Skip to contentSkip to conversation

erofs: use kmap_local_page() only for erofs_bread()

PATCH In review Gao Xiang 5 participants 14 messages Oct 18, 2022 – Dec 7, 2022 EROFS file system

lore
Share & tools
lore

Gao Xiang

author
#0 lore raw

PATCH v2 1/1 erofs: use kmap_local_page() only for erofs_bread()

Convert all mapped erofs_bread() users to use kmap_local_page() instead of kmap() or kmap_atomic().

---
 fs/erofs/data.c     | 8 ++------
 fs/erofs/internal.h | 3 +--
 fs/erofs/xattr.c    | 8 ++++----
 fs/erofs/zmap.c     | 4 ++--
 4 files changed, 9 insertions(+), 14 deletions(-)
fs/erofs/data.c +2 -6
index fe8ac0e163f7..fe1ae80284bf 100644
@@ -13,9 +13,7 @@
 void erofs_unmap_metabuf(struct erofs_buf *buf)
 {
 	if (buf->kmap_type == EROFS_KMAP)
-		kunmap(buf->page);
-	else if (buf->kmap_type == EROFS_KMAP_ATOMIC)
-		kunmap_atomic(buf->base);
+		kunmap_local(buf->base);
 	buf->base = NULL;
 	buf->kmap_type = EROFS_NO_KMAP;
 }
@@ -54,9 +52,7 @@ void *erofs_bread(struct erofs_buf *buf, struct inode *inode,
 	}
 	if (buf->kmap_type == EROFS_NO_KMAP) {
 		if (type == EROFS_KMAP)
-			buf->base = kmap(page);
-		else if (type == EROFS_KMAP_ATOMIC)
-			buf->base = kmap_atomic(page);
+			buf->base = kmap_local_page(page);
 		buf->kmap_type = type;
 	} else if (buf->kmap_type != type) {
 		DBG_BUGON(1);
fs/erofs/internal.h +1 -2
index 1701df48c446..67dc8e177211 100644
@@ -253,8 +253,7 @@ static inline int erofs_wait_on_workgroup_freezed(struct erofs_workgroup *grp)
 
 enum erofs_kmap_type {
 	EROFS_NO_KMAP,		/* don't map the buffer */
-	EROFS_KMAP,		/* use kmap() to map the buffer */
-	EROFS_KMAP_ATOMIC,	/* use kmap_atomic() to map the buffer */
+	EROFS_KMAP,		/* use kmap_local_page() to map the buffer */
 };
 
 struct erofs_buf {
fs/erofs/xattr.c +4 -4
index 8106bcb5a38d..a62fb8a3318a 100644
@@ -148,7 +148,7 @@ static inline int xattr_iter_fixup(struct xattr_iter *it)
 
 	it->blkaddr += erofs_blknr(it->ofs);
 	it->kaddr = erofs_read_metabuf(&it->buf, it->sb, it->blkaddr,
-				       EROFS_KMAP_ATOMIC);
+				       EROFS_KMAP);
 	if (IS_ERR(it->kaddr))
 		return PTR_ERR(it->kaddr);
 	it->ofs = erofs_blkoff(it->ofs);
@@ -174,7 +174,7 @@ static int inline_xattr_iter_begin(struct xattr_iter *it,
 	it->ofs = erofs_blkoff(iloc(sbi, vi->nid) + inline_xattr_ofs);
 
 	it->kaddr = erofs_read_metabuf(&it->buf, inode->i_sb, it->blkaddr,
-				       EROFS_KMAP_ATOMIC);
+				       EROFS_KMAP);
 	if (IS_ERR(it->kaddr))
 		return PTR_ERR(it->kaddr);
 	return vi->xattr_isize - xattr_header_sz;
@@ -368,7 +368,7 @@ static int shared_getxattr(struct inode *inode, struct getxattr_iter *it)
 
 		it->it.ofs = xattrblock_offset(sbi, vi->xattr_shared_xattrs[i]);
 		it->it.kaddr = erofs_read_metabuf(&it->it.buf, sb, blkaddr,
-						  EROFS_KMAP_ATOMIC);
+						  EROFS_KMAP);
 		if (IS_ERR(it->it.kaddr))
 			return PTR_ERR(it->it.kaddr);
 		it->it.blkaddr = blkaddr;
@@ -580,7 +580,7 @@ static int shared_listxattr(struct listxattr_iter *it)
 
 		it->it.ofs = xattrblock_offset(sbi, vi->xattr_shared_xattrs[i]);
 		it->it.kaddr = erofs_read_metabuf(&it->it.buf, sb, blkaddr,
-						  EROFS_KMAP_ATOMIC);
+						  EROFS_KMAP);
 		if (IS_ERR(it->it.kaddr))
 			return PTR_ERR(it->it.kaddr);
 		it->it.blkaddr = blkaddr;
fs/erofs/zmap.c +2 -2
index 0bb66927e3d0..749a5ac943f4 100644
@@ -178,7 +178,7 @@ static int legacy_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 	unsigned int advise, type;
 
 	m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
-				      erofs_blknr(pos), EROFS_KMAP_ATOMIC);
+				      erofs_blknr(pos), EROFS_KMAP);
 	if (IS_ERR(m->kaddr))
 		return PTR_ERR(m->kaddr);
 
@@ -416,7 +416,7 @@ static int compacted_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 out:
 	pos += lcn * (1 << amortizedshift);
 	m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
-				      erofs_blknr(pos), EROFS_KMAP_ATOMIC);
+				      erofs_blknr(pos), EROFS_KMAP);
 	if (IS_ERR(m->kaddr))
 		return PTR_ERR(m->kaddr);
 	return unpack_compacted_index(m, amortizedshift, pos, lookahead);
Signature · 1 line
2.24.4
#1 lore raw

fs/erofs/data.c void *erofs_bread(struct erofs_buf *buf, struct inode

Quoted 27 lines · CEST Gao Xiang

Convert all mapped erofs_bread() users to use kmap_local_page() instead of kmap() or kmap_atomic().

