libnvdimm/nsio: differentiate between probe mapping and runtime mapping
Subsystems (4)
PATCH v3 1/1 libnvdimm/nsio: differentiate between probe mapping and runtime mapping
nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserver block area to check for pfn superblock type and map the full namespace resource only before using the namespace.
---
Changes from v2:
* update changelog
Changes from V1:
* update changelog
* update patch based on review feedback.
drivers/dax/pmem/core.c | 2 +- drivers/nvdimm/claim.c | 7 +++---- drivers/nvdimm/nd.h | 4 ++-- drivers/nvdimm/pfn.h | 6 ++++++ drivers/nvdimm/pfn_devs.c | 5 ----- drivers/nvdimm/pmem.c | 15 ++++++++++++--- 6 files changed, 24 insertions(+), 15 deletions(-)
drivers/dax/pmem/core.c +1 -1
@@ -28,7 +28,7 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_nsio_enable (dev, nsio,info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);
drivers/nvdimm/claim.c +3 -4
@@ -300,12 +300,12 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc; } -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common; - nsio->size =resource_size (res); + nsio->size = size;if (!devm_request_mem_region (dev, res->start,resource_size (res),dev_name (&ndns->dev))) {dev_warn (dev,"could not reserve region %pR\n" , res);
@@ -318,8 +318,7 @@ int devm_nsio_enable(struct device *dev, struct nd_namespace_io *nsio)nvdimm_badblocks_populate (to_nd_region (ndns->dev.parent), &nsio->bb, &nsio->res); - nsio->addr =devm_memremap (dev, res->start,resource_size (res), -ARCH_MEMREMAP_PMEM ); + nsio->addr =devm_memremap (dev, res->start, size,ARCH_MEMREMAP_PMEM );return PTR_ERR_OR_ZERO (nsio->addr); }
drivers/nvdimm/nd.h +2 -2
@@ -376,7 +376,7 @@ void nvdimm_badblocks_populate(struct nd_region *nd_region,#define MAX_STRUCT_PAGE_SIZE 64int nvdimm_setup_pfn (struct nd_pfn *nd_pfn,struct dev_pagemap *pgmap); -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio); +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size);void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio);#else static inline int nvdimm_setup_pfn(struct nd_pfn *nd_pfn,
@@ -385,7 +385,7 @@ static inline int nvdimm_setup_pfn(struct nd_pfn *nd_pfn,return -ENXIO ; }static inline int devm_nsio_enable (struct device *dev, -struct nd_namespace_io *nsio) +struct nd_namespace_io *nsio,unsigned long size) {return -ENXIO ; }
drivers/nvdimm/pfn.h +6 -0
@@ -36,4 +36,10 @@ struct nd_pfn_sb { __le64 checksum; }; +static inline u32 info_block_reserve (void ) +{ +return ALIGN (SZ_8K ,PAGE_SIZE ); +} + +#endif /* __NVDIMM_PFN_H */
drivers/nvdimm/pfn_devs.c +0 -5
@@ -635,11 +635,6 @@ int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns) }EXPORT_SYMBOL (nd_pfn_probe); -static u32 info_block_reserve (void ) -{ -return ALIGN (SZ_8K ,PAGE_SIZE ); -} -/* * We hotplug memory at sub-section granularity, pad the reserved area * from the previous section base to the namespace base address.
drivers/nvdimm/pmem.c +12 -3
@@ -491,17 +491,26 @@ static int pmem_attach_disk(struct device *dev,static int nd_pmem_probe (struct device *dev) {int ret; +struct nd_namespace_io *nsio;struct nd_namespace_common *ndns; ndns =nvdimm_namespace_common_probe (dev);if (IS_ERR (ndns))return PTR_ERR (ndns); -if (devm_nsio_enable (dev,to_nd_namespace_io (&ndns->dev))) -return -ENXIO ; + nsio =to_nd_namespace_io (&ndns->dev); -if (is_nd_btt (dev)) +if (is_nd_btt (dev)) { +/* +* Map with resource size +*/ +if (devm_nsio_enable (dev, nsio,resource_size (&nsio->res))) +return -ENXIO ;return nvdimm_namespace_attach_btt (ndns); + } + +if (devm_nsio_enable (dev, nsio,info_block_reserve ())) +return -ENXIO ;if (is_nd_pfn (dev))return pmem_attach_disk (dev, ndns);
Signature · 4 lines
2.21.0 _______________________________________________ Linux-nvdimm mailing list -- [email protected] To unsubscribe send an email to [email protected]
Quoted 11 lines · Aneesh Kumar K.V
nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserver block area to
s/reserver/reserved/
Quoted 2 lines
check for pfn superblock type and map the full namespace resource only before using the namespace.
Are you going to follow up with the BTT patch that uses this new facility?
