1) Partial failure of ASLR on 32-bit Fedora
My Fedora is a couple of releases behind, so no idea if it's been fixed. It seems that the desire to pack all the shared libraries into virtual address 0x00nnnnnn has a catastrophic failure mode when there are too many libraries: something always ends up at 0x00110000. You can see it with repeated invocations of
ldd /opt/google/chrome/chrome|grep 0x0011:libglib-2.0.so.0 => /lib/libglib-2.0.so.0 (0x00110000)
libXext.so.6 => /usr/lib/libXext.so.6 (0x00110000)
libdl.so.2 => /lib/libdl.so.2 (0x00110000)
Which exact library is placed at the fixed address is random. However, any fixed address can be a real problem to ASLR. For example, in the browser context, take a bug such as Chris Rohlf's older but interesting CSS type confusion. Without a fixed address, a crash is a likely outcome. With a fixed address, the exact library mapped at the fixed address could easily be fingerprinted, and the BSS section read to leak heap pointers (e.g. via singleton patterns). Bye bye to both NX and ASLR.
Aside: in the 32-bit browser context with plenty of physical memory, a Javascript-based heap spray could easily fill most of the address space such that the attacker's deference has a low chance of failure.
Aside #2: my guess is that this scheme is designed to prevent a return-to-glibc attack vs. strcpy(), by making sure that all executable addresses contain a NULL byte. I'm probably missing something, but it seems like the fact that strcpy() NULL-terminates, combined with the little-endianness of Intel, makes this not so strong.
2) Missed opportunity to use more entropy on 64-bit
If you look at the maps of a 64-bit process, you'll see most virtual memory areas correspond to the formula 0x7fnnxxxxxxxx where all your stuff is piled together in xxxxxxxx and nn is random. At least, nothing is in or near a predictable location. One way to look at how this could be better is this: If you emit a 4GB heap spray, you have a ~1/256 chance of guessing where it is. Using the additional 7 bits of entropy might be useful, especially for the heap.
3) Bad mmap() randomization
Although the stack, heap and binary are placed at reasonably random locations, unhinted mmap() chunks are sort of just piled up adjacent, typically in a descending-vm-address fashion. This can lead to problems where a buffer overflow crashes into a sensitive mapping -- such a JIT mapping. (This is one reason JIT mappings have their own randomizing allocator in v8).
4) Heap / stack collision likely with ASLR binary
On a 32-bit kernel you might see:
b8105000-b8124000 rw-p 00000000 00:00 0 [heap]
bfae5000-bfb0a000 rw-p 00000000 00:00 0 [stack]
Or on a 64-bit kernel running a 32-bit process:
f7c52000-f7c73000 rw-p 00000000 00:00 0 [heap]
ff948000-ff96d000 rw-p 00000000 00:00 0 [stack]
In both cases, the heap doesn't have to grow too large before it cannot grow any larger. When this happens, most heap implementations fall back to mmap() allocations, and suffer the problems of 3) above. These things chained together with a very minor infoleak such as my cross-browser XSLT heap address leak could in fact leak the location of the executable, leading to a full NX/ASLR bypass.
Conclusion
A 32-bit address space just isn't very big any more, compared with todays large binaries, large number of shared library dependencies and large heaps. It's no surprise that everything is looking a little crammed in. The good news is that there are no obvious and severe problems with the 64-bit situation, although the full entropy isn't used. Applications (such as v8 / Chromium) can and do fix that situation for the most sensitive mappings themselves.