Skip to contentSkip to conversation

fs/proc/kcore: reinstate bounce buffer for KCORE_TEXT regions

PATCH In review Lorenzo Stoakes 5 participants 11 messages Jul 31, 2023 – Aug 1, 2023

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs/proc/kcore: reinstate bounce buffer for KCORE_TEXT regions

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
 fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +25 -1
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
 
 static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
 	loff_t *fpos = &iocb->ki_pos;
 	size_t phdrs_offset, notes_offset, data_offset;
 	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 			fallthrough;
 		case KCORE_VMEMMAP:
 		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
+				}
 			/*
 			 * We use _copy_to_iter() to bypass usermode hardening
 			 * which would otherwise prevent this operation.
 			 */
-			if (_copy_to_iter((char *)start, tsz, iter) != tsz) {
+			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
 				ret = -EFAULT;
 				goto out;
 			}
@@ -595,6 +608,10 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	if (ret)
 		return ret;
 
+	filp->private_data = kmalloc(PAGE_SIZE, GFP_KERNEL);
+	if (!filp->private_data)
+		return -ENOMEM;
+
 	if (kcore_need_update)
 		kcore_update_ram();
 	if (i_size_read(inode) != proc_root_kcore->size) {
@@ -605,9 +622,16 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	return 0;
 }
 
+static int release_kcore(struct inode *inode, struct file *file)
+{
+	kfree(file->private_data);
+	return 0;
+}
+
 static const struct proc_ops kcore_proc_ops = {
 	.proc_read_iter	= read_kcore_iter,
 	.proc_open	= open_kcore,
+	.proc_release	= release_kcore,
 	.proc_lseek	= default_llseek,
 };
 
Signature · 1 line
2.41.0
Tested-by on v1 1/1
#1 lore raw
Quoted 9 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

Quoted 6 lines

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")

it fixed my issue, thanks

jirka

Quoted 75 lines
---
 fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +25 -1
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
 
 static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
 	loff_t *fpos = &iocb->ki_pos;
 	size_t phdrs_offset, notes_offset, data_offset;
 	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 			fallthrough;
 		case KCORE_VMEMMAP:
 		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
+				}
 			/*
 			 * We use _copy_to_iter() to bypass usermode hardening
 			 * which would otherwise prevent this operation.
 			 */
-			if (_copy_to_iter((char *)start, tsz, iter) != tsz) {
+			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
 				ret = -EFAULT;
 				goto out;
 			}
@@ -595,6 +608,10 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	if (ret)
 		return ret;
 
+	filp->private_data = kmalloc(PAGE_SIZE, GFP_KERNEL);
+	if (!filp->private_data)
+		return -ENOMEM;
+
 	if (kcore_need_update)
 		kcore_update_ram();
 	if (i_size_read(inode) != proc_root_kcore->size) {
@@ -605,9 +622,16 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	return 0;
 }
 
+static int release_kcore(struct inode *inode, struct file *file)
+{
+	kfree(file->private_data);
+	return 0;
+}
+
 static const struct proc_ops kcore_proc_ops = {
 	.proc_read_iter	= read_kcore_iter,
 	.proc_open	= open_kcore,
+	.proc_release	= release_kcore,
 	.proc_lseek	= default_llseek,
 };
 
Signature · 1 line
2.41.0
Tested-by on v1 1/1
#2 lore raw
In reply to Lorenzo Stoakes
Quoted 13 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Quoted 8 lines
Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
 fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)

I can confirm this fixes the arm64 issue reported by Mike over at [1].

Cheers,

Will

[1] https://lore.kernel.org/r/[email protected]

#3 lore raw
In reply to Lorenzo Stoakes
Quoted 44 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
  fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
  1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +7 -0
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
 
  static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
  {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
  	loff_t *fpos = &iocb->ki_pos;
  	size_t phdrs_offset, notes_offset, data_offset;
  	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
  			fallthrough;
  		case KCORE_VMEMMAP:
  		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.

fs/proc/kcore.c static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter) · line 573

Quoted 10 lines
fs/proc/kcore.c +6 -0
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
+				}
  			/*
  			 * We use _copy_to_iter() to bypass usermode hardening
  			 * which would otherwise prevent this operation.
  			 */

Having a comment at this indentation level looks for the else case looks kind of weird.

