On Tue, Sep 10, 2024 at 05:45:07PM -0700, Charlie Jenkins wrote:
On Tue, Sep 10, 2024 at 03:08:14PM -0400, Liam R. Howlett wrote:
- Catalin Marinas catalin.marinas@arm.com [240906 07:44]:
On Fri, Sep 06, 2024 at 09:55:42AM +0000, Arnd Bergmann wrote:
On Fri, Sep 6, 2024, at 09:14, Guo Ren wrote:
On Fri, Sep 6, 2024 at 3:18 PM Arnd Bergmann arnd@arndb.de wrote:
It's also unclear to me how we want this flag to interact with the existing logic in arch_get_mmap_end(), which attempts to limit the default mapping to a 47-bit address space already.
To optimize RISC-V progress, I recommend:
Step 1: Approve the patch. Step 2: Update Go and OpenJDK's RISC-V backend to utilize it. Step 3: Wait approximately several iterations for Go & OpenJDK Step 4: Remove the 47-bit constraint in arch_get_mmap_end()
Point 4 is an ABI change. What guarantees that there isn't still software out there that relies on the old behaviour?
I really want to first see a plausible explanation about why RISC-V can't just implement this using a 47-bit DEFAULT_MAP_WINDOW like all the other major architectures (x86, arm64, powerpc64),
FWIW arm64 actually limits DEFAULT_MAP_WINDOW to 48-bit in the default configuration. We end up with a 47-bit with 16K pages but for a different reason that has to do with LPA2 support (I doubt we need this for the user mapping but we need to untangle some of the macros there; that's for a separate discussion).
That said, we haven't encountered any user space problems with a 48-bit DEFAULT_MAP_WINDOW. So I also think RISC-V should follow a similar approach (47 or 48 bit default limit). Better to have some ABI consistency between architectures. One can still ask for addresses above this default limit via mmap().
I think that is best as well.
Can we please just do what x86 and arm64 does?
I responded to Arnd in the other thread, but I am still not convinced that the solution that x86 and arm64 have selected is the best solution. The solution of defaulting to 47 bits does allow applications the ability to get addresses that are below 47 bits. However, due to differences across architectures it doesn't seem possible to have all architectures default to the same value. Additionally, this flag will be able to help users avoid potential bugs where a hint address is passed that causes upper bits of a VA to be used.
The reason we added this limit on arm64 is that we noticed programs using the top 8 bits of a 64-bit pointer for additional information. IIRC, it wasn't even openJDK but some JavaScript JIT. We could have taught those programs of a new flag but since we couldn't tell how many are out there, it was the safest to default to a smaller limit and opt in to the higher one. Such opt-in is via mmap() but if you prefer a prctl() flag, that's fine by me as well (though I think this should be opt-in to higher addresses rather than opt-out of the higher addresses).