Recent changes (master)

index 886140d..3295824 100644
@@ -235,6 +235,7 @@ void convert_thread_options_to_cpu(struct thread_options *o,
 	o->new_group = le32_to_cpu(top->new_group);
 	o->numjobs = le32_to_cpu(top->numjobs);
 	o->cpus_allowed_policy = le32_to_cpu(top->cpus_allowed_policy);
+	o->gpu_dev_id = le32_to_cpu(top->gpu_dev_id);
 	o->iolog = le32_to_cpu(top->iolog);
 	o->rwmixcycle = le32_to_cpu(top->rwmixcycle);
 	o->nice = le32_to_cpu(top->nice);

@@ -420,6 +421,7 @@ void convert_thread_options_to_net(struct thread_options_pack *top,
 	top->new_group = cpu_to_le32(o->new_group);
 	top->numjobs = cpu_to_le32(o->numjobs);
 	top->cpus_allowed_policy = cpu_to_le32(o->cpus_allowed_policy);
+	top->gpu_dev_id = cpu_to_le32(o->gpu_dev_id);
 	top->iolog = cpu_to_le32(o->iolog);
 	top->rwmixcycle = cpu_to_le32(o->rwmixcycle);
 	top->nice = cpu_to_le32(o->nice);

index b67613e..6b2b669 100644
@@ -59,6 +59,10 @@
 #define MPOL_LOCAL MPOL_MAX
 #endif
 
+#ifdef CONFIG_CUDA
+#include <cuda.h>
+#endif
+
 /*
  * offset generator types
  */

@@ -408,6 +412,18 @@ struct thread_data {
 	struct steadystate_data ss;
 
 	char verror[FIO_VERROR_SIZE];
+
+#ifdef CONFIG_CUDA
+	/*
+	 * for GPU memory management
+	 */
+	int gpu_dev_cnt;
+	int gpu_dev_id;
+	CUdevice  cu_dev;
+	CUcontext cu_ctx;
+	CUdeviceptr dev_mem_ptr;
+#endif	
+
 };
 
 /*

index 9aa452d..52a5f03 100644
@@ -1070,6 +1070,9 @@ static void init_flags(struct thread_data *td)
 
 	if (o->verify_async || o->io_submit_mode == IO_MODE_OFFLOAD)
 		td->flags |= TD_F_NEED_LOCK;
+
+	if (o->mem_type == MEM_CUDA_MALLOC)
+		td->flags &= ~TD_F_SCRAMBLE_BUFFERS;
 }
 
 static int setup_random_seeds(struct thread_data *td)

index 88f35c9..fd63119 100644
@@ -1674,8 +1674,10 @@ out:
 	if (!td_io_prep(td, io_u)) {
 		if (!td->o.disable_lat)
 			fio_gettime(&io_u->start_time, NULL);
+
 		if (do_scramble)
 			small_content_scramble(io_u);
+
 		return io_u;
 	}
 err_put:

@@ -2043,6 +2045,9 @@ void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
 {
 	struct thread_options *o = &td->o;
 
+	if (o->mem_type == MEM_CUDA_MALLOC)
+		return;
+
 	if (o->compress_percentage || o->dedupe_percentage) {
 		unsigned int perc = td->o.compress_percentage;
 		struct frand_state *rs;

index 0aec744..2fb6fe7 100644
@@ -4,7 +4,7 @@
 #ifdef CONFIG_GETMNTENT
 #include <mntent.h>
 
-#include "lib/mountcheck.h"
+#include "mountcheck.h"
 
 #define MTAB	"/etc/mtab"
 

index 448d3ff..8d08841 100644
@@ -10,7 +10,7 @@
 /**
  * num2str() - Cheesy number->string conversion, complete with carry rounding error.
  * @num: quantity (e.g., number of blocks, bytes or bits)
- * @maxlen: max number of digits in the output string (not counting prefix and units)
+ * @maxlen: max number of digits in the output string (not counting prefix and units, but counting .)
  * @base: multiplier for num (e.g., if num represents Ki, use 1024)
  * @pow2: select unit prefix - 0=power-of-10 decimal SI, nonzero=power-of-2 binary IEC
  * @units: select units - N2S_* macros defined in num2str.h

