Skip to contentSkip to conversation

memremap fix for 4.3

GIT PULL Williams, Dan J 3 participants 4 messages Oct 29, 2015 MM core

lore
Share & tools
lore
#0 lore raw

Hi Linus, please pull from:

git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm libnvdimm-fixes

...to receive a small fixlet for 4.3.

The new memremap() api introduced in the 4.3 cycle to unify/replace ioremap_cache() and ioremap_wt() is mishandling the highmem case. This patch has received a build success notification from a 0day-kbuild-robot run and has been out for a review for a day. Russell has not had a chance to weigh in on it yet.

I do not think the usage of kmap is strictly necessary as we should be able to fall back to ioremap_cache(), but I include it for two reasons:

1/ ARM ioremap() will WARN if passed a pfn_valid() address.

2/ acpi_map() carries a similar kmap fallback, and that quirk can now be centrally carried in this common routine.

---

The following changes since commit 25cb62b76430a91cc6195f902e61c2cb84ade622:

Linux 4.3-rc5 (2015-10-11 11:09:45 -0700)

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm libnvdimm-fixes

for you to fetch changes up to bcaa4236b558092dd5ff14ea943e89fd944fcd28:

memremap: fix highmem support (2015-10-26 16:55:56 -0400)

----------------------------------------------------------------
Dan Williams (1):
memremap: fix highmem support

 include/linux/highmem.h | 12 ++++++++++++
 kernel/memremap.c       | 28 +++++++++++++++++++++++++---
 2 files changed, 37 insertions(+), 3 deletions(-)

commit bcaa4236b558092dd5ff14ea943e89fd944fcd28

Author: Dan Williams <[email protected]>
Date:   Mon Oct 26 16:55:56 2015 -0400

memremap: fix highmem support

Currently memremap checks if the range is "System RAM" and returns the kernel linear address. This is broken for highmem platforms where a range may be "System RAM", but is not part of the kernel linear mapping. Similar to acpi_map(), use kmap() for PAGE_SIZE memremap() requests for highmem, and fall back to ioremap_cache() otherwise.

The impact of this bug is low for now since the pmem driver is the only user of memremap(), but this is important to fix before more conversions to memremap arrive in 4.4.

include/linux/highmem.h +12 -0
index 6aefcd0031a6..c20cf24c76dd 100644
@@ -41,6 +41,13 @@ void kmap_flush_unused(void);
 
 struct page *kmap_to_page(void *addr);
 
+static inline bool is_kmap_addr(const void *x)
+{
+	unsigned long addr = (unsigned long) x;
+
+	return addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP);
+}
+
 #else /* CONFIG_HIGHMEM */
 
 static inline unsigned int nr_free_highpages(void) { return 0; }
@@ -50,6 +57,11 @@ static inline struct page *kmap_to_page(void *addr)
 	return virt_to_page(addr);
 }
 
+static inline bool is_kmap_addr(const void *x)
+{
+	return false;
+}
+
 #define totalhigh_pages 0UL
 
 #ifndef ARCH_HAS_KMAP
kernel/memremap.c +25 -3
index 72b0c66628b6..901d7ec3583a 100644
@@ -10,6 +10,7 @@
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  * General Public License for more details.
  */
+#include <linux/highmem.h>
 #include <linux/device.h>
 #include <linux/types.h>
 #include <linux/io.h>
@@ -24,6 +25,25 @@ __weak void __iomem *ioremap_cache(resource_size_t offset, unsigned long size)
 }
 #endif
 
+static void *try_ram_remap(resource_size_t offset, size_t size)
+{
+	struct page *page = pfn_to_page(offset >> PAGE_SHIFT);
+	unsigned int pg_off = offset & ~PAGE_MASK;
+
+	/* In the simple case just return the existing linear address */
+	if (!PageHighMem(page))
+		return __va(offset);
+
+	/*
+	 * Try kmap first since some arch ioremap implementations fail when
+	 * being passed a ram address.
+	 */
+	if (pg_off + size <= PAGE_SIZE)
+		return kmap(page) + pg_off;
+
+	return NULL;
+}
+
 /**
  * memremap() - remap an iomem_resource as cacheable memory
  * @offset: iomem resource start address
@@ -66,8 +86,8 @@ void *memremap(resource_size_t offset, size_t size, unsigned long flags)
 		 * the requested range is potentially in "System RAM"
 		 */
 		if (is_ram == REGION_INTERSECTS)
-			addr = __va(offset);
-		else
+			addr = try_ram_remap(offset, size);
+		if (!addr)
 			addr = ioremap_cache(offset, size);
 	}
 
