| Age | Commit message (Collapse) | Author |
|---|
|
* 'fix_devdax' of https://github.com/harish-24/fio:
Fix: dev-dax engine building with make
|
|
With recent changes to Makefile dev-dax skips building. This patch
fixes that.
Signed-off-by: Harish <harish@linux.ibm.com>
|
|
* 'io_uring-opt' of https://github.com/antonblanchard/fio:
io_uring: Avoid needless update of completion queue head pointer
|
|
I'm seeing a slowdown in io_uring performance on a POWER9 box when
the userspace and kernel polling threads are on two cores that
share an L2 cache.
fio_ioring_cqring_reap() always stores to the completion queue head
pointer, even if nothing was reaped and the value hasn't changed.
Changing this to only update the head pointer when it changes results
in a 95% improvement in performance on this particular test.
Signed-off-by: Anton Blanchard <anton@ozlabs.org>
|
|
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
* 'windows' of https://github.com/bvanassche/fio:
os/windows/posix.c: Use INVALID_SOCKET instead of < 0
os/windows/posix.c: Strip trailing whitespace
|
|
From
https://docs.microsoft.com/en-us/windows/win32/winsock/socket-data-type-2:
"Because the SOCKET type is unsigned, compiling existing source code from,
for example, a UNIX environment may lead to compiler warnings about
signed/unsigned data type mismatches.
This means, for example, that checking for errors when the socket and
accept functions return should not be done by comparing the return value
with -1, or seeing if the value is negative (both common and legal
approaches in UNIX). Instead, an application should use the manifest
constant INVALID_SOCKET as defined in the Winsock2.h header file."
This patch fixes the following build error for the latest version of
the MinGW64 cross-compiler on Cygwin:
os/windows/posix.c:886:17: error: comparison of unsigned expression < 0
is always false
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
* 'workqueue' of https://github.com/bvanassche/fio:
workqueue: Fix race conditions in the workqueue mechanism
workqueue: Rework while loop locking
|
|
From https://pubs.opengroup.org/onlinepubs/9699919799/functions/pthread_cond_wait.html:
"The application shall ensure that these functions are called with mutex
locked by the calling thread; otherwise, an error (for
PTHREAD_MUTEX_ERRORCHECK and robust mutexes) or undefined behavior (for
other mutexes) results.
From https://pubs.opengroup.org/onlinepubs/9699919799/functions/pthread_cond_signal.html:
"however, if predictable scheduling behavior is required, then that mutex
shall be locked by the thread calling pthread_cond_broadcast() or
pthread_cond_signal()."
Hence always hold the associated mutex around pthread_cond_wait() and
pthread_signal() calls.
This patch fixes the hangs reported by Travis and Appveyor for test case
t0013.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Reduce the number of pthread_mutex_lock() and unlock calls by moving
these outside while loops. Other than protecting the all_sw_idle() call with
the wq->flush_lock, this patch does not change any functionality.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
* 'overlap' of https://github.com/bvanassche/fio:
Add a test for serialize_overlap=1
Enable error checking for the mutex that serializes overlapping I/O
Fix a potentially infinite loop in check_overlap()
t/run-fio-tests.py: Increase IOPS tolerance further
.appveyor.yml: Select the latest software image
|
|
This test causes fio to call check_overlap().
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Since it is nontrivial to verify that pthread_mutex_lock() and unlock
calls are paired for the overlap_check mutex, enable runtime verification.
The PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP initializer causes
pthread_mutex_lock() to fail with EDEADLK if the mutex was already locked
and causes pthread_mutex_unlock() to fail with EPERM if the mutex is owned
by another thread.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
If the following happens:
* check_overlap() finds an overlap.
* All other threads finish after the overlap has been found and before
the next iteration of the do/while loop starts.
Then the do/while loop in check_overlap() will iterate forever. Fix this
by rewriting check_overlap() such that this cannot happen.
This patch fixes the following Coverity complaint:
CID 184174 (#2 of 2): Double lock (LOCK)
Fixes: c06379a65d46 ("fio: enable overlap checking with offload submission")
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
From an Appveyor build:
DEBUG:root:Test 12: iops1: 984.676882
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This is necessary to make the overlap test (t0013) pass.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Jens suggested we provide the user with the likely reasons why we failed
to load the given engine.
This should help better resolve these issues.
Signed-off-by: Yigal Korman <ykorman@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
When enabled, some of the more dependency-heavy internal engines are
converted to "plugin" engines, i.e. they are built into separate object
files and are loaded by fio on demand.
This helps downstream distros package these engines separately and not
force a long list of package dependencies from the base fio package.
Signed-off-by: Yigal Korman <ykorman@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Some engines were adding their LIBS in the Makefile and some were
declaring them in the configure script.
This is inconsistent and confusing and prevents easy bulk changes to the
engines.
Signed-off-by: Yigal Korman <ykorman@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
ioengine_load will try to load a given engine and if it doesn't match
the current one (default is always psync), it will remove the current
one and try again.
This makes sense when the engine is loaded successfully, but if it's
not, we try to load it twice for no reason.
Fix by failing if we were unable to load requested engine and not
retrying.
