| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CAB=BE-Qh72tB+z9zRxqiC7Jg+RVcip1+vHJXopuzykR08FYiCA@mail.gmail.com>
Date: Mon, 26 Dec 2022 10:38:23 -0800
From: Sandeep Dhavale <dhavale@...gle.com>
To: Sandeep Dhavale <dhavale@...gle.com>, Gao Xiang <xiang@...nel.org>,
Chao Yu <chao@...nel.org>, Yue Hu <huyue2@...lpad.com>,
Jeffle Xu <jefflexu@...ux.alibaba.com>,
kernel-team@...roid.com, linux-erofs@...ts.ozlabs.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH] EROFS: Replace erofs_unzipd workqueue with per-cpu threads
On Wed, Dec 21, 2022 at 6:48 PM Gao Xiang <xiang@...nel.org> wrote:
>
> Hi Sandeen,
>
> On Thu, Dec 22, 2022 at 01:15:29AM +0000, Sandeep Dhavale wrote:
> > Using per-cpu thread pool we can reduce the scheduling latency compared
> > to workqueue implementation. With this patch scheduling latency and
> > variation is reduced as per-cpu threads are SCHED_FIFO kthread_workers.
> >
> > The results were evaluated on arm64 Android devices running 5.10 kernel.
> >
> > The table below shows resulting improvements of total scheduling latency
> > for the same app launch benchmark runs with 50 iterations. Scheduling
> > latency is the latency between when the task (workqueue kworker vs
> > kthread_worker) became eligible to run to when it actually started
> > running.
> > +-------------------------+-----------+----------------+---------+
> > | | workqueue | kthread_worker | diff |
> > +-------------------------+-----------+----------------+---------+
> > | Average (us) | 15253 | 2914 | -80.89% |
> > | Median (us) | 14001 | 2912 | -79.20% |
> > | Minimum (us) | 3117 | 1027 | -67.05% |
> > | Maximum (us) | 30170 | 3805 | -87.39% |
> > | Standard deviation (us) | 7166 | 359 | |
> > +-------------------------+-----------+----------------+---------+
> >
> > Signed-off-by: Sandeep Dhavale <dhavale@...gle.com>
>
> Thanks for the patch. Generally, This path looks good to me (compared
> with softirq context.)
>
> With the background at LPC 22, I can see how important such low latency
> requirement is needed for Android upstream and AOSP. However, could you
> add some link or some brief background to other folks without such
> impression?
>
> > ---
> > fs/erofs/zdata.c | 84 +++++++++++++++++++++++++++++++++++-------------
> > fs/erofs/zdata.h | 4 ++-
> > 2 files changed, 64 insertions(+), 24 deletions(-)
> >
> > diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
> > index ccf7c55d477f..646667dbe615 100644
> > --- a/fs/erofs/zdata.c
> > +++ b/fs/erofs/zdata.c
> > @@ -8,6 +8,7 @@
> > #include "compress.h"
> > #include <linux/prefetch.h>
> > #include <linux/psi.h>
> > +#include <linux/slab.h>
> >
> > #include <trace/events/erofs.h>
> >
> > @@ -184,26 +185,56 @@ typedef tagptr1_t compressed_page_t;
> > #define tag_compressed_page_justfound(page) \
> > tagptr_fold(compressed_page_t, page, 1)
> >
> > -static struct workqueue_struct *z_erofs_workqueue __read_mostly;
> > +static struct kthread_worker **z_erofs_kthread_pool;
> >
> > -void z_erofs_exit_zip_subsystem(void)
> > +static void z_erofs_destroy_kthread_pool(void)
> > {
> > - destroy_workqueue(z_erofs_workqueue);
> > - z_erofs_destroy_pcluster_pool();
> > + unsigned long cpu;
> > +
> > + for_each_possible_cpu(cpu) {
> > + if (z_erofs_kthread_pool[cpu]) {
> > + kthread_destroy_worker(z_erofs_kthread_pool[cpu]);
> > + z_erofs_kthread_pool[cpu] = NULL;
> > + }
> > + }
> > + kfree(z_erofs_kthread_pool);
> > }
> >
> > -static inline int z_erofs_init_workqueue(void)
> > +static int z_erofs_create_kthread_workers(void)
> > {
> > - const unsigned int onlinecpus = num_possible_cpus();
> > + unsigned long cpu;
> > + struct kthread_worker *worker;
> > +
> > + for_each_possible_cpu(cpu) {
> > + worker = kthread_create_worker_on_cpu(cpu, 0, "z_erofs/%ld", cpu);
>
> how about calling them as erofs_worker/%ld, since in the future they
> can also be used for other uses (like verification or decryption).
