Skip to contentSkip to conversation

kexec: Replace kmap() with kmap_local_page()

PATCH In review Fabio M. De Francesco 3 participants 7 messages Jul 7, 2022 – Aug 2, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 kexec: Replace kmap() with kmap_local_page()

The use of kmap() and kmap_atomic() are being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and not globally visible. Furthermore, the mappings can be acquired from any context (including interrupts).

Therefore, use kmap_local_page() in kexec_core.c because these mappings are per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

---

v1->v2: A sentence of the commit message contained an error due to a mistake in copy-pasting from a previous patch. Replace "aio.c" with "kexec_core.c".

 kernel/kexec_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/kexec_core.c +4 -4
index 4d34c78334ce..6f98274765d4 100644
@@ -814,7 +814,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 		if (result < 0)
 			goto out;
 
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		/* Start with a clear page */
 		clear_page(ptr);
 		ptr += maddr & ~PAGE_MASK;
@@ -827,7 +827,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 			memcpy(ptr, kbuf, uchunk);
 		else
 			result = copy_from_user(ptr, buf, uchunk);
-		kunmap(page);
+		kunmap_local(ptr);
 		if (result) {
 			result = -EFAULT;
 			goto out;
@@ -878,7 +878,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 			goto out;
 		}
 		arch_kexec_post_alloc_pages(page_address(page), 1, 0);
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		ptr += maddr & ~PAGE_MASK;
 		mchunk = min_t(size_t, mbytes,
 				PAGE_SIZE - (maddr & ~PAGE_MASK));
@@ -894,7 +894,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 		else
 			result = copy_from_user(ptr, buf, uchunk);
 		kexec_flush_icache_page(page);
-		kunmap(page);
+		kunmap_local(ptr);
 		arch_kexec_pre_free_pages(page_address(page), 1);
 		if (result) {
 			result = -EFAULT;
Signature · 1 line
2.36.1
Reviewed-by on v2 1/1
#1 lore raw
Quoted 7 lines · Fabio M. De Francesco

The use of kmap() and kmap_atomic() are being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and not globally visible. Furthermore, the mappings can be acquired from any context (including interrupts).

Quoted 7 lines

Therefore, use kmap_local_page() in kexec_core.c because these mappings are per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

Quoted 53 lines

---

v1->v2: A sentence of the commit message contained an error due to a mistake in copy-pasting from a previous patch. Replace "aio.c" with "kexec_core.c".

 kernel/kexec_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/kexec_core.c +4 -4
index 4d34c78334ce..6f98274765d4 100644
@@ -814,7 +814,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 		if (result < 0)
 			goto out;
 
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		/* Start with a clear page */
 		clear_page(ptr);
 		ptr += maddr & ~PAGE_MASK;
@@ -827,7 +827,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 			memcpy(ptr, kbuf, uchunk);
 		else
 			result = copy_from_user(ptr, buf, uchunk);
-		kunmap(page);
+		kunmap_local(ptr);
 		if (result) {
 			result = -EFAULT;
 			goto out;
@@ -878,7 +878,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 			goto out;
 		}
 		arch_kexec_post_alloc_pages(page_address(page), 1, 0);
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		ptr += maddr & ~PAGE_MASK;
 		mchunk = min_t(size_t, mbytes,
 				PAGE_SIZE - (maddr & ~PAGE_MASK));
@@ -894,7 +894,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 		else
 			result = copy_from_user(ptr, buf, uchunk);
 		kexec_flush_icache_page(page);
-		kunmap(page);
+		kunmap_local(ptr);
 		arch_kexec_pre_free_pages(page_address(page), 1);
 		if (result) {
 			result = -EFAULT;
Signature · 1 line
2.36.1
Reviewed-by on v2 1/1
#2 lore raw
In reply to Fabio M. De Francesco
Quoted 7 lines · Fabio M. De Francesco

The use of kmap() and kmap_atomic() are being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and not globally visible. Furthermore, the mappings can be acquired from any context (including interrupts).

Quoted 8 lines

Therefore, use kmap_local_page() in kexec_core.c because these mappings are per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

Quoted 52 lines

---

v1->v2: A sentence of the commit message contained an error due to a mistake in copy-pasting from a previous patch. Replace "aio.c" with "kexec_core.c".

 kernel/kexec_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/kexec_core.c +4 -4
index 4d34c78334ce..6f98274765d4 100644
@@ -814,7 +814,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 		if (result < 0)
 			goto out;
 
