diff mbox series

[v2] PCI: Prevent device from doing RPM when it's unplugged

Message ID 20231212043808.212754-1-kai.heng.feng@canonical.com (mailing list archive)
State Not Applicable
Headers show
Series [v2] PCI: Prevent device from doing RPM when it's unplugged | expand

Checks

Context Check Description
netdev/tree_selection success Not a local patch

Commit Message

Kai-Heng Feng Dec. 12, 2023, 4:38 a.m. UTC
When inserting an SD7.0 card to Realtek card reader, the card reader
unplugs itself and morph into a NVMe device. The slot Link down on hot
unplugged can cause the following error:
[   63.898861] pcieport 0000:00:1c.0: pciehp: Slot(8): Link Down
[   63.912118] BUG: unable to handle page fault for address: ffffb24d403e5010
[   63.912122] #PF: supervisor read access in kernel mode
[   63.912125] #PF: error_code(0x0000) - not-present page
[   63.912126] PGD 100000067 P4D 100000067 PUD 1001fe067 PMD 100d97067 PTE 0
[   63.912131] Oops: 0000 [#1] PREEMPT SMP PTI
[   63.912134] CPU: 3 PID: 534 Comm: kworker/3:10 Not tainted 6.4.0 #6
[   63.912137] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./H370M Pro4, BIOS P3.40 10/25/2018
[   63.912138] Workqueue: pm pm_runtime_work
[   63.912144] RIP: 0010:ioread32+0x2e/0x70
[   63.912148] Code: ff 03 00 77 25 48 81 ff 00 00 01 00 77 14 8b 15 08 d9 54 01 b8 ff ff ff ff 85 d2 75 14 c3 cc cc cc cc 89 fa ed c3 cc cc cc cc <8b> 07 c3 cc cc cc cc 55 83 ea 01 48 89 fe 48 c7 c7 98 6f 15 99 48
[   63.912150] RSP: 0018:ffffb24d40a5bd78 EFLAGS: 00010296
[   63.912152] RAX: ffffb24d403e5000 RBX: 0000000000000152 RCX: 000000000000007f
[   63.912153] RDX: 000000000000ff00 RSI: ffffb24d403e5010 RDI: ffffb24d403e5010
[   63.912155] RBP: ffffb24d40a5bd98 R08: ffffb24d403e5010 R09: 0000000000000000
[   63.912156] R10: ffff9074cd95e7f4 R11: 0000000000000003 R12: 000000000000007f
[   63.912158] R13: ffff9074e1a68c00 R14: ffff9074e1a68d00 R15: 0000000000009003
[   63.912159] FS:  0000000000000000(0000) GS:ffff90752a180000(0000) knlGS:0000000000000000
[   63.912161] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[   63.912162] CR2: ffffb24d403e5010 CR3: 0000000152832006 CR4: 00000000003706e0
[   63.912164] Call Trace:
[   63.912165]  <TASK>
[   63.912167]  ? show_regs+0x68/0x70
[   63.912171]  ? __die_body+0x20/0x70
[   63.912173]  ? __die+0x2b/0x40
[   63.912175]  ? page_fault_oops+0x160/0x480
[   63.912177]  ? search_bpf_extables+0x63/0x90
[   63.912180]  ? ioread32+0x2e/0x70
[   63.912183]  ? search_exception_tables+0x5f/0x70
[   63.912186]  ? kernelmode_fixup_or_oops+0xa2/0x120
[   63.912189]  ? __bad_area_nosemaphore+0x179/0x230
[   63.912191]  ? bad_area_nosemaphore+0x16/0x20
[   63.912193]  ? do_kern_addr_fault+0x8b/0xa0
[   63.912195]  ? exc_page_fault+0xe5/0x180
[   63.912198]  ? asm_exc_page_fault+0x27/0x30
[   63.912203]  ? ioread32+0x2e/0x70
[   63.912206]  ? rtsx_pci_write_register+0x5b/0x90 [rtsx_pci]
[   63.912217]  rtsx_set_l1off_sub+0x1c/0x30 [rtsx_pci]
[   63.912226]  rts5261_set_l1off_cfg_sub_d0+0x36/0x40 [rtsx_pci]
[   63.912234]  rtsx_pci_runtime_idle+0xc7/0x160 [rtsx_pci]
[   63.912243]  ? __pfx_pci_pm_runtime_idle+0x10/0x10
[   63.912246]  pci_pm_runtime_idle+0x34/0x70
[   63.912248]  rpm_idle+0xc4/0x2b0
[   63.912251]  pm_runtime_work+0x93/0xc0
[   63.912254]  process_one_work+0x21a/0x430
[   63.912258]  worker_thread+0x4a/0x3c0
[   63.912261]  ? __pfx_worker_thread+0x10/0x10
[   63.912263]  kthread+0x106/0x140
[   63.912266]  ? __pfx_kthread+0x10/0x10
[   63.912268]  ret_from_fork+0x29/0x50
[   63.912273]  </TASK>
[   63.912274] Modules linked in: nvme nvme_core snd_hda_codec_hdmi snd_sof_pci_intel_cnl snd_sof_intel_hda_common snd_hda_codec_realtek snd_hda_codec_generic snd_soc_hdac_hda soundwire_intel ledtrig_audio nls_iso8859_1 soundwire_generic_allocation soundwire_cadence snd_sof_intel_hda_mlink snd_sof_intel_hda snd_sof_pci snd_sof_xtensa_dsp snd_sof snd_sof_utils snd_hda_ext_core snd_soc_acpi_intel_match snd_soc_acpi soundwire_bus snd_soc_core snd_compress ac97_bus snd_pcm_dmaengine snd_hda_intel i915 snd_intel_dspcfg snd_intel_sdw_acpi intel_rapl_msr snd_hda_codec intel_rapl_common snd_hda_core x86_pkg_temp_thermal intel_powerclamp snd_hwdep coretemp snd_pcm kvm_intel drm_buddy ttm mei_hdcp kvm drm_display_helper snd_seq_midi snd_seq_midi_event cec crct10dif_pclmul ghash_clmulni_intel sha512_ssse3 aesni_intel crypto_simd rc_core cryptd rapl snd_rawmidi drm_kms_helper binfmt_misc intel_cstate i2c_algo_bit joydev snd_seq snd_seq_device syscopyarea wmi_bmof snd_timer sysfillrect input_leds snd ee1004 sysimgblt mei_me soundcore
[   63.912324]  mei intel_pch_thermal mac_hid acpi_tad acpi_pad sch_fq_codel msr parport_pc ppdev lp ramoops drm parport reed_solomon efi_pstore ip_tables x_tables autofs4 hid_generic usbhid hid rtsx_pci_sdmmc crc32_pclmul ahci e1000e i2c_i801 i2c_smbus rtsx_pci xhci_pci libahci xhci_pci_renesas video wmi
[   63.912346] CR2: ffffb24d403e5010
[   63.912348] ---[ end trace 0000000000000000 ]---

