fio: add fdp support for io_uring_cmd nvme engine

Add support for NVMe TP4146 Flexible Data Placemen, allowing placement
identifiers in write commands. The user can enabled this with the new
"fdp=1" parameter for fio's io_uring_cmd ioengine. By default, the fio
jobs will cycle through all the namespace's available placement
identifiers for write commands. The user can limit which placement
identifiers can be used with additional parameter, "fdp_pli=<list,>",
which can be used to separate write intensive jobs from less intensive
ones.

Setting up your namespace for FDP is outside the scope of 'fio', so this
assumes the namespace is already properly configured for the mode.

Link: https://lore.kernel.org/fio/CAKi7+wfX-eaUD5pky5cJ824uCzsQ4sPYMZdp3AuCUZOA1TQrYw@mail.gmail.com/T/#m056018eb07229bed00d4e589f9760b2a2aa009fc
Based-on-a-patch-by: Ankit Kumar <[email protected]>
Signed-off-by: Keith Busch <[email protected]>
Reviewed-by: Damien Le Moal <[email protected]>
[Vincent: fold in sfree fix from Ankit]
Signed-off-by: Vincent Fu <[email protected]>

Author: keithbusch

Diff for: HOWTO.rst

@@ -2423,6 +2423,18 @@ with the caveat that when used on the command line, they must come after the
 	For direct I/O, requests will only succeed if cache invalidation isn't required,
 	file blocks are fully allocated and the disk request could be issued immediately.
 
+.. option:: fdp=bool : [io_uring_cmd]
+
+	Enable Flexible Data Placement mode for write commands.
+
+.. option:: fdp_pli=str : [io_uring_cmd]
+
+	Select which Placement ID Index/Indicies this job is allowed to use for
+	writes. By default, the job will cycle through all available Placement
+        IDs, so use this to isolate these identifiers to specific jobs. If you
+        want fio to use placement identifier only at indices 0, 2 and 5 specify
+        ``fdp_pli=0,2,5``.
+
 .. option:: cpuload=int : [cpuio]
 
 	Attempt to use the specified percentage of CPU cycles. This is a mandatory

Diff for: Makefile

@@ -62,7 +62,7 @@ SOURCE :=	$(sort $(patsubst $(SRCDIR)/%,%,$(wildcard $(SRCDIR)/crc/*.c)) \
 		gettime-thread.c helpers.c json.c idletime.c td_error.c \
 		profiles/tiobench.c profiles/act.c io_u_queue.c filelock.c \
 		workqueue.c rate-submit.c optgroup.c helper_thread.c \
-		steadystate.c zone-dist.c zbd.c dedupe.c
+		steadystate.c zone-dist.c zbd.c dedupe.c fdp.c
 
 ifdef CONFIG_LIBHDFS
   HDFSFLAGS= -I $(JAVA_HOME)/include -I $(JAVA_HOME)/include/linux -I $(FIO_LIBHDFS_INCLUDE)

Diff for: cconv.c

@@ -349,6 +349,11 @@ int convert_thread_options_to_cpu(struct thread_options *o,
 
 	for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++)
 		o->merge_blktrace_iters[i].u.f = fio_uint64_to_double(le64_to_cpu(top->merge_blktrace_iters[i].u.i));
+
+	o->fdp = le32_to_cpu(top->fdp);
+	o->fdp_nrpli = le32_to_cpu(top->fdp_nrpli);
+	for (i = 0; i < o->fdp_nrpli; i++)
+		o->fdp_plis[i] = le32_to_cpu(top->fdp_plis[i]);
 #if 0
 	uint8_t cpumask[FIO_TOP_STR_MAX];
 	uint8_t verify_cpumask[FIO_TOP_STR_MAX];
@@ -638,6 +643,11 @@ void convert_thread_options_to_net(struct thread_options_pack *top,
 
 	for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++)
 		top->merge_blktrace_iters[i].u.i = __cpu_to_le64(fio_double_to_uint64(o->merge_blktrace_iters[i].u.f));
+
+	top->fdp = cpu_to_le32(o->fdp);
+	top->fdp_nrpli = cpu_to_le32(o->fdp_nrpli);
+	for (i = 0; i < o->fdp_nrpli; i++)
+		top->fdp_plis[i] = cpu_to_le32(o->fdp_plis[i]);
 #if 0
 	uint8_t cpumask[FIO_TOP_STR_MAX];
 	uint8_t verify_cpumask[FIO_TOP_STR_MAX];

