Re: [PATCH v5 5/7] pmem: add pmem_recovery_write() dax op

[Jane Chu <jane.chu@oracle.com>]

On 2/3/2022 10:21 PM, Dan Williams wrote:
> On Fri, Jan 28, 2022 at 1:32 PM Jane Chu <jane.chu@oracle.com> wrote:
>>
>> pmem_recovery_write() consists of clearing poison via DSM,
>> clearing page HWPoison bit, re-state _PAGE_PRESENT bit,
>> cacheflush, write, and finally clearing bad-block record.
>>
>> A competing pread thread is held off during recovery write
>> by the presence of bad-block record. A competing recovery_write
>> thread is serialized by a lock.
>>
>> Signed-off-by: Jane Chu <jane.chu@oracle.com>
>> ---
>>   drivers/nvdimm/pmem.c | 82 +++++++++++++++++++++++++++++++++++++++----
>>   drivers/nvdimm/pmem.h |  1 +
>>   2 files changed, 77 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
>> index 638e64681db9..f2d6b34d48de 100644
>> --- a/drivers/nvdimm/pmem.c
>> +++ b/drivers/nvdimm/pmem.c
>> @@ -69,6 +69,14 @@ static void hwpoison_clear(struct pmem_device *pmem,
>>          }
>>   }
>>
>> +static void pmem_clear_badblocks(struct pmem_device *pmem, sector_t sector,
>> +                               long cleared_blks)
>> +{
>> +       badblocks_clear(&pmem->bb, sector, cleared_blks);
>> +       if (pmem->bb_state)
>> +               sysfs_notify_dirent(pmem->bb_state);
>> +}
>> +
>>   static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
>>                  phys_addr_t offset, unsigned int len)
>>   {
>> @@ -88,9 +96,7 @@ static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
>>                  dev_dbg(dev, "%#llx clear %ld sector%s\n",
>>                                  (unsigned long long) sector, cleared,
>>                                  cleared > 1 ? "s" : "");
>> -               badblocks_clear(&pmem->bb, sector, cleared);
>> -               if (pmem->bb_state)
>> -                       sysfs_notify_dirent(pmem->bb_state);
>> +               pmem_clear_badblocks(pmem, sector, cleared);
>>          }
>>
>>          arch_invalidate_pmem(pmem->virt_addr + offset, len);
>> @@ -257,10 +263,15 @@ static int pmem_rw_page(struct block_device *bdev, sector_t sector,
>>   __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
>>                  long nr_pages, void **kaddr, pfn_t *pfn)
>>   {
>> +       bool bad_pmem;
>> +       bool do_recovery = false;
>>          resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
>>
>> -       if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
>> -                                       PFN_PHYS(nr_pages))))
>> +       bad_pmem = is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
>> +                               PFN_PHYS(nr_pages));
>> +       if (bad_pmem && kaddr)
>> +               do_recovery = dax_recovery_started(pmem->dax_dev, kaddr);
>> +       if (bad_pmem && !do_recovery)
>>                  return -EIO;
>>
>>          if (kaddr)
>> @@ -301,10 +312,68 @@ static long pmem_dax_direct_access(struct dax_device *dax_dev,
>>          return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
>>   }
>>
>> +/*
>> + * The recovery write thread started out as a normal pwrite thread and
>> + * when the filesystem was told about potential media error in the
>> + * range, filesystem turns the normal pwrite to a dax_recovery_write.
>> + *
>> + * The recovery write consists of clearing poison via DSM, clearing page
>> + * HWPoison bit, reenable page-wide read-write permission, flush the
>> + * caches and finally write.  A competing pread thread needs to be held
>> + * off during the recovery process since data read back might not be valid.
>> + * And that's achieved by placing the badblock records clearing after
>> + * the completion of the recovery write.
>> + *
>> + * Any competing recovery write thread needs to be serialized, and this is
>> + * done via pmem device level lock .recovery_lock.
>> + */
>>   static size_t pmem_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
>>                  void *addr, size_t bytes, struct iov_iter *i)
>>   {
>> -       return 0;
>> +       size_t rc, len, off;
>> +       phys_addr_t pmem_off;
>> +       struct pmem_device *pmem = dax_get_private(dax_dev);
>> +       struct device *dev = pmem->bb.dev;
>> +       sector_t sector;
>> +       long cleared, cleared_blk;
>> +
>> +       mutex_lock(&pmem->recovery_lock);
>> +
>> +       /* If no poison found in the range, go ahead with write */
>> +       off = (unsigned long)addr & ~PAGE_MASK;
>> +       len = PFN_PHYS(PFN_UP(off + bytes));
>> +       if (!is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512, len)) {
>> +               rc = _copy_from_iter_flushcache(addr, bytes, i);
>> +               goto write_done;
>> +       }
> 
> is_bad_pmem() takes a seqlock so it should be safe to move the
> recovery_lock below this point.