(does that comment still apply?)

Signature · 3 lines
Cheers,

David / dhildenb
#4 lore raw
Quoted 49 lines · David Hildenbrand
Quoted 48 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
  fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
  1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +12 -0
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
  static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
  {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
  	loff_t *fpos = &iocb->ki_pos;
  	size_t phdrs_offset, notes_offset, data_offset;
  	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
  			fallthrough;
  		case KCORE_VMEMMAP:
  		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
Quoted 8 lines

fs/proc/kcore.c static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter) · line 573

Quoted 5 lines
fs/proc/kcore.c +1 -0
+				}
  			/*
  			 * We use _copy_to_iter() to bypass usermode hardening
  			 * which would otherwise prevent this operation.
  			 */

Having a comment at this indentation level looks for the else case looks kind of weird.

Yeah, but having it indented again would be weird and seem like it doesn't apply to the block below, there's really no good spot for it and checkpatch.pl doesn't mind so I think this is ok :)

Quoted 2 lines

(does that comment still apply?)

Hm good point, actually, now we're using the bounce buffer we don't need to avoid usermode hardening any more.

However since we've established a bounce buffer ourselves its still appropriate to use _copy_to_iter() as we know the source region is good to copy from.

To make life easy I'll just respin with an updated comment :)

Quoted 7 lines

--
Cheers,

David / dhildenb

#5 lore raw
Quoted 66 lines · Lorenzo Stoakes
Quoted 49 lines · David Hildenbrand
Quoted 48 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
   fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
   1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +12 -0
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
   static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
   {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
   	loff_t *fpos = &iocb->ki_pos;
   	size_t phdrs_offset, notes_offset, data_offset;
   	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
   			fallthrough;
   		case KCORE_VMEMMAP:
   		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
Quoted 8 lines

fs/proc/kcore.c static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter) · line 573

Quoted 5 lines
fs/proc/kcore.c +1 -0
+				}
   			/*
   			 * We use _copy_to_iter() to bypass usermode hardening
   			 * which would otherwise prevent this operation.
   			 */

Having a comment at this indentation level looks for the else case looks kind of weird.

Yeah, but having it indented again would be weird and seem like it doesn't apply to the block below, there's really no good spot for it and checkpatch.pl doesn't mind so I think this is ok :)

Quoted 2 lines

(does that comment still apply?)

Quoted 8 lines

Hm good point, actually, now we're using the bounce buffer we don't need to avoid usermode hardening any more.

However since we've established a bounce buffer ourselves its still appropriate to use _copy_to_iter() as we know the source region is good to copy from.

To make life easy I'll just respin with an updated comment :)

I'm not too picky this time, no need to resend if everybody else is fine :P

Signature · 3 lines
Cheers,

David / dhildenb
#6 lore raw
Quoted 70 lines · David Hildenbrand
Quoted 69 lines · Lorenzo Stoakes
Quoted 49 lines · David Hildenbrand
Quoted 48 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
   fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
   1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +12 -0
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
   static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
   {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
   	loff_t *fpos = &iocb->ki_pos;
   	size_t phdrs_offset, notes_offset, data_offset;
   	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
   			fallthrough;
   		case KCORE_VMEMMAP:
   		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
Quoted 8 lines

fs/proc/kcore.c static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter) · line 573

Quoted 5 lines
fs/proc/kcore.c +1 -0
+				}
   			/*
   			 * We use _copy_to_iter() to bypass usermode hardening
   			 * which would otherwise prevent this operation.
   			 */

Having a comment at this indentation level looks for the else case looks kind of weird.

Yeah, but having it indented again would be weird and seem like it doesn't apply to the block below, there's really no good spot for it and checkpatch.pl doesn't mind so I think this is ok :)

Quoted 2 lines

(does that comment still apply?)

