[fio.git] / memory.c

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

#include "fio.h"
#ifndef FIO_NO_HAVE_SHM_H
#include <sys/shm.h>
#endif

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)
{
#ifndef CONFIG_NO_SHM
	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;
#else
	log_err("fio: shm not supported\n");
	return 1;
#endif
}

static void free_mem_shm(struct thread_data *td)
{
#ifndef CONFIG_NO_SHM
	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);
#endif
}

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 %llu/%d %p\n", (unsigned long long) 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 != 1 && td->mmapfd != -1) {
			close(td->mmapfd);
			if (td->o.mmapfile)
				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 %llu %p\n", (unsigned long long) total_mem,
						td->orig_buffer);
	munmap(td->orig_buffer, td->orig_buffer_size);
	if (td->o.mmapfile) {
		if (td->mmapfd != -1)
			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 %llu %p\n", (unsigned long long) 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->o.oatomic ||
	    (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 %llu for buffers\n", (unsigned long long) 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 || td->o.oatomic)
		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>
2f9ade3c JA	8	#include <sys/mman.h>
	9
	10	#include "fio.h"
a5e0ee11 O	11	#ifndef FIO_NO_HAVE_SHM_H
	12	#include <sys/shm.h>
	13	#endif
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	{
91e47529	66	#ifndef CONFIG_NO_SHM
03e20d68	67	int flags = IPC_CREAT \| S_IRUSR \| S_IWUSR;
b6f9676e	68
a1242a20 JA	69	if (td->o.mem_type == MEM_SHMHUGE) {
	70	unsigned long mask = td->o.hugepage_size - 1;
	71
b6f9676e	72	flags \|= SHM_HUGETLB;
a1242a20 JA	73	total_mem = (total_mem + mask) & ~mask;
a1242a20 JA	74	}
b6f9676e	75
829a602c JA	76	td->shm_id = shmget(IPC_PRIVATE, total_mem, flags);
829a602c JA	77	dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id);
b6f9676e JA	78	if (td->shm_id < 0) {
b6f9676e JA	79	td_verror(td, errno, "shmget");
da7d79b0	80	if (geteuid() != 0 && (errno == ENOMEM \|\| errno == EPERM))
b6f9676e	81	log_err("fio: you may need to run this job as root\n");
886b878a	82	if (td->o.mem_type == MEM_SHMHUGE) {
5ec10eaa JA	83	if (errno == EINVAL) {
	84	log_err("fio: check that you have free huge"
	85	" pages and that hugepage-size is"
	86	" correct.\n");
	87	} else if (errno == ENOSYS) {
	88	log_err("fio: your system does not appear to"
	89	" support huge pages.\n");
	90	} else if (errno == ENOMEM) {
	91	log_err("fio: no huge pages available, do you"
	92	" need to alocate some? See HOWTO.\n");
	93	}
d8602dd0	94	}
5ec10eaa	95
b6f9676e JA	96	return 1;
b6f9676e JA	97	}
2f9ade3c	98
b6f9676e	99	td->orig_buffer = shmat(td->shm_id, NULL, 0);
ee56ad50	100	dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
b6f9676e JA	101	if (td->orig_buffer == (void *) -1) {
	102	td_verror(td, errno, "shmat");
	103	td->orig_buffer = NULL;
	104	return 1;
	105	}
	106
	107	return 0;
91e47529 JA	108	#else
	109	log_err("fio: shm not supported\n");
	110	return 1;
	111	#endif
b6f9676e JA	112	}
b6f9676e JA	113
829a602c JA	114	static void free_mem_shm(struct thread_data *td)
829a602c JA	115	{
91e47529	116	#ifndef CONFIG_NO_SHM
829a602c JA	117	struct shmid_ds sbuf;
	118
	119	dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id, td->orig_buffer);
	120	shmdt(td->orig_buffer);
	121	shmctl(td->shm_id, IPC_RMID, &sbuf);
91e47529	122	#endif
829a602c JA	123	}
829a602c JA	124
0f805c00	125	static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	126	{
d9759b1e	127	int flags = 0;
b6f9676e	128
4a995dda	129	td->mmapfd = -1;
b6f9676e	130
d6dc02fb JA	131	if (td->o.mem_type == MEM_MMAPHUGE) {
	132	unsigned long mask = td->o.hugepage_size - 1;
	133
d9759b1e	134	/* TODO: make sure the file is a real hugetlbfs file */
836fcc0f	135	if (!td->o.mmapfile)
d9759b1e	136	flags \|= MAP_HUGETLB;
d6dc02fb JA	137	total_mem = (total_mem + mask) & ~mask;
	138	}
	139
83ea422a JA	140	if (td->o.mmapfile) {
83ea422a JA	141	td->mmapfd = open(td->o.mmapfile, O_RDWR\|O_CREAT, 0644);
b6f9676e JA	142
	143	if (td->mmapfd < 0) {