---
 fs/erofs/data.c     | 8 ++------
 fs/erofs/internal.h | 3 +--
 fs/erofs/xattr.c    | 8 ++++----
 fs/erofs/zmap.c     | 4 ++--
 4 files changed, 9 insertions(+), 14 deletions(-)
fs/erofs/data.c +1 -3
index fe8ac0e163f7..fe1ae80284bf 100644
@@ -13,9 +13,7 @@
 void erofs_unmap_metabuf(struct erofs_buf *buf)
 {
 	if (buf->kmap_type == EROFS_KMAP)
-		kunmap(buf->page);
-	else if (buf->kmap_type == EROFS_KMAP_ATOMIC)
-		kunmap_atomic(buf->base);
+		kunmap_local(buf->base);
 	buf->base = NULL;
 	buf->kmap_type = EROFS_NO_KMAP;
 }
@@ -54,9 +52,7 @@ void *erofs_bread(struct erofs_buf *buf, struct inode 

*inode,

fs/erofs/internal.h static inline int erofs_wait_on_workgroup_freezed(struct

Quoted 15 lines
fs/erofs/data.c +1 -3
 	}
 	if (buf->kmap_type == EROFS_NO_KMAP) {
 		if (type == EROFS_KMAP)
-			buf->base = kmap(page);
-		else if (type == EROFS_KMAP_ATOMIC)
-			buf->base = kmap_atomic(page);
+			buf->base = kmap_local_page(page);
 		buf->kmap_type = type;
 	} else if (buf->kmap_type != type) {
 		DBG_BUGON(1);
fs/erofs/internal.h +0 -0
index 1701df48c446..67dc8e177211 100644
@@ -253,8 +253,7 @@ static inline int erofs_wait_on_workgroup_freezed(struct 

erofs_workgroup *grp)

fs/erofs/internal.h static inline int erofs_wait_on_workgroup_freezed(struct erofs_workgroup *grp)

Quoted 6 lines
fs/erofs/internal.h +1 -2
 enum erofs_kmap_type {
 	EROFS_NO_KMAP,		/* don't map the buffer */
-	EROFS_KMAP,		/* use kmap() to map the buffer */
-	EROFS_KMAP_ATOMIC,	/* use kmap_atomic() to map the buffer */
+	EROFS_KMAP,		/* use kmap_local_page() to map the 

buffer */

fs/erofs/xattr.c static int inline_xattr_iter_begin(struct xattr_iter

Quoted 17 lines
fs/erofs/internal.h +0 -0
 };
 
 struct erofs_buf {
fs/erofs/xattr.c +1 -1
index 8106bcb5a38d..a62fb8a3318a 100644
@@ -148,7 +148,7 @@ static inline int xattr_iter_fixup(struct xattr_iter *it)
 
 	it->blkaddr += erofs_blknr(it->ofs);
 	it->kaddr = erofs_read_metabuf(&it->buf, it->sb, it->blkaddr,
-				       EROFS_KMAP_ATOMIC);
+				       EROFS_KMAP);
 	if (IS_ERR(it->kaddr))
 		return PTR_ERR(it->kaddr);
 	it->ofs = erofs_blkoff(it->ofs);
@@ -174,7 +174,7 @@ static int inline_xattr_iter_begin(struct xattr_iter 

*it,

fs/erofs/xattr.c static int shared_getxattr(struct inode *inode, struct

Quoted 9 lines
fs/erofs/xattr.c +1 -1
 	it->ofs = erofs_blkoff(iloc(sbi, vi->nid) + inline_xattr_ofs);
 
 	it->kaddr = erofs_read_metabuf(&it->buf, inode->i_sb, it->blkaddr,
-				       EROFS_KMAP_ATOMIC);
+				       EROFS_KMAP);
 	if (IS_ERR(it->kaddr))
 		return PTR_ERR(it->kaddr);
 	return vi->xattr_isize - xattr_header_sz;
@@ -368,7 +368,7 @@ static int shared_getxattr(struct inode *inode, struct 

getxattr_iter *it)

fs/erofs/xattr.c static int shared_getxattr(struct inode *inode, struct · line 370

Quoted 4 lines
fs/erofs/xattr.c +0 -0
 		it->it.ofs = xattrblock_offset(sbi, vi-
 xattr_shared_xattrs[i]);
 		it->it.kaddr = erofs_read_metabuf(&it->it.buf, sb, 

blkaddr,

fs/erofs/xattr.c static int shared_getxattr(struct inode *inode, struct · line 371

