X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=memory.c;h=fe657225a2739bf878b2ddfb0909771ccb6e21ad;hp=e969221f36f86a8a1c74b502152a97e75f81c219;hb=035538530f86a5b3ecb8c9d0ec031562ec9b6028;hpb=95820b6e6c92025df8d89c0bf39b174e53137c41

diff --git a/memory.c b/memory.c
index e969221f..fe657225 100644
--- a/memory.c
+++ b/memory.c
@@ -5,12 +5,12 @@
 #include <sys/stat.h>
 #include <fcntl.h>
 #include <unistd.h>
-#ifndef FIO_NO_HAVE_SHM_H
-#include <sys/shm.h>
-#endif
 #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)
 {
@@ -33,13 +33,13 @@ int fio_pin_memory(struct thread_data *td)
 	dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem);
 
 	/*
-	 * Don't allow mlock of more than real_mem-128MB
+	 * 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 %lluMB\n",
+			log_info("fio: limiting mlocked memory to %lluMiB\n",
 							td->o.lockmem >> 20);
 		}
 	}
@@ -63,6 +63,7 @@ int fio_pin_memory(struct thread_data *td)
 
 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) {
@@ -88,7 +89,7 @@ static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
 					" support huge pages.\n");
 			} else if (errno == ENOMEM) {
 				log_err("fio: no huge pages available, do you"
-					" need to alocate some? See HOWTO.\n");
+					" need to allocate some? See HOWTO.\n");
 			}
 		}
 
@@ -104,22 +105,28 @@ static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
 	}
 
 	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;
+	td->mmapfd = -1;
 
 	if (td->o.mem_type == MEM_MMAPHUGE) {
 		unsigned long mask = td->o.hugepage_size - 1;
@@ -139,12 +146,14 @@ static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
 			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)
+		if (td->o.mem_type == MEM_MMAPHUGE ||
+		    td->o.mem_type == MEM_MMAPSHARED)
 			flags |= MAP_SHARED;
 		else
 			flags |= MAP_PRIVATE;
@@ -153,14 +162,15 @@ static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
 
 	td->orig_buffer = mmap(NULL, total_mem, PROT_READ | PROT_WRITE, flags,
 				td->mmapfd, 0);
-	dprint(FD_MEM, "mmap %u/%d %p\n", total_mem, td->mmapfd,
-						td->orig_buffer);
+	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) {
+		if (td->mmapfd != 1 && td->mmapfd != -1) {
 			close(td->mmapfd);
-			unlink(td->o.mmapfile);
+			if (td->o.mmapfile)
+				unlink(td->o.mmapfile);
 		}
 
 		return 1;
@@ -171,10 +181,12 @@ static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
 
 static void free_mem_mmap(struct thread_data *td, size_t total_mem)
 {
-	dprint(FD_MEM, "munmap %u %p\n", total_mem, td->orig_buffer);
+	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) {
-		close(td->mmapfd);
+		if (td->mmapfd != -1)
+			close(td->mmapfd);
 		unlink(td->o.mmapfile);
 		free(td->o.mmapfile);
 	}
@@ -183,7 +195,8 @@ static void free_mem_mmap(struct thread_data *td, size_t total_mem)
 static int alloc_mem_malloc(struct thread_data *td, size_t total_mem)
 {
 	td->orig_buffer = malloc(total_mem);
-	dprint(FD_MEM, "malloc %u %p\n", total_mem, td->orig_buffer);
+	dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem,
+							td->orig_buffer);
 
 	return td->orig_buffer == NULL;
 }
@@ -194,6 +207,75 @@ static void free_mem_malloc(struct thread_data *td)
 	free(td->orig_buffer);
 }
 
+#ifdef CONFIG_CUDA
+
+static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem)
+{
+	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;
+}
+
+static void free_mem_cudamalloc(struct thread_data *td)
+{
+	if ((void *) 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.
  */
@@ -202,26 +284,41 @@ int allocate_io_mem(struct thread_data *td)
 	size_t total_mem;
 	int ret = 0;
 
-	if (td->io_ops->flags & FIO_NOIO)
+	if (td_ioengine_flagged(td, FIO_NOIO))
 		return 0;
 
 	total_mem = td->orig_buffer_size;
 
-	if (td->o.odirect || td->o.mem_align ||
-	    (td->io_ops->flags & FIO_MEMALIGN)) {
+	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 %lu for buffers\n", (size_t) total_mem);
+	dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem);
 
-	if (td->o.mem_type == MEM_MALLOC)
+	/*
+	 * 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)
+	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);
+#ifdef CONFIG_CUDA
+	else if (td->o.mem_type == MEM_CUDA_MALLOC)
+		ret = alloc_mem_cudamalloc(td, total_mem);
+#endif
 	else {
 		log_err("fio: bad mem type: %d\n", td->o.mem_type);
 		ret = 1;
@@ -238,15 +335,23 @@ void free_io_mem(struct thread_data *td)
 	unsigned int total_mem;
 
 	total_mem = td->orig_buffer_size;
-	if (td->o.odirect)
+	if (td->o.odirect || td->o.oatomic)
 		total_mem += page_mask;
 
-	if (td->o.mem_type == MEM_MALLOC)
+	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)
+	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);
+#ifdef CONFIG_CUDA
+	else if (td->o.mem_type == MEM_CUDA_MALLOC)
+		free_mem_cudamalloc(td);
+#endif
 	else
 		log_err("Bad memory type %u\n", td->o.mem_type);