]> pilppa.com Git - linux-2.6-omap-h63xx.git/commitdiff
ftrace: convert single large buffer into single pages.
authorSteven Rostedt <srostedt@redhat.com>
Mon, 12 May 2008 19:20:43 +0000 (21:20 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Fri, 23 May 2008 18:38:51 +0000 (20:38 +0200)
Allocating large buffers for the tracer may fail easily.
This patch converts the buffer from a large ordered allocation
to single pages. It uses the struct page LRU field to link the
pages together.

Later patches may also implement dynamic increasing and decreasing
of the trace buffers.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
kernel/trace/trace.c
kernel/trace/trace.h

index 1b8eca7650d4c116bea1a1e984941287d4c1b311..d7ad030a4c49874ca874795be94d972426770726 100644 (file)
@@ -15,6 +15,7 @@
 #include <linux/kallsyms.h>
 #include <linux/seq_file.h>
 #include <linux/debugfs.h>
+#include <linux/pagemap.h>
 #include <linux/hardirq.h>
 #include <linux/linkage.h>
 #include <linux/uaccess.h>
@@ -49,7 +50,7 @@ static struct trace_array     max_tr;
 static DEFINE_PER_CPU(struct trace_array_cpu, max_data);
 
 static int                     tracer_enabled;
-static unsigned long           trace_nr_entries = 4096UL;
+static unsigned long           trace_nr_entries = 16384UL;
 
 static struct tracer           *trace_types __read_mostly;
 static struct tracer           *current_trace __read_mostly;
@@ -57,6 +58,8 @@ static int                    max_tracer_type_len;
 
 static DEFINE_MUTEX(trace_types_lock);
 
+#define ENTRIES_PER_PAGE (PAGE_SIZE / sizeof(struct trace_entry))
+
 static int __init set_nr_entries(char *str)
 {
        if (!str)
@@ -103,6 +106,7 @@ static const char *trace_options[] = {
 
 static unsigned trace_flags;
 
+static DEFINE_SPINLOCK(ftrace_max_lock);
 
 /*
  * Copy the new maximum trace into the separate maximum-trace
@@ -136,17 +140,23 @@ update_max_tr(struct trace_array *tr, struct task_struct *tsk, int cpu)
 {
        struct trace_array_cpu *data;
        void *save_trace;
+       struct list_head save_pages;
        int i;
 
+       WARN_ON_ONCE(!irqs_disabled());
+       spin_lock(&ftrace_max_lock);
        /* clear out all the previous traces */
        for_each_possible_cpu(i) {
                data = tr->data[i];
                save_trace = max_tr.data[i]->trace;
+               save_pages = max_tr.data[i]->trace_pages;
                memcpy(max_tr.data[i], data, sizeof(*data));
                data->trace = save_trace;
+               data->trace_pages = save_pages;
        }
 
        __update_max_tr(tr, tsk, cpu);
+       spin_unlock(&ftrace_max_lock);
 }
 
 /**
@@ -160,16 +170,22 @@ update_max_tr_single(struct trace_array *tr, struct task_struct *tsk, int cpu)
 {
        struct trace_array_cpu *data = tr->data[cpu];
        void *save_trace;
+       struct list_head save_pages;
        int i;
 
+       WARN_ON_ONCE(!irqs_disabled());
+       spin_lock(&ftrace_max_lock);
        for_each_possible_cpu(i)
                tracing_reset(max_tr.data[i]);
 
        save_trace = max_tr.data[cpu]->trace;
+       save_pages = max_tr.data[cpu]->trace_pages;
        memcpy(max_tr.data[cpu], data, sizeof(*data));
        data->trace = save_trace;
+       data->trace_pages = save_pages;
 
        __update_max_tr(tr, tsk, cpu);
+       spin_unlock(&ftrace_max_lock);
 }
 
 int register_tracer(struct tracer *type)
@@ -236,7 +252,8 @@ void unregister_tracer(struct tracer *type)
 void notrace tracing_reset(struct trace_array_cpu *data)
 {
        data->trace_idx = 0;
-       atomic_set(&data->underrun, 0);
+       data->trace_current = data->trace;
+       data->trace_current_idx = 0;
 }
 
 #ifdef CONFIG_FTRACE
@@ -367,21 +384,27 @@ tracing_get_trace_entry(struct trace_array *tr,
 {
        unsigned long idx, idx_next;
        struct trace_entry *entry;
+       struct page *page;
+       struct list_head *next;
 
-       idx = data->trace_idx;
+       data->trace_idx++;
+       idx = data->trace_current_idx;
        idx_next = idx + 1;
 
