Message ID | 1372444009-11544-6-git-send-email-pbonzini@redhat.com |
---|---|
State | New |
Headers | show |
Paolo Bonzini <pbonzini@redhat.com> writes: > The next patch will change qemu/tls.h to support more platforms, but at > some performance cost. Declare cpu_single_env directly instead of using > the tls.h abstractions. > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Reviewed-by: Anthony Liguori <aliguori@us.ibm.com> Regards, Anthony Liguori > --- > exec.c | 10 ++++++++-- > include/exec/cpu-all.h | 14 +++++++++++--- > include/qemu/tls.h | 52 -------------------------------------------------- > 3 files changed, 19 insertions(+), 57 deletions(-) > delete mode 100644 include/qemu/tls.h > > diff --git a/exec.c b/exec.c > index d28403b..a788981 100644 > --- a/exec.c > +++ b/exec.c > @@ -70,9 +70,15 @@ static MemoryRegion io_mem_unassigned; > #endif > > CPUArchState *first_cpu; > + > /* current CPU in the current thread. It is only valid inside > - cpu_exec() */ > -DEFINE_TLS(CPUArchState *,cpu_single_env); > + * cpu_exec(). See comment in include/exec/cpu-all.h. */ > +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) > +__thread CPUArchState *cpu_single_env; > +#else > +CPUArchState *cpu_single_env; > +#endif > + > /* 0 = Do not count executed instructions. > 1 = Precise instruction counting. > 2 = Adaptive rate instruction counting. */ > diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h > index e9c3717..2202ba3 100644 > --- a/include/exec/cpu-all.h > +++ b/include/exec/cpu-all.h > @@ -20,7 +20,6 @@ > #define CPU_ALL_H > > #include "qemu-common.h" > -#include "qemu/tls.h" > #include "exec/cpu-common.h" > #include "qemu/thread.h" > > @@ -368,8 +367,17 @@ void cpu_dump_statistics(CPUArchState *env, FILE *f, fprintf_function cpu_fprint > void QEMU_NORETURN cpu_abort(CPUArchState *env, const char *fmt, ...) > GCC_FMT_ATTR(2, 3); > extern CPUArchState *first_cpu; > -DECLARE_TLS(CPUArchState *,cpu_single_env); > -#define cpu_single_env tls_var(cpu_single_env) > + > +/* This is thread-local depending on __linux__ because: > + * - the only -user mode supporting multiple VCPU threads is linux-user > + * - TCG system mode is single-threaded regarding VCPUs > + * - KVM system mode is multi-threaded but limited to Linux > + */ > +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) > +extern __thread CPUArchState *cpu_single_env; > +#else > +extern CPUArchState *cpu_single_env; > +#endif > > /* Flags for use in ENV->INTERRUPT_PENDING. > > diff --git a/include/qemu/tls.h b/include/qemu/tls.h > deleted file mode 100644 > index b92ea9d..0000000 > --- a/include/qemu/tls.h > +++ /dev/null > @@ -1,52 +0,0 @@ > -/* > - * Abstraction layer for defining and using TLS variables > - * > - * Copyright (c) 2011 Red Hat, Inc > - * Copyright (c) 2011 Linaro Limited > - * > - * Authors: > - * Paolo Bonzini <pbonzini@redhat.com> > - * Peter Maydell <peter.maydell@linaro.org> > - * > - * This program is free software; you can redistribute it and/or > - * modify it under the terms of the GNU General Public License as > - * published by the Free Software Foundation; either version 2 of > - * the License, or (at your option) any later version. > - * > - * This program is distributed in the hope that it will be useful, > - * but WITHOUT ANY WARRANTY; without even the implied warranty of > - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > - * GNU General Public License for more details. > - * > - * You should have received a copy of the GNU General Public License along > - * with this program; if not, see <http://www.gnu.org/licenses/>. > - */ > - > -#ifndef QEMU_TLS_H > -#define QEMU_TLS_H > - > -/* Per-thread variables. Note that we only have implementations > - * which are really thread-local on Linux; the dummy implementations > - * define plain global variables. > - * > - * This means that for the moment use should be restricted to > - * per-VCPU variables, which are OK because: > - * - the only -user mode supporting multiple VCPU threads is linux-user > - * - TCG system mode is single-threaded regarding VCPUs > - * - KVM system mode is multi-threaded but limited to Linux > - * > - * TODO: proper implementations via Win32 .tls sections and > - * POSIX pthread_getspecific. > - */ > -#ifdef __linux__ > -#define DECLARE_TLS(type, x) extern DEFINE_TLS(type, x) > -#define DEFINE_TLS(type, x) __thread __typeof__(type) tls__##x > -#define tls_var(x) tls__##x > -#else > -/* Dummy implementations which define plain global variables */ > -#define DECLARE_TLS(type, x) extern DEFINE_TLS(type, x) > -#define DEFINE_TLS(type, x) __typeof__(type) tls__##x > -#define tls_var(x) tls__##x > -#endif > - > -#endif > -- > 1.8.1.4
