lib/hexdump: introduce DUMP_PREFIX_UNHASHED for unhashed addresses
Subsystems (2)
PATCH 1/1 lib/hexdump: introduce DUMP_PREFIX_UNHASHED for unhashed addresses
Hashed addresses are useless in hexdumps unless you're comparing with other hashed addresses, which is unlikely. However, there's no need to break existing code, so introduce a new prefix type that prints unhashed addresses.
--- include/linux/printk.h | 3 ++- lib/hexdump.c | 9 +++++++-- 2 files changed, 9 insertions(+), 3 deletions(-)
include/linux/printk.h +2 -1
@@ -567,7 +567,8 @@ extern const struct file_operations kmsg_fops;enum {DUMP_PREFIX_NONE ,DUMP_PREFIX_ADDRESS , -DUMP_PREFIX_OFFSET +DUMP_PREFIX_OFFSET , +DUMP_PREFIX_UNHASHED , }; extern int hex_dump_to_buffer(const void *buf, size_t len, int rowsize, int groupsize, char *linebuf, size_t linebuflen,
lib/hexdump.c +7 -2
@@ -211,8 +211,9 @@ EXPORT_SYMBOL(hex_dump_to_buffer);* @level: kernel log level (e.g. KERN_DEBUG) * @prefix_str: string to prefix each line with; * caller supplies trailing spaces for alignment if desired -* @prefix_type: controls whether prefix of an offset, address, or none -* is printed (%DUMP_PREFIX_OFFSET, %DUMP_PREFIX_ADDRESS, %DUMP_PREFIX_NONE) +* @prefix_type: controls whether prefix of an offset, hashed address, +* unhashed address, or none is printed (%DUMP_PREFIX_OFFSET, +* %DUMP_PREFIX_ADDRESS, %DUMP_PREFIX_UNHASHED, %DUMP_PREFIX_NONE) * @rowsize: number of bytes to print per line; must be 16 or 32 * @groupsize: number of bytes to print at a time (1, 2, 4, 8; default = 1) * @buf: data blob to dump
@@ -256,6 +257,10 @@ void print_hex_dump(const char *level, const char *prefix_str, int prefix_type, linebuf, sizeof(linebuf), ascii);switch (prefix_type) { +case DUMP_PREFIX_UNHASHED : +printk ("%s%s%px: %s\n" , + level, prefix_str, ptr + i, linebuf); +break ;case DUMP_PREFIX_ADDRESS :printk ("%s%s%p: %s\n" , level, prefix_str, ptr + i, linebuf);
Signature · 1 line
2.25.1
Adding Kees into CC because it is security related. Adding Andrew into CC because he usually takes patches for hexdump.
Quoted 7 lines · Timur Tabi
Hashed addresses are useless in hexdumps unless you're comparing with other hashed addresses, which is unlikely. However, there's no need to break existing code, so introduce a new prefix type that prints unhashed addresses.
Signed-off-by Timur Tabi <[email protected]>Cc Roman Fietze <[email protected]>
I agree that there should be way to print the real address.
And it is sane to add a new mode so that the current users stay hashed.
Quoted 48 lines
--- include/linux/printk.h | 3 ++- lib/hexdump.c | 9 +++++++-- 2 files changed, 9 insertions(+), 3 deletions(-)include/linux/printk.h +2
-1@@ -567,7 +567,8 @@ extern const struct file_operations kmsg_fops;enum {DUMP_PREFIX_NONE ,DUMP_PREFIX_ADDRESS , -DUMP_PREFIX_OFFSET +DUMP_PREFIX_OFFSET , +DUMP_PREFIX_UNHASHED , }; extern int hex_dump_to_buffer(const void *buf, size_t len, int rowsize, int groupsize, char *linebuf, size_t linebuflen,lib/hexdump.c +7
-2@@ -211,8 +211,9 @@ EXPORT_SYMBOL(hex_dump_to_buffer);* @level: kernel log level (e.g. KERN_DEBUG) * @prefix_str: string to prefix each line with; * caller supplies trailing spaces for alignment if desired -* @prefix_type: controls whether prefix of an offset, address, or none -* is printed (%DUMP_PREFIX_OFFSET, %DUMP_PREFIX_ADDRESS, %DUMP_PREFIX_NONE) +* @prefix_type: controls whether prefix of an offset, hashed address, +* unhashed address, or none is printed (%DUMP_PREFIX_OFFSET, +* %DUMP_PREFIX_ADDRESS, %DUMP_PREFIX_UNHASHED, %DUMP_PREFIX_NONE) * @rowsize: number of bytes to print per line; must be 16 or 32 * @groupsize: number of bytes to print at a time (1, 2, 4, 8; default = 1) * @buf: data blob to dump @@ -256,6 +257,10 @@ void print_hex_dump(const char *level, const char *prefix_str, int prefix_type, linebuf, sizeof(linebuf), ascii);switch (prefix_type) { +case DUMP_PREFIX_UNHASHED : +printk ("%s%s%px: %s\n" , + level, prefix_str, ptr + i, linebuf); +break ;case DUMP_PREFIX_ADDRESS :printk ("%s%s%p: %s\n" , level, prefix_str, ptr + i, linebuf);Signature · 1 line
2.25.1
Andrew Morton
maintainerQuoted 9 lines · On Mon, 11 Jan 2021 11:10:56 +0100 Petr Mladek <[email protected]> wrote:
Adding Kees into CC because it is security related. Adding Andrew into CC because he usually takes patches for hexdump.