Quoted 1 line
fs/erofs/xattr.c +0 -1
-						  

EROFS_KMAP_ATOMIC);

fs/erofs/xattr.c static int shared_getxattr(struct inode *inode, struct · line 371

Quoted 1 line
fs/erofs/xattr.c +1 -0
+						  

EROFS_KMAP);

fs/erofs/xattr.c static int shared_listxattr(struct listxattr_iter *it) · line 583

Quoted 8 lines
fs/erofs/xattr.c +0 -0
 		if (IS_ERR(it->it.kaddr))
 			return PTR_ERR(it->it.kaddr);
 		it->it.blkaddr = blkaddr;
@@ -580,7 +580,7 @@ static int shared_listxattr(struct listxattr_iter *it)
 
 		it->it.ofs = xattrblock_offset(sbi, vi-
 xattr_shared_xattrs[i]);
 		it->it.kaddr = erofs_read_metabuf(&it->it.buf, sb, 

blkaddr,

fs/erofs/xattr.c static int shared_listxattr(struct listxattr_iter *it) · line 584

Quoted 1 line
fs/erofs/xattr.c +0 -1
-						  

EROFS_KMAP_ATOMIC);

fs/erofs/xattr.c static int shared_listxattr(struct listxattr_iter *it) · line 584

Quoted 1 line
fs/erofs/xattr.c +1 -0
+						  

EROFS_KMAP);

fs/erofs/zmap.c static int legacy_load_cluster_from_disk(struct

Quoted 8 lines
fs/erofs/xattr.c +0 -0
 		if (IS_ERR(it->it.kaddr))
 			return PTR_ERR(it->it.kaddr);
fs/erofs/xattr.c +0 -0
 		it->it.blkaddr = blkaddr;
fs/erofs/zmap.c +0 -0
index 0bb66927e3d0..749a5ac943f4 100644
@@ -178,7 +178,7 @@ static int legacy_load_cluster_from_disk(struct 

z_erofs_maprecorder *m,

fs/erofs/zmap.c static int legacy_load_cluster_from_disk(struct z_erofs_maprecorder *m,

Quoted 4 lines
fs/erofs/zmap.c +0 -1
 	unsigned int advise, type;
 
 	m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
-				      erofs_blknr(pos), 

EROFS_KMAP_ATOMIC);

fs/erofs/zmap.c static int legacy_load_cluster_from_disk(struct · line 181

Quoted 1 line
fs/erofs/zmap.c +1 -0
+				      erofs_blknr(pos), 

EROFS_KMAP);

fs/erofs/zmap.c static int compacted_load_cluster_from_disk(struct

Quoted 4 lines
fs/erofs/zmap.c +0 -0
 	if (IS_ERR(m->kaddr))
 		return PTR_ERR(m->kaddr);
 
@@ -416,7 +416,7 @@ static int compacted_load_cluster_from_disk(struct 

z_erofs_maprecorder *m,

fs/erofs/zmap.c static int compacted_load_cluster_from_disk(struct z_erofs_maprecorder *m,

Quoted 4 lines
fs/erofs/zmap.c +0 -1
 out:
 	pos += lcn * (1 << amortizedshift);
 	m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
-				      erofs_blknr(pos), 

EROFS_KMAP_ATOMIC);

fs/erofs/zmap.c static int compacted_load_cluster_from_disk(struct · line 419

Quoted 1 line
fs/erofs/zmap.c +1 -0
+				      erofs_blknr(pos), 

EROFS_KMAP);

Quoted 7 lines
fs/erofs/zmap.c +0 -0
 	if (IS_ERR(m->kaddr))
 		return PTR_ERR(m->kaddr);
 	return unpack_compacted_index(m, amortizedshift, pos, lookahead);
Signature · 1 line
2.24.4

Can you please say whether or not you ran some tests on a 32 bits machine, possibly QEMU/KVM x86_32 with enough RAM (I'd suggest 6GB), booting a kernel with HIGHMEM4G or HIGHMEM64G enabled?

Did you check whether or not the pointers are handed to other contexts?

You didn't say anything about the safety of these conversions. And you also didn't say why you did them (I don't care but many might ignore why).

Therefore, can you please say at least how you checked that these conversions are safe so that they won't break the filesystem in 32bit machines?

Thanks,

Fabio M. De Francesco

#2 lore raw
In reply to Gao Xiang
Quoted 11 lines · CEST Gao Xiang

Convert all mapped erofs_bread() users to use kmap_local_page() instead of kmap() or kmap_atomic().

---
 fs/erofs/data.c     | 8 ++------
 fs/erofs/internal.h | 3 +--
 fs/erofs/xattr.c    | 8 ++++----
 fs/erofs/zmap.c     | 4 ++--
 4 files changed, 9 insertions(+), 14 deletions(-)

I just realized that you know the code of fs/erofs very well. I saw a Gao Xiang in MAINTAINERS, although having a different email address.

Therefore, I'm sure that everybody can trust that you checked everything is needed to assure the safety of the conversions.

However, an extended commit message would have prevented me to send you the previous email with all those questions / objections.

Thanks,

Fabio

author
#3 lore raw

Hi Fabio,

Quoted 13 lines · Fabio M. De Francesco
Quoted 11 lines · CEST Gao Xiang

Convert all mapped erofs_bread() users to use kmap_local_page() instead of kmap() or kmap_atomic().

