The current way we deal with PtrAuth is a bit heavy handed:

- We forcefully save the host's keys on each vcpu_load()
- Handling the PtrAuth trap forces us to go all the way back
  to the exit handling code to just set the HCR bits

Overall, this is pretty cumbersome. A better approach would be
to handle it the same way we deal with the FPSIMD registers:

- On vcpu_load() disable PtrAuth for the guest
- On first use, save the host's keys, enable PtrAuth in the
  guest

Crucially, this can happen as a fixup, which is done very early
on exit. We can then reenter the guest immediately without
leaving the hypervisor role.

Another thing is that it simplify the rest of the host handling:
exiting all the way to the host means that the only possible
outcome for this trap is to inject an UNDEF.
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

When using the PtrAuth feature in a guest, we need to save the host's
keys before allowing the guest to program them. For that, we dump
them in a per-CPU data structure (the so called host context).

But both call sites that do this are in preemptible context,
which may end up in disaster should the vcpu thread get preempted
before reentering the guest.

Instead, save the keys eagerly on each vcpu_load(). This has an
increased overhead, but is at least safe.

Cc: stable@vger.kernel.org
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

On a system with FWB, we don't need to unmap Stage-2 on reboot,
as even if userspace takes this opportunity to repaint the whole
of memory, FWB ensures that the data side stays consistent even
if the guest uses non-cacheable mappings.

However, the I-side is not necessarily coherent with the D-side
if CTR_EL0.DIC is 0. In this case, invalidate the i-cache to
preserve coherency.
Reported-by: Alexandru Elisei <alexandru.elisei@arm.com>
Reviewed-by: Alexandru Elisei <alexandru.elisei@arm.com>
Fixes: 892713e9 ("KVM: arm64: Sidestep stage2_unmap_vm() on vcpu reset when S2FWB is supported")
Signed-off-by: Marc Zyngier <maz@kernel.org>

The general comment about keeping the enum order in sync
with the save/restore code has been obsolete for many years now.

Just drop it.

Note that there are other ordering requirements in the enum,
such as the PtrAuth and PMU registers, which are still valid.
Reported-by: James Morse <james.morse@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

We currently assume that an exception is delivered to EL1, always.
Once we emulate EL2, this no longer will be the case. To prepare
for this, add a target_mode parameter.

While we're at it, merge the computing of the target PC and PSTATE in
a single function that updates both PC and CPSR after saving their
previous values in the corresponding ELR/SPSR. This ensures that they
are updated in the correct order (a pretty common source of bugs...).
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

Keeping empty structure as the vcpu state initializer is slightly
wasteful: we only want to set pstate, and zero everything else.
Just do that.
Signed-off-by: Marc Zyngier <maz@kernel.org>

Our sysreg reset check has become a bit silly, as it only checks whether
a reset callback actually exists for a given sysreg entry, and apply the
method if available. Doing the check at each vcpu reset is pretty dumb,
as the tables never change. It is thus perfectly possible to do the same
checks at boot time.

This also allows us to introduce a sparse sys_regs[] array, something
that will be required with ARMv8.4-NV.
Signed-off-by: Marc Zyngier <maz@kernel.org>

As we're about to become a bit more harsh when it comes to the lack of
reset callbacks, let's add the missing PMU reset handlers. Note that
these only cover *CLR registers that were always covered by their *SET
counterpart, so there is no semantic change here.
Reviewed-by: James Morse <james.morse@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

Extract the direct HW accessors for later reuse.
Reviewed-by: James Morse <james.morse@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

If we move the used_lrs field to the version-specific cpu interface
structure, the following functions only operate on the struct
vgic_v3_cpu_if and not the full vcpu:

  __vgic_v3_save_state
  __vgic_v3_restore_state
  __vgic_v3_activate_traps
  __vgic_v3_deactivate_traps
  __vgic_v3_save_aprs
  __vgic_v3_restore_aprs

This is going to be very useful for nested virt, so move the used_lrs
field and change the prototypes and implementations of these functions to
take the cpu_if parameter directly.

