[fio.git] / memory.c

/*
 * Memory helpers
 */
#include <unistd.h>
#include <sys/shm.h>
#include <sys/mman.h>

#include "fio.h"

static void *pinned_mem;

void fio_unpin_memory(void)
{
	if (pinned_mem) {
		dprint(FD_MEM, "unpinning %llu bytes\n", mlock_size);
		if (munlock(pinned_mem, mlock_size) < 0)
			perror("munlock");
		munmap(pinned_mem, mlock_size);
		pinned_mem = NULL;
	}
}

int fio_pin_memory(void)
{
	unsigned long long phys_mem;

	if (!mlock_size)
		return 0;

	dprint(FD_MEM, "pinning %llu bytes\n", mlock_size);

	/*
	 * Don't allow mlock of more than real_mem-128MB
	 */
	phys_mem = os_phys_mem();
	if (phys_mem) {
		if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
			mlock_size = phys_mem - 128 * 1024 * 1024;
			log_info("fio: limiting mlocked memory to %lluMiB\n",
							mlock_size >> 20);
		}
	}

	pinned_mem = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE,
				MAP_PRIVATE | OS_MAP_ANON, 0, 0);
	if (pinned_mem == MAP_FAILED) {
		perror("malloc locked mem");
		pinned_mem = NULL;
		return 1;
	}
	if (mlock(pinned_mem, mlock_size) < 0) {
		perror("mlock");
		munmap(pinned_mem, mlock_size);
		pinned_mem = NULL;
		return 1;
	}

	return 0;
}

static int alloc_mem_shm(struct thread_data *td)
{
	int flags = IPC_CREAT | SHM_R | SHM_W;

	if (td->o.mem_type == MEM_SHMHUGE)
		flags |= SHM_HUGETLB;

	td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, flags);
	dprint(FD_MEM, "shmget %zu, %d\n", td->orig_buffer_size, td->shm_id);
	if (td->shm_id < 0) {
		td_verror(td, errno, "shmget");
		if (geteuid() != 0 && errno == ENOMEM)
			log_err("fio: you may need to run this job as root\n");
		if (td->o.mem_type == MEM_SHMHUGE) {
			if (errno == EINVAL) {
				log_err("fio: check that you have free huge"
					" pages and that hugepage-size is"
					" correct.\n");
			} else if (errno == ENOSYS) {
				log_err("fio: your system does not appear to"
					" support huge pages.\n");
			} else if (errno == ENOMEM) {
				log_err("fio: no huge pages available, do you"
					" need to alocate some? See HOWTO.\n");
			}
		}

		return 1;
	}

	td->orig_buffer = shmat(td->shm_id, NULL, 0);
	dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
	if (td->orig_buffer == (void *) -1) {
		td_verror(td, errno, "shmat");
		td->orig_buffer = NULL;
		return 1;
	}

	return 0;
}

static int alloc_mem_mmap(struct thread_data *td)
{
	int flags = MAP_PRIVATE;

	td->mmapfd = 0;

	if (td->mmapfile) {
		td->mmapfd = open(td->mmapfile, O_RDWR|O_CREAT, 0644);

		if (td->mmapfd < 0) {
			td_verror(td, errno, "open mmap file");
			td->orig_buffer = NULL;
			return 1;
		}
		if (ftruncate(td->mmapfd, td->orig_buffer_size) < 0) {
			td_verror(td, errno, "truncate mmap file");
			td->orig_buffer = NULL;
			return 1;
		}
	} else
		flags |= OS_MAP_ANON;

	td->orig_buffer = mmap(NULL, td->orig_buffer_size,
				PROT_READ | PROT_WRITE, flags, td->mmapfd, 0);
	dprint(FD_MEM, "mmap %zu/%d %p\n", td->orig_buffer_size, td->mmapfd,
							td->orig_buffer);
	if (td->orig_buffer == MAP_FAILED) {
		td_verror(td, errno, "mmap");
		td->orig_buffer = NULL;
		if (td->mmapfd) {
			close(td->mmapfd);
			unlink(td->mmapfile);
		}

		return 1;
	}

	return 0;
}

static int alloc_mem_malloc(struct thread_data *td)
{
	unsigned int bsize = td->orig_buffer_size;

	if (td->o.odirect)
		bsize += page_mask;

	td->orig_buffer = malloc(bsize);
	dprint(FD_MEM, "malloc %u %p\n", bsize, td->orig_buffer);
	if (td->orig_buffer)
		return 0;

	return 1;
}

/*
 * Setup the buffer area we need for io.
 */
int allocate_io_mem(struct thread_data *td)
{
	int ret = 0;

	if (td->io_ops->flags & FIO_NOIO)
		return 0;

	if (td->o.mem_type == MEM_MALLOC)
		ret = alloc_mem_malloc(td);
	else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
		ret = alloc_mem_shm(td);
	else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE)
		ret = alloc_mem_mmap(td);
	else {
		log_err("fio: bad mem type: %d\n", td->o.mem_type);
		ret = 1;
	}

