Skip to contentSkip to conversation

Prevent mmap command to map beyond EOF

PATCH In review Carlos Maiolino 2 participants 6 messages Oct 4, 2021 – Oct 7, 2021

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Prevent mmap command to map beyond EOF

Attempting to access a mmapp'ed region that does not correspond to the file results in a SIGBUS, so prevent xfs_io to even attempt to mmap() a region beyond EOF.

---

There is a caveat about this patch though. It is possible to mmap() a non-existent file region, extent the file to go beyond such region, and run operations in this mmapped region without such operations triggering a SIGBUS (excluding the file corruption factor here :). So, I'm not quite sure if it would be ok to check for this in mmap_f() as this patch does, or create a helper to check for such condition, and use it on the other operations (mread_f, mwrite_f, etc). What you folks think?

 io/mmap.c | 7 +++++++
 1 file changed, 7 insertions(+)
io/mmap.c +7 -0
index 9816cf68..77c5f2b6 100644
@@ -242,6 +242,13 @@ mmap_f(
 		return 0;
 	}
 
+	/* Check if we are mmapping beyond EOF */
+	if ((offset + length) > filesize()) {
+		printf(_("Attempting to mmap() beyond EOF\n"));
+		exitcode = 1;
+		return 0;
+	}
+
 	/*
 	 * mmap and munmap memory area of length2 region is helpful to
 	 * make a region of extendible free memory. It's generally used
Signature · 1 line
2.31.1
#1 lore raw
Quoted 7 lines · Carlos Maiolino

Attempting to access a mmapp'ed region that does not correspond to the file results in a SIGBUS, so prevent xfs_io to even attempt to mmap() a region beyond EOF.

---

Quoted 7 lines

There is a caveat about this patch though. It is possible to mmap() a non-existent file region, extent the file to go beyond such region, and run operations in this mmapped region without such operations triggering a SIGBUS (excluding the file corruption factor here :). So, I'm not quite sure if it would be ok to check for this in mmap_f() as this patch does, or create a helper to check for such condition, and use it on the other operations (mread_f, mwrite_f, etc). What you folks think?

What's the motivation for checking this in userspace? Programs are allowed to set up this (admittedly minimally functional) configuration, or even set it up after the mmap by truncating the file.

OTOH if your goal is to write a test to check the SIGBUS functionality, you could install a sigbus handler to report the signal to stderr, which would avoid bash writing junk about the sigbus to the terminal.

--D

Quoted 25 lines
 io/mmap.c | 7 +++++++
 1 file changed, 7 insertions(+)
io/mmap.c +7 -0
index 9816cf68..77c5f2b6 100644
@@ -242,6 +242,13 @@ mmap_f(
 		return 0;
 	}
 
+	/* Check if we are mmapping beyond EOF */
+	if ((offset + length) > filesize()) {
+		printf(_("Attempting to mmap() beyond EOF\n"));
+		exitcode = 1;
+		return 0;
+	}
+
 	/*
 	 * mmap and munmap memory area of length2 region is helpful to
 	 * make a region of extendible free memory. It's generally used
Signature · 1 line
2.31.1
#2 lore raw
Quoted 16 lines · Darrick J. Wong
Quoted 7 lines · Carlos Maiolino

Attempting to access a mmapp'ed region that does not correspond to the file results in a SIGBUS, so prevent xfs_io to even attempt to mmap() a region beyond EOF.

---

Quoted 7 lines

There is a caveat about this patch though. It is possible to mmap() a non-existent file region, extent the file to go beyond such region, and run operations in this mmapped region without such operations triggering a SIGBUS (excluding the file corruption factor here :). So, I'm not quite sure if it would be ok to check for this in mmap_f() as this patch does, or create a helper to check for such condition, and use it on the other operations (mread_f, mwrite_f, etc). What you folks think?

Quoted 3 lines

What's the motivation for checking this in userspace? Programs are allowed to set up this (admittedly minimally functional) configuration, or even set it up after the mmap by truncating the file.