No functional change.
Reviewed-by: James Morse <james.morse@arm.com>
Signed-off-by: Christoffer Dall <christoffer.dall@arm.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

This abstraction was introduced to hide the difference between arm and
arm64 but, with the former no longer supported, this abstraction can be
removed and the canonical kernel API used directly instead.
Signed-off-by: Andrew Scull <ascull@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
CC: Marc Zyngier <maz@kernel.org>
CC: James Morse <james.morse@arm.com>
CC: Suzuki K Poulose <suzuki.poulose@arm.com>
Link: https://lore.kernel.org/r/20200519104036.259917-1-ascull@google.com

The comment used to say that kvm_get_hyp_vector is only called on VHE systems.
In fact, it is also called from the nVHE init function cpu_init_hyp_mode().
Fix the comment to stop confusing devs.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200515152550.83810-1-dbrazdil@google.com

Pull bits of code to the only place where it is used. Remove empty function
__cpu_init_stage2(). Remove redundant has_vhe() check since this function is
nVHE-only. No functional changes intended.
Signed-off-by: David Brazdil <dbrazdil@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200515152056.83158-1-dbrazdil@google.com

KVM_CAP_MAX_VCPUS always return the maximum possible number of
VCPUs, irrespective of the selected interrupt controller. This
is pretty misleading for userspace that selects a GICv2 on a GICv3
system that supports v2 compat: It always gets a maximum of 512
VCPUs, even if the effective limit is 8. The 9th VCPU will fail
to be created, which is unexpected as far as userspace is concerned.

Fortunately, we already have the right information stashed in the
kvm structure, and we can return it as requested.
Reported-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Tested-by: Alexandru Elisei <alexandru.elisei@arm.com>
Reviewed-by: Alexandru Elisei <alexandru.elisei@arm.com>
Link: https://lore.kernel.org/r/20200427141507.284985-1-maz@kernel.org

There is already support of enabling dirty log gradually in small chunks
for x86 in commit 3c9bd400 ("KVM: x86: enable dirty log gradually in
small chunks"). This adds support for arm64.

x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
is enabled. However, for arm64, both huge pages and normal pages can be
write protected gradually by userspace.

Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
Linux VMs with different page size. The memory pressure is 127G in each
case. The time taken of memory_global_dirty_log_start in QEMU is listed
below:

Page Size      Before    After Optimization
  4K            650ms         1.8ms
  2M             4ms          1.8ms
  1G             2ms          1.8ms

Besides the time reduction, the biggest improvement is that we will minimize
the performance side effect (because of dissolving huge pages and marking
memslots dirty) on guest after enabling dirty log.
Signed-off-by: Keqian Zhu <zhukeqian1@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200413122023.52583-1-zhukeqian1@huawei.com

We support mapping host memory backed by PMD transparent hugepages
at stage2 as huge pages. However the checks are now spread across
two different places. Let us unify the handling of the THPs to
keep the code cleaner (and future proof for PUD THP support).
This patch moves transparent_hugepage_adjust() closer to the caller
to avoid a forward declaration for fault_supports_stage2_huge_mappings().

Also, since we already handle the case where the host VA and the guest
PA may not be aligned, the explicit VM_BUG_ON() is not required.
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200507123546.1875-3-yuzenghui@huawei.com

If we are checking whether the stage2 can map PAGE_SIZE,
we don't have to do the boundary checks as both the host
VMA and the guest memslots are page aligned. Bail the case
easily.

While we're at it, fixup a typo in the comment below.
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200507123546.1875-2-yuzenghui@huawei.com

Do cond_resched_lock() in stage2_flush_memslot() like what is done in
unmap_stage2_range() and other places holding mmu_lock while processing
a possibly large range of memory.
Signed-off-by: Jiang Yi <giangyi@amazon.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Link: https://lore.kernel.org/r/20200415084229.29992-1-giangyi@amazon.com

stage2_unmap_vm() was introduced to unmap user RAM region in the stage2
page table to make the caches coherent. E.g., a guest reboot with stage1
MMU disabled will access memory using non-cacheable attributes. If the
RAM and caches are not coherent at this stage, some evicted dirty cache
line may go and corrupt guest data in RAM.

Since ARMv8.4, S2FWB feature is mandatory and KVM will take advantage
of it to configure the stage2 page table and the attributes of memory
access. So we ensure that guests always access memory using cacheable
attributes and thus, the caches always be coherent.

