[PATCH for-next v8 0/5] fixed-buffer for uring-cmd/passthru

[PATCH for-next v8 0/5] fixed-buffer for uring-cmd/passthru

[Kanchan Joshi]

Currently uring-cmd lacks the ability to leverage the pre-registered
buffers. This series adds that support in uring-cmd, and plumbs
nvme passthrough to work with it.

Using registered-buffers showed IOPS hike from 1.9M to 2.2M in my tests.

Patch 1, 3, 4 = prep
Patch 2 = expand io_uring command to use registered-buffers
Patch 5 = expand nvme passthrough to use registered-buffers

Changes since v7:
- Patch 3: added many cleanups/refactoring suggested by Christoph
- Patch 4: added copying-pages fallback for bounce-buffer/dma-alignment case
  (Christoph)

Changes since v6:
- Patch 1: fix warning for io_uring_cmd_import_fixed (robot)
-
Changes since v5:
- Patch 4: newly addd, to split a nvme function into two
- Patch 3: folded cleanups in bio_map_user_iov (Chaitanya, Pankaj)
- Rebase to latest for-next

Changes since v4:
- Patch 1, 2: folded all review comments of Jens

Changes since v3:
- uring_cmd_flags, change from u16 to u32 (Jens)
- patch 3, add another helper to reduce code-duplication (Jens)

Changes since v2:
- Kill the new opcode, add a flag instead (Pavel)
- Fix standalone build issue with patch 1 (Pavel)

Changes since v1:
- Fix a naming issue for an exported helper

Anuj Gupta (2):
  io_uring: add io_uring_cmd_import_fixed
  io_uring: introduce fixed buffer support for io_uring_cmd

Kanchan Joshi (3):
  nvme: refactor nvme_alloc_user_request
  block: add helper to map bvec iterator for passthrough
  nvme: wire up fixed buffer support for nvme passthrough

 block/blk-map.c               | 111 +++++++++++++++++++++++---
 drivers/nvme/host/ioctl.c     | 141 ++++++++++++++++++++--------------
 include/linux/blk-mq.h        |   1 +
 include/linux/io_uring.h      |  10 ++-
 include/uapi/linux/io_uring.h |   9 +++
 io_uring/uring_cmd.c          |  26 ++++++-
 6 files changed, 230 insertions(+), 68 deletions(-)

-- 
2.25.1

[PATCH for-next v8 1/5] io_uring: add io_uring_cmd_import_fixed

[Kanchan Joshi]

From: Anuj Gupta <anuj20.g@samsung.com>

This is a new helper that callers can use to obtain a bvec iterator for
the previously mapped buffer. This is preparatory work to enable
fixed-buffer support for io_uring_cmd.

Signed-off-by: Anuj Gupta <anuj20.g@samsung.com>
Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
---
 include/linux/io_uring.h |  8 ++++++++
 io_uring/uring_cmd.c     | 10 ++++++++++
 2 files changed, 18 insertions(+)

diff --git a/include/linux/io_uring.h b/include/linux/io_uring.h
index 58676c0a398f..1dbf51115c30 100644
--- a/include/linux/io_uring.h
+++ b/include/linux/io_uring.h
@@ -4,6 +4,7 @@
 
 #include <linux/sched.h>
 #include <linux/xarray.h>
+#include <uapi/linux/io_uring.h>
 
 enum io_uring_cmd_flags {
 	IO_URING_F_COMPLETE_DEFER	= 1,
@@ -32,6 +33,8 @@ struct io_uring_cmd {
 };
 
 #if defined(CONFIG_IO_URING)
+int io_uring_cmd_import_fixed(u64 ubuf, unsigned long len, int rw,
+			      struct iov_iter *iter, void *ioucmd);
 void io_uring_cmd_done(struct io_uring_cmd *cmd, ssize_t ret, ssize_t res2);
 void io_uring_cmd_complete_in_task(struct io_uring_cmd *ioucmd,
 			void (*task_work_cb)(struct io_uring_cmd *));
@@ -59,6 +62,11 @@ static inline void io_uring_free(struct task_struct *tsk)
 		__io_uring_free(tsk);
 }
 #else
+static int io_uring_cmd_import_fixed(u64 ubuf, unsigned long len, int rw,
+			      struct iov_iter *iter, void *ioucmd)
+{
+	return -EOPNOTSUPP;
+}
 static inline void io_uring_cmd_done(struct io_uring_cmd *cmd, ssize_t ret,
 		ssize_t ret2)
 {
diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
index f3ed61e9bd0f..6a6d69523d75 100644
--- a/io_uring/uring_cmd.c
+++ b/io_uring/uring_cmd.c
@@ -8,6 +8,7 @@
 #include <uapi/linux/io_uring.h>
 
 #include "io_uring.h"
+#include "rsrc.h"
 #include "uring_cmd.h"
 
 static void io_uring_cmd_work(struct io_kiocb *req, bool *locked)
@@ -129,3 +130,12 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
 
 	return IOU_ISSUE_SKIP_COMPLETE;
 }
+
+int io_uring_cmd_import_fixed(u64 ubuf, unsigned long len, int rw,
+			      struct iov_iter *iter, void *ioucmd)
+{
+	struct io_kiocb *req = cmd_to_io_kiocb(ioucmd);
+
+	return io_import_fixed(rw, iter, req->imu, ubuf, len);
+}
+EXPORT_SYMBOL_GPL(io_uring_cmd_import_fixed);
-- 
2.25.1

[PATCH for-next v8 2/5] io_uring: introduce fixed buffer support for io_uring_cmd

[Kanchan Joshi]

From: Anuj Gupta <anuj20.g@samsung.com>

Add IORING_URING_CMD_FIXED flag that is to be used for sending io_uring
command with previously registered buffers. User-space passes the buffer
index in sqe->buf_index, same as done in read/write variants that uses
fixed buffers.