Diff for: engines/io_uring.c

@@ -1262,6 +1262,29 @@ static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td,
 	return fio_nvme_get_max_open_zones(td, f, max_open_zones);
 }
 
+static int fio_ioring_cmd_fetch_ruhs(struct thread_data *td, struct fio_file *f,
+				     struct fio_ruhs_info *fruhs_info)
+{
+	struct nvme_fdp_ruh_status *ruhs;
+	int bytes, ret, i;
+
+	bytes = sizeof(*ruhs) + 128 * sizeof(struct nvme_fdp_ruh_status_desc);
+	ruhs = scalloc(1, bytes);
+	if (!ruhs)
+		return -ENOMEM;
+
+	ret = fio_nvme_iomgmt_ruhs(td, f, ruhs, bytes);
+	if (ret)
+		goto free;
+
+	fruhs_info->nr_ruhs = le16_to_cpu(ruhs->nruhsd);
+	for (i = 0; i < fruhs_info->nr_ruhs; i++)
+		fruhs_info->plis[i] = le16_to_cpu(ruhs->ruhss[i].pid);
+free:
+	sfree(ruhs);
+	return ret;
+}
+
 static struct ioengine_ops ioengine_uring = {
 	.name			= "io_uring",
 	.version		= FIO_IOOPS_VERSION,
@@ -1307,6 +1330,7 @@ static struct ioengine_ops ioengine_uring_cmd = {
 	.get_max_open_zones	= fio_ioring_cmd_get_max_open_zones,
 	.options		= options,
 	.option_struct_size	= sizeof(struct ioring_options),
+	.fdp_fetch_ruhs		= fio_ioring_cmd_fetch_ruhs,
 };
 
 static void fio_init fio_ioring_register(void)

Diff for: engines/nvme.c

@@ -28,7 +28,8 @@ int fio_nvme_uring_cmd_prep(struct nvme_uring_cmd *cmd, struct io_u *io_u,
 	cmd->cdw10 = slba & 0xffffffff;
 	cmd->cdw11 = slba >> 32;
 	/* cdw12 represent number of lba's for read/write */
-	cmd->cdw12 = nlb;
+	cmd->cdw12 = nlb | (io_u->dtype << 20);
+	cmd->cdw13 = io_u->dspec << 16;
 	if (iov) {
 		iov->iov_base = io_u->xfer_buf;
 		iov->iov_len = io_u->xfer_buflen;
@@ -345,3 +346,40 @@ int fio_nvme_get_max_open_zones(struct thread_data *td, struct fio_file *f,
 	close(fd);
 	return ret;
 }
+
+static inline int nvme_fdp_reclaim_unit_handle_status(int fd, __u32 nsid,
+						      __u32 data_len, void *data)
+{
+	struct nvme_passthru_cmd cmd = {
+		.opcode		= nvme_cmd_io_mgmt_recv,
+		.nsid		= nsid,
+		.addr		= (__u64)(uintptr_t)data,
+		.data_len 	= data_len,
+		.cdw10		= 1,
+		.cdw11		= (data_len >> 2) - 1,
+	};
+
+	return ioctl(fd, NVME_IOCTL_IO_CMD, &cmd);
+}
+
+int fio_nvme_iomgmt_ruhs(struct thread_data *td, struct fio_file *f,
+			 struct nvme_fdp_ruh_status *ruhs, __u32 bytes)
+{
+	struct nvme_data *data = FILE_ENG_DATA(f);
+	int fd, ret;
+
+	fd = open(f->file_name, O_RDONLY | O_LARGEFILE);
+	if (fd < 0)
+		return -errno;
+
+	ret = nvme_fdp_reclaim_unit_handle_status(fd, data->nsid, bytes, ruhs);
+	if (ret) {
+		log_err("%s: nvme_fdp_reclaim_unit_handle_status failed, err=%d\n",
+			f->file_name, ret);
+		errno = ENOTSUP;
+	} else
+		errno = 0;
+
+	close(fd);
+	return -errno;
+}