Quoted 37 lines
Signed-off-by Aneesh Kumar K.V <[email protected]>---
Changes from v2:
* update changelogChanges from V1:
* update changelog
* update patch based on review feedback.drivers/dax/pmem/core.c | 2 +- drivers/nvdimm/claim.c | 7 +++---- drivers/nvdimm/nd.h | 4 ++-- drivers/nvdimm/pfn.h | 6 ++++++ drivers/nvdimm/pfn_devs.c | 5 ----- drivers/nvdimm/pmem.c | 15 ++++++++++++--- 6 files changed, 24 insertions(+), 15 deletions(-)drivers/dax/pmem/core.c +1
-1@@ -28,7 +28,7 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_nsio_enable (dev, nsio,info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);drivers/nvdimm/claim.c +0
-0@@ -300,12 +300,12 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc;
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 308
Quoted 10 lines
drivers/nvdimm/claim.c +2
-2} -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common; - nsio->size =resource_size (res); + nsio->size = size;
This needs a:
if (nsio->size)
devm_memunmap(dev, nsio->addr);
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 310
Quoted 2 lines
drivers/nvdimm/claim.c +0
-0if (!devm_request_mem_region (dev, res->start,resource_size (res),dev_name (&ndns->dev))) {
Also don't repeat the devm_request_mem_region() if one was already done.
Another thing to check is if nsio->size gets cleared when the namespace is disabled, if not that well need to be added in the shutdown path.
Quoted 99 lines
drivers/nvdimm/claim.c +1
-2dev_warn (dev,"could not reserve region %pR\n" , res);@@ -318,8 +318,7 @@ int devm_nsio_enable(struct device *dev, struct nd_namespace_io *nsio)nvdimm_badblocks_populate (to_nd_region (ndns->dev.parent), &nsio->bb, &nsio->res); - nsio->addr =devm_memremap (dev, res->start,resource_size (res), -ARCH_MEMREMAP_PMEM ); + nsio->addr =devm_memremap (dev, res->start, size,ARCH_MEMREMAP_PMEM );return PTR_ERR_OR_ZERO (nsio->addr); }drivers/nvdimm/nd.h +2
-2@@ -376,7 +376,7 @@ void nvdimm_badblocks_populate(struct nd_region *nd_region,#define MAX_STRUCT_PAGE_SIZE 64int nvdimm_setup_pfn (struct nd_pfn *nd_pfn,struct dev_pagemap *pgmap); -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio); +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size);void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio);#else static inline int nvdimm_setup_pfn(struct nd_pfn *nd_pfn,@@ -385,7 +385,7 @@ static inline int nvdimm_setup_pfn(struct nd_pfn *nd_pfn,return -ENXIO ; }static inline int devm_nsio_enable (struct device *dev, -struct nd_namespace_io *nsio) +struct nd_namespace_io *nsio,unsigned long size) {return -ENXIO ; }drivers/nvdimm/pfn.h +6
-0@@ -36,4 +36,10 @@ struct nd_pfn_sb { __le64 checksum; }; +static inline u32 info_block_reserve (void ) +{ +return ALIGN (SZ_8K ,PAGE_SIZE ); +} + +#endif /* __NVDIMM_PFN_H */ drivers/nvdimm/pfn_devs.c +0
-5@@ -635,11 +635,6 @@ int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns) }EXPORT_SYMBOL (nd_pfn_probe); -static u32 info_block_reserve (void ) -{ -return ALIGN (SZ_8K ,PAGE_SIZE ); -} -/* * We hotplug memory at sub-section granularity, pad the reserved area * from the previous section base to the namespace base address. drivers/nvdimm/pmem.c +12
-4@@ -491,17 +491,26 @@ static int pmem_attach_disk(struct device *dev,static int nd_pmem_probe (struct device *dev) {int ret; +struct nd_namespace_io *nsio;struct nd_namespace_common *ndns; ndns =nvdimm_namespace_common_probe (dev);if (IS_ERR (ndns))return PTR_ERR (ndns); -if (devm_nsio_enable (dev,to_nd_namespace_io (&ndns->dev))) -return -ENXIO ; + nsio =to_nd_namespace_io (&ndns->dev); -if (is_nd_btt (dev)) +if (is_nd_btt (dev)) { +/* +* Map with resource size +*/ +if (devm_nsio_enable (dev, nsio,resource_size (&nsio->res))) +return -ENXIO ;return nvdimm_namespace_attach_btt (ndns); + } + +if (devm_nsio_enable (dev, nsio,info_block_reserve ())) +return -ENXIO ;if (is_nd_pfn (dev))return pmem_attach_disk (dev, ndns); --
Quoted 2 lines
2.21.0
_______________________________________________
Linux-nvdimm mailing list -- [email protected] To unsubscribe send an email to [email protected]
Quoted 7 lines · Dan Williams
Quoted 5 lines · Aneesh Kumar K.V
nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
Quoted 8 lines
Quoted 6 lines
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserver block area to
s/reserver/reserved/
Will fix
Quoted 6 lines
Quoted 3 lines
check for pfn superblock type and map the full namespace resource only before using the namespace.
Are you going to follow up with the BTT patch that uses this new facility?
