[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)
{
	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 %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) {
			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
	1	/*
	2	* Memory helpers
	3	*/
	4	#include <sys/types.h>
	5	#include <sys/stat.h>
	6	#include <fcntl.h>
	7	#include <unistd.h>
	8	#include <sys/mman.h>
	9
	10	#include "fio.h"
	11	#ifndef FIO_NO_HAVE_SHM_H
	12	#include <sys/shm.h>
	13	#endif
	14
	15	void fio_unpin_memory(struct thread_data *td)
	16	{
	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)
	20	perror("munlock");
	21	munmap(td->pinned_mem, td->o.lockmem);
	22	td->pinned_mem = NULL;
	23	}
	24	}
	25
	26	int fio_pin_memory(struct thread_data *td)
	27	{
	28	unsigned long long phys_mem;
	29
	30	if (!td->o.lockmem)
	31	return 0;
	32
	33	dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
	34
	35	/*
	36	* Don't allow mlock of more than real_mem-128MB
	37	*/
	38	phys_mem = os_phys_mem();
	39	if (phys_mem) {
	40	if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) {
	41	td->o.lockmem = phys_mem - 128 * 1024 * 1024;
	42	log_info("fio: limiting mlocked memory to %lluMB\n",
	43	td->o.lockmem >> 20);
	44	}
	45	}
	46
	47	td->pinned_mem = mmap(NULL, td->o.lockmem, PROT_READ \| PROT_WRITE,
	48	MAP_PRIVATE \| OS_MAP_ANON, -1, 0);
	49	if (td->pinned_mem == MAP_FAILED) {
	50	perror("malloc locked mem");
	51	td->pinned_mem = NULL;
	52	return 1;
	53	}
	54	if (mlock(td->pinned_mem, td->o.lockmem) < 0) {
	55	perror("mlock");
	56	munmap(td->pinned_mem, td->o.lockmem);
	57	td->pinned_mem = NULL;
	58	return 1;
	59	}
	60
	61	return 0;
	62	}
	63
	64	static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
	65	{
	66	int flags = IPC_CREAT \| S_IRUSR \| S_IWUSR;
	67
	68	if (td->o.mem_type == MEM_SHMHUGE) {
	69	unsigned long mask = td->o.hugepage_size - 1;
	70
	71	flags \|= SHM_HUGETLB;
	72	total_mem = (total_mem + mask) & ~mask;
	73	}
	74
	75	td->shm_id = shmget(IPC_PRIVATE, total_mem, flags);
	76	dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id);
	77	if (td->shm_id < 0) {
	78	td_verror(td, errno, "shmget");
	79	if (geteuid() != 0 && (errno == ENOMEM \|\| errno == EPERM))
	80	log_err("fio: you may need to run this job as root\n");
	81	if (td->o.mem_type == MEM_SHMHUGE) {
	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	}
	93	}
	94
	95	return 1;
	96	}
	97
	98	td->orig_buffer = shmat(td->shm_id, NULL, 0);
	99	dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer);
	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
	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
	118	static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
	119	{
	120	int flags = 0;
	121
	122	td->mmapfd = -1;
	123
	124	if (td->o.mem_type == MEM_MMAPHUGE) {
	125	unsigned long mask = td->o.hugepage_size - 1;
	126
	127	/* TODO: make sure the file is a real hugetlbfs file */
	128	if (!td->o.mmapfile)
	129	flags \|= MAP_HUGETLB;
	130	total_mem = (total_mem + mask) & ~mask;
	131	}
	132
	133	if (td->o.mmapfile) {
	134	td->mmapfd = open(td->o.mmapfile, O_RDWR\|O_CREAT, 0644);
	135
	136	if (td->mmapfd < 0) {
	137	td_verror(td, errno, "open mmap file");
	138	td->orig_buffer = NULL;
	139	return 1;
	140	}
