[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-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 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	35	/*
420b104a	36	* Don't allow mlock of more than real_mem-128MiB
2f9ade3c JA	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;
420b104a	42	log_info("fio: limiting mlocked memory to %lluMiB\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"
dd117eb6	92	" need to allocate some? See HOWTO.\n");
5ec10eaa	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	140	if (td->o.mmapfile) {
f9cfc7d4 SB	141	if (access(td->o.mmapfile, F_OK) == 0)
	142	td->flags \|= TD_F_MMAP_KEEP;
	143
83ea422a	144	td->mmapfd = open(td->o.mmapfile, O_RDWR\|O_CREAT, 0644);
b6f9676e JA	145
	146	if (td->mmapfd < 0) {
	147	td_verror(td, errno, "open mmap file");
2f9ade3c JA	148	td->orig_buffer = NULL;
	149	return 1;
	150	}
d9759b1e	151	if (td->o.mem_type != MEM_MMAPHUGE &&
217b0f1d	152	td->o.mem_type != MEM_MMAPSHARED &&
d9759b1e	153	ftruncate(td->mmapfd, total_mem) < 0) {
b6f9676e	154	td_verror(td, errno, "truncate mmap file");
2f9ade3c JA	155	td->orig_buffer = NULL;
	156	return 1;
	157	}
217b0f1d LG	158	if (td->o.mem_type == MEM_MMAPHUGE \|\|
217b0f1d LG	159	td->o.mem_type == MEM_MMAPSHARED)
d9759b1e SL	160	flags \|= MAP_SHARED;
	161	else
	162	flags \|= MAP_PRIVATE;
b6f9676e	163	} else
d9759b1e	164	flags \|= OS_MAP_ANON \| MAP_PRIVATE;
b6f9676e	165
829a602c JA	166	td->orig_buffer = mmap(NULL, total_mem, PROT_READ \| PROT_WRITE, flags,
829a602c JA	167	td->mmapfd, 0);
4b91ee8f JA	168	dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem,
4b91ee8f JA	169	td->mmapfd, td->orig_buffer);
b6f9676e JA	170	if (td->orig_buffer == MAP_FAILED) {
	171	td_verror(td, errno, "mmap");
	172	td->orig_buffer = NULL;
9ce94349	173	if (td->mmapfd != 1 && td->mmapfd != -1) {
b6f9676e	174	close(td->mmapfd);
f9cfc7d4	175	if (td->o.mmapfile && !(td->flags & TD_F_MMAP_KEEP))
b3493a7a	176	unlink(td->o.mmapfile);
b6f9676e	177	}
5ec10eaa	178
b6f9676e	179	return 1;
2f9ade3c JA	180	}
	181
	182	return 0;
	183	}
	184
0f805c00	185	static void free_mem_mmap(struct thread_data *td, size_t total_mem)
b6f9676e	186	{
4b91ee8f JA	187	dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem,
4b91ee8f JA	188	td->orig_buffer);
829a602c	189	munmap(td->orig_buffer, td->orig_buffer_size);
83ea422a	190	if (td->o.mmapfile) {
4a995dda JA	191	if (td->mmapfd != -1)
4a995dda JA	192	close(td->mmapfd);
f9cfc7d4 SB	193	if (!(td->flags & TD_F_MMAP_KEEP))
f9cfc7d4 SB	194	unlink(td->o.mmapfile);
83ea422a	195	free(td->o.mmapfile);
829a602c JA	196	}
829a602c JA	197	}
d87612ac	198
0f805c00	199	static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
829a602c JA	200	{
829a602c JA	201	td->orig_buffer = malloc(total_mem);
4b91ee8f JA	202	dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
4b91ee8f JA	203	td->orig_buffer);
5ec10eaa	204
829a602c JA	205	return td->orig_buffer == NULL;
829a602c JA	206	}
b6f9676e	207
829a602c JA	208	static void free_mem_malloc(struct thread_data *td)
	209	{
	210	dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer);
	211	free(td->orig_buffer);
b6f9676e JA	212	}
b6f9676e JA	213
03553853 YR	214	static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
03553853 YR	215	{
15600335	216	#ifdef CONFIG_CUDA
03553853 YR	217	CUresult ret;
	218	char name[128];
	219
	220	ret = cuInit(0);
	221	if (ret != CUDA_SUCCESS) {
	222	log_err("fio: failed initialize cuda driver api\n");
	223	return 1;
	224	}
	225
	226	ret = cuDeviceGetCount(&td->gpu_dev_cnt);
	227	if (ret != CUDA_SUCCESS) {
	228	log_err("fio: failed get device count\n");
	229	return 1;
	230	}
	231	dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt);
	232
	233	if (td->gpu_dev_cnt == 0) {
	234	log_err("fio: no GPU device found. "
	235	"Can not perform GPUDirect RDMA.\n");
	236	return 1;
	237	}
	238
	239	td->gpu_dev_id = td->o.gpu_dev_id;
	240	ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id);
	241	if (ret != CUDA_SUCCESS) {
	242	log_err("fio: failed get GPU device\n");
	243	return 1;
	244	}
	245
	246	ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id);
	247	if (ret != CUDA_SUCCESS) {
	248	log_err("fio: failed get device name\n");
	249	return 1;
	250	}
	251	dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \
	252	td->gpu_dev_id, name);
	253
	254	ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev);
	255	if (ret != CUDA_SUCCESS) {
	256	log_err("fio: failed to create cuda context: %d\n", ret);
	257	return 1;
	258	}
	259
	260	ret = cuMemAlloc(&td->dev_mem_ptr, total_mem);
	261	if (ret != CUDA_SUCCESS) {
	262	log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem);
	263	return 1;
	264	}
	265	td->orig_buffer = (void *) td->dev_mem_ptr;
	266
	267	dprint(FD_MEM, "cudaMalloc %llu %p\n", \
	268	(unsigned long long) total_mem, td->orig_buffer);
	269	return 0;
15600335 JA	270	#else
	271	return -EINVAL;
	272	#endif
03553853 YR	273	}
	274
	275	static void free_mem_cudamalloc(struct thread_data *td)
	276	{
15600335 JA	277	#ifdef CONFIG_CUDA
15600335 JA	278	if (td->dev_mem_ptr != NULL)
03553853 YR	279	cuMemFree(td->dev_mem_ptr);
	280
	281	if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS)
	282	log_err("fio: failed to destroy cuda context\n");
03553853	283	#endif
15600335	284	}
03553853	285
b6f9676e	286	/*
03e20d68	287	* Set up the buffer area we need for io.
