[fio.git] / engines / librpma_fio.h

/*
 * librpma_fio: librpma_apm and librpma_gpspm engines' common header.
 *
 * Copyright 2021, Intel Corporation
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License,
 * version 2 as published by the Free Software Foundation..
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
 * GNU General Public License for more details.
 */

#ifndef LIBRPMA_FIO_H
#define LIBRPMA_FIO_H 1

#include "../fio.h"
#include "../optgroup.h"

#include <librpma.h>

/* servers' and clients' common */

#define librpma_td_verror(td, err, func) \
	td_vmsg((td), (err), rpma_err_2str(err), (func))

/* ceil(a / b) = (a + b - 1) / b */
#define LIBRPMA_FIO_CEIL(a, b) (((a) + (b) - 1) / (b))

/* common option structure for server and client */
struct librpma_fio_options_values {
	/*
	 * FIO considers .off1 == 0 absent so the first meaningful field has to
	 * have padding ahead of it.
	 */
	void *pad;
	char *server_ip;
	/* base server listening port */
	char *port;
	/* Direct Write to PMem is possible */
	unsigned int direct_write_to_pmem;
	/* Set to 0 to wait for completion instead of busy-wait polling completion. */
	unsigned int busy_wait_polling;
};

extern struct fio_option librpma_fio_options[];

/*
 * Limited by the maximum length of the private data
 * for rdma_connect() in case of RDMA_PS_TCP (28 bytes).
 */
#define LIBRPMA_FIO_DESCRIPTOR_MAX_SIZE 24

struct librpma_fio_workspace {
	uint16_t max_msg_num;	/* # of RQ slots */
	uint8_t direct_write_to_pmem; /* Direct Write to PMem is possible */
	uint8_t mr_desc_size;	/* size of mr_desc in descriptor[] */
	/* buffer containing mr_desc */
	char descriptor[LIBRPMA_FIO_DESCRIPTOR_MAX_SIZE];
};

#define LIBRPMA_FIO_PORT_STR_LEN_MAX 12

int librpma_fio_td_port(const char *port_base_str, struct thread_data *td,
		char *port_out);

struct librpma_fio_mem {
	/* memory buffer */
	char *mem_ptr;

	/* size of the mapped persistent memory */
	size_t size_mmap;
};

char *librpma_fio_allocate_dram(struct thread_data *td, size_t size,
		struct librpma_fio_mem *mem);

char *librpma_fio_allocate_pmem(struct thread_data *td, struct fio_file *f,
		size_t size, struct librpma_fio_mem *mem);

void librpma_fio_free(struct librpma_fio_mem *mem);

/* clients' common */

typedef int (*librpma_fio_flush_t)(struct thread_data *td,
		struct io_u *first_io_u, struct io_u *last_io_u,
		unsigned long long int len);

/*
 * RETURN VALUE
 * - ( 1) - on success
 * - ( 0) - skip
 * - (-1) - on error
 */
typedef int (*librpma_fio_get_io_u_index_t)(struct ibv_wc *wc,
		unsigned int *io_u_index);

struct librpma_fio_client_data {
	struct rpma_peer *peer;
	struct rpma_conn *conn;
	struct rpma_cq *cq;

	/* aligned td->orig_buffer */
	char *orig_buffer_aligned;

	/* ious's base address memory registration (cd->orig_buffer_aligned) */
	struct rpma_mr_local *orig_mr;

	struct librpma_fio_workspace *ws;

	/* a server's memory representation */
	struct rpma_mr_remote *server_mr;
	enum rpma_flush_type server_mr_flush_type;

	/* remote workspace description */
	size_t ws_size;

	/* in-memory queues */
	struct io_u **io_us_queued;
	int io_u_queued_nr;
	struct io_u **io_us_flight;
	int io_u_flight_nr;
	struct io_u **io_us_completed;
	int io_u_completed_nr;

