On 5/29/19 8:20 AM, Catalin Marinas wrote:
> Hi Khalid,
>
> On Tue, May 28, 2019 at 05:33:04PM -0600, Khalid Aziz wrote:
>> On Tue, 2019-05-28 at 16:40 +0100, Catalin Marinas wrote:
>>> I think another aspect is how we define the ABI. Is allowing tags to
>>> mlock() for example something specific to arm64 or would sparc ADI
>>> need the same? In the absence of other architectures defining such
>>> ABI, my preference would be to keep the wrappers in the arch code.
>>>
>>> Assuming sparc won't implement untagged_addr(), we can place the
>>> macros back in the generic code but, as per the review here, we need
>>> to be more restrictive on where we allow tagged addresses. For
>>> example, if mmap() gets a tagged address with MAP_FIXED, is it
>>> expected to return the tag?
>>
>> I would recommend against any ABI differences between ARM64 MTE/TBI and
>> sparc ADI unless it simply can not be helped. My understanding of
>> MTE/TBI is limited, so I will explain how sparc ADI works. On sparc, a
>> tagged address has no meaning until following steps happen:
>
> Before we go into the MTE/ADI similarities or differences, just to
> clarify that TBI is something that we supported from the start of the
> arm64 kernel port. TBI (top byte ignore) allows a user pointer to have
> non-zero top byte and dereference it without causing a fault (the
> hardware masks it out). The user/kernel ABI does not allow such tagged
> pointers into the kernel, nor would the kernel return any such tagged
> addresses.
>
> With MTE (memory tagging extensions), the top-byte meaning is changed
> from no longer being ignored to actually being checked against a tag in
> the physical RAM (we call it allocation tag).
>
>> 1. set the user mode PSTATE.mcde bit. This acts as the master switch to
>> enable ADI for a process.
>>
>> 2. set TTE.mcd bit on TLB entries that match the address range ADI is
>> being enabled on.
>
> Something close enough for MTE, with the difference that enabling it is
> not a PSTATE bit but rather a system control bit (SCTLR_EL1 register),
> so only the kernel can turn it on/off for the user.
>
>> 3. Store version tag for the range of addresses userspace wants ADI
>> enabled on using "stxa" instruction. These tags are stored in physical
>> memory by the memory controller.
>
> Do you have an "ldxa" instruction to load the tags from physical memory?
Yes, "ldxa" can be used to read current tag for any memory location.
Kernel uses it to read the tags for a physical page being swapped out
and restores those tags when the page is swapped back in.
>
>> Steps 1 and 2 are accomplished by userspace by calling mprotect() with
>> PROT_ADI. Tags are set by storing tags in a loop, for example:
>>
>> version = 10;
>> tmp_addr = shmaddr;
>> end = shmaddr + BUFFER_SIZE;
>> while (tmp_addr < end) {
>> asm volatile(
>> "stxa %1, [%0]0x90\n\t"
>> :
>> : "r" (tmp_addr), "r" (version));
>> tmp_addr += adi_blksz;
>> }
>
> On arm64, a sequence similar to the above would live in the libc. So a
> malloc() call will tag the memory and return the tagged address to thePre-coloring could easily be done by
> user.
>
> We were not planning for a PROT_ADI/MTE but rather have MTE enabled for
> all user memory ranges. We may revisit this before we upstream the MTE
> support (probably some marginal benefit for the hardware not fetching
> the tags from memory if we don't need to, e.g. code sections).
>
> Given that we already have the TBI feature and with MTE enabled the top
> byte is no longer ignored, we are planning for an explicit opt-in by the
> user via prctl() to enable MTE.
OK. I had initially proposed enabling ADI for a process using prctl().
Feedback I got was prctl was not a desirable interface and I ended up
making mprotect() with PROT_ADI enable ADI on the process instead. Just
something to keep in mind.
>
>> With these semantics, giving mmap() or shamat() a tagged address is
>> meaningless since no tags have been stored at the addresses mmap() will
>> allocate and one can not store tags before memory range has been
>> allocated. If we choose to allow tagged addresses to come into mmap()
>> and shmat(), sparc code can strip the tags unconditionally and that may
>> help simplify ABI and/or code.
>
> We could say that with TBI (pre-MTE support), the top byte is actually
> ignored on mmap(). Now, if you pass a MAP_FIXED with a tagged address,
> should the user expect the same tagged address back or stripping the tag
> is acceptable? If we want to keep the current mmap() semantics, I'd say
> the same tag is returned. However, with MTE this also implies that the
> memory was coloured.