Diff for: engines/nvme.h

@@ -67,6 +67,7 @@ enum nvme_admin_opcode {
 enum nvme_io_opcode {
 	nvme_cmd_write			= 0x01,
 	nvme_cmd_read			= 0x02,
+	nvme_cmd_io_mgmt_recv		= 0x12,
 	nvme_zns_cmd_mgmt_send		= 0x79,
 	nvme_zns_cmd_mgmt_recv		= 0x7a,
 };
@@ -192,6 +193,23 @@ struct nvme_zone_report {
 	struct nvme_zns_desc	entries[];
 };
 
+struct nvme_fdp_ruh_status_desc {
+	__u16 pid;
+	__u16 ruhid;
+	__u32 earutr;
+	__u64 ruamw;
+	__u8  rsvd16[16];
+};
+
+struct nvme_fdp_ruh_status {
+	__u8  rsvd0[14];
+	__le16 nruhsd;
+	struct nvme_fdp_ruh_status_desc ruhss[];
+};
+
+int fio_nvme_iomgmt_ruhs(struct thread_data *td, struct fio_file *f,
+			 struct nvme_fdp_ruh_status *ruhs, __u32 bytes);
+
 int fio_nvme_get_info(struct fio_file *f, __u32 *nsid, __u32 *lba_sz,
 		      __u64 *nlba);
 

Diff for: examples/uring-cmd-fdp.fio

@@ -0,0 +1,37 @@
+# io_uring_cmd I/O engine for nvme-ns generic character device with FDP enabled
+# This assumes the namespace is already configured with FDP support and has at
+# least 8 available reclaim units.
+#
+# Each job targets different ranges of LBAs with different placement
+# identifiers, and has different write intensity.
+
+[global]
+filename=/dev/ng0n1
+ioengine=io_uring_cmd
+cmd_type=nvme
+iodepth=32
+bs=4K
+fdp=1
+time_based=1
+runtime=1000
+
+[write-heavy]
+rw=randrw
+rwmixwrite=90
+fdp_pli=0,1,2,3
+offset=0%
+size=30%
+
+[write-mid]
+rw=randrw
+rwmixwrite=30
+fdp_pli=4,5
+offset=30%
+size=30%
+
+[write-light]
+rw=randrw
+rwmixwrite=10
+fdp_pli=6
+offset=60%
+size=30%

