Skip to contentSkip to conversation

Mini-OS update of events initialisation

PATCH In review Grzegorz Milos 2 participants 2 messages Nov 15, 2006 – Nov 16, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Mini-OS update of events initialisation

This patch moves initialisation of events (masking event channels) earlier during the boot process. Otherwise 64bit guests would sometimes crash.

Admittedly, I cannot explain the behavior I observed fully. This patch seems to fix it well, but I'd be grateful if someone on the list could explain it.

The problem only appears in 64 bit guest, and only if MiniOS keeps accessing the console ring page (which it does, as all printks write to the console, even before events are initialised. Of course without sending notifications).
What happens is that after a while, usually in memory initialisation stage, but at fairly random place MiniOS gets "restarted" i.e. start_kernel is called again. This leads to a crash as soon as MiniOS tries to remap shared info page. It looks like if Xen or xend decided that MiniOS domain should have masked the event channels before writing to the console page, and reboot it ...

Any ideas?

Thanks
Gregor

extras/mini-os/include/x86/os.h +2 -0
diff -r 8ed53f24279e -r 2ba83e480bd6 extras/mini-os/include/x86/os.h
@@ -18,6 +18,8 @@
 #ifndef __ASSEMBLY__
 #include <types.h>
 #include <hypervisor.h>
+
+#define USED    __attribute__ ((used))
 
 extern void do_exit(void);
 #define BUG do_exit
extras/mini-os/kernel.c +3 -3
diff -r 8ed53f24279e -r 2ba83e480bd6 extras/mini-os/kernel.c
@@ -116,6 +116,9 @@ void start_kernel(start_info_t *si)
     printk("  cmd_line:   %s\n",  
            si->cmd_line ? (const char *)si->cmd_line : "NULL");
 
+    /* Set up events. */
+    init_events();
+    
     arch_print_info();
 
     setup_xen_features();
@@ -123,9 +126,6 @@ void start_kernel(start_info_t *si)
     /* Init memory management. */
     init_mm();
 
-    /* Set up events. */
-    init_events();
-    
     /* Init time and timers. */
     init_time();
 
extras/mini-os/mm.c +2 -5
diff -r 8ed53f24279e -r 2ba83e480bd6 extras/mini-os/mm.c
@@ -148,7 +148,7 @@ static chunk_head_t  free_tail[FREELIST_
  * Prints allocation[0/1] for @nr_pages, starting at @start
  * address (virtual).
  */
-static void print_allocation(void *start, int nr_pages)
+USED static void print_allocation(void *start, int nr_pages)
 {
     unsigned long pfn_start = virt_to_pfn(start);
     int count;
@@ -163,7 +163,7 @@ static void print_allocation(void *start
  * Prints chunks (making them with letters) for @nr_pages starting
  * at @start (virtual).
  */
-static void print_chunks(void *start, int nr_pages)
+USED static void print_chunks(void *start, int nr_pages)
 {
     char chunks[1001], current='A';
     int order, count;
@@ -408,7 +408,6 @@ void new_pt_frame(unsigned long *pt_pfn,
          do_exit();
          break;
     }
-
     /* Update the entry */
 #if defined(__x86_64__)
     tab = pte_to_virt(tab[l4_table_offset(pt_page)]);
@@ -446,7 +445,6 @@ void new_pt_frame(unsigned long *pt_pfn,
        printk("ERROR: mmu_update failed\n");
        do_exit();
     }
-
     *pt_pfn += 1;
 }
 
@@ -581,7 +579,6 @@ void build_pagetable(unsigned long *star
         }
         start_address += PAGE_SIZE;
     }
-
     *start_pfn = pt_pfn;
 }
 

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#1 lore raw
Quoted 6 lines · Grzegorz Milos

What happens is that after a while, usually in memory initialisation stage, but at fairly random place MiniOS gets "restarted" i.e. start_kernel is called again. This leads to a crash as soon as MiniOS tries to remap shared info page. It looks like if Xen or xend decided that MiniOS domain should have masked the event channels before writing to the console page, and reboot it ...

If you enable event delivery before registering a callback handler then your domain will crash on first event delivery. But then you would be restarted in a fresh domain with a new shared_info page; wou wouldn't jump back to the entry code of the existing domain.

-- Keir