>
Is assigning a tag aprivileged operationon ARM64? I am thinking not
since you mentioned libc could do it in a loop for malloc'd memory.
mmap() can return the same tagged address but I am uneasy about kernel
pre-coloring the pages. Database can mmap 100's of GB of memory. That is
lot of work being offloaded to the kernel to pre-color the page even if
done in batches as pages are faulted in.
>>> My thoughts on allowing tags (quick look):
>>>
>>> brk - no
>>> get_mempolicy - yes
>>> madvise - yes
>>> mbind - yes
>>> mincore - yes
>>> mlock, mlock2, munlock - yes
>>> mmap - no (we may change this with MTE but not for TBI)
>>> mmap_pgoff - not used on arm64
>>> mprotect - yes
>>> mremap - yes for old_address, no for new_address (on par with mmap)
>>> msync - yes
>>> munmap - probably no (mmap does not return tagged ptrs)
>>> remap_file_pages - no (also deprecated syscall)
>>> shmat, shmdt - shall we allow tagged addresses on shared memory?
>>>
>>> The above is only about the TBI ABI while ignoring hardware MTE. For
>>> the latter, we may want to change the mmap() to allow pre-colouring
>>> on page fault which means that munmap()/mprotect() should also
>>> support tagged pointers. Possibly mremap() as well but we need to
>>> decide whether it should allow re-colouring the page (probably no,
>>> in which case old_address and new_address should have the same tag).
>>> For some of these we'll end up with arm64 specific wrappers again,
>>> unless sparc ADI adopts exactly the same ABI restrictions.
>>
>> Let us keep any restrictions common across ARM64 and sparc. pre-
>> coloring on sparc in the kernel would mean kernel will have to execute
>> stxa instructions in a loop for each page being faulted in.
>
> Since the user can probe the pre-existing colour in a faulted-in page
> (either with some 'ldxa' instruction or by performing a tag-checked
> access), the kernel should always pre-colour (even if colour 0) any
> allocated page. There might not be an obvious security risk but I feel
> uneasy about letting colours leak between address spaces (different user
> processes or between kernel and user).
On sparc, tags 0 and 15 are special in that 0 means untagged memory and
15 means match any tag in the address. Colour 0 is the default for any
newly faulted in page on sparc.
>
> Since we already need such loop in the kernel, we might as well allow
> user space to require a certain colour. This comes in handy for large
> malloc() and another advantage is that the C library won't be stuck
> trying to paint the whole range (think GB).
If kernel is going to pre-color all pages in a vma, we will need to
store the default tag in the vma. It will add more time to page fault
handling code. On sparc M7, kernel will need to execute additional 128
stxa instructions to set the tags on a normal page.
>
>> Not that big a deal but doesn't that assume the entire page has the
>> same tag which is dedcued from the upper bits of address? Shouldn't we
>> support tags at the same granularity level as what the hardware
>> supports?
>
> That's mostly about large malloc() optimisation via mmap(), the latter
> working on page granularity already. There is another use-case for
> pre-coloured thread stacks, also allocated via anonymous mmap().
>
>> We went through this discussion for sparc and decision was to support
>> tags at the same granularity as hardware. That means we can not deduce
>> tags from the first address that pioints into an mmap or shmat region.
>> Those tags and the upper bytes of colored address could change for
>> every cacheline sized block (64-bytes on sparc M7).
>
> It's 16-byte for arm64, so smaller than the cacheline.
>
>> We can try to store tags for an entire region in vma but that is
>> expensive, plus on sparc tags are set in userspace with no
>> participation from kernel and now we need a way for userspace to
>> communicate the tags to kernel.
>
> We can't support finer granularity through the mmap() syscall and, as
> you said, the vma is not the right thing to store the individual tags.
> With the above extension to mmap(), we'd have to store a colour per vma
> and prevent merging if different colours (we could as well use the
> pkeys mechanism we already have in the kernel but use a colour per vma
> instead of a key).
Since tags can change on any part of mmap region on sparc at any time
without kernel being involved, I am not sure I see much reason for
kernel to enforce any tag related restrictions.
>
> Of course, the user is allowed to change the in-memory colours at a
> finer granularity and the kernel will preserve them during swapping
> out/in, page migration etc. The above mmap() proposal is just for the
> first fault-in of a page in a given range/vma.
>
>> From sparc point of view, making kernel responsible for assigning tags
>> to a page on page fault is full of pitfalls.
>
> This could be just some arm64-specific but if you plan to deploy it more
> generically for sparc (at the C library level), you may find this
> useful.
>
Common semantics from app developer point of view will be very useful to
maintain. If arm64 says mmap with MAP_FIXED and a tagged address will
return a pre-colored page, I would rather have it be the same on any
architecture. Is there a use case that justifies kernel doing this extra
work?
--
Khalid