	144	td_verror(td, errno, "open mmap file");
2f9ade3c JA	145	td->orig_buffer = NULL;
	146	return 1;
	147	}
d9759b1e SL	148	if (td->o.mem_type != MEM_MMAPHUGE &&
d9759b1e SL	149	ftruncate(td->mmapfd, total_mem) < 0) {
b6f9676e	150	td_verror(td, errno, "truncate mmap file");
2f9ade3c JA	151	td->orig_buffer = NULL;
	152	return 1;
	153	}
d9759b1e SL	154	if (td->o.mem_type == MEM_MMAPHUGE)
	155	flags \|= MAP_SHARED;
	156	else
	157	flags \|= MAP_PRIVATE;
b6f9676e	158	} else
d9759b1e	159	flags \|= OS_MAP_ANON \| MAP_PRIVATE;
b6f9676e	160
829a602c JA	161	td->orig_buffer = mmap(NULL, total_mem, PROT_READ \| PROT_WRITE, flags,
829a602c JA	162	td->mmapfd, 0);
4b91ee8f JA	163	dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem,
4b91ee8f JA	164	td->mmapfd, td->orig_buffer);
b6f9676e JA	165	if (td->orig_buffer == MAP_FAILED) {
	166	td_verror(td, errno, "mmap");
	167	td->orig_buffer = NULL;
9ce94349	168	if (td->mmapfd != 1 && td->mmapfd != -1) {
b6f9676e	169	close(td->mmapfd);
b3493a7a JA	170	if (td->o.mmapfile)
b3493a7a JA	171	unlink(td->o.mmapfile);
b6f9676e	172	}
5ec10eaa	173
b6f9676e	174	return 1;
2f9ade3c JA	175	}
	176
	177	return 0;
	178	}
	179
0f805c00	180	static void free_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	181	{
4b91ee8f JA	182	dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem,
4b91ee8f JA	183	td->orig_buffer);
829a602c	184	munmap(td->orig_buffer, td->orig_buffer_size);
83ea422a	185	if (td->o.mmapfile) {
4a995dda JA	186	if (td->mmapfd != -1)
4a995dda JA	187	close(td->mmapfd);
83ea422a JA	188	unlink(td->o.mmapfile);
83ea422a JA	189	free(td->o.mmapfile);
829a602c JA	190	}
829a602c JA	191	}
d87612ac	192
0f805c00	193	static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
829a602c JA	194	{
829a602c JA	195	td->orig_buffer = malloc(total_mem);
4b91ee8f JA	196	dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
4b91ee8f JA	197	td->orig_buffer);
5ec10eaa	198
829a602c JA	199	return td->orig_buffer == NULL;
829a602c JA	200	}
b6f9676e	201
829a602c JA	202	static void free_mem_malloc(struct thread_data *td)
	203	{
	204	dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
	205	free(td->orig_buffer);
b6f9676e JA	206	}
	207
	208	/*
03e20d68	209	* Set up the buffer area we need for io.
b6f9676e JA	210	*/
	211	int allocate_io_mem(struct thread_data *td)
	212	{
0f805c00	213	size_t total_mem;
b6f9676e JA	214	int ret = 0;
b6f9676e JA	215
b4c5e1ac JA	216	if (td->io_ops->flags & FIO_NOIO)
	217	return 0;
	218
829a602c	219	total_mem = td->orig_buffer_size;
d529ee19	220
d01612f3	221	if (td->o.odirect \|\| td->o.mem_align \|\| td->o.oatomic \|\|
ca7e0ddb	222	(td->io_ops->flags & FIO_MEMALIGN)) {
829a602c	223	total_mem += page_mask;
d529ee19 JA	224	if (td->o.mem_align && td->o.mem_align > page_size)
	225	total_mem += td->o.mem_align - page_size;
	226	}
829a602c	227
4b91ee8f	228	dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
0f805c00	229
b6f9676e	230	if (td->o.mem_type == MEM_MALLOC)
829a602c	231	ret = alloc_mem_malloc(td, total_mem);
b6f9676e	232	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
829a602c	233	ret = alloc_mem_shm(td, total_mem);
b6f9676e	234	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
829a602c	235	ret = alloc_mem_mmap(td, total_mem);
b6f9676e JA	236	else {
	237	log_err("fio: bad mem type: %d\n", td->o.mem_type);
	238	ret = 1;
	239	}
	240
3deb3101 JA	241	if (ret)
	242	td_verror(td, ENOMEM, "iomem allocation");
	243
b6f9676e JA	244	return ret;
	245	}
	246
2f9ade3c JA	247	void free_io_mem(struct thread_data *td)
2f9ade3c JA	248	{
829a602c JA	249	unsigned int total_mem;
	250
	251	total_mem = td->orig_buffer_size;
d01612f3	252	if (td->o.odirect \|\| td->o.oatomic)
829a602c JA	253	total_mem += page_mask;
	254
	255	if (td->o.mem_type == MEM_MALLOC)
	256	free_mem_malloc(td);
	257	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	258	free_mem_shm(td);
	259	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
	260	free_mem_mmap(td, total_mem);
	261	else
2dc1bbeb	262	log_err("Bad memory type %u\n", td->o.mem_type);
2f9ade3c JA	263
2f9ade3c JA	264	td->orig_buffer = NULL;
829a602c	265	td->orig_buffer_size = 0;
2f9ade3c	266	}