-       if (unlikely(idx_next >= tr->entries)) {
-               atomic_inc(&data->underrun);
+       entry = data->trace_current + idx * TRACE_ENTRY_SIZE;
+
+       if (unlikely(idx_next >= ENTRIES_PER_PAGE)) {
+               page = virt_to_page(data->trace_current);
+               if (unlikely(&page->lru == data->trace_pages.prev))
+                       next = data->trace_pages.next;
+               else
+                       next = page->lru.next;
+               page = list_entry(next, struct page, lru);
+               data->trace_current = page_address(page);
                idx_next = 0;
        }
 
-       data->trace_idx = idx_next;
-
-       if (unlikely(idx_next != 0 && atomic_read(&data->underrun)))
-               atomic_inc(&data->underrun);
-
-       entry = data->trace + idx * TRACE_ENTRY_SIZE;
+       data->trace_current_idx = idx_next;
 
        return entry;
 }
@@ -442,21 +465,38 @@ enum trace_file_type {
 };
 
 static struct trace_entry *
-trace_entry_idx(struct trace_array *tr, unsigned long idx, int cpu)
+trace_entry_idx(struct trace_array *tr, struct trace_array_cpu *data,
+               struct trace_iterator *iter, int cpu)
 {
-       struct trace_entry *array = tr->data[cpu]->trace;
-       unsigned long underrun;
+       struct page *page;
+       struct trace_entry *array;
 
-       if (idx >= tr->entries)
+       if (iter->next_idx[cpu] >= tr->entries ||
+           iter->next_idx[cpu] >= data->trace_idx)
                return NULL;
 
-       underrun = atomic_read(&tr->data[cpu]->underrun);
-       if (underrun)
-               idx = ((underrun - 1) + idx) % tr->entries;
-       else if (idx >= tr->data[cpu]->trace_idx)
-               return NULL;
+       if (!iter->next_page[cpu]) {
+               /*
+                * Initialize. If the count of elements in
+                * this buffer is greater than the max entries
+                * we had an underrun. Which means we looped around.
+                * We can simply use the current pointer as our
+                * starting point.
+                */
+               if (data->trace_idx >= tr->entries) {
+                       page = virt_to_page(data->trace_current);
+                       iter->next_page[cpu] = &page->lru;
+                       iter->next_page_idx[cpu] = data->trace_current_idx;
+               } else {
+                       iter->next_page[cpu] = data->trace_pages.next;
+                       iter->next_page_idx[cpu] = 0;
+               }
+       }
 
-       return &array[idx];
+       page = list_entry(iter->next_page[cpu], struct page, lru);
+       array = page_address(page);
+
+       return &array[iter->next_page_idx[cpu]];
 }
 
 static struct notrace trace_entry *
@@ -470,7 +510,7 @@ find_next_entry(struct trace_iterator *iter, int *ent_cpu)
        for_each_possible_cpu(cpu) {
                if (!tr->data[cpu]->trace)
                        continue;
-               ent = trace_entry_idx(tr, iter->next_idx[cpu], cpu);
+               ent = trace_entry_idx(tr, tr->data[cpu], iter, cpu);
                if (ent &&
                    (!next || (long)(next->idx - ent->idx) > 0)) {
                        next = ent;
@@ -492,8 +532,19 @@ static void *find_next_entry_inc(struct trace_iterator *iter)
        next = find_next_entry(iter, &next_cpu);
 
        if (next) {
-               iter->next_idx[next_cpu]++;
                iter->idx++;
+               iter->next_idx[next_cpu]++;
+               iter->next_page_idx[next_cpu]++;
+               if (iter->next_page_idx[next_cpu] >= ENTRIES_PER_PAGE) {
+                       struct trace_array_cpu *data = iter->tr->data[next_cpu];
+
+                       iter->next_page_idx[next_cpu] = 0;
+                       iter->next_page[next_cpu] =
+                               iter->next_page[next_cpu]->next;
+                       if (iter->next_page[next_cpu] == &data->trace_pages)
+                               iter->next_page[next_cpu] =
+                                       data->trace_pages.next;
+               }
        }
        iter->ent = next;
        iter->cpu = next_cpu;
@@ -554,14 +605,16 @@ static void *s_start(struct seq_file *m, loff_t *pos)
                iter->cpu = 0;
                iter->idx = -1;
 
-               for (i = 0; i < NR_CPUS; i++)
+               for_each_possible_cpu(i) {
                        iter->next_idx[i] = 0;
+                       iter->next_page[i] = NULL;
+               }
 
                for (p = iter; p && l < *pos; p = s_next(m, p, &l))
                        ;
 
        } else {
-               l = *pos;
+               l = *pos - 1;
                p = s_next(m, p, &l);
        }
 