I am not yet sure about that. ioremap/vmap area is the way we get a kernel mapping without struct page backing. What we are suggesting hee is the ability to have a direct mapped mapping without struct page. I need to look at closely about what that imply.
Quoted 45 lines
Quoted 45 lines
Signed-off-by Aneesh Kumar K.V <[email protected]>---
Changes from v2:
* update changelogChanges from V1:
* update changelog
* update patch based on review feedback.drivers/dax/pmem/core.c | 2 +- drivers/nvdimm/claim.c | 7 +++---- drivers/nvdimm/nd.h | 4 ++-- drivers/nvdimm/pfn.h | 6 ++++++ drivers/nvdimm/pfn_devs.c | 5 ----- drivers/nvdimm/pmem.c | 15 ++++++++++++--- 6 files changed, 24 insertions(+), 15 deletions(-)drivers/dax/pmem/core.c +1
-1@@ -28,7 +28,7 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_nsio_enable (dev, nsio,info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);drivers/nvdimm/claim.c +1
-1@@ -300,12 +300,12 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc; } -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common;
Quoted 7 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 308
Quoted 2 lines
drivers/nvdimm/claim.c +1
-1- nsio->size =resource_size (res); + nsio->size = size;This needs a:
if (
nsio->size)devm_memunmap(dev, nsio->addr);
Why ? we should not be calling devm_nsio_enable twice now.
Quoted 11 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 310
Quoted 4 lines
drivers/nvdimm/claim.c +0
-0if (!devm_request_mem_region (dev, res->start,resource_size (res),dev_name (&ndns->dev))) {Also don't repeat the
devm_request_mem_region()if one was already done.Another thing to check is if
nsio->sizegets cleared when the namespace is disabled, if not that well need to be added in the shutdown path.
Not sure I understand that. So with this patch when we probe we always use info_block_reserve() size irrespective of the device. This probe will result in us adding a btt/pfn/dax device and we return -ENXIO after this probe. This return value will cause us to destroy the I/O remap mmapping we did with info_block_reserve() size. Also the nsio data structure is also freed.
nd_pmem_probe is then again called with btt device type and in that case we do devm_memremap again.
if (btt)
remap the full namespace size.
else
remap the info_block_reserve_size.
This infor block reserve mapping is unmapped in pmem_attach_disk()
Anything i am missing in the above flow?
Quoted 97 lines
Quoted 97 lines
drivers/nvdimm/claim.c +1
-2dev_warn (dev,"could not reserve region %pR\n" , res);@@ -318,8 +318,7 @@ int devm_nsio_enable(struct device *dev, struct nd_namespace_io *nsio)nvdimm_badblocks_populate (to_nd_region (ndns->dev.parent), &nsio->bb, &nsio->res); - nsio->addr =devm_memremap (dev, res->start,resource_size (res), -ARCH_MEMREMAP_PMEM ); + nsio->addr =devm_memremap (dev, res->start, size,ARCH_MEMREMAP_PMEM );return PTR_ERR_OR_ZERO (nsio->addr); }drivers/nvdimm/nd.h +2
-2@@ -376,7 +376,7 @@ void nvdimm_badblocks_populate(struct nd_region *nd_region,#define MAX_STRUCT_PAGE_SIZE 64int nvdimm_setup_pfn (struct nd_pfn *nd_pfn,struct dev_pagemap *pgmap); -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio); +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size);void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio);#else static inline int nvdimm_setup_pfn(struct nd_pfn *nd_pfn,@@ -385,7 +385,7 @@ static inline int nvdimm_setup_pfn(struct nd_pfn *nd_pfn,return -ENXIO ; }static inline int devm_nsio_enable (struct device *dev, -struct nd_namespace_io *nsio) +struct nd_namespace_io *nsio,unsigned long size) {return -ENXIO ; }drivers/nvdimm/pfn.h +6
-0@@ -36,4 +36,10 @@ struct nd_pfn_sb { __le64 checksum; }; +static inline u32 info_block_reserve (void ) +{ +return ALIGN (SZ_8K ,PAGE_SIZE ); +} + +#endif /* __NVDIMM_PFN_H */ drivers/nvdimm/pfn_devs.c +0
-5@@ -635,11 +635,6 @@ int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns) }EXPORT_SYMBOL (nd_pfn_probe); -static u32 info_block_reserve (void ) -{ -return ALIGN (SZ_8K ,PAGE_SIZE ); -} -/* * We hotplug memory at sub-section granularity, pad the reserved area * from the previous section base to the namespace base address. drivers/nvdimm/pmem.c +12
-3@@ -491,17 +491,26 @@ static int pmem_attach_disk(struct device *dev,static int nd_pmem_probe (struct device *dev) {int ret; +struct nd_namespace_io *nsio;struct nd_namespace_common *ndns; ndns =nvdimm_namespace_common_probe (dev);if (IS_ERR (ndns))return PTR_ERR (ndns); -if (devm_nsio_enable (dev,to_nd_namespace_io (&ndns->dev))) -return -ENXIO ; + nsio =to_nd_namespace_io (&ndns->dev); -if (is_nd_btt (dev)) +if (is_nd_btt (dev)) { +/* +* Map with resource size +*/ +if (devm_nsio_enable (dev, nsio,resource_size (&nsio->res))) +return -ENXIO ;return nvdimm_namespace_attach_btt (ndns); + } + +if (devm_nsio_enable (dev, nsio,info_block_reserve ())) +return -ENXIO ;
Quoted 5 lines
Quoted 5 lines
drivers/nvdimm/pmem.c +0
-1if (is_nd_pfn (dev))return pmem_attach_disk (dev, ndns); --2.21.0
_______________________________________________
Linux-nvdimm mailing list -- [email protected] To unsubscribe send an email to [email protected]
Quoted 118 lines · Aneesh Kumar K.V
Quoted 7 lines · Dan Williams
Quoted 5 lines · Aneesh Kumar K.V
nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
Quoted 8 lines
Quoted 6 lines
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserver block area to
s/reserver/reserved/
Will fix
Quoted 6 lines
Quoted 3 lines
check for pfn superblock type and map the full namespace resource only before using the namespace.
