Re: [RFC PATCH v3 4/5] riscv/cmpxchg: Implement cmpxchg for variables of size 1 and 2

[Leonardo Bras Soares Passos]

v5: https://lore.kernel.org/all/20230810040349.92279-2-leobras@xxxxxxxxxx/

On Tue, Aug 8, 2023 at 11:02 PM Leonardo Bras Soares Passos
<leobras@xxxxxxxxxx> wrote:
>
> Hello Guo Ren,
>
> On Mon, Aug 7, 2023 at 1:17 PM Leonardo Bras Soares Passos
> <leobras@xxxxxxxxxx> wrote:
> >
> > On Sat, Aug 5, 2023 at 1:24 AM Guo Ren <guoren@xxxxxxxxxx> wrote:
> > >
> > > On Sat, Aug 5, 2023 at 11:14 AM Leonardo Bras Soares Passos
> > > <leobras@xxxxxxxxxx> wrote:
> > > >
> > > > Hello Guo Ren, thanks for the feedback!
> > > >
> > > > On Fri, Aug 4, 2023 at 2:45 PM Guo Ren <guoren@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Aug 4, 2023 at 4:49 AM Leonardo Bras <leobras@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > cmpxchg for variables of size 1-byte and 2-bytes is not yet available for
> > > > > > riscv, even though its present in other architectures such as arm64 and
> > > > > > x86. This could lead to not being able to implement some locking mechanisms
> > > > > > or requiring some rework to make it work properly.
> > > > > >
> > > > > > Implement 1-byte and 2-bytes cmpxchg in order to achieve parity with other
> > > > > > architectures.
> > > > > >
> > > > > > Signed-off-by: Leonardo Bras <leobras@xxxxxxxxxx>
> > > > > > ---
> > > > > >  arch/riscv/include/asm/cmpxchg.h | 35 ++++++++++++++++++++++++++++++++
> > > > > >  1 file changed, 35 insertions(+)
> > > > > >
> > > > > > diff --git a/arch/riscv/include/asm/cmpxchg.h b/arch/riscv/include/asm/cmpxchg.h
> > > > > > index 5a07646fae65..dfb433ac544f 100644
> > > > > > --- a/arch/riscv/include/asm/cmpxchg.h
> > > > > > +++ b/arch/riscv/include/asm/cmpxchg.h
> > > > > > @@ -72,6 +72,36 @@
> > > > > >   * indicated by comparing RETURN with OLD.
> > > > > >   */
> > > > > >
> > > > > > +#define __arch_cmpxchg_mask(sc_sfx, prepend, append, r, p, o, n)       \
> > > > > > +({                                                                     \
> > > > > > +       /* Depends on 2-byte variables being 2-byte aligned */          \
> > > > > > +       ulong __s = ((ulong)(p) & 0x3) * BITS_PER_BYTE;                 \
> > > > > > +       ulong __mask = GENMASK(((sizeof(*p)) * BITS_PER_BYTE) - 1, 0)   \
> > > > > > +                       << __s;                                         \
> > > > > > +       ulong __newx = (ulong)(n) << __s;                               \
> > > > > > +       ulong __oldx = (ulong)(o) << __s;                               \
> > > > > > +       ulong __retx;                                                   \
> > > > > > +       register unsigned int __rc;                                     \
> > > > > > +                                                                       \
> > > > > > +       __asm__ __volatile__ (                                          \
> > > > > > +               prepend                                                 \
> > > > > > +               "0:     lr.w %0, %2\n"                                  \
>
>       bne  %0, %z3, 1f\n"                             \
> > > >
> > > > > bug:
> > > > > -               "       and  %0, %0, %z5\n"                             \
> > > > > -               "       bne  %0, %z3, 1f\n"                             \
> > > > > +               "       and  %1, %0, %z5\n"                             \
> > > > > +               "       bne  %1, %z3, 1f\n"                             \
> > > > > Your code breaks the %0.
> > > >
> > > > What do you mean by breaks here?
