[fio.git] / memory.c

/*
 * Memory helpers
 */
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
#ifndef FIO_NO_HAVE_SHM_H
#include <sys/shm.h>
#endif
#include <sys/mman.h>

#include "fio.h"

void fio_unpin_memory(struct thread_data *td)
{
	if (td->pinned_mem) {
		dprint(FD_MEM, "unpinning %llu bytes\n", td->o.lockmem);
		if (munlock(td->pinned_mem, td->o.lockmem) < 0)
			perror("munlock");
		munmap(td->pinned_mem, td->o.lockmem);
		td->pinned_mem = NULL;
	}
}

int fio_pin_memory(struct thread_data *td)
{
	unsigned long long phys_mem;

	if (!td->o.lockmem)
		return 0;

	dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);

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

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

	return 0;
}

static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
{
	int flags = IPC_CREAT | S_IRUSR | S_IWUSR;

	if (td->o.mem_type == MEM_SHMHUGE) {
		unsigned long mask = td->o.hugepage_size - 1;

		flags |= SHM_HUGETLB;
		total_mem = (total_mem + mask) & ~mask;
	}

	td->shm_id = shmget(IPC_PRIVATE, total_mem, flags);
	dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id);
	if (td->shm_id < 0) {
		td_verror(td, errno, "shmget");
		if (geteuid() != 0 && (errno == ENOMEM || errno == EPERM))
			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 void free_mem_shm(struct thread_data *td)
{
	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);
}

static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
{
	int flags = 0;

	td->mmapfd = 1;

	if (td->o.mem_type == MEM_MMAPHUGE) {
		unsigned long mask = td->o.hugepage_size - 1;

		/* TODO: make sure the file is a real hugetlbfs file */
		if (!td->o.mmapfile)
			flags |= MAP_HUGETLB;
		total_mem = (total_mem + mask) & ~mask;
	}

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

		if (td->mmapfd < 0) {
			td_verror(td, errno, "open mmap file");
			td->orig_buffer = NULL;
			return 1;
		}
		if (td->o.mem_type != MEM_MMAPHUGE &&
		    ftruncate(td->mmapfd, total_mem) < 0) {
			td_verror(td, errno, "truncate mmap file");
			td->orig_buffer = NULL;
			return 1;
		}
		if (td->o.mem_type == MEM_MMAPHUGE)
			flags |= MAP_SHARED;
		else
			flags |= MAP_PRIVATE;
	} else
		flags |= OS_MAP_ANON | MAP_PRIVATE;

	td->orig_buffer = mmap(NULL, total_mem, PROT_READ | PROT_WRITE, flags,
				td->mmapfd, 0);
	dprint(FD_MEM, "mmap %u/%d %p\n", total_mem, 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->o.mmapfile);
		}

		return 1;
	}

	return 0;
}

static void free_mem_mmap(struct thread_data *td, size_t total_mem)
{
	dprint(FD_MEM, "munmap %u %p\n", total_mem, td->orig_buffer);
	munmap(td->orig_buffer, td->orig_buffer_size);
	if (td->o.mmapfile) {
		close(td->mmapfd);
		unlink(td->o.mmapfile);
		free(td->o.mmapfile);
	}
}

static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
{
	td->orig_buffer = malloc(total_mem);
	dprint(FD_MEM, "malloc %u %p\n", total_mem, td->orig_buffer);

	return td->orig_buffer == NULL;
}

static void free_mem_malloc(struct thread_data *td)
{
	dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
	free(td->orig_buffer);
}

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

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

	total_mem = td->orig_buffer_size;

	if (td->o.odirect || td->o.mem_align ||
	    (td->io_ops->flags & FIO_MEMALIGN)) {
		total_mem += page_mask;
		if (td->o.mem_align && td->o.mem_align > page_size)
			total_mem += td->o.mem_align - page_size;
	}

	dprint(FD_MEM, "Alloc %lu for buffers\n", (size_t) total_mem);

	if (td->o.mem_type == MEM_MALLOC)
		ret = alloc_mem_malloc(td, total_mem);
	else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
		ret = alloc_mem_shm(td, total_mem);
	else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE)
		ret = alloc_mem_mmap(td, total_mem);
	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)
{
	unsigned int total_mem;