---
 fs/erofs/data.c     | 8 ++------
 fs/erofs/internal.h | 3 +--
 fs/erofs/xattr.c    | 8 ++++----
 fs/erofs/zmap.c     | 4 ++--
 4 files changed, 9 insertions(+), 14 deletions(-)
Quoted 8 lines

I just realized that you know the code of fs/erofs very well. I saw a Gao Xiang in MAINTAINERS, although having a different email address.

Therefore, I'm sure that everybody can trust that you checked everything is needed to assure the safety of the conversions.

However, an extended commit message would have prevented me to send you the previous email with all those questions / objections.

Thanks for your suggestion. Yeah, this conversion looks trivial [since most paths for erofs_bread() don't have more restriction in principle so we can just disable migration. One of what I need to care is nested kmap() usage, some unmap/remap order cannot be simply converted to kmap_local() but I think it's not the case for erofs_bread(). Actually EROFS has one of such nested kmap() usage, but I don't really care its performance on HIGHMEM platforms, so I think kmap() is still somewhat useful compared to kmap_local() from this point of view], but in order to make it all work properly, I will try to do stress test with 32-bit platform later. Since it targets for the next cycle 6.2, I will do a full stress test in the next following weeks.

Thanks,
Gao Xiang

Quoted 5 lines

Thanks,

Fabio

#4 lore raw
Quoted 20 lines · CEST Gao Xiang

Hi Fabio,

Quoted 17 lines · Fabio M. De Francesco
Quoted 11 lines · CEST Gao Xiang

Convert all mapped erofs_bread() users to use kmap_local_page() instead of kmap() or kmap_atomic().

---
 fs/erofs/data.c     | 8 ++------
 fs/erofs/internal.h | 3 +--
 fs/erofs/xattr.c    | 8 ++++----
 fs/erofs/zmap.c     | 4 ++--
 4 files changed, 9 insertions(+), 14 deletions(-)

I just realized that you know the code of fs/erofs very well. I saw a Gao Xiang in MAINTAINERS, although having a different email address.

Therefore, I'm sure that everybody can trust that you checked everything

is

Quoted 3 lines
Quoted 3 lines

needed to assure the safety of the conversions.

However, an extended commit message would have prevented me to send you

the

Quoted 6 lines
Quoted 1 line

previous email with all those questions / objections.

Thanks for your suggestion.
Yeah, this conversion looks trivial [since most
paths for erofs_bread() don't have more restriction in principle so we can just disable migration.

Not sure about what you mean by "restrictions". Few months ago I updated the kmap_local_page() documentantation (highmem.rst). Please take a look at it, so that you may check if what you call restrictions are intended the way you mean.

The two most important rules are (1) that users cannot hand the virtual kernel addresses returned by kmap_local_page() to other contexts (that is why they are thread local) and (2) how to nest mappings /unmappings.

Quoted 2 lines

One of what I need to care is nested kmap() usage,
some unmap/remap order cannot be simply converted to kmap_local()

Correct about nesting. If we map A and then B, we must unmap B and then A.

However, as you seem to convey, not always unmappings in right order (stack based) is possible, sometimes because very long functions have many loop's breaks and many goto exit labels.

Quoted 4 lines

but I think
it's not the case for erofs_bread(). Actually EROFS has one of such nested kmap() usage, but I don't really care its performance on HIGHMEM platforms, so I think kmap() is still somewhat useful compared to kmap_local() from

this

Quoted 1 line

point of view],

In Btrfs I solved (thanks to David S.' advice) by mapping only one of two pages, only the one coming from the page cache.

The other page didn't need the use of kmap_local_page() because it was allocated in the filesystem with "alloc_page(GFP_NOFS)". GFP_NOFS won't ever allocate from ZONE_HIGHMEM, therefore a direct page_address() could avoid the mapping and the nesting issues.

Did you check if you may solve the same way?

A little group of people are working to remove all kmap() and kmap_atomic() we meet across the whole kernel. I have not yet encountered conversions which cannot be made. Sometimes we may refactor, if what I said above cannot apply.

Quoted 2 lines

but in order to make it all work properly, I will try to do stress test with 32-bit platform later.

I use QEMU/KVM x86_32 VM, 6GB RAM, and a kernel with HIGHMEM64 enabled and an openSUSE Tumbleweed 32 distro. I've heard that Debian too provides an x86_32 distribution.

Quoted 6 lines

Since it targets for the next cycle
6.2, I will do a full stress test in the next following weeks.

Thanks,
Gao Xiang

Thanks,

Fabio

author
#5 lore raw
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

...

Quoted 20 lines
Quoted 3 lines

One of what I need to care is nested kmap() usage,
some unmap/remap order cannot be simply converted to kmap_local()

Correct about nesting. If we map A and then B, we must unmap B and then A.

However, as you seem to convey, not always unmappings in right order (stack based) is possible, sometimes because very long functions have many loop's breaks and many goto exit labels.

Quoted 4 lines

but I think
it's not the case for erofs_bread(). Actually EROFS has one of such nested kmap() usage, but I don't really care its performance on HIGHMEM platforms, so I think kmap() is still somewhat useful compared to kmap_local() from

this

Quoted 1 line

point of view],

In Btrfs I solved (thanks to David S.' advice) by mapping only one of two pages, only the one coming from the page cache.

