Whamcloud - gitweb
Landing b_hd_newconfig on HEAD
[fs/lustre-release.git] / lnet / libcfs / tracefile.h
index f581257..f3568e9 100644 (file)
@@ -1,5 +1,17 @@
-#ifndef __PORTALS_TRACEFILE_H
-#define __PORTALS_TRACEFILE_H
+#ifndef __LIBCFS_TRACEFILE_H__
+#define __LIBCFS_TRACEFILE_H__
+
+#include <libcfs/libcfs.h>
+
+/* trace file lock routines */
+
+int  tracefile_init_arch(void);
+void tracefile_fini_arch(void);
+
+void tracefile_read_lock(void);
+void tracefile_read_unlock(void);
+void tracefile_write_lock(void);
+void tracefile_write_unlock(void);
 
 int tracefile_dump_all_pages(char *filename);
 void trace_debug_print(void);
@@ -19,4 +31,180 @@ int trace_read_debug_mb(char *page, char **start, off_t off, int count,
 int trace_dk(struct file *file, const char *buffer, unsigned long count,
              void *data);
 
-#endif /* __PORTALS_TRACEFILE_H */
+extern void libcfs_debug_dumplog_internal(void *arg);
+extern void libcfs_register_panic_notifier(void);
+extern void libcfs_unregister_panic_notifier(void);
+extern int  libcfs_panic_in_progress;
+
+#ifdef LUSTRE_TRACEFILE_PRIVATE
+/*
+ * Private declare for tracefile
+ */
+#define TCD_MAX_PAGES (5 << (20 - CFS_PAGE_SHIFT))
+#define TCD_STOCK_PAGES (TCD_MAX_PAGES)
+
+#define TRACEFILE_SIZE (500 << 20)
+
+/* Size of a buffer for sprinting console messages to in IRQ context (no
+ * logging in IRQ context) */
+#define TRACE_CONSOLE_BUFFER_SIZE   1024
+
+union trace_data_union {
+       struct trace_cpu_data {
+               /*
+                * pages with trace records not yet processed by tracefiled.
+                */
+               struct list_head        tcd_pages;
+               /* number of pages on ->tcd_pages */
+               unsigned long           tcd_cur_pages;
+
+               /*
+                * pages with trace records already processed by
+                * tracefiled. These pages are kept in memory, so that some
+                * portion of log can be written in the event of LBUG. This
+                * list is maintained in LRU order.
+                *
+                * Pages are moved to ->tcd_daemon_pages by tracefiled()
+                * (put_pages_on_daemon_list()). LRU pages from this list are
+                * discarded when list grows too large.
+                */
+               struct list_head        tcd_daemon_pages;
+               /* number of pages on ->tcd_daemon_pages */
+               unsigned long           tcd_cur_daemon_pages;
+
+               /*
+                * Maximal number of pages allowed on ->tcd_pages and
+                * ->tcd_daemon_pages each. Always TCD_MAX_PAGES in current
+                * implementation.
+                */
+               unsigned long           tcd_max_pages;
+
+               /*
+                * preallocated pages to write trace records into. Pages from
+                * ->tcd_stock_pages are moved to ->tcd_pages by
+                * portals_debug_msg().
+                *
+                * This list is necessary, because on some platforms it's
+                * impossible to perform efficient atomic page allocation in a
+                * non-blockable context.
+                *
+                * Such platforms fill ->tcd_stock_pages "on occasion", when
+                * tracing code is entered in blockable context.
+                *
+                * trace_get_tage_try() tries to get a page from
+                * ->tcd_stock_pages first and resorts to atomic page
+                * allocation only if this queue is empty. ->tcd_stock_pages
+                * is replenished when tracing code is entered in blocking
+                * context (darwin-tracefile.c:trace_get_tcd()). We try to
+                * maintain TCD_STOCK_PAGES (40 by default) pages in this
+                * queue. Atomic allocation is only required if more than
+                * TCD_STOCK_PAGES pagesful are consumed by trace records all
+                * emitted in non-blocking contexts. Which is quite unlikely.