Signed-off-by: Anuj Gupta <anuj20.g@samsung.com>
Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
---
 include/linux/io_uring.h      |  2 +-
 include/uapi/linux/io_uring.h |  9 +++++++++
 io_uring/uring_cmd.c          | 16 +++++++++++++++-
 3 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/include/linux/io_uring.h b/include/linux/io_uring.h
index 1dbf51115c30..e10c5cc81082 100644
--- a/include/linux/io_uring.h
+++ b/include/linux/io_uring.h
@@ -28,7 +28,7 @@ struct io_uring_cmd {
 		void *cookie;
 	};
 	u32		cmd_op;
-	u32		pad;
+	u32		flags;
 	u8		pdu[32]; /* available inline for free use */
 };
 
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 92f29d9505a6..ab7458033ee3 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -56,6 +56,7 @@ struct io_uring_sqe {
 		__u32		hardlink_flags;
 		__u32		xattr_flags;
 		__u32		msg_ring_flags;
+		__u32		uring_cmd_flags;
 	};
 	__u64	user_data;	/* data to be passed back at completion time */
 	/* pack this to avoid bogus arm OABI complaints */
@@ -219,6 +220,14 @@ enum io_uring_op {
 	IORING_OP_LAST,
 };
 
+/*
+ * sqe->uring_cmd_flags
+ * IORING_URING_CMD_FIXED	use registered buffer; pass thig flag
+ *				along with setting sqe->buf_index.
+ */
+#define IORING_URING_CMD_FIXED	(1U << 0)
+
+
 /*
  * sqe->fsync_flags
  */
diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
index 6a6d69523d75..faefa9f6f259 100644
--- a/io_uring/uring_cmd.c
+++ b/io_uring/uring_cmd.c
@@ -4,6 +4,7 @@
 #include <linux/file.h>
 #include <linux/io_uring.h>
 #include <linux/security.h>
+#include <linux/nospec.h>
 
 #include <uapi/linux/io_uring.h>
 
@@ -77,8 +78,21 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 {
 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
 
-	if (sqe->rw_flags || sqe->__pad1)
+	if (sqe->__pad1)
 		return -EINVAL;
+
+	ioucmd->flags = READ_ONCE(sqe->uring_cmd_flags);
+	if (ioucmd->flags & IORING_URING_CMD_FIXED) {
+		struct io_ring_ctx *ctx = req->ctx;
+		u16 index;
+
+		req->buf_index = READ_ONCE(sqe->buf_index);
+		if (unlikely(req->buf_index >= ctx->nr_user_bufs))
+			return -EFAULT;
+		index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
+		req->imu = ctx->user_bufs[index];
+		io_req_set_rsrc_node(req, ctx, 0);
+	}
 	ioucmd->cmd = sqe->cmd;
 	ioucmd->cmd_op = READ_ONCE(sqe->cmd_op);
 	return 0;
-- 
2.25.1

[PATCH for-next v8 3/5] nvme: refactor nvme_alloc_user_request

[Kanchan Joshi]

Separate this out to two functions with reduced number of arguments.
While at it, do bit of refactoring in nvme_add_user_metadata too.
_
Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
---
 drivers/nvme/host/ioctl.c | 132 +++++++++++++++++++++-----------------
 1 file changed, 74 insertions(+), 58 deletions(-)

diff --git a/drivers/nvme/host/ioctl.c b/drivers/nvme/host/ioctl.c
index 548aca8b5b9f..9537991deac9 100644
--- a/drivers/nvme/host/ioctl.c
+++ b/drivers/nvme/host/ioctl.c
@@ -20,19 +20,20 @@ static void __user *nvme_to_user_ptr(uintptr_t ptrval)
 	return (void __user *)ptrval;
 }
 