@@ -94,7 +114,9 @@ EXPORT_SYMBOL(memremap);
 
 void memunmap(void *addr)
 {
-	if (is_vmalloc_addr(addr))
+	if (is_kmap_addr(addr))
+		kunmap(addr);
+	else if (is_vmalloc_addr(addr))
 		iounmap((void __iomem *) addr);
 }
 EXPORT_SYMBOL(memunmap);
#1 lore raw
Quoted 12 lines · Dan J

Hi Linus, please pull from:

git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm libnvdimm-fixes

...to receive a small fixlet for 4.3.

The new memremap() api introduced in the 4.3 cycle to unify/replace ioremap_cache() and ioremap_wt() is mishandling the highmem case. This patch has received a build success notification from a 0day-kbuild-robot run and has been out for a review for a day. Russell has not had a chance to weigh in on it yet.

Quoted 5 lines

I do not think the usage of kmap is strictly necessary as we should be able to fall back to ioremap_cache(), but I include it for two reasons:

1/ ARM ioremap() will WARN if passed a pfn_valid() address.

It will also return NULL

Quoted 3 lines

2/ acpi_map() carries a similar kmap fallback, and that quirk can now be centrally carried in this common routine.

I *really* think we should remove the kmap() fallback. memremap() is intended as a drop-in replacement for ioremap_cache(), and if the latter does not allow pfn_valid() pages to be ioremap'ed() on ARM, they shouldn't silently end up being kmap()'ed at some random time in the future when a certain user of ioremap_cache() gets upgraded to memremap().

Could we not merge this for 4.3 please, and give Russell and the other ARM folks some time to chime in? In the mean time, we could fix just the bug by adding the !PageHighmem() test in the code path that ends up returning the __va() result of the input physical address.

Regards,
Ard.

Quoted 134 lines

---

The following changes since commit 25cb62b76430a91cc6195f902e61c2cb84ade622:

Linux 4.3-rc5 (2015-10-11 11:09:45 -0700)

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm libnvdimm-fixes

for you to fetch changes up to bcaa4236b558092dd5ff14ea943e89fd944fcd28:

memremap: fix highmem support (2015-10-26 16:55:56 -0400)

----------------------------------------------------------------
Dan Williams (1):
memremap: fix highmem support

 include/linux/highmem.h | 12 ++++++++++++
 kernel/memremap.c       | 28 +++++++++++++++++++++++++---
 2 files changed, 37 insertions(+), 3 deletions(-)

commit bcaa4236b558092dd5ff14ea943e89fd944fcd28

Author: Dan Williams <[email protected]>
Date:   Mon Oct 26 16:55:56 2015 -0400

memremap: fix highmem support

Currently memremap checks if the range is "System RAM" and returns the kernel linear address. This is broken for highmem platforms where a range may be "System RAM", but is not part of the kernel linear mapping. Similar to acpi_map(), use kmap() for PAGE_SIZE memremap() requests for highmem, and fall back to ioremap_cache() otherwise.

The impact of this bug is low for now since the pmem driver is the only user of memremap(), but this is important to fix before more conversions to memremap arrive in 4.4.

include/linux/highmem.h +12 -0
index 6aefcd0031a6..c20cf24c76dd 100644
@@ -41,6 +41,13 @@ void kmap_flush_unused(void);
 
 struct page *kmap_to_page(void *addr);
 
+static inline bool is_kmap_addr(const void *x)
+{
+       unsigned long addr = (unsigned long) x;
+
+       return addr >= PKMAP_ADDR(0) && addr < PKMAP_ADDR(LAST_PKMAP);
+}
+
 #else /* CONFIG_HIGHMEM */
 
 static inline unsigned int nr_free_highpages(void) { return 0; }
@@ -50,6 +57,11 @@ static inline struct page *kmap_to_page(void *addr)
        return virt_to_page(addr);
 }
 
+static inline bool is_kmap_addr(const void *x)
+{
+       return false;
+}
+
 #define totalhigh_pages 0UL
 