Are you going to follow up with the BTT patch that uses this new facility?
I am not yet sure about that. ioremap/vmap area is the way we get a kernel mapping without struct page backing. What we are suggesting hee is the ability to have a direct mapped mapping without struct page. I need to look at closely about what that imply.
Quoted 45 lines
Quoted 45 lines
Signed-off-by Aneesh Kumar K.V <[email protected]>---
Changes from v2:
* update changelogChanges from V1:
* update changelog
* update patch based on review feedback.drivers/dax/pmem/core.c | 2 +- drivers/nvdimm/claim.c | 7 +++---- drivers/nvdimm/nd.h | 4 ++-- drivers/nvdimm/pfn.h | 6 ++++++ drivers/nvdimm/pfn_devs.c | 5 ----- drivers/nvdimm/pmem.c | 15 ++++++++++++--- 6 files changed, 24 insertions(+), 15 deletions(-)drivers/dax/pmem/core.c +1
-1@@ -28,7 +28,7 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_nsio_enable (dev, nsio,info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);drivers/nvdimm/claim.c +1
-1@@ -300,12 +300,12 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc; } -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common;Quoted 7 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 308
Quoted 2 lines
drivers/nvdimm/claim.c +1
-1- nsio->size =resource_size (res); + nsio->size = size;This needs a:
if (
nsio->size)devm_memunmap(dev, nsio->addr);Why ? we should not be calling
devm_nsio_enabletwice now.Quoted 11 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 310
Quoted 4 lines
drivers/nvdimm/claim.c +0
-0if (!devm_request_mem_region (dev, res->start,resource_size (res),dev_name (&ndns->dev))) {Also don't repeat the
devm_request_mem_region()if one was already done.Another thing to check is if
nsio->sizegets cleared when the namespace is disabled, if not that well need to be added in the shutdown path.Not sure I understand that. So with this patch when we probe we always use
info_block_reserve()size irrespective of the device. This probe will result in us adding a btt/pfn/dax device and we return-ENXIOafter this probe. This return value will cause us to destroy the I/O remap mmapping we did withinfo_block_reserve()size. Also the nsio data structure is also freed.
nd_pmem_probeis then again called with btt device type and in that case we dodevm_memremapagain.if (btt)
remap the full namespace size.
else
remap theinfo_block_reserve_size.
Quoted 4 lines
This infor block reserve mapping is unmapped in
pmem_attach_disk()Anything i am missing in the above flow?
Oh no, you're right, this does make the implementation only call devm_nsio_enable() once. However, now that I look I want to suggest some small reorganizations of where devm_nsio_enable() is called. I'll take a stab at folding some changes on top of your patch.
_______________________________________________
Linux-nvdimm mailing list -- [email protected]
To unsubscribe send an email to [email protected]
Quoted 124 lines · Dan Williams
Quoted 121 lines · Aneesh Kumar K.V
Quoted 7 lines · Dan Williams
Quoted 5 lines · Aneesh Kumar K.V
nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
Quoted 8 lines
Quoted 6 lines
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserver block area to
s/reserver/reserved/
Will fix
Quoted 6 lines
Quoted 3 lines
check for pfn superblock type and map the full namespace resource only before using the namespace.
Are you going to follow up with the BTT patch that uses this new facility?
I am not yet sure about that. ioremap/vmap area is the way we get a kernel mapping without struct page backing. What we are suggesting hee is the ability to have a direct mapped mapping without struct page. I need to look at closely about what that imply.