My biggest motivation was actually seeing xfs_io crashing due a sigbus while running generic/172 and generic/173. And personally, I'd rather see an error message like "attempt to mmap/mwrite beyond EOF" than seeing it crash. Also, as you mentioned, programs are allowed to set up such kind of configuration (IIUC what you mean, mixing mmap, extend, truncate, etc), so, I believe such userspace programs should also ensure they are not attempting to write to invalid memory.

Quoted 4 lines

OTOH if your goal is to write a test to check the SIGBUS functionality, you could install a sigbus handler to report the signal to stderr, which would avoid bash writing junk about the sigbus to the terminal.

No, I'm just trying to avoid xfs_io crashing if we point it to invalid memory :)

Cheers.

Quoted 29 lines

--D

Quoted 25 lines
 io/mmap.c | 7 +++++++
 1 file changed, 7 insertions(+)
io/mmap.c +7 -0
index 9816cf68..77c5f2b6 100644
@@ -242,6 +242,13 @@ mmap_f(
 		return 0;
 	}
 
+	/* Check if we are mmapping beyond EOF */
+	if ((offset + length) > filesize()) {
+		printf(_("Attempting to mmap() beyond EOF\n"));
+		exitcode = 1;
+		return 0;
+	}
+
 	/*
 	 * mmap and munmap memory area of length2 region is helpful to
 	 * make a region of extendible free memory. It's generally used
Signature · 1 line
2.31.1
Signature · 1 line
Carlos
#3 lore raw
Quoted 21 lines · Carlos Maiolino
Quoted 16 lines · Darrick J. Wong
Quoted 7 lines · Carlos Maiolino

Attempting to access a mmapp'ed region that does not correspond to the file results in a SIGBUS, so prevent xfs_io to even attempt to mmap() a region beyond EOF.

---

Quoted 7 lines

There is a caveat about this patch though. It is possible to mmap() a non-existent file region, extent the file to go beyond such region, and run operations in this mmapped region without such operations triggering a SIGBUS (excluding the file corruption factor here :). So, I'm not quite sure if it would be ok to check for this in mmap_f() as this patch does, or create a helper to check for such condition, and use it on the other operations (mread_f, mwrite_f, etc). What you folks think?

Quoted 3 lines

What's the motivation for checking this in userspace? Programs are allowed to set up this (admittedly minimally functional) configuration, or even set it up after the mmap by truncating the file.

Quoted 7 lines

My biggest motivation was actually seeing xfs_io crashing due a sigbus while running generic/172 and generic/173. And personally, I'd rather see an error message like "attempt to mmap/mwrite beyond EOF" than seeing it crash. Also, as you mentioned, programs are allowed to set up such kind of configuration (IIUC what you mean, mixing mmap, extend, truncate, etc), so, I believe such userspace programs should also ensure they are not attempting to write to invalid memory.

This patch would /also/ prevent us from writing an fstest to check that a process /does/ get SIGBUS when writing to a mapping beyond EOF. Huh, we don't have a test for that...

Also, where does generic/173 write to a mapping beyond EOF? It sets up a file of blksz*nr_blks bytes, clones it, fills the fs to full, and then writes that number of bytes to the mmap region to trigger SIGBUS when the COW fails due to ENOSPC.

--D

Quoted 41 lines
Quoted 3 lines

OTOH if your goal is to write a test to check the SIGBUS functionality, you could install a sigbus handler to report the signal to stderr, which would avoid bash writing junk about the sigbus to the terminal.

No, I'm just trying to avoid xfs_io crashing if we point it to invalid memory :)

Cheers.

Quoted 29 lines

--D

Quoted 25 lines
 io/mmap.c | 7 +++++++
 1 file changed, 7 insertions(+)
io/mmap.c +7 -0
index 9816cf68..77c5f2b6 100644
@@ -242,6 +242,13 @@ mmap_f(
 		return 0;
 	}
 
+	/* Check if we are mmapping beyond EOF */
+	if ((offset + length) > filesize()) {
+		printf(_("Attempting to mmap() beyond EOF\n"));
+		exitcode = 1;
+		return 0;
+	}
+
 	/*
 	 * mmap and munmap memory area of length2 region is helpful to
 	 * make a region of extendible free memory. It's generally used
Signature · 1 line
2.31.1
Signature · 1 line
Carlos
#4 lore raw
Quoted 11 lines
Quoted 7 lines

My biggest motivation was actually seeing xfs_io crashing due a sigbus while running generic/172 and generic/173. And personally, I'd rather see an error message like "attempt to mmap/mwrite beyond EOF" than seeing it crash. Also, as you mentioned, programs are allowed to set up such kind of configuration (IIUC what you mean, mixing mmap, extend, truncate, etc), so, I believe such userspace programs should also ensure they are not attempting to write to invalid memory.