Quoted 5 lines · Timur Tabi
Hashed addresses are useless in hexdumps unless you're comparing with other hashed addresses, which is unlikely. However, there's no need to break existing code, so introduce a new prefix type that prints unhashed addresses.
Quoted 8 lines
Quoted 2 lines
Signed-off-by Timur Tabi <[email protected]>Cc Roman Fietze <[email protected]>I agree that there should be way to print the real address.
And it is sane to add a new mode so that the current users stay hashed.
I doubt if Kees (or I or anyone else) can review this change because there are no callers which actually use the new DUMP_PREFIX_UNHASHED. Is it intended that some other places in the kernel be changed to use this? If so, please describe where and why, so that others can better understand both the requirement and the security implications.
If it is intended that this be used mainly for developer debug and not to be shipped in the mainline kernel then let's get this info into the changelog as well.
Quoted 5 lines · Andrew Morton
I doubt if Kees (or I or anyone else) can review this change because there are no callers which actually use the new
DUMP_PREFIX_UNHASHED. Is it intended that some other places in the kernel be changed to use this? If so, please describe where and why, so that others can better understand both the requirement and the security implications.
In my opinion, hashed addresses make no sense in a hexdump, so I would say that ALL callers should change. But none of the drivers I've written call print_hex_dump(), so I can't make those changes myself.
Quoted 3 lines
If it is intended that this be used mainly for developer debug and not to be shipped in the mainline kernel then let's get this info into the changelog as well.
I definitely want this patch included in the mainline kernel. Just because there aren't any users today doesn't mean that there won't be. In fact, I suspect that most current users haven't noticed that the addresses have changed or don't care any more, but if they were to write the code today, they would use unhashed addresses.
If you want, I can include a patch that changes a few callers of print_hex_dump() to use DUMP_PREFIX_UNHASHED, based on what I think would be useful.
Petr Mladek
maintainerQuoted 10 lines · Timur Tabi
Quoted 5 lines · Andrew Morton
I doubt if Kees (or I or anyone else) can review this change because there are no callers which actually use the new
DUMP_PREFIX_UNHASHED. Is it intended that some other places in the kernel be changed to use this? If so, please describe where and why, so that others can better understand both the requirement and the security implications.In my opinion, hashed addresses make no sense in a hexdump, so I would say that ALL callers should change. But none of the drivers I've written call
print_hex_dump(), so I can't make those changes myself.
I know that you probably know it because you introduced new mode instead of updating the existing one. But to be sure.
We need to be careful here. The hashed pointer has been introduced for a reason. It prevents leaking pointers and helping bad guys.
The original plan was to introduce %pK. It was supposed to prevent non-privileged users from seeing the real pointer value. It did not really worked because it was only rarely used. The plain %p was heavily used in historical and even in a new code.
By other words, every print_hex_dump() used need to be reviewed in which context might be called.
Quoted 9 lines
Quoted 3 lines
If it is intended that this be used mainly for developer debug and not to be shipped in the mainline kernel then let's get this info into the changelog as well.
I definitely want this patch included in the mainline kernel. Just because there aren't any users today doesn't mean that there won't be. In fact, I suspect that most current users haven't noticed that the addresses have changed or don't care any more, but if they were to write the code today, they would use unhashed addresses.
I am pretty sure that will look for this functionality sooner or later. The hashed pointers make debugging really complicated.
Quoted 3 lines
If you want, I can include a patch that changes a few callers of
print_hex_dump()to useDUMP_PREFIX_UNHASHED, based on what I think would be useful.
It would be nice.
Best Regards,
Petr
Quoted 2 lines · Petr Mladek
By other words, every
print_hex_dump()used need to be reviewed in which context might be called.
I did a rough analysis of all current usage of DUMP_PREFIX_ADDRESS in the kernel, compared to the introduction of %px and hashed addresses, using git-blame to find the most recent commit that modifies a line that contains _ADDRESS. Of 150 cases, 110 of them are newer than %px, so I'm assuming that these developers chose _ADDRESS instead of _OFFSET knowing that the addresses are hashed.
Quoted 5 lines
Quoted 3 lines
If you want, I can include a patch that changes a few callers of
print_hex_dump()to useDUMP_PREFIX_UNHASHED, based on what I think would be useful.It would be nice.
I have a v2 that I'm about to post, and I included a patch that modifies check_poison_mem() in mm/page_poison.c. I chose this file because I figured actual addresses would probably be necessary for tracking memory-related errors. Also, that call to print_hex_dump() was added back in 2009, so it predates the introduction of %px.