libnvdimm, pfn: during init, clear errors in the metadata area
Vishal Verma
maintainerPATCH 1/1 libnvdimm, pfn: during init, clear errors in the metadata area
If there are badblocks present in the 'struct page' area for pfn namespaces, until now, the only way to clear them has been to force the namespace into raw mode, clear the errors, and re-enable the fsdax mode. This is clunky, given that it should be easy enough for the pfn driver to do the same.
Add a new helper that uses the most recently available badblocks list to check whether there are any badblocks that lie in the volatile struct page area. If so, before initializing the struct pages, send down targeted writes via nvdimm_write_bytes to write zeroes to the affected blocks, and thus clear errors.
--- drivers/nvdimm/pfn_devs.c | 57 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 57 insertions(+)
drivers/nvdimm/pfn_devs.c +57 -0
@@ -361,8 +361,59 @@ struct device *nd_pfn_create(struct nd_region *nd_region)return dev; } +static int nd_pfn_clear_meta_errors (struct nd_pfn *nd_pfn) +{ +struct nd_region *nd_region =to_nd_region (nd_pfn->dev.parent); +struct nd_namespace_common *ndns = nd_pfn->ndns; +void *zero_page =page_address (ZERO_PAGE (0 )); +struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; +sector_t first_bad, meta_start; +struct nd_namespace_io *nsio; +int num_bad, meta_num, rc; +bool bb_present; + + nsio =to_nd_namespace_io (&ndns->dev); + meta_start = (SZ_4K +sizeof (*pfn_sb)) >>9 ; + meta_num = (le64_to_cpu (pfn_sb->dataoff) >>9 ) - meta_start; + +do { +unsigned long zero_len; +u64 nsoff; + + bb_present = !!badblocks_check (&nd_region->bb, meta_start, + meta_num, &first_bad, &num_bad); +if (bb_present) { +dev_dbg (&nd_pfn->dev,"meta: %x badblocks at %lx\n" , + num_bad, first_bad); + nsoff = (nd_region->ndr_start + (first_bad <<9 )) - + nsio->res.start; + zero_len = num_bad <<9 ; +while (zero_len) { +unsigned long chunk =min (zero_len,PAGE_SIZE ); + + rc =nvdimm_write_bytes (ndns, nsoff, zero_page, + chunk,0 ); +if (rc) +break ; + + zero_len -= chunk; + nsoff += chunk; + } +if (rc) { +dev_err (&nd_pfn->dev, +"error clearing %x badblocks at %lx\n" , + num_bad, first_bad); +return rc; + } + } + }while (bb_present); + +return 0 ; +} +int nd_pfn_validate (struct nd_pfn *nd_pfn,const char *sig) { +int rc;u64 checksum, offset;enum nd_pfn_mode mode;struct nd_namespace_io *nsio;
@@ -477,6 +528,12 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)return -ENXIO ; } +if (mode ==PFN_MODE_PMEM ) { + rc =nd_pfn_clear_meta_errors (nd_pfn); +if (rc) +return rc; + } +return 0 ; }EXPORT_SYMBOL (nd_pfn_validate);
Signature · 6 lines
2.14.4 _______________________________________________ Linux-nvdimm mailing list [email protected] https://lists.01.org/mailman/listinfo/linux-nvdimm
Quoted 24 lines · Vishal Verma
If there are badblocks present in the 'struct page' area for pfn namespaces, until now, the only way to clear them has been to force the namespace into raw mode, clear the errors, and re-enable the fsdax mode. This is clunky, given that it should be easy enough for the pfn driver to do the same.
Add a new helper that uses the most recently available badblocks list to check whether there are any badblocks that lie in the volatile struct page area. If so, before initializing the struct pages, send down targeted writes via
nvdimm_write_bytesto write zeroes to the affected blocks, and thus clear errors.Cc Dan Williams <[email protected]>Signed-off-by Vishal Verma <[email protected]>--- drivers/nvdimm/pfn_devs.c | 57 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 57 insertions(+)drivers/nvdimm/pfn_devs.c +0
-0
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 363
Quoted 3 lines
drivers/nvdimm/pfn_devs.c +0
-0@@ -361,8 +361,59 @@ struct device *nd_pfn_create(struct nd_region *nd_region)return dev; }
Perhaps a comment here about the fact that we are clearing the volatile memmap metadata space associated with a pfn instance.
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 364
Quoted 1 line
drivers/nvdimm/pfn_devs.c +1
-0+static int nd_pfn_clear_meta_errors (struct nd_pfn *nd_pfn)
Let's just say "memmap" rather than "meta" to be more explicit.