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		/* Start with a clear page */
 		clear_page(ptr);
 		ptr += maddr & ~PAGE_MASK;
@@ -827,7 +827,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 			memcpy(ptr, kbuf, uchunk);
 		else
 			result = copy_from_user(ptr, buf, uchunk);
-		kunmap(page);
+		kunmap_local(ptr);
 		if (result) {
 			result = -EFAULT;
 			goto out;
@@ -878,7 +878,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 			goto out;
 		}
 		arch_kexec_post_alloc_pages(page_address(page), 1, 0);
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		ptr += maddr & ~PAGE_MASK;
 		mchunk = min_t(size_t, mbytes,
 				PAGE_SIZE - (maddr & ~PAGE_MASK));
@@ -894,7 +894,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 		else
 			result = copy_from_user(ptr, buf, uchunk);
 		kexec_flush_icache_page(page);
-		kunmap(page);
+		kunmap_local(ptr);
 		arch_kexec_pre_free_pages(page_address(page), 1);
 		if (result) {
 			result = -EFAULT;
Signature · 1 line
2.36.1
Acked-by on v2 1/1
#3 lore raw
In reply to Fabio M. De Francesco
Quoted 7 lines · Fabio M. De Francesco

The use of kmap() and kmap_atomic() are being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and not globally visible. Furthermore, the mappings can be acquired from any context (including interrupts).

Quoted 5 lines

Therefore, use kmap_local_page() in kexec_core.c because these mappings are per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

Wondering what arch you tested with.

This looks good, but may not benefit much. Say so because I doubt how many 32bit systems are using kexec/kdump mechanism.

Anyway,

Quoted 61 lines

---

v1->v2: A sentence of the commit message contained an error due to a mistake in copy-pasting from a previous patch. Replace "aio.c" with "kexec_core.c".

 kernel/kexec_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/kexec_core.c +4 -4
index 4d34c78334ce..6f98274765d4 100644
@@ -814,7 +814,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 		if (result < 0)
 			goto out;
 
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		/* Start with a clear page */
 		clear_page(ptr);
 		ptr += maddr & ~PAGE_MASK;
@@ -827,7 +827,7 @@ static int kimage_load_normal_segment(struct kimage *image,
 			memcpy(ptr, kbuf, uchunk);
 		else
 			result = copy_from_user(ptr, buf, uchunk);
-		kunmap(page);
+		kunmap_local(ptr);
 		if (result) {
 			result = -EFAULT;
 			goto out;
@@ -878,7 +878,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 			goto out;
 		}
 		arch_kexec_post_alloc_pages(page_address(page), 1, 0);
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		ptr += maddr & ~PAGE_MASK;
 		mchunk = min_t(size_t, mbytes,
 				PAGE_SIZE - (maddr & ~PAGE_MASK));
@@ -894,7 +894,7 @@ static int kimage_load_crash_segment(struct kimage *image,
 		else
 			result = copy_from_user(ptr, buf, uchunk);
 		kexec_flush_icache_page(page);
-		kunmap(page);
+		kunmap_local(ptr);
 		arch_kexec_pre_free_pages(page_address(page), 1);
 		if (result) {
 			result = -EFAULT;
Signature · 7 lines
2.36.1


_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec

_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec

#4 lore raw
Quoted 9 lines · On martedì 2 agosto 2022 05:06:54 CEST Baoquan He wrote:
Quoted 8 lines · Fabio M. De Francesco

The use of kmap() and kmap_atomic() are being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and not globally visible. Furthermore, the mappings can be acquired from any context (including interrupts).

Therefore, use kmap_local_page() in kexec_core.c because these mappings

are

Quoted 6 lines
Quoted 4 lines

per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

Wondering what arch you tested with.

I'm sorry, I forgot to say that I use x86_32 with 4GB to 6GB RAM. This is usually an information that I add in the commit messages of all the recent conversions I'm working on across the entire kernel.

Another important information (overlooked again this time) is that (1) kmap() comes with an overhead as mapping space is restricted and protected by a global lock for synchronization and (2) it also requires global TLB invalidation when the kmap’s pool wraps and it might block when the mapping space is fully utilized until a slot becomes available.

More information about why these kmap() to kmap_local_page() conversions are needed / preferred can be found in the recent changes I made to highmem.rst. They are already in mainline since about two months.

A second round of additional changes has been taken by Andrew M. just few days ago.

My goal is to convert the most of the kmap() call sites that are still left across the whole kernel. I'm not yet sure that these kinds of conversions can be done everywhere, especially if the kernel virtual address of the mapping is handed to other contexts, because this would invalidate the pointer returned by kmap_local_page().

Quoted 2 lines

This looks good, but may not benefit much. Say so because I doubt how many 32bit systems are using kexec/kdump mechanism.

I really cannot say nothing about how many 32 bits systems are using kexec/ kdump mechanism, however I still think that the conversions are worth everywhere.

Quoted 4 lines

Anyway,

Thank you so much!

Fabio

Quoted 9 lines
Quoted 9 lines

---

v1->v2: A sentence of the commit message contained an error due to a mistake in copy-pasting from a previous patch. Replace "aio.c" with "kexec_core.c".