Quoted 6 lines

The other page didn't need the use of kmap_local_page() because it was allocated in the filesystem with "alloc_page(GFP_NOFS)". GFP_NOFS won't ever allocate from ZONE_HIGHMEM, therefore a direct page_address() could avoid the mapping and the nesting issues.

Did you check if you may solve the same way?

That is not simple. Currently we have compressed pages and decompressed pages (page cache or others), and they can be unmapped when either data is all consumed, so compressed pages can be unmapped first, or decompressed pages can be unmapped first. That quite depends on which pages goes first.

I think such usage is a quite common pattern for decoder or encoder, you could take a look at z_erofs_lzma_decompress() in fs/erofs/decompressor_lzma.c. So kmap() is still useful for such cases since I don't really care the HIGHMEM performance but correctness, but other alternative could churn/complex the map/unmap/remap pattern.

Thanks,
Gao Xiang

Quoted 23 lines

A little group of people are working to remove all kmap() and kmap_atomic() we meet across the whole kernel. I have not yet encountered conversions which cannot be made. Sometimes we may refactor, if what I said above cannot apply.

Quoted 2 lines

but in order to make it all work properly, I will try to do stress test with 32-bit platform later.

I use QEMU/KVM x86_32 VM, 6GB RAM, and a kernel with HIGHMEM64 enabled and an openSUSE Tumbleweed 32 distro. I've heard that Debian too provides an x86_32 distribution.

Quoted 6 lines

Since it targets for the next cycle
6.2, I will do a full stress test in the next following weeks.

Thanks,
Gao Xiang

Thanks,

Fabio

#6 lore raw
Quoted 4 lines · Gao Xiang
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

...

[snip]

Quoted 17 lines
Quoted 10 lines

In Btrfs I solved (thanks to David S.' advice) by mapping only one of two pages, only the one coming from the page cache.

The other page didn't need the use of kmap_local_page() because it was allocated in the filesystem with "alloc_page(GFP_NOFS)". GFP_NOFS won't ever allocate from ZONE_HIGHMEM, therefore a direct page_address() could avoid the mapping and the nesting issues.

Did you check if you may solve the same way?

That is not simple. Currently we have compressed pages and decompressed pages (page cache or others), and they can be unmapped when either data is all consumed, so compressed pages can be unmapped first, or decompressed pages can be unmapped first. That quite depends on which pages goes first.

Quoted 6 lines

I think such usage is a quite common pattern for decoder or encoder, you could take a look at z_erofs_lzma_decompress() in fs/erofs/decompressor_lzma.c. So kmap() is still useful for such cases since I don't really care the HIGHMEM performance but correctness, but other alternative could churn/complex the map/unmap/remap pattern.

When you say kmap() is still useful is this because of the map/unmap ordering restrictions or because the address is required in different threads?

Ira

author
#7 lore raw

Hi Ira,

Quoted 32 lines · Ira Weiny
Quoted 4 lines · Gao Xiang
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

...

[snip]

Quoted 17 lines
Quoted 10 lines

In Btrfs I solved (thanks to David S.' advice) by mapping only one of two pages, only the one coming from the page cache.

The other page didn't need the use of kmap_local_page() because it was allocated in the filesystem with "alloc_page(GFP_NOFS)". GFP_NOFS won't ever allocate from ZONE_HIGHMEM, therefore a direct page_address() could avoid the mapping and the nesting issues.

Did you check if you may solve the same way?

That is not simple. Currently we have compressed pages and decompressed pages (page cache or others), and they can be unmapped when either data is all consumed, so compressed pages can be unmapped first, or decompressed pages can be unmapped first. That quite depends on which pages goes first.

Quoted 6 lines

I think such usage is a quite common pattern for decoder or encoder, you could take a look at z_erofs_lzma_decompress() in fs/erofs/decompressor_lzma.c. So kmap() is still useful for such cases since I don't really care the HIGHMEM performance but correctness, but other alternative could churn/complex the map/unmap/remap pattern.

Quoted 2 lines

When you say kmap() is still useful is this because of the map/unmap ordering restrictions or because the address is required in different threads?

... mainly due to map/unmap ordering restriction. I think the decompressor here could still be a simple dependency. I'm not sure if there are more complicated cases (like multiple decoding/encoding sources into target pages) though..

Thanks,
Gao Xiang

Quoted 2 lines

Ira

#8 lore raw
In reply to Gao Xiang
Quoted 6 lines · CEST Gao Xiang
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

...

Quoted 1 line
Quoted 6 lines
Quoted 6 lines
Quoted 2 lines

One of what I need to care is nested kmap() usage,
some unmap/remap order cannot be simply converted to kmap_local()

Correct about nesting. If we map A and then B, we must unmap B and then A.

However, as you seem to convey, not always unmappings in right order

(stack

Quoted 5 lines
Quoted 5 lines

based) is possible, sometimes because very long functions have many loop's breaks and many goto exit labels.

