bpf: Disable preemption when increasing per-cpu map_locked

Per-cpu htab->map_locked is used to prohibit the concurrent accesses
from both NMI and non-NMI contexts. But since commit 74d862b6
("sched: Make migrate_disable/enable() independent of RT"),
migrate_disable() is also preemptible under CONFIG_PREEMPT case, so now
map_locked also disallows concurrent updates from normal contexts
(e.g. userspace processes) unexpectedly as shown below:

process A                      process B

htab_map_update_elem()
  htab_lock_bucket()
    migrate_disable()
    /* return 1 */
    __this_cpu_inc_return()
    /* preempted by B */

                               htab_map_update_elem()
                                 /* the same bucket as A */
                                 htab_lock_bucket()
                                   migrate_disable()
                                   /* return 2, so lock fails */
                                   __this_cpu_inc_return()
                                   return -EBUSY

A fix that seems feasible is using in_nmi() in htab_lock_bucket() and
only checking the value of map_locked for nmi context. But it will
re-introduce dead-lock on bucket lock if htab_lock_bucket() is re-entered
through non-tracing program (e.g. fentry program).

One cannot use preempt_disable() to fix this issue as htab_use_raw_lock
being false causes the bucket lock to be a spin lock which can sleep and
does not work with preempt_disable().

Therefore, use migrate_disable() when using the spinlock instead of
preempt_disable() and defer fixing concurrent updates to when the kernel
has its own BPF memory allocator.

Fixes: 74d862b6 ("sched: Make migrate_disable/enable() independent of RT")
Reviewed-by: Hao Luo <haoluo@google.com>
Signed-off-by: Hou Tao <houtao1@huawei.com>
Link: https://lore.kernel.org/r/20220831042629.130006-2-houtao@huaweicloud.com


Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>