Quoted 10 lines
drivers/nvdimm/pfn_devs.c +10
-0+{ +struct nd_region *nd_region =to_nd_region (nd_pfn->dev.parent); +struct nd_namespace_common *ndns = nd_pfn->ndns; +void *zero_page =page_address (ZERO_PAGE (0 )); +struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; +sector_t first_bad, meta_start; +struct nd_namespace_io *nsio; +int num_bad, meta_num, rc; +bool bb_present; +
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 384
Quoted 10 lines
drivers/nvdimm/pfn_devs.c +10
-0+ nsio =to_nd_namespace_io (&ndns->dev); + meta_start = (SZ_4K +sizeof (*pfn_sb)) >>9 ; + meta_num = (le64_to_cpu (pfn_sb->dataoff) >>9 ) - meta_start; + + do { + unsigned long zero_len; + u64 nsoff; + + bb_present = !!badblocks_check(&nd_region->bb, meta_start, + meta_num, &first_bad, &num_bad);
The !! throws me off, I don't think it's necessary. You could just make bb_present an int.
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 389
Quoted 5 lines
drivers/nvdimm/pfn_devs.c +5
-0+if (bb_present) { +dev_dbg (&nd_pfn->dev,"meta: %x badblocks at %lx\n" , + num_bad, first_bad); + nsoff = (nd_region->ndr_start + (first_bad <<9 )) - + nsio->res.start;
Perhaps this should be ALIGN_DOWN((nd_region->ndr_start + (first_bad << 9)) - nsio->res.start, PAGE_SIZE)...
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 390
Quoted 1 line
drivers/nvdimm/pfn_devs.c +1
-0+ zero_len = num_bad <<9 ;
...and then make this ALIGN(num_bad << 9, PAGE_SIZE), or otherwise a comment about zero_len being zero...
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 392
Quoted 2 lines
drivers/nvdimm/pfn_devs.c +2
-0+while (zero_len) { +unsigned long chunk =min (zero_len,PAGE_SIZE );
...because the min(x, PAGE_SIZE) seems arbitrary.
Aligned clearing may be preferable considering we may be needing to fixup page protections.
Quoted 28 lines
drivers/nvdimm/pfn_devs.c +22
-0+ + rc =nvdimm_write_bytes (ndns, nsoff, zero_page, + chunk,0 ); +if (rc) +break ; + + zero_len -= chunk; + nsoff += chunk; + } +if (rc) { +dev_err (&nd_pfn->dev, +"error clearing %x badblocks at %lx\n" , + num_bad, first_bad); +return rc; + } + } + }while (bb_present); + +return 0 ; +} +int nd_pfn_validate (struct nd_pfn *nd_pfn,const char *sig) { +int rc;u64 checksum, offset;enum nd_pfn_mode mode;struct nd_namespace_io *nsio;
drivers/nvdimm/pfn_devs.c int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig) · line 534
Quoted 7 lines
drivers/nvdimm/pfn_devs.c +4
-0@@ -477,6 +528,12 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)return -ENXIO ; } +if (mode ==PFN_MODE_PMEM ) { + rc =nd_pfn_clear_meta_errors (nd_pfn); +if (rc) +return rc;
I think this can just be "return nd_pfn_clear_meta_errors()", and maybe move the mode check inside nd_pfn_clear_meta_errors() to return early if no volatile metadata to clear.
_______________________________________________
Linux-nvdimm mailing list
[email protected]
https://lists.01.org/mailman/listinfo/linux-nvdimm
Verma, Vishal L
maintainerQuoted 51 lines · Dan Williams
Quoted 24 lines · Vishal Verma
If there are badblocks present in the 'struct page' area for pfn namespaces, until now, the only way to clear them has been to force the namespace into raw mode, clear the errors, and re-enable the fsdax mode. This is clunky, given that it should be easy enough for the pfn driver to do the same.
Add a new helper that uses the most recently available badblocks list to check whether there are any badblocks that lie in the volatile struct page area. If so, before initializing the struct pages, send down targeted writes via
nvdimm_write_bytesto write zeroes to the affected blocks, and thus clear errors.Cc Dan Williams <[email protected]>Signed-off-by Vishal Verma <[email protected]>--- drivers/nvdimm/pfn_devs.c | 57 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 57 insertions(+)drivers/nvdimm/pfn_devs.c +0
-0drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 363
Quoted 3 lines
drivers/nvdimm/pfn_devs.c +0
-0@@ -361,8 +361,59 @@ struct device *nd_pfn_create(struct nd_region *nd_region)return dev; }Perhaps a comment here about the fact that we are clearing the volatile memmap metadata space associated with a pfn instance.