So on CPUs that support S2FWB, we can safely reset the vcpu without a
heavy stage2 unmapping.
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200415072835.1164-1-yuzenghui@huawei.com

Fix spelling and typos (e.g., repeated words) in comments.
Signed-off-by: Fuad Tabba <tabba@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200401140310.29701-1-tabba@google.com

By the time we start using the has_vhe() helper, we have long
discovered whether we are running VHE or not. It thus makes
sense to use cpus_have_final_cap() instead of cpus_have_const_cap(),
which leads to a small text size reduction.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Acked-by: David Brazdil <dbrazdil@google.com>
Link: https://lore.kernel.org/r/20200513103828.74580-1-maz@kernel.org

Now that this function isn't constrained by the 32bit PCS,
let's simplify it by taking a single 64bit offset instead
of two 32bit parameters.
Signed-off-by: Marc Zyngier <maz@kernel.org>

Consolidate references to the CONFIG_KVM configuration item to encompass
entire folders rather than per line.
Signed-off-by: Fuad Tabba <tabba@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20200505154520.194120-5-tabba@google.com

Changing CONFIG_KVM to be a 'menuconfig' entry in Kconfig mean that we
can straightforwardly enumerate optional features, such as the virtual
PMU device as dependent options.
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Fuad Tabba <tabba@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200505154520.194120-4-tabba@google.com

arm64 KVM supports 16k pages since 02e0b760
("arm64: kvm: Add support for 16K pages"), so update the Kconfig help
text accordingly.
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Fuad Tabba <tabba@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200505154520.194120-3-tabba@google.com

CONFIG_KVM_ARM_HOST is just a proxy for CONFIG_KVM, so remove it in favour
of the latter.
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Fuad Tabba <tabba@google.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200505154520.194120-2-tabba@google.com

Now that the 32bit KVM/arm host is a distant memory, let's move the
whole of the KVM/arm64 code into the arm64 tree.

As they said in the song: Welcome Home (Sanitarium).
Signed-off-by: Marc Zyngier <maz@kernel.org>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20200513104034.74741-1-maz@kernel.org

Pull x86 fixes from Thomas Gleixner:
 "A set of fixes for x86:

   - Ensure that direct mapping alias is always flushed when changing
     page attributes. The optimization for small ranges failed to do so
     when the virtual address was in the vmalloc or module space.

   - Unbreak the trace event registration for syscalls without arguments
     caused by the refactoring of the SYSCALL_DEFINE0() macro.

   - Move the printk in the TSC deadline timer code to a place where it
     is guaranteed to only be called once during boot and cannot be
     rearmed by clearing warn_once after boot. If it's invoked post boot
     then lockdep rightfully complains about a potential deadlock as the
     calling context is different.

   - A series of fixes for objtool and the ORC unwinder addressing
     variety of small issues:

       - Stack offset tracking for indirect CFAs in objtool ignored
         subsequent pushs and pops

       - Repair the unwind hints in the register clearing entry ASM code

       - Make the unwinding in the low level exit to usermode code stop
         after switching to the trampoline stack. The unwind hint is no
         longer valid and the ORC unwinder emits a warning as it can't
         find the registers anymore.

       - Fix unwind hints in switch_to_asm() and rewind_stack_do_exit()
         which caused objtool to generate bogus ORC data.

       - Prevent unwinder warnings when dumping the stack of a
         non-current task as there is no way to be sure about the
         validity because the dumped stack can be a moving target.

       - Make the ORC unwinder behave the same way as the frame pointer
         unwinder when dumping an inactive tasks stack and do not skip
         the first frame.

       - Prevent ORC unwinding before ORC data has been initialized

       - Immediately terminate unwinding when a unknown ORC entry type
         is found.

       - Prevent premature stop of the unwinder caused by IRET frames.

       - Fix another infinite loop in objtool caused by a negative
         offset which was not catched.

       - Address a few build warnings in the ORC unwinder and add
         missing static/ro_after_init annotations"

* tag 'x86-urgent-2020-05-10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86/unwind/orc: Move ORC sorting variables under !CONFIG_MODULES
  x86/apic: Move TSC deadline timer debug printk
  ftrace/x86: Fix trace event registration for syscalls without arguments
  x86/mm/cpa: Flush direct map alias during cpa
  objtool: Fix infinite loop in for_offset_range()
  x86/unwind/orc: Fix premature unwind stoppage due to IRET frames
  x86/unwind/orc: Fix error path for bad ORC entry type
  x86/unwind/orc: Prevent unwinding before ORC initialization
  x86/unwind/orc: Don't skip the first frame for inactive tasks
  x86/unwind: Prevent false warnings for non-current tasks
  x86/unwind/orc: Convert global variables to static
  x86/entry/64: Fix unwind hints in rewind_stack_do_exit()
  x86/entry/64: Fix unwind hints in __switch_to_asm()
  x86/entry/64: Fix unwind hints in kernel exit path
  x86/entry/64: Fix unwind hints in register clearing code
  objtool: Fix stack offset tracking for indirect CFAs

Pull objtool fix from Thomas Gleixner:
 "A single fix for objtool to prevent an infinite loop in the
  jump table search which can be triggered when building the
  kernel with '-ffunction-sections'"

* tag 'objtool-urgent-2020-05-10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  objtool: Fix infinite loop in find_jump_table()

Pull locking fix from Thomas Gleixner:
 "A single fix for the fallout of the recent futex uacess rework.

