RFR: 8220095: Assertion failure when symlink (with different name) is used for lib/modules file
Jiangli Zhou
jianglizhou at google.com
Wed Mar 13 17:22:00 UTC 2019
On Tue, Mar 12, 2019 at 4:42 PM Jiangli Zhou <jianglizhou at google.com> wrote:
> One of my early experimental revision actually simply used strcmp to
> compare the full path and to also recorded the full path for the canonical
> modules name.
>
> I decide to not go with that because there is no guarantee that the path
> given to is_modules_image check is in canonical form and it diverts from
> the existing assumptions about the usages of this API (checking the file
> name only). The "/dev/fd/0" example you have is not a direct result of the
> current fix but is an inherent issue of the modules check. Any file with
> the same designated modules name would pass the check. If we want to safely
> use the full path comparison, we need to closely exam all the call sites
> and make sure all the places use canonical paths. I'll some analysis on
> that.
>
I've double checked all existing call sites of is_modules_image(const
char*). No extra processing is done (e.g. extracting the file name) on the
path strings when passing to is_modules_image. Using the full string
comparison appears to be safe so I have no objection. I've reverted my
change to record the complete canonical path as the modules identity and
changed to do full path string comparison:
http://cr.openjdk.java.net/~jiangli/8220095/webrev.02/
Local testing looks good. I've also submitted a submit-repo testing. Please
help rerun tier2, tier3 and maybe some of the higher tier tests due to the
semantics change of is_modules_image.
Thanks,
Jiangli
>
> Thanks,
> Jiangli
>
> On Tue, Mar 12, 2019 at 4:13 PM Ioi Lam <ioi.lam at oracle.com> wrote:
>
>> This seems more complicated that it needs to be.
>>
>> + The VM can have at most one modules file.
>>
>> + The only reason the VM calls is_modules_image(p) is to check whether p,
>> a full path, is THE modules file. In all cases, p is actually derived from
>> the singleton ClassPathImageEntry::name(). It's just that in some cases,
>> such as in the logging code, we have lost the pointer to the
>> ClassPathImageEntry, and just have the full path.
>>
>> + On line 722 of
>> http://cr.openjdk.java.net/~jiangli/8220095/webrev.01/src/hotspot/share/classfile/classLoader.cpp.sdiff.html,
>> you already have a full path of THE modules file.
>>
>> Why can't you just do a strcmp of the full path? That's much simpler, and
>> more importantly, correct. Your version will return true for "/dev/fd/0"
>> for the test case described in the bug report. That doesn't seem to be
>> correct.
>>
>>
>> Thanks
>> - Ioi
>>
>>
>>
>> On 3/12/19 9:10 AM, Jiangli Zhou wrote:
>>
>> Hi Ioi,
>>
>> Thanks for looking into this also.
>>
>> On Mon, Mar 11, 2019 at 8:19 PM Ioi Lam <ioi.lam at oracle.com> wrote:
>>
>>>
>>>
>>> On 3/11/19 7:30 PM, Ioi Lam wrote:
>>> > I think this patch should be further improved.
>>> >
>>> > First of all, the original code is problematic:
>>> >
>>> > static bool is_modules_image(const char* name) { return
>>> > string_ends_with(name, MODULES_IMAGE_NAME); }
>>> >
>>> > So if we have a file called /foo/bar/xxxmodules, this function returns
>>> > true!
>>> >
>>>
>>> BTW, If you have a symlink JAVA_HOME/lib/modules -> JAVA_HOME/lib/r,
>>> then your new version of is_modules_image will return true for any jar
>>> file.
>>>
>>> See
>>>
>>> http://hg.openjdk.java.net/jdk/jdk/file/f984aca565c1/src/hotspot/share/oops/instanceKlass.cpp#l3370
>>>
>>> if (ClassLoader::is_modules_image(cfs->source())) {
>>> info_stream.print(" source: jrt:/%s", module_name);
>>> }
>>>
>>> This will cause all classes loaded from modular JAR files be incorrectly
>>> printed as "jrt://"
>>>
>>> Even if you also use the last "/" path separator, it's still not safe to
>>> use the filename alone to determine whether a file is indeed the modules
>>> image file. I.e., what if you have JAVA_HOME/lib/modules ->
>>> JAVA_HOME/lib/test.jar?
>>>
>>
>> The current bug addresses the practical issue when the modules is a sym
>> link to a unique content identifier. Using modules as a link to any
>> arbitrary file is not addressed by this bug. It's in the same area but a
>> different issue.
>>
>> If lib/modules ->lib/test.jar, or lib/modules -> ../ (or any other
>> arbitrary file/dir) is done, the VM fails differently. This is the same as
>> if you replace libjvm.so with another random file.