-static void *nvme_add_user_metadata(struct bio *bio, void __user *ubuf,
-		unsigned len, u32 seed, bool write)
+static void *nvme_add_user_metadata(struct request *req, void __user *ubuf,
+		unsigned len, u32 seed)
 {
 	struct bio_integrity_payload *bip;
 	int ret = -ENOMEM;
 	void *buf;
+	struct bio *bio = req->bio;
 
 	buf = kmalloc(len, GFP_KERNEL);
 	if (!buf)
 		goto out;
 
 	ret = -EFAULT;
-	if (write && copy_from_user(buf, ubuf, len))
+	if ((req_op(req) == REQ_OP_DRV_OUT) && copy_from_user(buf, ubuf, len))
 		goto out_free_meta;
 
 	bip = bio_integrity_alloc(bio, GFP_KERNEL, 1);
@@ -45,8 +46,10 @@ static void *nvme_add_user_metadata(struct bio *bio, void __user *ubuf,
 	bip->bip_iter.bi_sector = seed;
 	ret = bio_integrity_add_page(bio, virt_to_page(buf), len,
 			offset_in_page(buf));
-	if (ret == len)
+	if (ret == len) {
+		req->cmd_flags |= REQ_INTEGRITY;
 		return buf;
+	}
 	ret = -ENOMEM;
 out_free_meta:
 	kfree(buf);
@@ -65,70 +68,67 @@ static int nvme_finish_user_metadata(struct request *req, void __user *ubuf,
 }
 
 static struct request *nvme_alloc_user_request(struct request_queue *q,
-		struct nvme_command *cmd, void __user *ubuffer,
-		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
-		u32 meta_seed, void **metap, unsigned timeout, bool vec,
-		blk_opf_t rq_flags, blk_mq_req_flags_t blk_flags)
+		struct nvme_command *cmd, blk_opf_t rq_flags,
+		blk_mq_req_flags_t blk_flags)
 {
-	bool write = nvme_is_write(cmd);
-	struct nvme_ns *ns = q->queuedata;
-	struct block_device *bdev = ns ? ns->disk->part0 : NULL;
 	struct request *req;
-	struct bio *bio = NULL;
-	void *meta = NULL;
-	int ret;
 
 	req = blk_mq_alloc_request(q, nvme_req_op(cmd) | rq_flags, blk_flags);
 	if (IS_ERR(req))
 		return req;
 	nvme_init_request(req, cmd);
-
-	if (timeout)
-		req->timeout = timeout;
 	nvme_req(req)->flags |= NVME_REQ_USERCMD;
+	return req;
+}
 
-	if (ubuffer && bufflen) {
-		if (!vec)
-			ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen,
-				GFP_KERNEL);
-		else {
-			struct iovec fast_iov[UIO_FASTIOV];
-			struct iovec *iov = fast_iov;
-			struct iov_iter iter;
-
-			ret = import_iovec(rq_data_dir(req), ubuffer, bufflen,
-					UIO_FASTIOV, &iov, &iter);
-			if (ret < 0)
-				goto out;
-			ret = blk_rq_map_user_iov(q, req, NULL, &iter,
-					GFP_KERNEL);
-			kfree(iov);
-		}
-		if (ret)
+static int nvme_map_user_request(struct request *req, void __user *ubuffer,
+		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
+		u32 meta_seed, void **metap, bool vec)
+{
+	struct request_queue *q = req->q;
+	struct nvme_ns *ns = q->queuedata;
+	struct block_device *bdev = ns ? ns->disk->part0 : NULL;
+	struct bio *bio = NULL;
+	void *meta = NULL;
+	int ret;
+
+	if (!vec)
+		ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen,
+			GFP_KERNEL);
+	else {
+		struct iovec fast_iov[UIO_FASTIOV];
+		struct iovec *iov = fast_iov;
+		struct iov_iter iter;
+
+		ret = import_iovec(rq_data_dir(req), ubuffer, bufflen,
+				UIO_FASTIOV, &iov, &iter);
+		if (ret < 0)
 			goto out;
-		bio = req->bio;
-		if (bdev)
-			bio_set_dev(bio, bdev);
-		if (bdev && meta_buffer && meta_len) {
-			meta = nvme_add_user_metadata(bio, meta_buffer, meta_len,
-					meta_seed, write);
-			if (IS_ERR(meta)) {
-				ret = PTR_ERR(meta);
-				goto out_unmap;
-			}
-			req->cmd_flags |= REQ_INTEGRITY;
-			*metap = meta;
+		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
+		kfree(iov);
+	}
+	if (ret)
+		goto out;
+	bio = req->bio;
+	if (bdev)
+		bio_set_dev(bio, bdev);
+	if (bdev && meta_buffer && meta_len) {
+		meta = nvme_add_user_metadata(req, meta_buffer, meta_len,
+				meta_seed);
+		if (IS_ERR(meta)) {
+			ret = PTR_ERR(meta);
+			goto out_unmap;
 		}
+		*metap = meta;
 	}
 
-	return req;
+	return ret;
 
 out_unmap:
 	if (bio)
 		blk_rq_unmap_user(bio);
 out:
-	blk_mq_free_request(req);
-	return ERR_PTR(ret);
+	return ret;
 }
 
 static int nvme_submit_user_cmd(struct request_queue *q,
@@ -141,13 +141,19 @@ static int nvme_submit_user_cmd(struct request_queue *q,
 	struct bio *bio;
 	int ret;
 
-	req = nvme_alloc_user_request(q, cmd, ubuffer, bufflen, meta_buffer,
-			meta_len, meta_seed, &meta, timeout, vec, 0, 0);
+	req = nvme_alloc_user_request(q, cmd, 0, 0);
 	if (IS_ERR(req))
 		return PTR_ERR(req);
 
