Re: [PATCH 3/3] nfit: do an ARS rescan on hitting a latent media error

[Vishal Verma <vishal.l.verma@intel.com>]

On 07/19, Dan Williams wrote:
> On Mon, Jul 18, 2016 at 5:45 PM, Vishal Verma <vishal.l.verma@intel.com> wrote:
> > When a latent (unknown to 'badblocks') error is encountered, it will
> > trigger a machine check exception. On a system with machine check
> > recovery, this will only SIGBUS the process(es) which had the bad page
> > mapped (as opposed to a kernel panic on platforms without machine
> > check recovery features). In the former case, we want to trigger a full
> > rescan of that nvdimm bus. This will allow any additional, new errors
> > to be captured in the block devices' badblocks lists, and offending
> > operations on them can be trapped early, avoiding machine checks.
> >
> > This is done by registering a callback function with the
> > x86_mce_decoder_chain and calling the new ars_rescan functionality with
> > the address in the mce notificatiion.
> >
> > Cc: Dan Williams <dan.j.williams@intel.com>
> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > Cc: Tony Luck <tony.luck@intel.com>
> > Cc: <linux-acpi@vger.kernel.org>
> > Cc: <linux-nvdimm@lists.01.org>
> > Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
> > ---
> >  drivers/acpi/nfit.c | 102 ++++++++++++++++++++++++++++++++++++++++++++++++++++
> >  drivers/acpi/nfit.h |   1 +
> >  2 files changed, 103 insertions(+)
> >
> > diff --git a/drivers/acpi/nfit.c b/drivers/acpi/nfit.c
> > index def9505..0d2d7a3 100644
> > --- a/drivers/acpi/nfit.c
> > +++ b/drivers/acpi/nfit.c
> > @@ -12,6 +12,7 @@
> >   */
> >  #include <linux/list_sort.h>
> >  #include <linux/libnvdimm.h>
> > +#include <linux/notifier.h>
> >  #include <linux/module.h>
> >  #include <linux/mutex.h>
> >  #include <linux/ndctl.h>
> > @@ -23,6 +24,7 @@
> >  #include <linux/io.h>
> >  #include <linux/nd.h>
> >  #include <asm/cacheflush.h>
> > +#include <asm/mce.h>
> >  #include "nfit.h"
> >
> >  /*
> > @@ -50,6 +52,9 @@ module_param(disable_vendor_specific, bool, S_IRUGO);
> >  MODULE_PARM_DESC(disable_vendor_specific,
> >                 "Limit commands to the publicly specified set\n");
> >
> > +static LIST_HEAD(acpi_descs);
> > +static DEFINE_MUTEX(acpi_desc_lock);
> > +
> >  static struct workqueue_struct *nfit_wq;
> >
> >  struct nfit_table_prev {
> > @@ -2382,6 +2387,7 @@ static int acpi_nfit_check_deletions(struct acpi_nfit_desc *acpi_desc,
> >
> >  int acpi_nfit_init(struct acpi_nfit_desc *acpi_desc, acpi_size sz)
> >  {
> > +       struct acpi_nfit_desc *acpi_desc_entry;
> >         struct device *dev = acpi_desc->dev;
> >         struct nfit_table_prev prev;
> >         const void *end;
> > @@ -2439,6 +2445,25 @@ int acpi_nfit_init(struct acpi_nfit_desc *acpi_desc, acpi_size sz)
> >
> >         rc = acpi_nfit_register_regions(acpi_desc);
> >
> > +       /*
> > +        * We may get here due to an update of the nfit via _FIT.
> > +        * Check if the acpi_desc we're (re)initializing is already
> > +        * present in the list, and if so, don't re-add it
> > +        */
> > +       mutex_lock(&acpi_desc_lock);
> > +       if (list_empty(&acpi_descs))
> > +               list_add_tail(&acpi_desc->list, &acpi_descs);
> 
> No need to special case list_empty(), it's covered below and this
> isn't a fast path.
> 
> > +       else {
> > +               int found = 0;
> > +
> > +               list_for_each_entry(acpi_desc_entry, &acpi_descs, list)
> > +                       if (acpi_desc_entry == acpi_desc)
> > +                               found = 1;
> > +               if (found == 0)
> > +                       list_add_tail(&acpi_desc->list, &acpi_descs);
> > +       }
> > +       mutex_unlock(&acpi_desc_lock);
> > +
> >   out_unlock:
> >         mutex_unlock(&acpi_desc->init_mutex);
> >         return rc;
> > @@ -2522,6 +2547,69 @@ static int acpi_nfit_ars_rescan(struct nvdimm_bus_descriptor *nd_desc)
> >         return 0;
> >  }
> >
> > +static int nfit_handle_mce(struct notifier_block *nb, unsigned long val,
> > +                       void *data)
> > +{
> > +       struct mce *mce = (struct mce *)data;
> > +       struct acpi_nfit_desc *acpi_desc;
> > +       struct nfit_spa *nfit_spa;
> > +
> > +       /* We only care about memory errors */
> > +       if (!(mce->status & MCACOD))
> > +               return NOTIFY_DONE;
> > +
> > +       /*
> > +        * mce->addr contains the physical addr accessed that caused the
