[PATCH] Split out the memory handling from fio.c
In the process also fix some bugs in the memory pinning.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
diff --git a/Makefile b/Makefile
index a3e215e..111d5cd 100644
--- a/Makefile
+++ b/Makefile
@@ -2,6 +2,8 @@
CFLAGS = -W -Wall -O2 -g -D_GNU_SOURCE -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64
PROGS = fio
SCRIPTS = fio_generate_plots
+OBJS = fio.o ioengines.o init.o stat.o log.o time.o md5.o crc32.o \
+ filesetup.o eta.o verify.o memory.o
INSTALL = install
prefix = /usr/local
@@ -15,7 +17,7 @@
all: depend $(PROGS) $(SCRIPTS)
$(MAKE) -C engines
-fio: fio.o ioengines.o init.o stat.o log.o time.o md5.o crc32.o filesetup.o eta.o verify.o
+fio: $(OBJS)
$(CC) $(CFLAGS) -o $@ $(filter %.o,$^) -lpthread -lm -ldl
clean:
diff --git a/fio.c b/fio.c
index 624c959..8d2a8e5 100644
--- a/fio.c
+++ b/fio.c
@@ -45,8 +45,6 @@
int temp_stall_ts;
char *fio_inst_prefix = _INST_PREFIX;
-extern unsigned long long mlock_size;
-
#define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
static volatile int startup_sem;
@@ -237,8 +235,6 @@
rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
if (rate < td->ratemin) {
fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
- if (rate_quit)
- terminate_threads(td->groupid);
return 1;
}
}
@@ -753,6 +749,8 @@
rate_throttle(td, usec, icd.bytes_done[td->ddir]);
if (check_min_rate(td, &e)) {
+ if (rate_quit)
+ terminate_threads(td->groupid);
td_verror(td, ENOMEM);
break;
}
@@ -800,19 +798,7 @@
free(io_u);
}
- if (td->mem_type == MEM_MALLOC)
- free(td->orig_buffer);
- else if (td->mem_type == MEM_SHM) {
- struct shmid_ds sbuf;
-
- shmdt(td->orig_buffer);
- shmctl(td->shm_id, IPC_RMID, &sbuf);
- } else if (td->mem_type == MEM_MMAP)
- munmap(td->orig_buffer, td->orig_buffer_size);
- else
- log_err("Bad memory type %d\n", td->mem_type);
-
- td->orig_buffer = NULL;
+ free_io_mem(td);
}
static int init_io_u(struct thread_data *td)
@@ -831,32 +817,8 @@
td->orig_buffer_size = td->max_bs * max_units + MASK;
- if (td->mem_type == MEM_MALLOC)
- td->orig_buffer = malloc(td->orig_buffer_size);
- else if (td->mem_type == MEM_SHM) {
- td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
- if (td->shm_id < 0) {
- td_verror(td, errno);
- perror("shmget");
- return 1;
- }
-
- td->orig_buffer = shmat(td->shm_id, NULL, 0);
- if (td->orig_buffer == (void *) -1) {
- td_verror(td, errno);
- perror("shmat");
- td->orig_buffer = NULL;
- return 1;
- }
- } else if (td->mem_type == MEM_MMAP) {
- td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
- if (td->orig_buffer == MAP_FAILED) {
- td_verror(td, errno);
- perror("mmap");
- td->orig_buffer = NULL;
- return 1;
- }
- }
+ if (allocate_io_mem(td))
+ return 1;
p = ALIGN(td->orig_buffer);
for (i = 0; i < max_units; i++) {
@@ -1126,48 +1088,6 @@
terminate_threads(TERMINATE_ALL);
}
-static void fio_unpin_memory(void *pinned)
-{
- if (pinned) {
- if (munlock(pinned, mlock_size) < 0)
- perror("munlock");
- munmap(pinned, mlock_size);
- }
-}
-
-static void *fio_pin_memory(void)
-{
- unsigned long long phys_mem;
- void *ptr;
-
- if (!mlock_size)
- return NULL;
-
- /*
- * Don't allow mlock of more than real_mem-128MB
- */
- phys_mem = os_phys_mem();
- if (phys_mem) {
- if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
- mlock_size = phys_mem - 128 * 1024 * 1024;
- fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
- }
- }
-
- ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
- if (!ptr) {
- perror("malloc locked mem");
- return NULL;
- }
- if (mlock(ptr, mlock_size) < 0) {
- munmap(ptr, mlock_size);
- perror("mlock");
- return NULL;
- }
-
- return ptr;
-}
-
/*
* Main function for kicking off and reaping jobs, as needed.