> > > >
> > > > In the end of this macro, I intended  to have __retx = (*p & __mask)
> > > > which means the value is clean to be rotated at the end of the macro
> > > > (no need to apply the mask again): r = __ret >> __s;
> > > >
> > > > Also, I assumed we are supposed to return the same variable type
> > > > as the pointer, so this is valid:
> > > > u8 a, *b, c;
> > > > a = xchg(b, c);
> > > >
> > > > Is this correct?
> > > I missed your removing "__ret & mask" at the end. So this may not the problem.
>
> It was actually the problem :)
> Even though I revised a lot, I was missing this which was made clear
> by test-running on a VM:
>
>         "0:    lr.w %0, %2\n"                    \
>         "    and  %0, %0, %z5\n"                \
>         "    bne  %0, %z3, 1f\n"                \
>         "    and  %1, %0, %z6\n"                \
>         "    or   %1, %1, %z4\n"                \
>         "    sc.w" sc_sfx " %1, %1, %2\n"            \
>         "    bnez %1, 0b\n"
>
> If I do "and %0, %0, %z5" there, the line  "and  %1, %0, %z6"  will
> output %1 = 0, which is not the intended.
> You were right! The above fix solves the issue, and I just have to
> mask it after.
>
> Sorry it took me a few days to realize this.
>
>
> > >
> > > Your patch can't boot. After chewing your code for several hours, I
> > > found a problem:
> > > diff --git a/arch/riscv/include/asm/cmpxchg.h b/arch/riscv/include/asm/cmpxchg.h
> > > index 943f094375c7..67bcce63b267 100644
> > > --- a/arch/riscv/include/asm/cmpxchg.h
> > > +++ b/arch/riscv/include/asm/cmpxchg.h
> > > @@ -14,6 +14,7 @@
> > >  #define __arch_xchg_mask(prepend, append, r, p, n)                     \
> > >  ({                                                                     \
> > >         /* Depends on 2-byte variables being 2-byte aligned */          \
> > > +       volatile ulong *__p = (ulong *)((ulong)(p) & ~0x3);             \
> >
> > Wow, I totally missed this one: I should not assume the processor
> > behavior on unaligned lr/sc.
> > Thanks for spotting this :)
> >
> > >         ulong __s = ((ulong)(p) & 0x3) * BITS_PER_BYTE;                 \
> > >         ulong __mask = GENMASK(((sizeof(*p)) * BITS_PER_BYTE) - 1, 0)   \
> > >                         << __s;                                         \
> > > @@ -29,7 +30,7 @@
> > >                "        sc.w %1, %1, %2\n"                              \
> > >                "        bnez %1, 0b\n"                                  \
> > >                append                                                   \
> > > -              : "=&r" (__retx), "=&r" (__rc), "+A" (*(p))              \
> > > +              : "=&r" (__retx), "=&r" (__rc), "+A" (*(__p))            \
> > >                : "rJ" (__newx), "rJ" (~__mask)                          \
> > >                : "memory");                                             \
> > >                                                                         \
> > > @@ -106,6 +107,7 @@
> > >  #define __arch_cmpxchg_mask(sc_sfx, prepend, append, r, p, o, n)       \
> > >  ({                                                                     \
> > >         /* Depends on 2-byte variables being 2-byte aligned */          \
> > > +       volatile ulong *__p = (ulong *)((ulong)(p) & ~0x3);             \
> > >         ulong __s = ((ulong)(p) & 0x3) * BITS_PER_BYTE;                 \
> > >         ulong __mask = GENMASK(((sizeof(*p)) * BITS_PER_BYTE) - 1, 0)   \
> > >                         << __s;                                         \
> > > @@ -125,7 +127,7 @@
> > >                 "       bnez %1, 0b\n"                                  \
> > >                 append                                                  \
> > >                 "1:\n"                                                  \
> > > -               : "=&r" (__retx), "=&r" (__rc), "+A" (*(p))             \
> > > +               : "=&r" (__retx), "=&r" (__rc), "+A" (*(__p))           \
> > >                 : "rJ" ((long)__oldx), "rJ" (__newx),                   \
> > >                   "rJ" (__mask), "rJ" (~__mask)                         \
> > >                 : "memory");                                            \
> > >
> > > But the lkvm-static still can't boot with paravirt_spinlock .... Are
> > > there any atomic tests in the Linux?