	/* SQ control. Note: all of them have to be kept in sync. */
	uint32_t op_send_posted;
	uint32_t op_send_completed;
	uint32_t op_recv_completed;

	librpma_fio_flush_t flush;
	librpma_fio_get_io_u_index_t get_io_u_index;

	/* engine-specific client data */
	void *client_data;
};

int librpma_fio_client_init(struct thread_data *td,
		struct rpma_conn_cfg *cfg);
void librpma_fio_client_cleanup(struct thread_data *td);

int librpma_fio_file_nop(struct thread_data *td, struct fio_file *f);
int librpma_fio_client_get_file_size(struct thread_data *td,
		struct fio_file *f);

int librpma_fio_client_post_init(struct thread_data *td);

enum fio_q_status librpma_fio_client_queue(struct thread_data *td,
		struct io_u *io_u);

int librpma_fio_client_commit(struct thread_data *td);

int librpma_fio_client_getevents(struct thread_data *td, unsigned int min,
		unsigned int max, const struct timespec *t);

struct io_u *librpma_fio_client_event(struct thread_data *td, int event);

char *librpma_fio_client_errdetails(struct io_u *io_u);

static inline int librpma_fio_client_io_read(struct thread_data *td,
		struct io_u *io_u, int flags)
{
	struct librpma_fio_client_data *ccd = td->io_ops_data;
	size_t dst_offset = (char *)(io_u->xfer_buf) - ccd->orig_buffer_aligned;
	size_t src_offset = io_u->offset;
	int ret;

	if ((ret = rpma_read(ccd->conn, ccd->orig_mr, dst_offset,
			ccd->server_mr, src_offset, io_u->xfer_buflen,
			flags, (void *)(uintptr_t)io_u->index))) {
		librpma_td_verror(td, ret, "rpma_read");
		return -1;
	}

	return 0;
}

static inline int librpma_fio_client_io_write(struct thread_data *td,
		struct io_u *io_u)
{
	struct librpma_fio_client_data *ccd = td->io_ops_data;
	size_t src_offset = (char *)(io_u->xfer_buf) - ccd->orig_buffer_aligned;
	size_t dst_offset = io_u->offset;
	int ret;

	if ((ret = rpma_write(ccd->conn, ccd->server_mr, dst_offset,
			ccd->orig_mr, src_offset, io_u->xfer_buflen,
			RPMA_F_COMPLETION_ON_ERROR,
			(void *)(uintptr_t)io_u->index))) {
		librpma_td_verror(td, ret, "rpma_write");
		return -1;
	}

	return 0;
}

static inline int librpma_fio_client_io_complete_all_sends(
		struct thread_data *td)
{
	struct librpma_fio_client_data *ccd = td->io_ops_data;
	struct ibv_wc wc;
	int ret;

	while (ccd->op_send_posted != ccd->op_send_completed) {
		/* get a completion */
		ret = rpma_cq_get_wc(ccd->cq, 1, &wc, NULL);
		if (ret == RPMA_E_NO_COMPLETION) {
			/* lack of completion is not an error */
			continue;
		} else if (ret != 0) {
			/* an error occurred */
			librpma_td_verror(td, ret, "rpma_cq_get_wc");
			break;
		}

		if (wc.status != IBV_WC_SUCCESS)
			return -1;

		if (wc.opcode == IBV_WC_SEND)
			++ccd->op_send_completed;
		else {
			log_err(
				"A completion other than IBV_WC_SEND got during cleaning up the CQ from SENDs\n");
			return -1;
		}
	}

	/*
	 * All posted SENDs are completed and RECVs for them (responses) are
	 * completed. This is the initial situation so the counters are reset.
	 */
	if (ccd->op_send_posted == ccd->op_send_completed &&
			ccd->op_send_completed == ccd->op_recv_completed) {
		ccd->op_send_posted = 0;
		ccd->op_send_completed = 0;
		ccd->op_recv_completed = 0;
	}

	return 0;
}

/* servers' common */

typedef int (*librpma_fio_prepare_connection_t)(
		struct thread_data *td,
		struct rpma_conn_req *conn_req);

struct librpma_fio_server_data {
	struct rpma_peer *peer;

