debug: add PR_SET_VMA_ANON_NAME support
v2 cover debug: add PR_SET_VMA_ANON_NAME support
Add PR_SET_VMA_ANON_NAME support to name mmap'd regions of memory. This is only supported on Linux >= 5.17, but allows QEMU to communicate names for each VMA segment to linux, where they are visible in /proc/$pid/[s]maps and utilities like pmap.
v2:
- add va_end()
- fix commit message on 1/7
- remove redundant call in 4/7
- remove RFC comment in 5/7
John Snow (7):
debug/mem: add utility to name anonymous memory mappings debug/mem: add tracing to qemu_vname_vma() debug/mem: name <reserved> regions in mmap-alloc debug/mem: name pthreads
debug/mem: name coroutine stacks
debug/mem: name qemu_anon_ram_alloc/qemu_ram_mmap pages debug/mem: HACK add qemu_name_buffer
meson.build | 8 ++++ include/qemu/bitmap.h | 6 +++ include/qemu/mmap-alloc.h | 3 +- include/qemu/mmap-vma.h | 21 ++++++++ include/qemu/osdep.h | 2 +- hw/core/loader.c | 3 ++ hw/display/virtio-gpu-gl.c | 2 +- hw/i386/x86-common.c | 3 ++ migration/ram.c | 3 +- system/device_tree.c | 2 + system/physmem.c | 15 ++++-- target/i386/kvm/tdx.c | 3 +- util/mmap-alloc.c | 17 +++++-- util/mmap-vma.c | 98 ++++++++++++++++++++++++++++++++++++++ util/oslib-posix.c | 7 ++- util/oslib-win32.c | 2 +- util/qemu-thread-posix.c | 30 ++++++++++++ util/meson.build | 1 + util/trace-events | 3 ++ 19 files changed, 213 insertions(+), 16 deletions(-) create mode 100644 include/qemu/mmap-vma.h create mode 100644 util/mmap-vma.c
Signature · 1 line
2.56.1
PATCH v2 1/7 debug/mem: add utility to name anonymous memory mappings
Add support to name mmap'd memory regions using PR_SET_VMA_ANON_NAME. On unsupported platforms, these QEMU utility functions are no-ops.
This is useful primarily for debugging and profiling QEMU; being able to see and correlate each VMA segment with a specific use within QEMU.
Support for PR_SET_VMA_ANON_NAME was added in Linux 5.17, and requires CONFIG_ANON_VMA_NAME=y to be set in the kernel configuration. Fedora (and RHEL) do not enable this feature by default. See also
https://man7.org/linux/man-pages/man2/pr_set_vma.2const.html
--- include/qemu/mmap-vma.h | 18 +++++++++++ util/mmap-vma.c | 71 +++++++++++++++++++++++++++++++++++++++++ util/meson.build | 1 + 3 files changed, 90 insertions(+) create mode 100644 include/qemu/mmap-vma.h create mode 100644 util/mmap-vma.c
include/qemu/mmap-vma.h new file +18 -0
@@ -0,0 +1,18 @@ +/* +* Utility for naming VMA allocations in linux +* +* Copyright (c) Red Hat, Inc 2026 +* +* SPDX-License-Identifier: LGPL-2.1-or-later +*/ + +#ifndef __QEMU_MMAP_VMA_H +#define __QEMU_MMAP_VMA_H + +#include <stddef.h> +#include "osdep.h" + +int qemu_name_vma(void *ptr, size_t sz, const char *fmt, ...) + G_GNUC_PRINTF(3, 4); + +#endif
util/mmap-vma.c new file +71 -0
71 lines not on this page · raw message
util/meson.build +1 -0
@@ -1,5 +1,6 @@ util_ss.add(files('osdep.c', 'cutils.c', 'unicode.c', 'qemu-timer-common.c')) util_ss.add(files('thread-context.c'), numa) +util_ss.add(files('mmap-vma.c')) if host_os != 'windows' util_ss.add(files('aio-posix.c')) util_ss.add(files('fdmon-poll.c'))
Signature · 1 line
2.56.1
Re: [PATCH v2 1/7] debug/mem: add utility to name anonymous memory mappings
Quoted 75 lines · John Snow
Add support to name mmap'd memory regions using
PR_SET_VMA_ANON_NAME. On unsupported platforms, these QEMU utility functions are no-ops.This is useful primarily for debugging and profiling QEMU; being able to see and correlate each VMA segment with a specific use within QEMU.
Support for
PR_SET_VMA_ANON_NAMEwas added in Linux 5.17, and requiresCONFIG_ANON_VMA_NAME=yto be set in the kernel configuration. Fedora (and RHEL) do not enable this feature by default. See also
https://man7.org/linux/man-pages/man2/pr_set_vma.2const.htmlSigned-off-by John Snow <[email protected]>--- include/qemu/mmap-vma.h | 18 +++++++++++ util/mmap-vma.c | 71 +++++++++++++++++++++++++++++++++++++++++ util/meson.build | 1 + 3 files changed, 90 insertions(+) create mode 100644 include/qemu/mmap-vma.h create mode 100644 util/mmap-vma.cinclude/qemu/mmap-vma.h new file +18
-0@@ -0,0 +1,18 @@ +/* +* Utility for naming VMA allocations in linux +* +* Copyright (c) Red Hat, Inc 2026 +* +* SPDX-License-Identifier: LGPL-2.1-or-later +*/ + +#ifndef __QEMU_MMAP_VMA_H +#define __QEMU_MMAP_VMA_H + +#include <stddef.h> +#include "osdep.h" + +int qemu_name_vma(void *ptr, size_t sz, const char *fmt, ...) + G_GNUC_PRINTF(3, 4); + +#endif util/mmap-vma.c new file +24
-0@@ -0,0 +1,71 @@ +/* +* Utility for naming VMA allocations in linux +* +* Copyright (c) Red Hat, Inc 2026 +* +* SPDX-License-Identifier: LGPL-2.1-or-later +*/ +#include "qemu/mmap-vma.h" +#include "qemu/osdep.h" + +#include <errno.h> +#include <stddef.h> +#include <stdio.h> + +#if defined(__linux__) +#include <sys/prctl.h> +#endif + +#define VMA_PREFIX " qemu: " +#define VMA_MAXLEN 80 + +/* +* Give a human-readable name to a previously mmap'd memory segment. +* This feature is only available on Linux >= 5.17, and only when the
util/mmap-vma.c line 34
Quoted 10 lines
util/mmap-vma.c new file +10
-0+* kernel was configured with CONFIG_ANON_VMA_NAME=y. +* +* For systems without this option, prctl() will return -1 and set +* errno to EINVAL, and can be safely ignored. +* +* For systems with this option, these human-readable names will be +* visible in /proc/$pid/[s]maps, pmap output, etc. +*/ +G_GNUC_PRINTF(3, 0) +static int qemu_vname_vma (void *ptr,size_t sz,const char *fmt,va_list args)
As the comment says above, we can't assume the kernel will have the feature enabled.