-	bio = req->bio;
+	req->timeout = timeout;
+	if (ubuffer && bufflen) {
+		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
+			meta_len, meta_seed, &meta, vec);
+		if (ret)
+			goto out;
+	}
 
+	bio = req->bio;
 	ret = nvme_execute_passthru_rq(req);
 
 	if (result)
@@ -157,6 +163,7 @@ static int nvme_submit_user_cmd(struct request_queue *q,
 						meta_len, ret);
 	if (bio)
 		blk_rq_unmap_user(bio);
+out:
 	blk_mq_free_request(req);
 	return ret;
 }
@@ -418,6 +425,7 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
 	blk_opf_t rq_flags = 0;
 	blk_mq_req_flags_t blk_flags = 0;
 	void *meta = NULL;
+	int ret;
 
 	if (!capable(CAP_SYS_ADMIN))
 		return -EACCES;
@@ -457,13 +465,18 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
 		rq_flags |= REQ_POLLED;
 
 retry:
-	req = nvme_alloc_user_request(q, &c, nvme_to_user_ptr(d.addr),
-			d.data_len, nvme_to_user_ptr(d.metadata),
-			d.metadata_len, 0, &meta, d.timeout_ms ?
-			msecs_to_jiffies(d.timeout_ms) : 0, vec, rq_flags,
-			blk_flags);
+	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
 	if (IS_ERR(req))
 		return PTR_ERR(req);
+	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
+
+	if (d.addr && d.data_len) {
+		ret = nvme_map_user_request(req, nvme_to_user_ptr(d.addr),
+			d.data_len, nvme_to_user_ptr(d.metadata),
+			d.metadata_len, 0, &meta, vec);
+		if (ret)
+			goto out_err;
+	}
 	req->end_io = nvme_uring_cmd_end_io;
 	req->end_io_data = ioucmd;
 
@@ -486,6 +499,9 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
 
 	blk_execute_rq_nowait(req, false);
 	return -EIOCBQUEUED;
+out_err:
+	blk_mq_free_request(req);
+	return ret;
 }
 
 static bool is_ctrl_ioctl(unsigned int cmd)
-- 
2.25.1

Re: [PATCH for-next v8 3/5] nvme: refactor nvme_alloc_user_request

[Christoph Hellwig]

> -static void *nvme_add_user_metadata(struct bio *bio, void __user *ubuf,
> -		unsigned len, u32 seed, bool write)
> +static void *nvme_add_user_metadata(struct request *req, void __user *ubuf,
> +		unsigned len, u32 seed)

Please split out a separate patch just for the changes to
nvme_add_user_metadata.

> -	if (ret == len)
> +	if (ret == len) {
> +		req->cmd_flags |= REQ_INTEGRITY;
>  		return buf;
> +	}
>  	ret = -ENOMEM;

Nit: Please keep the successful path in line and branch out for the
errors, i.e.


	if (ret != len) {
	 	ret = -ENOMEM;
		goto out_free_meta;
	}

	req->cmd_flags |= REQ_INTEGRITY;
	return buf;

We should have done this already with the old code, but now that more
is added to the success path it becomes even more important:.

> +	else {
> +		struct iovec fast_iov[UIO_FASTIOV];
> +		struct iovec *iov = fast_iov;
> +		struct iov_iter iter;
> +
> +		ret = import_iovec(rq_data_dir(req), ubuffer, bufflen,
> +				UIO_FASTIOV, &iov, &iter);
> +		if (ret < 0)
>  			goto out;
> +		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
> +		kfree(iov);
> +	}

While you touch this:  I think thi block of code would also be a good
separate helper.  Maybe even in the block layer given the the scsi
ioctl code and sg duplicate it, and already missed the fast_iov
treatment due to the duplication.  Having this in a separate function
is also nice to keep the fast_iov stack footprint isolated.

> +	if (ret)
> +		goto out;
> +	bio = req->bio;

I think we can also do away with this bio local variable now.

> +	if (bdev)
> +		bio_set_dev(bio, bdev);

We don't need the bio_set_dev here as mentioned last time, so I think
we should remove it in a prep patch.

> +		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
> +			meta_len, meta_seed, &meta, vec);

Nit: Add an extra tab for indentation here please.

>  	if (IS_ERR(req))
>  		return PTR_ERR(req);
> +	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;

	if (d.timeout_ms)
		req->timeout = msecs_to_jiffies(d.timeout_ms);

Re: [PATCH for-next v8 3/5] nvme: refactor nvme_alloc_user_request

[Kanchan Joshi]

[-- Attachment #1: Type: text/plain, Size: 1324 bytes --]

On Fri, Sep 23, 2022 at 05:38:19PM +0200, Christoph Hellwig wrote:
>
>> +	else {
>> +		struct iovec fast_iov[UIO_FASTIOV];
>> +		struct iovec *iov = fast_iov;
>> +		struct iov_iter iter;
>> +
>> +		ret = import_iovec(rq_data_dir(req), ubuffer, bufflen,
>> +				UIO_FASTIOV, &iov, &iter);
>> +		if (ret < 0)
>>  			goto out;
>> +		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
>> +		kfree(iov);
>> +	}
>
>While you touch this:  I think thi block of code would also be a good
>separate helper.  Maybe even in the block layer given the the scsi
>ioctl code and sg duplicate it, and already missed the fast_iov
>treatment due to the duplication.  Having this in a separate function
>is also nice to keep the fast_iov stack footprint isolated.

