Message-ID: <20171017155318.5rcoprfwxs27k2qq@linux-n805>
Date:   Tue, 17 Oct 2017 08:53:18 -0700
From:   Davidlohr Bueso <dave@...olabs.net>
To:     Jason Baron <jbaron@...mai.com>
Cc:     Waiman Long <longman@...hat.com>, akpm@...ux-foundation.org,
        Alexander Viro <viro@...iv.linux.org.uk>,
        Jan Kara <jack@...e.com>,
        Jeff Layton <jlayton@...chiereds.net>,
        "J. Bruce Fields" <bfields@...ldses.org>,
        Tejun Heo <tj@...nel.org>,
        Christoph Lameter <cl@...ux-foundation.org>,
        linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org,
        Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Andi Kleen <andi@...stfloor.org>,
        Dave Chinner <dchinner@...hat.com>,
        Boqun Feng <boqun.feng@...il.com>
Subject: Re: [PATCH v7 7/6] fs/epoll: scale nested callbacks

On Mon, 16 Oct 2017, Jason Baron wrote:

>Hi,
>
>I posted a patch to completely remove the lock here from the
>ep_poll_safewake() patch here:
>
>http://lkml.iu.edu/hypermail/linux/kernel/1710.1/05834.html

Kernel development never ceases to amaze me. Two complementary
fixes to a 8+ y/o performance issue on the same day - not that
nested epolls are that common, but it also comes from two real
workloads...

Getting rid of the lock altogether is always the best way.

>
>So these are going to conflict. The reason its safe to remove the lock
>is that there are loop and depth checks now that are performed during
>EPOLL_CTL_ADD. Specifically, ep_loop_check(). I would prefer to these
>checks once add add time as opposed to at each wakeup (even if they can
>be scaled better).

Wrt conflicts, no worries, I'll rebase -- and hopefully we can get
the dlock stuff in for v4.15 as well.

>
>I also have a pending patch to do something similar for
>poll_readywalk_ncalls, but I think that poll_safewake_ncalls is the most
>egregious one here?

The customer's workload issues are for the loop_ncalls and readywalk_ncalls
lists, so I'd be interested in seeing your patch for the later. The reason
your patch above is likely not to help my scenario is that most of the time
is spent at a dispatcher level doing epoll_wait, not too many EPOLL_CTL_ADDs
going on afaict.