Quoted 45 lines
Quoted 45 lines
Signed-off-by Aneesh Kumar K.V <[email protected]>---
Changes from v2:
* update changelogChanges from V1:
* update changelog
* update patch based on review feedback.drivers/dax/pmem/core.c | 2 +- drivers/nvdimm/claim.c | 7 +++---- drivers/nvdimm/nd.h | 4 ++-- drivers/nvdimm/pfn.h | 6 ++++++ drivers/nvdimm/pfn_devs.c | 5 ----- drivers/nvdimm/pmem.c | 15 ++++++++++++--- 6 files changed, 24 insertions(+), 15 deletions(-)drivers/dax/pmem/core.c +1
-1@@ -28,7 +28,7 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_nsio_enable (dev, nsio,info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);drivers/nvdimm/claim.c +1
-1@@ -300,12 +300,12 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc; } -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common;Quoted 7 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 308
Quoted 2 lines
drivers/nvdimm/claim.c +1
-1- nsio->size =resource_size (res); + nsio->size = size;This needs a:
if (
nsio->size)devm_memunmap(dev, nsio->addr);Why ? we should not be calling
devm_nsio_enabletwice now.Quoted 11 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 310
Quoted 4 lines
drivers/nvdimm/claim.c +0
-0if (!devm_request_mem_region (dev, res->start,resource_size (res),dev_name (&ndns->dev))) {Also don't repeat the
devm_request_mem_region()if one was already done.Another thing to check is if
nsio->sizegets cleared when the namespace is disabled, if not that well need to be added in the shutdown path.Not sure I understand that. So with this patch when we probe we always use
info_block_reserve()size irrespective of the device. This probe will result in us adding a btt/pfn/dax device and we return-ENXIOafter this probe. This return value will cause us to destroy the I/O remap mmapping we did withinfo_block_reserve()size. Also the nsio data structure is also freed.
nd_pmem_probeis then again called with btt device type and in that case we dodevm_memremapagain.if (btt)
remap the full namespace size.
else
remap theinfo_block_reserve_size.This infor block reserve mapping is unmapped in
pmem_attach_disk()
Quoted 6 lines
Quoted 1 line
Anything i am missing in the above flow?
Oh no, you're right, this does make the implementation only call
devm_nsio_enable()once. However, now that I look I want to suggest some small reorganizations of wheredevm_nsio_enable()is called. I'll take a stab at folding some changes on top of your patch.
This change is causing the "pfn-meta-errors.sh" test to fail. It is due to the fact the info_block_reserve() is not a sufficient reservation for errors in the page metadata area.
_______________________________________________
Linux-nvdimm mailing list -- [email protected] To unsubscribe send an email to [email protected]
Quoted 127 lines · Dan Williams
Quoted 126 lines · Dan Williams
Quoted 122 lines · Aneesh Kumar K.V
Quoted 7 lines · Dan Williams
Quoted 5 lines · Aneesh Kumar K.V
nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
Quoted 8 lines
Quoted 6 lines
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserver block area to
s/reserver/reserved/
Will fix
Quoted 6 lines
Quoted 3 lines
check for pfn superblock type and map the full namespace resource only before using the namespace.
Are you going to follow up with the BTT patch that uses this new facility?
I am not yet sure about that. ioremap/vmap area is the way we get a kernel mapping without struct page backing. What we are suggesting hee is the ability to have a direct mapped mapping without struct page. I need to look at closely about what that imply.
Quoted 45 lines
Quoted 45 lines
Signed-off-by Aneesh Kumar K.V <[email protected]>---
Changes from v2:
* update changelogChanges from V1:
* update changelog
* update patch based on review feedback.drivers/dax/pmem/core.c | 2 +- drivers/nvdimm/claim.c | 7 +++---- drivers/nvdimm/nd.h | 4 ++-- drivers/nvdimm/pfn.h | 6 ++++++ drivers/nvdimm/pfn_devs.c | 5 ----- drivers/nvdimm/pmem.c | 15 ++++++++++++--- 6 files changed, 24 insertions(+), 15 deletions(-)drivers/dax/pmem/core.c +1
-1@@ -28,7 +28,7 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_nsio_enable (dev, nsio,info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);drivers/nvdimm/claim.c +1
-1@@ -300,12 +300,12 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc; } -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio,unsigned long size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common;Quoted 7 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 308
Quoted 2 lines
drivers/nvdimm/claim.c +1
-1- nsio->size =resource_size (res); + nsio->size = size;This needs a:
if (
nsio->size)devm_memunmap(dev, nsio->addr);Why ? we should not be calling
devm_nsio_enabletwice now.Quoted 11 lines
drivers/nvdimm/claim.c static int nsio_rw_bytes(struct nd_namespace_common *ndns, · line 310
Quoted 4 lines
drivers/nvdimm/claim.c +0
-0if (!devm_request_mem_region (dev, res->start,resource_size (res),dev_name (&ndns->dev))) {Also don't repeat the
devm_request_mem_region()if one was already done.Another thing to check is if
nsio->sizegets cleared when the namespace is disabled, if not that well need to be added in the shutdown path.Not sure I understand that. So with this patch when we probe we always use
info_block_reserve()size irrespective of the device. This probe will result in us adding a btt/pfn/dax device and we return-ENXIOafter this probe. This return value will cause us to destroy the I/O remap mmapping we did withinfo_block_reserve()size. Also the nsio data structure is also freed.
nd_pmem_probeis then again called with btt device type and in that case we dodevm_memremapagain.if (btt)
remap the full namespace size.
else
remap theinfo_block_reserve_size.This infor block reserve mapping is unmapped in
pmem_attach_disk()Anything i am missing in the above flow?