> >
> > I recall reading something about 'locktorture' or so, not sure if useful:
> >
> > https://docs.kernel.org/locking/locktorture.html
> >
> > >
> > > I found you use some "register int variables". Would it cause the problem?
> >
> > Honestly, not sure.
> > I will try inspecting the generated asm for any unexpected changes
> > compared to your version.
>
> changed type to ulong, no changes detected in asm
>
> >
> > >
> > > You can reference this file, and it has passed the lock torture test:
> > > https://github.com/guoren83/linux/blob/sg2042-master-qspinlock-64ilp32_v4/arch/riscv/include/asm/cmpxchg.h
> > >
> > > I also merged your patches with the qspinlock series: (Use the above
> > > cmpxchg.h the lkvm would run normally.)
> > > https://github.com/guoren83/linux/tree/qspinlock_v11
> >
> > Thanks!
> >
> > I should reply as soon as I find anything.
> >
>
> Some changes found (for xchg16):
> - pointer alignment wrong (need mask with ~0x3) : solved. Doesn't seem
> to need to be volatile, though.
>
> - AND with 0x3 for getting shift: I was trusting 2-byte vars to be
> 2-byte aligned , ie. p & 0x1 == 0: may not be true. Solved (it's
> simple)
>
> - on your code, IIUC you also assume ptr & 0x1 == 0, on
>         ulong *__ptr = (ulong *)((ulong)ptr & ~2);
> if it's not aligned, you may have __ptr = 0xYYYYYY01
> (Not saying it's wrong to assume this, just listing It differs on my code)
>
> - also verified your code generates addw and sllw on some points mine
> generates add and sll, don't see any issue there
>
> - There are extra sll and srl in my asm, because I end up casting to
> typeof(*ptr) at the end.
>
> changes found in cmpxchg (__cmpxchg_small_relaxed)
> - mostly the same as above,
>
> - the beginning of my function ends up being smaller, since I do my
> masking differently.
>
> I fixed the above issues, and  ran some userspace debugging tests on a
> VM, and all seems to be correct now.
>
> I then fetched your github repo, replaced my patches v3->v4 on your
> tree, added linux boot parameter "qspinlock" on my qemu cmdline
> (-machine virt): booting and working just fine.
>
> I will send a v4 soon, please give it a test when possible.
>
> Best regards,
> Leo
>
>
>
> > >
> > >
> > >
> > > >
> > > > > > +               append                                                  \
> > > > > > +               "1:\n"                                                  \
> > > > > > +               : "=&r" (__retx), "=&r" (__rc), "+A" (*(p))             \
> > > > > > +               : "rJ" ((long)__oldx), "rJ" (__newx),                   \
> > > > > > +                 "rJ" (__mask), "rJ" (~__mask)                         \
> > > > > > +               : "memory");                                            \
> > > > > > +                                                                       \
> > > > > > +       r = (__typeof__(*(p)))(__retx >> __s);                          \
> > > > > > +})
> > > > > > +
> > > > > >
> > > > > >  #define __arch_cmpxchg(lr_sfx, sc_sfx, prepend, append, r, p, co, o, n)        \
> > > > > >  ({                                                                     \
> > > > > > @@ -98,6 +128,11 @@
> > > > > >         __typeof__(*(ptr)) __ret;                                       \
> > > > > >                                                                         \
> > > > > >         switch (sizeof(*__ptr)) {                                       \
> > > > > > +       case 1:                                                         \
> > > > > > +       case 2:                                                         \
> > > > > > +               __arch_cmpxchg_mask(sc_sfx, prepend, append,            \
> > > > > > +                                       __ret, __ptr, __old, __new);    \
> > > > > > +               break;                                                  \
> > > > > >         case 4:                                                         \
> > > > > >                 __arch_cmpxchg(".w", ".w" sc_sfx, prepend, append,      \
> > > > > >                                 __ret, __ptr, (long), __old, __new);    \
> > > > > > --
> > > > > > 2.41.0
> > > > > >
> > > > >
> > > > >
> > > > > --
> > > > > Best Regards
> > > > >  Guo Ren
> > > > >
> > > >
> > >
> > >
> > > --
> > > Best Regards
> > >  Guo Ren
> > >