[fio.git] / memory.c

/*
 * Memory helpers
 */
#include <fcntl.h>
#include <unistd.h>
#include <sys/mman.h>
#include <sys/stat.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-128MiB
	 */
	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 %lluMiB\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 allocate 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) {
		if (access(td->o.mmapfile, F_OK) == 0)
			td->flags |= TD_F_MMAP_KEEP;

		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 &&
		    td->o.mem_type != MEM_MMAPSHARED &&
		    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 ||
		    td->o.mem_type == MEM_MMAPSHARED)
			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 && !(td->flags & TD_F_MMAP_KEEP))
				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);
		if (!(td->flags & TD_F_MMAP_KEEP))
			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);
}

static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
{
#ifdef CONFIG_CUDA
	CUresult ret;
	char name[128];

	ret = cuInit(0);
	if (ret != CUDA_SUCCESS) {
		log_err("fio: failed initialize cuda driver api\n");
		return 1;
	}

	ret = cuDeviceGetCount(&td->gpu_dev_cnt);
	if (ret != CUDA_SUCCESS) {
		log_err("fio: failed get device count\n");
		return 1;
	}
	dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt);

	if (td->gpu_dev_cnt == 0) {
		log_err("fio: no GPU device found. "
			"Can not perform GPUDirect RDMA.\n");
		return 1;
	}

	td->gpu_dev_id = td->o.gpu_dev_id;
	ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id);
	if (ret != CUDA_SUCCESS) {
		log_err("fio: failed get GPU device\n");
		return 1;
	}

	ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id);
	if (ret != CUDA_SUCCESS) {
		log_err("fio: failed get device name\n");
		return 1;
	}
	dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \
	       td->gpu_dev_id, name);

	ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev);
	if (ret != CUDA_SUCCESS) {
		log_err("fio: failed to create cuda context: %d\n", ret);
		return 1;
	}

	ret = cuMemAlloc(&td->dev_mem_ptr, total_mem);
	if (ret != CUDA_SUCCESS) {
		log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem);
		return 1;
	}
	td->orig_buffer = (void *) td->dev_mem_ptr;

	dprint(FD_MEM, "cudaMalloc %llu %p\n",				\
	       (unsigned long long) total_mem, td->orig_buffer);
	return 0;
#else
	return -EINVAL;
#endif
}

static void free_mem_cudamalloc(struct thread_data *td)
{
#ifdef CONFIG_CUDA
	if (td->dev_mem_ptr != NULL)
		cuMemFree(td->dev_mem_ptr);

	if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS)
		log_err("fio: failed to destroy cuda context\n");
#endif
}

/*
 * 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_ioengine_flagged(td, FIO_NOIO))
		return 0;

	total_mem = td->orig_buffer_size;