	if (ret)
		td_verror(td, ENOMEM, "iomem allocation");

	return ret;
}

void free_io_mem(struct thread_data *td)
{
	if (td->o.mem_type == MEM_MALLOC) {
		dprint(FD_MEM, "free mem %p\n", td->orig_buffer);
		free(td->orig_buffer);
	} else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE) {
		struct shmid_ds sbuf;

		dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id,
							td->orig_buffer);
		shmdt(td->orig_buffer);
		shmctl(td->shm_id, IPC_RMID, &sbuf);
	} else if (td->o.mem_type == MEM_MMAP ||
		   td->o.mem_type == MEM_MMAPHUGE) {
		dprint(FD_MEM, "munmap %zu %p\n", td->orig_buffer_size,
							td->orig_buffer);
		munmap(td->orig_buffer, td->orig_buffer_size);
		if (td->mmapfile) {
			close(td->mmapfd);
			unlink(td->mmapfile);
			free(td->mmapfile);
		}
	} else
		log_err("Bad memory type %u\n", td->o.mem_type);

	td->orig_buffer = NULL;
}
Commit	Line	Data
2f9ade3c JA	1	/*
	2	* Memory helpers
	3	*/
	4	#include <unistd.h>
	5	#include <sys/shm.h>
	6	#include <sys/mman.h>
	7
	8	#include "fio.h"
2f9ade3c	9
2f9ade3c JA	10	static void *pinned_mem;
	11
	12	void fio_unpin_memory(void)
	13	{
	14	if (pinned_mem) {
ee56ad50	15	dprint(FD_MEM, "unpinning %llu bytes\n", mlock_size);
2f9ade3c JA	16	if (munlock(pinned_mem, mlock_size) < 0)
	17	perror("munlock");
	18	munmap(pinned_mem, mlock_size);
	19	pinned_mem = NULL;
	20	}
	21	}
	22
	23	int fio_pin_memory(void)
	24	{
	25	unsigned long long phys_mem;
	26
	27	if (!mlock_size)
	28	return 0;
	29
ee56ad50 JA	30	dprint(FD_MEM, "pinning %llu bytes\n", mlock_size);
ee56ad50 JA	31
2f9ade3c JA	32	/*
	33	* Don't allow mlock of more than real_mem-128MB
	34	*/
	35	phys_mem = os_phys_mem();
	36	if (phys_mem) {
	37	if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
	38	mlock_size = phys_mem - 128 * 1024 * 1024;
5ec10eaa JA	39	log_info("fio: limiting mlocked memory to %lluMiB\n",
5ec10eaa JA	40	mlock_size >> 20);
2f9ade3c JA	41	}
	42	}
	43
5ec10eaa JA	44	pinned_mem = mmap(NULL, mlock_size, PROT_READ \| PROT_WRITE,
5ec10eaa JA	45	MAP_PRIVATE \| OS_MAP_ANON, 0, 0);
2f9ade3c JA	46	if (pinned_mem == MAP_FAILED) {
	47	perror("malloc locked mem");
	48	pinned_mem = NULL;
	49	return 1;
	50	}
	51	if (mlock(pinned_mem, mlock_size) < 0) {
	52	perror("mlock");
	53	munmap(pinned_mem, mlock_size);
	54	pinned_mem = NULL;
	55	return 1;
	56	}
	57
	58	return 0;
	59	}
	60
b6f9676e	61	static int alloc_mem_shm(struct thread_data *td)
2f9ade3c	62	{
b6f9676e JA	63	int flags = IPC_CREAT \| SHM_R \| SHM_W;
	64
	65	if (td->o.mem_type == MEM_SHMHUGE)
	66	flags \|= SHM_HUGETLB;
	67
	68	td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, flags);
ee56ad50	69	dprint(FD_MEM, "shmget %zu, %d\n", td->orig_buffer_size, td->shm_id);
b6f9676e JA	70	if (td->shm_id < 0) {
	71	td_verror(td, errno, "shmget");
	72	if (geteuid() != 0 && errno == ENOMEM)
	73	log_err("fio: you may need to run this job as root\n");
886b878a	74	if (td->o.mem_type == MEM_SHMHUGE) {
5ec10eaa JA	75	if (errno == EINVAL) {
	76	log_err("fio: check that you have free huge"
	77	" pages and that hugepage-size is"
	78	" correct.\n");
	79	} else if (errno == ENOSYS) {
	80	log_err("fio: your system does not appear to"
	81	" support huge pages.\n");
	82	} else if (errno == ENOMEM) {
	83	log_err("fio: no huge pages available, do you"
	84	" need to alocate some? See HOWTO.\n");
	85	}
d8602dd0	86	}
5ec10eaa	87
b6f9676e JA	88	return 1;
b6f9676e JA	89	}
2f9ade3c	90
b6f9676e	91	td->orig_buffer = shmat(td->shm_id, NULL, 0);
ee56ad50	92	dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