@@ -654,9 +707,8 @@ print_trace_header(struct seq_file *m, struct trace_iterator *iter)
        struct trace_array *tr = iter->tr;
        struct trace_array_cpu *data = tr->data[tr->cpu];
        struct tracer *type = current_trace;
-       unsigned long underruns = 0;
-       unsigned long underrun;
-       unsigned long entries   = 0;
+       unsigned long total   = 0;
+       unsigned long entries = 0;
        int cpu;
        const char *name = "preemption";
 
@@ -665,11 +717,10 @@ print_trace_header(struct seq_file *m, struct trace_iterator *iter)
 
        for_each_possible_cpu(cpu) {
                if (tr->data[cpu]->trace) {
-                       underrun = atomic_read(&tr->data[cpu]->underrun);
-                       if (underrun) {
-                               underruns += underrun;
+                       total += tr->data[cpu]->trace_idx;
+                       if (tr->data[cpu]->trace_idx > tr->entries)
                                entries += tr->entries;
-                       else
+                       else
                                entries += tr->data[cpu]->trace_idx;
                }
        }
@@ -682,7 +733,7 @@ print_trace_header(struct seq_file *m, struct trace_iterator *iter)
                   " (M:%s VP:%d, KP:%d, SP:%d HP:%d",
                   data->saved_latency,
                   entries,
-                  (entries + underruns),
+                  total,
                   tr->cpu,
 #if defined(CONFIG_PREEMPT_NONE)
                   "server",
@@ -882,8 +933,7 @@ static int trace_empty(struct trace_iterator *iter)
                data = iter->tr->data[cpu];
 
                if (data->trace &&
-                   (data->trace_idx ||
-                    atomic_read(&data->underrun)))
+                   data->trace_idx)
                        return 0;
        }
        return 1;
@@ -1464,42 +1514,109 @@ static struct tracer no_tracer __read_mostly =
        .name = "none",
 };
 
-static inline notrace int page_order(const unsigned long size)
+static int trace_alloc_page(void)
 {
-       const unsigned long nr_pages = DIV_ROUND_UP(size, PAGE_SIZE);
-       return ilog2(roundup_pow_of_two(nr_pages));
+       struct trace_array_cpu *data;
+       void *array;
+       struct page *page, *tmp;
+       LIST_HEAD(pages);
+       int i;
+
+       /* first allocate a page for each CPU */
+       for_each_possible_cpu(i) {
+               array = (void *)__get_free_page(GFP_KERNEL);
+               if (array == NULL) {
+                       printk(KERN_ERR "tracer: failed to allocate page"
+                              "for trace buffer!\n");
+                       goto free_pages;
+               }
+
+               page = virt_to_page(array);
+               list_add(&page->lru, &pages);
+
+/* Only allocate if we are actually using the max trace */
+#ifdef CONFIG_TRACER_MAX_TRACE
+               array = (void *)__get_free_page(GFP_KERNEL);
+               if (array == NULL) {
+                       printk(KERN_ERR "tracer: failed to allocate page"
+                              "for trace buffer!\n");
+                       goto free_pages;
+               }
+               page = virt_to_page(array);
+               list_add(&page->lru, &pages);
+#endif
+       }
+
+       /* Now that we successfully allocate a page per CPU, add them */
+       for_each_possible_cpu(i) {
+               data = global_trace.data[i];
+               page = list_entry(pages.next, struct page, lru);
+               list_del(&page->lru);
+               list_add_tail(&page->lru, &data->trace_pages);
+               ClearPageLRU(page);
+
+#ifdef CONFIG_TRACER_MAX_TRACE
+               data = max_tr.data[i];
+               page = list_entry(pages.next, struct page, lru);
+               list_del(&page->lru);
+               list_add_tail(&page->lru, &data->trace_pages);
+               SetPageLRU(page);
+#endif
+       }
+       global_trace.entries += ENTRIES_PER_PAGE;
+
+       return 0;
+
+ free_pages:
+       list_for_each_entry_safe(page, tmp, &pages, lru) {
+               list_del(&page->lru);
+               __free_page(page);
+       }
+       return -ENOMEM;
 }
 
 __init static int tracer_alloc_buffers(void)
 {
-       const int order = page_order(trace_nr_entries * TRACE_ENTRY_SIZE);
-       const unsigned long size = (1UL << order) << PAGE_SHIFT;
-       struct trace_entry *array;
+       struct trace_array_cpu *data;
+       void *array;
+       struct page *page;
+       int pages = 0;
        int i;
 
+       /* Allocate the first page for all buffers */
        for_each_possible_cpu(i) {
-               global_trace.data[i] = &per_cpu(global_trace_cpu, i);
+               data = global_trace.data[i] = &per_cpu(global_trace_cpu, i);
                max_tr.data[i] = &per_cpu(max_data, i);
 