Quoted 9 lines
Quoted 4 lines
Oh no, you're right, this does make the implementation only call
devm_nsio_enable()once. However, now that I look I want to suggest some small reorganizations of wheredevm_nsio_enable()is called. I'll take a stab at folding some changes on top of your patch.This change is causing the "
pfn-meta-errors.sh" test to fail. It is due to the fact theinfo_block_reserve()is not a sufficient reservation for errors in the page metadata area.
Can you share the call stack? we clear bad blocks on write isn't? and we do only read while probing?
We are still working on enabling ndctl test to run on POWER. Hence was not able capture these issues. Will try to get that resolved soon.
-aneesh
_______________________________________________
Linux-nvdimm mailing list -- [email protected] To unsubscribe send an email to [email protected]
On Wed, Oct 30, 2019 at 9:10 PM Aneesh Kumar K.V
<[email protected]> wrote:
[..]
Quoted 7 lines
Quoted 1 line
Can you share the call stack? we clear bad blocks on write isn't? and we do only read while probing?
We are still working on enabling
ndctl testto run on POWER. Hence was not able capture these issues. Will try to get that resolved soon.
Here's the splat. I've also attached a reworked version of the patch that has code organization fixups. It might be the case that it needs to fully move to a model where each nd_*_probe() routine does its own management of devm_namespace_enable() based on the sizes it needs to complete the probe process.
nd_pmem namespace3.0: request out of range
WARNING: CPU: 16 PID: 1551 at tools/testing/nvdimm/../../../drivers/nvdimm/claim.c:264 nsio_rw_bytes+0x14c/0x260 [libnvdimm] RIP: 0010:nsio_rw_bytes+0x14c/0x260 [libnvdimm] Call Trace: nd_pfn_validate+0x3d9/0x560 [libnvdimm] nd_pfn_probe+0xc0/0x150 [libnvdimm] ? nd_btt_probe+0x1e/0x270 [libnvdimm] nd_pmem_probe+0x63/0xc0 [nd_pmem] nvdimm_bus_probe+0x91/0x310 [libnvdimm]
That nd_pfn_validate+0x3d9 corresponds to the nvdimm_write_bytes() in nd_pfn_clear_memmap_errors().
libnvdimm/namespace: Differentiate between probe mapping and runtime mapping
From: Aneesh Kumar K.V <[email protected]>
The nvdimm core currently maps the full namespace to an ioremap range while probing the namespace mode. This can result in probe failures on architectures that have limited ioremap space.
For example, with a large btt namespace that consumes most of I/O remap range, depending on the sequence of namespace initialization, the user can find a pfn namespace initialization failure due to unavailable I/O remap space which nvdimm core uses for temporary mapping.
nvdimm core can avoid this failure by only mapping the reserved info block area to check for pfn superblock type and map the full namespace resource only before using the namespace.
Given that personalities like BTT can be layered on top of any namespace type create a generic form of devm_nsio_enable (devm_namespace_enable) and use it inside the per-personality attach routines. Now devm_namespace_enable() is always paired with disable unless the mapping is going to be used for long term runtime access.
--- drivers/dax/pmem/core.c | 6 +++--- drivers/nvdimm/btt.c | 10 ++++++++-- drivers/nvdimm/claim.c | 14 ++++++-------- drivers/nvdimm/namespace_devs.c | 17 +++++++++++++++++ drivers/nvdimm/nd-core.h | 16 ++++++++++++++++ drivers/nvdimm/nd.h | 22 +++++++++------------- drivers/nvdimm/pfn_devs.c | 9 ++------- drivers/nvdimm/pmem.c | 17 +++++++++++++---- 8 files changed, 74 insertions(+), 37 deletions(-)
drivers/dax/pmem/core.c +3 -3
@@ -25,20 +25,20 @@ struct dev_dax *__dax_pmem_probe(struct device *dev, enum dev_dax_subsys subsys) ndns =nvdimm_namespace_common_probe (dev);if (IS_ERR (ndns))return ERR_CAST (ndns); - nsio =to_nd_namespace_io (&ndns->dev);/* parse the 'pfn' info block via ->rw_bytes */ - rc =devm_nsio_enable (dev, nsio); + rc =devm_namespace_enable (dev, ndns,nd_info_block_reserve ());if (rc)return ERR_PTR (rc); rc =nvdimm_setup_pfn (nd_pfn, &pgmap);if (rc)return ERR_PTR (rc); -devm_nsio_disable (dev, nsio); +devm_namespace_disable (dev, ndns);/* reserve the metadata area, device-dax will reserve the data */ pfn_sb = nd_pfn->pfn_sb; offset =le64_to_cpu (pfn_sb->dataoff); + nsio =to_nd_namespace_io (&ndns->dev);if (!devm_request_mem_region (dev, nsio->res.start, offset,dev_name (&ndns->dev))) {dev_warn (dev,"could not reserve metadata\n" );