On 28 June 2013 14:26, Paolo Bonzini <pbonzini@redhat.com> wrote: > > +/* This is thread-local depending on __linux__ because: Is the comment perhaps unchanged from an earlier revision that used a different test? It seems odd to me to reference __linux__ here. > + * - the only -user mode supporting multiple VCPU threads is linux-user > + * - TCG system mode is single-threaded regarding VCPUs > + * - KVM system mode is multi-threaded but limited to Linux > + */ > +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) Also, in discussion on the FreeBSD bsd-user patch set the suggestion was made that we do away with a flag, and just have thread support always enabled. Would you suggest this test then become KVM || (USER_ONLY && (USE_NPTL || __FreeBSD__))?
On 28 June 2013 19:26, Paolo Bonzini <pbonzini@redhat.com> wrote: > The next patch will change qemu/tls.h to support more platforms, but at > some performance cost. Declare cpu_single_env directly instead of using > the tls.h abstractions. > > Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> > --- > exec.c | 10 ++++++++-- > include/exec/cpu-all.h | 14 +++++++++++--- > include/qemu/tls.h | 52 -------------------------------------------------- > 3 files changed, 19 insertions(+), 57 deletions(-) > delete mode 100644 include/qemu/tls.h > > diff --git a/exec.c b/exec.c > index d28403b..a788981 100644 > --- a/exec.c > +++ b/exec.c > @@ -70,9 +70,15 @@ static MemoryRegion io_mem_unassigned; > #endif > > CPUArchState *first_cpu; > + > /* current CPU in the current thread. It is only valid inside > - cpu_exec() */ > -DEFINE_TLS(CPUArchState *,cpu_single_env); > + * cpu_exec(). See comment in include/exec/cpu-all.h. */ > +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) > +__thread CPUArchState *cpu_single_env; > +#else > +CPUArchState *cpu_single_env; > +#endif I don't like having the semantics of this variable differ depending on whether CONFIG_KVM was defined. In particular this means that the variable is per-thread if you're running TCG on a QEMU that was configured with KVM support, but not per-thread if you're running TCG on a QEMU that was configured without per-thread support. That's just bizarre and a recipe for confusion and for bugs creeping in in the less-well-tested config combinations. We should just be consistent and always make this be per-thread. thanks -- PMM
Il 29/06/2013 01:53, Ed Maste ha scritto: > On 28 June 2013 14:26, Paolo Bonzini <pbonzini@redhat.com> wrote: >> >> +/* This is thread-local depending on __linux__ because: > > Is the comment perhaps unchanged from an earlier revision that used a > different test? It seems odd to me to reference __linux__ here. > >> + * - the only -user mode supporting multiple VCPU threads is linux-user >> + * - TCG system mode is single-threaded regarding VCPUs >> + * - KVM system mode is multi-threaded but limited to Linux >> + */ >> +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) > > Also, in discussion on the FreeBSD bsd-user patch set the suggestion > was made that we do away with a flag, and just have thread support > always enabled. Would you suggest this test then become KVM || > (USER_ONLY && (USE_NPTL || __FreeBSD__))? I would suggest that you have something like CONFIG_USER_THREADS that can be used by both linux-user and bsd-user. Paolo
Il 29/06/2013 12:55, Peter Maydell ha scritto: > On 28 June 2013 19:26, Paolo Bonzini <pbonzini@redhat.com> wrote: >> The next patch will change qemu/tls.h to support more platforms, but at >> some performance cost. Declare cpu_single_env directly instead of using >> the tls.h abstractions. >> >> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> >> --- >> exec.c | 10 ++++++++-- >> include/exec/cpu-all.h | 14 +++++++++++--- >> include/qemu/tls.h | 52 -------------------------------------------------- >> 3 files changed, 19 insertions(+), 57 deletions(-) >> delete mode 100644 include/qemu/tls.h >> >> diff --git a/exec.c b/exec.c >> index d28403b..a788981 100644 >> --- a/exec.c >> +++ b/exec.c >> @@ -70,9 +70,15 @@ static MemoryRegion io_mem_unassigned; >> #endif >> >> CPUArchState *first_cpu; >> + >> /* current CPU in the current thread. It is only valid inside >> - cpu_exec() */ >> -DEFINE_TLS(CPUArchState *,cpu_single_env); >> + * cpu_exec(). See comment in include/exec/cpu-all.h. */ >> +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) >> +__thread CPUArchState *cpu_single_env; >> +#else >> +CPUArchState *cpu_single_env; >> +#endif > > I don't like having the semantics of this variable differ > depending on whether CONFIG_KVM was defined. In particular > this means that the variable is per-thread if you're running > TCG on a QEMU that was configured with KVM support, but > not per-thread if you're running TCG on a QEMU that was > configured without per-thread support. That's just bizarre > and a recipe for confusion and for bugs creeping in in the > less-well-tested config combinations. > > We should just be consistent and always make this be > per-thread. If it's okay to make cpu_single_env accesses more expensive by a factor of 4 on TLS-deficient hosts (at least OpenBSD; do Darwin and NetBSD support thread-local storage?), I'm all for it. I (and I guess Stefan too) do not want to introduce performance regressions in these patches. Making it simpler is something that one would do after having tested at least one of OpenBSD/Darwin/whatever. This patch does not make things worse than before. If anything, it's better because *more* targets have non-TLS semantics: namely non-KVM targets on Linux become non-TLS. Paolo
On 1 July 2013 11:45, Paolo Bonzini <pbonzini@redhat.com> wrote: > Il 29/06/2013 12:55, Peter Maydell ha scritto: >> We should just be consistent and always make this be >> per-thread. > > If it's okay to make cpu_single_env accesses more expensive by a factor > of 4 on TLS-deficient hosts (at least OpenBSD; do Darwin and NetBSD > support thread-local storage?), I'm all for it. I (and I guess Stefan > too) do not want to introduce performance regressions in these patches. > Making it simpler is something that one would do after having tested at > least one of OpenBSD/Darwin/whatever. MacOSX 10.8 supports __thread if you build with clang (not if you build with gcc), though I can't speak to its performance since it compiles to a function call to get the variable's address. > This patch does not make things worse than before. If anything, it's > better because *more* targets have non-TLS semantics: namely non-KVM > targets on Linux become non-TLS. That is making things worse! Non-TLS is the untested minority case, we want to be taking things out of it, not pushing configs that were previously TLS into it. -- PMM
Il 01/07/2013 13:05, Peter Maydell ha scritto: >> > This patch does not make things worse than before. If anything, it's >> > better because *more* targets have non-TLS semantics: namely non-KVM >> > targets on Linux become non-TLS. > That is making things worse! Non-TLS is the untested minority > case, we want to be taking things out of it, not pushing > configs that were previously TLS into it. I think we should strive for one of these two: (1) all targets are TLS; (2) all targets are non-TLS if this is possible. Either maximizes the homogeneity across platforms. Paolo
On 1 July 2013 17:21, Paolo Bonzini <pbonzini@redhat.com> wrote: > I think we should strive for one of these two: > > (1) all targets are TLS; > > (2) all targets are non-TLS if this is possible. > > Either maximizes the homogeneity across platforms. Since the two largest cases are both "cpu_single_env must be TLS" (ie (a) system emulation built with KVM support and (b) linux-user), the set of targets which can be non-TLS is really really small, and I think (1) makes much more sense. (I'm assuming you don't want to try to support cpu_single_env being both per-thread and not-per-thread in a single binary depending on whether the user passes -enable-kvm or not.) -- PMM
Il 01/07/2013 18:26, Peter Maydell ha scritto: > > I think we should strive for one of these two: > > > > (1) all targets are TLS; > > > > (2) all targets are non-TLS if this is possible. > > > > Either maximizes the homogeneity across platforms. > > Since the two largest cases are both "cpu_single_env must be TLS" > (ie (a) system emulation built with KVM support and (b) linux-user), > the set of targets which can be non-TLS is really really small, > and I think (1) makes much more sense. Not many linux-user targets support threads (including not i386). > (I'm assuming you don't want to try to support cpu_single_env > being both per-thread and not-per-thread in a single binary > depending on whether the user passes -enable-kvm or not.) No, of course not.