-               array = (struct trace_entry *)
-                         __get_free_pages(GFP_KERNEL, order);
+               array = (void *)__get_free_page(GFP_KERNEL);
                if (array == NULL) {
-                       printk(KERN_ERR "tracer: failed to allocate"
-                              " %ld bytes for trace buffer!\n", size);
+                       printk(KERN_ERR "tracer: failed to allocate page"
+                              "for trace buffer!\n");
                        goto free_buffers;
                }
-               global_trace.data[i]->trace = array;
+               data->trace = array;
+
+               /* set the array to the list */
+               INIT_LIST_HEAD(&data->trace_pages);
+               page = virt_to_page(array);
+               list_add(&page->lru, &data->trace_pages);
+               /* use the LRU flag to differentiate the two buffers */
+               ClearPageLRU(page);
 
 /* Only allocate if we are actually using the max trace */
 #ifdef CONFIG_TRACER_MAX_TRACE
-               array = (struct trace_entry *)
-                         __get_free_pages(GFP_KERNEL, order);
+               array = (void *)__get_free_page(GFP_KERNEL);
                if (array == NULL) {
-                       printk(KERN_ERR "wakeup tracer: failed to allocate"
-                              " %ld bytes for trace buffer!\n", size);
+                       printk(KERN_ERR "tracer: failed to allocate page"
+                              "for trace buffer!\n");
                        goto free_buffers;
                }
                max_tr.data[i]->trace = array;
+
+               INIT_LIST_HEAD(&max_tr.data[i]->trace_pages);
+               page = virt_to_page(array);
+               list_add(&page->lru, &max_tr.data[i]->trace_pages);
+               SetPageLRU(page);
 #endif
        }
 
@@ -1507,11 +1624,18 @@ __init static int tracer_alloc_buffers(void)
         * Since we allocate by orders of pages, we may be able to
         * round up a bit.
         */
-       global_trace.entries = size / TRACE_ENTRY_SIZE;
+       global_trace.entries = ENTRIES_PER_PAGE;
        max_tr.entries = global_trace.entries;
+       pages++;
+
+       while (global_trace.entries < trace_nr_entries) {
+               if (trace_alloc_page())
+                       break;
+               pages++;
+       }
 
-       pr_info("tracer: %ld bytes allocated for %ld",
-               size, trace_nr_entries);
+       pr_info("tracer: %d pages allocated for %ld",
+               pages, trace_nr_entries);
        pr_info(" entries of %ld bytes\n", (long)TRACE_ENTRY_SIZE);
        pr_info("   actual entries %ld\n", global_trace.entries);
 
@@ -1526,17 +1650,26 @@ __init static int tracer_alloc_buffers(void)
 
  free_buffers:
        for (i-- ; i >= 0; i--) {
+               struct page *page, *tmp;
                struct trace_array_cpu *data = global_trace.data[i];
 
                if (data && data->trace) {
-                       free_pages((unsigned long)data->trace, order);
+                       list_for_each_entry_safe(page, tmp,
+                                                &data->trace_pages, lru) {
+                               list_del(&page->lru);
+                               __free_page(page);
+                       }
                        data->trace = NULL;
                }
 
 #ifdef CONFIG_TRACER_MAX_TRACE
                data = max_tr.data[i];
                if (data && data->trace) {
-                       free_pages((unsigned long)data->trace, order);
+                       list_for_each_entry_safe(page, tmp,
+                                                &data->trace_pages, lru) {
+                               list_del(&page->lru);
+                               __free_page(page);
+                       }
                        data->trace = NULL;
                }
 #endif
index 3173a93561d4ef24428f81ca25c5a699bee8ae3f..83e257e38084d6c1371a238482f9d96b63c29bf8 100644 (file)
@@ -54,9 +54,11 @@ struct trace_entry {
  */
 struct trace_array_cpu {
        void                    *trace;
+       void                    *trace_current;
+       unsigned                trace_current_idx;
+       struct list_head        trace_pages;
        unsigned long           trace_idx;
        atomic_t                disabled;
-       atomic_t                underrun;
        unsigned long           saved_latency;
        unsigned long           critical_start;
        unsigned long           critical_end;
@@ -112,8 +114,10 @@ struct trace_iterator {
        unsigned long           iter_flags;
        loff_t                  pos;
        unsigned long           next_idx[NR_CPUS];
+       struct list_head        *next_page[NR_CPUS];
+       unsigned                next_page_idx[NR_CPUS];
+       long                    idx;
        int                     cpu;
-       int                     idx;
 };
 
 void notrace tracing_reset(struct trace_array_cpu *data);