drivers/nvdimm/btt.c +8 -2
@@ -1674,7 +1674,8 @@ int nvdimm_namespace_attach_btt(struct nd_namespace_common *ndns)struct nd_region *nd_region;struct btt_sb *btt_sb;struct btt *btt; -size_t rawsize; +size_t size, rawsize; +int rc;if (!nd_btt->uuid || !nd_btt->ndns || !nd_btt->lbasize) {dev_dbg (&nd_btt->dev,"incomplete btt configuration\n" );
@@ -1685,6 +1686,11 @@ int nvdimm_namespace_attach_btt(struct nd_namespace_common *ndns)if (!btt_sb)return -ENOMEM ; + size =nvdimm_namespace_capacity (ndns); + rc =devm_namespace_enable (&nd_btt->dev, ndns, size); +if (rc) +return rc; +/* * If this returns < 0, that is ok as it just means there wasn't * an existing BTT, and we're creating a new one. We still need to
@@ -1693,7 +1699,7 @@ int nvdimm_namespace_attach_btt(struct nd_namespace_common *ndns)*/ nd_btt_version (nd_btt, ndns, btt_sb); - rawsize =nvdimm_namespace_capacity (ndns) - nd_btt->initial_offset; + rawsize = size - nd_btt->initial_offset;if (rawsize <ARENA_MIN_SIZE ) { dev_dbg(&nd_btt->dev, "%s must be at least %ld bytes\n", dev_name(&ndns->dev),
drivers/nvdimm/claim.c +6 -8
@@ -300,13 +300,14 @@ static int nsio_rw_bytes(struct nd_namespace_common *ndns,return rc; } -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio, +resource_size_t size) {struct resource *res = &nsio->res;struct nd_namespace_common *ndns = &nsio->common; - nsio->size =resource_size (res); -if (!devm_request_mem_region (dev, res->start,resource_size (res), + nsio->size = size; +if (!devm_request_mem_region (dev, res->start, size,dev_name (&ndns->dev))) {dev_warn (dev,"could not reserve region %pR\n" , res);return -EBUSY ;
@@ -318,12 +319,10 @@ int devm_nsio_enable(struct device *dev, struct nd_namespace_io *nsio)nvdimm_badblocks_populate (to_nd_region (ndns->dev.parent), &nsio->bb, &nsio->res); - nsio->addr =devm_memremap (dev, res->start,resource_size (res), -ARCH_MEMREMAP_PMEM ); + nsio->addr =devm_memremap (dev, res->start, size,ARCH_MEMREMAP_PMEM );return PTR_ERR_OR_ZERO (nsio->addr); } -EXPORT_SYMBOL_GPL (devm_nsio_enable);void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio) {
@@ -331,6 +330,5 @@ void devm_nsio_disable(struct device *dev, struct nd_namespace_io *nsio)devm_memunmap (dev, nsio->addr);devm_exit_badblocks (dev, &nsio->bb); -devm_release_mem_region (dev, res->start,resource_size (res)); +devm_release_mem_region (dev, res->start, nsio->size); } -EXPORT_SYMBOL_GPL (devm_nsio_disable);
drivers/nvdimm/namespace_devs.c +17 -0
@@ -1763,6 +1763,23 @@ struct nd_namespace_common *nvdimm_namespace_common_probe(struct device *dev) }EXPORT_SYMBOL (nvdimm_namespace_common_probe); +int devm_namespace_enable (struct device *dev,struct nd_namespace_common *ndns, +resource_size_t size) +{ +if (is_namespace_blk (&ndns->dev)) +return 0 ; +return devm_nsio_enable (dev,to_nd_namespace_io (&ndns->dev), size); +} +EXPORT_SYMBOL_GPL (devm_namespace_enable); + +void devm_namespace_disable (struct device *dev,struct nd_namespace_common *ndns) +{ +if (is_namespace_blk (&ndns->dev)) +return ; +devm_nsio_disable (dev,to_nd_namespace_io (&ndns->dev)); +} +EXPORT_SYMBOL_GPL (devm_namespace_disable); +static struct device **create_namespace_io (struct nd_region *nd_region) {struct nd_namespace_io *nsio;
drivers/nvdimm/nd-core.h +16 -0
@@ -171,6 +171,22 @@ ssize_t nd_namespace_store(struct device *dev,struct nd_pfn *to_nd_pfn_safe (struct device *dev);bool is_nvdimm_bus (struct device *dev); +#if IS_ENABLED (CONFIG_ND_CLAIM ) +int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio, +resource_size_t size); +void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio); +#else +static inline int devm_nsio_enable (struct device *dev, +struct nd_namespace_io *nsio,resource_size_t size) +{ +return -ENXIO ; +} + +void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio) +{ +} +#endif +#ifdef CONFIG_PROVE_LOCKING extern struct class *nd_class;