Totally agree on goodness. 
I think instead of new helper this seems suited to go inside
blk_rq_map_user_iov itself. That will make it symmetric to
blk_rq_map_user which also combines import + mapping.

But if I go that route now, I will have to alter parameters of
blk_rq_map_user_iov, and that will make it mandatory to change the
callers (scsi-ioctl, sg) too. Nothing hairy, but that means further
growth of unrelated elements in this series. Hope you agree that
separate series is much better, which I will post after this.

Will fold all other changes you pointed.

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]

Re: [PATCH for-next v8 3/5] nvme: refactor nvme_alloc_user_request

[Kanchan Joshi]

[-- Attachment #1: Type: text/plain, Size: 464 bytes --]

>> +	if (ret)
>> +		goto out;
>> +	bio = req->bio;
>
>I think we can also do away with this bio local variable now.
>
>> +	if (bdev)
>> +		bio_set_dev(bio, bdev);
>
>We don't need the bio_set_dev here as mentioned last time, so I think
>we should remove it in a prep patch.

we miss completing polled io with this change.
bdev needs to be put in bio to complete polled passthrough IO.
nvme_ns_chr_uring_cmd_iopoll uses bio_poll and that in turn makes use of
this.

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]

Re: [PATCH for-next v8 3/5] nvme: refactor nvme_alloc_user_request

[Christoph Hellwig]

On Mon, Sep 26, 2022 at 01:13:54AM +0530, Kanchan Joshi wrote:
>>> +	if (ret)
>>> +		goto out;
>>> +	bio = req->bio;
>>
>> I think we can also do away with this bio local variable now.
>>
>>> +	if (bdev)
>>> +		bio_set_dev(bio, bdev);
>>
>> We don't need the bio_set_dev here as mentioned last time, so I think
>> we should remove it in a prep patch.
>
> we miss completing polled io with this change.
> bdev needs to be put in bio to complete polled passthrough IO.
> nvme_ns_chr_uring_cmd_iopoll uses bio_poll and that in turn makes use of
> this.

Oh, indeed - polling is another and someone unexpected user in
addition to the I/O accounting that does not apply to passthrough
requests.  That also means we can't poll admin commands at all.

Re: [PATCH for-next v8 3/5] nvme: refactor nvme_alloc_user_request

[Kanchan Joshi]

[-- Attachment #1: Type: text/plain, Size: 1076 bytes --]

On Mon, Sep 26, 2022 at 04:51:59PM +0200, Christoph Hellwig wrote:
>On Mon, Sep 26, 2022 at 01:13:54AM +0530, Kanchan Joshi wrote:
>>>> +	if (ret)
>>>> +		goto out;
>>>> +	bio = req->bio;
>>>
>>> I think we can also do away with this bio local variable now.
>>>
>>>> +	if (bdev)
>>>> +		bio_set_dev(bio, bdev);
>>>
>>> We don't need the bio_set_dev here as mentioned last time, so I think
>>> we should remove it in a prep patch.
>>
>> we miss completing polled io with this change.
>> bdev needs to be put in bio to complete polled passthrough IO.
>> nvme_ns_chr_uring_cmd_iopoll uses bio_poll and that in turn makes use of
>> this.
>
>Oh, indeed - polling is another and someone unexpected user in
>addition to the I/O accounting that does not apply to passthrough
>requests.  That also means we can't poll admin commands at all.

Yes. That falls back to IRQ completions.

I think it should be possible to support if we use request-only
interface. Most of the information in bio-poll interface comes from
request.
But I doubt if polling for admin command is a useful thing.

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]

[PATCH for-next v8 4/5] block: add helper to map bvec iterator for passthrough

[Kanchan Joshi]

Add blk_rq_map_user_bvec which maps the bvec iterator into a bio and
places that into the request. This helper will be used in nvme for
uring-passthrough with fixed-buffer.
While at it, create another helper bio_map_get to reduce the code
duplication.

Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
Signed-off-by: Anuj Gupta <anuj20.g@samsung.com>
---
 block/blk-map.c        | 111 +++++++++++++++++++++++++++++++++++++----
 include/linux/blk-mq.h |   1 +
 2 files changed, 102 insertions(+), 10 deletions(-)

diff --git a/block/blk-map.c b/block/blk-map.c
index 7693f8e3c454..d6265d49b15b 100644
--- a/block/blk-map.c
+++ b/block/blk-map.c
@@ -241,17 +241,10 @@ static void bio_map_put(struct bio *bio)
 	}
 }
 