This patch would /also/ prevent us from writing an fstest to check that a process /does/ get SIGBUS when writing to a mapping beyond EOF. Huh, we don't have a test for that...

TBH, this kind of test in xfstests didn't pass through my mind, but, this is mmap (and its mapped regions) behavior, I don't see why we would want to test this in xfstests, but, well, I am neither a memory expert to say what it should/shouldn't do (I'm just following the man page), nor I have any authority to say what we can/can't do in xfstests :)

Quoted 5 lines

Also, where does generic/173 write to a mapping beyond EOF? It sets up a file of blksz*nr_blks bytes, clones it, fills the fs to full, and then writes that number of bytes to the mmap region to trigger SIGBUS when the COW fails due to ENOSPC.

The whole command that ends up receiving a SIGBUS is:

xfs_io -i -f -c 'mmap -rw 0 41943040' -c 'mwrite -S 0x62 0 41943040'

I have a slightly suspicious this is happening because it's trying to access the very last byte of the file, but I do need to check again the machine where I was running g/173. At a later point, I just did some tests using an empty, 0 sized file. So I do not recall from the top of my head the behavior from g/173. I can do it later, now I need to run to a doc appt.

Quoted 45 lines

--D

Quoted 41 lines
Quoted 3 lines

OTOH if your goal is to write a test to check the SIGBUS functionality, you could install a sigbus handler to report the signal to stderr, which would avoid bash writing junk about the sigbus to the terminal.

No, I'm just trying to avoid xfs_io crashing if we point it to invalid memory :)

Cheers.

Quoted 29 lines

--D

Quoted 25 lines
 io/mmap.c | 7 +++++++
 1 file changed, 7 insertions(+)
io/mmap.c +7 -0
index 9816cf68..77c5f2b6 100644
@@ -242,6 +242,13 @@ mmap_f(
 		return 0;
 	}
 
+	/* Check if we are mmapping beyond EOF */
+	if ((offset + length) > filesize()) {
+		printf(_("Attempting to mmap() beyond EOF\n"));
+		exitcode = 1;
+		return 0;
+	}
+
 	/*
 	 * mmap and munmap memory area of length2 region is helpful to
 	 * make a region of extendible free memory. It's generally used
Signature · 1 line
2.31.1
Signature · 1 line
Carlos
Signature · 1 line
Carlos
#5 lore raw
Quoted 11 lines · Carlos Maiolino
Quoted 11 lines
Quoted 7 lines

My biggest motivation was actually seeing xfs_io crashing due a sigbus while running generic/172 and generic/173. And personally, I'd rather see an error message like "attempt to mmap/mwrite beyond EOF" than seeing it crash. Also, as you mentioned, programs are allowed to set up such kind of configuration (IIUC what you mean, mixing mmap, extend, truncate, etc), so, I believe such userspace programs should also ensure they are not attempting to write to invalid memory.

This patch would /also/ prevent us from writing an fstest to check that a process /does/ get SIGBUS when writing to a mapping beyond EOF. Huh, we don't have a test for that...

After looking closer into g/173, I see what you mean now.

Quoted 4 lines

The whole command that ends up receiving a SIGBUS is:

xfs_io -i -f -c 'mmap -rw 0 41943040' -c 'mwrite -S 0x62 0 41943040'

And...

Quoted 2 lines

At a later point, I just did some tests using an empty, 0 sized file.

And playing around with a 0 sized file without coming back to look closer into g/173 was what had been blinding me.
Please just disregard this patch and my apologies for the noise.

Signature · 1 line
Carlos