configure: update compiler probing
[fio.git] / memory.c
index 257914b..04dc3be 100644 (file)
--- a/memory.c
+++ b/memory.c
@@ -1,75 +1,83 @@
 /*
  * Memory helpers
  */
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
 #include <unistd.h>
-#include <sys/shm.h>
 #include <sys/mman.h>
 
 #include "fio.h"
+#ifndef FIO_NO_HAVE_SHM_H
+#include <sys/shm.h>
+#endif
 
-static void *pinned_mem;
-
-void fio_unpin_memory(void)
+void fio_unpin_memory(struct thread_data *td)
 {
-       if (pinned_mem) {
-               dprint(FD_MEM, "unpinning %llu bytes\n", mlock_size);
-               if (munlock(pinned_mem, mlock_size) < 0)
+       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(pinned_mem, mlock_size);
-               pinned_mem = NULL;
+               munmap(td->pinned_mem, td->o.lockmem);
+               td->pinned_mem = NULL;
        }
 }
 
-int fio_pin_memory(void)
+int fio_pin_memory(struct thread_data *td)
 {
        unsigned long long phys_mem;
 
-       if (!mlock_size)
+       if (!td->o.lockmem)
                return 0;
 
-       dprint(FD_MEM, "pinning %llu bytes\n", mlock_size);
+       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 ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
-                       mlock_size = phys_mem - 128 * 1024 * 1024;
+               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",
-                                                       mlock_size >> 20);
+                                                       td->o.lockmem >> 20);
                }
        }
 
-       pinned_mem = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE,
-                               MAP_PRIVATE | OS_MAP_ANON, 0, 0);
-       if (pinned_mem == MAP_FAILED) {
+       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");
-               pinned_mem = NULL;
+               td->pinned_mem = NULL;
                return 1;
        }
-       if (mlock(pinned_mem, mlock_size) < 0) {
+       if (mlock(td->pinned_mem, td->o.lockmem) < 0) {
                perror("mlock");
-               munmap(pinned_mem, mlock_size);
-               pinned_mem = NULL;
+               munmap(td->pinned_mem, td->o.lockmem);
+               td->pinned_mem = NULL;
                return 1;
        }
 
        return 0;
 }
 
-static int alloc_mem_shm(struct thread_data *td)
+static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem)
 {
-       int flags = IPC_CREAT | SHM_R | SHM_W;
+#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;
 
-       if (td->o.mem_type == MEM_SHMHUGE)
                flags |= SHM_HUGETLB;
+               total_mem = (total_mem + mask) & ~mask;
+       }
 
-       td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, flags);
-       dprint(FD_MEM, "shmget %zu, %d\n", td->orig_buffer_size, td->shm_id);
+       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)
+               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) {
@@ -81,7 +89,7 @@ static int alloc_mem_shm(struct thread_data *td)
                                        " 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");
                        }
                }
 
@@ -97,40 +105,75 @@ static int alloc_mem_shm(struct thread_data *td)
        }
 
        return 0;
+#else
+       log_err("fio: shm not supported\n");
+       return 1;
+#endif
 }
 
-static int alloc_mem_mmap(struct thread_data *td)
+static void free_mem_shm(struct thread_data *td)
 {
-       int flags = MAP_PRIVATE;
+#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
+}
 
-       td->mmapfd = 0;
+static int alloc_mem_mmap(struct thread_data *td, size_t total_mem)
+{
+       int flags = 0;
 
-       if (td->mmapfile) {
-               td->mmapfd = open(td->mmapfile, O_RDWR|O_CREAT, 0644);
+       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 (ftruncate(td->mmapfd, td->orig_buffer_size) < 0) {
+               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;
+               flags |= OS_MAP_ANON | MAP_PRIVATE;
 
-       td->orig_buffer = mmap(NULL, td->orig_buffer_size,
-                               PROT_READ | PROT_WRITE, flags, td->mmapfd, 0);
-       dprint(FD_MEM, "mmap %zu/%d %p\n", td->orig_buffer_size, td->mmapfd,
-                                                       td->orig_buffer);
+       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) {
+               if (td->mmapfd != 1 && td->mmapfd != -1) {
                        close(td->mmapfd);
-                       unlink(td->mmapfile);
+                       if (td->o.mmapfile && !(td->flags & TD_F_MMAP_KEEP))
+                               unlink(td->o.mmapfile);
                }
 
                return 1;
@@ -139,37 +182,148 @@ static int alloc_mem_mmap(struct thread_data *td)
        return 0;
 }
 
-static int alloc_mem_malloc(struct thread_data *td)
+static void free_mem_mmap(struct thread_data *td, size_t total_mem)
 {
-       unsigned int bsize = td->orig_buffer_size;
+       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);
+       }
+}
 
-       if (td->o.odirect)
-               bsize += page_mask;
+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);
 
-       td->orig_buffer = malloc(bsize);
-       dprint(FD_MEM, "malloc %u %p\n", bsize, td->orig_buffer);
-       if (td->orig_buffer)
-               return 0;
+       return td->orig_buffer == NULL;
+}
 
-       return 1;
+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
 }
 
 /*
- * Setup the buffer area we need for io.
+ * 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)
+       if (td_ioengine_flagged(td, FIO_NOIO))
                return 0;
 
-       if (td->o.mem_type == MEM_MALLOC)
-               ret = alloc_mem_malloc(td);
+       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);
-       else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE)
-               ret = alloc_mem_mmap(td);
+               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;
@@ -183,28 +337,27 @@ int allocate_io_mem(struct thread_data *td)
 
 void free_io_mem(struct thread_data *td)
 {
-       if (td->o.mem_type == MEM_MALLOC) {
-               dprint(FD_MEM, "free mem %p\n", td->orig_buffer);
-               free(td->orig_buffer);
-       } else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE) {
-               struct shmid_ds sbuf;
+       unsigned int total_mem;
 
-               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);
-       } else if (td->o.mem_type == MEM_MMAP ||
-                  td->o.mem_type == MEM_MMAPHUGE) {
-               dprint(FD_MEM, "munmap %zu %p\n", td->orig_buffer_size,
-                                                       td->orig_buffer);
-               munmap(td->orig_buffer, td->orig_buffer_size);
-               if (td->mmapfile) {
-                       close(td->mmapfd);
-                       unlink(td->mmapfile);
-                       free(td->mmapfile);
-               }
-       } else
+       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;
 }