-static int bio_map_user_iov(struct request *rq, struct iov_iter *iter,
+static struct bio *bio_map_get(struct request *rq, unsigned int nr_vecs,
 		gfp_t gfp_mask)
 {
-	unsigned int max_sectors = queue_max_hw_sectors(rq->q);
-	unsigned int nr_vecs = iov_iter_npages(iter, BIO_MAX_VECS);
 	struct bio *bio;
-	int ret;
-	int j;
-
-	if (!iov_iter_count(iter))
-		return -EINVAL;
 
 	if (rq->cmd_flags & REQ_POLLED) {
 		blk_opf_t opf = rq->cmd_flags | REQ_ALLOC_CACHE;
@@ -259,13 +252,31 @@ static int bio_map_user_iov(struct request *rq, struct iov_iter *iter,
 		bio = bio_alloc_bioset(NULL, nr_vecs, opf, gfp_mask,
 					&fs_bio_set);
 		if (!bio)
-			return -ENOMEM;
+			return NULL;
 	} else {
 		bio = bio_kmalloc(nr_vecs, gfp_mask);
 		if (!bio)
-			return -ENOMEM;
+			return NULL;
 		bio_init(bio, NULL, bio->bi_inline_vecs, nr_vecs, req_op(rq));
 	}
+	return bio;
+}
+
+static int bio_map_user_iov(struct request *rq, struct iov_iter *iter,
+		gfp_t gfp_mask)
+{
+	unsigned int max_sectors = queue_max_hw_sectors(rq->q);
+	unsigned int nr_vecs = iov_iter_npages(iter, BIO_MAX_VECS);
+	struct bio *bio;
+	int ret;
+	int j;
+
+	if (!iov_iter_count(iter))
+		return -EINVAL;
+
+	bio = bio_map_get(rq, nr_vecs, gfp_mask);
+	if (bio == NULL)
+		return -ENOMEM;
 
 	while (iov_iter_count(iter)) {
 		struct page **pages, *stack_pages[UIO_FASTIOV];
@@ -611,6 +622,86 @@ int blk_rq_map_user(struct request_queue *q, struct request *rq,
 }
 EXPORT_SYMBOL(blk_rq_map_user);
 
+/* Prepare bio for passthrough IO given an existing bvec iter */
+int blk_rq_map_user_bvec(struct request *rq, struct iov_iter *iter)
+{
+	struct request_queue *q = rq->q;
+	size_t nr_iter, nr_segs, i;
+	struct bio *bio = NULL;
+	struct bio_vec *bv, *bvecs, *bvprvp = NULL;
+	struct queue_limits *lim = &q->limits;
+	unsigned int nsegs = 0, bytes = 0;
+	bool copy = false;
+	int ret;
+	unsigned long align = q->dma_pad_mask | queue_dma_alignment(q);
+
+	/* see if we need to copy pages due to any weird situation */
+	if (blk_queue_may_bounce(q))
+		copy = true;
+	else if (iov_iter_alignment(iter) & align)
+		copy = true;
+
+	if (copy) {
+		do {
+			ret = bio_copy_user_iov(rq, NULL, iter, GFP_KERNEL);
+			if (ret) {
+				blk_rq_unmap_user(bio);
+				rq->bio = NULL;
+				break;
+			}
+			if (!bio)
+				bio = rq->bio;
+		} while (iov_iter_count(iter));
+
+		return ret;
+	}
+	/* common (non-copy) case handling */
+	nr_iter = iov_iter_count(iter);
+	nr_segs = iter->nr_segs;
+
+	if (!nr_iter || (nr_iter >> SECTOR_SHIFT) > queue_max_hw_sectors(q))
+		return -EINVAL;
+	if (nr_segs > queue_max_segments(q))
+		return -EINVAL;
+
+	/* no iovecs to alloc, as we already have a BVEC iterator */
+	bio = bio_map_get(rq, 0, GFP_KERNEL);
+	if (bio == NULL)
+		return -ENOMEM;
+
+	bio_iov_bvec_set(bio, iter);
+	blk_rq_bio_prep(rq, bio, nr_segs);
+
+	/* loop to perform a bunch of sanity checks */
+	bvecs = (struct bio_vec *)iter->bvec;
+	for (i = 0; i < nr_segs; i++) {
+		bv = &bvecs[i];
+		/*
+		 * If the queue doesn't support SG gaps and adding this
+		 * offset would create a gap, disallow it.
+		 */
+		if (bvprvp && bvec_gap_to_prev(lim, bvprvp, bv->bv_offset))
+			goto put_bio;
+
+		/* check full condition */
+		if (nsegs >= nr_segs || bytes > UINT_MAX - bv->bv_len)
+			goto put_bio;
+		if (bytes + bv->bv_len > nr_iter)
+			goto put_bio;
+		if (bv->bv_offset + bv->bv_len > PAGE_SIZE)
+			goto put_bio;
+
+		nsegs++;
+		bytes += bv->bv_len;
+		bvprvp = bv;
+	}
+	return 0;
+put_bio:
+	bio_map_put(bio);
+	return -EINVAL;
+}
+EXPORT_SYMBOL_GPL(blk_rq_map_user_bvec);
+
 /**
  * blk_rq_unmap_user - unmap a request with user data
  * @bio:	       start of bio list
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 00a15808c137..1a9ae17e49be 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -977,6 +977,7 @@ struct rq_map_data {
 	bool from_user;
 };
 