+                */
+               struct list_head        tcd_stock_pages;
+               /* number of pages on ->tcd_stock_pages */
+               unsigned long           tcd_cur_stock_pages;
+
+               int                     tcd_shutting_down;
+               int                     tcd_cpu;
+       } tcd;
+       char __pad[SMP_CACHE_BYTES];
+};
+
+/* XXX nikita: this declaration is internal to tracefile.c and should probably
+ * be moved there */
+struct page_collection {
+       struct list_head        pc_pages;
+       /*
+        * spin-lock protecting ->pc_pages. It is taken by smp_call_function()
+        * call-back functions. XXX nikita: Which is horrible: all processors
+        * receive NMI at the same time only to be serialized by this
+        * lock. Probably ->pc_pages should be replaced with an array of
+        * NR_CPUS elements accessed locklessly.
+        */
+       spinlock_t              pc_lock;
+       /*
+        * if this flag is set, collect_pages() will spill both
+        * ->tcd_daemon_pages and ->tcd_pages to the ->pc_pages. Otherwise,
+        * only ->tcd_pages are spilled.
+        */
+       int                     pc_want_daemon_pages;
+};
+
+/* XXX nikita: this declaration is internal to tracefile.c and should probably
+ * be moved there */
+struct tracefiled_ctl {
+       struct completion       tctl_start;
+       struct completion       tctl_stop;
+       cfs_waitq_t             tctl_waitq;
+       pid_t                   tctl_pid;
+       atomic_t                tctl_shutdown;
+};
+
+/*
+ * small data-structure for each page owned by tracefiled.
+ */
+/* XXX nikita: this declaration is internal to tracefile.c and should probably
+ * be moved there */
+struct trace_page {
+       /*
+        * page itself
+        */
+       cfs_page_t      *page;
+       /*
+        * linkage into one of the lists in trace_data_union or
+        * page_collection
+        */
+       struct list_head linkage;
+       /*
+        * number of bytes used within this page
+        */
+       unsigned int     used;
+       /*
+        * cpu that owns this page
+        */
+       int              cpu;
+};
+
+extern void set_ptldebug_header(struct ptldebug_header *header,
+                          int subsys, int mask, const int line,
+                          unsigned long stack);
+extern void print_to_console(struct ptldebug_header *hdr, int mask, const char *buf,
+                            int len, const char *file, const char *fn);
+
+extern struct trace_cpu_data *trace_get_tcd(void);
+extern void trace_put_tcd(struct trace_cpu_data *tcd);
+extern char *trace_get_console_buffer(void);
+extern void trace_put_console_buffer(char *buffer);
+
+extern void trace_call_on_all_cpus(void (*fn)(void *arg), void *arg);
+
+int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
+                      struct list_head *stock);
+
+
+int tcd_owns_tage(struct trace_cpu_data *tcd, struct trace_page *tage);
+
+extern void trace_assertion_failed(const char *str, const char *fn,
+                                  const char *file, int line);
+
+/* ASSERTION that is safe to use within the debug system */
+#define __LASSERT(cond)                                                                \
+({                                                                             \
+       if (unlikely(!(cond))) {                                                \
+                trace_assertion_failed("ASSERTION("#cond") failed",            \
+                                      __FUNCTION__, __FILE__, __LINE__);       \
+       }                                                                       \
+})
+
+#define __LASSERT_TAGE_INVARIANT(tage)                 \
+({                                                     \
+        __LASSERT(tage != NULL);                       \
+        __LASSERT(tage->page != NULL);                 \
+        __LASSERT(tage->used <= CFS_PAGE_SIZE);                \
+        __LASSERT(cfs_page_count(tage->page) > 0);     \
+})
+
+#endif /* LUSTRE_TRACEFILE_PRIVATE */
+
+#endif /* __LIBCFS_TRACEFILE_H__ */