@@ -23,9 +23,9 @@ char *num2str(uint64_t num, int maxlen, int base, int pow2, int units)
 	const char **unitprefix;
 	const char *unitstr[] = { "", "/s", "B", "bit", "B/s", "bit/s" };
 	const unsigned int thousand[] = { 1000, 1024 };
-	unsigned int modulo, decimals;
+	unsigned int modulo;
 	int unit_index = 0, post_index, carry = 0;
-	char tmp[32];
+	char tmp[32], fmt[32];
 	char *buf;
 
 	compiletime_assert(sizeof(sistr) == sizeof(iecstr), "unit prefix arrays must be identical sizes");

@@ -62,6 +62,9 @@ char *num2str(uint64_t num, int maxlen, int base, int pow2, int units)
 		break;
 	}
 
+	/*
+	 * Divide by K/Ki until string length of num <= maxlen.
+	 */
 	modulo = -1U;
 	while (post_index < sizeof(sistr)) {
 		sprintf(tmp, "%llu", (unsigned long long) num);

@@ -74,6 +77,9 @@ char *num2str(uint64_t num, int maxlen, int base, int pow2, int units)
 		post_index++;
 	}
 
+	/*
+	 * If no modulo, then we're done.
+	 */
 	if (modulo == -1U) {
 done:
 		if (post_index >= ARRAY_SIZE(sistr))

@@ -84,23 +90,25 @@ done:
 		return buf;
 	}
 
+	/*
+	 * If no room for decimals, then we're done.
+	 */
 	sprintf(tmp, "%llu", (unsigned long long) num);
-	decimals = maxlen - strlen(tmp);
-	if ((int)decimals <= 1) {
+	if ((int)(maxlen - strlen(tmp)) <= 1) {
 		if (carry)
 			num++;
 		goto done;
 	}
 
-	do {
-		sprintf(tmp, "%u", modulo);
-		if (strlen(tmp) <= decimals - 1)
-			break;
-
-		modulo = (modulo + 9) / 10;
-	} while (1);
+	/*
+	 * Fill in everything and return the result.
+	 */
+	assert(maxlen - strlen(tmp) - 1 > 0);
+	assert(modulo < thousand[!!pow2]);
+	sprintf(fmt, "%%.%df", (int)(maxlen - strlen(tmp) - 1));
+	sprintf(tmp, fmt, (double)modulo / (double)thousand[!!pow2]);
 
-	sprintf(buf, "%llu.%u%s%s", (unsigned long long) num, modulo,
+	sprintf(buf, "%llu.%s%s%s", (unsigned long long) num, &tmp[2],
 			unitprefix[post_index], unitstr[unit_index]);
 	return buf;
 }

index b8ae809..0aeb935 100644
@@ -1,6 +1,13 @@
-#include "fio.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include <errno.h>
+#include <assert.h>
+
 #include "strntol.h"
 #include "pattern.h"
+#include "../minmax.h"
 #include "../oslib/strcasestr.h"
 
 /**

index 9c3e0d6..3f60a67 100644
@@ -36,7 +36,7 @@
 #include <string.h>
 #include <assert.h>
 #include "rand.h"
-#include "lib/pattern.h"
+#include "pattern.h"
 #include "../hash.h"
 
 int arch_random;

48 lines not on this page · raw message

index adf45bd..f622c8d 100644
@@ -2,7 +2,7 @@
 #include <stdlib.h>
 #include <limits.h>
 
-#include "lib/strntol.h"
+#include "strntol.h"
 
 long strntol(const char *str, size_t sz, char **end, int base)
 {

index 681df70..3d535c7 100644
@@ -6,7 +6,6 @@
 #include <sys/types.h>
 #include <fcntl.h>
 #include "ieee754.h"
-#include "../log.h"
 #include "zipf.h"
 #include "../minmax.h"
 #include "../hash.h"

94 lines not on this page · raw message

index e0deab0..85574d7 100644
@@ -2604,6 +2604,12 @@ struct fio_option fio_options[FIO_MAX_OPTS] = {
 			    .help = "Like mmap, but use huge pages",
 			  },
 #endif
+#ifdef CONFIG_CUDA
+			  { .ival = "cudamalloc",
+			    .oval = MEM_CUDA_MALLOC,
+			    .help = "Allocate GPU device memory for GPUDirect RDMA",
+			  },
+#endif
 		  },
 	},
 	{