>
Sure, it makes sense.
> > + if (IS_ERR(worker)) {
> > + z_erofs_destroy_kthread_pool();
> > + return -ENOMEM;
> > + }
> > + sched_set_fifo(worker->task);
>
> Could we add some kernel configuration option to enable/disable this,
> since I'm not sure if all users need RT threads.
>
Ok, will do.
> > + z_erofs_kthread_pool[cpu] = worker;
> > + }
> > + return 0;
> > +}
> >
> > - /*
> > - * no need to spawn too many threads, limiting threads could minimum
> > - * scheduling overhead, perhaps per-CPU threads should be better?
> > - */
> > - z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
> > - WQ_UNBOUND | WQ_HIGHPRI,
> > - onlinecpus + onlinecpus / 4);
> > - return z_erofs_workqueue ? 0 : -ENOMEM;
> > +static int z_erofs_init_kthread_pool(void)
> > +{
> > + int err;
> > +
> > + z_erofs_kthread_pool = kcalloc(num_possible_cpus(),
> > + sizeof(struct kthread_worker *), GFP_ATOMIC);
> > + if (!z_erofs_kthread_pool)
> > + return -ENOMEM;
> > + err = z_erofs_create_kthread_workers();
> > +
> > + return err;
> > +}
> > +
> > +
> > +void z_erofs_exit_zip_subsystem(void)
> > +{
> > + z_erofs_destroy_kthread_pool();
> > + z_erofs_destroy_pcluster_pool();
> > }
> >
> > int __init z_erofs_init_zip_subsystem(void)
> > @@ -211,10 +242,16 @@ int __init z_erofs_init_zip_subsystem(void)
> > int err = z_erofs_create_pcluster_pool();
> >
> > if (err)
> > - return err;
> > - err = z_erofs_init_workqueue();
> > + goto out_error_pcluster_pool;
> > +
> > + err = z_erofs_init_kthread_pool();
> > if (err)
> > - z_erofs_destroy_pcluster_pool();
> > + goto out_error_kthread_pool;
> > +
> > + return err;
> > +out_error_kthread_pool:
> > + z_erofs_destroy_pcluster_pool();
> > +out_error_pcluster_pool:
> > return err;
> > }
> >
> > @@ -1143,7 +1180,7 @@ static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
> > }
> > }
> >
> > -static void z_erofs_decompressqueue_work(struct work_struct *work)
> > +static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
> > {
> > struct z_erofs_decompressqueue *bgq =
> > container_of(work, struct z_erofs_decompressqueue, u.work);
> > @@ -1170,15 +1207,16 @@ static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
> >
> > if (atomic_add_return(bios, &io->pending_bios))
> > return;
> > - /* Use workqueue and sync decompression for atomic contexts only */
> > + /* Use kthread_workers and sync decompression for atomic contexts only */
> > if (in_atomic() || irqs_disabled()) {
> > - queue_work(z_erofs_workqueue, &io->u.work);
> > + kthread_queue_work(z_erofs_kthread_pool[raw_smp_processor_id()],
> > + &io->u.work);
>
> Should we need to handle cpu online/offline as well?
>
Ok, let me try to add cpuhp support. I will work on V2.
> Thanks,
> Gao Xiang
Thanks,
Sandeep.
Powered by blists - more mailing lists