> > +        * machine check. We need to walk through the list of NFITs, and see
> > +        * if any of them matches that address, and only then start a scrub.
> > +        */
> > +       mutex_lock(&acpi_desc_lock);
> > +       if (list_empty(&acpi_descs))
> > +               goto out;
> 
> Again, no need to check for empty, list_for_each_entry() already does that...
> 
> > +
> > +       list_for_each_entry(acpi_desc, &acpi_descs, list) {
> > +               struct device *dev = acpi_desc->dev;
> > +               int found_match = 0;
> > +
> > +               list_for_each_entry(nfit_spa, &acpi_desc->spas, list) {
> > +                       struct acpi_nfit_system_address *spa = nfit_spa->spa;
> > +
> > +                       if (nfit_spa_type(spa) != NFIT_SPA_PM)
> > +                               continue;
> > +                       /* find the spa that covers the mce addr */
> > +                       if (spa->address > mce->addr)
> > +                               continue;
> > +                       if ((spa->address + spa->length - 1) < mce->addr)
> > +                               continue;
> > +                       found_match = 1;
> > +                       dev_dbg(dev, "%s: addr in SPA %d (0x%llx, 0x%llx)\n",
> > +                               __func__, spa->range_index, spa->address,
> > +                               spa->length);
> > +                       /*
> > +                        * We can break at the first match because we're going
> > +                        * to rescan all the SPA ranges. There shouldn't be any
> > +                        * aliasing anyway.
> > +                        */
> > +                       break;
> > +               }
> > +
> > +               /*
> > +                * We can ignore an -EBUSY here because if an ARS is already
> > +                * in progress, just let that be the last authoritative one
> > +                */
> > +               if (found_match)
> > +                       acpi_nfit_ars_rescan(&acpi_desc->nd_desc);
> > +       }
> > +
> > + out:
> > +       mutex_unlock(&acpi_desc_lock);
> > +       return NOTIFY_DONE;
> > +}
> > +
> > +static struct notifier_block nfit_mce_dec = {
> > +       .notifier_call  = nfit_handle_mce,
> > +};
> > +
> >  void acpi_nfit_desc_init(struct acpi_nfit_desc *acpi_desc, struct device *dev)
> >  {
> >         struct nvdimm_bus_descriptor *nd_desc;
> > @@ -2616,6 +2704,9 @@ static int acpi_nfit_remove(struct acpi_device *adev)
> >         acpi_desc->cancel = 1;
> >         flush_workqueue(nfit_wq);
> >         nvdimm_bus_unregister(acpi_desc->nvdimm_bus);
> > +       mutex_lock(&acpi_desc_lock);
> > +       list_del(&acpi_desc->list);
> > +       mutex_unlock(&acpi_desc_lock);
> >         return 0;
> >  }
> >
> > @@ -2725,13 +2816,24 @@ static __init int nfit_init(void)
> >         if (!nfit_wq)
> >                 return -ENOMEM;
> >
> > +       INIT_LIST_HEAD(&acpi_descs);
> > +       mce_register_decode_chain(&nfit_mce_dec);
> > +
> >         return acpi_bus_register_driver(&acpi_nfit_driver);
> >  }
> >
> >  static __exit void nfit_exit(void)
> >  {
> > +       struct acpi_nfit_desc *acpi_desc, *next;
> > +
> > +       mce_unregister_decode_chain(&nfit_mce_dec);
> >         acpi_bus_unregister_driver(&acpi_nfit_driver);
> >         destroy_workqueue(nfit_wq);
> > +       mutex_lock(&acpi_desc_lock);
> > +       if (list_empty(&acpi_descs))
> > +               list_for_each_entry_safe(acpi_desc, next, &acpi_descs, list)
> > +                       list_del(&acpi_desc->list);
> 
> We should WARN here, since there should be no way, outside of a bug,
> that 'acpi_descs' is still populated after
> acpi_bus_unregister_driver().

Agreed, also just spotted another bug - it should've been
if (!list_empty()) ...

> 
> > +       mutex_unlock(&acpi_desc_lock);
> >  }
> >
> >  module_init(nfit_init);
> > diff --git a/drivers/acpi/nfit.h b/drivers/acpi/nfit.h
> > index db95c5d..cf4d42d 100644
> > --- a/drivers/acpi/nfit.h
> > +++ b/drivers/acpi/nfit.h
> > @@ -147,6 +147,7 @@ struct acpi_nfit_desc {
> >         struct nd_cmd_ars_status *ars_status;
> >         size_t ars_status_size;
> >         struct work_struct work;
> > +       struct list_head list;
> >         unsigned int cancel:1;
> >         unsigned long dimm_cmd_force_en;
> >         unsigned long bus_cmd_force_en;
> 
> Outside of the minor comments above, this looks good to me.

Ok, I'll fix these up and resend. Thanks!