@@ -3563,6 +3569,18 @@ struct fio_option fio_options[FIO_MAX_OPTS] = {
 		.help	= "Build fio with libnuma-dev(el) to enable this option",
 	},
 #endif
+#ifdef CONFIG_CUDA
+	{
+		.name	= "gpu_dev_id",
+		.lname	= "GPU device ID",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct thread_options, gpu_dev_id),
+		.help	= "Set GPU device ID for GPUDirect RDMA",
+		.def    = "0",
+		.category = FIO_OPT_C_GENERAL,
+		.group	= FIO_OPT_G_INVALID,
+	},
+#endif
 	{
 		.name	= "end_fsync",
 		.lname	= "End fsync",

index 2bbd14a..5fbccd3 100644
@@ -5,8 +5,7 @@
 #include <stdio.h>
 #include <unistd.h>
 
-#include "../os/os.h"
-#include "oslib/linux-dev-lookup.h"
+#include "linux-dev-lookup.h"
 
 int blktrace_lookup_device(const char *redirect, char *path, unsigned int maj,
 			   unsigned int min)

index 3329b83..3b33d0e 100644
@@ -1,5 +1,5 @@
 #include <string.h>
-#include "oslib/strlcat.h"
+#include "strlcat.h"
 
 size_t strlcat(char *dst, const char *src, size_t size)
 {

index fd5605f..4d4fddd 100644
@@ -135,6 +135,7 @@ static unsigned long long get_mult_time(const char *str, int len,
 	const char *p = str;
 	char *c;
 	unsigned long long mult = 1;
+	int i;
 
 	/*
          * Go forward until we hit a non-digit, or +/- sign

@@ -153,7 +154,7 @@ static unsigned long long get_mult_time(const char *str, int len,
 	}
 
 	c = strdup(p);
-	for (int i = 0; i < strlen(c); i++)
+	for (i = 0; i < strlen(c); i++)
 		c[i] = tolower(c[i]);
 
 	if (!strncmp("us", c, 2) || !strncmp("usec", c, 4))

index 798d5a8..5c720d4 100644
@@ -49,7 +49,7 @@ struct fio_net_cmd_reply {
 };
 
 enum {
-	FIO_SERVER_VER			= 60,
+	FIO_SERVER_VER			= 61,
 
 	FIO_SERVER_MAX_FRAGMENT_PDU	= 1024,
 	FIO_SERVER_MAX_CMD_MB		= 2048,

155 lines not on this page · raw message

index 2b2df33..d0f3fe9 100644
@@ -20,6 +20,7 @@ enum fio_memtype {
 	MEM_MMAP,	/* use anonynomous mmap */
 	MEM_MMAPHUGE,	/* memory mapped huge file */
 	MEM_MMAPSHARED, /* use mmap with shared flag */
+	MEM_CUDA_MALLOC,/* use GPU memory */
 };
 
 #define ERROR_STR_MAX	128

@@ -198,6 +199,8 @@ struct thread_options {
 	unsigned short numa_mem_mode;
 	unsigned int numa_mem_prefer_node;
 	char *numa_memnodes;
+	unsigned int gpu_dev_id;
+
 	unsigned int iolog;
 	unsigned int rwmixcycle;
 	unsigned int rwmix[DDIR_RWDIR_CNT];

@@ -336,7 +339,6 @@ struct thread_options_pack {
 	uint32_t iodepth_batch;
 	uint32_t iodepth_batch_complete_min;
 	uint32_t iodepth_batch_complete_max;
-	uint32_t __proper_alignment_for_64b;
 
 	uint64_t size;
 	uint64_t io_size;

@@ -411,7 +413,6 @@ struct thread_options_pack {
 	uint32_t bs_unaligned;
 	uint32_t fsync_on_close;
 	uint32_t bs_is_seq_rand;
-	uint32_t pad1;
 
 	uint32_t random_distribution;
 	uint32_t exitall_error;

@@ -467,6 +468,8 @@ struct thread_options_pack {
 	uint8_t verify_cpumask[FIO_TOP_STR_MAX];
 	uint8_t log_gz_cpumask[FIO_TOP_STR_MAX];
 #endif
+	uint32_t gpu_dev_id;
+	uint32_t pad;
 	uint32_t cpus_allowed_policy;
 	uint32_t iolog;
 	uint32_t rwmixcycle;