Thanks,
Davidlohr

>
>Thanks,
>
>-Jason
>
>On 10/13/2017 11:45 AM, Davidlohr Bueso wrote:
>> A customer reported massive contention on the ncalls->lock in which
>> the workload is designed around nested epolls (where the fd is already
>> an epoll).
>>
>> 83.49%  [kernel]               [k] __pv_queued_spin_lock_slowpath
>>  2.70%  [kernel]               [k] ep_call_nested.constprop.13
>>  1.94%  [kernel]               [k] _raw_spin_lock_irqsave
>>  1.83%  [kernel]               [k]
>> __raw_callee_save___pv_queued_spin_unlock
>>  1.45%  [kernel]               [k] _raw_spin_unlock_irqrestore
>>  0.41%  [kernel]               [k] ep_scan_ready_list.isra.8
>>  0.36%  [kernel]               [k] pvclock_clocksource_read
>>
>> The application running on kvm, is using a shared memory IPC communication
>> with a pipe wakeup mechanism, and a two level dispatcher both built around
>> 'epoll_wait'. There is one process per physical core and a full mesh of
>> pipes
>> between them, so on a 18 core system (the actual case), there are 18*18
>> pipes
>> for the IPCs alone.
>>
>> This patch proposes replacing the nested calls global linked list with a
>> dlock
>> interface, for which we can benefit from pcpu lists when doing
>> ep_poll_safewake(),
>> and hashing for the current task, which provides two benefits:
>>
>> 1. Speeds up the process of loop and max-depth checks from O(N) lookups
>> to O(1)
>>   (albeit possible collisions, which we have to iterate); and,
>>
>> 2. Provides smaller lock granularity.
>>
>> cpus    before        after       diff
>> 1    1409370        1344804     -4.58%
>> 2    1015690        1337053     31.63%
>> 3     721009        1273254     76.59%
>> 4     380959        1128931    196.33%
>> 5     287603        1028362    257.56%
>> 6     221104         894943    304.76%
>> 7     173592         976395    462.46%
>> 8     145860         922584    532.51%
>> 9     127877         925900    624.05%
>> 10     112603         791456    602.87%
>> 11      97926         724296    639.63%
>> 12      80732         730485    804.82%
>>
>> With the exception of a single cpu, where the overhead of jhashing
>> influences), we
>> get some pretty good raw throughput increase. Similarly, the amount of
>> time spent
>> decreases immensely as well.
>>
>> Passes ltp related testcases.
>>
>> Signed-off-by: Davidlohr Bueso <dbueso@...e.de>
>> ---
>> fs/eventpoll.c | 88
>> +++++++++++++++++++++++++++++++++++-----------------------
>> 1 file changed, 53 insertions(+), 35 deletions(-)
>>
>> diff --git a/fs/eventpoll.c b/fs/eventpoll.c
>> index 2fabd19cdeea..675c97fdc5da 100644
>> --- a/fs/eventpoll.c
>> +++ b/fs/eventpoll.c
>> @@ -22,7 +22,7 @@
>> #include <linux/slab.h>
>> #include <linux/poll.h>
>> #include <linux/string.h>
>> -#include <linux/list.h>
>> +#include <linux/dlock-list.h>
>> #include <linux/hash.h>
>> #include <linux/spinlock.h>
>> #include <linux/syscalls.h>
>> @@ -119,7 +119,7 @@ struct epoll_filefd {
>>  * and loop cycles.
>>  */
>> struct nested_call_node {
>> -    struct list_head llink;
>> +    struct dlock_list_node llink;
>>     void *cookie;
>>     void *ctx;
>> };
>> @@ -129,8 +129,7 @@ struct nested_call_node {
>>  * maximum recursion dept and loop cycles.
>>  */
>> struct nested_calls {
>> -    struct list_head tasks_call_list;
>> -    spinlock_t lock;
>> +    struct dlock_list_heads tasks_call_list;
>> };
>>
>> /*
>> @@ -371,10 +370,14 @@ static inline int ep_op_has_event(int op)
>> }
>>
>> /* Initialize the poll safe wake up structure */
>> -static void ep_nested_calls_init(struct nested_calls *ncalls)
>> +static inline int ep_nested_calls_init(struct nested_calls *ncalls)
>> +{
>> +    return alloc_dlock_list_heads(&ncalls->tasks_call_list, true);
>> +}
>> +
>> +static inline void ep_nested_calls_free(struct nested_calls *ncalls)
>> {
>> -    INIT_LIST_HEAD(&ncalls->tasks_call_list);
>> -    spin_lock_init(&ncalls->lock);
>> +    free_dlock_list_heads(&ncalls->tasks_call_list);
>> }
>>
>> /**
>> @@ -483,45 +486,50 @@ static int ep_call_nested(struct nested_calls