	if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
	    td_ioengine_flagged(td, 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 the IO engine has hooks to allocate/free memory, use those. But
	 * error out if the user explicitly asked for something else.
	 */
	if (td->io_ops->iomem_alloc) {
		if (fio_option_is_set(&td->o, mem_type)) {
			log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
			ret = 1;
		} else
			ret = td->io_ops->iomem_alloc(td, total_mem);
	} else 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 ||
		 td->o.mem_type == MEM_MMAPSHARED)
		ret = alloc_mem_mmap(td, total_mem);
	else if (td->o.mem_type == MEM_CUDA_MALLOC)
		ret = alloc_mem_cudamalloc(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->io_ops->iomem_alloc) {
		if (td->io_ops->iomem_free)
			td->io_ops->iomem_free(td);
	} else 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 ||
		 td->o.mem_type == MEM_MMAPSHARED)
		free_mem_mmap(td, total_mem);
	else if (td->o.mem_type == MEM_CUDA_MALLOC)
		free_mem_cudamalloc(td);
	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	4	#include <fcntl.h>
2f9ade3c	5	#include <unistd.h>
2f9ade3c	6	#include <sys/mman.h>
3d2d14bc	7	#include <sys/stat.h>
2f9ade3c JA	8
2f9ade3c JA	9	#include "fio.h"
a5e0ee11 O	10	#ifndef FIO_NO_HAVE_SHM_H
	11	#include <sys/shm.h>
	12	#endif
2f9ade3c	13
1b79a070	14	void fio_unpin_memory(struct thread_data *td)
2f9ade3c	15	{
1b79a070 JA	16	if (td->pinned_mem) {
	17	dprint(FD_MEM, "unpinning %llu bytes\n", td->o.lockmem);
	18	if (munlock(td->pinned_mem, td->o.lockmem) < 0)
2f9ade3c	19	perror("munlock");
1b79a070 JA	20	munmap(td->pinned_mem, td->o.lockmem);
1b79a070 JA	21	td->pinned_mem = NULL;
2f9ade3c JA	22	}
	23	}
	24
1b79a070	25	int fio_pin_memory(struct thread_data *td)
2f9ade3c JA	26	{
	27	unsigned long long phys_mem;
	28
1b79a070	29	if (!td->o.lockmem)
2f9ade3c JA	30	return 0;
2f9ade3c JA	31
1b79a070	32	dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
ee56ad50	33
2f9ade3c	34	/*
420b104a	35	* Don't allow mlock of more than real_mem-128MiB
2f9ade3c JA	36	*/
	37	phys_mem = os_phys_mem();
	38	if (phys_mem) {
1b79a070 JA	39	if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) {
1b79a070 JA	40	td->o.lockmem = phys_mem - 128 * 1024 * 1024;
420b104a	41	log_info("fio: limiting mlocked memory to %lluMiB\n",
1b79a070	42	td->o.lockmem >> 20);
2f9ade3c JA	43	}
	44	}
	45
1b79a070	46	td->pinned_mem = mmap(NULL, td->o.lockmem, PROT_READ \| PROT_WRITE,
a55820db	47	MAP_PRIVATE \| OS_MAP_ANON, -1, 0);
1b79a070	48	if (td->pinned_mem == MAP_FAILED) {
2f9ade3c	49	perror("malloc locked mem");
1b79a070	50	td->pinned_mem = NULL;
2f9ade3c JA	51	return 1;
2f9ade3c JA	52	}
1b79a070	53	if (mlock(td->pinned_mem, td->o.lockmem) < 0) {
2f9ade3c	54	perror("mlock");
1b79a070 JA	55	munmap(td->pinned_mem, td->o.lockmem);
1b79a070 JA	56	td->pinned_mem = NULL;
2f9ade3c JA	57	return 1;
	58	}
	59
	60	return 0;
	61	}
	62
829a602c	63	static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
2f9ade3c	64	{
91e47529	65	#ifndef CONFIG_NO_SHM
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"
dd117eb6	91	" need to allocate some? See HOWTO.\n");
5ec10eaa	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;
91e47529 JA	107	#else
	108	log_err("fio: shm not supported\n");
	109	return 1;
	110	#endif
b6f9676e JA	111	}
b6f9676e JA	112
829a602c JA	113	static void free_mem_shm(struct thread_data *td)
829a602c JA	114	{
91e47529	115	#ifndef CONFIG_NO_SHM
829a602c JA	116	struct shmid_ds sbuf;
	117
	118	dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id, td->orig_buffer);
	119	shmdt(td->orig_buffer);
	120	shmctl(td->shm_id, IPC_RMID, &sbuf);
91e47529	121	#endif
829a602c JA	122	}
829a602c JA	123
0f805c00	124	static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	125	{
d9759b1e	126	int flags = 0;
b6f9676e	127
4a995dda	128	td->mmapfd = -1;
b6f9676e	129
d6dc02fb JA	130	if (td->o.mem_type == MEM_MMAPHUGE) {