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

	if (td->o.mem_type == MEM_MALLOC)
		free_mem_malloc(td);
	else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE)
		free_mem_shm(td);
	else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE)
		free_mem_mmap(td, total_mem);
	else
		log_err("Bad memory type %u\n", td->o.mem_type);

	td->orig_buffer = NULL;
	td->orig_buffer_size = 0;
}
Commit	Line	Data
2f9ade3c JA	1	/*
	2	* Memory helpers
	3	*/
5921e80c JA	4	#include <sys/types.h>
	5	#include <sys/stat.h>
	6	#include <fcntl.h>
2f9ade3c	7	#include <unistd.h>
ec5c6b12	8	#ifndef FIO_NO_HAVE_SHM_H
2f9ade3c	9	#include <sys/shm.h>
ec5c6b12	10	#endif
2f9ade3c JA	11	#include <sys/mman.h>
	12
	13	#include "fio.h"
2f9ade3c	14
1b79a070	15	void fio_unpin_memory(struct thread_data *td)
2f9ade3c	16	{
1b79a070 JA	17	if (td->pinned_mem) {
	18	dprint(FD_MEM, "unpinning %llu bytes\n", td->o.lockmem);
	19	if (munlock(td->pinned_mem, td->o.lockmem) < 0)
2f9ade3c	20	perror("munlock");
1b79a070 JA	21	munmap(td->pinned_mem, td->o.lockmem);
1b79a070 JA	22	td->pinned_mem = NULL;
2f9ade3c JA	23	}
	24	}
	25
1b79a070	26	int fio_pin_memory(struct thread_data *td)
2f9ade3c JA	27	{
	28	unsigned long long phys_mem;
	29
1b79a070	30	if (!td->o.lockmem)
2f9ade3c JA	31	return 0;
2f9ade3c JA	32
1b79a070	33	dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
ee56ad50	34
2f9ade3c JA	35	/*
	36	* Don't allow mlock of more than real_mem-128MB
	37	*/
	38	phys_mem = os_phys_mem();
	39	if (phys_mem) {
1b79a070 JA	40	if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) {
1b79a070 JA	41	td->o.lockmem = phys_mem - 128 * 1024 * 1024;
b22989b9	42	log_info("fio: limiting mlocked memory to %lluMB\n",
1b79a070	43	td->o.lockmem >> 20);
2f9ade3c JA	44	}
	45	}
	46
1b79a070	47	td->pinned_mem = mmap(NULL, td->o.lockmem, PROT_READ \| PROT_WRITE,
a55820db	48	MAP_PRIVATE \| OS_MAP_ANON, -1, 0);
1b79a070	49	if (td->pinned_mem == MAP_FAILED) {
2f9ade3c	50	perror("malloc locked mem");
1b79a070	51	td->pinned_mem = NULL;
2f9ade3c JA	52	return 1;
2f9ade3c JA	53	}
1b79a070	54	if (mlock(td->pinned_mem, td->o.lockmem) < 0) {
2f9ade3c	55	perror("mlock");
1b79a070 JA	56	munmap(td->pinned_mem, td->o.lockmem);
1b79a070 JA	57	td->pinned_mem = NULL;
2f9ade3c JA	58	return 1;
	59	}
	60
	61	return 0;
	62	}
	63
829a602c	64	static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
2f9ade3c	65	{
03e20d68	66	int flags = IPC_CREAT \| S_IRUSR \| S_IWUSR;
b6f9676e	67
a1242a20 JA	68	if (td->o.mem_type == MEM_SHMHUGE) {
	69	unsigned long mask = td->o.hugepage_size - 1;
	70
b6f9676e	71	flags \|= SHM_HUGETLB;
a1242a20 JA	72	total_mem = (total_mem + mask) & ~mask;
a1242a20 JA	73	}
b6f9676e	74
829a602c JA	75	td->shm_id = shmget(IPC_PRIVATE, total_mem, flags);
829a602c JA	76	dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id);
b6f9676e JA	77	if (td->shm_id < 0) {
b6f9676e JA	78	td_verror(td, errno, "shmget");