On 1 July 2013 21:52, Paolo Bonzini <pbonzini@redhat.com> wrote: > Il 01/07/2013 18:26, Peter Maydell ha scritto: >> Since the two largest cases are both "cpu_single_env must be TLS" >> (ie (a) system emulation built with KVM support and (b) linux-user), >> the set of targets which can be non-TLS is really really small, >> and I think (1) makes much more sense. > > Not many linux-user targets support threads (including not i386). i386 guest is a comparatively rare case for linux-user (because most people have an i386 box they can run them on). Also I'm hoping to get most of the linux-user guests up to the point where we can just have CONFIG_NPTL be true for all of them -- I have several patches on-list which are trying to head in that direction. thanks -- PMM
Am 01.07.2013 23:34, schrieb Peter Maydell: > On 1 July 2013 21:52, Paolo Bonzini <pbonzini@redhat.com> wrote: >> Il 01/07/2013 18:26, Peter Maydell ha scritto: >>> Since the two largest cases are both "cpu_single_env must be TLS" >>> (ie (a) system emulation built with KVM support and (b) linux-user), >>> the set of targets which can be non-TLS is really really small, >>> and I think (1) makes much more sense. >> >> Not many linux-user targets support threads (including not i386). > > i386 guest is a comparatively rare case for linux-user (because > most people have an i386 box they can run them on). Also I'm > hoping to get most of the linux-user guests up to the point where > we can just have CONFIG_NPTL be true for all of them -- I have > several patches on-list which are trying to head in that direction. Alex had posted a patch which implements NPTL for i386, tested with WINE, I believe. Andreas
On 07/02/2013 03:40 PM, Andreas Färber wrote: > Am 01.07.2013 23:34, schrieb Peter Maydell: >> On 1 July 2013 21:52, Paolo Bonzini<pbonzini@redhat.com> wrote: >>> Il 01/07/2013 18:26, Peter Maydell ha scritto: >>>> Since the two largest cases are both "cpu_single_env must be TLS" >>>> (ie (a) system emulation built with KVM support and (b) linux-user), >>>> the set of targets which can be non-TLS is really really small, >>>> and I think (1) makes much more sense. >>> Not many linux-user targets support threads (including not i386). >> i386 guest is a comparatively rare case for linux-user (because >> most people have an i386 box they can run them on). Also I'm >> hoping to get most of the linux-user guests up to the point where >> we can just have CONFIG_NPTL be true for all of them -- I have >> several patches on-list which are trying to head in that direction. > Alex had posted a patch which implements NPTL for i386, tested with > WINE, I believe. Not sure I posted it, but I do have it ready. I'll rebase it once the NPTL refactorings are in. They look very sane to me. Alex
diff --git a/exec.c b/exec.c index d28403b..a788981 100644 --- a/exec.c +++ b/exec.c @@ -70,9 +70,15 @@ static MemoryRegion io_mem_unassigned; #endif CPUArchState *first_cpu; + /* current CPU in the current thread. It is only valid inside - cpu_exec() */ -DEFINE_TLS(CPUArchState *,cpu_single_env); + * cpu_exec(). See comment in include/exec/cpu-all.h. */ +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) +__thread CPUArchState *cpu_single_env; +#else +CPUArchState *cpu_single_env; +#endif + /* 0 = Do not count executed instructions. 1 = Precise instruction counting. 