	131	unsigned long mask = td->o.hugepage_size - 1;
	132
d9759b1e	133	/* TODO: make sure the file is a real hugetlbfs file */
836fcc0f	134	if (!td->o.mmapfile)
d9759b1e	135	flags \|= MAP_HUGETLB;
d6dc02fb JA	136	total_mem = (total_mem + mask) & ~mask;
	137	}
	138
83ea422a	139	if (td->o.mmapfile) {
f9cfc7d4 SB	140	if (access(td->o.mmapfile, F_OK) == 0)
	141	td->flags \|= TD_F_MMAP_KEEP;
	142
83ea422a	143	td->mmapfd = open(td->o.mmapfile, O_RDWR\|O_CREAT, 0644);
b6f9676e JA	144
	145	if (td->mmapfd < 0) {
	146	td_verror(td, errno, "open mmap file");
2f9ade3c JA	147	td->orig_buffer = NULL;
	148	return 1;
	149	}
d9759b1e	150	if (td->o.mem_type != MEM_MMAPHUGE &&
217b0f1d	151	td->o.mem_type != MEM_MMAPSHARED &&
d9759b1e	152	ftruncate(td->mmapfd, total_mem) < 0) {
b6f9676e	153	td_verror(td, errno, "truncate mmap file");
2f9ade3c JA	154	td->orig_buffer = NULL;
	155	return 1;
	156	}
217b0f1d LG	157	if (td->o.mem_type == MEM_MMAPHUGE \|\|
217b0f1d LG	158	td->o.mem_type == MEM_MMAPSHARED)
d9759b1e SL	159	flags \|= MAP_SHARED;
	160	else
	161	flags \|= MAP_PRIVATE;
b6f9676e	162	} else
d9759b1e	163	flags \|= OS_MAP_ANON \| MAP_PRIVATE;
b6f9676e	164
829a602c JA	165	td->orig_buffer = mmap(NULL, total_mem, PROT_READ \| PROT_WRITE, flags,
829a602c JA	166	td->mmapfd, 0);
4b91ee8f JA	167	dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem,
4b91ee8f JA	168	td->mmapfd, td->orig_buffer);
b6f9676e JA	169	if (td->orig_buffer == MAP_FAILED) {
	170	td_verror(td, errno, "mmap");
	171	td->orig_buffer = NULL;
9ce94349	172	if (td->mmapfd != 1 && td->mmapfd != -1) {
b6f9676e	173	close(td->mmapfd);
f9cfc7d4	174	if (td->o.mmapfile && !(td->flags & TD_F_MMAP_KEEP))
b3493a7a	175	unlink(td->o.mmapfile);
b6f9676e	176	}
5ec10eaa	177
b6f9676e	178	return 1;
2f9ade3c JA	179	}
	180
	181	return 0;
	182	}
	183
0f805c00	184	static void free_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	185	{
4b91ee8f JA	186	dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem,
4b91ee8f JA	187	td->orig_buffer);
829a602c	188	munmap(td->orig_buffer, td->orig_buffer_size);
83ea422a	189	if (td->o.mmapfile) {
4a995dda JA	190	if (td->mmapfd != -1)
4a995dda JA	191	close(td->mmapfd);
f9cfc7d4 SB	192	if (!(td->flags & TD_F_MMAP_KEEP))
f9cfc7d4 SB	193	unlink(td->o.mmapfile);
83ea422a	194	free(td->o.mmapfile);
829a602c JA	195	}
829a602c JA	196	}
d87612ac	197
0f805c00	198	static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
829a602c JA	199	{
829a602c JA	200	td->orig_buffer = malloc(total_mem);
4b91ee8f JA	201	dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
4b91ee8f JA	202	td->orig_buffer);
5ec10eaa	203
829a602c JA	204	return td->orig_buffer == NULL;
829a602c JA	205	}
b6f9676e	206
829a602c JA	207	static void free_mem_malloc(struct thread_data *td)
	208	{
	209	dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
	210	free(td->orig_buffer);
b6f9676e JA	211	}
b6f9676e JA	212
03553853 YR	213	static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
03553853 YR	214	{
15600335	215	#ifdef CONFIG_CUDA
03553853 YR	216	CUresult ret;
	217	char name[128];
	218
	219	ret = cuInit(0);
	220	if (ret != CUDA_SUCCESS) {
	221	log_err("fio: failed initialize cuda driver api\n");
	222	return 1;
	223	}
	224
	225	ret = cuDeviceGetCount(&td->gpu_dev_cnt);
	226	if (ret != CUDA_SUCCESS) {
	227	log_err("fio: failed get device count\n");
	228	return 1;
	229	}
	230	dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt);
	231
	232	if (td->gpu_dev_cnt == 0) {
	233	log_err("fio: no GPU device found. "
	234	"Can not perform GPUDirect RDMA.\n");
	235	return 1;
	236	}
	237
	238	td->gpu_dev_id = td->o.gpu_dev_id;
	239	ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id);
	240	if (ret != CUDA_SUCCESS) {
	241	log_err("fio: failed get GPU device\n");
	242	return 1;
	243	}
	244
	245	ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id);
	246	if (ret != CUDA_SUCCESS) {
	247	log_err("fio: failed get device name\n");
	248	return 1;
	249	}
