thread overview]
Message-ID: <9dbb082763ba53ee6f2ceda51772ed9f6ca72383.1783616211.git.asml.silence@gmail.com> (raw)
Message-ID: <20260711104008.mnksQKrBkas1itNkDdYsDSVqFNxWdZgLF360zKLpEes@z> (raw)
In-Reply-To: <cover.1783616211.git.asml.silence@gmail.com>
It's not always possible for the user to predict during registration how
much memory zcrx will need to sustain the traffic. Allow to dynamically
add more areas with a new ctrl code ZCRX_CTRL_ADD_AREA.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
include/uapi/linux/io_uring/zcrx.h | 7 +++
io_uring/zcrx.c | 84 +++++++++++++++++++++++++-----
2 files changed, 79 insertions(+), 12 deletions(-)
diff --git a/include/uapi/linux/io_uring/zcrx.h b/include/uapi/linux/io_uring/zcrx.h
index 15c05c45ce36..08cdb173b04b 100644
--- a/include/uapi/linux/io_uring/zcrx.h
+++ b/include/uapi/linux/io_uring/zcrx.h
@@ -116,6 +116,7 @@ enum zcrx_ctrl_op {
ZCRX_CTRL_FLUSH_RQ,
ZCRX_CTRL_EXPORT,
ZCRX_CTRL_ARM_NOTIFICATION,
+ ZCRX_CTRL_ADD_AREA,
__ZCRX_CTRL_LAST,
};
@@ -134,6 +135,11 @@ struct zcrx_ctrl_arm_notif {
__u32 __resv[11];
};
+struct zcrx_ctrl_add_area {
+ __u64 area_ptr; /* pointer to struct io_uring_zcrx_area_reg */
+ __u64 __resv[5];
+};
+
struct zcrx_ctrl {
__u32 zcrx_id;
__u32 op; /* see enum zcrx_ctrl_op */
@@ -143,6 +149,7 @@ struct zcrx_ctrl {
struct zcrx_ctrl_export zc_export;
struct zcrx_ctrl_flush_rq zc_flush;
struct zcrx_ctrl_arm_notif zc_arm_notif;
+ struct zcrx_ctrl_add_area zc_area;
};
};
diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c
index 3f61f942c393..f7592a3c058d 100644
--- a/io_uring/zcrx.c
+++ b/io_uring/zcrx.c
@@ -36,6 +36,7 @@
#define ZCRX_REFILL_CAP MIN(64 * ZCRX_MAX_FRAGS_PER_PAGE, 1024)
#define IO_ZCRX_AREA_SUPPORTED_FLAGS (IORING_ZCRX_AREA_DMABUF)
+#define ZCRX_MAX_AREAS 1024
#define IO_DMA_ATTR (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING)
@@ -46,7 +47,7 @@ static inline u64 zcrx_area_id_to_token(u32 area_id)
static inline u32 zcrx_next_area_id(struct io_zcrx_ifq *zcrx)
{
- return zcrx->nr_areas;
+ return READ_ONCE(zcrx->nr_areas);
}
static inline struct io_zcrx_ifq *io_pp_to_ifq(struct page_pool *pp)
@@ -295,8 +296,6 @@ static int io_import_area(struct io_zcrx_ifq *ifq,
if (area_reg->flags & ~IO_ZCRX_AREA_SUPPORTED_FLAGS)
return -EINVAL;
- if (area_reg->rq_area_token)
- return -EINVAL;
if (area_reg->__resv2[0] || area_reg->__resv2[1])
return -EINVAL;
@@ -311,15 +310,11 @@ static int io_import_area(struct io_zcrx_ifq *ifq,
return io_import_umem(ifq, mem, area_reg);
}
-static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
- struct io_zcrx_area *area)
+static void __io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
+ struct io_zcrx_area *area)
{
int i;
- if (!area)
- return;
-
- guard(mutex)(&ifq->pp_lock);
if (!area->is_mapped)
return;
area->is_mapped = false;
@@ -337,6 +332,15 @@ static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
}
}
+static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq,
+ struct io_zcrx_area *area)
+{
+ if (!area)
+ return;
+ guard(mutex)(&ifq->pp_lock);
+ __io_zcrx_unmap_area(ifq, area);
+}
+
static void io_zcrx_unmap_areas(struct io_zcrx_ifq *ifq)
{
unsigned area_idx;
@@ -475,7 +479,9 @@ static int io_zcrx_append_area(struct io_zcrx_ifq *ifq,
struct io_zcrx_area **areas, **old_areas;
unsigned old_nr;
- if (WARN_ON_ONCE(ifq->kern_readable != kern_readable))
+ if (ifq->kern_readable != kern_readable)
+ return -EINVAL;
+ if (ifq->nr_areas + 1 > ZCRX_MAX_AREAS)
return -EINVAL;
if (WARN_ON_ONCE(area->area_id != zcrx_next_area_id(ifq)))
return -EINVAL;
@@ -516,7 +522,7 @@ static int __zcrx_create_area(struct io_zcrx_ifq *ifq,
return -EINVAL;
buf_size_shift = ilog2(rx_buf_len);
}
- if (WARN_ON_ONCE(ifq->niov_shift))
+ if (ifq->niov_shift && ifq->niov_shift != buf_size_shift)
return -EINVAL;
if (!ifq->dev && buf_size_shift != PAGE_SHIFT)
return -EOPNOTSUPP;
@@ -578,7 +584,7 @@ static int __zcrx_create_area(struct io_zcrx_ifq *ifq,
return 0;
err:
if (area) {
- io_zcrx_unmap_area(ifq, area);
+ __io_zcrx_unmap_area(ifq, area);
io_zcrx_free_area(ifq, area);
}
return ret;
@@ -1012,6 +1018,8 @@ int io_register_zcrx(struct io_ring_ctx *ctx,
if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area)))
return -EFAULT;
+ if (area.rq_area_token)
+ return -EINVAL;
memset(¬if, 0, sizeof(notif));
if (reg.notif_desc && copy_from_user(¬if, u64_to_user_ptr(reg.notif_desc),
@@ -1074,6 +1082,8 @@ int io_register_zcrx(struct io_ring_ctx *ctx,
goto err;
}
+ WARN_ON_ONCE(!ifq->niov_shift);
+
reg.zcrx_id = id;
scoped_guard(mutex, &ctx->mmap_lock) {