da7d79b0	79	if (geteuid() != 0 && (errno == ENOMEM \|\| errno == EPERM))
b6f9676e	80	log_err("fio: you may need to run this job as root\n");
886b878a	81	if (td->o.mem_type == MEM_SHMHUGE) {
5ec10eaa JA	82	if (errno == EINVAL) {
	83	log_err("fio: check that you have free huge"
	84	" pages and that hugepage-size is"
	85	" correct.\n");
	86	} else if (errno == ENOSYS) {
	87	log_err("fio: your system does not appear to"
	88	" support huge pages.\n");
	89	} else if (errno == ENOMEM) {
	90	log_err("fio: no huge pages available, do you"
	91	" need to alocate some? See HOWTO.\n");
	92	}
d8602dd0	93	}
5ec10eaa	94
b6f9676e JA	95	return 1;
b6f9676e JA	96	}
2f9ade3c	97
b6f9676e	98	td->orig_buffer = shmat(td->shm_id, NULL, 0);
ee56ad50	99	dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
b6f9676e JA	100	if (td->orig_buffer == (void *) -1) {
	101	td_verror(td, errno, "shmat");
	102	td->orig_buffer = NULL;
	103	return 1;
	104	}
	105
	106	return 0;
	107	}
	108
829a602c JA	109	static void free_mem_shm(struct thread_data *td)
	110	{
	111	struct shmid_ds sbuf;
	112
	113	dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id, td->orig_buffer);
	114	shmdt(td->orig_buffer);
	115	shmctl(td->shm_id, IPC_RMID, &sbuf);
	116	}
	117
0f805c00	118	static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	119	{
d9759b1e	120	int flags = 0;
b6f9676e	121
a55820db	122	td->mmapfd = 1;
b6f9676e	123
d6dc02fb JA	124	if (td->o.mem_type == MEM_MMAPHUGE) {
	125	unsigned long mask = td->o.hugepage_size - 1;
	126
d9759b1e	127	/* TODO: make sure the file is a real hugetlbfs file */
836fcc0f	128	if (!td->o.mmapfile)
d9759b1e	129	flags \|= MAP_HUGETLB;
d6dc02fb JA	130	total_mem = (total_mem + mask) & ~mask;
	131	}
	132
83ea422a JA	133	if (td->o.mmapfile) {
83ea422a JA	134	td->mmapfd = open(td->o.mmapfile, O_RDWR\|O_CREAT, 0644);
b6f9676e JA	135
	136	if (td->mmapfd < 0) {
	137	td_verror(td, errno, "open mmap file");
2f9ade3c JA	138	td->orig_buffer = NULL;
	139	return 1;
	140	}
d9759b1e SL	141	if (td->o.mem_type != MEM_MMAPHUGE &&
d9759b1e SL	142	ftruncate(td->mmapfd, total_mem) < 0) {
b6f9676e	143	td_verror(td, errno, "truncate mmap file");
2f9ade3c JA	144	td->orig_buffer = NULL;
	145	return 1;
	146	}
d9759b1e SL	147	if (td->o.mem_type == MEM_MMAPHUGE)
	148	flags \|= MAP_SHARED;
	149	else
	150	flags \|= MAP_PRIVATE;
b6f9676e	151	} else
d9759b1e	152	flags \|= OS_MAP_ANON \| MAP_PRIVATE;
b6f9676e	153
829a602c JA	154	td->orig_buffer = mmap(NULL, total_mem, PROT_READ \| PROT_WRITE, flags,
	155	td->mmapfd, 0);
	156	dprint(FD_MEM, "mmap %u/%d %p\n", total_mem, td->mmapfd,
	157	td->orig_buffer);
b6f9676e JA	158	if (td->orig_buffer == MAP_FAILED) {
	159	td_verror(td, errno, "mmap");
	160	td->orig_buffer = NULL;
	161	if (td->mmapfd) {
	162	close(td->mmapfd);
83ea422a	163	unlink(td->o.mmapfile);
b6f9676e	164	}
5ec10eaa	165
b6f9676e	166	return 1;