>>
>> The canonical modules name is obtained from the actual modules image
>> during setup_boot_search_path, so it's guaranteed to refer to the correct
>> identifier. Checking the file name after '/' is safe.
>>
>> The substring issue that you pointed out has not been observed, as that's
>> prevented by the naming method of the content identifiers. The substring
>> issue however can be solved relatively inexpensively by making sure the
>> check covers the entire file name after '/'.
>>
>>
>>>
>>>
>>> Also, I think we can have a version that's simpler than what I proposed
>>> below:
>>>
>>> + We can have at most a single ClassPathImageEntry.
>>> + All calls to ClassLoader::is_modules_image() are made after the
>>> ClassPathImageEntry is created
>>>
>>> so:
>>>
>>> static bool is_modules_image(const char* name) {
>>> if (ClassPathImageEntry::singleton() != NULL) {
>>> return strcmp(name, ClassPathImageEntry::singleton()->name()) ==
>>> 0);
>>> } else {
>>> assert(_exploded_entries != NULL, "must be running with exploded
>>> modules");
>>> return false;
>>> }
>>> }
>>>
>>
>> The argument passed can be a path to the modules, so strcmp cannot be
>> simply used for 'name'.
>>
>> static bool is_modules_image(const char* name) {
>>
>> if (Arguments::has_jimage()) {
>>
>> assert(_modules_image_name != NULL, "must be set");
>>
>> const char* p = skip_directories(name);
>>
>> return strcmp(p, _modules_image_name) == 0;
>>
>> }
>>
>> return false;
>>
>> }
>>
>>
>> I've revised the is_modules_image() to also cover the substring issue.
>>
>>
>>
>>
>>>
>>> This function should be changed to always return true
>>>
>>> bool ClassPathImageEntry::is_modules_image() const {
>>> - return ClassLoader::is_modules_image(name());
>>> + return true;
>>> }
>>>
>>
>> Leaving ClassPathImageEntry::is_modules_image() without changing is
>> safer, so it can make sure the ClassPathImageEntry name agrees with the
>> canonical modules name.
>>
>>
>>>
>>>
>>> To make sure we don't call is_modules_image before ClassPathImageEntry
>>> is created, we could do
>>>
>>> static bool is_modules_image(const char* name) {
>>> DEBUG_ONLY(_is_modules_image_called = true;)
>>> ...
>>> }
>>>
>>
>> The 'assert(_modules_image_name != NULL, "must be set")' check makes sure
>> is_modules_image(const char*) is not called before setup_boot_search_path.
>> There is no need to add additional flag.
>>
>> Thanks,
>> Jiangli
>>
>>
>>>
>>>
>>> ClassPathImageEntry::ClassPathImageEntry() {
>>> assert(_singleton == NULL, "...");
>>> assert(!_is_modules_image_called, "Don't call
>>> ClassLoader::is_modules_image() before ClassPathImageEntry is
>>> allocated");
>>> ...
>>>
>>>
>>> Thanks
>>> - Ioi
>>>
>>>
>>> > It seems the patch simply makes the code more convoluted. Instead,
>>> > shouldn't we just open the file and check the magic number?
>>> >
>>> > If performance is important, we can use a cache, because we allow a
>>> > single module image:
>>> >
>>> > static bool is_modules_image(const char* name) {
>>> > static const char* singleton_module_name = NULL;
>>> >
>>> > if (singleton_module_name != NULL) {
>>> > return strcmp(name, singleton_module_name) == 0;
>>> > }
>>> > // open <name> and check magic number
>>> > if (good) {
>>> > singleton_module_name = os::strdup(name);
>>> > return true;
>>> > }
>>> > return false;
>>> > }
>>> >
>>> > Thanks
>>> > - Ioi
>>> >
>>> > On 3/11/19 10:51 AM, Jiangli Zhou wrote:
>>> >> On Mon, Mar 11, 2019 at 8:44 AM Calvin Cheung <
>>> calvin.cheung at oracle.com>
>>> >> wrote:
>>> >>
>>> >>> Hi Jiangli,
>>> >>>
>>> >>> The updated webrev looks good.
>>> >>>
>>> >>> On 3/10/19, 6:41 PM, Jiangli Zhou wrote:
>>> >>>> Here is the updated webrev:
>>> >>>> http://cr.openjdk.java.net/~jiangli/8220095/webrev.01/
>>> >>>> <http://cr.openjdk.java.net/%7Ejiangli/8220095/webrev.01/>.
>>> >>>>
>>> >>>> - Incorporated all suggestions except the test platforms for
>>> >>>> ModulesSymLink.java. I changed to requires linux, mac and solaris
>>> >>>> explicitly instead to avoid any potential issue with other
>>> >>>> non-mainline testing platforms.
