On Tue, 15 Mar 2022 22:58:48 GMT, Tyler Steele <d...@openjdk.java.net> wrote:
> As described in the linked issue, NullClassBytesTest fails due an > OutOfMemoryError produced on AIX when the test calls defineClass with a byte > array of size of 0. The native implementation of defineClass then calls > malloc with a size of 0. On AIX malloc(0) returns NULL, while on other > platforms it return a valid address. When NULL is produced by malloc for this > reason, ClassLoader.c incorrectly interprets this as a failure due to a lack > of memory. > > This PR modifies ClassLoader.c to produce an OutOfMemoryError only when > `errno == ENOMEM` and to produce a ClassFormatError with the message > "ClassLoader internal allocation failure" in all other cases (in which malloc > returns NULL). > > In addition, I performed some minor tidy-up work in ClassLoader.c by changing > instances of `return 0` to `return NULL`, and `if (some_ptr == 0)` to `if > (some_ptr == NULL)`. This was done to improve the clarity of the code in > ClassLoader.c, but didn't feel worthy of opening a separate issue. > > ### Alternatives > > It would be possible to address this failure by modifying the test to accept > the OutOfMemoryError on AIX. I thought it was a better solution to modify > ClassLoader.c to produce an OutOfMemoryError only when the system is actually > out of memory. > > ### Testing > > This change has been tested on AIX and Linux/x86. Hi Tyler, The way we solve this usually is by homogenizing malloc behavior across all platforms with `if (len == 0) len=1;`, see e.g. https://github.com/openjdk/jdk/blob/08cadb4754da0d5e68ee2df45f4098d203d14115/src/hotspot/share/runtime/os.cpp#L644-L647 I suggest you do that here too since it would be a far less intrusive change. If you don't want to fix up all three malloc call sites, just reroute them to a local malloc stub (`my_malloc()`) that corrects length and calls the real malloc. If you want, you can `#ifdef _AIX` that length correction too, e.g.: static void* my_malloc(size_t l) { #ifdef _AIX if (l == 0) l = 1; #endif return malloc(l); } That would be a far smaller change, and it would not introduce a new behavioral difference (because with your change, AIX now throws ClassFormatError where other platforms don't). Side note: nothing against changing 0 to NULL, but please in a separate cleanup patch. Cheers, Thomas Btw, which malloc call was the problematic exactly? Cannot be the one in getUTF, since that one already adds len + 1 and never gets called with a zero length anyway. ------------- Changes requested by stuefe (Reviewer). PR: https://git.openjdk.java.net/jdk/pull/7829