On 1/20/24 14:49, Charlie Jenkins wrote:
On Sat, Jan 20, 2024 at 02:13:14PM +0800, Yangyu Chen wrote:
Thanks for your reply.
On 1/20/24 09:34, Charlie Jenkins wrote:
On Sun, Jan 14, 2024 at 01:26:57AM +0800, Yangyu Chen wrote:
Hi, Charlie
Although this patchset has been merged I still have some questions about this patchset. Because it breaks regular mmap if address >= 38 bits on sv48 / sv57 capable systems like qemu. For example, If a userspace program wants to mmap an anonymous page to addr=(1<<45) on an sv48 capable system, it will fail and kernel will mmaped to another sv39 address since it does
Thank you for raising this concern. To make sure I am understanding correctly, you are passing a hint address of (1<<45) and expecting mmap to return 1<<45 and if it returns a different address you are describing mmap as failing? If you want an address that is in the sv48 space you can pass in an address that is greater than 1<<47.
not meet the requirement to use sv48 as you wrote:
else if ((((_addr) >= VA_USER_SV48)) && (VA_BITS >= VA_BITS_SV48)) \ mmap_end = VA_USER_SV48; \ else \ mmap_end = VA_USER_SV39; \
Then, How can a userspace program create a mmap with a hint if the address
= (1<<38) after your patch without MAP_FIXED? The only way to do this is
to pass a hint >= (1<<47) on mmap syscall then kernel will return a random address in sv48 address space but the hint address gets lost. I think this
In order to force mmap to return the address provided you must use MAP_FIXED. Otherwise, the address is a "hint" and has no guarantees. The hint address on riscv is used to mean "don't give me an address that uses more bits than this". This behavior is not unique to riscv, arm64 and powerpc use a similar scheme. In arch/arm64/include/asm/processor.h there is the following code:
#define arch_get_mmap_base(addr, base) ((addr > DEFAULT_MAP_WINDOW) ? \ base + TASK_SIZE - DEFAULT_MAP_WINDOW :\ base)
arm64/powerpc are only concerned with a single boundary so the code is simpler.
As you say, this code in arm64/powerpc will not meet the issue I address. For example, If the addr here is (1<<50) on arm64, the arch_get_mmap_base will return base+TASK_SIZE-DEFAULT_MAP_WINDOW which is (1<<vabits_actual). And this behavior on arm64/powerpc/x86 does not break anything since we will use a larger address space if the hint address is specified on the address > DEFAULT_MAP_WINDOW. The corresponding behavior on RISC-V should be if the hint address > BIT(47) then use Sv57 address space and use Sv48 when the hint address > BIT(38) if we want Sv39 by default.
However, your patch needs the address >= BIT(47) rather than BIT(38) to use Sv48 and address >= BIT(56) to use Sv57, thus breaking existing userspace software to create mapping on the hint address without MAP_FIXED set.
Code that needs mmap to provide a specific address must use MAP_FIXED. On riscv, it was decided that the address returned from mmap cannot be greater than the hint address. This is currently implemented by using the largest address space that can fit into the hint address. It may be possible that this range can be extended to use all of the addresses that are less than or equal to the hint address.
So this decision might be wrong. It requires some userspace software to modify their mmap flags to fit with this. For example, a binary translate JIT compiler already probes this platform is capable with Sv48, then want to create mapping on some address specified on the mmap hint to align with foreign binary native address but also provide a fallback path with performance overhead. Your patch here will always let userspace software use a fallback path with performance overhead until the userspace software changes its syscall to use MAP_FIXED. But it is not required in x86, arm64, powerpc.
From reading the code even on arm64 if you pass an address that is greater than DEFAULT_MAP_WINDOW it is not guaranteed that mmap will return an address that is greater than DEFAULT_MAP_WINDOW. It may still be provide an address that is less than DEFAULT_MAP_WINDOW if it fails to find an address above. This seems like this would also break your use case.
Yeah. As I said before, this patch will always let userspace software use a fallback path and this only happens in RISC-V. Make default sv48 is right, but RISC-V implementation for this and changing the hint address behavior might be wrong. And x86, arm64, powerpc already use 48-bit address space by default but do not change the meaning of hint address on mmap.
violate the principle of mmap syscall as kernel should take the hint and attempt to create the mapping there.
Although the man page for mmap does say "on Linux, the kernel will pick a nearby page boundary" it is still a hint address so there is no strict requirement (and the precedent has already been set by arm64/powerpc).
Yeah. There is no strict requirement. But currently x86/arm64/powerpc works in this situation well. The hint address on these ISAs is not used as the upper bound to allocating the address. However, on RISC-V, you treat this as the upper bound.
I don't think patching in this way is right. However, if we only revert this patch, some programs relying on mmap to return address with effective bits <= 48 will still be an issue and it might expand to other ISAs if they implement larger virtual address space like RISC-V sv57. A better way to solve this might be adding a MAP_48BIT flag to mmap like MAP_32BIT has been introduced for decades.
Thanks, Yangyu Chen
- Charlie