It is harmless if we try to set the VMA name and it fails though. We merely loose a bit of debug info. Thus I'd suggest we make the function return type 'void', so that it is clear that we do NOT expect callers to check for errors.
Quoted 51 lines
util/mmap-vma.c new file +37
-0+{ +int rc =0 ; + +#if defined(PR_SET_VMA_ANON_NAME ) +char strbuf[VMA_MAXLEN ]; + +assert (sizeof (VMA_PREFIX ) <sizeof (strbuf)); +strncpy (strbuf,VMA_PREFIX ,sizeof (strbuf)); + rc =vsnprintf (strbuf +sizeof (VMA_PREFIX ), +sizeof (strbuf) -sizeof (VMA_PREFIX ), fmt, args); +if (rc <0 ) { +return -errno; + }else if (rc >= (sizeof (strbuf) -sizeof (VMA_PREFIX ))) { +sprintf (strbuf +sizeof (strbuf) -4 ,"..." ); + } + + rc =prctl (PR_SET_VMA ,PR_SET_VMA_ANON_NAME , ptr, sz, strbuf); +if (rc) { +return -errno; + } +#endif + +return rc; +} + +G_GNUC_PRINTF(3, 4) +int qemu_name_vma (void *ptr,size_t sz,const char *fmt, ...) +{ +int rc; +va_list args; + +va_start (args, fmt); + rc =qemu_vname_vma (ptr, sz, fmt, args); +va_end (args); + +return rc; +}util/meson.build +1
-0@@ -1,5 +1,6 @@ util_ss.add(files('osdep.c', 'cutils.c', 'unicode.c', 'qemu-timer-common.c')) util_ss.add(files('thread-context.c'), numa) +util_ss.add(files('mmap-vma.c')) if host_os != 'windows' util_ss.add(files('aio-posix.c')) util_ss.add(files('fdmon-poll.c'))Signature · 1 line
2.56.1
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
Re: [PATCH v2 1/7] debug/mem: add utility to name anonymous memory mappings
Quoted 91 lines · Daniel P. Berrangé
Quoted 75 lines · John Snow
Add support to name mmap'd memory regions using
PR_SET_VMA_ANON_NAME. On unsupported platforms, these QEMU utility functions are no-ops.This is useful primarily for debugging and profiling QEMU; being able to see and correlate each VMA segment with a specific use within QEMU.
Support for
PR_SET_VMA_ANON_NAMEwas added in Linux 5.17, and requiresCONFIG_ANON_VMA_NAME=yto be set in the kernel configuration. Fedora (and RHEL) do not enable this feature by default. See also
https://man7.org/linux/man-pages/man2/pr_set_vma.2const.htmlSigned-off-by John Snow <[email protected]>--- include/qemu/mmap-vma.h | 18 +++++++++++ util/mmap-vma.c | 71 +++++++++++++++++++++++++++++++++++++++++ util/meson.build | 1 + 3 files changed, 90 insertions(+) create mode 100644 include/qemu/mmap-vma.h create mode 100644 util/mmap-vma.cinclude/qemu/mmap-vma.h new file +18
-0@@ -0,0 +1,18 @@ +/* +* Utility for naming VMA allocations in linux +* +* Copyright (c) Red Hat, Inc 2026 +* +* SPDX-License-Identifier: LGPL-2.1-or-later +*/ + +#ifndef __QEMU_MMAP_VMA_H +#define __QEMU_MMAP_VMA_H + +#include <stddef.h> +#include "osdep.h" + +int qemu_name_vma(void *ptr, size_t sz, const char *fmt, ...) + G_GNUC_PRINTF(3, 4); + +#endif util/mmap-vma.c new file +24
-0@@ -0,0 +1,71 @@ +/* +* Utility for naming VMA allocations in linux +* +* Copyright (c) Red Hat, Inc 2026 +* +* SPDX-License-Identifier: LGPL-2.1-or-later +*/ +#include "qemu/mmap-vma.h" +#include "qemu/osdep.h" + +#include <errno.h> +#include <stddef.h> +#include <stdio.h> + +#if defined(__linux__) +#include <sys/prctl.h> +#endif + +#define VMA_PREFIX " qemu: " +#define VMA_MAXLEN 80 + +/* +* Give a human-readable name to a previously mmap'd memory segment. +* This feature is only available on Linux >= 5.17, and only when the util/mmap-vma.c line 34
Quoted 10 lines
util/mmap-vma.c new file +10
-0+* kernel was configured with CONFIG_ANON_VMA_NAME=y. +* +* For systems without this option, prctl() will return -1 and set +* errno to EINVAL, and can be safely ignored. +* +* For systems with this option, these human-readable names will be +* visible in /proc/$pid/[s]maps, pmap output, etc. +*/ +G_GNUC_PRINTF(3, 0) +static int qemu_vname_vma (void *ptr,size_t sz,const char *fmt,va_list args)As the comment says above, we can't assume the kernel will have the feature enabled.
Quoted 5 lines
It is harmless if we try to set the VMA name and it fails though. We merely loose a bit of debug info. Thus I'd suggest we make the function return type 'void', so that it is clear that we do NOT expect callers to check for errors.
You're probably right, I think I was holding out for the possibility of callers being able to check if they got the memory region boundaries correct (in the case of naming ptmalloc buffers), but.... with the amount of extra work it'd take to make a robust system for identifying actual mmap regions possible, it'd be just as easy to add a simple int retcode back in, so... void for now it is, sure.
--js
Quoted 60 lines
Quoted 50 lines
util/mmap-vma.c new file +37
-0+{ +int rc =0 ; + +#if defined(PR_SET_VMA_ANON_NAME ) +char strbuf[VMA_MAXLEN ]; + +assert (sizeof (VMA_PREFIX ) <sizeof (strbuf)); +strncpy (strbuf,VMA_PREFIX ,sizeof (strbuf)); + rc =vsnprintf (strbuf +sizeof (VMA_PREFIX ), +sizeof (strbuf) -sizeof (VMA_PREFIX ), fmt, args); +if (rc <0 ) { +return -errno; + }else if (rc >= (sizeof (strbuf) -sizeof (VMA_PREFIX ))) { +sprintf (strbuf +sizeof (strbuf) -4 ,"..." ); + } + + rc =prctl (PR_SET_VMA ,PR_SET_VMA_ANON_NAME , ptr, sz, strbuf); +if (rc) { +return -errno; + } +#endif + +return rc; +} + +G_GNUC_PRINTF(3, 4) +int qemu_name_vma (void *ptr,size_t sz,const char *fmt, ...) +{ +int rc; +va_list args; + +va_start (args, fmt); + rc =qemu_vname_vma (ptr, sz, fmt, args); +va_end (args); + +return rc; +}util/meson.build +1
-1@@ -1,5 +1,6 @@ util_ss.add(files('osdep.c', 'cutils.c', 'unicode.c', 'qemu-timer-common.c')) util_ss.add(files('thread-context.c'), numa) +util_ss.add(files('mmap-vma.c')) if host_os != 'windows' util_ss.add(files('aio-posix.c')) util_ss.add(files('fdmon-poll.c')) --Quoted 2 lines
2.56.1
With regards,
Daniel
--|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
PATCH v2 2/7 debug/mem: add tracing to qemu_vname_vma()
In reply to John Snow--- util/mmap-vma.c | 6 +++++- util/trace-events | 3 +++ 2 files changed, 8 insertions(+), 1 deletion(-)
util/mmap-vma.c +5 -1
@@ -7,6 +7,7 @@*/ #include "qemu/mmap-vma.h" #include "qemu/osdep.h" +#include "trace.h" #include <errno.h> #include <stddef.h>
@@ -36,6 +37,7 @@ static int qemu_vname_vma(void *ptr, size_t sz, const char *fmt, va_list args)int rc =0 ;#if defined(PR_SET_VMA_ANON_NAME ) +int err;char strbuf[VMA_MAXLEN ];assert (sizeof (VMA_PREFIX ) <sizeof (strbuf));
@@ -49,8 +51,10 @@ static int qemu_vname_vma(void *ptr, size_t sz, const char *fmt, va_list args) } rc =prctl (PR_SET_VMA ,PR_SET_VMA_ANON_NAME , ptr, sz, strbuf); + err = errno; +trace_qemu_vname_vma (strbuf, ptr, sz, rc, err);if (rc) { -return -errno; +return -err; }#endif
util/trace-events +3 -0
@@ -117,3 +117,6 @@ module_lookup_object_type(const char *name) "name %s" # datadir.c load_file(const char *name, const char *path) "name %s location %s" + +# mmap-vma.c +qemu_vname_vma(const char *name, void *addr, size_t sz, int rc, int err) "name %s addr %p size %zu: rc %i errno %i"
Signature · 1 line
2.56.1
PATCH v2 3/7 debug/mem: name <reserved> regions in mmap-alloc
In reply to John SnowFor mmap memory reserved by QEMU, give it the name "<reserved>". This will show up in maps output as "anon: qemu: <reserved>".
--- util/mmap-alloc.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
util/mmap-alloc.c +6 -1
@@ -22,6 +22,7 @@#include "qemu/host-utils.h" #include "qemu/cutils.h" #include "qemu/error-report.h" +#include "qemu/mmap-vma.h" #define HUGETLBFS_MAGIC 0x958458f6
@@ -152,6 +153,7 @@ static bool map_noreserve_effective(int fd, uint32_t qemu_map_flags)static void *mmap_reserve (size_t size,int fd) {int flags =MAP_PRIVATE ; +void *ptr;#if defined(__powerpc64__) && defined(__linux__)/*
@@ -174,7 +176,10 @@ static void *mmap_reserve(size_t size, int fd) flags |=MAP_ANONYMOUS ;#endif -return mmap (0 , size,PROT_NONE , flags, fd,0 ); + ptr =mmap (0 , size,PROT_NONE , flags, fd,0 ); +qemu_name_vma (ptr, size,"<reserved>" ); + +return ptr; }/*
Signature · 1 line
2.56.1
PATCH v2 4/7 debug/mem: name pthreads
In reply to John Snow--- meson.build | 8 ++++++++ util/qemu-thread-posix.c | 30 ++++++++++++++++++++++++++++++ 2 files changed, 38 insertions(+)
meson.build +8 -0
@@ -2878,7 +2878,15 @@ config_host_data.set('CONFIG_PTHREAD_GET_NAME_NP', cc.links(osdep_prefix + ''' pthread_get_name_np(pthread_self(), buf, sizeof(buf)); return 0; }''', dependencies: threads)) +config_host_data.set('CONFIG_PTHREAD_GETATTR_NP', cc.links(osdep_prefix + ''' + #include <pthread.h> + int main(void) + { + pthread_attr_t attr; + pthread_getattr_np(pthread_self(), &attr); + return 0; + }''', dependencies: threads)) config_host_data.set('CONFIG_PTHREAD_CONDATTR_SETCLOCK', cc.links(osdep_prefix + ''' #include <pthread.h>
util/qemu-thread-posix.c +30 -0
@@ -17,6 +17,7 @@#include "qemu-thread-common.h" #include "qemu/tsan.h" #include "qemu/bitmap.h" +#include "qemu/mmap-vma.h" #if defined(CONFIG_PTHREAD_SET_NAME_NP ) || defined(CONFIG_PTHREAD_GET_NAME_NP )#include <pthread_np.h>
@@ -362,6 +363,13 @@ static void qemu_thread_atexit_notify(void *arg)void qemu_thread_set_name (const char *name) { +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +pthread_attr_t real_attr; +void *pthread_addr; +size_t stacksize; +size_t guardsize; +# endif +/* * Attempt to set the thread name; note that this is for debug, so * we're not going to fail if we can't set it.
@@ -373,6 +381,28 @@ void qemu_thread_set_name(const char *name)# elif defined(CONFIG_PTHREAD_SET_NAME_NP )pthread_set_name_np (pthread_self (), name);# endif + +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +if (pthread_getattr_np (pthread_self (), &real_attr)) { +return ; + } + +if (pthread_attr_getguardsize (&real_attr, &guardsize)) { + goto cleanup; + } + +if (pthread_attr_getstack (&real_attr, &pthread_addr, &stacksize)) { + goto cleanup; + } + +/* Though there's a guard page, it isn't actually marked RO, so +* just label the entire region */ +qemu_name_vma (pthread_addr - guardsize, stacksize + guardsize, +"pthread (%s)" , name); + + cleanup: +pthread_attr_destroy (&real_attr); +# endif } typedef struct {
Signature · 1 line
2.56.1
Re: [PATCH v2 4/7] debug/mem: name pthreads
Quoted 64 lines · John Snow
Signed-off-by John Snow <[email protected]>--- meson.build | 8 ++++++++ util/qemu-thread-posix.c | 30 ++++++++++++++++++++++++++++++ 2 files changed, 38 insertions(+)meson.build +8
-0@@ -2878,7 +2878,15 @@ config_host_data.set('CONFIG_PTHREAD_GET_NAME_NP', cc.links(osdep_prefix + ''' pthread_get_name_np(pthread_self(), buf, sizeof(buf)); return 0; }''', dependencies: threads)) +config_host_data.set('CONFIG_PTHREAD_GETATTR_NP', cc.links(osdep_prefix + ''' + #include <pthread.h> + int main(void) + { + pthread_attr_t attr; + pthread_getattr_np(pthread_self(), &attr); + return 0; + }''', dependencies: threads)) config_host_data.set('CONFIG_PTHREAD_CONDATTR_SETCLOCK', cc.links(osdep_prefix + ''' #include <pthread.h>util/qemu-thread-posix.c +16
-0@@ -17,6 +17,7 @@#include "qemu-thread-common.h" #include "qemu/tsan.h" #include "qemu/bitmap.h" +#include "qemu/mmap-vma.h" #if defined(CONFIG_PTHREAD_SET_NAME_NP ) || defined(CONFIG_PTHREAD_GET_NAME_NP )#include <pthread_np.h> @@ -362,6 +363,13 @@ static void qemu_thread_atexit_notify(void *arg)void qemu_thread_set_name (const char *name) { +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +pthread_attr_t real_attr; +void *pthread_addr; +size_t stacksize; +size_t guardsize; +# endif +/* * Attempt to set the thread name; note that this is for debug, so * we're not going to fail if we can't set it. @@ -373,6 +381,28 @@ void qemu_thread_set_name(const char *name)# elif defined(CONFIG_PTHREAD_SET_NAME_NP )pthread_set_name_np (pthread_self (), name);# endif + +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +if (pthread_getattr_np (pthread_self (), &real_attr)) { +return ; + } + +if (pthread_attr_getguardsize (&real_attr, &guardsize)) { + goto cleanup;
util/qemu-thread-posix.c void qemu_thread_set_name(const char *name) · line 401
Quoted 10 lines
util/qemu-thread-posix.c +10
-0+ } + +if (pthread_attr_getstack (&real_attr, &pthread_addr, &stacksize)) { + goto cleanup; + } + +/* Though there's a guard page, it isn't actually marked RO, so +* just label the entire region */ +qemu_name_vma (pthread_addr - guardsize, stacksize + guardsize, +"pthread (%s)" , name);
On the one hand this code is pretty sketchy as it is making some assumptions about glibc allocating thread stack. On the other hand qemu_name_vma() is a no-op on non-Linux platforms, and we ignore fails, so we'll degrade nicely if any C libraries (musl) deal with thread stack differently.
I feel that glibc itself really ought to set the VMA name in response to pthread_set_name_np but I'm open to us doing this in QEMU though as I think it is safe enough.
Quoted 10 lines
util/qemu-thread-posix.c +4
-0+ + cleanup: +pthread_attr_destroy (&real_attr); +# endif } typedef struct {Signature · 1 line
2.56.1
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
Daniel P. Berrangé
Reviewed-byRe: [PATCH v2 4/7] debug/mem: name pthreads
Quoted 82 lines · Daniel P. Berrangé
Quoted 64 lines · John Snow
Signed-off-by John Snow <[email protected]>--- meson.build | 8 ++++++++ util/qemu-thread-posix.c | 30 ++++++++++++++++++++++++++++++ 2 files changed, 38 insertions(+)meson.build +8
-0@@ -2878,7 +2878,15 @@ config_host_data.set('CONFIG_PTHREAD_GET_NAME_NP', cc.links(osdep_prefix + ''' pthread_get_name_np(pthread_self(), buf, sizeof(buf)); return 0; }''', dependencies: threads)) +config_host_data.set('CONFIG_PTHREAD_GETATTR_NP', cc.links(osdep_prefix + ''' + #include <pthread.h> + int main(void) + { + pthread_attr_t attr; + pthread_getattr_np(pthread_self(), &attr); + return 0; + }''', dependencies: threads)) config_host_data.set('CONFIG_PTHREAD_CONDATTR_SETCLOCK', cc.links(osdep_prefix + ''' #include <pthread.h>util/qemu-thread-posix.c +16
-0@@ -17,6 +17,7 @@#include "qemu-thread-common.h" #include "qemu/tsan.h" #include "qemu/bitmap.h" +#include "qemu/mmap-vma.h" #if defined(CONFIG_PTHREAD_SET_NAME_NP ) || defined(CONFIG_PTHREAD_GET_NAME_NP )#include <pthread_np.h> @@ -362,6 +363,13 @@ static void qemu_thread_atexit_notify(void *arg)void qemu_thread_set_name (const char *name) { +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +pthread_attr_t real_attr; +void *pthread_addr; +size_t stacksize; +size_t guardsize; +# endif +/* * Attempt to set the thread name; note that this is for debug, so * we're not going to fail if we can't set it. @@ -373,6 +381,28 @@ void qemu_thread_set_name(const char *name)# elif defined(CONFIG_PTHREAD_SET_NAME_NP )pthread_set_name_np (pthread_self (), name);# endif + +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +if (pthread_getattr_np (pthread_self (), &real_attr)) { +return ; + } + +if (pthread_attr_getguardsize (&real_attr, &guardsize)) { + goto cleanup;util/qemu-thread-posix.c void qemu_thread_set_name(const char *name) · line 401
Quoted 10 lines
util/qemu-thread-posix.c +10
-0+ } + +if (pthread_attr_getstack (&real_attr, &pthread_addr, &stacksize)) { + goto cleanup; + } + +/* Though there's a guard page, it isn't actually marked RO, so +* just label the entire region */ +qemu_name_vma (pthread_addr - guardsize, stacksize + guardsize, +"pthread (%s)" , name);On the one hand this code is pretty sketchy as it is making some assumptions about glibc allocating thread stack. On the other hand
qemu_name_vma()is a no-op on non-Linux platforms, and we ignore fails, so we'll degrade nicely if any C libraries (musl) deal with thread stack differently.
Quoted 3 lines
I feel that glibc itself really ought to set the VMA name in response to
pthread_set_name_npbut I'm open to us doing this in QEMU though as I think it is safe enough.
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
Re: [PATCH v2 4/7] debug/mem: name pthreads
Quoted 84 lines · Daniel P. Berrangé
Quoted 82 lines · Daniel P. Berrangé
Quoted 64 lines · John Snow
Signed-off-by John Snow <[email protected]>--- meson.build | 8 ++++++++ util/qemu-thread-posix.c | 30 ++++++++++++++++++++++++++++++ 2 files changed, 38 insertions(+)meson.build +8
-0@@ -2878,7 +2878,15 @@ config_host_data.set('CONFIG_PTHREAD_GET_NAME_NP', cc.links(osdep_prefix + ''' pthread_get_name_np(pthread_self(), buf, sizeof(buf)); return 0; }''', dependencies: threads)) +config_host_data.set('CONFIG_PTHREAD_GETATTR_NP', cc.links(osdep_prefix + ''' + #include <pthread.h> + int main(void) + { + pthread_attr_t attr; + pthread_getattr_np(pthread_self(), &attr); + return 0; + }''', dependencies: threads)) config_host_data.set('CONFIG_PTHREAD_CONDATTR_SETCLOCK', cc.links(osdep_prefix + ''' #include <pthread.h>util/qemu-thread-posix.c +16
-0@@ -17,6 +17,7 @@#include "qemu-thread-common.h" #include "qemu/tsan.h" #include "qemu/bitmap.h" +#include "qemu/mmap-vma.h" #if defined(CONFIG_PTHREAD_SET_NAME_NP ) || defined(CONFIG_PTHREAD_GET_NAME_NP )#include <pthread_np.h> @@ -362,6 +363,13 @@ static void qemu_thread_atexit_notify(void *arg)void qemu_thread_set_name (const char *name) { +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +pthread_attr_t real_attr; +void *pthread_addr; +size_t stacksize; +size_t guardsize; +# endif +/* * Attempt to set the thread name; note that this is for debug, so * we're not going to fail if we can't set it. @@ -373,6 +381,28 @@ void qemu_thread_set_name(const char *name)# elif defined(CONFIG_PTHREAD_SET_NAME_NP )pthread_set_name_np (pthread_self (), name);# endif + +# if defined(CONFIG_PTHREAD_GETATTR_NP ) +if (pthread_getattr_np (pthread_self (), &real_attr)) { +return ; + } + +if (pthread_attr_getguardsize (&real_attr, &guardsize)) { + goto cleanup;util/qemu-thread-posix.c void qemu_thread_set_name(const char *name) · line 401
Quoted 10 lines
util/qemu-thread-posix.c +10
-0+ } + +if (pthread_attr_getstack (&real_attr, &pthread_addr, &stacksize)) { + goto cleanup; + } + +/* Though there's a guard page, it isn't actually marked RO, so +* just label the entire region */ +qemu_name_vma (pthread_addr - guardsize, stacksize + guardsize, +"pthread (%s)" , name);On the one hand this code is pretty sketchy as it is making some assumptions about glibc allocating thread stack. On the other hand
qemu_name_vma()is a no-op on non-Linux platforms, and we ignore fails, so we'll degrade nicely if any C libraries (musl) deal with thread stack differently.
Quoted 3 lines
Quoted 3 lines
I feel that glibc itself really ought to set the VMA name in response to
pthread_set_name_npbut I'm open to us doing this in QEMU though as I think it is safe enough.
libc will label this, but it will give it a very generic name like "malloc" or something like that. This is a "free upgrade", but you're right that it is porcelain as it requires us to know a little bit about the memory layout.
I'll add a better comment in-line here to warn folks that this is not necessarily *super* reliable, though in testing it works great...
One concern I have is that even though there is a guard page allocation here, nothing appears to set permissions on said guard
page: it remains fully RW. That's not a particularly effective guard page, I'd have thought...
Quoted 11 lines
Reviewed-by Daniel P. Berrangé <[email protected]>With regards,
Daniel
--|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
PATCH v2 5/7 debug/mem: name coroutine stacks
In reply to John Snow--- util/oslib-posix.c | 3 +++ 1 file changed, 3 insertions(+)
util/oslib-posix.c +3 -0
@@ -60,6 +60,7 @@#include "qemu/memalign.h" #include "qemu/mmap-alloc.h" +#include "qemu/mmap-vma.h" #define MAX_MEM_PREALLOC_THREAD_COUNT 32
@@ -742,6 +743,8 @@ void *qemu_alloc_stack(size_t *sz)abort (); } +qemu_name_vma (ptr, *sz,"coroutine-stack" ); +/* Stack grows down -- guard page at the bottom. */ if (mprotect (ptr, pagesz,PROT_NONE ) !=0 ) {perror ("failed to set up stack guard page" );
Signature · 1 line
2.56.1
Re: [PATCH v2 5/7] debug/mem: name coroutine stacks
Quoted 4 lines · John Snow
Signed-off-by John Snow <[email protected]>--- util/oslib-posix.c | 3 +++ 1 file changed, 3 insertions(+)
Quoted 14 lines
util/oslib-posix.c +1
-0@@ -60,6 +60,7 @@#include "qemu/memalign.h" #include "qemu/mmap-alloc.h" +#include "qemu/mmap-vma.h" #define MAX_MEM_PREALLOC_THREAD_COUNT 32
util/oslib-posix.c void *qemu_alloc_stack(size_t *sz) · line 746
Quoted 4 lines
util/oslib-posix.c +1
-0@@ -742,6 +743,8 @@ void *qemu_alloc_stack(size_t *sz)abort (); } +qemu_name_vma (ptr, *sz,"coroutine-stack" );
Our coroutines are crying out to be named in the future :-)
Quoted 7 lines
util/oslib-posix.c +1
-0+/* Stack grows down -- guard page at the bottom. */ if (mprotect (ptr, pagesz,PROT_NONE ) !=0 ) {perror ("failed to set up stack guard page" );Signature · 1 line
2.56.1
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
Re: [PATCH v2 5/7] debug/mem: name coroutine stacks
Quoted 22 lines · Daniel P. Berrangé
Quoted 4 lines · John Snow
Signed-off-by John Snow <[email protected]>--- util/oslib-posix.c | 3 +++ 1 file changed, 3 insertions(+)Reviewed-by Daniel P. Berrangé <[email protected]>Quoted 13 lines
util/oslib-posix.c +1
-0@@ -60,6 +60,7 @@#include "qemu/memalign.h" #include "qemu/mmap-alloc.h" +#include "qemu/mmap-vma.h" #define MAX_MEM_PREALLOC_THREAD_COUNT 32
Quoted 7 lines
util/oslib-posix.c void *qemu_alloc_stack(size_t *sz) · line 746
Quoted 5 lines
util/oslib-posix.c +1
-0@@ -742,6 +743,8 @@ void *qemu_alloc_stack(size_t *sz)abort (); } +qemu_name_vma (ptr, *sz,"coroutine-stack" );Our coroutines are crying out to be named in the future :-)
I know :)
I looked for five minutes but wasn't exactly sure how to name them. My understanding is each coroutine stack could theoretically service different routines, but only one at a time. I am not so clear on our model that I felt I could confidently give them meaningful names, so I left it as an exercise to a future reader.
(Naughty, naughty.)
Since we're in freeze and it seems unlikely we'll agree this is rc1 fodder, I'm happy to take some extra time and improve the helpers, but some pointed suggestions would not go wrong :)
--js
Quoted 16 lines
Quoted 8 lines
util/oslib-posix.c +1
-1+/* Stack grows down -- guard page at the bottom. */ if (mprotect (ptr, pagesz,PROT_NONE ) !=0 ) {perror ("failed to set up stack guard page" ); --2.56.1
With regards,
Daniel
--|: https://berrange.com ~~ https://hachyderm.io/@berrange :| |: https://libvirt.org ~~ https://entangle-photo.org :| |: https://pixelfed.art/berrange ~~ https://fstop138.berrange.com :|
PATCH v2 6/7 debug/mem: name qemu_anon_ram_alloc/qemu_ram_mmap pages
In reply to John SnowThis patch adds a nickname parameter to qemu_anon_ram_alloc, which passes it to qemu_ram_mmap, which passes it in turn to mmap_activate, where it is used to name the region being allocated.
--- include/qemu/mmap-alloc.h | 3 ++- include/qemu/osdep.h | 2 +- hw/display/virtio-gpu-gl.c | 2 +- migration/ram.c | 3 ++- system/physmem.c | 7 +++++-- target/i386/kvm/tdx.c | 3 ++- util/mmap-alloc.c | 10 +++++++--- util/oslib-posix.c | 4 ++-- util/oslib-win32.c | 2 +- 9 files changed, 23 insertions(+), 13 deletions(-)
include/qemu/mmap-alloc.h +2 -1
@@ -36,7 +36,8 @@ void *qemu_ram_mmap(int fd,size_t size, size_t align, uint32_t qemu_map_flags, - off_t map_offset); + off_t map_offset, + const char *nickname);void qemu_ram_munmap (int fd,void *ptr,size_t size);
include/qemu/osdep.h +1 -1
@@ -552,7 +552,7 @@ void QEMU_ERROR("code path is reachable")int qemu_daemon (int nochdir,int noclose);void *qemu_anon_ram_alloc (size_t size,uint64_t *align,bool shared, -bool noreserve); +bool noreserve,const char *nickname);void qemu_anon_ram_free (void *ptr,size_t size);int qemu_shm_alloc (size_t size,Error **errp);
hw/display/virtio-gpu-gl.c +1 -1
@@ -145,7 +145,7 @@ static void virtio_gpu_gl_device_realize(DeviceState *qdev, Error **errp)#if !defined(CONFIG_WIN32 )if (virtio_gpu_hostmem_enabled (b->conf)) { map =qemu_ram_mmap (-1 , b->conf.hostmem,qemu_real_host_page_size (), -0 ,0 ); +0 ,0 ,"virgl-hostmem" );if (map ==MAP_FAILED ) {error_setg_errno (errp, errno,"virgl hostmem region could not be initialized" );
migration/ram.c +2 -1
@@ -3720,7 +3720,8 @@ int colo_init_ram_cache(Error **errp) WITH_RCU_READ_LOCK_GUARD() {RAMBLOCK_FOREACH_NOT_IGNORED (block) { block->colo_cache =qemu_anon_ram_alloc (block->used_length, -NULL , false, false); +NULL , false, false, +"colo cache" );if (!block->colo_cache) { error_setg(errp, "Can't alloc memory for COLO cache of " "block %s, size 0x" RAM_ADDR_FMT,
system/physmem.c +5 -2
@@ -1767,7 +1767,8 @@ static void *file_ram_alloc(RAMBlock *block, qemu_map_flags |= (block->flags &RAM_SHARED ) ?QEMU_MAP_SHARED :0 ; qemu_map_flags |= (block->flags &RAM_PMEM ) ?QEMU_MAP_SYNC :0 ; qemu_map_flags |= (block->flags &RAM_NORESERVE ) ?QEMU_MAP_NORESERVE :0 ; - area =qemu_ram_mmap (fd, memory, block->mr->align, qemu_map_flags, offset); + area =qemu_ram_mmap (fd, memory, block->mr->align, qemu_map_flags, offset, +"file_ram_alloc" );if (area ==MAP_FAILED ) {error_setg_errno (errp, errno,"unable to map backing store for guest RAM" );
@@ -2169,7 +2170,9 @@ static void ram_block_add(RAMBlock *new_block, Error **errp) }else { new_block->host =qemu_anon_ram_alloc (new_block->max_length, &new_block->mr->align, - shared, noreserve); + shared, noreserve, + new_block->mr->name); +if (!new_block->host) { g_autofree char *size_s = size_to_str(new_block->max_length); error_setg_errno(errp, errno,
target/i386/kvm/tdx.c +2 -1
@@ -354,7 +354,8 @@ static void tdx_finalize_vm(Notifier *notifier, void *unused)case TDVF_SECTION_TYPE_TD_HOB :case TDVF_SECTION_TYPE_TEMP_MEM : entry->mem_ptr =qemu_ram_mmap (-1 , entry->size, -qemu_real_host_page_size (),0 ,0 ); +qemu_real_host_page_size (),0 ,0 , +"tdx" );if (entry->mem_ptr ==MAP_FAILED ) {error_report ("Failed to mmap memory for TDVF section %d" , entry->type);
util/mmap-alloc.c +7 -3
@@ -187,7 +187,8 @@ static void *mmap_reserve(size_t size, int fd)* it accessible. */ static void *mmap_activate (void *ptr,size_t size,int fd, -uint32_t qemu_map_flags,off_t map_offset) +uint32_t qemu_map_flags,off_t map_offset, +const char *nickname) {const bool noreserve = qemu_map_flags &QEMU_MAP_NORESERVE ;const bool readonly = qemu_map_flags &QEMU_MAP_READONLY ;
@@ -236,6 +237,8 @@ static void *mmap_activate(void *ptr, size_t size, int fd,*/ activated_ptr =mmap (ptr, size, prot, flags, fd, map_offset); } + +qemu_name_vma (ptr, size,"%s" , nickname);return activated_ptr; }
@@ -253,7 +256,8 @@ void *qemu_ram_mmap(int fd,size_t size, size_t align, uint32_t qemu_map_flags, - off_t map_offset) + off_t map_offset, + const char *nickname) { const size_t guard_pagesize = mmap_guard_pagesize(fd);size_t offset, total;
@@ -277,7 +281,7 @@ void *qemu_ram_mmap(int fd, offset =QEMU_ALIGN_UP ((uintptr_t )guardptr, align) - (uintptr_t )guardptr; ptr =mmap_activate (guardptr + offset, size, fd, qemu_map_flags, - map_offset); + map_offset, nickname);if (ptr ==MAP_FAILED ) {munmap (guardptr, total);return MAP_FAILED ;
util/oslib-posix.c +2 -2
@@ -207,13 +207,13 @@ fail_close:/* alloc shared memory pages */ void *qemu_anon_ram_alloc (size_t size,uint64_t *alignment,bool shared, -bool noreserve) +bool noreserve,const char *nickname) {const uint32_t qemu_map_flags = (shared ?QEMU_MAP_SHARED :0 ) | (noreserve ?QEMU_MAP_NORESERVE :0 );size_t align =QEMU_VMALLOC_ALIGN ;#ifndef EMSCRIPTEN -void *ptr =qemu_ram_mmap (-1 , size, align, qemu_map_flags,0 ); +void *ptr =qemu_ram_mmap (-1 , size, align, qemu_map_flags,0 , nickname);if (ptr ==MAP_FAILED ) {return NULL ;
util/oslib-win32.c +1 -1
@@ -45,7 +45,7 @@ static int get_allocation_granularity(void) }void *qemu_anon_ram_alloc (size_t size,uint64_t *align,bool shared, -bool noreserve) +bool noreserve,const char *nickname) {void *ptr;
Signature · 1 line
2.56.1
PATCH v2 7/7 debug/mem: HACK add qemu_name_buffer
In reply to John SnowThis is a hack to allow us to name malloc'd buffers that are backed by anonymous mmap allocations.
Quite useful in my debugging, probably not quite appropriate for upstream unless we can figure out a more robust method for identifying when we get an mmap'd page from malloc and adjusting the pointer offset accordingly - I am currently subtracting a fixed 16 bytes because that "works on my system", and I have no clue if that's appropriate everywhere or not.
At least this way, my trick is published to the list for posterity.
--- include/qemu/bitmap.h | 6 ++++++ include/qemu/mmap-vma.h | 3 +++ hw/core/loader.c | 3 +++ hw/i386/x86-common.c | 3 +++ system/device_tree.c | 2 ++ system/physmem.c | 8 ++++++-- util/mmap-vma.c | 23 +++++++++++++++++++++++ 7 files changed, 46 insertions(+), 2 deletions(-)
include/qemu/bitmap.h +6 -0
@@ -104,6 +104,12 @@ static inline unsigned long *bitmap_try_new(long nbits) return g_try_new0(unsigned long, nelem); } +static inline size_t bitmap_size (long nbits) +{ +long nelem =BITS_TO_LONGS (nbits); +return sizeof (unsigned long ) * nelem; +} +static inline unsigned long *bitmap_new (long nbits) {long nelem =BITS_TO_LONGS (nbits);
include/qemu/mmap-vma.h +3 -0
@@ -15,4 +15,7 @@ int qemu_name_vma(void *ptr, size_t sz, const char *fmt, ...) G_GNUC_PRINTF(3, 4); +int qemu_name_buffer(void *ptr, size_t sz, const char *fmt, ...) + G_GNUC_PRINTF(3, 4); +#endif
hw/core/loader.c +3 -0
@@ -45,6 +45,7 @@#include "qemu/osdep.h" #include "qemu/datadir.h" #include "qemu/error-report.h" +#include "qemu/mmap-vma.h" #include "qapi/error.h" #include "qapi/qapi-commands-machine.h" #include "qapi/type-helpers.h"
@@ -1111,6 +1112,8 @@ ssize_t rom_add_file(const char *file, const char *fw_dir, goto err; } +qemu_name_buffer (rom->data, size,"%s" , file); +if (fw_dir) { rom->fw_dir =g_strdup (fw_dir); rom->fw_file =g_strdup (file);
hw/i386/x86-common.c +3 -0
@@ -25,6 +25,7 @@#include "qemu/cutils.h" #include "qemu/units.h" #include "qemu/datadir.h" +#include "qemu/mmap-vma.h" #include "qapi/error.h" #include "system/numa.h" #include "system/system.h"
@@ -914,6 +915,8 @@ void x86_load_linux(X86MachineState *x86ms, setup =g_malloc (setup_size); kernel =g_malloc (kernel_size); +qemu_name_buffer (kernel, kernel_size,"%s" , kernel_filename); +fseek (f,0 ,SEEK_SET );if (fread (setup,1 , setup_size, f) != setup_size) {fprintf (stderr,"fread() failed\n" );
system/device_tree.c +2 -0
@@ -27,6 +27,7 @@#include "hw/core/loader.h" #include "hw/core/boards.h" #include "qemu/config-file.h" +#include "qemu/mmap-vma.h" #include "qapi/qapi-commands-machine.h" #include "qobject/qdict.h"
@@ -41,6 +42,7 @@ void *create_device_tree(int *sizep) *sizep =FDT_MAX_SIZE ; fdt =g_malloc0 (FDT_MAX_SIZE ); +qemu_name_buffer (fdt,FDT_MAX_SIZE ,"device_tree" ); ret =fdt_create (fdt,FDT_MAX_SIZE );if (ret <0 ) { goto fail;
system/physmem.c +6 -2
@@ -26,6 +26,7 @@#include "qemu/hbitmap.h" #include "qemu/madvise.h" #include "qemu/lockable.h" +#include "qemu/mmap-vma.h" #ifdef CONFIG_TCG #include "accel/tcg/cpu-ops.h"
@@ -2105,7 +2106,7 @@ void qemu_ram_msync(RAMBlock *block, ram_addr_t start, ram_addr_t length) }/* Called with ram_list.mutex held */ -static void dirty_memory_extend (ram_addr_t new_ram_size) +static void dirty_memory_extend (ram_addr_t new_ram_size,const char *nickname) {unsigned int old_num_blocks = ram_list.num_dirty_blocks; unsigned int new_num_blocks = DIV_ROUND_UP(new_ram_size,
@@ -2133,6 +2134,9 @@ static void dirty_memory_extend(ram_addr_t new_ram_size)for (j = old_num_blocks; j < new_num_blocks; j++) { new_blocks->blocks[j] =bitmap_new (DIRTY_MEMORY_BLOCK_SIZE ); +qemu_name_buffer (new_blocks->blocks[j], +bitmap_size (DIRTY_MEMORY_BLOCK_SIZE ), +"bitmap<%s>" , nickname); }qatomic_rcu_set (&ram_list.dirty_memory[i], new_blocks);
@@ -2254,7 +2258,7 @@ static void ram_block_add(RAMBlock *new_block, Error **errp) } ram_size = (new_block->offset + new_block->max_length) >>TARGET_PAGE_BITS ; -dirty_memory_extend (ram_size); +dirty_memory_extend (ram_size, new_block->mr->name);/* Keep the list sorted from biggest to smallest block. Unlike QTAILQ, * QLIST (which has an RCU-friendly variant) does not have insertion at * tail, so save the last element in last_block.
util/mmap-vma.c +23 -0
@@ -73,3 +73,26 @@ int qemu_name_vma(void *ptr, size_t sz, const char *fmt, ...)return rc; } + +G_GNUC_PRINTF(3, 4) +int qemu_name_buffer (void *ptr,size_t sz,const char *fmt, ...) +{ +/* +* this is hacky as hell, but ptmalloc will occasionally use +* mmap() for large buffers - when it does so, you get a pointer +* to 16 bytes after the actual mmap. This function compensates +* for that to guess at what the real mmap allocation was. +* +* This can be used unconditionally whether or not the allocation +* was actually mmap'd - if it wasn't, this just fails. Yes, it's +* still very ugly. +*/ +va_list args; +size_t pagesz =qemu_real_host_page_size (); + + ptr = ptr -0x10 ; + sz =ROUND_UP (sz +0x10 , pagesz); + +va_start (args, fmt); +return qemu_vname_vma (ptr, sz, fmt, args); +}
Signature · 1 line
2.56.1