*/
@@ -1176,9 +1096,9 @@
struct thread_data *td;
unsigned long spent;
int i, todo, nr_running, m_rate, t_rate, nr_started;
- void *mlocked_mem;
- mlocked_mem = fio_pin_memory();
+ if (fio_pin_memory())
+ return;
if (!terse_output) {
printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
@@ -1336,7 +1256,7 @@
}
update_io_ticks();
- fio_unpin_memory(mlocked_mem);
+ fio_unpin_memory();
}
int main(int argc, char *argv[])
diff --git a/fio.h b/fio.h
index 4af6a74..b9ae15b 100644
--- a/fio.h
+++ b/fio.h
@@ -480,6 +480,14 @@
extern int do_io_u_verify(struct thread_data *, struct io_u **);
/*
+ * Memory helpers
+ */
+extern int fio_pin_memory(void);
+extern void fio_unpin_memory(void);
+extern int allocate_io_mem(struct thread_data *);
+extern void free_io_mem(struct thread_data *);
+
+/*
* This is a pretty crappy semaphore implementation, but with the use that fio
* has (just signalling start/go conditions), it doesn't have to be better.
* Naturally this would not work for any type of contended semaphore or
diff --git a/memory.c b/memory.c
new file mode 100644
index 0000000..d8924a8
--- /dev/null
+++ b/memory.c
@@ -0,0 +1,108 @@
+/*
+ * Memory helpers
+ */
+#include <unistd.h>
+#include <sys/shm.h>
+#include <sys/mman.h>
+
+#include "fio.h"
+#include "os.h"
+
+extern unsigned long long mlock_size;
+static void *pinned_mem;
+
+void fio_unpin_memory(void)
+{
+ if (pinned_mem) {
+ if (munlock(pinned_mem, mlock_size) < 0)
+ perror("munlock");
+ munmap(pinned_mem, mlock_size);
+ pinned_mem = NULL;
+ }
+}
+
+int fio_pin_memory(void)
+{
+ unsigned long long phys_mem;
+
+ if (!mlock_size)
+ return 0;
+
+ /*
+ * Don't allow mlock of more than real_mem-128MB
+ */
+ phys_mem = os_phys_mem();
+ if (phys_mem) {
+ if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
+ mlock_size = phys_mem - 128 * 1024 * 1024;
+ fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
+ }
+ }
+
+ pinned_mem = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+ if (pinned_mem == MAP_FAILED) {
+ perror("malloc locked mem");
+ pinned_mem = NULL;
+ return 1;
+ }
+ if (mlock(pinned_mem, mlock_size) < 0) {
+ perror("mlock");
+ munmap(pinned_mem, mlock_size);
+ pinned_mem = NULL;
+ return 1;
+ }
+
+ return 0;
+}
+
+/*
+ * Setup the buffer area we need for io.
+ */
+int allocate_io_mem(struct thread_data *td)
+{
+ if (td->mem_type == MEM_MALLOC)
+ td->orig_buffer = malloc(td->orig_buffer_size);
+ else if (td->mem_type == MEM_SHM) {
+ td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
+ if (td->shm_id < 0) {
+ td_verror(td, errno);
+ perror("shmget");
+ return 1;
+ }
+
+ td->orig_buffer = shmat(td->shm_id, NULL, 0);
+ if (td->orig_buffer == (void *) -1) {
+ td_verror(td, errno);
+ perror("shmat");
+ td->orig_buffer = NULL;
+ return 1;
+ }
+ } else if (td->mem_type == MEM_MMAP) {
+ td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+ if (td->orig_buffer == MAP_FAILED) {
+ td_verror(td, errno);
+ perror("mmap");
+ td->orig_buffer = NULL;
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+void free_io_mem(struct thread_data *td)
+{
+ if (td->mem_type == MEM_MALLOC)
+ free(td->orig_buffer);
+ else if (td->mem_type == MEM_SHM) {
+ struct shmid_ds sbuf;
+
+ shmdt(td->orig_buffer);
+ shmctl(td->shm_id, IPC_RMID, &sbuf);
+ } else if (td->mem_type == MEM_MMAP)
+ munmap(td->orig_buffer, td->orig_buffer_size);
+ else
+ log_err("Bad memory type %d\n", td->mem_type);
+
+ td->orig_buffer = NULL;
+}