Diff for: fdp.c

@@ -0,0 +1,119 @@
+/*
+ * Note: This is similar to a very basic setup
+ * of ZBD devices
+ *
+ * Specify fdp=1 (With char devices /dev/ng0n1)
+ */
+
+#include <errno.h>
+#include <string.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include "file.h"
+#include "fio.h"
+
+#include "pshared.h"
+#include "fdp.h"
+
+static int fdp_ruh_info(struct thread_data *td, struct fio_file *f,
+			struct fio_ruhs_info *ruhs)
+{
+	int ret = -EINVAL;
+
+	if (td->io_ops && td->io_ops->fdp_fetch_ruhs) {
+		ret = td->io_ops->fdp_fetch_ruhs(td, f, ruhs);
+		if (ret < 0) {
+			td_verror(td, errno, "fdp fetch ruhs failed");
+			log_err("%s: fdp fetch ruhs failed (%d)\n",
+				f->file_name, errno);
+		}
+	} else
+		log_err("%s: engine (%s) lacks fetch ruhs\n",
+			f->file_name, td->io_ops->name);
+
+	return ret;
+}
+
+static int init_ruh_info(struct thread_data *td, struct fio_file *f)
+{
+	struct fio_ruhs_info *ruhs, *tmp;
+	int i, ret;
+
+	ruhs = scalloc(1, sizeof(*ruhs) + 128 * sizeof(*ruhs->plis));
+	if (!ruhs)
+		return -ENOMEM;
+
+	ret = fdp_ruh_info(td, f, ruhs);
+	if (ret) {
+		log_info("fio: ruh info failed for %s (%d)\n",
+			 f->file_name, -ret);
+		goto out;
+	}
+
+	if (ruhs->nr_ruhs > 128)
+		ruhs->nr_ruhs = 128;
+
+	if (td->o.fdp_nrpli == 0) {
+		f->ruhs_info = ruhs;
+		return 0;
+	}
+
+	for (i = 0; i < td->o.fdp_nrpli; i++) {
+		if (td->o.fdp_plis[i] > ruhs->nr_ruhs) {
+			ret = -EINVAL;
+			goto out;
+		}
+	}
+
+	tmp = scalloc(1, sizeof(*tmp) + ruhs->nr_ruhs * sizeof(*tmp->plis));
+	if (!tmp) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	tmp->nr_ruhs = td->o.fdp_nrpli;
+	for (i = 0; i < td->o.fdp_nrpli; i++)
+		tmp->plis[i] = ruhs->plis[td->o.fdp_plis[i]];
+	f->ruhs_info = tmp;
+out:
+	sfree(ruhs);
+	return ret;
+}
+
+int fdp_init(struct thread_data *td)
+{
+	struct fio_file *f;
+	int i, ret = 0;
+
+	for_each_file(td, f, i) {
+		ret = init_ruh_info(td, f);
+		if (ret)
+			break;
+	}
+	return ret;
+}
+
+void fdp_free_ruhs_info(struct fio_file *f)
+{
+	if (!f->ruhs_info)
+		return;
+	sfree(f->ruhs_info);
+	f->ruhs_info = NULL;
+}
+
+void fdp_fill_dspec_data(struct thread_data *td, struct io_u *io_u)
+{
+	struct fio_file *f = io_u->file;
+	struct fio_ruhs_info *ruhs = f->ruhs_info;
+	int dspec;
+
+	if (!ruhs || io_u->ddir != DDIR_WRITE) {
+		io_u->dtype = 0;
+		io_u->dspec = 0;
+		return;
+	}
+
+	dspec = ruhs->plis[ruhs->pli_loc++ % ruhs->nr_ruhs];
+	io_u->dtype = 2;
+	io_u->dspec = dspec;
+}

Diff for: fdp.h

@@ -0,0 +1,16 @@
+#ifndef FIO_FDP_H
+#define FIO_FDP_H
+
+#include "io_u.h"
+
+struct fio_ruhs_info {
+	uint32_t nr_ruhs;
+	uint32_t pli_loc;
+	uint16_t plis[];
+};
+
+int fdp_init(struct thread_data *td);
+void fdp_free_ruhs_info(struct fio_file *f);
+void fdp_fill_dspec_data(struct thread_data *td, struct io_u *io_u);
+
+#endif /* FIO_FDP_H */

Diff for: file.h

@@ -12,6 +12,7 @@
 
 /* Forward declarations */
 struct zoned_block_device_info;
+struct fdp_ruh_info;
 
 /*
  * The type of object we are working on
@@ -101,6 +102,8 @@ struct fio_file {
 	uint64_t file_offset;
 	uint64_t io_size;
 
+	struct fio_ruhs_info *ruhs_info;
+
 	/*
 	 * Zoned block device information. See also zonemode=zbd.
 	 */

Diff for: filesetup.c

@@ -1407,6 +1407,12 @@ int setup_files(struct thread_data *td)
 
 	td_restore_runstate(td, old_state);
 
+	if (td->o.fdp) {
+		err = fdp_init(td);
+		if (err)
+			goto err_out;
+	}
+
 	return 0;
 
 err_offset:
@@ -1584,6 +1590,8 @@ void fio_file_free(struct fio_file *f)
 {
 	if (fio_file_axmap(f))
 		axmap_free(f->io_axmap);
+	if (f->ruhs_info)
+		sfree(f->ruhs_info);
 	if (!fio_file_smalloc(f)) {
 		free(f->file_name);
 		free(f);
@@ -1617,6 +1625,7 @@ void close_and_free_files(struct thread_data *td)
 		}
 
 		zbd_close_file(f);
+		fdp_free_ruhs_info(f);
 		fio_file_free(f);
 	}