Signed-off-by: Yigal Korman <ykorman@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
* 'num2str' of https://github.com/bvanassche/fio:
num2str(): Add the E (exa) prefix
Add a num2str() unit test
num2str(): Fix overflow handling
num2str(): Remove the fmt[] array
num2str(): Use asprintf() instead of malloc()
|
|
This change guarantees that all 64-bit integers that should be assigned a
multiplier are assigned a multiplier.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This patch changes the output for the arguments (UINT64_MAX, 4, 1, 0,
N2S_NONE, "18.4") from "18.4(null)" into "18.4".
This patch fixes the following Coverity complaint:
CID 169307 (#1 of 1): Out-of-bounds read (OVERRUN)
27. overrun-local: Overrunning array of 6 8-byte elements at element
index 6 (byte offset 55) by dereferencing pointer unitprefix + post_index.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This patch does not change any functionality.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This patch reduces the size of the buffer allocated by num2str().
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
* 'pmemblk' of https://github.com/bvanassche/fio:
Unbreak the pmemblk engine
|
|
Reported-by: Yi Zhang <yi.zhang@redhat.com>
Tested-by: Yi Zhang <yi.zhang@redhat.com>
Fixes: e9c7be0e32e6 ("pmemblk: Fix a memory leak")
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
* 'compiler' of https://github.com/bvanassche/fio:
Merge compiler/gcc4.h into compiler/compiler.h
|
|
Follow the example of the Linux kernel and merge the file compiler/gcc4.h
into compiler/gcc.h. Additionally, enable __must_check,
__compiletime_warning() and __compiletime_error() for clang.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
* 'master' of https://github.com/safl/fio:
Changed signedness of seqlock.sequence fixing comparison-warning
|
|
Signed-off-by: Simon A. F. Lund <simon.lund@samsung.com>
|
|
Fixes: baa16486051f ("compiler/compiler.h: minimum GCC version is 4.9")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
GCC 10.x reports:
oslib/linux-blkzoned.c: In function ‘blkzoned_report_zones’:
oslib/linux-blkzoned.c:146:18: warning: taking address of packed member of ‘struct blk_zone_report’ may result in an unaligned pointer value [-Waddress-of-packed-member]
146 | blkz = (void *) &hdr->zones[0];
which is totally fine, but we can easily avoid this warning by just
setting the 'blkz' pointer to the end of 'hdr'.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
* 'atomics' of https://github.com/bvanassche/fio:
Optimize fio_gettime_offload()
Optimize the seqlock implementation
Add a test that sets gtod_cpu=1
t/run-fio-tests.py: Increase IOPS tolerance further
fio: Use atomic_load_acquire() and atomic_store_release() where appropriate
engines/io_uring: Use atomic_{load_acquire,store_release}()
engines/libaio: Use atomic_store_release() instead of read_barrier()
arch/arch.h: Introduce atomic_{load_acquire,store_release}()
fio_sem: Remove a read_barrier() call
Make __rand_0_1() compatible with clang
configure: Use -Wimplicit-fallthrough=2 instead of -Wimplicit-fallthrough=3
|
|
This patch not only an optimization but also a bug fix because it guarantees
that fio_gettime_offload() never returns an inconsistent timestamp, something
that could happen before. In this context inconsistent means returning the
.tv_sec value from one gettimeofday() call and the .tv_usec value from another
gettimeofday() call.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Use atomic_load_acquire() and atomic_store_release() instead of barriers
where appropriate.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
None of the existing tests sets gtod_cpu=1. Hence add a test that sets
gtod_cpu=1.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
The write_barrier() in io_completed() and also the read barriers in
verify.c are misplaced: the write barrier should occur before the flags
update instead of after and the read barriers should occur after the flags
read instead of before. Fix this by using atomic_{load_acquire,
store_release}() instead of read and write barriers.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This patch improves performance by using acquire and release semantics
instead of barriers and also brings the io_uring engine code closer to that
of liburing.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This is a micro-optimization that does not affect the functionality of the
code.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Implement atomic_load_acquire() and atomic_store_release() with C11
atomic operations. These two primitives will be used in later patches.
This patch increases the minimum requirement for the compiler fio is
built with from C90 to C11.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Since pthread_mutex_lock() already provides load-acquire semantics, calling
read_barrier() just after pthread_mutex_lock() is not necessary. See also
commit 4d4e80f2b426 ("Revamp file locking").
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This patch fixes the following clang compiler error:
crc/../arch/../lib/rand.h:109:25: error: implicit conversion from
'unsigned long long' to 'double' changes value from 18446744073709551615
to 18446744073709551616 [-Werror,-Wimplicit-int-float-conversion]
return (val + 1.0) / (FRAND64_MAX + 1.0);
^~~~~~~~~~~ ~
crc/../arch/../lib/rand.h:9:22: note: expanded from macro 'FRAND64_MAX'
#define FRAND64_MAX (-1ULL)
^~~~~
Fixes: e7b240474543 ("Fixups for poisson rate")
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
This makes clang accept /* fallthrough */ as a fallthrough annotation.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
|
|
Updates the description of log file formats to include and
describe the priority bit entry.
Signed-off-by: Phillip Chen <phillip.a.chen@seagate.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