>>> >>>> - Reverted the changes in sharedPathsMiscInfo.cpp. Further testing
>>> >>>> showed SharedPathsMiscInfo::check() were dealing the system boot
>>> path
>>> >>>> string set up by os::set_boot_path. Lois comments reminded me to
>>> >>>> double check with this. Thanks!
>>> >>>>
>>> >>>> Please help fire tier2 and tier3 runs with mach5.
>>> >>> I ran tier2 and tier3 tests with your latest changes and saw no test
>>> >>> failures.
>>> >>>
>>> >> Great, thanks a lot!
>>> >>
>>> >>
>>> >>> BTW, I was unable to download the jdk.patch - got "403 - forbidden"
>>> >>> error.
>>> >>> Same error was seen with individual patch files.
>>> >>> I used wget to get all the raw files into my repo and then submitted
>>> >>> the
>>> >>> build and test job.
>>> >>> Could you check the file permissions?
>>> >>>
>>> >> Yes, that's the file permission issue. I fixed the permission for the
>>> >> jdk.patch.
>>> >>
>>> >> Thanks!
>>> >> Jiangli
>>> >>
>>> >>
>>> >>> thanks,
>>> >>> Calvin
>>> >>>> Thanks!
>>> >>>> Jiangli
>>> >>>>
>>> >>>> On Thu, Mar 7, 2019 at 5:11 AM Lois Foltan <lois.foltan at oracle.com
>>> >>>> <mailto:lois.foltan at oracle.com>> wrote:
>>> >>>>
>>> >>>> On 3/6/2019 6:20 PM, Jiangli Zhou wrote:
>>> >>>> > Please review the following fix for 8220095.
>>> >>>> >
>>> >>>> > webrev:
>>> http://cr.openjdk.java.net/~jiangli/8220095/webrev.00/
>>> >>>> <http://cr.openjdk.java.net/%7Ejiangli/8220095/webrev.00/>
>>> >>>> > bug: https://bugs.openjdk.java.net/browse/JDK-8220095
>>> >>>> >
>>> >>>> > Symbolic links may be used commonly in some cloud
>>> environments.
>>> >>>> The target
>>> >>>> > files might have different names than the linked names. The
>>> VM
>>> >>>> crashes when
>>> >>>> > 'lib/modules' links to a target file with a different name,
>>> for
>>> >>>> example,
>>> >>>> > lib/modules -> lib/0.
>>> >>>> >
>>> >>>> > The usage of the hard-coded MODULES_IMAGE_NAME (used by
>>> >>>> > ClassLoader::is_modules_image(const char*)) can become
>>> >>>> problematic in this
>>> >>>> > case and leads to assertion failures and crashes in this
>>> >>>> case. The
>>> >>>> > is_modules_image()
>>> >>>> > checks always fail even the actual file is the runtime image
>>> >>>> because the
>>> >>>> > canonical paths (which might not end with 'modules') are
>>> passed
>>> >>>> in for the
>>> >>>> > check. The fix is to obtain the real name from the canonical
>>> >>>> path early
>>> >>>> > during initialization and use that for the is_modules_image()
>>> >>> check.
>>> >>>> >
>>> >>>> > Tested with submit repo testing, tier1, tier2, and hotspot
>>> >>>> runtime tests
>>> >>>> > locally.
>>> >>>> >
>>> >>>> > Thanks!
>>> >>>> > Jiangli
>>> >>>>
>>> >>>> Hi Jiangli,
>>> >>>>
>>> >>>> This change looks good! A couple of minor comments:
>>> >>>>
>>> >>>> - classLoader.hpp
>>> >>>> line #39 - I think you should expand on the comment and
>>> add
>>> >>>> why in
>>> >>>> the case of the canonical path name MODULES_IMAGE_NAME can not
>>> be
>>> >>>> used
>>> >>>> as-is but if one is dealing with the system boot path string
>>> set
>>> >>>> up by
>>> >>>> os::set_boot_path, than MODULES_IMAGE_NAME should be used.
>>> >>>>
>>> >>>> - classLoader.cpp
>>> >>>> line #702 - Can you add the same explanation comment as in
>>> >>>> classLoader.hpp ahead of the new
>>> >>>> ClassLoader::init_modules_image_name()
>>> >>>> method? That would be helpful.
>>> >>>> line #710 - ahead of setting _modules_image_name please
>>> >>>> consider
>>> >>>> adding an assert that p1's length is greater than 0?
>>> >>>> Maybe there should there be an error
>>> case
>>> >>>> test
>>> >>>> where "lib/modules -> lib/"
>>> >>>>
>>> >>>> Thanks,
>>> >>>> Lois
>>> >>>>
>>> >>>> -
>>> >>>>
>>> >
>>>
>>>
>>
More information about the hotspot-runtime-dev
mailing list