io_setup() makes the ring at least 4 events per possible CPU of the
node. The ring is mapped into the task memory, and CRIU restores it with
io_setup() and moves the new mapping to the address of the old one, so
it needs a ring of the same size. On a node with more possible CPUs it
gets a bigger ring, and the restore fails. E.g. io_setup(1) makes a
ring of one page, 127 events, on 4 CPUs, and of two pages, 255 events,
on 16 CPUs.

Skip the CPU heuristic while the VE is pseudosuper, which it is almost
only for the restore. Keep the other adjustments, the doubling and the
rounding up to pages: they don't depend on the node, and the queue depth
CRIU estimates from the dumped ring already offsets them, so the restore
gets the ring of the dumped size.

A ring restored below the minimum of this node doesn't use the per-CPU
caches of free slots: it has too few slots to spare for them.

https://virtuozzo.atlassian.net/browse/VSTOR-154129

Feature: ve: CRIU support
Signed-off-by: Pavel Tikhomirov <[email protected]>
---
 fs/aio.c | 21 ++++++++++++++++++++-
 1 file changed, 20 insertions(+), 1 deletion(-)

diff --git a/fs/aio.c b/fs/aio.c
index 901cb622fe2c4..2ff1758507611 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -109,6 +109,7 @@ struct kioctx {
         * counter at a time:
         */
        unsigned                req_batch;
+       bool                    no_percpu_reqs;
        /*
         * This is what userspace passed to io_setup(), it's not used for
         * anything but counting against the global max_reqs quota.
@@ -745,6 +746,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
        struct mm_struct *mm = current->mm;
        struct kioctx *ctx;
        struct ve_struct *ve = get_exec_env();
+       bool no_percpu_reqs = false;
        int err = -ENOMEM;
 
        /*
@@ -761,8 +763,16 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
         * and unavailable, double nr_events so userspace sees what they
         * expected: additionally, we move req_batch slots to/from percpu
         * counters at a time, so make sure that isn't 0:
+        *
+        * CRIU restores a ring with io_setup() and needs the size it had on
+        * dump, which can be below the minimum of this node due to the CPU
+        * heuristic below, so skip the heuristic on restore. Such a ring
+        * can't spare slots for the percpu counters, so it doesn't use them.
         */
-       nr_events = max(nr_events, num_possible_cpus() * 4);
+       if (ve_is_super(ve) || !ve->is_pseudosuper)
+               nr_events = max(nr_events, num_possible_cpus() * 4);
+       else if (nr_events < num_possible_cpus() * 4)
+               no_percpu_reqs = true;
        nr_events *= 2;
 
        /* Prevent overflows */
@@ -779,6 +789,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
                return ERR_PTR(-ENOMEM);
 
        ctx->max_reqs = max_reqs;
+       ctx->no_percpu_reqs = no_percpu_reqs;
        ctx->ve = get_ve(ve);
 
        spin_lock_init(&ctx->ctx_lock);
@@ -950,6 +961,11 @@ static void put_reqs_available(struct kioctx *ctx, 
unsigned nr)
        struct kioctx_cpu *kcpu;
        unsigned long flags;
 
+       if (ctx->no_percpu_reqs) {
+               atomic_add(nr, &ctx->reqs_available);
+               return;
+       }
+
        local_irq_save(flags);
        kcpu = this_cpu_ptr(ctx->cpu);
        kcpu->reqs_available += nr;
@@ -968,6 +984,9 @@ static bool __get_reqs_available(struct kioctx *ctx)
        bool ret = false;
        unsigned long flags;
 
+       if (ctx->no_percpu_reqs)
+               return atomic_dec_if_positive(&ctx->reqs_available) >= 0;
+
        local_irq_save(flags);
        kcpu = this_cpu_ptr(ctx->cpu);
        if (!kcpu->reqs_available) {

-- 
2.56.0

_______________________________________________
Devel mailing list
[email protected]
https://lists.openvz.org/mailman/listinfo/devel

Reply via email to