	/* resources of an incoming connection */
	struct rpma_conn *conn;
	struct rpma_cq *cq;

	char *ws_ptr;
	struct rpma_mr_local *ws_mr;
	struct librpma_fio_mem mem;

	/* engine-specific server data */
	void *server_data;

	librpma_fio_prepare_connection_t prepare_connection;
};

int librpma_fio_server_init(struct thread_data *td);

void librpma_fio_server_cleanup(struct thread_data *td);

int librpma_fio_server_open_file(struct thread_data *td,
		struct fio_file *f, struct rpma_conn_cfg *cfg);

int librpma_fio_server_close_file(struct thread_data *td,
		struct fio_file *f);

#endif /* LIBRPMA_FIO_H */
Commit	Line	Data
e4c4625f JM	1	/*
	2	* librpma_fio: librpma_apm and librpma_gpspm engines' common header.
	3	*
	4	* Copyright 2021, Intel Corporation
	5	*
	6	* This program is free software; you can redistribute it and/or
	7	* modify it under the terms of the GNU General Public License,
	8	* version 2 as published by the Free Software Foundation..
	9	*
	10	* This program is distributed in the hope that it will be useful,
	11	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	12	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	13	* GNU General Public License for more details.
	14	*/
	15
	16	#ifndef LIBRPMA_FIO_H
	17	#define LIBRPMA_FIO_H 1
	18
	19	#include "../fio.h"
	20	#include "../optgroup.h"
	21
	22	#include <librpma.h>
	23
	24	/* servers' and clients' common */
	25
	26	#define librpma_td_verror(td, err, func) \
	27	td_vmsg((td), (err), rpma_err_2str(err), (func))
	28
	29	/* ceil(a / b) = (a + b - 1) / b */
	30	#define LIBRPMA_FIO_CEIL(a, b) (((a) + (b) - 1) / (b))
	31
	32	/* common option structure for server and client */
	33	struct librpma_fio_options_values {
	34	/*
	35	* FIO considers .off1 == 0 absent so the first meaningful field has to
	36	* have padding ahead of it.
	37	*/
	38	void *pad;
	39	char *server_ip;
	40	/* base server listening port */
	41	char *port;
	42	/* Direct Write to PMem is possible */
	43	unsigned int direct_write_to_pmem;
6a229978 OS	44	/* Set to 0 to wait for completion instead of busy-wait polling completion. */
6a229978 OS	45	unsigned int busy_wait_polling;
e4c4625f JM	46	};
	47
	48	extern struct fio_option librpma_fio_options[];
	49
	50	/*
	51	* Limited by the maximum length of the private data
	52	* for rdma_connect() in case of RDMA_PS_TCP (28 bytes).
	53	*/
	54	#define LIBRPMA_FIO_DESCRIPTOR_MAX_SIZE 24
	55
	56	struct librpma_fio_workspace {
	57	uint16_t max_msg_num; /* # of RQ slots */
	58	uint8_t direct_write_to_pmem; /* Direct Write to PMem is possible */
	59	uint8_t mr_desc_size; /* size of mr_desc in descriptor[] */
	60	/* buffer containing mr_desc */
	61	char descriptor[LIBRPMA_FIO_DESCRIPTOR_MAX_SIZE];
	62	};
	63
	64	#define LIBRPMA_FIO_PORT_STR_LEN_MAX 12
	65
	66	int librpma_fio_td_port(const char port_base_str, struct thread_data td,
	67	char *port_out);
	68
	69	struct librpma_fio_mem {
	70	/* memory buffer */
	71	char *mem_ptr;
	72
	73	/* size of the mapped persistent memory */
	74	size_t size_mmap;
	75	};
	76
	77	char librpma_fio_allocate_dram(struct thread_data td, size_t size,
