src/patches/suse-2.6.27.31/patches.arch/ppc-pseries-migration_hang_fix.patch

   1 Subject: Fix partition migration hang under load
   2 From: Brian King <brking@linux.vnet.ibm.com>
   3 References: 470563 - LTC51153
   4
   5 While testing partition migration with heavy CPU load using
   6 shared processors, it was observed that sometimes the migration
   7 would never complete and would appear to hang. Currently, the
   8 migration code assumes that if H_SUCCESS is returned from the H_JOIN
   9 then the migration is complete and the processor is waking up on
  10 the target system. If there was an outstanding PROD to the processor
  11 when the H_JOIN is called, however, it will return H_SUCCESS on the source
  12 system, causing the migration to hang, or in some scenarios cause
  13 the kernel to crash on the complete call waking the caller
  14 of rtas_percpu_suspend_me.
  15
  16 Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
  17 Signed-off-by: Olaf Hering <olh@suse.de>
  18 ---
  19
  20  arch/powerpc/kernel/rtas.c |   10 ++++++++--
  21  1 file changed, 8 insertions(+), 2 deletions(-)
  22
  23 --- a/arch/powerpc/kernel/rtas.c
  24 +++ b/arch/powerpc/kernel/rtas.c
  25 @@ -46,6 +46,7 @@ EXPORT_SYMBOL(rtas);
  26
  27  struct rtas_suspend_me_data {
  28         atomic_t working; /* number of cpus accessing this struct */
  29 +       atomic_t done;
  30         int token; /* ibm,suspend-me */
  31         int error;
  32         struct completion *complete; /* wait on this until working == 0 */
  33 @@ -663,7 +664,7 @@ static int ibm_suspend_me_token = RTAS_U
  34  #ifdef CONFIG_PPC_PSERIES
  35  static void rtas_percpu_suspend_me(void *info)
  36  {
  37 -       long rc;
  38 +       long rc = H_SUCCESS;
  39         unsigned long msr_save;
  40         int cpu;
  41         struct rtas_suspend_me_data *data =
  42 @@ -675,7 +676,8 @@ static void rtas_percpu_suspend_me(void
  43         msr_save = mfmsr();
  44         mtmsr(msr_save & ~(MSR_EE));
  45
  46 -       rc = plpar_hcall_norets(H_JOIN);
  47 +       while (rc == H_SUCCESS && !atomic_read(&data->done))
  48 +               rc = plpar_hcall_norets(H_JOIN);
  49
  50         mtmsr(msr_save);
  51
  52 @@ -698,6 +700,9 @@ static void rtas_percpu_suspend_me(void
  53                        smp_processor_id(), rc);
  54                 data->error = rc;
  55         }
  56 +
  57 +       atomic_set(&data->done, 1);
  58 +
  59         /* This cpu did the suspend or got an error; in either case,
  60          * we need to prod all other other cpus out of join state.
  61          * Extra prods are harmless.
  62 @@ -740,6 +745,7 @@ static int rtas_ibm_suspend_me(struct rt
  63         }
  64
  65         atomic_set(&data.working, 0);
  66 +       atomic_set(&data.done, 0);
  67         data.token = rtas_token("ibm,suspend-me");
  68         data.error = 0;
  69         data.complete = &done;