rpma: add support for File System DAX

File System DAX is handled in a different way than Device DAX:

1) In case of File System DAX, each thread uses a separate file
from this file system and no offset is needed. In case of Device DAX,
each thread uses a separate offset within the same Device DAX.

2) File System DAX requires rpma_mr_advise(3)(ibv_advise_mr(3))
to be called for the registered memory to avoid page faults
and degraded performance.

Ref: https://github.com/axboe/fio/issues/1238

Signed-off-by: Wang, Long <long1.wang@intel.com>

diff --git a/engines/librpma_fio.c b/engines/librpma_fio.c
index 3d605ed6c3721dcb5866d3bbe6b4754a3cb9e3dd..9d6ebf38ee865130ae9450dc79719dbbecc3f419 100644 (file)
--- a/engines/librpma_fio.c
+++ b/engines/librpma_fio.c
@@ -108,7 +108,7 @@ char *librpma_fio_allocate_dram(struct thread_data *td, size_t size,
        return mem_ptr;
 }
 
-char *librpma_fio_allocate_pmem(struct thread_data *td, const char *filename,
+char *librpma_fio_allocate_pmem(struct thread_data *td, struct fio_file *f,
                size_t size, struct librpma_fio_mem *mem)
 {
        size_t size_mmap = 0;
@@ -122,18 +122,24 @@ char *librpma_fio_allocate_pmem(struct thread_data *td, const char *filename,
                return NULL;
        }
 
-       ws_offset = (td->thread_number - 1) * size;
+       if (f->filetype == FIO_TYPE_CHAR) {
+               /* Each thread uses a separate offset within DeviceDAX. */
+               ws_offset = (td->thread_number - 1) * size;
+       } else {
+               /* Each thread uses a separate FileSystemDAX file. No offset is needed. */
+               ws_offset = 0;
+       }
 
-       if (!filename) {
+       if (!f->file_name) {
                log_err("fio: filename is not set\n");
                return NULL;
        }
 
        /* map the file */
-       mem_ptr = pmem_map_file(filename, 0 /* len */, 0 /* flags */,
+       mem_ptr = pmem_map_file(f->file_name, 0 /* len */, 0 /* flags */,
                        0 /* mode */, &size_mmap, &is_pmem);
        if (mem_ptr == NULL) {
-               log_err("fio: pmem_map_file(%s) failed\n", filename);
+               log_err("fio: pmem_map_file(%s) failed\n", f->file_name);
                /* pmem_map_file() sets errno on failure */
                td_verror(td, errno, "pmem_map_file");
                return NULL;
@@ -142,7 +148,7 @@ char *librpma_fio_allocate_pmem(struct thread_data *td, const char *filename,
        /* pmem is expected */
        if (!is_pmem) {
                log_err("fio: %s is not located in persistent memory\n",
-                       filename);
+                       f->file_name);
                goto err_unmap;
        }
 
@@ -150,12 +156,12 @@ char *librpma_fio_allocate_pmem(struct thread_data *td, const char *filename,
        if (size_mmap < ws_offset + size) {
                log_err(
                        "fio: %s is too small to handle so many threads (%zu < %zu)\n",
-                       filename, size_mmap, ws_offset + size);
+                       f->file_name, size_mmap, ws_offset + size);
                goto err_unmap;
        }
 
        log_info("fio: size of memory mapped from the file %s: %zu\n",
-               filename, size_mmap);
+               f->file_name, size_mmap);
 
        mem->mem_ptr = mem_ptr;
        mem->size_mmap = size_mmap;
@@ -893,6 +899,7 @@ int librpma_fio_server_open_file(struct thread_data *td, struct fio_file *f,
        size_t mem_size = td->o.size;
        size_t mr_desc_size;
        void *ws_ptr;
+       bool is_dram;
        int usage_mem_type;
        int ret;
 
@@ -910,14 +917,14 @@ int librpma_fio_server_open_file(struct thread_data *td, struct fio_file *f,
                return -1;
        }
 
-       if (strcmp(f->file_name, "malloc") == 0) {
+       is_dram = !strcmp(f->file_name, "malloc");
+       if (is_dram) {
                /* allocation from DRAM using posix_memalign() */
                ws_ptr = librpma_fio_allocate_dram(td, mem_size, &csd->mem);
                usage_mem_type = RPMA_MR_USAGE_FLUSH_TYPE_VISIBILITY;
        } else {
                /* allocation from PMEM using pmem_map_file() */
-               ws_ptr = librpma_fio_allocate_pmem(td, f->file_name,
-                               mem_size, &csd->mem);
+               ws_ptr = librpma_fio_allocate_pmem(td, f, mem_size, &csd->mem);
                usage_mem_type = RPMA_MR_USAGE_FLUSH_TYPE_PERSISTENT;
        }
 
@@ -934,6 +941,21 @@ int librpma_fio_server_open_file(struct thread_data *td, struct fio_file *f,
                goto err_free;
        }
 
+       if (!is_dram && f->filetype == FIO_TYPE_FILE) {
+               ret = rpma_mr_advise(mr, 0, mem_size,
+                               IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE,
+                               IBV_ADVISE_MR_FLAG_FLUSH);
+               if (ret) {
+                       librpma_td_verror(td, ret, "rpma_mr_advise");
+                       /* an invalid argument is an error */
+                       if (ret == RPMA_E_INVAL)
+                               goto err_mr_dereg;
+
+                       /* log_err used instead of log_info to avoid corruption of the JSON output */
+                       log_err("Note: having rpma_mr_advise(3) failed because of RPMA_E_NOSUPP or RPMA_E_PROVIDER may come with a performance penalty, but it is not a blocker for running the benchmark.\n");
+               }
+       }
+
        /* get size of the memory region's descriptor */
        if ((ret = rpma_mr_get_descriptor_size(mr, &mr_desc_size))) {
                librpma_td_verror(td, ret, "rpma_mr_get_descriptor_size");

diff --git a/engines/librpma_fio.h b/engines/librpma_fio.h
index fb89d99d696cd26be4061ec30865782051ac16ae..2c507e9c5c1c74290a61d10ba04b30d1e3a01d62 100644 (file)
--- a/engines/librpma_fio.h
+++ b/engines/librpma_fio.h
@@ -77,7 +77,7 @@ struct librpma_fio_mem {
 char *librpma_fio_allocate_dram(struct thread_data *td, size_t size,
                struct librpma_fio_mem *mem);
 
-char *librpma_fio_allocate_pmem(struct thread_data *td, const char *filename,
+char *librpma_fio_allocate_pmem(struct thread_data *td, struct fio_file *f,
                size_t size, struct librpma_fio_mem *mem);
 
 void librpma_fio_free(struct librpma_fio_mem *mem);