dax: Fix xarray entry association for mixed mappings
Jan Kara
reviewerPATCH 1/1 dax: Fix xarray entry association for mixed mappings
When inserting entry into xarray, we store mapping and index in corresponding struct pages for memory error handling. When it happened that one process was mapping file at PMD granularity while another process at PTE granularity, we could wrongly deassociate PMD range and then reassociate PTE range leaving the rest of struct pages in PMD range without mapping information which could later cause missed notifications about memory errors. Fix the problem by calling the association / deassociation code if and only if we are really going to update the xarray (deassociating and associating zero or empty entries is just no-op so there's no reason to complicate the code with trying to avoid the calls for these cases).
--- fs/dax.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)
fs/dax.c +4 -5
@@ -728,12 +728,11 @@ static void *dax_insert_entry(struct xa_state *xas,xas_reset (xas);xas_lock_irq (xas); -if (dax_entry_size (entry) !=dax_entry_size (new_entry)) { +if (dax_is_zero_entry (entry) ||dax_is_empty_entry (entry)) { +void *old; +dax_disassociate_entry (entry, mapping, false);dax_associate_entry (new_entry, mapping, vmf->vma, vmf->address); - } - -if (dax_is_zero_entry (entry) ||dax_is_empty_entry (entry)) {/* * Only swap our new entry into the page cache if the current * entry is a zero page or an empty entry. If a normal PTE or
@@ -742,7 +741,7 @@ static void *dax_insert_entry(struct xa_state *xas,* existing entry is a PMD, we will just leave the PMD in the * tree and dirty it if necessary. */ -void *old =dax_lock_entry (xas, new_entry); + old =dax_lock_entry (xas, new_entry);WARN_ON_ONCE (old !=xa_mk_value (xa_to_value (entry) |DAX_LOCKED )); entry = new_entry;
Signature · 1 line
2.16.4
Quoted 12 lines · Jan Kara
When inserting entry into xarray, we store mapping and index in corresponding struct pages for memory error handling. When it happened that one process was mapping file at PMD granularity while another process at PTE granularity, we could wrongly deassociate PMD range and then reassociate PTE range leaving the rest of struct pages in PMD range without mapping information which could later cause missed notifications about memory errors. Fix the problem by calling the association / deassociation code if and only if we are really going to update the xarray (deassociating and associating zero or empty entries is just no-op so there's no reason to complicate the code with trying to avoid the calls for these cases).
Looks good to me, I assume this also needs:
collides with a busy page")
Quoted 36 lines
Signed-off-by Jan Kara <[email protected]>--- fs/dax.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)fs/dax.c +4
-6@@ -728,12 +728,11 @@ static void *dax_insert_entry(struct xa_state *xas,xas_reset (xas);xas_lock_irq (xas); -if (dax_entry_size (entry) !=dax_entry_size (new_entry)) { +if (dax_is_zero_entry (entry) ||dax_is_empty_entry (entry)) { +void *old; +dax_disassociate_entry (entry, mapping, false);dax_associate_entry (new_entry, mapping, vmf->vma, vmf->address); - } - -if (dax_is_zero_entry (entry) ||dax_is_empty_entry (entry)) {/* * Only swap our new entry into the page cache if the current * entry is a zero page or an empty entry. If a normal PTE or @@ -742,7 +741,7 @@ static void *dax_insert_entry(struct xa_state *xas,* existing entry is a PMD, we will just leave the PMD in the * tree and dirty it if necessary. */ -void *old =dax_lock_entry (xas, new_entry); + old =dax_lock_entry (xas, new_entry);WARN_ON_ONCE (old !=xa_mk_value (xa_to_value (entry) |DAX_LOCKED )); entry = new_entry; --
Quoted 2 lines
2.16.4
_______________________________________________
Linux-nvdimm mailing list
[email protected]
https://lists.01.org/mailman/listinfo/linux-nvdimm
Jan Kara
reviewerQuoted 14 lines · Dan Williams
Quoted 12 lines · Jan Kara
When inserting entry into xarray, we store mapping and index in corresponding struct pages for memory error handling. When it happened that one process was mapping file at PMD granularity while another process at PTE granularity, we could wrongly deassociate PMD range and then reassociate PTE range leaving the rest of struct pages in PMD range without mapping information which could later cause missed notifications about memory errors. Fix the problem by calling the association / deassociation code if and only if we are really going to update the xarray (deassociating and associating zero or empty entries is just no-op so there's no reason to complicate the code with trying to avoid the calls for these cases).
Quoted 5 lines
Looks good to me, I assume this also needs:
Cc <[email protected]>Fixes d2c997c0f145 ("fs, dax: use page->mapping to warn if truncatecollides with a busy page")
Yes, thanks for that.
Honza
Quoted 39 lines
Quoted 39 lines
Signed-off-by Jan Kara <[email protected]>--- fs/dax.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)fs/dax.c +4
-6@@ -728,12 +728,11 @@ static void *dax_insert_entry(struct xa_state *xas,xas_reset (xas);xas_lock_irq (xas); -if (dax_entry_size (entry) !=dax_entry_size (new_entry)) { +if (dax_is_zero_entry (entry) ||dax_is_empty_entry (entry)) { +void *old; +dax_disassociate_entry (entry, mapping, false);dax_associate_entry (new_entry, mapping, vmf->vma, vmf->address); - } - -if (dax_is_zero_entry (entry) ||dax_is_empty_entry (entry)) {/* * Only swap our new entry into the page cache if the current * entry is a zero page or an empty entry. If a normal PTE or @@ -742,7 +741,7 @@ static void *dax_insert_entry(struct xa_state *xas,* existing entry is a PMD, we will just leave the PMD in the * tree and dirty it if necessary. */ -void *old =dax_lock_entry (xas, new_entry); + old =dax_lock_entry (xas, new_entry);WARN_ON_ONCE (old !=xa_mk_value (xa_to_value (entry) |DAX_LOCKED )); entry = new_entry; --2.16.4
Signature · 6 lines
Jan Kara <[email protected]> SUSE Labs, CR _______________________________________________ Linux-nvdimm mailing list [email protected] https://lists.01.org/mailman/listinfo/linux-nvdimm