 #ifndef ARCH_HAS_KMAP
kernel/memremap.c +25 -3
index 72b0c66628b6..901d7ec3583a 100644
@@ -10,6 +10,7 @@
  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  * General Public License for more details.
  */
+#include <linux/highmem.h>
 #include <linux/device.h>
 #include <linux/types.h>
 #include <linux/io.h>
@@ -24,6 +25,25 @@ __weak void __iomem *ioremap_cache(resource_size_t offset, unsigned long size)
 }
 #endif
 
+static void *try_ram_remap(resource_size_t offset, size_t size)
+{
+       struct page *page = pfn_to_page(offset >> PAGE_SHIFT);
+       unsigned int pg_off = offset & ~PAGE_MASK;
+
+       /* In the simple case just return the existing linear address */
+       if (!PageHighMem(page))
+               return __va(offset);
+
+       /*
+        * Try kmap first since some arch ioremap implementations fail when
+        * being passed a ram address.
+        */
+       if (pg_off + size <= PAGE_SIZE)
+               return kmap(page) + pg_off;
+
+       return NULL;
+}
+
 /**
  * memremap() - remap an iomem_resource as cacheable memory
  * @offset: iomem resource start address
@@ -66,8 +86,8 @@ void *memremap(resource_size_t offset, size_t size, unsigned long flags)
                 * the requested range is potentially in "System RAM"
                 */
                if (is_ram == REGION_INTERSECTS)
-                       addr = __va(offset);
-               else
+                       addr = try_ram_remap(offset, size);
+               if (!addr)
                        addr = ioremap_cache(offset, size);
        }
 
@@ -94,7 +114,9 @@ EXPORT_SYMBOL(memremap);
 
 void memunmap(void *addr)
 {
-       if (is_vmalloc_addr(addr))
+       if (is_kmap_addr(addr))
+               kunmap(addr);
+       else if (is_vmalloc_addr(addr))
                iounmap((void __iomem *) addr);
 }
 EXPORT_SYMBOL(memunmap);
#2 lore raw
In reply to Williams, Dan J
Quoted 11 lines · Dan J

Hi Linus, please pull from:

git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm libnvdimm-fixes

...to receive a small fixlet for 4.3.

The new memremap() api introduced in the 4.3 cycle to unify/replace ioremap_cache() and ioremap_wt() is mishandling the highmem case. This patch has received a build success notification from a 0day-kbuild-robot run and has been out for a review for a day. Russell has not had a chance to weigh in on it yet.

Oh, was this merged for 4.3-rc1? I haven't noticed any problems if it has.

Quoted 4 lines

I do not think the usage of kmap is strictly necessary as we should be able to fall back to ioremap_cache(), but I include it for two reasons:

1/ ARM ioremap() will WARN if passed a pfn_valid() address.

We don't support ioremap() on system RAM on ARM, period. That's because ioremap() sets up page tables with incompatible attributes compared to those which are/will be setup by the lowmem/kmap* mappings, which leads to "unpredictable" behaviour.

The only time RAM is mappable with ioremap() is if it's stolen from the kernel at boot time, which prevents the kernel from managing it and setting up memory-like mappings.

Signature · 2 lines
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
#3 lore raw
Quoted 21 lines · Russell King - ARM Linux
Quoted 11 lines · Dan J

Hi Linus, please pull from:

git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm libnvdimm-fixes

...to receive a small fixlet for 4.3.

The new memremap() api introduced in the 4.3 cycle to unify/replace ioremap_cache() and ioremap_wt() is mishandling the highmem case. This patch has received a build success notification from a 0day-kbuild-robot run and has been out for a review for a day. Russell has not had a chance to weigh in on it yet.

Oh, was this merged for 4.3-rc1? I haven't noticed any problems if it has.

Quoted 4 lines

I do not think the usage of kmap is strictly necessary as we should be able to fall back to ioremap_cache(), but I include it for two reasons:

1/ ARM ioremap() will WARN if passed a pfn_valid() address.

Quoted 9 lines

We don't support ioremap() on system RAM on ARM, period. That's because ioremap() sets up page tables with incompatible attributes compared to those which are/will be setup by the lowmem/kmap* mappings, which leads to "unpredictable" behaviour.

The only time RAM is mappable with ioremap() is if it's stolen from the kernel at boot time, which prevents the kernel from managing it and setting up memory-like mappings.

Ok, I read that as: "if someone calls memremap() on a 'System RAM' address on ARM and we can't find the kernel linear address then just pass it through to arch level remap code where it should rightly WARN." I'll reflow the patch with that change.