	78	struct librpma_fio_mem *mem);
	79
a6becc33	80	char librpma_fio_allocate_pmem(struct thread_data td, struct fio_file *f,
e4c4625f JM	81	size_t size, struct librpma_fio_mem *mem);
	82
	83	void librpma_fio_free(struct librpma_fio_mem *mem);
	84
	85	/* clients' common */
	86
	87	typedef int (librpma_fio_flush_t)(struct thread_data td,
	88	struct io_u first_io_u, struct io_u last_io_u,
	89	unsigned long long int len);
	90
	91	/*
	92	* RETURN VALUE
	93	* - ( 1) - on success
	94	* - ( 0) - skip
	95	* - (-1) - on error
	96	*/
4ef7dd21	97	typedef int (librpma_fio_get_io_u_index_t)(struct ibv_wc wc,
e4c4625f JM	98	unsigned int *io_u_index);
	99
	100	struct librpma_fio_client_data {
	101	struct rpma_peer *peer;
	102	struct rpma_conn *conn;
4ef7dd21	103	struct rpma_cq *cq;
e4c4625f JM	104
	105	/* aligned td->orig_buffer */
	106	char *orig_buffer_aligned;
	107
	108	/* ious's base address memory registration (cd->orig_buffer_aligned) */
	109	struct rpma_mr_local *orig_mr;
	110
	111	struct librpma_fio_workspace *ws;
	112
	113	/* a server's memory representation */
	114	struct rpma_mr_remote *server_mr;
	115	enum rpma_flush_type server_mr_flush_type;
	116
	117	/* remote workspace description */
	118	size_t ws_size;
	119
	120	/* in-memory queues */
	121	struct io_u **io_us_queued;
	122	int io_u_queued_nr;
	123	struct io_u **io_us_flight;
	124	int io_u_flight_nr;
	125	struct io_u **io_us_completed;
	126	int io_u_completed_nr;
	127
	128	/* SQ control. Note: all of them have to be kept in sync. */
	129	uint32_t op_send_posted;
	130	uint32_t op_send_completed;
	131	uint32_t op_recv_completed;
	132
	133	librpma_fio_flush_t flush;
	134	librpma_fio_get_io_u_index_t get_io_u_index;
	135
	136	/* engine-specific client data */
	137	void *client_data;
	138	};
	139
	140	int librpma_fio_client_init(struct thread_data *td,
	141	struct rpma_conn_cfg *cfg);
	142	void librpma_fio_client_cleanup(struct thread_data *td);
	143
	144	int librpma_fio_file_nop(struct thread_data td, struct fio_file f);
	145	int librpma_fio_client_get_file_size(struct thread_data *td,
	146	struct fio_file *f);
	147
	148	int librpma_fio_client_post_init(struct thread_data *td);
	149
	150	enum fio_q_status librpma_fio_client_queue(struct thread_data *td,
	151	struct io_u *io_u);
	152
	153	int librpma_fio_client_commit(struct thread_data *td);
	154
	155	int librpma_fio_client_getevents(struct thread_data *td, unsigned int min,
	156	unsigned int max, const struct timespec *t);
	157
	158	struct io_u librpma_fio_client_event(struct thread_data td, int event);
	159
	160	char librpma_fio_client_errdetails(struct io_u io_u);
	161
	162	static inline int librpma_fio_client_io_read(struct thread_data *td,
	163	struct io_u *io_u, int flags)
	164	{
	165	struct librpma_fio_client_data *ccd = td->io_ops_data;
	166	size_t dst_offset = (char *)(io_u->xfer_buf) - ccd->orig_buffer_aligned;
	167	size_t src_offset = io_u->offset;
168	int ret;
169
170	if ((ret = rpma_read(ccd->conn, ccd->orig_mr, dst_offset,
171	ccd->server_mr, src_offset, io_u->xfer_buflen,