Hm good point, actually, now we're using the bounce buffer we don't need to avoid usermode hardening any more.

Quoted 8 lines
Quoted 5 lines

However since we've established a bounce buffer ourselves its still appropriate to use _copy_to_iter() as we know the source region is good to copy from.

To make life easy I'll just respin with an updated comment :)

I'm not too picky this time, no need to resend if everybody else is fine :P

Haha you know the classic Lorenzo respin spiral and want to avoid it I see ;)

The comment is actually inaccurate now, so to avoid noise + make life easy (maybe) for Andrew here's a fix patch that just corrects the comment:-

----8<----

From d2b8fb271f21b79048e5630699133f77a93d0481 Mon Sep 17 00:00:00 2001
From: Lorenzo Stoakes <[email protected]>
Date: Tue, 1 Aug 2023 17:36:08 +0100
Subject: [PATCH] fs/proc/kcore: correct comment

Correct comment to be strictly correct about reasoning.

---
 fs/proc/kcore.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
fs/proc/kcore.c +2 -2
index 3bc689038232..23fc24d16b31 100644
@@ -568,8 +568,8 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 					goto out;
 				}
 			/*
-			 * We use _copy_to_iter() to bypass usermode hardening
-			 * which would otherwise prevent this operation.
+			 * We know the bounce buffer is safe to copy from, so
+			 * use _copy_to_iter() directly.
 			 */
 			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
 				ret = -EFAULT;

--
2.41.0

#7 lore raw
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

Hm good point, actually, now we're using the bounce buffer we don't need to avoid usermode hardening any more.

However since we've established a bounce buffer ourselves its still appropriate to use _copy_to_iter() as we know the source region is good to copy from.

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

To make life easy I'll just respin with an updated comment :)

I'm not too picky this time, no need to resend if everybody else is fine :P

Haha you know the classic Lorenzo respin spiral and want to avoid it I see ;)

Don't want to make your apparently stressful week more stressful. Not this time ;)

Quoted 34 lines

The comment is actually inaccurate now, so to avoid noise + make life easy (maybe) for Andrew here's a fix patch that just corrects the comment:-

----8<----

From d2b8fb271f21b79048e5630699133f77a93d0481 Mon Sep 17 00:00:00 2001
From: Lorenzo Stoakes <[email protected]>
Date: Tue, 1 Aug 2023 17:36:08 +0100
Subject: [PATCH] fs/proc/kcore: correct comment

Correct comment to be strictly correct about reasoning.

---
  fs/proc/kcore.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
fs/proc/kcore.c +2 -3
index 3bc689038232..23fc24d16b31 100644
@@ -568,8 +568,8 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
  					goto out;
  				}
  			/*
-			 * We use _copy_to_iter() to bypass usermode hardening
-			 * which would otherwise prevent this operation.
+			 * We know the bounce buffer is safe to copy from, so
+			 * use _copy_to_iter() directly.
  			 */
  			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
  				ret = -EFAULT;
--
Quoted 2 lines

2.41.0

Thanks!

Signature · 3 lines
Cheers,

David / dhildenb
Tested-by Reviewed-by on v1 1/1
#8 lore raw
In reply to Lorenzo Stoakes
Quoted 76 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
 fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +20 -1
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
 
 static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
 	loff_t *fpos = &iocb->ki_pos;
 	size_t phdrs_offset, notes_offset, data_offset;
 	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 			fallthrough;
 		case KCORE_VMEMMAP:
 		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
+				}
 			/*
 			 * We use _copy_to_iter() to bypass usermode hardening
 			 * which would otherwise prevent this operation.
 			 */
-			if (_copy_to_iter((char *)start, tsz, iter) != tsz) {
+			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
 				ret = -EFAULT;
 				goto out;
 			}
@@ -595,6 +608,10 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	if (ret)
 		return ret;
 
+	filp->private_data = kmalloc(PAGE_SIZE, GFP_KERNEL);
+	if (!filp->private_data)
+		return -ENOMEM;
+
 	if (kcore_need_update)
 		kcore_update_ram();
 	if (i_size_read(inode) != proc_root_kcore->size) {
@@ -605,9 +622,16 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	return 0;
 }
 
+static int release_kcore(struct inode *inode, struct file *file)
+{

fs/proc/kcore.c static int open_kcore(struct inode *inode, struct file *filp) · line 636

Quoted 10 lines
fs/proc/kcore.c +5 -0
+	kfree(file->private_data);
+	return 0;
+}
+
 static const struct proc_ops kcore_proc_ops = {
 	.proc_read_iter	= read_kcore_iter,
 	.proc_open	= open_kcore,
+	.proc_release	= release_kcore,
 	.proc_lseek	= default_llseek,
 };

On 6.5-rc4, the failures can be reproduced stably on a arm64 machine. With patch applied, both makedumpfile and objdump test cases passed.

And the code change looks good to me, thanks.

===============================================
[root@ ~]# makedumpfile --mem-usage /proc/kcore
The kernel version is not supported.
The makedumpfile operation may be incomplete.

TYPE PAGES EXCLUDABLE DESCRIPTION
----------------------------------------------------------------------
ZERO 76234 yes Pages filled with zero
NON_PRI_CACHE 147613 yes Cache pages without private flag
PRI_CACHE 3847 yes Cache pages with private flag
USER 15276 yes User process pages
FREE 15809884 yes Free pages
KERN_DATA 459950 no Dumpable kernel data

page size: 4096
Total pages on system: 16512804
Total size on system: 67636445184 Byte

[root@ ~]# objdump -d --start-address=0x^C
[root@ ~]# cat /proc/kallsyms | grep ksys_read
ffffab3be77229d8 T ksys_readahead
ffffab3be782a700 T ksys_read
[root@ ~]# objdump -d --start-address=0xffffab3be782a700 --stop-address=0xffffab3be782a710 /proc/kcore

/proc/kcore: file format elf64-littleaarch64

Disassembly of section load1:

ffffab3be782a700 <load1+0x41a700>:
ffffab3be782a700: aa1e03e9 mov x9, x30
ffffab3be782a704: d503201f nop
ffffab3be782a708: d503233f paciasp
ffffab3be782a70c: a9bc7bfd stp x29, x30, [sp, #-64]!
objdump: error: /proc/kcore(load2) is too large (0x7bff70000000 bytes)
objdump: Reading section load2 failed because: memory exhausted

#9 lore raw
Quoted 126 lines · Baoquan He
Quoted 76 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
 fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +20 -1
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
 
 static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
 	loff_t *fpos = &iocb->ki_pos;
 	size_t phdrs_offset, notes_offset, data_offset;
 	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 			fallthrough;
 		case KCORE_VMEMMAP:
 		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
+				}
 			/*
 			 * We use _copy_to_iter() to bypass usermode hardening
 			 * which would otherwise prevent this operation.
 			 */
-			if (_copy_to_iter((char *)start, tsz, iter) != tsz) {
+			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
 				ret = -EFAULT;
 				goto out;
 			}
@@ -595,6 +608,10 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	if (ret)
 		return ret;
 
+	filp->private_data = kmalloc(PAGE_SIZE, GFP_KERNEL);
+	if (!filp->private_data)
+		return -ENOMEM;
+
 	if (kcore_need_update)
 		kcore_update_ram();
 	if (i_size_read(inode) != proc_root_kcore->size) {
@@ -605,9 +622,16 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	return 0;
 }
 
+static int release_kcore(struct inode *inode, struct file *file)
+{

fs/proc/kcore.c static int open_kcore(struct inode *inode, struct file *filp) · line 636

Quoted 10 lines
fs/proc/kcore.c +5 -0
+	kfree(file->private_data);
+	return 0;
+}
+
 static const struct proc_ops kcore_proc_ops = {
 	.proc_read_iter	= read_kcore_iter,
 	.proc_open	= open_kcore,
+	.proc_release	= release_kcore,
 	.proc_lseek	= default_llseek,
 };

On 6.5-rc4, the failures can be reproduced stably on a arm64 machine. With patch applied, both makedumpfile and objdump test cases passed.

And the code change looks good to me, thanks.

===============================================
[root@ ~]# makedumpfile --mem-usage /proc/kcore
The kernel version is not supported.
The makedumpfile operation may be incomplete.

TYPE PAGES EXCLUDABLE DESCRIPTION
----------------------------------------------------------------------
ZERO 76234 yes Pages filled with zero
NON_PRI_CACHE 147613 yes Cache pages without private flag
PRI_CACHE 3847 yes Cache pages with private flag
USER 15276 yes User process pages
FREE 15809884 yes Free pages
KERN_DATA 459950 no Dumpable kernel data

page size: 4096
Total pages on system: 16512804
Total size on system: 67636445184 Byte

[root@ ~]# objdump -d --start-address=0x^C
[root@ ~]# cat /proc/kallsyms | grep ksys_read
ffffab3be77229d8 T ksys_readahead
ffffab3be782a700 T ksys_read
[root@ ~]# objdump -d --start-address=0xffffab3be782a700 --stop-address=0xffffab3be782a710 /proc/kcore

/proc/kcore: file format elf64-littleaarch64

Disassembly of section load1:

Quoted 7 lines

ffffab3be782a700 <load1+0x41a700>:
ffffab3be782a700: aa1e03e9 mov x9, x30
ffffab3be782a704: d503201f nop
ffffab3be782a708: d503233f paciasp
ffffab3be782a70c: a9bc7bfd stp x29, x30, [sp, #-64]!
objdump: error: /proc/kcore(load2) is too large (0x7bff70000000 bytes)
objdump: Reading section load2 failed because: memory exhausted

By the way, I can still see the objdump error saying kcore is too large as above, at the same time there's console printing as below. Haven't checked it's objdump's issue or kernel's.

[ 6631.575800] __vm_enough_memory: pid: 5321, comm: objdump, not enough memory for the allocation
[ 6631.584469] __vm_enough_memory: pid: 5321, comm: objdump, not enough memory for the allocation
#10 lore raw
Quoted 92 lines · Baoquan He
Quoted 91 lines · Baoquan He
Quoted 76 lines · Lorenzo Stoakes

Some architectures do not populate the entire range categorised by KCORE_TEXT, so we must ensure that the kernel address we read from is valid.

Unfortunately there is no solution currently available to do so with a purely iterator solution so reinstate the bounce buffer in this instance so we can use copy_from_kernel_nofault() in order to avoid page faults when regions are unmapped.

This change partly reverts commit 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data"), reinstating the bounce buffer, but adapts the code to continue to use an iterator.

Fixes 2e1c0170771e ("fs/proc/kcore: avoid bounce buffer for ktext data")
---
 fs/proc/kcore.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
fs/proc/kcore.c +20 -1
index 9cb32e1a78a0..3bc689038232 100644
@@ -309,6 +309,8 @@ static void append_kcore_note(char *notes, size_t *i, const char *name,
 
 static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 {
+	struct file *file = iocb->ki_filp;
+	char *buf = file->private_data;
 	loff_t *fpos = &iocb->ki_pos;
 	size_t phdrs_offset, notes_offset, data_offset;
 	size_t page_offline_frozen = 1;
@@ -554,11 +556,22 @@ static ssize_t read_kcore_iter(struct kiocb *iocb, struct iov_iter *iter)
 			fallthrough;
 		case KCORE_VMEMMAP:
 		case KCORE_TEXT:
+			/*
+			 * Sadly we must use a bounce buffer here to be able to
+			 * make use of copy_from_kernel_nofault(), as these
+			 * memory regions might not always be mapped on all
+			 * architectures.
+			 */
+			if (copy_from_kernel_nofault(buf, (void *)start, tsz)) {
+				if (iov_iter_zero(tsz, iter) != tsz) {
+					ret = -EFAULT;
+					goto out;
+				}
 			/*
 			 * We use _copy_to_iter() to bypass usermode hardening
 			 * which would otherwise prevent this operation.
 			 */
-			if (_copy_to_iter((char *)start, tsz, iter) != tsz) {
+			} else if (_copy_to_iter(buf, tsz, iter) != tsz) {
 				ret = -EFAULT;
 				goto out;
 			}
@@ -595,6 +608,10 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	if (ret)
 		return ret;
 
+	filp->private_data = kmalloc(PAGE_SIZE, GFP_KERNEL);
+	if (!filp->private_data)
+		return -ENOMEM;
+
 	if (kcore_need_update)
 		kcore_update_ram();
 	if (i_size_read(inode) != proc_root_kcore->size) {
@@ -605,9 +622,16 @@ static int open_kcore(struct inode *inode, struct file *filp)
 	return 0;
 }
 
+static int release_kcore(struct inode *inode, struct file *file)
+{

fs/proc/kcore.c static int open_kcore(struct inode *inode, struct file *filp) · line 636

Quoted 10 lines
fs/proc/kcore.c +5 -0
+	kfree(file->private_data);
+	return 0;
+}
+
 static const struct proc_ops kcore_proc_ops = {
 	.proc_read_iter	= read_kcore_iter,
 	.proc_open	= open_kcore,
+	.proc_release	= release_kcore,
 	.proc_lseek	= default_llseek,
 };

On 6.5-rc4, the failures can be reproduced stably on a arm64 machine. With patch applied, both makedumpfile and objdump test cases passed.

Quoted 4 lines
Quoted 4 lines

And the code change looks good to me, thanks.

Thanks!

Quoted 39 lines
Quoted 31 lines

===============================================
[root@ ~]# makedumpfile --mem-usage /proc/kcore
The kernel version is not supported.
The makedumpfile operation may be incomplete.

TYPE PAGES EXCLUDABLE DESCRIPTION
----------------------------------------------------------------------
ZERO 76234 yes Pages filled with zero
NON_PRI_CACHE 147613 yes Cache pages without private flag
PRI_CACHE 3847 yes Cache pages with private flag
USER 15276 yes User process pages
FREE 15809884 yes Free pages
KERN_DATA 459950 no Dumpable kernel data

page size: 4096
Total pages on system: 16512804
Total size on system: 67636445184 Byte

[root@ ~]# objdump -d --start-address=0x^C
[root@ ~]# cat /proc/kallsyms | grep ksys_read
ffffab3be77229d8 T ksys_readahead
ffffab3be782a700 T ksys_read
[root@ ~]# objdump -d --start-address=0xffffab3be782a700 --stop-address=0xffffab3be782a710 /proc/kcore

/proc/kcore: file format elf64-littleaarch64

Disassembly of section load1:

Quoted 7 lines

ffffab3be782a700 <load1+0x41a700>:
ffffab3be782a700: aa1e03e9 mov x9, x30
ffffab3be782a704: d503201f nop
ffffab3be782a708: d503233f paciasp
ffffab3be782a70c: a9bc7bfd stp x29, x30, [sp, #-64]!
objdump: error: /proc/kcore(load2) is too large (0x7bff70000000 bytes)
objdump: Reading section load2 failed because: memory exhausted

Quoted 7 lines

By the way, I can still see the objdump error saying kcore is too large as above, at the same time there's console printing as below. Haven't checked it's objdump's issue or kernel's.

[ 6631.575800] __vm_enough_memory: pid: 5321, comm: objdump, not enough memory for the allocation
[ 6631.584469] __vm_enough_memory: pid: 5321, comm: objdump, not enough memory for the allocation

Yeah this issue existed before this patch was applied on arm64, apparently an ancient objdump bug according to the other thread [0]. I confirmed it exists on v6.0 kernel for instance.

[0]:https://lore.kernel.org/all/[email protected]/