Quoted 2 lines

but I think
it's not the case for erofs_bread(). Actually EROFS has one of such

nested

Quoted 1 line
Quoted 1 line
Quoted 1 line

kmap() usage, but I don't really care its performance on HIGHMEM

platforms,

Quoted 4 lines
Quoted 4 lines
Quoted 1 line

so I think kmap() is still somewhat useful compared to kmap_local() from

this

Quoted 1 line

point of view],

fs/erofs conversions are in our (Ira's and my) list. So I'm am happy to see that we can delete some entries because of your changes. :-)

Quoted 5 lines
Quoted 5 lines

In Btrfs I solved (thanks to David S.' advice) by mapping only one of two pages, only the one coming from the page cache.

The other page didn't need the use of kmap_local_page() because it was allocated in the filesystem with "alloc_page(GFP_NOFS)". GFP_NOFS won't

ever

Quoted 1 line
Quoted 1 line

allocate from ZONE_HIGHMEM, therefore a direct page_address() could avoid

the

Quoted 10 lines
Quoted 3 lines

mapping and the nesting issues.

Did you check if you may solve the same way?

That is not simple. Currently we have compressed pages and decompressed pages (page cache or others), and they can be unmapped when either data is all consumed, so compressed pages can be unmapped first, or decompressed pages can be unmapped first. That quite depends on which pages goes first.

Quoted 3 lines

I think such usage is a quite common pattern for decoder or encoder, you could take a look at z_erofs_lzma_decompress() in fs/erofs/decompressor_lzma.c.

I haven't yet read that code, however I may attempt to propose a pattern for solving this kinds of issue, I mean where you don't know which page got mapped last...

It's not elegant but it may work. You have compressed and decompressed pages and you can't know in advance what page should be unmapped first because you can't know in which order they where mapped, right?

I'd use a variable to save two different values, each representing the last page mapped. When the code gets to the unmapping block (perhaps in an "out" label), just check what that variable contains. Depending on that value, say 'c' or 'd', you will be able to know what must be unmapped for first. An "if / else" can do the work.

What do you think of this?

Quoted 3 lines

So kmap() is still useful for such cases
since I don't really care the HIGHMEM performance but correctness, but other alternative could churn/complex the map/unmap/remap pattern.

Sooner or later someone will have to address this issue and remove those kmap() call sites. We are working on this and hope to always figure out a way to work it out.

I hope that what I wrote above may help, although I'm writing on a purely theoretically bases, since, as said, I haven't yet seen that code.

If due to my poor English I've not been able to convey my thoughts please let me know, so that I'll try to reword.

Thanks,

Fabio

Quoted 5 lines

Thanks,
Gao Xiang

Quoted 2 lines

A little group of people are working to remove all kmap() and

kmap_atomic() we

Quoted 2 lines
Quoted 2 lines

meet across the whole kernel. I have not yet encountered conversions which cannot be made. Sometimes we may refactor, if what I said above cannot

apply.

Quoted 5 lines
Quoted 5 lines
Quoted 3 lines

but in order to make it all work properly, I will try to do stress test with 32-bit platform later.

I use QEMU/KVM x86_32 VM, 6GB RAM, and a kernel with HIGHMEM64G enabled

and an

Quoted 1 line
Quoted 1 line

openSUSE Tumbleweed 32 distro. I've heard that Debian too provides an

x86_32

Quoted 14 lines
Quoted 13 lines

distribution.

Quoted 6 lines

Since it targets for the next cycle
6.2, I will do a full stress test in the next following weeks.

Thanks,
Gao Xiang

Thanks,

Fabio

author
#9 lore raw
Quoted 52 lines · Fabio M. De Francesco
Quoted 6 lines · CEST Gao Xiang
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

...

Quoted 1 line
Quoted 6 lines
Quoted 6 lines
Quoted 2 lines

One of what I need to care is nested kmap() usage,
some unmap/remap order cannot be simply converted to kmap_local()

Correct about nesting. If we map A and then B, we must unmap B and then A.

However, as you seem to convey, not always unmappings in right order

(stack

Quoted 5 lines
Quoted 5 lines

based) is possible, sometimes because very long functions have many loop's breaks and many goto exit labels.

Quoted 2 lines

but I think
it's not the case for erofs_bread(). Actually EROFS has one of such

nested

Quoted 1 line
Quoted 1 line
Quoted 1 line

kmap() usage, but I don't really care its performance on HIGHMEM

platforms,

Quoted 4 lines
Quoted 4 lines
Quoted 1 line

so I think kmap() is still somewhat useful compared to kmap_local() from

this

Quoted 1 line

point of view],

fs/erofs conversions are in our (Ira's and my) list. So I'm am happy to see that we can delete some entries because of your changes. :-)

Quoted 5 lines
Quoted 5 lines

In Btrfs I solved (thanks to David S.' advice) by mapping only one of two pages, only the one coming from the page cache.

The other page didn't need the use of kmap_local_page() because it was allocated in the filesystem with "alloc_page(GFP_NOFS)". GFP_NOFS won't

ever

Quoted 1 line
Quoted 1 line

allocate from ZONE_HIGHMEM, therefore a direct page_address() could avoid

the

Quoted 10 lines
Quoted 3 lines

mapping and the nesting issues.

Did you check if you may solve the same way?

That is not simple. Currently we have compressed pages and decompressed pages (page cache or others), and they can be unmapped when either data is all consumed, so compressed pages can be unmapped first, or decompressed pages can be unmapped first. That quite depends on which pages goes first.