172	flags, (void *)(uintptr_t)io_u->index))) {
173	librpma_td_verror(td, ret, "rpma_read");
174	return -1;
175	}
176
177	return 0;
178	}
179
180	static inline int librpma_fio_client_io_write(struct thread_data *td,
181	struct io_u *io_u)
182	{
183	struct librpma_fio_client_data *ccd = td->io_ops_data;
184	size_t src_offset = (char *)(io_u->xfer_buf) - ccd->orig_buffer_aligned;
185	size_t dst_offset = io_u->offset;
186	int ret;
187
188	if ((ret = rpma_write(ccd->conn, ccd->server_mr, dst_offset,
189	ccd->orig_mr, src_offset, io_u->xfer_buflen,
190	RPMA_F_COMPLETION_ON_ERROR,
191	(void *)(uintptr_t)io_u->index))) {
192	librpma_td_verror(td, ret, "rpma_write");
193	return -1;
194	}
195
196	return 0;
197	}
198
199	static inline int librpma_fio_client_io_complete_all_sends(
200	struct thread_data *td)
201	{
202	struct librpma_fio_client_data *ccd = td->io_ops_data;
4ef7dd21	203	struct ibv_wc wc;
e4c4625f JM	204	int ret;
	205
	206	while (ccd->op_send_posted != ccd->op_send_completed) {
	207	/* get a completion */
4ef7dd21	208	ret = rpma_cq_get_wc(ccd->cq, 1, &wc, NULL);
e4c4625f JM	209	if (ret == RPMA_E_NO_COMPLETION) {
	210	/* lack of completion is not an error */
	211	continue;
	212	} else if (ret != 0) {
	213	/* an error occurred */
4ef7dd21	214	librpma_td_verror(td, ret, "rpma_cq_get_wc");
e4c4625f JM	215	break;
	216	}
	217
4ef7dd21	218	if (wc.status != IBV_WC_SUCCESS)
e4c4625f JM	219	return -1;
e4c4625f JM	220
4ef7dd21	221	if (wc.opcode == IBV_WC_SEND)
e4c4625f JM	222	++ccd->op_send_completed;
	223	else {
	224	log_err(
4ef7dd21	225	"A completion other than IBV_WC_SEND got during cleaning up the CQ from SENDs\n");
e4c4625f JM	226	return -1;
	227	}
	228	}
	229
	230	/*
	231	* All posted SENDs are completed and RECVs for them (responses) are
	232	* completed. This is the initial situation so the counters are reset.
	233	*/
	234	if (ccd->op_send_posted == ccd->op_send_completed &&
	235	ccd->op_send_completed == ccd->op_recv_completed) {
	236	ccd->op_send_posted = 0;
	237	ccd->op_send_completed = 0;
	238	ccd->op_recv_completed = 0;
	239	}
	240
	241	return 0;
	242	}
	243
	244	/* servers' common */
	245
	246	typedef int (*librpma_fio_prepare_connection_t)(
	247	struct thread_data *td,
	248	struct rpma_conn_req *conn_req);
	249
	250	struct librpma_fio_server_data {
	251	struct rpma_peer *peer;
	252
	253	/* resources of an incoming connection */
	254	struct rpma_conn *conn;
4ef7dd21	255	struct rpma_cq *cq;
e4c4625f JM	256
	257	char *ws_ptr;
	258	struct rpma_mr_local *ws_mr;
	259	struct librpma_fio_mem mem;
	260
	261	/* engine-specific server data */
	262	void *server_data;
	263
	264	librpma_fio_prepare_connection_t prepare_connection;
	265	};
	266
	267	int librpma_fio_server_init(struct thread_data *td);
	268
	269	void librpma_fio_server_cleanup(struct thread_data *td);
	270
	271	int librpma_fio_server_open_file(struct thread_data *td,
	272	struct fio_file f, struct rpma_conn_cfg cfg);
	273
	274	int librpma_fio_server_close_file(struct thread_data *td,
	275	struct fio_file *f);
	276
	277	#endif /* LIBRPMA_FIO_H */