Date:   Sun, 24 Apr 2022 10:58:10 +0800
From:   Zhangfei Gao <zhangfei.gao@...aro.org>
To:     "zhangfei.gao@...mail.com" <zhangfei.gao@...mail.com>
Cc:     Jean-Philippe Brucker <jean-philippe@...aro.org>,
        Fenghua Yu <fenghua.yu@...el.com>,
        Dave Hansen <dave.hansen@...el.com>,
        Joerg Roedel <joro@...tes.org>,
        Ravi V Shankar <ravi.v.shankar@...el.com>,
        Tony Luck <tony.luck@...el.com>,
        Ashok Raj <ashok.raj@...el.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Dave Hansen <dave.hansen@...ux.intel.com>,
        x86 <x86@...nel.org>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        iommu <iommu@...ts.linux-foundation.org>,
        Ingo Molnar <mingo@...hat.com>, Borislav Petkov <bp@...en8.de>,
        Andy Lutomirski <luto@...nel.org>,
        Josh Poimboeuf <jpoimboe@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>, will@...nel.org,
        robin.murphy@....com
Subject: Re: [PATCH v4 05/11] iommu/sva: Assign a PASID to mm on PASID
 allocation and free it on mm exit

On Sat, 23 Apr 2022 at 19:13, zhangfei.gao@...mail.com
<zhangfei.gao@...mail.com> wrote:
>
> Hi, Jean
>
> On 2022/4/22 下午11:50, Jean-Philippe Brucker wrote:
> > On Fri, Apr 22, 2022 at 09:15:01PM +0800, zhangfei.gao@...mail.com wrote:
> >>> I'm trying to piece together what happens from the kernel point of view.
> >>>
> >>> * master process with mm A opens a queue fd through uacce, which calls
> >>>     iommu_sva_bind_device(dev, A) -> PASID 1
> >>>
> >>> * master forks and exits. Child (daemon) gets mm B, inherits the queue fd.
> >>>     The device is still bound to mm A with PASID 1, since the queue fd is
> >>>     still open.
> >>> We discussed this before, but I don't remember where we left off. The
> >>> child can't use the queue because its mappings are not copied on fork(),
> >>> and the queue is still bound to the parent mm A. The child either needs to
> >>> open a new queue or take ownership of the old one with a new uacce ioctl.
> >> Yes, currently nginx aligned with the case.
> >> Child process (worker process) reopen uacce,
> >>
> >> Master process (do init) open uacce, iommu_sva_bind_device(dev, A) -> PASID
> >> 1
> >> Master process fork Child (daemon) and exit.
> >>
> >> Child (daemon)  does not use PASID 1 any more, only fork and manage worker
> >> process.
> >> Worker process reopen uacce, iommu_sva_bind_device(dev, B) PASID 2
> >>
> >> So it is expected.
> > Yes, that's fine
> >
> >>> Is that the "IMPLEMENT_DYNAMIC_BIND_FN()" you mention, something out of
> >>> tree?  This operation should unbind from A before binding to B, no?
> >>> Otherwise we leak PASID 1.
> >> In 5.16 PASID 1 from master is hold until nginx service stop.
> >> nginx start
> >> master:
> >> iommu_sva_alloc_pasid mm->pasid=1      // master process
> >>
> >> lynx https start:
> >> iommu_sva_alloc_pasid mm->pasid=2    //worker process
> >>
> >> nginx stop:  from fops_release
> >> iommu_sva_free_pasid mm->pasid=2   // worker process
> >> iommu_sva_free_pasid mm->pasid=1  // master process
> > That's the expected behavior (master could close its fd before forking, in
> > order to free things up earlier, but it's not mandatory)
> Currently we unbind in fops_release, so the ioasid allocated in master
> can only be freed when nginx stop,
> when all forked fd are closed.
>
> >
> >> Have one silly question.
> >>
> >> kerne driver
> >> fops_open
> >> iommu_sva_bind_device
> >>
> >> fops_release
> >> iommu_sva_unbind_device
> >>
> >> application
> >> main()
> >> fd = open
> >> return;
> >>
> >> Application exit but not close(fd), is it expected fops_release will be
> >> called automatically by system?
> > Yes, the application doesn't have to call close() explicitly, the file
> > descriptor is closed automatically on exit. Note that the fd is copied on
> > fork(), so it is only released once parent and all child processes exit.
> Yes, in case the application ended unexpected, like ctrl+c
> >
> >> On 5.17
> >> fops_release is called automatically, as well as iommu_sva_unbind_device.
> >> On 5.18-rc1.
> >> fops_release is not called, have to manually call close(fd)
> > Right that's weird
> Looks it is caused by the fix patch, via mmget, which may add refcount
> of fd.
>
> Some experiments
> 1. 5.17, everything works well.
>
> 2. 5.17 + patchset of "iommu/sva: Assign a PASID to mm on PASID
> allocation and free it on mm exit"
>
> Test application, exit without close uacce fd
> First time:  fops_release can be called automatically.
>
> log:
> ioasid_alloc ioasid=1
> iommu_sva_alloc_pasid pasid=1
> iommu_sva_bind_device handle=00000000263a2ee8
> ioasid_free ioasid=1
> uacce_fops_release q=0000000055ca3cdf
> iommu_sva_unbind_device handle=00000000263a2ee8
>
> Second time: hardware reports error
>
> uacce_fops_open q=000000008e4d6f78
> ioasid_alloc ioasid=1
> iommu_sva_alloc_pasid pasid=1
> iommu_sva_bind_device handle=00000000cfd11788
> // Haredware reports error
> hisi_sec2 0000:b6:00.0: qm_acc_do_task_timeout [error status=0x20] found
> hisi_sec2 0000:b6:00.0: qm_acc_wb_not_ready_timeout [error status=0x40]
> found
> hisi_sec2 0000:b6:00.0: sec_fsm_hbeat_rint [error status=0x20] found
> hisi_sec2 0000:b6:00.0: Controller resetting...
> hisi_sec2 0000:b6:00.0: QM mailbox operation timeout!
> hisi_sec2 0000:b6:00.0: Failed to dump sqc!
> hisi_sec2 0000:b6:00.0: Failed to drain out data for stopping!
> hisi_sec2 0000:b6:00.0: Bus lock! Please reset system.
> hisi_sec2 0000:b6:00.0: Controller reset failed (-110)
> hisi_sec2 0000:b6:00.0: controller reset failed (-110)
>
> 3. Add the fix patch of using mmget in bind.
> Test application, exit without close uacce fd
>
> fops_release can NOT be called automatically, looks mmget adds refcount
> of fd.