  With those changes GCC9 fails to analyze arch_futex_atomic_op_inuser()
  correctly and emits a 'maybe unitialized' warning. While we usually
  ignore compiler stupidity the conditional store is pointless anyway
  because the correct case has to store. For the fault case the extra
  store does no harm"

* tag 'locking-urgent-2020-05-10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  ARM: futex: Address build warning

Pull iommu fixes from Joerg Roedel:

 - Race condition fixes for the AMD IOMMU driver.

   These are five patches fixing two race conditions around
   increase_address_space(). The first race condition was around the
   non-atomic update of the domain page-table root pointer and the
   variable containing the page-table depth (called mode). This is fixed
   now be merging page-table root and mode into one 64-bit field which
   is read/written atomically.

   The second race condition was around updating the page-table root
   pointer and making it public before the hardware caches were flushed.
   This could cause addresses to be mapped and returned to drivers which
   are not reachable by IOMMU hardware yet, causing IO page-faults. This
   is fixed too by adding the necessary flushes before a new page-table
   root is published.

   Related to the race condition fixes these patches also add a missing
   domain_flush_complete() barrier to update_domain() and a fix to bail
   out of the loop which tries to increase the address space when the
   call to increase_address_space() fails.

   Qian was able to trigger the race conditions under high load and
   memory pressure within a few days of testing. He confirmed that he
   has seen no issues anymore with the fixes included here.

 - Fix for a list-handling bug in the VirtIO IOMMU driver.

* tag 'iommu-fixes-v5.7-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu:
  iommu/virtio: Reverse arguments to list_add
  iommu/amd: Do not flush Device Table in iommu_map_page()
  iommu/amd: Update Device Table in increase_address_space()
  iommu/amd: Call domain_flush_complete() in update_domain()
  iommu/amd: Do not loop forever when trying to increase address space
  iommu/amd: Fix race in increase_address_space()/fetch_pte()

Pull block fixes from Jens Axboe:

 - a small series fixing a use-after-free of bdi name (Christoph,Yufen)

 - NVMe fix for a regression with the smaller CQ update (Alexey)

 - NVMe fix for a hang at namespace scanning error recovery (Sagi)

 - fix race with blk-iocost iocg->abs_vdebt updates (Tejun)

* tag 'block-5.7-2020-05-09' of git://git.kernel.dk/linux-block:
  nvme: fix possible hang when ns scanning fails during error recovery
  nvme-pci: fix "slimmer CQ head update"
  bdi: add a ->dev_name field to struct backing_dev_info
  bdi: use bdi_dev_name() to get device name
  bdi: move bdi_dev_name out of line
  vboxsf: don't use the source name in the bdi name
  iocost: protect iocg->abs_vdebt with iocg->waitq.lock

It seems that for whatever reason, gcc-10 ends up not inlining a couple
of functions that used to be inlined before.  Even if they only have one
single callsite - it looks like gcc may have decided that the code was
unlikely, and not worth inlining.

The code generation difference is harmless, but caused a few new section
mismatch errors, since the (now no longer inlined) function wasn't in
the __init section, but called other init functions:

   Section mismatch in reference from the function kexec_free_initrd() to the function .init.text:free_initrd_mem()
   Section mismatch in reference from the function tpm2_calc_event_log_size() to the function .init.text:early_memremap()
   Section mismatch in reference from the function tpm2_calc_event_log_size() to the function .init.text:early_memunmap()

So add the appropriate __init annotation to make modpost not complain.
In both cases there were trivially just a single callsite from another
__init function.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Pull RISC-V fixes from Palmer Dabbelt:
 "A smattering of fixes and cleanups:

   - Dead code removal.

   - Exporting riscv_cpuid_to_hartid_mask for modules.

   - Per-CPU tracking of ISA features.

   - Setting max_pfn correctly when probing memory.

   - Adding a note to the VDSO so glibc can check the kernel's version
     without a uname().