b6f9676e JA	288	*/
	289	int allocate_io_mem(struct thread_data *td)
	290	{
0f805c00	291	size_t total_mem;
b6f9676e JA	292	int ret = 0;
b6f9676e JA	293
9b87f09b	294	if (td_ioengine_flagged(td, FIO_NOIO))
b4c5e1ac JA	295	return 0;
b4c5e1ac JA	296
829a602c	297	total_mem = td->orig_buffer_size;
d529ee19	298
d01612f3	299	if (td->o.odirect \|\| td->o.mem_align \|\| td->o.oatomic \|\|
9b87f09b	300	td_ioengine_flagged(td, FIO_MEMALIGN)) {
829a602c	301	total_mem += page_mask;
d529ee19 JA	302	if (td->o.mem_align && td->o.mem_align > page_size)
	303	total_mem += td->o.mem_align - page_size;
	304	}
829a602c	305
4b91ee8f	306	dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
0f805c00	307
a6cb85e2 JA	308	/*
	309	* If the IO engine has hooks to allocate/free memory, use those. But
	310	* error out if the user explicitly asked for something else.
	311	*/
	312	if (td->io_ops->iomem_alloc) {
	313	if (fio_option_is_set(&td->o, mem_type)) {
	314	log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n");
	315	ret = 1;
	316	} else
	317	ret = td->io_ops->iomem_alloc(td, total_mem);
	318	} else if (td->o.mem_type == MEM_MALLOC)
829a602c	319	ret = alloc_mem_malloc(td, total_mem);
b6f9676e	320	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
829a602c	321	ret = alloc_mem_shm(td, total_mem);
217b0f1d LG	322	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE \|\|
217b0f1d LG	323	td->o.mem_type == MEM_MMAPSHARED)
829a602c	324	ret = alloc_mem_mmap(td, total_mem);
03553853 YR	325	else if (td->o.mem_type == MEM_CUDA_MALLOC)
03553853 YR	326	ret = alloc_mem_cudamalloc(td, total_mem);
b6f9676e JA	327	else {
	328	log_err("fio: bad mem type: %d\n", td->o.mem_type);
	329	ret = 1;
	330	}
	331
3deb3101 JA	332	if (ret)
	333	td_verror(td, ENOMEM, "iomem allocation");
	334
b6f9676e JA	335	return ret;
	336	}
	337
2f9ade3c JA	338	void free_io_mem(struct thread_data *td)
2f9ade3c JA	339	{
829a602c JA	340	unsigned int total_mem;
	341
	342	total_mem = td->orig_buffer_size;
d01612f3	343	if (td->o.odirect \|\| td->o.oatomic)
829a602c JA	344	total_mem += page_mask;
829a602c JA	345
a6cb85e2 JA	346	if (td->io_ops->iomem_alloc) {
	347	if (td->io_ops->iomem_free)
	348	td->io_ops->iomem_free(td);
	349	} else if (td->o.mem_type == MEM_MALLOC)
829a602c JA	350	free_mem_malloc(td);
	351	else if (td->o.mem_type == MEM_SHM \|\| td->o.mem_type == MEM_SHMHUGE)
	352	free_mem_shm(td);
217b0f1d LG	353	else if (td->o.mem_type == MEM_MMAP \|\| td->o.mem_type == MEM_MMAPHUGE \|\|
217b0f1d LG	354	td->o.mem_type == MEM_MMAPSHARED)
829a602c	355	free_mem_mmap(td, total_mem);
03553853 YR	356	else if (td->o.mem_type == MEM_CUDA_MALLOC)
03553853 YR	357	free_mem_cudamalloc(td);
829a602c	358	else
2dc1bbeb	359	log_err("Bad memory type %u\n", td->o.mem_type);
2f9ade3c JA	360
2f9ade3c JA	361	td->orig_buffer = NULL;
829a602c	362	td->orig_buffer_size = 0;
2f9ade3c	363	}