@@ -1559,6 +1569,54 @@ static int zcrx_arm_notif(struct io_ring_ctx *ctx, struct io_zcrx_ifq *zcrx,
return 0;
}
+static int zcrx_ctrl_add_area(struct io_ring_ctx *ctx, struct io_zcrx_ifq *ifq,
+ struct zcrx_ctrl *ctrl)
+{
+ struct zcrx_ctrl_add_area *ctrl_add = &ctrl->zc_area;
+ struct io_uring_zcrx_area_reg __user *area_uptr;
+ struct io_uring_zcrx_area_reg area_reg;
+ struct io_zcrx_area *area = NULL;
+ int ret;
+
+ area_uptr = u64_to_user_ptr(ctrl_add->area_ptr);
+ if (copy_from_user(&area_reg, area_uptr, sizeof(area_reg)))
+ return -EFAULT;
+ if (!mem_is_zero(&ctrl_add->__resv, sizeof(ctrl_add->__resv)))
+ return -EINVAL;
+ if (area_reg.rq_area_token)
+ return -EINVAL;
+
+ while (true) {
+ u32 area_id = zcrx_next_area_id(ifq);
+
+ /*
+ * It's hard to roll back append and page faults under
+ * ->pp_lock is a bad idea. Grab and post an unstable area id
+ * first, and then check-retry under the lock.
+ */
+ area_reg.rq_area_token = zcrx_area_id_to_token(area_id);
+ if (copy_to_user(area_uptr, &area_reg, sizeof(area_reg)))
+ return -EFAULT;
+
+ guard(mutex)(&ifq->pp_lock);
+ if (area_id != zcrx_next_area_id(ifq))
+ continue;
+
+ ret = __zcrx_create_area(ifq, &area_reg, &area,
+ 1U << ifq->niov_shift, area_id);
+ if (ret)
+ break;
+
+ ret = io_zcrx_append_area(ifq, area);
+ if (ret)
+ __io_zcrx_unmap_area(ifq, area);
+ break;
+ }
+ if (ret && area)
+ io_zcrx_free_area(ifq, area);
+ return ret;
+}
+
int io_zcrx_ctrl(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
{
struct zcrx_ctrl ctrl;
@@ -1585,6 +1643,8 @@ int io_zcrx_ctrl(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
return zcrx_export(ctx, zcrx, &ctrl, arg);
case ZCRX_CTRL_ARM_NOTIFICATION:
return zcrx_arm_notif(ctx, zcrx, &ctrl);
+ case ZCRX_CTRL_ADD_AREA:
+ return zcrx_ctrl_add_area(ctx, zcrx, &ctrl);
}
return -EOPNOTSUPP;
--
2.54.0
next prev parent reply other threads:[~2026-07-11 9:12 UTC|newest]
Thread overview: 36+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-07-11 9:11 [PATCH review-only 00/17] zcrx RQ improvements and dynamic memory provisioning Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 01/17] io_uring/zcrx: scale refilling with large pages Pavel Begunkov
2026-07-11 10:39 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 02/17] io_uring/zcrx: move RQ head/tail to separate cache lines Pavel Begunkov
2026-07-11 10:39 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 03/17] io_uring/zcrx: add RQ iterator Pavel Begunkov
2026-07-11 10:39 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 04/17] io_uring/zcrx: cache RQ tail Pavel Begunkov
2026-07-11 10:39 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 05/17] io_uring/zcrx: coalesce same-niov RQEs on refill Pavel Begunkov
2026-07-11 10:39 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 06/17] io_uring/zcrx: constify area_reg on import Pavel Begunkov
2026-07-11 10:39 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 07/17] io_uring/zcrx: add helper for deriving area token Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 08/17] io_uring/zcrx: don't pass ifq_reg to area creation Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 09/17] io_uring/zcrx: split dmabuf unmap and release Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 10/17] io_uring/zcrx: unmap under netdev lock Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 11/17] io_uring/zcrx: split append out of area creation Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 12/17] io_uring/zcrx: move freelist lock to struct zcrx Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 13/17] io_uring/zcrx: array of areas Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 14/17] io_uring/zcrx: pass area_id to __zcrx_create_area() Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 9:11 ` Pavel Begunkov [this message]
2026-07-11 10:40 ` [PATCH review-only 15/17] io_uring/zcrx: add dynamic area creation Pavel Begunkov
2026-07-11 9:11 ` [PATCH review-only 16/17] io_urint/zcrx: narrow var scope in io_zcrx_recv_skb() Pavel Begunkov
2026-07-11 10:40 ` Pavel Begunkov
2026-07-11 10:39 ` [PATCH review-only 00/17] zcrx RQ improvements and dynamic memory provisioning Pavel Begunkov
2026-07-11 10:39 ` [PATCH RESEND " Pavel Begunkov
2026-07-11 10:40 ` [PATCH review-only 17/17] io_uring/zcrx: don't reload skb_shinfo Pavel Begunkov
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=9dbb082763ba53ee6f2ceda51772ed9f6ca72383.1783616211.git.asml.silence@gmail.com \
--to=asml.silence@gmail.com \
--cc=io-uring@vger.kernel.org \
--cc=netdev@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox