[PATCH v1 0/2] Initialize backend memory objects in parallel

[PATCH v1 0/2] Initialize backend memory objects in parallel

[Mark Kanda]

QEMU initializes preallocated backend memory when parsing the corresponding
objects from the command line. In certain scenarios, such as memory being
preallocated across multiple numa nodes, this approach is not optimal due to
the unnecessary serialization.

This series addresses this issue by initializing the backend memory objects in
parallel.

Mark Kanda (2):
  oslib-posix: refactor memory prealloc threads
  oslib-posix: initialize backend memory objects in parallel

 include/qemu/osdep.h |   6 ++
 system/vl.c          |   2 +
 util/oslib-posix.c   | 150 +++++++++++++++++++++++++++++--------------
 util/oslib-win32.c   |   5 ++
 4 files changed, 116 insertions(+), 47 deletions(-)

-- 
2.39.3

[PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[Mark Kanda]

Refactor the memory prealloc threads support:
- Make memset context a global qlist
- Move the memset thread join/cleanup code to a separate routine

This is functionally equivalent and facilitates multiple memset contexts
(used in a subsequent patch).

Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
---
 util/oslib-posix.c | 104 +++++++++++++++++++++++++++++----------------
 1 file changed, 68 insertions(+), 36 deletions(-)

diff --git a/util/oslib-posix.c b/util/oslib-posix.c
index e86fd64e09..293297ac6c 100644
--- a/util/oslib-posix.c
+++ b/util/oslib-posix.c
@@ -63,11 +63,15 @@
 
 struct MemsetThread;
 
+static QLIST_HEAD(, MemsetContext) memset_contexts =
+    QLIST_HEAD_INITIALIZER(memset_contexts);
+
 typedef struct MemsetContext {
     bool all_threads_created;
     bool any_thread_failed;
     struct MemsetThread *threads;
     int num_threads;
+    QLIST_ENTRY(MemsetContext) next;
 } MemsetContext;
 
 struct MemsetThread {
@@ -81,7 +85,7 @@ struct MemsetThread {
 typedef struct MemsetThread MemsetThread;
 
 /* used by sigbus_handler() */
-static MemsetContext *sigbus_memset_context;
+static bool sigbus_memset_context;
 struct sigaction sigbus_oldact;
 static QemuMutex sigbus_mutex;
 
@@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
 #endif /* CONFIG_LINUX */
 {
     int i;
+    MemsetContext *context;
 
     if (sigbus_memset_context) {
-        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
-            MemsetThread *thread = &sigbus_memset_context->threads[i];
+        QLIST_FOREACH(context, &memset_contexts, next) {
+            for (i = 0; i < context->num_threads; i++) {
+                MemsetThread *thread = &context->threads[i];
 
-            if (qemu_thread_is_self(&thread->pgthread)) {
-                siglongjmp(thread->env, 1);
+                if (qemu_thread_is_self(&thread->pgthread)) {
+                    siglongjmp(thread->env, 1);
+                }
             }
         }
     }
@@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages,
                            bool use_madv_populate_write)
 {
     static gsize initialized = 0;
-    MemsetContext context = {
-        .num_threads = get_memset_num_threads(hpagesize, numpages, max_threads),
-    };
+    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
     size_t numpages_per_thread, leftover;
     void *(*touch_fn)(void *);
-    int ret = 0, i = 0;
+    int i = 0;
     char *addr = area;
 
+    context->num_threads =
+        get_memset_num_threads(hpagesize, numpages, max_threads);
+
     if (g_once_init_enter(&initialized)) {
         qemu_mutex_init(&page_mutex);
         qemu_cond_init(&page_cond);
@@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages,
 
     if (use_madv_populate_write) {
         /* Avoid creating a single thread for MADV_POPULATE_WRITE */
-        if (context.num_threads == 1) {
+        if (context->num_threads == 1) {
             if (qemu_madvise(area, hpagesize * numpages,
                              QEMU_MADV_POPULATE_WRITE)) {
                 return -errno;
@@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages,
         touch_fn = do_touch_pages;
     }
 
-    context.threads = g_new0(MemsetThread, context.num_threads);
-    numpages_per_thread = numpages / context.num_threads;
-    leftover = numpages % context.num_threads;
-    for (i = 0; i < context.num_threads; i++) {
-        context.threads[i].addr = addr;
-        context.threads[i].numpages = numpages_per_thread + (i < leftover);
-        context.threads[i].hpagesize = hpagesize;
-        context.threads[i].context = &context;
+    context->threads = g_new0(MemsetThread, context->num_threads);
+    numpages_per_thread = numpages / context->num_threads;
+    leftover = numpages % context->num_threads;
+    for (i = 0; i < context->num_threads; i++) {
+        context->threads[i].addr = addr;
+        context->threads[i].numpages = numpages_per_thread + (i < leftover);
+        context->threads[i].hpagesize = hpagesize;
+        context->threads[i].context = context;
         if (tc) {
-            thread_context_create_thread(tc, &context.threads[i].pgthread,
+            thread_context_create_thread(tc, &context->threads[i].pgthread,
                                          "touch_pages",
-                                         touch_fn, &context.threads[i],
+                                         touch_fn, &context->threads[i],
                                          QEMU_THREAD_JOINABLE);
         } else {
-            qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
-                               touch_fn, &context.threads[i],
+            qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
+                               touch_fn, &context->threads[i],
                                QEMU_THREAD_JOINABLE);
         }
-        addr += context.threads[i].numpages * hpagesize;
+        addr += context->threads[i].numpages * hpagesize;
     }
 
+    QLIST_INSERT_HEAD(&memset_contexts, context, next);
+
     if (!use_madv_populate_write) {
-        sigbus_memset_context = &context;
+        sigbus_memset_context = true;
     }
 
+    return 0;
+}
+
+static int wait_mem_prealloc(void)
+{
+    int i, ret = 0;
+    MemsetContext *context, *next_context;
     qemu_mutex_lock(&page_mutex);
-    context.all_threads_created = true;
+    QLIST_FOREACH(context, &memset_contexts, next) {
+        context->all_threads_created = true;
+    }
     qemu_cond_broadcast(&page_cond);
     qemu_mutex_unlock(&page_mutex);
 
-    for (i = 0; i < context.num_threads; i++) {
-        int tmp = (uintptr_t)qemu_thread_join(&context.threads[i].pgthread);
+    QLIST_FOREACH(context, &memset_contexts, next) {
+        for (i = 0; i < context->num_threads; i++) {
+            int tmp =
+                (uintptr_t)qemu_thread_join(&context->threads[i].pgthread);
 
-        if (tmp) {
-            ret = tmp;
+            if (tmp) {
+                ret = tmp;
+            }
         }
     }
 
-    if (!use_madv_populate_write) {
-        sigbus_memset_context = NULL;
+    if (sigbus_oldact.sa_handler) {
+        /* restore the previous sighandler */
+        if (sigaction(SIGBUS, &sigbus_oldact, NULL)) {
+            /* Terminate QEMU since it can't recover from error */
+            perror("wait_mem_prealloc: failed to reinstall signal handler");
+            exit(1);
+        }
+        memset(&sigbus_oldact, 0, sizeof(sigbus_oldact));
     }
-    g_free(context.threads);
+    sigbus_memset_context = false;
 
+    QLIST_FOREACH_SAFE(context, &memset_contexts, next, next_context) {
+        QLIST_REMOVE(context, next);
+        g_free(context->threads);
+        g_free(context);
+    }
     return ret;
 }
 
@@ -547,11 +580,10 @@ void qemu_prealloc_mem(int fd, char *area, size_t sz, int max_threads,
     }
 
     if (!use_madv_populate_write) {
-        ret = sigaction(SIGBUS, &sigbus_oldact, NULL);
+        ret = wait_mem_prealloc();
         if (ret) {
-            /* Terminate QEMU since it can't recover from error */
-            perror("qemu_prealloc_mem: failed to reinstall signal handler");
-            exit(1);
+            error_setg_errno(errp, -ret,
+                "qemu_prealloc_mem: failed waiting for memory prealloc");
         }
         qemu_mutex_unlock(&sigbus_mutex);
     }
-- 
2.39.3

[PATCH v1 2/2] oslib-posix: initialize backend memory objects in parallel

[Mark Kanda]

QEMU initializes preallocated backend memory as the objects are parsed from
the command line. This is not optimal in some cases (e.g. memory spanning
multiple numa nodes) because the memory objects are initialized in series.

Allow the initialization to occur in parallel. The performance increase is
significant and scales with the number of objects. On a 2 socket Skylake VM
with 128GB and 2 init threads per socket (256GB total), the memory init time
decreases from ~27 seconds to ~14 seconds.

Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
---
 include/qemu/osdep.h |  6 ++++++
 system/vl.c          |  2 ++
 util/oslib-posix.c   | 46 +++++++++++++++++++++++++++++++++-----------
 util/oslib-win32.c   |  5 +++++
 4 files changed, 48 insertions(+), 11 deletions(-)

diff --git a/include/qemu/osdep.h b/include/qemu/osdep.h
index d30ba73eda..57185e6309 100644
--- a/include/qemu/osdep.h
+++ b/include/qemu/osdep.h
@@ -682,6 +682,12 @@ typedef struct ThreadContext ThreadContext;
 void qemu_prealloc_mem(int fd, char *area, size_t sz, int max_threads,
                        ThreadContext *tc, Error **errp);
 
+/**
+ * Wait for any outstanding memory prealloc initialization
+ * to complete.
+ */
+void wait_mem_prealloc_init(void);
+
 /**
  * qemu_get_pid_name:
  * @pid: pid of a process
diff --git a/system/vl.c b/system/vl.c
index 6b87bfa32c..9e04acbb2c 100644
--- a/system/vl.c
+++ b/system/vl.c
@@ -2010,6 +2010,8 @@ static void qemu_create_late_backends(void)
 
     object_option_foreach_add(object_create_late);
 
+    wait_mem_prealloc_init();
+
     if (tpm_init() < 0) {
         exit(1);
     }
diff --git a/util/oslib-posix.c b/util/oslib-posix.c
index 293297ac6c..667d2d960c 100644
--- a/util/oslib-posix.c
+++ b/util/oslib-posix.c
@@ -91,6 +91,7 @@ static QemuMutex sigbus_mutex;
 
 static QemuMutex page_mutex;
 static QemuCond page_cond;
+static bool prealloc_init;
 
 int qemu_get_thread_id(void)
 {
@@ -487,6 +488,12 @@ static int wait_mem_prealloc(void)
 {
     int i, ret = 0;
     MemsetContext *context, *next_context;
+
+    /* Return if memory prealloc isn't enabled or active */
+    if (QLIST_EMPTY(&memset_contexts) || !prealloc_init) {
+        return 0;
+    }
+
     qemu_mutex_lock(&page_mutex);
     QLIST_FOREACH(context, &memset_contexts, next) {
         context->all_threads_created = true;
@@ -553,21 +560,23 @@ void qemu_prealloc_mem(int fd, char *area, size_t sz, int max_threads,
         }
 
         qemu_mutex_lock(&sigbus_mutex);
-        memset(&act, 0, sizeof(act));
+        if (!sigbus_oldact.sa_handler) {
+            memset(&act, 0, sizeof(act));
 #ifdef CONFIG_LINUX
-        act.sa_sigaction = &sigbus_handler;
-        act.sa_flags = SA_SIGINFO;
+            act.sa_sigaction = &sigbus_handler;
+            act.sa_flags = SA_SIGINFO;
 #else /* CONFIG_LINUX */
-        act.sa_handler = &sigbus_handler;
-        act.sa_flags = 0;
+            act.sa_handler = &sigbus_handler;
+            act.sa_flags = 0;
 #endif /* CONFIG_LINUX */
 
-        ret = sigaction(SIGBUS, &act, &sigbus_oldact);
-        if (ret) {
-            qemu_mutex_unlock(&sigbus_mutex);
-            error_setg_errno(errp, errno,
-                "qemu_prealloc_mem: failed to install signal handler");
-            return;
+            ret = sigaction(SIGBUS, &act, &sigbus_oldact);
+            if (ret) {
+                qemu_mutex_unlock(&sigbus_mutex);
+                error_setg_errno(errp, errno,
+                    "qemu_prealloc_mem: failed to install signal handler");
+                return;
+            }
         }
     }
 
@@ -589,6 +598,21 @@ void qemu_prealloc_mem(int fd, char *area, size_t sz, int max_threads,
     }
 }
 
+void wait_mem_prealloc_init(void)
+{
+    /*
+     * Set prealloc_init true to make wait_mem_prealloc() wait for the
+     * initialization to complete.
+     */
+    prealloc_init = true;
+
+    /* Wait for any outstanding init to complete */
+    if (wait_mem_prealloc()) {
+        perror("wait_mem_prealloc_init: failed waiting for memory prealloc");
+        exit(1);
+    }
+}
+
 char *qemu_get_pid_name(pid_t pid)
 {
     char *name = NULL;
diff --git a/util/oslib-win32.c b/util/oslib-win32.c
index 55b0189dc3..72e050bee1 100644
--- a/util/oslib-win32.c
+++ b/util/oslib-win32.c
@@ -276,6 +276,11 @@ void qemu_prealloc_mem(int fd, char *area, size_t sz, int max_threads,
     }
 }
 
+void wait_mem_prealloc_init(void)
+{
+    /* not supported */
+}
+
 char *qemu_get_pid_name(pid_t pid)
 {
     /* XXX Implement me */
-- 
2.39.3

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[David Hildenbrand]

On 08.01.24 16:10, Mark Kanda wrote:
> Refactor the memory prealloc threads support:
> - Make memset context a global qlist
> - Move the memset thread join/cleanup code to a separate routine
> 
> This is functionally equivalent and facilitates multiple memset contexts
> (used in a subsequent patch).
> 
> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
> ---
>   util/oslib-posix.c | 104 +++++++++++++++++++++++++++++----------------
>   1 file changed, 68 insertions(+), 36 deletions(-)
> 
> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
> index e86fd64e09..293297ac6c 100644
> --- a/util/oslib-posix.c
> +++ b/util/oslib-posix.c
> @@ -63,11 +63,15 @@
>   
>   struct MemsetThread;
>   
> +static QLIST_HEAD(, MemsetContext) memset_contexts =
> +    QLIST_HEAD_INITIALIZER(memset_contexts);
> +
>   typedef struct MemsetContext {
>       bool all_threads_created;
>       bool any_thread_failed;
>       struct MemsetThread *threads;
>       int num_threads;
> +    QLIST_ENTRY(MemsetContext) next;
>   } MemsetContext;
>   
>   struct MemsetThread {
> @@ -81,7 +85,7 @@ struct MemsetThread {
>   typedef struct MemsetThread MemsetThread;
>   
>   /* used by sigbus_handler() */
> -static MemsetContext *sigbus_memset_context;
> +static bool sigbus_memset_context;
>   struct sigaction sigbus_oldact;
>   static QemuMutex sigbus_mutex;
>   
> @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>   #endif /* CONFIG_LINUX */
>   {
>       int i;
> +    MemsetContext *context;
>   
>       if (sigbus_memset_context) {
> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
> -            MemsetThread *thread = &sigbus_memset_context->threads[i];
> +        QLIST_FOREACH(context, &memset_contexts, next) {
> +            for (i = 0; i < context->num_threads; i++) {
> +                MemsetThread *thread = &context->threads[i];
>   
> -            if (qemu_thread_is_self(&thread->pgthread)) {
> -                siglongjmp(thread->env, 1);
> +                if (qemu_thread_is_self(&thread->pgthread)) {
> +                    siglongjmp(thread->env, 1);
> +                }
>               }
>           }
>       }
> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages,
>                              bool use_madv_populate_write)
>   {
>       static gsize initialized = 0;
> -    MemsetContext context = {
> -        .num_threads = get_memset_num_threads(hpagesize, numpages, max_threads),
> -    };
> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>       size_t numpages_per_thread, leftover;
>       void *(*touch_fn)(void *);
> -    int ret = 0, i = 0;
> +    int i = 0;
>       char *addr = area;
>   
> +    context->num_threads =
> +        get_memset_num_threads(hpagesize, numpages, max_threads);
> +
>       if (g_once_init_enter(&initialized)) {
>           qemu_mutex_init(&page_mutex);
>           qemu_cond_init(&page_cond);
> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages,
>   
>       if (use_madv_populate_write) {
>           /* Avoid creating a single thread for MADV_POPULATE_WRITE */
> -        if (context.num_threads == 1) {
> +        if (context->num_threads == 1) {
>               if (qemu_madvise(area, hpagesize * numpages,
>                                QEMU_MADV_POPULATE_WRITE)) {
>                   return -errno;
> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages,
>           touch_fn = do_touch_pages;
>       }
>   
> -    context.threads = g_new0(MemsetThread, context.num_threads);
> -    numpages_per_thread = numpages / context.num_threads;
> -    leftover = numpages % context.num_threads;
> -    for (i = 0; i < context.num_threads; i++) {
> -        context.threads[i].addr = addr;
> -        context.threads[i].numpages = numpages_per_thread + (i < leftover);
> -        context.threads[i].hpagesize = hpagesize;
> -        context.threads[i].context = &context;
> +    context->threads = g_new0(MemsetThread, context->num_threads);
> +    numpages_per_thread = numpages / context->num_threads;
> +    leftover = numpages % context->num_threads;
> +    for (i = 0; i < context->num_threads; i++) {
> +        context->threads[i].addr = addr;
> +        context->threads[i].numpages = numpages_per_thread + (i < leftover);
> +        context->threads[i].hpagesize = hpagesize;
> +        context->threads[i].context = context;
>           if (tc) {
> -            thread_context_create_thread(tc, &context.threads[i].pgthread,
> +            thread_context_create_thread(tc, &context->threads[i].pgthread,
>                                            "touch_pages",
> -                                         touch_fn, &context.threads[i],
> +                                         touch_fn, &context->threads[i],
>                                            QEMU_THREAD_JOINABLE);
>           } else {
> -            qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
> -                               touch_fn, &context.threads[i],
> +            qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
> +                               touch_fn, &context->threads[i],
>                                  QEMU_THREAD_JOINABLE);
>           }
> -        addr += context.threads[i].numpages * hpagesize;
> +        addr += context->threads[i].numpages * hpagesize;
>       }
>   
> +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
> +
>       if (!use_madv_populate_write) {
> -        sigbus_memset_context = &context;
> +        sigbus_memset_context = true;

Could we just use the sigbus handling alone and support parallel init 
only when using MADV_POPULATE_WRITE where we don't have to mess with 
signal handlers?

Further, how do you changes interact with other callers of 
qemu_prealloc_mem(), like virtio-mem?

-- 
Cheers,

David / dhildenb

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[Mark Kanda]

On 1/8/24 9:40 AM, David Hildenbrand wrote:
> On 08.01.24 16:10, Mark Kanda wrote:
>> Refactor the memory prealloc threads support:
>> - Make memset context a global qlist
>> - Move the memset thread join/cleanup code to a separate routine
>>
>> This is functionally equivalent and facilitates multiple memset contexts
>> (used in a subsequent patch).
>>
>> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
>> ---
>>   util/oslib-posix.c | 104 +++++++++++++++++++++++++++++----------------
>>   1 file changed, 68 insertions(+), 36 deletions(-)
>>
>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
>> index e86fd64e09..293297ac6c 100644
>> --- a/util/oslib-posix.c
>> +++ b/util/oslib-posix.c
>> @@ -63,11 +63,15 @@
>>     struct MemsetThread;
>>   +static QLIST_HEAD(, MemsetContext) memset_contexts =
>> +    QLIST_HEAD_INITIALIZER(memset_contexts);
>> +
>>   typedef struct MemsetContext {
>>       bool all_threads_created;
>>       bool any_thread_failed;
>>       struct MemsetThread *threads;
>>       int num_threads;
>> +    QLIST_ENTRY(MemsetContext) next;
>>   } MemsetContext;
>>     struct MemsetThread {
>> @@ -81,7 +85,7 @@ struct MemsetThread {
>>   typedef struct MemsetThread MemsetThread;
>>     /* used by sigbus_handler() */
>> -static MemsetContext *sigbus_memset_context;
>> +static bool sigbus_memset_context;
>>   struct sigaction sigbus_oldact;
>>   static QemuMutex sigbus_mutex;
>>   @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>>   #endif /* CONFIG_LINUX */
>>   {
>>       int i;
>> +    MemsetContext *context;
>>         if (sigbus_memset_context) {
>> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
>> -            MemsetThread *thread = &sigbus_memset_context->threads[i];
>> +        QLIST_FOREACH(context, &memset_contexts, next) {
>> +            for (i = 0; i < context->num_threads; i++) {
>> +                MemsetThread *thread = &context->threads[i];
>>   -            if (qemu_thread_is_self(&thread->pgthread)) {
>> -                siglongjmp(thread->env, 1);
>> +                if (qemu_thread_is_self(&thread->pgthread)) {
>> +                    siglongjmp(thread->env, 1);
>> +                }
>>               }
>>           }
>>       }
>> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t 
>> hpagesize, size_t numpages,
>>                              bool use_madv_populate_write)
>>   {
>>       static gsize initialized = 0;
>> -    MemsetContext context = {
>> -        .num_threads = get_memset_num_threads(hpagesize, numpages, 
>> max_threads),
>> -    };
>> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>>       size_t numpages_per_thread, leftover;
>>       void *(*touch_fn)(void *);
>> -    int ret = 0, i = 0;
>> +    int i = 0;
>>       char *addr = area;
>>   +    context->num_threads =
>> +        get_memset_num_threads(hpagesize, numpages, max_threads);
>> +
>>       if (g_once_init_enter(&initialized)) {
>>           qemu_mutex_init(&page_mutex);
>>           qemu_cond_init(&page_cond);
>> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t 
>> hpagesize, size_t numpages,
>>         if (use_madv_populate_write) {
>>           /* Avoid creating a single thread for MADV_POPULATE_WRITE */
>> -        if (context.num_threads == 1) {
>> +        if (context->num_threads == 1) {
>>               if (qemu_madvise(area, hpagesize * numpages,
>>                                QEMU_MADV_POPULATE_WRITE)) {
>>                   return -errno;
>> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t 
>> hpagesize, size_t numpages,
>>           touch_fn = do_touch_pages;
>>       }
>>   -    context.threads = g_new0(MemsetThread, context.num_threads);
>> -    numpages_per_thread = numpages / context.num_threads;
>> -    leftover = numpages % context.num_threads;
>> -    for (i = 0; i < context.num_threads; i++) {
>> -        context.threads[i].addr = addr;
>> -        context.threads[i].numpages = numpages_per_thread + (i < 
>> leftover);
>> -        context.threads[i].hpagesize = hpagesize;
>> -        context.threads[i].context = &context;
>> +    context->threads = g_new0(MemsetThread, context->num_threads);
>> +    numpages_per_thread = numpages / context->num_threads;
>> +    leftover = numpages % context->num_threads;
>> +    for (i = 0; i < context->num_threads; i++) {
>> +        context->threads[i].addr = addr;
>> +        context->threads[i].numpages = numpages_per_thread + (i < 
>> leftover);
>> +        context->threads[i].hpagesize = hpagesize;
>> +        context->threads[i].context = context;
>>           if (tc) {
>> -            thread_context_create_thread(tc, 
>> &context.threads[i].pgthread,
>> +            thread_context_create_thread(tc, 
>> &context->threads[i].pgthread,
>>                                            "touch_pages",
>> -                                         touch_fn, &context.threads[i],
>> +                                         touch_fn, 
>> &context->threads[i],
>> QEMU_THREAD_JOINABLE);
>>           } else {
>> - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
>> -                               touch_fn, &context.threads[i],
>> + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
>> +                               touch_fn, &context->threads[i],
>>                                  QEMU_THREAD_JOINABLE);
>>           }
>> -        addr += context.threads[i].numpages * hpagesize;
>> +        addr += context->threads[i].numpages * hpagesize;
>>       }
>>   +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
>> +
>>       if (!use_madv_populate_write) {
>> -        sigbus_memset_context = &context;
>> +        sigbus_memset_context = true;
>
Thanks David,

> Could we just use the sigbus handling alone and support parallel init 
> only when using MADV_POPULATE_WRITE where we don't have to mess with 
> signal handlers?
>

Ideally, we're hoping to support this with earlier kernels which don't 
support MADV_POPULATE _WRITE. But, I will check to see if we really need it.

> Further, how do you changes interact with other callers of 
> qemu_prealloc_mem(), like virtio-mem?
>

I'm not familiar with the intricacies of virtio-mem, but the basic idea 
of this series is to *only* allow parallel init during the start up 
phase (while prealloc_init == false). Once we have parsed all the 
command line args, we set prealloc_init = true 
(wait_mem_prealloc_init()) which causes all subsequent calls to 
qemu_prealloc_mem() to perform initialization synchronously. So, I 
*think* this covers the virtio-mem use case. Am I missing something?

Thanks/regards,
-Mark

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[David Hildenbrand]

On 08.01.24 19:40, Mark Kanda wrote:
> On 1/8/24 9:40 AM, David Hildenbrand wrote:
>> On 08.01.24 16:10, Mark Kanda wrote:
>>> Refactor the memory prealloc threads support:
>>> - Make memset context a global qlist
>>> - Move the memset thread join/cleanup code to a separate routine
>>>
>>> This is functionally equivalent and facilitates multiple memset contexts
>>> (used in a subsequent patch).
>>>
>>> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
>>> ---
>>>    util/oslib-posix.c | 104 +++++++++++++++++++++++++++++----------------
>>>    1 file changed, 68 insertions(+), 36 deletions(-)
>>>
>>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
>>> index e86fd64e09..293297ac6c 100644
>>> --- a/util/oslib-posix.c
>>> +++ b/util/oslib-posix.c
>>> @@ -63,11 +63,15 @@
>>>      struct MemsetThread;
>>>    +static QLIST_HEAD(, MemsetContext) memset_contexts =
>>> +    QLIST_HEAD_INITIALIZER(memset_contexts);
>>> +
>>>    typedef struct MemsetContext {
>>>        bool all_threads_created;
>>>        bool any_thread_failed;
>>>        struct MemsetThread *threads;
>>>        int num_threads;
>>> +    QLIST_ENTRY(MemsetContext) next;
>>>    } MemsetContext;
>>>      struct MemsetThread {
>>> @@ -81,7 +85,7 @@ struct MemsetThread {
>>>    typedef struct MemsetThread MemsetThread;
>>>      /* used by sigbus_handler() */
>>> -static MemsetContext *sigbus_memset_context;
>>> +static bool sigbus_memset_context;
>>>    struct sigaction sigbus_oldact;
>>>    static QemuMutex sigbus_mutex;
>>>    @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>>>    #endif /* CONFIG_LINUX */
>>>    {
>>>        int i;
>>> +    MemsetContext *context;
>>>          if (sigbus_memset_context) {
>>> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
>>> -            MemsetThread *thread = &sigbus_memset_context->threads[i];
>>> +        QLIST_FOREACH(context, &memset_contexts, next) {
>>> +            for (i = 0; i < context->num_threads; i++) {
>>> +                MemsetThread *thread = &context->threads[i];
>>>    -            if (qemu_thread_is_self(&thread->pgthread)) {
>>> -                siglongjmp(thread->env, 1);
>>> +                if (qemu_thread_is_self(&thread->pgthread)) {
>>> +                    siglongjmp(thread->env, 1);
>>> +                }
>>>                }
>>>            }
>>>        }
>>> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t
>>> hpagesize, size_t numpages,
>>>                               bool use_madv_populate_write)
>>>    {
>>>        static gsize initialized = 0;
>>> -    MemsetContext context = {
>>> -        .num_threads = get_memset_num_threads(hpagesize, numpages,
>>> max_threads),
>>> -    };
>>> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>>>        size_t numpages_per_thread, leftover;
>>>        void *(*touch_fn)(void *);
>>> -    int ret = 0, i = 0;
>>> +    int i = 0;
>>>        char *addr = area;
>>>    +    context->num_threads =
>>> +        get_memset_num_threads(hpagesize, numpages, max_threads);
>>> +
>>>        if (g_once_init_enter(&initialized)) {
>>>            qemu_mutex_init(&page_mutex);
>>>            qemu_cond_init(&page_cond);
>>> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t
>>> hpagesize, size_t numpages,
>>>          if (use_madv_populate_write) {
>>>            /* Avoid creating a single thread for MADV_POPULATE_WRITE */
>>> -        if (context.num_threads == 1) {
>>> +        if (context->num_threads == 1) {
>>>                if (qemu_madvise(area, hpagesize * numpages,
>>>                                 QEMU_MADV_POPULATE_WRITE)) {
>>>                    return -errno;
>>> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t
>>> hpagesize, size_t numpages,
>>>            touch_fn = do_touch_pages;
>>>        }
>>>    -    context.threads = g_new0(MemsetThread, context.num_threads);
>>> -    numpages_per_thread = numpages / context.num_threads;
>>> -    leftover = numpages % context.num_threads;
>>> -    for (i = 0; i < context.num_threads; i++) {
>>> -        context.threads[i].addr = addr;
>>> -        context.threads[i].numpages = numpages_per_thread + (i <
>>> leftover);
>>> -        context.threads[i].hpagesize = hpagesize;
>>> -        context.threads[i].context = &context;
>>> +    context->threads = g_new0(MemsetThread, context->num_threads);
>>> +    numpages_per_thread = numpages / context->num_threads;
>>> +    leftover = numpages % context->num_threads;
>>> +    for (i = 0; i < context->num_threads; i++) {
>>> +        context->threads[i].addr = addr;
>>> +        context->threads[i].numpages = numpages_per_thread + (i <
>>> leftover);
>>> +        context->threads[i].hpagesize = hpagesize;
>>> +        context->threads[i].context = context;
>>>            if (tc) {
>>> -            thread_context_create_thread(tc,
>>> &context.threads[i].pgthread,
>>> +            thread_context_create_thread(tc,
>>> &context->threads[i].pgthread,
>>>                                             "touch_pages",
>>> -                                         touch_fn, &context.threads[i],
>>> +                                         touch_fn,
>>> &context->threads[i],
>>> QEMU_THREAD_JOINABLE);
>>>            } else {
>>> - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
>>> -                               touch_fn, &context.threads[i],
>>> + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
>>> +                               touch_fn, &context->threads[i],
>>>                                   QEMU_THREAD_JOINABLE);
>>>            }
>>> -        addr += context.threads[i].numpages * hpagesize;
>>> +        addr += context->threads[i].numpages * hpagesize;
>>>        }
>>>    +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
>>> +
>>>        if (!use_madv_populate_write) {
>>> -        sigbus_memset_context = &context;
>>> +        sigbus_memset_context = true;
>>
> Thanks David,
> 
>> Could we just use the sigbus handling alone and support parallel init
>> only when using MADV_POPULATE_WRITE where we don't have to mess with
>> signal handlers?
>>
> 
> Ideally, we're hoping to support this with earlier kernels which don't
> support MADV_POPULATE _WRITE. But, I will check to see if we really need it.

That's around since Linux 5.14, so please try simplifying.

> 
>> Further, how do you changes interact with other callers of
>> qemu_prealloc_mem(), like virtio-mem?
>>
> 
> I'm not familiar with the intricacies of virtio-mem, but the basic idea
> of this series is to *only* allow parallel init during the start up
> phase (while prealloc_init == false). Once we have parsed all the
> command line args, we set prealloc_init = true
> (wait_mem_prealloc_init()) which causes all subsequent calls to
> qemu_prealloc_mem() to perform initialization synchronously. So, I
> *think* this covers the virtio-mem use case. Am I missing something?

Good, so this should likely not affect virtio-mem (which also ends up 
preallocating memory when loading from a vmstate).

To make this all a bit clearer, what about the following to make this:

* Optimize for MADV_POPULATE_WRITE. If we really need support for
   !MADV_POPULATE_WRITE, this is better added on top later.
* Pass in via a parameter that the caller requests async handling. "bool
   async" should be good enough. Then, pass that only from the memory
   backend call, while QEMU is still initializing (we can find a way to
   make that work).
* Provide a function that waits for any remaining async os-mem-prealloc
   activity. That is essentially "wait_mem_prealloc_init", but without
   the special internal flag handling.

Further, I do wonder if we want to make that behavior configurable. For 
example, one might want to initialize backends sequentially using 16 
threads max, instead of consuming multiple times 16 threads 
concurrently. Could either be a machine option or a memory backend 
option (async-prealloc=on). Once hotplugging such backends, we would 
disable it for now, but could allow it in the future.

-- 
Cheers,

David / dhildenb

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[Daniel P. Berrangé]

On Tue, Jan 09, 2024 at 03:02:00PM +0100, David Hildenbrand wrote:
> On 08.01.24 19:40, Mark Kanda wrote:
> > On 1/8/24 9:40 AM, David Hildenbrand wrote:
> > > On 08.01.24 16:10, Mark Kanda wrote:
> > > > Refactor the memory prealloc threads support:
> > > > - Make memset context a global qlist
> > > > - Move the memset thread join/cleanup code to a separate routine
> > > > 
> > > > This is functionally equivalent and facilitates multiple memset contexts
> > > > (used in a subsequent patch).
> > > > 
> > > > Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
> > > > ---
> > > >    util/oslib-posix.c | 104 +++++++++++++++++++++++++++++----------------
> > > >    1 file changed, 68 insertions(+), 36 deletions(-)
> > > > 
> > > > diff --git a/util/oslib-posix.c b/util/oslib-posix.c
> > > > index e86fd64e09..293297ac6c 100644
> > > > --- a/util/oslib-posix.c
> > > > +++ b/util/oslib-posix.c
> > > > @@ -63,11 +63,15 @@
> > > >      struct MemsetThread;
> > > >    +static QLIST_HEAD(, MemsetContext) memset_contexts =
> > > > +    QLIST_HEAD_INITIALIZER(memset_contexts);
> > > > +
> > > >    typedef struct MemsetContext {
> > > >        bool all_threads_created;
> > > >        bool any_thread_failed;
> > > >        struct MemsetThread *threads;
> > > >        int num_threads;
> > > > +    QLIST_ENTRY(MemsetContext) next;
> > > >    } MemsetContext;
> > > >      struct MemsetThread {
> > > > @@ -81,7 +85,7 @@ struct MemsetThread {
> > > >    typedef struct MemsetThread MemsetThread;
> > > >      /* used by sigbus_handler() */
> > > > -static MemsetContext *sigbus_memset_context;
> > > > +static bool sigbus_memset_context;
> > > >    struct sigaction sigbus_oldact;
> > > >    static QemuMutex sigbus_mutex;
> > > >    @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
> > > >    #endif /* CONFIG_LINUX */
> > > >    {
> > > >        int i;
> > > > +    MemsetContext *context;
> > > >          if (sigbus_memset_context) {
> > > > -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
> > > > -            MemsetThread *thread = &sigbus_memset_context->threads[i];
> > > > +        QLIST_FOREACH(context, &memset_contexts, next) {
> > > > +            for (i = 0; i < context->num_threads; i++) {
> > > > +                MemsetThread *thread = &context->threads[i];
> > > >    -            if (qemu_thread_is_self(&thread->pgthread)) {
> > > > -                siglongjmp(thread->env, 1);
> > > > +                if (qemu_thread_is_self(&thread->pgthread)) {
> > > > +                    siglongjmp(thread->env, 1);
> > > > +                }
> > > >                }
> > > >            }
> > > >        }
> > > > @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t
> > > > hpagesize, size_t numpages,
> > > >                               bool use_madv_populate_write)
> > > >    {
> > > >        static gsize initialized = 0;
> > > > -    MemsetContext context = {
> > > > -        .num_threads = get_memset_num_threads(hpagesize, numpages,
> > > > max_threads),
> > > > -    };
> > > > +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
> > > >        size_t numpages_per_thread, leftover;
> > > >        void *(*touch_fn)(void *);
> > > > -    int ret = 0, i = 0;
> > > > +    int i = 0;
> > > >        char *addr = area;
> > > >    +    context->num_threads =
> > > > +        get_memset_num_threads(hpagesize, numpages, max_threads);
> > > > +
> > > >        if (g_once_init_enter(&initialized)) {
> > > >            qemu_mutex_init(&page_mutex);
> > > >            qemu_cond_init(&page_cond);
> > > > @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t
> > > > hpagesize, size_t numpages,
> > > >          if (use_madv_populate_write) {
> > > >            /* Avoid creating a single thread for MADV_POPULATE_WRITE */
> > > > -        if (context.num_threads == 1) {
> > > > +        if (context->num_threads == 1) {
> > > >                if (qemu_madvise(area, hpagesize * numpages,
> > > >                                 QEMU_MADV_POPULATE_WRITE)) {
> > > >                    return -errno;
> > > > @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t
> > > > hpagesize, size_t numpages,
> > > >            touch_fn = do_touch_pages;
> > > >        }
> > > >    -    context.threads = g_new0(MemsetThread, context.num_threads);
> > > > -    numpages_per_thread = numpages / context.num_threads;
> > > > -    leftover = numpages % context.num_threads;
> > > > -    for (i = 0; i < context.num_threads; i++) {
> > > > -        context.threads[i].addr = addr;
> > > > -        context.threads[i].numpages = numpages_per_thread + (i <
> > > > leftover);
> > > > -        context.threads[i].hpagesize = hpagesize;
> > > > -        context.threads[i].context = &context;
> > > > +    context->threads = g_new0(MemsetThread, context->num_threads);
> > > > +    numpages_per_thread = numpages / context->num_threads;
> > > > +    leftover = numpages % context->num_threads;
> > > > +    for (i = 0; i < context->num_threads; i++) {
> > > > +        context->threads[i].addr = addr;
> > > > +        context->threads[i].numpages = numpages_per_thread + (i <
> > > > leftover);
> > > > +        context->threads[i].hpagesize = hpagesize;
> > > > +        context->threads[i].context = context;
> > > >            if (tc) {
> > > > -            thread_context_create_thread(tc,
> > > > &context.threads[i].pgthread,
> > > > +            thread_context_create_thread(tc,
> > > > &context->threads[i].pgthread,
> > > >                                             "touch_pages",
> > > > -                                         touch_fn, &context.threads[i],
> > > > +                                         touch_fn,
> > > > &context->threads[i],
> > > > QEMU_THREAD_JOINABLE);
> > > >            } else {
> > > > - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
> > > > -                               touch_fn, &context.threads[i],
> > > > + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
> > > > +                               touch_fn, &context->threads[i],
> > > >                                   QEMU_THREAD_JOINABLE);
> > > >            }
> > > > -        addr += context.threads[i].numpages * hpagesize;
> > > > +        addr += context->threads[i].numpages * hpagesize;
> > > >        }
> > > >    +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
> > > > +
> > > >        if (!use_madv_populate_write) {
> > > > -        sigbus_memset_context = &context;
> > > > +        sigbus_memset_context = true;
> > > 
> > Thanks David,
> > 
> > > Could we just use the sigbus handling alone and support parallel init
> > > only when using MADV_POPULATE_WRITE where we don't have to mess with
> > > signal handlers?
> > > 
> > 
> > Ideally, we're hoping to support this with earlier kernels which don't
> > support MADV_POPULATE _WRITE. But, I will check to see if we really need it.
> 
> That's around since Linux 5.14, so please try simplifying.
> 
> > 
> > > Further, how do you changes interact with other callers of
> > > qemu_prealloc_mem(), like virtio-mem?
> > > 
> > 
> > I'm not familiar with the intricacies of virtio-mem, but the basic idea
> > of this series is to *only* allow parallel init during the start up
> > phase (while prealloc_init == false). Once we have parsed all the
> > command line args, we set prealloc_init = true
> > (wait_mem_prealloc_init()) which causes all subsequent calls to
> > qemu_prealloc_mem() to perform initialization synchronously. So, I
> > *think* this covers the virtio-mem use case. Am I missing something?
> 
> Good, so this should likely not affect virtio-mem (which also ends up
> preallocating memory when loading from a vmstate).
> 
> To make this all a bit clearer, what about the following to make this:
> 
> * Optimize for MADV_POPULATE_WRITE. If we really need support for
>   !MADV_POPULATE_WRITE, this is better added on top later.
> * Pass in via a parameter that the caller requests async handling. "bool
>   async" should be good enough. Then, pass that only from the memory
>   backend call, while QEMU is still initializing (we can find a way to
>   make that work).
> * Provide a function that waits for any remaining async os-mem-prealloc
>   activity. That is essentially "wait_mem_prealloc_init", but without
>   the special internal flag handling.
> 
> Further, I do wonder if we want to make that behavior configurable. For
> example, one might want to initialize backends sequentially using 16 threads
> max, instead of consuming multiple times 16 threads concurrently.

Seems to me that no matter what parallelisation we use (within
mem regions, or across mem regions, or a mix of both), we should
never use more threads than there are host CPUs.

Can we have a pool of threads sized per available host CPUs that
QEMU can access. Then for each memory backend fire off a set of
init jobs that get distributed to "appropriate" threads in the
pool. By appropriate I mean threads with the same NUMA affinity
as the memory backend. This would give parallelisation both
within a single large memory region, as well as across memory
regions.


eg 4 host CPUs, 3 memory regions (1GB for 1st numa node, 1GB
for 2nd numa node, and 1 GB with no numa affinity). If we
spread the init work then we end up with

 1st thread gets 500MB to init from 1st memory region, and
     250MB to init from 3rd memory region

 2st thread gets 500MB to init from 1st memory region, and
     250MB to init from 3rd memory region

 3st thread gets 500MB to init from 2nd memory region, and
     250MB to init from 3rd memory region

 4th thread gets 500MB to init from 2nd memory region, and
     250MB to init from 3rd memory region

>                                                                    Could
> either be a machine option or a memory backend option (async-prealloc=on).
> Once hotplugging such backends, we would disable it for now, but could allow
> it in the future.

IMHO "do the right thing" without user config is preferrable to adding
yet more cli options here.

With regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[David Hildenbrand]

On 09.01.24 15:15, Daniel P. Berrangé wrote:
> On Tue, Jan 09, 2024 at 03:02:00PM +0100, David Hildenbrand wrote:
>> On 08.01.24 19:40, Mark Kanda wrote:
>>> On 1/8/24 9:40 AM, David Hildenbrand wrote:
>>>> On 08.01.24 16:10, Mark Kanda wrote:
>>>>> Refactor the memory prealloc threads support:
>>>>> - Make memset context a global qlist
>>>>> - Move the memset thread join/cleanup code to a separate routine
>>>>>
>>>>> This is functionally equivalent and facilitates multiple memset contexts
>>>>> (used in a subsequent patch).
>>>>>
>>>>> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
>>>>> ---
>>>>>     util/oslib-posix.c | 104 +++++++++++++++++++++++++++++----------------
>>>>>     1 file changed, 68 insertions(+), 36 deletions(-)
>>>>>
>>>>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
>>>>> index e86fd64e09..293297ac6c 100644
>>>>> --- a/util/oslib-posix.c
>>>>> +++ b/util/oslib-posix.c
>>>>> @@ -63,11 +63,15 @@
>>>>>       struct MemsetThread;
>>>>>     +static QLIST_HEAD(, MemsetContext) memset_contexts =
>>>>> +    QLIST_HEAD_INITIALIZER(memset_contexts);
>>>>> +
>>>>>     typedef struct MemsetContext {
>>>>>         bool all_threads_created;
>>>>>         bool any_thread_failed;
>>>>>         struct MemsetThread *threads;
>>>>>         int num_threads;
>>>>> +    QLIST_ENTRY(MemsetContext) next;
>>>>>     } MemsetContext;
>>>>>       struct MemsetThread {
>>>>> @@ -81,7 +85,7 @@ struct MemsetThread {
>>>>>     typedef struct MemsetThread MemsetThread;
>>>>>       /* used by sigbus_handler() */
>>>>> -static MemsetContext *sigbus_memset_context;
>>>>> +static bool sigbus_memset_context;
>>>>>     struct sigaction sigbus_oldact;
>>>>>     static QemuMutex sigbus_mutex;
>>>>>     @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>>>>>     #endif /* CONFIG_LINUX */
>>>>>     {
>>>>>         int i;
>>>>> +    MemsetContext *context;
>>>>>           if (sigbus_memset_context) {
>>>>> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
>>>>> -            MemsetThread *thread = &sigbus_memset_context->threads[i];
>>>>> +        QLIST_FOREACH(context, &memset_contexts, next) {
>>>>> +            for (i = 0; i < context->num_threads; i++) {
>>>>> +                MemsetThread *thread = &context->threads[i];
>>>>>     -            if (qemu_thread_is_self(&thread->pgthread)) {
>>>>> -                siglongjmp(thread->env, 1);
>>>>> +                if (qemu_thread_is_self(&thread->pgthread)) {
>>>>> +                    siglongjmp(thread->env, 1);
>>>>> +                }
>>>>>                 }
>>>>>             }
>>>>>         }
>>>>> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t
>>>>> hpagesize, size_t numpages,
>>>>>                                bool use_madv_populate_write)
>>>>>     {
>>>>>         static gsize initialized = 0;
>>>>> -    MemsetContext context = {
>>>>> -        .num_threads = get_memset_num_threads(hpagesize, numpages,
>>>>> max_threads),
>>>>> -    };
>>>>> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>>>>>         size_t numpages_per_thread, leftover;
>>>>>         void *(*touch_fn)(void *);
>>>>> -    int ret = 0, i = 0;
>>>>> +    int i = 0;
>>>>>         char *addr = area;
>>>>>     +    context->num_threads =
>>>>> +        get_memset_num_threads(hpagesize, numpages, max_threads);
>>>>> +
>>>>>         if (g_once_init_enter(&initialized)) {
>>>>>             qemu_mutex_init(&page_mutex);
>>>>>             qemu_cond_init(&page_cond);
>>>>> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t
>>>>> hpagesize, size_t numpages,
>>>>>           if (use_madv_populate_write) {
>>>>>             /* Avoid creating a single thread for MADV_POPULATE_WRITE */
>>>>> -        if (context.num_threads == 1) {
>>>>> +        if (context->num_threads == 1) {
>>>>>                 if (qemu_madvise(area, hpagesize * numpages,
>>>>>                                  QEMU_MADV_POPULATE_WRITE)) {
>>>>>                     return -errno;
>>>>> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t
>>>>> hpagesize, size_t numpages,
>>>>>             touch_fn = do_touch_pages;
>>>>>         }
>>>>>     -    context.threads = g_new0(MemsetThread, context.num_threads);
>>>>> -    numpages_per_thread = numpages / context.num_threads;
>>>>> -    leftover = numpages % context.num_threads;
>>>>> -    for (i = 0; i < context.num_threads; i++) {
>>>>> -        context.threads[i].addr = addr;
>>>>> -        context.threads[i].numpages = numpages_per_thread + (i <
>>>>> leftover);
>>>>> -        context.threads[i].hpagesize = hpagesize;
>>>>> -        context.threads[i].context = &context;
>>>>> +    context->threads = g_new0(MemsetThread, context->num_threads);
>>>>> +    numpages_per_thread = numpages / context->num_threads;
>>>>> +    leftover = numpages % context->num_threads;
>>>>> +    for (i = 0; i < context->num_threads; i++) {
>>>>> +        context->threads[i].addr = addr;
>>>>> +        context->threads[i].numpages = numpages_per_thread + (i <
>>>>> leftover);
>>>>> +        context->threads[i].hpagesize = hpagesize;
>>>>> +        context->threads[i].context = context;
>>>>>             if (tc) {
>>>>> -            thread_context_create_thread(tc,
>>>>> &context.threads[i].pgthread,
>>>>> +            thread_context_create_thread(tc,
>>>>> &context->threads[i].pgthread,
>>>>>                                              "touch_pages",
>>>>> -                                         touch_fn, &context.threads[i],
>>>>> +                                         touch_fn,
>>>>> &context->threads[i],
>>>>> QEMU_THREAD_JOINABLE);
>>>>>             } else {
>>>>> - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
>>>>> -                               touch_fn, &context.threads[i],
>>>>> + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
>>>>> +                               touch_fn, &context->threads[i],
>>>>>                                    QEMU_THREAD_JOINABLE);
>>>>>             }
>>>>> -        addr += context.threads[i].numpages * hpagesize;
>>>>> +        addr += context->threads[i].numpages * hpagesize;
>>>>>         }
>>>>>     +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
>>>>> +
>>>>>         if (!use_madv_populate_write) {
>>>>> -        sigbus_memset_context = &context;
>>>>> +        sigbus_memset_context = true;
>>>>
>>> Thanks David,
>>>
>>>> Could we just use the sigbus handling alone and support parallel init
>>>> only when using MADV_POPULATE_WRITE where we don't have to mess with
>>>> signal handlers?
>>>>
>>>
>>> Ideally, we're hoping to support this with earlier kernels which don't
>>> support MADV_POPULATE _WRITE. But, I will check to see if we really need it.
>>
>> That's around since Linux 5.14, so please try simplifying.
>>
>>>
>>>> Further, how do you changes interact with other callers of
>>>> qemu_prealloc_mem(), like virtio-mem?
>>>>
>>>
>>> I'm not familiar with the intricacies of virtio-mem, but the basic idea
>>> of this series is to *only* allow parallel init during the start up
>>> phase (while prealloc_init == false). Once we have parsed all the
>>> command line args, we set prealloc_init = true
>>> (wait_mem_prealloc_init()) which causes all subsequent calls to
>>> qemu_prealloc_mem() to perform initialization synchronously. So, I
>>> *think* this covers the virtio-mem use case. Am I missing something?
>>
>> Good, so this should likely not affect virtio-mem (which also ends up
>> preallocating memory when loading from a vmstate).
>>
>> To make this all a bit clearer, what about the following to make this:
>>
>> * Optimize for MADV_POPULATE_WRITE. If we really need support for
>>    !MADV_POPULATE_WRITE, this is better added on top later.
>> * Pass in via a parameter that the caller requests async handling. "bool
>>    async" should be good enough. Then, pass that only from the memory
>>    backend call, while QEMU is still initializing (we can find a way to
>>    make that work).
>> * Provide a function that waits for any remaining async os-mem-prealloc
>>    activity. That is essentially "wait_mem_prealloc_init", but without
>>    the special internal flag handling.
>>
>> Further, I do wonder if we want to make that behavior configurable. For
>> example, one might want to initialize backends sequentially using 16 threads
>> max, instead of consuming multiple times 16 threads concurrently.
> 
> Seems to me that no matter what parallelisation we use (within
> mem regions, or across mem regions, or a mix of both), we should
> never use more threads than there are host CPUs.

Yes. It gets tricky with multipe NUMA nodes, though. And that's what's 
being requested here.

> 
> Can we have a pool of threads sized per available host CPUs that
> QEMU can access. Then for each memory backend fire off a set of
> init jobs that get distributed to "appropriate" threads in the
> pool. By appropriate I mean threads with the same NUMA affinity
> as the memory backend. This would give parallelisation both
> within a single large memory region, as well as across memory
> regions.
> 
> 
> eg 4 host CPUs, 3 memory regions (1GB for 1st numa node, 1GB
> for 2nd numa node, and 1 GB with no numa affinity). If we
> spread the init work then we end up with

I'll note that having a mixture of numa + no numa is not a common config 
we should try to optimize.

> 
>   1st thread gets 500MB to init from 1st memory region, and
>       250MB to init from 3rd memory region
> 
>   2st thread gets 500MB to init from 1st memory region, and
>       250MB to init from 3rd memory region
> 
>   3st thread gets 500MB to init from 2nd memory region, and
>       250MB to init from 3rd memory region
> 
>   4th thread gets 500MB to init from 2nd memory region, and
>       250MB to init from 3rd memory region

We do have prealloc contexts that are used to craft new CPUs with the 
right NUMA affinity. Prealloc contexts are set for memory backends. So 
extending prealloc context to limit/reuse threads and do the pooling 
might be possible.

> 
>>                                                                     Could
>> either be a machine option or a memory backend option (async-prealloc=on).
>> Once hotplugging such backends, we would disable it for now, but could allow
>> it in the future.
> 
> IMHO "do the right thing" without user config is preferrable to adding
> yet more cli options here.

Sure, if there is an easy way.

-- 
Cheers,

David / dhildenb

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[Mark Kanda]

On 1/9/24 8:25 AM, David Hildenbrand wrote:
> On 09.01.24 15:15, Daniel P. Berrangé wrote:
>> On Tue, Jan 09, 2024 at 03:02:00PM +0100, David Hildenbrand wrote:
>>> On 08.01.24 19:40, Mark Kanda wrote:
>>>> On 1/8/24 9:40 AM, David Hildenbrand wrote:
>>>>> On 08.01.24 16:10, Mark Kanda wrote:
>>>>>> Refactor the memory prealloc threads support:
>>>>>> - Make memset context a global qlist
>>>>>> - Move the memset thread join/cleanup code to a separate routine
>>>>>>
>>>>>> This is functionally equivalent and facilitates multiple memset 
>>>>>> contexts
>>>>>> (used in a subsequent patch).
>>>>>>
>>>>>> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
>>>>>> ---
>>>>>>     util/oslib-posix.c | 104 
>>>>>> +++++++++++++++++++++++++++++----------------
>>>>>>     1 file changed, 68 insertions(+), 36 deletions(-)
>>>>>>
>>>>>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
>>>>>> index e86fd64e09..293297ac6c 100644
>>>>>> --- a/util/oslib-posix.c
>>>>>> +++ b/util/oslib-posix.c
>>>>>> @@ -63,11 +63,15 @@
>>>>>>       struct MemsetThread;
>>>>>>     +static QLIST_HEAD(, MemsetContext) memset_contexts =
>>>>>> +    QLIST_HEAD_INITIALIZER(memset_contexts);
>>>>>> +
>>>>>>     typedef struct MemsetContext {
>>>>>>         bool all_threads_created;
>>>>>>         bool any_thread_failed;
>>>>>>         struct MemsetThread *threads;
>>>>>>         int num_threads;
>>>>>> +    QLIST_ENTRY(MemsetContext) next;
>>>>>>     } MemsetContext;
>>>>>>       struct MemsetThread {
>>>>>> @@ -81,7 +85,7 @@ struct MemsetThread {
>>>>>>     typedef struct MemsetThread MemsetThread;
>>>>>>       /* used by sigbus_handler() */
>>>>>> -static MemsetContext *sigbus_memset_context;
>>>>>> +static bool sigbus_memset_context;
>>>>>>     struct sigaction sigbus_oldact;
>>>>>>     static QemuMutex sigbus_mutex;
>>>>>>     @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>>>>>>     #endif /* CONFIG_LINUX */
>>>>>>     {
>>>>>>         int i;
>>>>>> +    MemsetContext *context;
>>>>>>           if (sigbus_memset_context) {
>>>>>> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
>>>>>> -            MemsetThread *thread = 
>>>>>> &sigbus_memset_context->threads[i];
>>>>>> +        QLIST_FOREACH(context, &memset_contexts, next) {
>>>>>> +            for (i = 0; i < context->num_threads; i++) {
>>>>>> +                MemsetThread *thread = &context->threads[i];
>>>>>>     -            if (qemu_thread_is_self(&thread->pgthread)) {
>>>>>> -                siglongjmp(thread->env, 1);
>>>>>> +                if (qemu_thread_is_self(&thread->pgthread)) {
>>>>>> +                    siglongjmp(thread->env, 1);
>>>>>> +                }
>>>>>>                 }
>>>>>>             }
>>>>>>         }
>>>>>> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t
>>>>>> hpagesize, size_t numpages,
>>>>>>                                bool use_madv_populate_write)
>>>>>>     {
>>>>>>         static gsize initialized = 0;
>>>>>> -    MemsetContext context = {
>>>>>> -        .num_threads = get_memset_num_threads(hpagesize, numpages,
>>>>>> max_threads),
>>>>>> -    };
>>>>>> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>>>>>>         size_t numpages_per_thread, leftover;
>>>>>>         void *(*touch_fn)(void *);
>>>>>> -    int ret = 0, i = 0;
>>>>>> +    int i = 0;
>>>>>>         char *addr = area;
>>>>>>     +    context->num_threads =
>>>>>> +        get_memset_num_threads(hpagesize, numpages, max_threads);
>>>>>> +
>>>>>>         if (g_once_init_enter(&initialized)) {
>>>>>>             qemu_mutex_init(&page_mutex);
>>>>>>             qemu_cond_init(&page_cond);
>>>>>> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t
>>>>>> hpagesize, size_t numpages,
>>>>>>           if (use_madv_populate_write) {
>>>>>>             /* Avoid creating a single thread for 
>>>>>> MADV_POPULATE_WRITE */
>>>>>> -        if (context.num_threads == 1) {
>>>>>> +        if (context->num_threads == 1) {
>>>>>>                 if (qemu_madvise(area, hpagesize * numpages,
>>>>>> QEMU_MADV_POPULATE_WRITE)) {
>>>>>>                     return -errno;
>>>>>> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t
>>>>>> hpagesize, size_t numpages,
>>>>>>             touch_fn = do_touch_pages;
>>>>>>         }
>>>>>>     -    context.threads = g_new0(MemsetThread, 
>>>>>> context.num_threads);
>>>>>> -    numpages_per_thread = numpages / context.num_threads;
>>>>>> -    leftover = numpages % context.num_threads;
>>>>>> -    for (i = 0; i < context.num_threads; i++) {
>>>>>> -        context.threads[i].addr = addr;
>>>>>> -        context.threads[i].numpages = numpages_per_thread + (i <
>>>>>> leftover);
>>>>>> -        context.threads[i].hpagesize = hpagesize;
>>>>>> -        context.threads[i].context = &context;
>>>>>> +    context->threads = g_new0(MemsetThread, context->num_threads);
>>>>>> +    numpages_per_thread = numpages / context->num_threads;
>>>>>> +    leftover = numpages % context->num_threads;
>>>>>> +    for (i = 0; i < context->num_threads; i++) {
>>>>>> +        context->threads[i].addr = addr;
>>>>>> +        context->threads[i].numpages = numpages_per_thread + (i <
>>>>>> leftover);
>>>>>> +        context->threads[i].hpagesize = hpagesize;
>>>>>> +        context->threads[i].context = context;
>>>>>>             if (tc) {
>>>>>> -            thread_context_create_thread(tc,
>>>>>> &context.threads[i].pgthread,
>>>>>> +            thread_context_create_thread(tc,
>>>>>> &context->threads[i].pgthread,
>>>>>> "touch_pages",
>>>>>> -                                         touch_fn, 
>>>>>> &context.threads[i],
>>>>>> +                                         touch_fn,
>>>>>> &context->threads[i],
>>>>>> QEMU_THREAD_JOINABLE);
>>>>>>             } else {
>>>>>> - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
>>>>>> -                               touch_fn, &context.threads[i],
>>>>>> + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
>>>>>> +                               touch_fn, &context->threads[i],
>>>>>> QEMU_THREAD_JOINABLE);
>>>>>>             }
>>>>>> -        addr += context.threads[i].numpages * hpagesize;
>>>>>> +        addr += context->threads[i].numpages * hpagesize;
>>>>>>         }
>>>>>>     +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
>>>>>> +
>>>>>>         if (!use_madv_populate_write) {
>>>>>> -        sigbus_memset_context = &context;
>>>>>> +        sigbus_memset_context = true;
>>>>>
>>>> Thanks David,
>>>>
>>>>> Could we just use the sigbus handling alone and support parallel init
>>>>> only when using MADV_POPULATE_WRITE where we don't have to mess with
>>>>> signal handlers?
>>>>>
>>>>
>>>> Ideally, we're hoping to support this with earlier kernels which don't
>>>> support MADV_POPULATE _WRITE. But, I will check to see if we really 
>>>> need it.
>>>
>>> That's around since Linux 5.14, so please try simplifying.
>>>
>>>>
>>>>> Further, how do you changes interact with other callers of
>>>>> qemu_prealloc_mem(), like virtio-mem?
>>>>>
>>>>
>>>> I'm not familiar with the intricacies of virtio-mem, but the basic 
>>>> idea
>>>> of this series is to *only* allow parallel init during the start up
>>>> phase (while prealloc_init == false). Once we have parsed all the
>>>> command line args, we set prealloc_init = true
>>>> (wait_mem_prealloc_init()) which causes all subsequent calls to
>>>> qemu_prealloc_mem() to perform initialization synchronously. So, I
>>>> *think* this covers the virtio-mem use case. Am I missing something?
>>>
>>> Good, so this should likely not affect virtio-mem (which also ends up
>>> preallocating memory when loading from a vmstate).
>>>
>>> To make this all a bit clearer, what about the following to make this:
>>>
>>> * Optimize for MADV_POPULATE_WRITE. If we really need support for
>>>    !MADV_POPULATE_WRITE, this is better added on top later.
>>> * Pass in via a parameter that the caller requests async handling. 
>>> "bool
>>>    async" should be good enough. Then, pass that only from the memory
>>>    backend call, while QEMU is still initializing (we can find a way to
>>>    make that work).
>>> * Provide a function that waits for any remaining async os-mem-prealloc
>>>    activity. That is essentially "wait_mem_prealloc_init", but without
>>>    the special internal flag handling.
>>>
>>> Further, I do wonder if we want to make that behavior configurable. For
>>> example, one might want to initialize backends sequentially using 16 
>>> threads
>>> max, instead of consuming multiple times 16 threads concurrently.
>>
>> Seems to me that no matter what parallelisation we use (within
>> mem regions, or across mem regions, or a mix of both), we should
>> never use more threads than there are host CPUs.
>
> Yes. It gets tricky with multipe NUMA nodes, though. And that's what's 
> being requested here.
>
>>
>> Can we have a pool of threads sized per available host CPUs that
>> QEMU can access. Then for each memory backend fire off a set of
>> init jobs that get distributed to "appropriate" threads in the
>> pool. By appropriate I mean threads with the same NUMA affinity
>> as the memory backend. This would give parallelisation both
>> within a single large memory region, as well as across memory
>> regions.
>>
>>
>> eg 4 host CPUs, 3 memory regions (1GB for 1st numa node, 1GB
>> for 2nd numa node, and 1 GB with no numa affinity). If we
>> spread the init work then we end up with
>
> I'll note that having a mixture of numa + no numa is not a common 
> config we should try to optimize.
>
>>
>>   1st thread gets 500MB to init from 1st memory region, and
>>       250MB to init from 3rd memory region
>>
>>   2st thread gets 500MB to init from 1st memory region, and
>>       250MB to init from 3rd memory region
>>
>>   3st thread gets 500MB to init from 2nd memory region, and
>>       250MB to init from 3rd memory region
>>
>>   4th thread gets 500MB to init from 2nd memory region, and
>>       250MB to init from 3rd memory region
>
> We do have prealloc contexts that are used to craft new CPUs with the 
> right NUMA affinity. Prealloc contexts are set for memory backends. So 
> extending prealloc context to limit/reuse threads and do the pooling 
> might be possible.
>
>>
>>> Could
>>> either be a machine option or a memory backend option 
>>> (async-prealloc=on).
>>> Once hotplugging such backends, we would disable it for now, but 
>>> could allow
>>> it in the future.
>>
>> IMHO "do the right thing" without user config is preferrable to adding
>> yet more cli options here.
>
> Sure, if there is an easy way.
>

Thank you David and Daniel for your feedback. I plan to address your 
suggestions in v2 (including optimizing for MADV_POPULATE_WRITE).

Best regards,
-Mark

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[Mark Kanda]

On 1/9/24 8:25 AM, David Hildenbrand wrote:
> On 09.01.24 15:15, Daniel P. Berrangé wrote:
>> On Tue, Jan 09, 2024 at 03:02:00PM +0100, David Hildenbrand wrote:
>>> On 08.01.24 19:40, Mark Kanda wrote:
>>>> On 1/8/24 9:40 AM, David Hildenbrand wrote:
>>>>> On 08.01.24 16:10, Mark Kanda wrote:
>>>>>> Refactor the memory prealloc threads support:
>>>>>> - Make memset context a global qlist
>>>>>> - Move the memset thread join/cleanup code to a separate routine
>>>>>>
>>>>>> This is functionally equivalent and facilitates multiple memset 
>>>>>> contexts
>>>>>> (used in a subsequent patch).
>>>>>>
>>>>>> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
>>>>>> ---
>>>>>>     util/oslib-posix.c | 104 
>>>>>> +++++++++++++++++++++++++++++----------------
>>>>>>     1 file changed, 68 insertions(+), 36 deletions(-)
>>>>>>
>>>>>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
>>>>>> index e86fd64e09..293297ac6c 100644
>>>>>> --- a/util/oslib-posix.c
>>>>>> +++ b/util/oslib-posix.c
>>>>>> @@ -63,11 +63,15 @@
>>>>>>       struct MemsetThread;
>>>>>>     +static QLIST_HEAD(, MemsetContext) memset_contexts =
>>>>>> +    QLIST_HEAD_INITIALIZER(memset_contexts);
>>>>>> +
>>>>>>     typedef struct MemsetContext {
>>>>>>         bool all_threads_created;
>>>>>>         bool any_thread_failed;
>>>>>>         struct MemsetThread *threads;
>>>>>>         int num_threads;
>>>>>> +    QLIST_ENTRY(MemsetContext) next;
>>>>>>     } MemsetContext;
>>>>>>       struct MemsetThread {
>>>>>> @@ -81,7 +85,7 @@ struct MemsetThread {
>>>>>>     typedef struct MemsetThread MemsetThread;
>>>>>>       /* used by sigbus_handler() */
>>>>>> -static MemsetContext *sigbus_memset_context;
>>>>>> +static bool sigbus_memset_context;
>>>>>>     struct sigaction sigbus_oldact;
>>>>>>     static QemuMutex sigbus_mutex;
>>>>>>     @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>>>>>>     #endif /* CONFIG_LINUX */
>>>>>>     {
>>>>>>         int i;
>>>>>> +    MemsetContext *context;
>>>>>>           if (sigbus_memset_context) {
>>>>>> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
>>>>>> -            MemsetThread *thread = 
>>>>>> &sigbus_memset_context->threads[i];
>>>>>> +        QLIST_FOREACH(context, &memset_contexts, next) {
>>>>>> +            for (i = 0; i < context->num_threads; i++) {
>>>>>> +                MemsetThread *thread = &context->threads[i];
>>>>>>     -            if (qemu_thread_is_self(&thread->pgthread)) {
>>>>>> -                siglongjmp(thread->env, 1);
>>>>>> +                if (qemu_thread_is_self(&thread->pgthread)) {
>>>>>> +                    siglongjmp(thread->env, 1);
>>>>>> +                }
>>>>>>                 }
>>>>>>             }
>>>>>>         }
>>>>>> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t
>>>>>> hpagesize, size_t numpages,
>>>>>>                                bool use_madv_populate_write)
>>>>>>     {
>>>>>>         static gsize initialized = 0;
>>>>>> -    MemsetContext context = {
>>>>>> -        .num_threads = get_memset_num_threads(hpagesize, numpages,
>>>>>> max_threads),
>>>>>> -    };
>>>>>> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>>>>>>         size_t numpages_per_thread, leftover;
>>>>>>         void *(*touch_fn)(void *);
>>>>>> -    int ret = 0, i = 0;
>>>>>> +    int i = 0;
>>>>>>         char *addr = area;
>>>>>>     +    context->num_threads =
>>>>>> +        get_memset_num_threads(hpagesize, numpages, max_threads);
>>>>>> +
>>>>>>         if (g_once_init_enter(&initialized)) {
>>>>>>             qemu_mutex_init(&page_mutex);
>>>>>>             qemu_cond_init(&page_cond);
>>>>>> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t
>>>>>> hpagesize, size_t numpages,
>>>>>>           if (use_madv_populate_write) {
>>>>>>             /* Avoid creating a single thread for 
>>>>>> MADV_POPULATE_WRITE */
>>>>>> -        if (context.num_threads == 1) {
>>>>>> +        if (context->num_threads == 1) {
>>>>>>                 if (qemu_madvise(area, hpagesize * numpages,
>>>>>> QEMU_MADV_POPULATE_WRITE)) {
>>>>>>                     return -errno;
>>>>>> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t
>>>>>> hpagesize, size_t numpages,
>>>>>>             touch_fn = do_touch_pages;
>>>>>>         }
>>>>>>     -    context.threads = g_new0(MemsetThread, 
>>>>>> context.num_threads);
>>>>>> -    numpages_per_thread = numpages / context.num_threads;
>>>>>> -    leftover = numpages % context.num_threads;
>>>>>> -    for (i = 0; i < context.num_threads; i++) {
>>>>>> -        context.threads[i].addr = addr;
>>>>>> -        context.threads[i].numpages = numpages_per_thread + (i <
>>>>>> leftover);
>>>>>> -        context.threads[i].hpagesize = hpagesize;
>>>>>> -        context.threads[i].context = &context;
>>>>>> +    context->threads = g_new0(MemsetThread, context->num_threads);
>>>>>> +    numpages_per_thread = numpages / context->num_threads;
>>>>>> +    leftover = numpages % context->num_threads;
>>>>>> +    for (i = 0; i < context->num_threads; i++) {
>>>>>> +        context->threads[i].addr = addr;
>>>>>> +        context->threads[i].numpages = numpages_per_thread + (i <
>>>>>> leftover);
>>>>>> +        context->threads[i].hpagesize = hpagesize;
>>>>>> +        context->threads[i].context = context;
>>>>>>             if (tc) {
>>>>>> -            thread_context_create_thread(tc,
>>>>>> &context.threads[i].pgthread,
>>>>>> +            thread_context_create_thread(tc,
>>>>>> &context->threads[i].pgthread,
>>>>>> "touch_pages",
>>>>>> -                                         touch_fn, 
>>>>>> &context.threads[i],
>>>>>> +                                         touch_fn,
>>>>>> &context->threads[i],
>>>>>> QEMU_THREAD_JOINABLE);
>>>>>>             } else {
>>>>>> - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
>>>>>> -                               touch_fn, &context.threads[i],
>>>>>> + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
>>>>>> +                               touch_fn, &context->threads[i],
>>>>>> QEMU_THREAD_JOINABLE);
>>>>>>             }
>>>>>> -        addr += context.threads[i].numpages * hpagesize;
>>>>>> +        addr += context->threads[i].numpages * hpagesize;
>>>>>>         }
>>>>>>     +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
>>>>>> +
>>>>>>         if (!use_madv_populate_write) {
>>>>>> -        sigbus_memset_context = &context;
>>>>>> +        sigbus_memset_context = true;
>>>>>
>>>> Thanks David,
>>>>
>>>>> Could we just use the sigbus handling alone and support parallel init
>>>>> only when using MADV_POPULATE_WRITE where we don't have to mess with
>>>>> signal handlers?
>>>>>
>>>>
>>>> Ideally, we're hoping to support this with earlier kernels which don't
>>>> support MADV_POPULATE _WRITE. But, I will check to see if we really 
>>>> need it.
>>>
>>> That's around since Linux 5.14, so please try simplifying.
>>>
>>>>
>>>>> Further, how do you changes interact with other callers of
>>>>> qemu_prealloc_mem(), like virtio-mem?
>>>>>
>>>>
>>>> I'm not familiar with the intricacies of virtio-mem, but the basic 
>>>> idea
>>>> of this series is to *only* allow parallel init during the start up
>>>> phase (while prealloc_init == false). Once we have parsed all the
>>>> command line args, we set prealloc_init = true
>>>> (wait_mem_prealloc_init()) which causes all subsequent calls to
>>>> qemu_prealloc_mem() to perform initialization synchronously. So, I
>>>> *think* this covers the virtio-mem use case. Am I missing something?
>>>
>>> Good, so this should likely not affect virtio-mem (which also ends up
>>> preallocating memory when loading from a vmstate).
>>>
>>> To make this all a bit clearer, what about the following to make this:
>>>
>>> * Optimize for MADV_POPULATE_WRITE. If we really need support for
>>>    !MADV_POPULATE_WRITE, this is better added on top later.
>>> * Pass in via a parameter that the caller requests async handling. 
>>> "bool
>>>    async" should be good enough. Then, pass that only from the memory
>>>    backend call, while QEMU is still initializing (we can find a way to
>>>    make that work).
>>> * Provide a function that waits for any remaining async os-mem-prealloc
>>>    activity. That is essentially "wait_mem_prealloc_init", but without
>>>    the special internal flag handling.
>>>
>>> Further, I do wonder if we want to make that behavior configurable. For
>>> example, one might want to initialize backends sequentially using 16 
>>> threads
>>> max, instead of consuming multiple times 16 threads concurrently.
>>
>> Seems to me that no matter what parallelisation we use (within
>> mem regions, or across mem regions, or a mix of both), we should
>> never use more threads than there are host CPUs.
>
> Yes. It gets tricky with multipe NUMA nodes, though. And that's what's 
> being requested here.
>
>>
>> Can we have a pool of threads sized per available host CPUs that
>> QEMU can access. Then for each memory backend fire off a set of
>> init jobs that get distributed to "appropriate" threads in the
>> pool. By appropriate I mean threads with the same NUMA affinity
>> as the memory backend. This would give parallelisation both
>> within a single large memory region, as well as across memory
>> regions.
>>
>>
>> eg 4 host CPUs, 3 memory regions (1GB for 1st numa node, 1GB
>> for 2nd numa node, and 1 GB with no numa affinity). If we
>> spread the init work then we end up with
>
> I'll note that having a mixture of numa + no numa is not a common 
> config we should try to optimize.
>
>>
>>   1st thread gets 500MB to init from 1st memory region, and
>>       250MB to init from 3rd memory region
>>
>>   2st thread gets 500MB to init from 1st memory region, and
>>       250MB to init from 3rd memory region
>>
>>   3st thread gets 500MB to init from 2nd memory region, and
>>       250MB to init from 3rd memory region
>>
>>   4th thread gets 500MB to init from 2nd memory region, and
>>       250MB to init from 3rd memory region
>
> We do have prealloc contexts that are used to craft new CPUs with the 
> right NUMA affinity. Prealloc contexts are set for memory backends. So 
> extending prealloc context to limit/reuse threads and do the pooling 
> might be possible.
>

 From a NUMA perspective, I feel prealloc contexts already supports what 
we want - i.e. we can set the number of prealloc threads and their node 
affinity per memory region. As such, unless you feel strongly about it, 
I'm inclined to leave additional enhancements (thread pool) as a future 
project.

Since multiple NUMA node configs are the primary motivation for parallel 
init, I do wonder if it should *only* be used in conjunction 
with prealloc context. Perhaps we should have parallel init only enabled 
for memory regions with prealloc context?

Thanks/regards,
-Mark

>>
>>> Could
>>> either be a machine option or a memory backend option 
>>> (async-prealloc=on).
>>> Once hotplugging such backends, we would disable it for now, but 
>>> could allow
>>> it in the future.
>>
>> IMHO "do the right thing" without user config is preferrable to adding
>> yet more cli options here.
>
> Sure, if there is an easy way.
>

Re: [PATCH v1 1/2] oslib-posix: refactor memory prealloc threads

[David Hildenbrand]

On 17.01.24 15:44, Mark Kanda wrote:
> On 1/9/24 8:25 AM, David Hildenbrand wrote:
>> On 09.01.24 15:15, Daniel P. Berrangé wrote:
>>> On Tue, Jan 09, 2024 at 03:02:00PM +0100, David Hildenbrand wrote:
>>>> On 08.01.24 19:40, Mark Kanda wrote:
>>>>> On 1/8/24 9:40 AM, David Hildenbrand wrote:
>>>>>> On 08.01.24 16:10, Mark Kanda wrote:
>>>>>>> Refactor the memory prealloc threads support:
>>>>>>> - Make memset context a global qlist
>>>>>>> - Move the memset thread join/cleanup code to a separate routine
>>>>>>>
>>>>>>> This is functionally equivalent and facilitates multiple memset
>>>>>>> contexts
>>>>>>> (used in a subsequent patch).
>>>>>>>
>>>>>>> Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
>>>>>>> ---
>>>>>>>      util/oslib-posix.c | 104
>>>>>>> +++++++++++++++++++++++++++++----------------
>>>>>>>      1 file changed, 68 insertions(+), 36 deletions(-)
>>>>>>>
>>>>>>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c
>>>>>>> index e86fd64e09..293297ac6c 100644
>>>>>>> --- a/util/oslib-posix.c
>>>>>>> +++ b/util/oslib-posix.c
>>>>>>> @@ -63,11 +63,15 @@
>>>>>>>        struct MemsetThread;
>>>>>>>      +static QLIST_HEAD(, MemsetContext) memset_contexts =
>>>>>>> +    QLIST_HEAD_INITIALIZER(memset_contexts);
>>>>>>> +
>>>>>>>      typedef struct MemsetContext {
>>>>>>>          bool all_threads_created;
>>>>>>>          bool any_thread_failed;
>>>>>>>          struct MemsetThread *threads;
>>>>>>>          int num_threads;
>>>>>>> +    QLIST_ENTRY(MemsetContext) next;
>>>>>>>      } MemsetContext;
>>>>>>>        struct MemsetThread {
>>>>>>> @@ -81,7 +85,7 @@ struct MemsetThread {
>>>>>>>      typedef struct MemsetThread MemsetThread;
>>>>>>>        /* used by sigbus_handler() */
>>>>>>> -static MemsetContext *sigbus_memset_context;
>>>>>>> +static bool sigbus_memset_context;
>>>>>>>      struct sigaction sigbus_oldact;
>>>>>>>      static QemuMutex sigbus_mutex;
>>>>>>>      @@ -295,13 +299,16 @@ static void sigbus_handler(int signal)
>>>>>>>      #endif /* CONFIG_LINUX */
>>>>>>>      {
>>>>>>>          int i;
>>>>>>> +    MemsetContext *context;
>>>>>>>            if (sigbus_memset_context) {
>>>>>>> -        for (i = 0; i < sigbus_memset_context->num_threads; i++) {
>>>>>>> -            MemsetThread *thread =
>>>>>>> &sigbus_memset_context->threads[i];
>>>>>>> +        QLIST_FOREACH(context, &memset_contexts, next) {
>>>>>>> +            for (i = 0; i < context->num_threads; i++) {
>>>>>>> +                MemsetThread *thread = &context->threads[i];
>>>>>>>      -            if (qemu_thread_is_self(&thread->pgthread)) {
>>>>>>> -                siglongjmp(thread->env, 1);
>>>>>>> +                if (qemu_thread_is_self(&thread->pgthread)) {
>>>>>>> +                    siglongjmp(thread->env, 1);
>>>>>>> +                }
>>>>>>>                  }
>>>>>>>              }
>>>>>>>          }
>>>>>>> @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t
>>>>>>> hpagesize, size_t numpages,
>>>>>>>                                 bool use_madv_populate_write)
>>>>>>>      {
>>>>>>>          static gsize initialized = 0;
>>>>>>> -    MemsetContext context = {
>>>>>>> -        .num_threads = get_memset_num_threads(hpagesize, numpages,
>>>>>>> max_threads),
>>>>>>> -    };
>>>>>>> +    MemsetContext *context = g_malloc0(sizeof(MemsetContext));
>>>>>>>          size_t numpages_per_thread, leftover;
>>>>>>>          void *(*touch_fn)(void *);
>>>>>>> -    int ret = 0, i = 0;
>>>>>>> +    int i = 0;
>>>>>>>          char *addr = area;
>>>>>>>      +    context->num_threads =
>>>>>>> +        get_memset_num_threads(hpagesize, numpages, max_threads);
>>>>>>> +
>>>>>>>          if (g_once_init_enter(&initialized)) {
>>>>>>>              qemu_mutex_init(&page_mutex);
>>>>>>>              qemu_cond_init(&page_cond);
>>>>>>> @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t
>>>>>>> hpagesize, size_t numpages,
>>>>>>>            if (use_madv_populate_write) {
>>>>>>>              /* Avoid creating a single thread for
>>>>>>> MADV_POPULATE_WRITE */
>>>>>>> -        if (context.num_threads == 1) {
>>>>>>> +        if (context->num_threads == 1) {
>>>>>>>                  if (qemu_madvise(area, hpagesize * numpages,
>>>>>>> QEMU_MADV_POPULATE_WRITE)) {
>>>>>>>                      return -errno;
>>>>>>> @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t
>>>>>>> hpagesize, size_t numpages,
>>>>>>>              touch_fn = do_touch_pages;
>>>>>>>          }
>>>>>>>      -    context.threads = g_new0(MemsetThread,
>>>>>>> context.num_threads);
>>>>>>> -    numpages_per_thread = numpages / context.num_threads;
>>>>>>> -    leftover = numpages % context.num_threads;
>>>>>>> -    for (i = 0; i < context.num_threads; i++) {
>>>>>>> -        context.threads[i].addr = addr;
>>>>>>> -        context.threads[i].numpages = numpages_per_thread + (i <
>>>>>>> leftover);
>>>>>>> -        context.threads[i].hpagesize = hpagesize;
>>>>>>> -        context.threads[i].context = &context;
>>>>>>> +    context->threads = g_new0(MemsetThread, context->num_threads);
>>>>>>> +    numpages_per_thread = numpages / context->num_threads;
>>>>>>> +    leftover = numpages % context->num_threads;
>>>>>>> +    for (i = 0; i < context->num_threads; i++) {
>>>>>>> +        context->threads[i].addr = addr;
>>>>>>> +        context->threads[i].numpages = numpages_per_thread + (i <
>>>>>>> leftover);
>>>>>>> +        context->threads[i].hpagesize = hpagesize;
>>>>>>> +        context->threads[i].context = context;
>>>>>>>              if (tc) {
>>>>>>> -            thread_context_create_thread(tc,
>>>>>>> &context.threads[i].pgthread,
>>>>>>> +            thread_context_create_thread(tc,
>>>>>>> &context->threads[i].pgthread,
>>>>>>> "touch_pages",
>>>>>>> -                                         touch_fn,
>>>>>>> &context.threads[i],
>>>>>>> +                                         touch_fn,
>>>>>>> &context->threads[i],
>>>>>>> QEMU_THREAD_JOINABLE);
>>>>>>>              } else {
>>>>>>> - qemu_thread_create(&context.threads[i].pgthread, "touch_pages",
>>>>>>> -                               touch_fn, &context.threads[i],
>>>>>>> + qemu_thread_create(&context->threads[i].pgthread, "touch_pages",
>>>>>>> +                               touch_fn, &context->threads[i],
>>>>>>> QEMU_THREAD_JOINABLE);
>>>>>>>              }
>>>>>>> -        addr += context.threads[i].numpages * hpagesize;
>>>>>>> +        addr += context->threads[i].numpages * hpagesize;
>>>>>>>          }
>>>>>>>      +    QLIST_INSERT_HEAD(&memset_contexts, context, next);
>>>>>>> +
>>>>>>>          if (!use_madv_populate_write) {
>>>>>>> -        sigbus_memset_context = &context;
>>>>>>> +        sigbus_memset_context = true;
>>>>>>
>>>>> Thanks David,
>>>>>
>>>>>> Could we just use the sigbus handling alone and support parallel init
>>>>>> only when using MADV_POPULATE_WRITE where we don't have to mess with
>>>>>> signal handlers?
>>>>>>
>>>>>
>>>>> Ideally, we're hoping to support this with earlier kernels which don't
>>>>> support MADV_POPULATE _WRITE. But, I will check to see if we really
>>>>> need it.
>>>>
>>>> That's around since Linux 5.14, so please try simplifying.
>>>>
>>>>>
>>>>>> Further, how do you changes interact with other callers of
>>>>>> qemu_prealloc_mem(), like virtio-mem?
>>>>>>
>>>>>
>>>>> I'm not familiar with the intricacies of virtio-mem, but the basic
>>>>> idea
>>>>> of this series is to *only* allow parallel init during the start up
>>>>> phase (while prealloc_init == false). Once we have parsed all the
>>>>> command line args, we set prealloc_init = true
>>>>> (wait_mem_prealloc_init()) which causes all subsequent calls to
>>>>> qemu_prealloc_mem() to perform initialization synchronously. So, I
>>>>> *think* this covers the virtio-mem use case. Am I missing something?
>>>>
>>>> Good, so this should likely not affect virtio-mem (which also ends up
>>>> preallocating memory when loading from a vmstate).
>>>>
>>>> To make this all a bit clearer, what about the following to make this:
>>>>
>>>> * Optimize for MADV_POPULATE_WRITE. If we really need support for
>>>>     !MADV_POPULATE_WRITE, this is better added on top later.
>>>> * Pass in via a parameter that the caller requests async handling.
>>>> "bool
>>>>     async" should be good enough. Then, pass that only from the memory
>>>>     backend call, while QEMU is still initializing (we can find a way to
>>>>     make that work).
>>>> * Provide a function that waits for any remaining async os-mem-prealloc
>>>>     activity. That is essentially "wait_mem_prealloc_init", but without
>>>>     the special internal flag handling.
>>>>
>>>> Further, I do wonder if we want to make that behavior configurable. For
>>>> example, one might want to initialize backends sequentially using 16
>>>> threads
>>>> max, instead of consuming multiple times 16 threads concurrently.
>>>
>>> Seems to me that no matter what parallelisation we use (within
>>> mem regions, or across mem regions, or a mix of both), we should
>>> never use more threads than there are host CPUs.
>>
>> Yes. It gets tricky with multipe NUMA nodes, though. And that's what's
>> being requested here.
>>
>>>
>>> Can we have a pool of threads sized per available host CPUs that
>>> QEMU can access. Then for each memory backend fire off a set of
>>> init jobs that get distributed to "appropriate" threads in the
>>> pool. By appropriate I mean threads with the same NUMA affinity
>>> as the memory backend. This would give parallelisation both
>>> within a single large memory region, as well as across memory
>>> regions.
>>>
>>>
>>> eg 4 host CPUs, 3 memory regions (1GB for 1st numa node, 1GB
>>> for 2nd numa node, and 1 GB with no numa affinity). If we
>>> spread the init work then we end up with
>>
>> I'll note that having a mixture of numa + no numa is not a common
>> config we should try to optimize.
>>
>>>
>>>    1st thread gets 500MB to init from 1st memory region, and
>>>        250MB to init from 3rd memory region
>>>
>>>    2st thread gets 500MB to init from 1st memory region, and
>>>        250MB to init from 3rd memory region
>>>
>>>    3st thread gets 500MB to init from 2nd memory region, and
>>>        250MB to init from 3rd memory region
>>>
>>>    4th thread gets 500MB to init from 2nd memory region, and
>>>        250MB to init from 3rd memory region
>>
>> We do have prealloc contexts that are used to craft new CPUs with the
>> right NUMA affinity. Prealloc contexts are set for memory backends. So
>> extending prealloc context to limit/reuse threads and do the pooling
>> might be possible.
>>
> 
>   From a NUMA perspective, I feel prealloc contexts already supports what
> we want - i.e. we can set the number of prealloc threads and their node
> affinity per memory region. As such, unless you feel strongly about it,
> I'm inclined to leave additional enhancements (thread pool) as a future
> project.

I tend to agree. It's uncommon to have many memory backends target the 
same NUMA node such that you would want to use pooling.

Usually, you'd expect in most vNUMA setups to have, say, 1 memory 
backend per vNUMA node that is mapped to some NUMA nodes in the parent.

Using prealloc contexts you can make sure that preallocating each 
individual one gets the right CPU affinity. Setting the prealloc threads 
you can limit the number of threads used per memory backend.

>  > Since multiple NUMA node configs are the primary motivation for parallel
> init, I do wonder if it should *only* be used in conjunction
> with prealloc context. Perhaps we should have parallel init only enabled
> for memory regions with prealloc context?

That sounds reasonable.

-- 
Cheers,

David / dhildenb