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 364
Quoted 1 line
drivers/nvdimm/pfn_devs.c +1
-0+static int nd_pfn_clear_meta_errors (struct nd_pfn *nd_pfn)Let's just say "memmap" rather than "meta" to be more explicit.
Quoted 15 lines
drivers/nvdimm/pfn_devs.c +15
-0+{ +struct nd_region *nd_region =to_nd_region (nd_pfn->dev.parent); +struct nd_namespace_common *ndns = nd_pfn->ndns; +void *zero_page =page_address (ZERO_PAGE (0 )); +struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb; +sector_t first_bad, meta_start; +struct nd_namespace_io *nsio; +int num_bad, meta_num, rc; +bool bb_present; + + nsio =to_nd_namespace_io (&ndns->dev); + meta_start = (SZ_4K +sizeof (*pfn_sb)) >>9 ; + meta_num = (le64_to_cpu (pfn_sb->dataoff) >>9 ) - meta_start; + + do {
Quoted 8 lines
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 384
Quoted 5 lines
drivers/nvdimm/pfn_devs.c +5
-0+unsigned long zero_len; +u64 nsoff; + + bb_present = !!badblocks_check (&nd_region->bb, meta_start, + meta_num, &first_bad, &num_bad);The !! throws me off, I don't think it's necessary. You could just make
bb_presentan int.
I was following the precedent of is_bad_pmem(), but int works too.
Quoted 15 lines
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 389
Quoted 6 lines
drivers/nvdimm/pfn_devs.c +5
-0+if (bb_present) { +dev_dbg (&nd_pfn->dev,"meta: %x badblocks at %lx\n" , + num_bad, first_bad); + nsoff = (nd_region->ndr_start + (first_bad <<9 )) - + nsio->res.start;Perhaps this should be
ALIGN_DOWN((nd_region->ndr_start + (first_bad << 9)) - nsio->res.start, PAGE_SIZE)...drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 390
Quoted 1 line
drivers/nvdimm/pfn_devs.c +1
-0+ zero_len = num_bad <<9 ;...and then make this
ALIGN(num_bad << 9, PAGE_SIZE), or otherwise a comment aboutzero_lenbeing zero...
Quoted 7 lines
drivers/nvdimm/pfn_devs.c struct device *nd_pfn_create(struct nd_region *nd_region) · line 392
Quoted 2 lines
drivers/nvdimm/pfn_devs.c +2
-0+while (zero_len) { +unsigned long chunk =min (zero_len,PAGE_SIZE );...because the
min(x, PAGE_SIZE)seems arbitrary.Aligned clearing may be preferable considering we may be needing to fixup page protections.
The min() is because our source of zeroes is PAGE_SIZE. zero_len *should* never be zero as a badblocks entry with a 'count' of zero would be a bug.. But aligned clearing makes sense too - we just potentially end up writing more than we absolutely need to.
What do you mean by fixup page protections?
I agree with the rest of the comment, and will fix those up.
Quoted 32 lines
Quoted 32 lines
drivers/nvdimm/pfn_devs.c +22
-0+ + rc =nvdimm_write_bytes (ndns, nsoff, zero_page, + chunk,0 ); +if (rc) +break ; + + zero_len -= chunk; + nsoff += chunk; + } +if (rc) { +dev_err (&nd_pfn->dev, +"error clearing %x badblocks at %lx\n" , + num_bad, first_bad); +return rc; + } + } + }while (bb_present); + +return 0 ; +} +int nd_pfn_validate (struct nd_pfn *nd_pfn,const char *sig) { +int rc;u64 checksum, offset;enum nd_pfn_mode mode;struct nd_namespace_io *nsio;@@ -477,6 +528,12 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)return -ENXIO ; }
Quoted 9 lines
drivers/nvdimm/pfn_devs.c int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig) · line 534
Quoted 4 lines
drivers/nvdimm/pfn_devs.c +4
-0+if (mode ==PFN_MODE_PMEM ) { + rc =nd_pfn_clear_meta_errors (nd_pfn); +if (rc) +return rc;I think this can just be "return
nd_pfn_clear_meta_errors()", and maybe move the mode check insidend_pfn_clear_meta_errors()to return early if no volatile metadata to clear.
_______________________________________________
Linux-nvdimm mailing list
[email protected]
https://lists.01.org/mailman/listinfo/linux-nvdimm