b6f9676e JA	93	if (td->orig_buffer == (void *) -1) {
	94	td_verror(td, errno, "shmat");
	95	td->orig_buffer = NULL;
	96	return 1;
	97	}
	98
	99	return 0;
	100	}
	101
	102	static int alloc_mem_mmap(struct thread_data *td)
	103	{
	104	int flags = MAP_PRIVATE;
	105
	106	td->mmapfd = 0;
	107
	108	if (td->mmapfile) {
	109	td->mmapfd = open(td->mmapfile, O_RDWR\|O_CREAT, 0644);
	110
	111	if (td->mmapfd < 0) {
	112	td_verror(td, errno, "open mmap file");
2f9ade3c JA	113	td->orig_buffer = NULL;
	114	return 1;
	115	}
b6f9676e JA	116	if (ftruncate(td->mmapfd, td->orig_buffer_size) < 0) {
b6f9676e JA	117	td_verror(td, errno, "truncate mmap file");
2f9ade3c JA	118	td->orig_buffer = NULL;
	119	return 1;
	120	}
b6f9676e JA	121	} else
	122	flags \|= OS_MAP_ANON;
	123
5ec10eaa JA	124	td->orig_buffer = mmap(NULL, td->orig_buffer_size,
5ec10eaa JA	125	PROT_READ \| PROT_WRITE, flags, td->mmapfd, 0);
ee56ad50 JA	126	dprint(FD_MEM, "mmap %zu/%d %p\n", td->orig_buffer_size, td->mmapfd,
ee56ad50 JA	127	td->orig_buffer);
b6f9676e JA	128	if (td->orig_buffer == MAP_FAILED) {
	129	td_verror(td, errno, "mmap");
	130	td->orig_buffer = NULL;
	131	if (td->mmapfd) {
	132	close(td->mmapfd);
	133	unlink(td->mmapfile);
	134	}
5ec10eaa	135
b6f9676e	136	return 1;
2f9ade3c JA	137	}
	138
	139	return 0;
	140	}
	141
b6f9676e JA	142	static int alloc_mem_malloc(struct thread_data *td)
b6f9676e JA	143	{
d87612ac JA	144	unsigned int bsize = td->orig_buffer_size;
	145
	146	if (td->o.odirect)
	147	bsize += page_mask;
5ec10eaa	148
d87612ac	149	td->orig_buffer = malloc(bsize);
ee56ad50	150	dprint(FD_MEM, "malloc %u %p\n", bsize, td->orig_buffer);
b6f9676e JA	151	if (td->orig_buffer)
	152	return 0;
	153
	154	return 1;
	155	}
	156
	157	/*
	158	* Setup the buffer area we need for io.
	159	*/
	160	int allocate_io_mem(struct thread_data *td)
	161	{
	162	int ret = 0;
	163
b4c5e1ac JA	164	if (td->io_ops->flags & FIO_NOIO)
	165	return 0;
	166
b6f9676e JA	167	if (td->o.mem_type == MEM_MALLOC)
	168	ret = alloc_mem_malloc(td);
	169	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	170	ret = alloc_mem_shm(td);
	171	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
	172	ret = alloc_mem_mmap(td);
	173	else {
	174	log_err("fio: bad mem type: %d\n", td->o.mem_type);
	175	ret = 1;
	176	}
	177
3deb3101 JA	178	if (ret)
	179	td_verror(td, ENOMEM, "iomem allocation");
	180
b6f9676e JA	181	return ret;
	182	}
	183
2f9ade3c JA	184	void free_io_mem(struct thread_data *td)
2f9ade3c JA	185	{
ee56ad50 JA	186	if (td->o.mem_type == MEM_MALLOC) {
ee56ad50 JA	187	dprint(FD_MEM, "free mem %p\n", td->orig_buffer);
2f9ade3c	188	free(td->orig_buffer);
ee56ad50	189	} else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE) {
2f9ade3c JA	190	struct shmid_ds sbuf;
2f9ade3c JA	191
5ec10eaa JA	192	dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id,
5ec10eaa JA	193	td->orig_buffer);
2f9ade3c JA	194	shmdt(td->orig_buffer);
2f9ade3c JA	195	shmctl(td->shm_id, IPC_RMID, &sbuf);
2dc1bbeb JA	196	} else if (td->o.mem_type == MEM_MMAP \|\|
2dc1bbeb JA	197	td->o.mem_type == MEM_MMAPHUGE) {
ee56ad50 JA	198	dprint(FD_MEM, "munmap %zu %p\n", td->orig_buffer_size,
ee56ad50 JA	199	td->orig_buffer);
2f9ade3c	200	munmap(td->orig_buffer, td->orig_buffer_size);
313cb206 JA	201	if (td->mmapfile) {
	202	close(td->mmapfd);
	203	unlink(td->mmapfile);
	204	free(td->mmapfile);
d0bdaf49 JA	205	}
d0bdaf49 JA	206	} else
2dc1bbeb	207	log_err("Bad memory type %u\n", td->o.mem_type);
2f9ade3c JA	208
	209	td->orig_buffer = NULL;
	210	}