+int blk_rq_map_user_bvec(struct request *rq, struct iov_iter *iter);
 int blk_rq_map_user(struct request_queue *, struct request *,
 		struct rq_map_data *, void __user *, unsigned long, gfp_t);
 int blk_rq_map_user_iov(struct request_queue *, struct request *,
-- 
2.25.1

Re: [PATCH for-next v8 4/5] block: add helper to map bvec iterator for passthrough

[Christoph Hellwig]

On Fri, Sep 23, 2022 at 02:58:53PM +0530, Kanchan Joshi wrote:
> Add blk_rq_map_user_bvec which maps the bvec iterator into a bio and
> places that into the request. This helper will be used in nvme for
> uring-passthrough with fixed-buffer.
> While at it, create another helper bio_map_get to reduce the code
> duplication.

As per discussion on the last round - this needs to into
blk_rq_map_user_iov.

[PATCH for-next v8 5/5] nvme: wire up fixed buffer support for nvme passthrough

[Kanchan Joshi]

if io_uring sends passthrough command with IORING_URING_CMD_FIXED flag,
use the pre-registered buffer to form the bio.
While at it, modify nvme_submit_user_cmd to take ubuffer as plain integer
argument, and do away with nvme_to_user_ptr conversion in callers.

Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
---
 drivers/nvme/host/ioctl.c | 39 ++++++++++++++++++++++++++-------------
 1 file changed, 26 insertions(+), 13 deletions(-)

diff --git a/drivers/nvme/host/ioctl.c b/drivers/nvme/host/ioctl.c
index 9537991deac9..0464a89c8f5a 100644
--- a/drivers/nvme/host/ioctl.c
+++ b/drivers/nvme/host/ioctl.c
@@ -81,9 +81,10 @@ static struct request *nvme_alloc_user_request(struct request_queue *q,
 	return req;
 }
 
-static int nvme_map_user_request(struct request *req, void __user *ubuffer,
+static int nvme_map_user_request(struct request *req, u64 ubuffer,
 		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
-		u32 meta_seed, void **metap, bool vec)
+		u32 meta_seed, void **metap, struct io_uring_cmd *ioucmd,
+		bool vec)
 {
 	struct request_queue *q = req->q;
 	struct nvme_ns *ns = q->queuedata;
@@ -91,17 +92,29 @@ static int nvme_map_user_request(struct request *req, void __user *ubuffer,
 	struct bio *bio = NULL;
 	void *meta = NULL;
 	int ret;
+	bool fixedbufs = ioucmd && (ioucmd->flags & IORING_URING_CMD_FIXED);
 
 	if (!vec)
-		ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen,
-			GFP_KERNEL);
+		if (!fixedbufs)
+			ret = blk_rq_map_user(q, req, NULL,
+					nvme_to_user_ptr(ubuffer), bufflen,
+					GFP_KERNEL);
+		else {
+			struct iov_iter iter;
+
+			ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
+					rq_data_dir(req), &iter, ioucmd);
+			if (ret < 0)
+				goto out;
+			ret = blk_rq_map_user_bvec(req, &iter);
+		}
 	else {
 		struct iovec fast_iov[UIO_FASTIOV];
 		struct iovec *iov = fast_iov;
 		struct iov_iter iter;
 
-		ret = import_iovec(rq_data_dir(req), ubuffer, bufflen,
-				UIO_FASTIOV, &iov, &iter);
+		ret = import_iovec(rq_data_dir(req), nvme_to_user_ptr(ubuffer),
+				bufflen, UIO_FASTIOV, &iov, &iter);
 		if (ret < 0)
 			goto out;
 		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
@@ -132,7 +145,7 @@ static int nvme_map_user_request(struct request *req, void __user *ubuffer,
 }
 
 static int nvme_submit_user_cmd(struct request_queue *q,
-		struct nvme_command *cmd, void __user *ubuffer,
+		struct nvme_command *cmd, u64 ubuffer,
 		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
 		u32 meta_seed, u64 *result, unsigned timeout, bool vec)
 {
@@ -148,7 +161,7 @@ static int nvme_submit_user_cmd(struct request_queue *q,
 	req->timeout = timeout;
 	if (ubuffer && bufflen) {
 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
-			meta_len, meta_seed, &meta, vec);
+			meta_len, meta_seed, &meta, NULL, vec);
 		if (ret)
 			goto out;
 	}
@@ -227,7 +240,7 @@ static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
 	c.rw.appmask = cpu_to_le16(io.appmask);
 
 	return nvme_submit_user_cmd(ns->queue, &c,
-			nvme_to_user_ptr(io.addr), length,
+			io.addr, length,
 			metadata, meta_len, lower_32_bits(io.slba), NULL, 0,
 			false);
 }
@@ -281,7 +294,7 @@ static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
 		timeout = msecs_to_jiffies(cmd.timeout_ms);
 
 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
-			nvme_to_user_ptr(cmd.addr), cmd.data_len,
+			cmd.addr, cmd.data_len,
 			nvme_to_user_ptr(cmd.metadata), cmd.metadata_len,
 			0, &result, timeout, false);
 