   - A fix to force the bootloader to initialize the boot spin tables,
     which still get used as a fallback when SBI-0.1 is enabled"

* tag 'riscv-for-linus-5.7-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/riscv/linux:
  RISC-V: Remove unused code from STRICT_KERNEL_RWX
  riscv: force __cpu_up_ variables to put in data section
  riscv: add Linux note to vdso
  riscv: set max_pfn to the PFN of the last page
  RISC-V: Remove N-extension related defines
  RISC-V: Add bitmap reprensenting ISA features common across CPUs
  RISC-V: Export riscv_cpuid_to_hartid_mask() API

gcc-10 has started warning about conflicting types for a few new
built-in functions, particularly 'free()'.

This results in warnings like:

   crypto/xts.c:325:13: warning: conflicting types for built-in function ‘free’; expected ‘void(void *)’ [-Wbuiltin-declaration-mismatch]

because the crypto layer had its local freeing functions called
'free()'.

Gcc-10 is in the wrong here, since that function is marked 'static', and
thus there is no chance of confusion with any standard library function
namespace.

But the simplest thing to do is to just use a different name here, and
avoid this gcc mis-feature.

[ Side note: gcc knowing about 'free()' is in itself not the
  mis-feature: the semantics of 'free()' are special enough that a
  compiler can validly do special things when seeing it.

  So the mis-feature here is that gcc thinks that 'free()' is some
  restricted name, and you can't shadow it as a local static function.

  Making the special 'free()' semantics be a function attribute rather
  than tied to the name would be the much better model ]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

gcc-10 now warns about passing aliasing pointers to functions that take
restricted pointers.

That's actually a great warning, and if we ever start using 'restrict'
in the kernel, it might be quite useful.  But right now we don't, and it
turns out that the only thing this warns about is an idiom where we have
declared a few functions to be "printf-like" (which seems to make gcc
pick up the restricted pointer thing), and then we print to the same
buffer that we also use as an input.

And people do that as an odd concatenation pattern, with code like this:

    #define sysfs_show_gen_prop(buffer, fmt, ...) \
        snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)

where we have 'buffer' as both the destination of the final result, and
as the initial argument.

Yes, it's a bit questionable.  And outside of the kernel, people do have
standard declarations like

    int snprintf( char *restrict buffer, size_t bufsz,
                  const char *restrict format, ... );

where that output buffer is marked as a restrict pointer that cannot
alias with any other arguments.

But in the context of the kernel, that 'use snprintf() to concatenate to
the end result' does work, and the pattern shows up in multiple places.
And we have not marked our own version of snprintf() as taking restrict
pointers, so the warning is incorrect for now, and gcc picks it up on
its own.

If we do start using 'restrict' in the kernel (and it might be a good
idea if people find places where it matters), we'll need to figure out
how to avoid this issue for snprintf and friends.  But in the meantime,
this warning is not useful.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

This is the final array bounds warning removal for gcc-10 for now.

Again, the warning is good, and we should re-enable all these warnings
when we have converted all the legacy array declaration cases to
flexible arrays. But in the meantime, it's just noise.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

When the controller is reconnecting, the host fails I/O and admin
commands as the host cannot reach the controller. ns scanning may
revalidate namespaces during that period and it is wrong to remove
namespaces due to these failures as we may hang (see 205da243).

One command that may fail is nvme_identify_ns_descs. Since we return
success due to having ns identify descriptor list optional, we continue
to compare ns identifiers in nvme_revalidate_disk, obviously fail and
return -ENODEV to nvme_validate_ns, which will remove the namespace.

Exactly what we don't want to happen.

Fixes: 22802bf7 ("nvme: Namepace identification descriptor list is optional")
Tested-by: Anton Eidelman <anton@lightbitslabs.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Pre-incrementing ->cq_head can't be done in memory because OOB value
can be observed by another context.

This devalues space savings compared to original code :-\

	$ ./scripts/bloat-o-meter ../vmlinux-000 ../obj/vmlinux
	add/remove: 0/0 grow/shrink: 0/4 up/down: 0/-32 (-32)
	Function                                     old     new   delta
	nvme_poll_irqdisable                         464     456      -8
	nvme_poll                                    455     447      -8
	nvme_irq                                     388     380      -8
	nvme_dev_disable                             955     947      -8

But the code is minimal now: one read for head, one read for q_depth,
one increment, one comparison, single instruction phase bit update and
one write for new head.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Reported-by: John Garry <john.garry@huawei.com>
Tested-by: John Garry <john.garry@huawei.com>
Fixes: e2a366a4 ("nvme-pci: slimmer CQ head update")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>