2f9ade3c JA	167	}
	168
	169	return 0;
	170	}
	171
0f805c00	172	static void free_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	173	{
829a602c JA	174	dprint(FD_MEM, "munmap %u %p\n", total_mem, td->orig_buffer);
829a602c JA	175	munmap(td->orig_buffer, td->orig_buffer_size);
83ea422a	176	if (td->o.mmapfile) {
829a602c	177	close(td->mmapfd);
83ea422a JA	178	unlink(td->o.mmapfile);
83ea422a JA	179	free(td->o.mmapfile);
829a602c JA	180	}
829a602c JA	181	}
d87612ac	182
0f805c00	183	static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
829a602c JA	184	{
	185	td->orig_buffer = malloc(total_mem);
	186	dprint(FD_MEM, "malloc %u %p\n", total_mem, td->orig_buffer);
5ec10eaa	187
829a602c JA	188	return td->orig_buffer == NULL;
829a602c JA	189	}
b6f9676e	190
829a602c JA	191	static void free_mem_malloc(struct thread_data *td)
	192	{
	193	dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
	194	free(td->orig_buffer);
b6f9676e JA	195	}
	196
	197	/*
03e20d68	198	* Set up the buffer area we need for io.
b6f9676e JA	199	*/
	200	int allocate_io_mem(struct thread_data *td)
	201	{
0f805c00	202	size_t total_mem;
b6f9676e JA	203	int ret = 0;
b6f9676e JA	204
b4c5e1ac JA	205	if (td->io_ops->flags & FIO_NOIO)
	206	return 0;
	207
829a602c	208	total_mem = td->orig_buffer_size;
d529ee19	209
ca7e0ddb JA	210	if (td->o.odirect \|\| td->o.mem_align \|\|
ca7e0ddb JA	211	(td->io_ops->flags & FIO_MEMALIGN)) {
829a602c	212	total_mem += page_mask;
d529ee19 JA	213	if (td->o.mem_align && td->o.mem_align > page_size)
	214	total_mem += td->o.mem_align - page_size;
	215	}
829a602c	216
0f805c00 JA	217	dprint(FD_MEM, "Alloc %lu for buffers\n", (size_t) total_mem);
0f805c00 JA	218
b6f9676e	219	if (td->o.mem_type == MEM_MALLOC)
829a602c	220	ret = alloc_mem_malloc(td, total_mem);
b6f9676e	221	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
829a602c	222	ret = alloc_mem_shm(td, total_mem);
b6f9676e	223	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
829a602c	224	ret = alloc_mem_mmap(td, total_mem);
b6f9676e JA	225	else {
	226	log_err("fio: bad mem type: %d\n", td->o.mem_type);
	227	ret = 1;
	228	}
	229
3deb3101 JA	230	if (ret)
	231	td_verror(td, ENOMEM, "iomem allocation");
	232
b6f9676e JA	233	return ret;
	234	}
	235
2f9ade3c JA	236	void free_io_mem(struct thread_data *td)
2f9ade3c JA	237	{
829a602c JA	238	unsigned int total_mem;
	239
	240	total_mem = td->orig_buffer_size;
	241	if (td->o.odirect)
	242	total_mem += page_mask;
	243
	244	if (td->o.mem_type == MEM_MALLOC)
	245	free_mem_malloc(td);
	246	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	247	free_mem_shm(td);
	248	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
	249	free_mem_mmap(td, total_mem);
	250	else
2dc1bbeb	251	log_err("Bad memory type %u\n", td->o.mem_type);
2f9ade3c JA	252
2f9ade3c JA	253	td->orig_buffer = NULL;
829a602c	254	td->orig_buffer_size = 0;
2f9ade3c	255	}