@@ -327,7 +340,7 @@ static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
 		timeout = msecs_to_jiffies(cmd.timeout_ms);
 
 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
-			nvme_to_user_ptr(cmd.addr), cmd.data_len,
+			cmd.addr, cmd.data_len,
 			nvme_to_user_ptr(cmd.metadata), cmd.metadata_len,
 			0, &cmd.result, timeout, vec);
 
@@ -471,9 +484,9 @@ static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
 
 	if (d.addr && d.data_len) {
-		ret = nvme_map_user_request(req, nvme_to_user_ptr(d.addr),
+		ret = nvme_map_user_request(req, d.addr,
 			d.data_len, nvme_to_user_ptr(d.metadata),
-			d.metadata_len, 0, &meta, vec);
+			d.metadata_len, 0, &meta, ioucmd, vec);
 		if (ret)
 			goto out_err;
 	}
-- 
2.25.1

Re: [PATCH for-next v8 5/5] nvme: wire up fixed buffer support for nvme passthrough

[Christoph Hellwig]

On Fri, Sep 23, 2022 at 02:58:54PM +0530, Kanchan Joshi wrote:
>  	if (!vec)
> +		if (!fixedbufs)
> +			ret = blk_rq_map_user(q, req, NULL,
> +					nvme_to_user_ptr(ubuffer), bufflen,
> +					GFP_KERNEL);
> +		else {
> +			struct iov_iter iter;
> +
> +			ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
> +					rq_data_dir(req), &iter, ioucmd);
> +			if (ret < 0)
> +				goto out;
> +			ret = blk_rq_map_user_bvec(req, &iter);
> +		}

Given that the fixedufs case doesn't handle the vec case we can
do with some untangling and a single level of indentation here.

Even with that a WARN_ON_ONCE() for that impossible case would be good
to have, though.

Re: [PATCH for-next v8 0/5] fixed-buffer for uring-cmd/passthru

[Jens Axboe]

On 9/23/22 3:28 AM, Kanchan Joshi wrote:
> Currently uring-cmd lacks the ability to leverage the pre-registered
> buffers. This series adds that support in uring-cmd, and plumbs
> nvme passthrough to work with it.
> 
> Using registered-buffers showed IOPS hike from 1.9M to 2.2M in my tests.

Ran my peak test on this, specifically:

t/io_uring -pX -d128 -b512 -s32 -c32 -F1 -B0 -R1 -X1 -n24 -P1 -u1 -O0 /dev/ng0n1 /dev/ng1n1 /dev/ng2n1 /dev/ng3n1 /dev/ng4n1 /dev/ng5n1 /dev/ng6n1 /dev/ng7n1 /dev/ng8n1 /dev/ng9n1 /dev/ng10n1 /dev/ng11n1 /dev/ng12n1 /dev/ng13n1 /dev/ng14n1 /dev/ng15n1 /dev/ng16n1 /dev/ng17n1 /dev/ng18n1 /dev/ng19n1 /dev/ng20n1 /dev/ng21n1 /dev/ng22n1 /dev/ng23n1

Before:

Polled (-p1): 96.8M IOPS
IRQ driven (-p0): 56.2M IOPS

With patches, set -B1 in the above:

Polled (-p1): 121.8M IOPS
IRQ driven (-p0): 68.7M IOPS

+22-26% improvement, which is not unexpected.

-- 
Jens Axboe

Re: [PATCH for-next v8 0/5] fixed-buffer for uring-cmd/passthru

[Kanchan Joshi]

[-- Attachment #1: Type: text/plain, Size: 1000 bytes --]

On Fri, Sep 23, 2022 at 08:15:28AM -0600, Jens Axboe wrote:
>On 9/23/22 3:28 AM, Kanchan Joshi wrote:
>> Currently uring-cmd lacks the ability to leverage the pre-registered
>> buffers. This series adds that support in uring-cmd, and plumbs
>> nvme passthrough to work with it.
>>
>> Using registered-buffers showed IOPS hike from 1.9M to 2.2M in my tests.
>
>Ran my peak test on this, specifically:
>
>t/io_uring -pX -d128 -b512 -s32 -c32 -F1 -B0 -R1 -X1 -n24 -P1 -u1 -O0 /dev/ng0n1 /dev/ng1n1 /dev/ng2n1 /dev/ng3n1 /dev/ng4n1 /dev/ng5n1 /dev/ng6n1 /dev/ng7n1 /dev/ng8n1 /dev/ng9n1 /dev/ng10n1 /dev/ng11n1 /dev/ng12n1 /dev/ng13n1 /dev/ng14n1 /dev/ng15n1 /dev/ng16n1 /dev/ng17n1 /dev/ng18n1 /dev/ng19n1 /dev/ng20n1 /dev/ng21n1 /dev/ng22n1 /dev/ng23n1
>
>Before:
>
>Polled (-p1): 96.8M IOPS
>IRQ driven (-p0): 56.2M IOPS
>
>With patches, set -B1 in the above:
>
>Polled (-p1): 121.8M IOPS
>IRQ driven (-p0): 68.7M IOPS
>
>+22-26% improvement, which is not unexpected.

Thanks for giving it a whirl.
>

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]