Okay, thanks!

> 
>> +
>> +       /* Not page-aligned range cannot be recovered */
>> +       if (off || !(PAGE_ALIGNED(bytes))) {
>> +               dev_warn(dev, "Found poison, but addr(%p) or bytes(%#lx) not page aligned\n",
>> +                       addr, bytes);
> 
> fs/dax.c will prevent this from happening, right? It seems like an
> upper layer bug if we get this far into the recovery process without
> checking if a full page is being overwritten.

Yes, at the start of each dax_iomap_iter, the buffer is page aligned. 
However, the underlying dax_copy_from_iter is allowed to return partial 
results, causing the subsequent 'while' loop within dax_iomap_iter to 
continue at not page aligned address. I ran into the situation when I 
played around dax_copy_from_iter, not sure in reality, partial result is 
legitimate, just thought to issue a warning should the situation happen.

> 
>> +               rc = (size_t) -EIO;
>> +               goto write_done;
>> +       }
>> +
>> +       pmem_off = PFN_PHYS(pgoff) + pmem->data_offset;
>> +       sector = (pmem_off - pmem->data_offset) / 512;
>> +       cleared = nvdimm_clear_poison(dev, pmem->phys_addr + pmem_off, len);
>> +       cleared_blk = cleared / 512;
>> +       if (cleared_blk > 0) {
>> +               hwpoison_clear(pmem, pmem->phys_addr + pmem_off, cleared);
>> +       } else {
>> +               dev_warn(dev, "pmem_recovery_write: cleared_blk: %ld\n",
>> +                       cleared_blk);
>> +               rc = (size_t) -EIO;
>> +               goto write_done;
>> +       }
>> +       arch_invalidate_pmem(pmem->virt_addr + pmem_off, bytes);
>> +       rc = _copy_from_iter_flushcache(addr, bytes, i);
>> +       pmem_clear_badblocks(pmem, sector, cleared_blk);
> 
> This duplicates pmem_clear_poison() can more code be shared between
> the 2 functions?

I'll look into refactoring pmem_clear_poison().
> 
> 
>> +
>> +write_done:
>> +       mutex_unlock(&pmem->recovery_lock);
>> +       return rc;
>>   }
>>
>>   static const struct dax_operations pmem_dax_ops = {
>> @@ -495,6 +564,7 @@ static int pmem_attach_disk(struct device *dev,
>>                  goto out_cleanup_dax;
>>          dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
>>          set_dax_recovery(dax_dev);
>> +       mutex_init(&pmem->recovery_lock);
>>          pmem->dax_dev = dax_dev;
>>
>>          rc = device_add_disk(dev, disk, pmem_attribute_groups);
>> diff --git a/drivers/nvdimm/pmem.h b/drivers/nvdimm/pmem.h
>> index 59cfe13ea8a8..971bff7552d6 100644
>> --- a/drivers/nvdimm/pmem.h
>> +++ b/drivers/nvdimm/pmem.h
>> @@ -24,6 +24,7 @@ struct pmem_device {
>>          struct dax_device       *dax_dev;
>>          struct gendisk          *disk;
>>          struct dev_pagemap      pgmap;
>> +       struct mutex            recovery_lock;
>>   };
>>
>>   long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
>> --
>> 2.18.4
>>

thanks!
-jane