Test application, exit without closing fd.
> >> kernel driver
> >> fops_open
> >> iommu_sva_bind_device
> >>
> >> fops_release
> >> iommu_sva_unbind_device

1.
5.17 kernel, no mmget & mmput

wd_release_queue no close
Compress bz=512000 nb=1×10, speed=139.5 MB/s (±0.0% N=1) overall=122.9
MB/s (±0.0%)
[   16.052989] do_exit current=d380000
[   16.053828] mmput atomic=1
[   16.054511]  __mmput atomic=0
[   16.070382] exit_task_work
[   16.070981] uacce_fops_release current=d380000
[   16.071999] CPU: 0 PID: 176 Comm: test_sva_perf Not tainted
5.16.0-rc1-27342-ge5f9f3f99a88-dirty #240
[   16.074007] Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0 02/06/2015
[   16.075530] Call trace:
[   16.076069]  dump_backtrace+0x0/0x1a0
[   16.076887]  show_stack+0x20/0x30
[   16.077629]  dump_stack_lvl+0x8c/0xb8
[   16.078441]  dump_stack+0x18/0x34
[   16.079176]  uacce_fops_release+0x44/0xdc
[   16.080060]  __fput+0x78/0x240
[   16.080743]  ____fput+0x18/0x28
[   16.081447]  task_work_run+0x88/0x160
[   16.082259]  do_exit+0x52c/0xa50
[   16.082974]  do_group_exit+0x84/0xa8
[   16.083768]  __wake_up_parent+0x0/0x38
[   16.084597]  invoke_syscall+0x4c/0x110
[   16.085435]  el0_svc_common.constprop.0+0x68/0x128
[   16.086501]  do_el0_svc+0x2c/0x90
[   16.087243]  el0_svc+0x24/0x70
[   16.087928]  el0t_64_sync_handler+0xb0/0xb8
[   16.088854]  el0t_64_sync+0x1a0/0x1a4
[   16.089775]  arm_smmu_sva_unbind
[   16.090577]  iommu_sva_free_pasid mm->pasid=1
[   16.091763] exit_task_work done

2. Add mmget in bind and mmput in unbind,
Since application do not close fd, so no unbind,& mmput
And fops_release is not called since mm_users account.

log:
[  101.642690] mmput atomic=3
wd_release_queue no close
Compress bz=512000 nb=1×10, speed=40.3 MB/s (±0.0% N=1) overall=38.7
MB/s (±0.0%)
[  101.671167] do_exit current=d9daf40
[  101.672003] mmput atomic=2
[  101.672712] exit_task_work
[  101.673292] exit_task_work done

Thanks



>
> So the fix method of using mmget blocks fops_release to be called once
> fd is closed without unbind.
>
> >> Since nginx may have a issue, it does not call close(fd) when nginx -s quit.
> > And you're sure that none of the processes are still alive or in zombie
> > state?  Just to cover every possibility.
> It can also reproduced by a simple application exit without close(uacce_fd)
>
> Thanks
>

Powered by blists - more mailing lists