2 = Adaptive rate instruction counting. */ diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h index e9c3717..2202ba3 100644 --- a/include/exec/cpu-all.h +++ b/include/exec/cpu-all.h @@ -20,7 +20,6 @@ #define CPU_ALL_H #include "qemu-common.h" -#include "qemu/tls.h" #include "exec/cpu-common.h" #include "qemu/thread.h" @@ -368,8 +367,17 @@ void cpu_dump_statistics(CPUArchState *env, FILE *f, fprintf_function cpu_fprint void QEMU_NORETURN cpu_abort(CPUArchState *env, const char *fmt, ...) GCC_FMT_ATTR(2, 3); extern CPUArchState *first_cpu; -DECLARE_TLS(CPUArchState *,cpu_single_env); -#define cpu_single_env tls_var(cpu_single_env) + +/* This is thread-local depending on __linux__ because: + * - the only -user mode supporting multiple VCPU threads is linux-user + * - TCG system mode is single-threaded regarding VCPUs + * - KVM system mode is multi-threaded but limited to Linux + */ +#if defined CONFIG_KVM || (defined CONFIG_USER_ONLY && defined CONFIG_USE_NPTL) +extern __thread CPUArchState *cpu_single_env; +#else +extern CPUArchState *cpu_single_env; +#endif /* Flags for use in ENV->INTERRUPT_PENDING. diff --git a/include/qemu/tls.h b/include/qemu/tls.h deleted file mode 100644 index b92ea9d..0000000 --- a/include/qemu/tls.h +++ /dev/null @@ -1,52 +0,0 @@ -/* - * Abstraction layer for defining and using TLS variables - * - * Copyright (c) 2011 Red Hat, Inc - * Copyright (c) 2011 Linaro Limited - * - * Authors: - * Paolo Bonzini <pbonzini@redhat.com> - * Peter Maydell <peter.maydell@linaro.org> - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License as - * published by the Free Software Foundation; either version 2 of - * the License, or (at your option) any later version. - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License along - * with this program; if not, see <http://www.gnu.org/licenses/>. - */ - -#ifndef QEMU_TLS_H -#define QEMU_TLS_H - -/* Per-thread variables. Note that we only have implementations - * which are really thread-local on Linux; the dummy implementations - * define plain global variables. - * - * This means that for the moment use should be restricted to - * per-VCPU variables, which are OK because: - * - the only -user mode supporting multiple VCPU threads is linux-user - * - TCG system mode is single-threaded regarding VCPUs - * - KVM system mode is multi-threaded but limited to Linux - * - * TODO: proper implementations via Win32 .tls sections and - * POSIX pthread_getspecific. - */ -#ifdef __linux__ -#define DECLARE_TLS(type, x) extern DEFINE_TLS(type, x) -#define DEFINE_TLS(type, x) __thread __typeof__(type) tls__##x -#define tls_var(x) tls__##x -#else -/* Dummy implementations which define plain global variables */ -#define DECLARE_TLS(type, x) extern DEFINE_TLS(type, x) -#define DEFINE_TLS(type, x) __typeof__(type) tls__##x -#define tls_var(x) tls__##x -#endif - -#endif
The next patch will change qemu/tls.h to support more platforms, but at some performance cost. Declare cpu_single_env directly instead of using the tls.h abstractions. Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> --- exec.c | 10 ++++++++-- include/exec/cpu-all.h | 14 +++++++++++--- include/qemu/tls.h | 52 -------------------------------------------------- 3 files changed, 19 insertions(+), 57 deletions(-) delete mode 100644 include/qemu/tls.h