Re: [PATCH v6 4/6] dax: add DAX_RECOVERY flag and .recovery_write dev_pgmap_ops

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

On 3/22/2022 2:01 AM, Christoph Hellwig wrote:
> On Sat, Mar 19, 2022 at 12:28:31AM -0600, Jane Chu wrote:
>> Introduce DAX_RECOVERY flag to dax_direct_access(). The flag is
>> not set by default in dax_direct_access() such that the helper
>> does not translate a pmem range to kernel virtual address if the
>> range contains uncorrectable errors.  When the flag is set,
>> the helper ignores the UEs and return kernel virtual adderss so
>> that the caller may get on with data recovery via write.
> 
> This DAX_RECOVERY doesn't actually seem to be used anywhere here or
> in the subsequent patches.  Did I miss something?

dax_iomap_iter() uses the flag in the same patch,
+               if ((map_len == -EIO) && (iov_iter_rw(iter) == WRITE)) {
+                       flags |= DAX_RECOVERY;
+                       map_len = dax_direct_access(dax_dev, pgoff, nrpg,
+                                               flags, &kaddr, NULL);


> 
>> Also introduce a new dev_pagemap_ops .recovery_write function.
>> The function is applicable to FSDAX device only. The device
>> page backend driver provides .recovery_write function if the
>> device has underlying mechanism to clear the uncorrectable
>> errors on the fly.
> 
> Why is this not in struct dax_operations?

Per Dan's comments to the v5 series, adding .recovery_write to
dax_operations causes a number of trivial dm targets changes.
Dan suggested that adding .recovery_write to pagemap_ops could
cut short the logistics of figuring out whether the driver backing
up a page is indeed capable of clearing poison. Please see
https://lkml.org/lkml/2022/2/4/31

> 
>>   
>> +size_t dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff,
>> +		void *addr, size_t bytes, struct iov_iter *iter)
>> +{
>> +	struct dev_pagemap *pgmap = dax_dev->pgmap;
>> +
>> +	if (!pgmap || !pgmap->ops->recovery_write)
>> +		return -EIO;
>> +	return pgmap->ops->recovery_write(pgmap, pgoff, addr, bytes,
>> +				(void *)iter);
> 
> No need to cast a type pointer to a void pointer.  But more importantly
> losing the type information here and passing it as void seems very
> wrong.

include/linux/memremap.h doesn't know struct iov_iter which is defined 
in include/linux/uio.h,  would you prefer to adding include/linux/uio.h 
to include/linux/memremap.h ?

> 
>> +static size_t pmem_recovery_write(struct dev_pagemap *pgmap, pgoff_t pgoff,
>> +		void *addr, size_t bytes, void *iter)
>> +{
>> +	struct pmem_device *pmem = pgmap->owner;
>> +
>> +	dev_warn(pmem->bb.dev, "%s: not yet implemented\n", __func__);
>> +
>> +	/* XXX more later */
>> +	return 0;
>> +}
> 
> This shuld not be added here - the core code can cope with a NULL
> method just fine.

Okay, will remove the XXX line.

> 
>> +		recov = 0;
>> +		flags = 0;
>> +		nrpg = PHYS_PFN(size);
> 
> Please spell out the words.  The recovery flag can also be
> a bool to make the code more readable.

Sure.

> 
>> +		map_len = dax_direct_access(dax_dev, pgoff, nrpg, flags,
>> +					&kaddr, NULL);
>> +		if ((map_len == -EIO) && (iov_iter_rw(iter) == WRITE)) {
> 
> No need for the inner braces.

Okay.

> 
>> +			flags |= DAX_RECOVERY;
>> +			map_len = dax_direct_access(dax_dev, pgoff, nrpg,
>> +						flags, &kaddr, NULL);
> 
> And noneed for the flags variable at all really.

Okay.
> 
>>   			xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
>>   					map_len, iter);
>>   		else
>> @@ -1271,6 +1286,11 @@ static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
>>   		length -= xfer;
>>   		done += xfer;
>>   
>> +		if (recov && (xfer == (ssize_t) -EIO)) {
>> +			pr_warn("dax_recovery_write failed\n");
>> +			ret = -EIO;
>> +			break;
> 
> And no, we can't just use an unsigned variable to communicate a
> negative error code.

Okay, will have dax_recovery_write return 0 in all error cases.

thanks!
-jane