Quoted 3 lines

I think such usage is a quite common pattern for decoder or encoder, you could take a look at z_erofs_lzma_decompress() in fs/erofs/decompressor_lzma.c.

Quoted 8 lines

I haven't yet read that code, however I may attempt to propose a pattern for solving this kinds of issue, I mean where you don't know which page got mapped last...

It's not elegant but it may work. You have compressed and decompressed pages and you can't know in advance what page should be unmapped first because you can't know in which order they where mapped, right?

Not really.

Quoted 5 lines

I'd use a variable to save two different values, each representing the last page mapped. When the code gets to the unmapping block (perhaps in an "out" label), just check what that variable contains. Depending on that value, say 'c' or 'd', you will be able to know what must be unmapped for first. An "if / else" can do the work.

That is not the simple nested unmapped case as you said above, I could take a very brief example:

1. map a decompresed page
2. map a compressed page
3. working
4. decompressed page is all consumed, unmap the current decompressed page
5. map the next decompressed page
6. working
7. decompressed page is all consumed, unmap the current decompressed page
8. map the next decompressed page
9. working
10. compressed page is all consumed, unmap the current compressed page
11. map the next compressed page
12. working
13. ... (anyway, unmap and remap a compressed page or a decompressed page in any order.)

until all process is finished. by using kmap(), it's much simple to implement this, but kmap_local(), it only complexes the code.

Thanks,
Gao Xiang

Quoted 3 lines

What do you think of this?

#10 lore raw
Quoted 4 lines · Gao Xiang
Quoted 3 lines · Fabio M. De Francesco
Quoted 2 lines · CEST Gao Xiang
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

[snip]

Quoted 3 lines

That is not the simple nested unmapped case as you said above, I could take a very brief example:

Building on this. The uncompressed pages always outnumber the compressed pages, right?

Quoted 3 lines

1. map a decompresed page
2. map a compressed page

First reverse these because you are going to need to map a new decompressed page before another compressed page. So:

1. map compressed
2. map decompressed

Then 4/5 and 7/8 become unmap/map new without issue.

Quoted 7 lines

3. working
4. decompressed page is all consumed, unmap the current decompressed page
5. map the next decompressed page
6. working
7. decompressed page is all consumed, unmap the current decompressed page
8. map the next decompressed page
9. working

This is more complicated but not overly so.

Simply

9.1 unmap decompressed

Quoted 2 lines

10. compressed page is all consumed, unmap the current compressed page
11. map the next compressed page

11.1 remap decompressed

Quoted 6 lines

12. working
13. ... (anyway, unmap and remap a compressed page or a decompressed page in any order.)

until all process is finished. by using kmap(), it's much simple to implement this, but kmap_local(), it only complexes the code.

Agreed kmap() is easier but I think this could work.

Basically you keep the compressed mapped first.

I also assume there is also a reverse of this so reverse the pages in that case.

Thoughts?
Ira

author
#11 lore raw
Quoted 6 lines · Ira Weiny
Quoted 4 lines · Gao Xiang
Quoted 3 lines · Fabio M. De Francesco
Quoted 2 lines · CEST Gao Xiang
Quoted 1 line · Fabio M. De Francesco

On Tuesday, October 18, 2022 11:29:21 PM CEST Gao Xiang wrote:

Quoted 8 lines

[snip]

Quoted 3 lines

That is not the simple nested unmapped case as you said above, I could take a very brief example:

Building on this. The uncompressed pages always outnumber the compressed pages, right?

Yes, it's always true for EROFS case.

I think if the locking order is reversed I could unmap and remap the pages in the correct order. But as you said below, it just could work but complex the code (I think if you have extra time you could check the code first.)

So as I said before, I don't really care HIGHMEM performance so here kmap_local() cannot benefit such case. Anyway, it'd be much better if kmap() is kept on my side anyway.

Thanks,
Gao Xiang

Quoted 48 lines
Quoted 4 lines

1. map a decompresed page
2. map a compressed page

First reverse these because you are going to need to map a new decompressed page before another compressed page. So:

1. map compressed
2. map decompressed

Then 4/5 and 7/8 become unmap/map new without issue.

Quoted 7 lines

3. working
4. decompressed page is all consumed, unmap the current decompressed page
5. map the next decompressed page
6. working
7. decompressed page is all consumed, unmap the current decompressed page
8. map the next decompressed page
9. working

This is more complicated but not overly so.

Simply

9.1 unmap decompressed

Quoted 2 lines

10. compressed page is all consumed, unmap the current compressed page
11. map the next compressed page

11.1 remap decompressed

Quoted 6 lines

12. working
13. ... (anyway, unmap and remap a compressed page or a decompressed page in any order.)

until all process is finished. by using kmap(), it's much simple to implement this, but kmap_local(), it only complexes the code.

Agreed kmap() is easier but I think this could work.

Basically you keep the compressed mapped first.

I also assume there is also a reverse of this so reverse the pages in that case.

Thoughts?
Ira

#12 lore raw
In reply to Gao Xiang
Quoted 2 lines · Gao Xiang

Convert all mapped erofs_bread() users to use kmap_local_page() instead of kmap() or kmap_atomic().