>> *ncalls, int max_nests,
>> {
>>     int error, call_nests = 0;
>>     unsigned long flags;
>> -    struct list_head *lsthead = &ncalls->tasks_call_list;
>> -    struct nested_call_node *tncur;
>> -    struct nested_call_node tnode;
>> +    struct dlock_list_head *head;
>> +    struct nested_call_node *tncur, tnode;
>>
>> -    spin_lock_irqsave(&ncalls->lock, flags);
>> +    head = dlock_list_hash(&ncalls->tasks_call_list, ctx);
>>
>> +    spin_lock_irqsave(&head->lock, flags);
>>     /*
>>      * Try to see if the current task is already inside this wakeup call.
>> -     * We use a list here, since the population inside this set is always
>> -     * very much limited.
>> +     *
>> +     * We use a chained hash table with linked lists here, and take the
>> +     * lock to avoid racing when collisions (where ctx pointer is the
>> key).
>> +     * Calls for which context is the cpu number, avoid hashing and act as
>> +     * pcpu add/removal.
>> +     *
>> +     * Since the population inside this set is always very much limited,
>> +     * list scanning should be short.
>>      */
>> -    list_for_each_entry(tncur, lsthead, llink) {
>> -        if (tncur->ctx == ctx &&
>> -            (tncur->cookie == cookie || ++call_nests > max_nests)) {
>> -            /*
>> -             * Ops ... loop detected or maximum nest level reached.
>> -             * We abort this wake by breaking the cycle itself.
>> -             */
>> -            error = -1;
>> -            goto out_unlock;
>> -        }
>> -    }
>> +    list_for_each_entry(tncur, &head->list, llink.list) {
>> +           if (tncur->ctx == ctx &&
>> +           (tncur->cookie == cookie || ++call_nests > EP_MAX_NESTS)) {
>> +               /*
>> +            * Ops ... loop detected or maximum nest level reached.
>> +            * We abort this wake by breaking the cycle itself.
>> +            */
>> +               error = -1;
>> +               spin_unlock_irqrestore(&head->lock, flags);
>> +               goto done;
>> +           }
>> +       }
>> +
>>
>>     /* Add the current task and cookie to the list */
>>     tnode.ctx = ctx;
>>     tnode.cookie = cookie;
>> -    list_add(&tnode.llink, lsthead);
>> -
>> -    spin_unlock_irqrestore(&ncalls->lock, flags);
>> +    tnode.llink.head = head;
>> +    list_add(&tnode.llink.list, &head->list);
>> +    spin_unlock_irqrestore(&head->lock, flags);
>>
>>     /* Call the nested function */
>>     error = (*nproc)(priv, cookie, call_nests);
>>
>>     /* Remove the current task from the list */
>> -    spin_lock_irqsave(&ncalls->lock, flags);
>> -    list_del(&tnode.llink);
>> -out_unlock:
>> -    spin_unlock_irqrestore(&ncalls->lock, flags);
>> -
>> +    dlock_lists_del(&tnode.llink);
>> +done:
>>     return error;
>> }
>>
>> @@ -2313,13 +2321,16 @@ static int __init eventpoll_init(void)
>>      * Initialize the structure used to perform epoll file descriptor
>>      * inclusion loops checks.
>>      */
>> -    ep_nested_calls_init(&poll_loop_ncalls);
>> +    if (ep_nested_calls_init(&poll_loop_ncalls))
>> +        goto err;
>>
>>     /* Initialize the structure used to perform safe poll wait head wake
>> ups */
>> -    ep_nested_calls_init(&poll_safewake_ncalls);
>> +    if (ep_nested_calls_init(&poll_safewake_ncalls))
>> +        goto err_free1;
>>
>>     /* Initialize the structure used to perform file's f_op->poll()
>> calls */
>> -    ep_nested_calls_init(&poll_readywalk_ncalls);
>> +    if (ep_nested_calls_init(&poll_readywalk_ncalls))
>> +        goto err_free0;
>>
>>     /*
>>      * We can have many thousands of epitems, so prevent this from
>> @@ -2336,5 +2347,12 @@ static int __init eventpoll_init(void)
>>             sizeof(struct eppoll_entry), 0, SLAB_PANIC, NULL);
>>
>>     return 0;
>> +
>> +err_free0:
>> +    ep_nested_calls_free(&poll_safewake_ncalls);
>> +err_free1:
>> +    ep_nested_calls_free(&poll_loop_ncalls);
>> +err:
>> +    return -ENOMEM;
>> }
>> fs_initcall(eventpoll_init);

Powered by blists - more mailing lists