37f9dac592
Internally, zvols are files exposed through the block device API. This is intended to reduce overhead when things require block devices. However, the ZoL zvol code emulates a traditional block device in that it has a top half and a bottom half. This is an unnecessary source of overhead that does not exist on any other OpenZFS platform does this. This patch removes it. Early users of this patch reported double digit performance gains in IOPS on zvols in the range of 50% to 80%. Comments in the code suggest that the current implementation was done to obtain IO merging from Linux's IO elevator. However, the DMU already does write merging while arc_read() should implicitly merge read IOs because only 1 thread is permitted to fetch the buffer into ARC. In addition, commercial ZFSOnLinux distributions report that regular files are more performant than zvols under the current implementation, and the main consumers of zvols are VMs and iSCSI targets, which have their own elevators to merge IOs. Some minor refactoring allows us to register zfs_request() as our ->make_request() handler in place of the generic_make_request() function. This eliminates the layer of code that broke IO requests on zvols into a top half and a bottom half. This has several benefits: 1. No per zvol spinlocks. 2. No redundant IO elevator processing. 3. Interrupts are disabled only when actually necessary. 4. No redispatching of IOs when all taskq threads are busy. 5. Linux's page out routines will properly block. 6. Many autotools checks become obsolete. An unfortunate consequence of eliminating the layer that generic_make_request() is that we no longer calls the instrumentation hooks for block IO accounting. Those hooks are GPL-exported, so we cannot call them ourselves and consequently, we lose the ability to do IO monitoring via iostat. Since zvols are internally files mapped as block devices, this should be okay. Anyone who is willing to accept the performance penalty for the block IO layer's accounting could use the loop device in between the zvol and its consumer. Alternatively, perf and ftrace likely could be used. Also, tools like latencytop will still work. Tools such as latencytop sometimes provide a better view of performance bottlenecks than the traditional block IO accounting tools do. Lastly, if direct reclaim occurs during spacemap loading and swap is on a zvol, this code will deadlock. That deadlock could already occur with sync=always on zvols. Given that swap on zvols is not yet production ready, this is not a blocker. Signed-off-by: Richard Yao <ryao@gentoo.org>
26 lines
844 B
Plaintext
26 lines
844 B
Plaintext
dnl #
|
|
dnl # Interface for issuing a discard bio:
|
|
dnl # 2.6.28-2.6.35: BIO_RW_DISCARD
|
|
dnl # 2.6.36-3.x: REQ_DISCARD
|
|
dnl #
|
|
|
|
dnl # Since REQ_DISCARD is a preprocessor definition, there is no need for an
|
|
dnl # autotools check for it. Also, REQ_DISCARD existed in the request layer
|
|
dnl # until torvalds/linux@7b6d91daee5cac6402186ff224c3af39d79f4a0e unified the
|
|
dnl # request layer and bio layer flags, so it would be wrong to assume that
|
|
dnl # the APIs are mutually exclusive contrary to the typical case.
|
|
AC_DEFUN([ZFS_AC_KERNEL_BIO_RW_DISCARD], [
|
|
AC_MSG_CHECKING([whether BIO_RW_DISCARD is defined])
|
|
ZFS_LINUX_TRY_COMPILE([
|
|
#include <linux/bio.h>
|
|
],[
|
|
int flags __attribute__ ((unused));
|
|
flags = BIO_RW_DISCARD;
|
|
],[
|
|
AC_MSG_RESULT(yes)
|
|
AC_DEFINE(HAVE_BIO_RW_DISCARD, 1, [BIO_RW_DISCARD is defined])
|
|
],[
|
|
AC_MSG_RESULT(no)
|
|
])
|
|
])
|