	141	if (td->o.mem_type != MEM_MMAPHUGE &&
	142	ftruncate(td->mmapfd, total_mem) < 0) {
	143	td_verror(td, errno, "truncate mmap file");
	144	td->orig_buffer = NULL;
	145	return 1;
	146	}
	147	if (td->o.mem_type == MEM_MMAPHUGE)
	148	flags \|= MAP_SHARED;
	149	else
	150	flags \|= MAP_PRIVATE;
	151	} else
	152	flags \|= OS_MAP_ANON \| MAP_PRIVATE;
	153
	154	td->orig_buffer = mmap(NULL, total_mem, PROT_READ \| PROT_WRITE, flags,
	155	td->mmapfd, 0);
	156	dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem,
	157	td->mmapfd, td->orig_buffer);
	158	if (td->orig_buffer == MAP_FAILED) {
	159	td_verror(td, errno, "mmap");
	160	td->orig_buffer = NULL;
	161	if (td->mmapfd != 1) {
	162	close(td->mmapfd);
	163	if (td->o.mmapfile)
	164	unlink(td->o.mmapfile);
	165	}
	166
	167	return 1;
	168	}
	169
	170	return 0;
	171	}
	172
	173	static void free_mem_mmap(struct thread_data *td, size_t total_mem)
	174	{
	175	dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem,
	176	td->orig_buffer);
	177	munmap(td->orig_buffer, td->orig_buffer_size);
	178	if (td->o.mmapfile) {
	179	if (td->mmapfd != -1)
	180	close(td->mmapfd);
	181	unlink(td->o.mmapfile);
	182	free(td->o.mmapfile);
	183	}
	184	}
	185
	186	static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
	187	{
	188	td->orig_buffer = malloc(total_mem);
	189	dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
	190	td->orig_buffer);
	191
	192	return td->orig_buffer == NULL;
	193	}
	194
	195	static void free_mem_malloc(struct thread_data *td)
	196	{
	197	dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
	198	free(td->orig_buffer);
	199	}
	200
	201	/*
	202	* Set up the buffer area we need for io.
	203	*/
	204	int allocate_io_mem(struct thread_data *td)
	205	{
	206	size_t total_mem;
	207	int ret = 0;
	208
	209	if (td->io_ops->flags & FIO_NOIO)
	210	return 0;
	211
	212	total_mem = td->orig_buffer_size;
	213
	214	if (td->o.odirect \|\| td->o.mem_align \|\| td->o.oatomic \|\|
	215	(td->io_ops->flags & FIO_MEMALIGN)) {
	216	total_mem += page_mask;
	217	if (td->o.mem_align && td->o.mem_align > page_size)
	218	total_mem += td->o.mem_align - page_size;
	219	}
	220
	221	dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
	222
	223	if (td->o.mem_type == MEM_MALLOC)
	224	ret = alloc_mem_malloc(td, total_mem);
	225	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	226	ret = alloc_mem_shm(td, total_mem);
	227	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
	228	ret = alloc_mem_mmap(td, total_mem);
	229	else {
	230	log_err("fio: bad mem type: %d\n", td->o.mem_type);
	231	ret = 1;
	232	}
	233
	234	if (ret)
	235	td_verror(td, ENOMEM, "iomem allocation");
	236
	237	return ret;
	238	}
	239
	240	void free_io_mem(struct thread_data *td)
	241	{
	242	unsigned int total_mem;
	243
	244	total_mem = td->orig_buffer_size;
	245	if (td->o.odirect \|\| td->o.oatomic)
	246	total_mem += page_mask;
	247
	248	if (td->o.mem_type == MEM_MALLOC)
	249	free_mem_malloc(td);
	250	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	251	free_mem_shm(td);
	252	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE)
	253	free_mem_mmap(td, total_mem);
	254	else
	255	log_err("Bad memory type %u\n", td->o.mem_type);
	256
	257	td->orig_buffer = NULL;
	258	td->orig_buffer_size = 0;
	259	}