Andreas reported that commit ee0a49a6 ("powerpc/uaccess: Switch
__put_user_size_allowed() to __put_user_asm_goto()") broke
CLONE_CHILD_SETTID.

Further inspection showed that the put_user() in schedule_tail() was
missing entirely, the store not emitted by the compiler.

  <.schedule_tail>:
    mflr    r0
    std     r0,16(r1)
    stdu    r1,-112(r1)
    bl      <.finish_task_switch>
    ld      r9,2496(r3)
    cmpdi   cr7,r9,0
    bne     cr7,<.schedule_tail+0x60>
    ld      r3,392(r13)
    ld      r9,1392(r3)
    cmpdi   cr7,r9,0
    beq     cr7,<.schedule_tail+0x3c>
    li      r4,0
    li      r5,0
    bl      <.__task_pid_nr_ns>
    nop
    bl      <.calculate_sigpending>
    nop
    addi    r1,r1,112
    ld      r0,16(r1)
    mtlr    r0
    blr
    nop
    nop
    nop
    bl      <.__balance_callback>
    b       <.schedule_tail+0x1c>

Notice there are no stores other than to the stack. There should be a
stw in there for the store to current->set_child_tid.

This is only seen with GCC 4.9 era compilers (tested with 4.9.3 and
4.9.4), and only when CONFIG_PPC_KUAP is disabled.

When CONFIG_PPC_KUAP=y, the inline asm that's part of the isync()
and mtspr() inlined via allow_user_access() seems to be enough to
avoid the bug.

We already have a macro to work around this (or a similar bug), called
asm_volatile_goto which includes an empty asm block to tickle the
compiler into generating the right code. So use that.

With this applied the code generation looks more like it will work:

  <.schedule_tail>:
    mflr    r0
    std     r31,-8(r1)
    std     r0,16(r1)
    stdu    r1,-144(r1)
    std     r3,112(r1)
    bl      <._mcount>
    nop
    ld      r3,112(r1)
    bl      <.finish_task_switch>
    ld      r9,2624(r3)
    cmpdi   cr7,r9,0
    bne     cr7,<.schedule_tail+0xa0>
    ld      r3,2408(r13)
    ld      r31,1856(r3)
    cmpdi   cr7,r31,0
    beq     cr7,<.schedule_tail+0x80>
    li      r4,0
    li      r5,0
    bl      <.__task_pid_nr_ns>
    nop
    li      r9,-1
    clrldi  r9,r9,12
    cmpld   cr7,r31,r9
    bgt     cr7,<.schedule_tail+0x80>
    lis     r9,16
    rldicr  r9,r9,32,31
    subf    r9,r31,r9
    cmpldi  cr7,r9,3
    ble     cr7,<.schedule_tail+0x80>
    li      r9,0
    stw     r3,0(r31)				<-- stw
    nop
    bl      <.calculate_sigpending>
    nop
    addi    r1,r1,144
    ld      r0,16(r1)
    ld      r31,-8(r1)
    mtlr    r0
    blr
    nop
    bl      <.__balance_callback>
    b       <.schedule_tail+0x30>

Fixes: ee0a49a6 ("powerpc/uaccess: Switch __put_user_size_allowed() to __put_user_asm_goto()")
Reported-by: Andreas Schwab <schwab@linux-m68k.org>
Tested-by: Andreas Schwab <schwab@linux-m68k.org>
Suggested-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201104111742.672142-1-mpe@ellerman.id.au