drivers/nvdimm/nd.h +9 -13
@@ -212,6 +212,11 @@ struct nd_dax {struct nd_pfn nd_pfn; }; +static inline u32 nd_info_block_reserve (void ) +{ +return ALIGN (SZ_8K ,PAGE_SIZE ); +} +enum nd_async_mode {ND_SYNC ,ND_ASYNC ,
@@ -373,29 +378,20 @@ const char *nvdimm_namespace_disk_name(struct nd_namespace_common *ndns,unsigned int pmem_sector_size (struct nd_namespace_common *ndns);void nvdimm_badblocks_populate (struct nd_region *nd_region,struct badblocks *bb,const struct resource *res); +int devm_namespace_enable (struct device *dev,struct nd_namespace_common *ndns, +resource_size_t size); +void devm_namespace_disable (struct device *dev, +struct nd_namespace_common *ndns);#if IS_ENABLED (CONFIG_ND_CLAIM ) -/* max struct page size independent of kernel config */ #define MAX_STRUCT_PAGE_SIZE 64 -int nvdimm_setup_pfn (struct nd_pfn *nd_pfn,struct dev_pagemap *pgmap); -int devm_nsio_enable (struct device *dev,struct nd_namespace_io *nsio); -void devm_nsio_disable (struct device *dev,struct nd_namespace_io *nsio);#else static inline int nvdimm_setup_pfn (struct nd_pfn *nd_pfn,struct dev_pagemap *pgmap) {return -ENXIO ; } -static inline int devm_nsio_enable (struct device *dev, -struct nd_namespace_io *nsio) -{ -return -ENXIO ; -} -static inline void devm_nsio_disable (struct device *dev, -struct nd_namespace_io *nsio) -{ -}#endif int nd_blk_region_init (struct nd_region *nd_region);int nd_region_activate (struct nd_region *nd_region);
drivers/nvdimm/pfn_devs.c +2 -7
@@ -627,11 +627,6 @@ int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns) }EXPORT_SYMBOL (nd_pfn_probe); -static u32 info_block_reserve (void ) -{ -return ALIGN (SZ_8K ,PAGE_SIZE ); -} -/* * We hotplug memory at sub-section granularity, pad the reserved area * from the previous section base to the namespace base address.
@@ -645,7 +640,7 @@ static unsigned long init_altmap_base(resource_size_t base)static unsigned long init_altmap_reserve (resource_size_t base) { -unsigned long reserve =info_block_reserve () >>PAGE_SHIFT ; +unsigned long reserve =nd_info_block_reserve () >>PAGE_SHIFT ;unsigned long base_pfn =PHYS_PFN (base); reserve += base_pfn -SUBSECTION_ALIGN_DOWN (base_pfn);
@@ -660,7 +655,7 @@ static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)u64 offset =le64_to_cpu (pfn_sb->dataoff);u32 start_pad =__le32_to_cpu (pfn_sb->start_pad);u32 end_trunc =__le32_to_cpu (pfn_sb->end_trunc); -u32 reserve =info_block_reserve (); +u32 reserve =nd_info_block_reserve ();struct nd_namespace_common *ndns = nd_pfn->ndns;struct nd_namespace_io *nsio =to_nd_namespace_io (&ndns->dev);resource_size_t base = nsio->res.start + start_pad;
drivers/nvdimm/pmem.c +13 -4
@@ -372,6 +372,10 @@ static int pmem_attach_disk(struct device *dev,if (!pmem)return -ENOMEM ; + rc =devm_namespace_enable (dev, ndns,nd_info_block_reserve ()); +if (rc) +return rc; +/* while nsio_rw_bytes is active, parse a pfn info block if present */ if (is_nd_pfn (dev)) { nd_pfn =to_nd_pfn (dev);
@@ -381,7 +385,7 @@ static int pmem_attach_disk(struct device *dev, }/* we're attaching a block device, disable raw namespace access */ -devm_nsio_disable (dev, nsio); +devm_namespace_disable (dev, ndns);dev_set_drvdata (dev, pmem); pmem->phys_addr = res->start;
@@ -497,15 +501,16 @@ static int nd_pmem_probe(struct device *dev)if (IS_ERR (ndns))return PTR_ERR (ndns); -if (devm_nsio_enable (dev,to_nd_namespace_io (&ndns->dev))) -return -ENXIO ; -if (is_nd_btt (dev))return nvdimm_namespace_attach_btt (ndns);if (is_nd_pfn (dev))return pmem_attach_disk (dev, ndns); + ret =devm_namespace_enable (dev, ndns,nd_info_block_reserve ()); +if (ret) +return ret; + ret =nd_btt_probe (dev, ndns);if (ret ==0 )return -ENXIO ;
@@ -532,6 +537,10 @@ static int nd_pmem_probe(struct device *dev)return -ENXIO ; else if (ret == -EOPNOTSUPP) return ret; + +/* probe complete, attach handles namespace enabling */ +devm_namespace_disable (dev ,ndns ); +return pmem_attach_disk (dev, ndns); }
_______________________________________________
Linux-nvdimm mailing list -- [email protected] To unsubscribe send an email to [email protected]