	250	dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \
	251	td->gpu_dev_id, name);
	252
	253	ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev);
	254	if (ret != CUDA_SUCCESS) {
	255	log_err("fio: failed to create cuda context: %d\n", ret);
	256	return 1;
	257	}
	258
	259	ret = cuMemAlloc(&td->dev_mem_ptr, total_mem);
	260	if (ret != CUDA_SUCCESS) {
	261	log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem);
	262	return 1;
	263	}
	264	td->orig_buffer = (void *) td->dev_mem_ptr;
	265
	266	dprint(FD_MEM, "cudaMalloc %llu %p\n", \
	267	(unsigned long long) total_mem, td->orig_buffer);
	268	return 0;
15600335 JA	269	#else
	270	return -EINVAL;
	271	#endif
03553853 YR	272	}
	273
	274	static void free_mem_cudamalloc(struct thread_data *td)
	275	{
15600335 JA	276	#ifdef CONFIG_CUDA
15600335 JA	277	if (td->dev_mem_ptr != NULL)
03553853 YR	278	cuMemFree(td->dev_mem_ptr);
	279
	280	if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS)
	281	log_err("fio: failed to destroy cuda context\n");
03553853	282	#endif
15600335	283	}
03553853	284
b6f9676e	285	/*
03e20d68	286	* Set up the buffer area we need for io.
b6f9676e JA	287	*/
	288	int allocate_io_mem(struct thread_data *td)
	289	{
0f805c00	290	size_t total_mem;
b6f9676e JA	291	int ret = 0;
b6f9676e JA	292
9b87f09b	293	if (td_ioengine_flagged(td, FIO_NOIO))
b4c5e1ac JA	294	return 0;
b4c5e1ac JA	295
829a602c	296	total_mem = td->orig_buffer_size;
d529ee19	297
d01612f3	298	if (td->o.odirect \|\| td->o.mem_align \|\| td->o.oatomic \|\|
9b87f09b	299	td_ioengine_flagged(td, FIO_MEMALIGN)) {
829a602c	300	total_mem += page_mask;
d529ee19 JA	301	if (td->o.mem_align && td->o.mem_align > page_size)
	302	total_mem += td->o.mem_align - page_size;
	303	}
829a602c	304
4b91ee8f	305	dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
0f805c00	306
a6cb85e2 JA	307	/*
	308	* If the IO engine has hooks to allocate/free memory, use those. But
	309	* error out if the user explicitly asked for something else.
	310	*/
	311	if (td->io_ops->iomem_alloc) {
	312	if (fio_option_is_set(&td->o, mem_type)) {
	313	log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
	314	ret = 1;
	315	} else
	316	ret = td->io_ops->iomem_alloc(td, total_mem);
	317	} else if (td->o.mem_type == MEM_MALLOC)
829a602c	318	ret = alloc_mem_malloc(td, total_mem);
b6f9676e	319	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
829a602c	320	ret = alloc_mem_shm(td, total_mem);
217b0f1d LG	321	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE \|\|
217b0f1d LG	322	td->o.mem_type == MEM_MMAPSHARED)
829a602c	323	ret = alloc_mem_mmap(td, total_mem);
03553853 YR	324	else if (td->o.mem_type == MEM_CUDA_MALLOC)
03553853 YR	325	ret = alloc_mem_cudamalloc(td, total_mem);
b6f9676e JA	326	else {
	327	log_err("fio: bad mem type: %d\n", td->o.mem_type);
	328	ret = 1;
	329	}
	330
3deb3101 JA	331	if (ret)
	332	td_verror(td, ENOMEM, "iomem allocation");
	333
b6f9676e JA	334	return ret;
	335	}
	336
2f9ade3c JA	337	void free_io_mem(struct thread_data *td)
2f9ade3c JA	338	{
829a602c JA	339	unsigned int total_mem;
	340
	341	total_mem = td->orig_buffer_size;
d01612f3	342	if (td->o.odirect \|\| td->o.oatomic)
829a602c JA	343	total_mem += page_mask;
829a602c JA	344
a6cb85e2 JA	345	if (td->io_ops->iomem_alloc) {
	346	if (td->io_ops->iomem_free)
	347	td->io_ops->iomem_free(td);
	348	} else if (td->o.mem_type == MEM_MALLOC)
829a602c JA	349	free_mem_malloc(td);
	350	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	351	free_mem_shm(td);
217b0f1d LG	352	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE \|\|
217b0f1d LG	353	td->o.mem_type == MEM_MMAPSHARED)
829a602c	354	free_mem_mmap(td, total_mem);
03553853 YR	355	else if (td->o.mem_type == MEM_CUDA_MALLOC)
03553853 YR	356	free_mem_cudamalloc(td);
829a602c	357	else
2dc1bbeb	358	log_err("Bad memory type %u\n", td->o.mem_type);
2f9ade3c JA	359
2f9ade3c JA	360	td->orig_buffer = NULL;
829a602c	361	td->orig_buffer_size = 0;
2f9ade3c	362	}