Reviewed-and-tested-by Jingbo Xu <[email protected]>
Quoted 111 lines
---
 fs/erofs/data.c     | 8 ++------
 fs/erofs/internal.h | 3 +--
 fs/erofs/xattr.c    | 8 ++++----
 fs/erofs/zmap.c     | 4 ++--
 4 files changed, 9 insertions(+), 14 deletions(-)
fs/erofs/data.c +2 -6
index fe8ac0e163f7..fe1ae80284bf 100644
@@ -13,9 +13,7 @@
 void erofs_unmap_metabuf(struct erofs_buf *buf)
 {
 	if (buf->kmap_type == EROFS_KMAP)
-		kunmap(buf->page);
-	else if (buf->kmap_type == EROFS_KMAP_ATOMIC)
-		kunmap_atomic(buf->base);
+		kunmap_local(buf->base);
 	buf->base = NULL;
 	buf->kmap_type = EROFS_NO_KMAP;
 }
@@ -54,9 +52,7 @@ void *erofs_bread(struct erofs_buf *buf, struct inode *inode,
 	}
 	if (buf->kmap_type == EROFS_NO_KMAP) {
 		if (type == EROFS_KMAP)
-			buf->base = kmap(page);
-		else if (type == EROFS_KMAP_ATOMIC)
-			buf->base = kmap_atomic(page);
+			buf->base = kmap_local_page(page);
 		buf->kmap_type = type;
 	} else if (buf->kmap_type != type) {
 		DBG_BUGON(1);
fs/erofs/internal.h +1 -2
index 1701df48c446..67dc8e177211 100644
@@ -253,8 +253,7 @@ static inline int erofs_wait_on_workgroup_freezed(struct erofs_workgroup *grp)
 
 enum erofs_kmap_type {
 	EROFS_NO_KMAP,		/* don't map the buffer */
-	EROFS_KMAP,		/* use kmap() to map the buffer */
-	EROFS_KMAP_ATOMIC,	/* use kmap_atomic() to map the buffer */
+	EROFS_KMAP,		/* use kmap_local_page() to map the buffer */
 };
 
 struct erofs_buf {
fs/erofs/xattr.c +4 -4
index 8106bcb5a38d..a62fb8a3318a 100644
@@ -148,7 +148,7 @@ static inline int xattr_iter_fixup(struct xattr_iter *it)
 
 	it->blkaddr += erofs_blknr(it->ofs);
 	it->kaddr = erofs_read_metabuf(&it->buf, it->sb, it->blkaddr,
-				       EROFS_KMAP_ATOMIC);
+				       EROFS_KMAP);
 	if (IS_ERR(it->kaddr))
 		return PTR_ERR(it->kaddr);
 	it->ofs = erofs_blkoff(it->ofs);
@@ -174,7 +174,7 @@ static int inline_xattr_iter_begin(struct xattr_iter *it,
 	it->ofs = erofs_blkoff(iloc(sbi, vi->nid) + inline_xattr_ofs);
 
 	it->kaddr = erofs_read_metabuf(&it->buf, inode->i_sb, it->blkaddr,
-				       EROFS_KMAP_ATOMIC);
+				       EROFS_KMAP);
 	if (IS_ERR(it->kaddr))
 		return PTR_ERR(it->kaddr);
 	return vi->xattr_isize - xattr_header_sz;
@@ -368,7 +368,7 @@ static int shared_getxattr(struct inode *inode, struct getxattr_iter *it)
 
 		it->it.ofs = xattrblock_offset(sbi, vi->xattr_shared_xattrs[i]);
 		it->it.kaddr = erofs_read_metabuf(&it->it.buf, sb, blkaddr,
-						  EROFS_KMAP_ATOMIC);
+						  EROFS_KMAP);
 		if (IS_ERR(it->it.kaddr))
 			return PTR_ERR(it->it.kaddr);
 		it->it.blkaddr = blkaddr;
@@ -580,7 +580,7 @@ static int shared_listxattr(struct listxattr_iter *it)
 
 		it->it.ofs = xattrblock_offset(sbi, vi->xattr_shared_xattrs[i]);
 		it->it.kaddr = erofs_read_metabuf(&it->it.buf, sb, blkaddr,
-						  EROFS_KMAP_ATOMIC);
+						  EROFS_KMAP);
 		if (IS_ERR(it->it.kaddr))
 			return PTR_ERR(it->it.kaddr);
 		it->it.blkaddr = blkaddr;
fs/erofs/zmap.c +2 -2
index 0bb66927e3d0..749a5ac943f4 100644
@@ -178,7 +178,7 @@ static int legacy_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 	unsigned int advise, type;
 
 	m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
-				      erofs_blknr(pos), EROFS_KMAP_ATOMIC);
+				      erofs_blknr(pos), EROFS_KMAP);
 	if (IS_ERR(m->kaddr))
 		return PTR_ERR(m->kaddr);
 
@@ -416,7 +416,7 @@ static int compacted_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 out:
 	pos += lcn * (1 << amortizedshift);
 	m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
-				      erofs_blknr(pos), EROFS_KMAP_ATOMIC);
+				      erofs_blknr(pos), EROFS_KMAP);
 	if (IS_ERR(m->kaddr))
 		return PTR_ERR(m->kaddr);
 	return unpack_compacted_index(m, amortizedshift, pos, lookahead);
Signature · 2 lines
Thanks,
Jingbo