Quoted 8 lines

kernel/kexec_core.c static int kimage_load_normal_segment(struct kimage

Quoted 8 lines
 kernel/kexec_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/kexec_core.c +0 -0
index 4d34c78334ce..6f98274765d4 100644
@@ -814,7 +814,7 @@ static int kimage_load_normal_segment(struct kimage 

*image,

Quoted 9 lines

kernel/kexec_core.c static int kimage_load_normal_segment(struct kimage

Quoted 9 lines
kernel/kexec_core.c +1 -1
 		if (result < 0)
 			goto out;
 
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		/* Start with a clear page */
 		clear_page(ptr);
 		ptr += maddr & ~PAGE_MASK;
@@ -827,7 +827,7 @@ static int kimage_load_normal_segment(struct kimage 

*image,

Quoted 9 lines

kernel/kexec_core.c static int kimage_load_crash_segment(struct kimage

Quoted 9 lines
kernel/kexec_core.c +1 -1
 			memcpy(ptr, kbuf, uchunk);
 		else
 			result = copy_from_user(ptr, buf, uchunk);
-		kunmap(page);
+		kunmap_local(ptr);
 		if (result) {
 			result = -EFAULT;
 			goto out;
@@ -878,7 +878,7 @@ static int kimage_load_crash_segment(struct kimage 

*image,

Quoted 9 lines

kernel/kexec_core.c static int kimage_load_crash_segment(struct kimage

Quoted 9 lines
kernel/kexec_core.c +1 -1
 			goto out;
 		}
 		arch_kexec_post_alloc_pages(page_address(page), 1, 0);
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		ptr += maddr & ~PAGE_MASK;
 		mchunk = min_t(size_t, mbytes,
 				PAGE_SIZE - (maddr & ~PAGE_MASK));
@@ -894,7 +894,7 @@ static int kimage_load_crash_segment(struct kimage 

*image,

Quoted 19 lines
Quoted 17 lines
kernel/kexec_core.c +1 -1
 		else
 			result = copy_from_user(ptr, buf, uchunk);
 		kexec_flush_icache_page(page);
-		kunmap(page);
+		kunmap_local(ptr);
 		arch_kexec_pre_free_pages(page_address(page), 1);
 		if (result) {
 			result = -EFAULT;
Signature · 7 lines
2.36.1


_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec
#5 lore raw
Quoted 22 lines · Fabio M. De Francesco
Quoted 9 lines · On martedì 2 agosto 2022 05:06:54 CEST Baoquan He wrote:
Quoted 8 lines · Fabio M. De Francesco

The use of kmap() and kmap_atomic() are being deprecated in favor of kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and not globally visible. Furthermore, the mappings can be acquired from any context (including interrupts).

Therefore, use kmap_local_page() in kexec_core.c because these mappings

are

Quoted 6 lines
Quoted 4 lines

per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

Wondering what arch you tested with.

I'm sorry, I forgot to say that I use x86_32 with 4GB to 6GB RAM. This is usually an information that I add in the commit messages of all the recent conversions I'm working on across the entire kernel.

Quoted 5 lines

Another important information (overlooked again this time) is that (1) kmap() comes with an overhead as mapping space is restricted and protected by a global lock for synchronization and (2) it also requires global TLB invalidation when the kmap’s pool wraps and it might block when the mapping space is fully utilized until a slot becomes available.

Thanks a lot for update with more details, Fabio. Maybe these can be added into log with v3 if you think they are worth adding, and with my Ack since no code change related. You decide.

Thanks again for the effort.

Quoted 14 lines

More information about why these kmap() to kmap_local_page() conversions are needed / preferred can be found in the recent changes I made to highmem.rst. They are already in mainline since about two months.

A second round of additional changes has been taken by Andrew M. just few days ago.

My goal is to convert the most of the kmap() call sites that are still left across the whole kernel. I'm not yet sure that these kinds of conversions can be done everywhere, especially if the kernel virtual address of the mapping is handed to other contexts, because this would invalidate the pointer returned by kmap_local_page().

Quoted 6 lines
Quoted 2 lines

This looks good, but may not benefit much. Say so because I doubt how many 32bit systems are using kexec/kdump mechanism.

I really cannot say nothing about how many 32 bits systems are using kexec/ kdump mechanism, however I still think that the conversions are worth everywhere.

Yes, agree.

Quoted 81 lines
Quoted 5 lines

Anyway,

Thank you so much!

Fabio

Quoted 9 lines
Quoted 9 lines

---

v1->v2: A sentence of the commit message contained an error due to a mistake in copy-pasting from a previous patch. Replace "aio.c" with "kexec_core.c".

Quoted 8 lines

kernel/kexec_core.c static int kimage_load_normal_segment(struct kimage

Quoted 8 lines
 kernel/kexec_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/kexec_core.c +0 -0
index 4d34c78334ce..6f98274765d4 100644
@@ -814,7 +814,7 @@ static int kimage_load_normal_segment(struct kimage 

*image,

Quoted 9 lines

kernel/kexec_core.c static int kimage_load_normal_segment(struct kimage

Quoted 9 lines
kernel/kexec_core.c +1 -1
 		if (result < 0)
 			goto out;
 
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		/* Start with a clear page */
 		clear_page(ptr);
 		ptr += maddr & ~PAGE_MASK;
@@ -827,7 +827,7 @@ static int kimage_load_normal_segment(struct kimage 

*image,

Quoted 9 lines

kernel/kexec_core.c static int kimage_load_crash_segment(struct kimage

Quoted 9 lines
kernel/kexec_core.c +1 -1
 			memcpy(ptr, kbuf, uchunk);
 		else
 			result = copy_from_user(ptr, buf, uchunk);
-		kunmap(page);
+		kunmap_local(ptr);
 		if (result) {
 			result = -EFAULT;
 			goto out;
@@ -878,7 +878,7 @@ static int kimage_load_crash_segment(struct kimage 

*image,

Quoted 9 lines

kernel/kexec_core.c static int kimage_load_crash_segment(struct kimage

Quoted 9 lines
kernel/kexec_core.c +1 -1
 			goto out;
 		}
 		arch_kexec_post_alloc_pages(page_address(page), 1, 0);
-		ptr = kmap(page);
+		ptr = kmap_local_page(page);
 		ptr += maddr & ~PAGE_MASK;
 		mchunk = min_t(size_t, mbytes,
 				PAGE_SIZE - (maddr & ~PAGE_MASK));
@@ -894,7 +894,7 @@ static int kimage_load_crash_segment(struct kimage 

*image,

Quoted 19 lines
Quoted 17 lines
kernel/kexec_core.c +1 -1
 		else
 			result = copy_from_user(ptr, buf, uchunk);
 		kexec_flush_icache_page(page);
-		kunmap(page);
+		kunmap_local(ptr);
 		arch_kexec_pre_free_pages(page_address(page), 1);
 		if (result) {
 			result = -EFAULT;
Signature · 7 lines
2.36.1


_______________________________________________
kexec mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kexec
#6 lore raw
Quoted 4 lines · On martedì 2 agosto 2022 09:32:46 CEST Baoquan He wrote:
Quoted 3 lines · Fabio M. De Francesco
Quoted 2 lines · On martedì 2 agosto 2022 05:06:54 CEST Baoquan He wrote:
Quoted 1 line · Fabio M. De Francesco

The use of kmap() and kmap_atomic() are being deprecated in favor

of

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

kmap_local_page().

With kmap_local_page(), the mappings are per thread, CPU local and

not

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

globally visible. Furthermore, the mappings can be acquired from

any

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

context (including interrupts).

Therefore, use kmap_local_page() in kexec_core.c because these

mappings

Quoted 10 lines
Quoted 10 lines

are

Quoted 6 lines
Quoted 4 lines

per thread, CPU local, and not globally visible.

Tested on a QEMU + KVM 32-bits VM booting a kernel with HIGHMEM64GB enabled.

Wondering what arch you tested with.

I'm sorry, I forgot to say that I use x86_32 with 4GB to 6GB RAM. This is usually an information that I add in the commit messages of all

the

Quoted 4 lines
Quoted 4 lines

recent conversions I'm working on across the entire kernel.

Another important information (overlooked again this time) is that (1) kmap() comes with an overhead as mapping space is restricted and

protected

Quoted 1 line
Quoted 1 line

by a global lock for synchronization and (2) it also requires global

TLB

Quoted 1 line
Quoted 1 line

invalidation when the kmap’s pool wraps and it might block when the

mapping

Quoted 3 lines
Quoted 1 line

space is fully utilized until a slot becomes available.

Thanks a lot for update with more details, Fabio.

You're welcome.

Quoted 3 lines

Maybe these can be
added into log with v3 if you think they are worth adding, and with my Ack since no code change related. You decide.

For weeks I've been thinking that these details were not necessary in the commit message of these conversions. Later I realized that a fair number of maintainers weren't aware of why we should change something that had been working "properly" for years. Perhaps they thought that we would face too many risks of breaking things versus little rewards.

After discussing with Greg K-H (who was initially not fond of these changes to core parts of firmware loading mechanism), he suggested that a proper commit log with the necessary information could have helped himself and the other maintainers who may not yet had time to deepen this topics.

This patch to "kexec" slipped in with no such additions.

Therefore, I agree with you in full: I'll add these information and send a new version ASAP.

Quoted 1 line

Thanks again for the effort.

Thanks so much for the "Acked-by:" tag from you.

Fabio

[snip]