This happens because scheduled pm_runtime_idle() is not cancelled.

So before releasing the device, stop all runtime power managements by
using pm_runtime_barrier() to fix the issue.

Link: https://lore.kernel.org/all/2ce258f371234b1f8a1a470d5488d00e@realtek.com/
Tested-by: Ricky Wu <ricky_wu@realtek.com>
---
v2:
  Cover more cases than just pciehp.
 
 drivers/pci/remove.c | 2 ++
 1 file changed, 2 insertions(+)

Comments

Bjorn Helgaas Dec. 14, 2023, 5:16 p.m. UTC | #1
[+cc Rafael, runtime PM expert :)]

On Tue, Dec 12, 2023 at 12:38:07PM +0800, Kai-Heng Feng wrote:
> When inserting an SD7.0 card to Realtek card reader, the card reader
> unplugs itself and morph into a NVMe device. The slot Link down on hot
> unplugged can cause the following error:

A page fault in a driver following a link down event sounds like
either a driver defect or a PCI core defect that could affect any
driver.  The rtsx power and ASPM management is very unusual, so I
don't feel super confident in it.

I guess the theory here is that while we're running
rtsx_pci_runtime_idle(), the link down event happens and we run
rtsx_pci_remove(), which unmaps the pcr->remap_addr page, and then
rtsx_pci_readl(RTSX_HAIMR) in the rtsx_pci_runtime_idle() path
references that unmapped page?

I looked through other drivers that use runtime PM.  The typical
pattern is:

  *_probe()
    pm_runtime_put
    pm_runtime_allow

  *_remove()
    pm_runtime_forbid
    pm_runtime_get

rtsx does the put/allow and forbid/get in the reverse order:

  rtsx_pci_probe()
    pm_runtime_allow
    pm_runtime_put

  rtsx_pci_remove()
    pm_runtime_get_sync
    pm_runtime_forbid
    iounmap(pcr->remap_addr)               # <-- unmap the page

  rtsx_pci_runtime_idle()
    ...
      ioread32(pcr->remap_addr + reg)      # <-- read from unmapped page

I don't know whether this is an issue, and isp_probe() and nhi_probe()
also use this reverse order, so maybe it's all fine.  But I do wonder
whether there's a reason to do it differently.

> [   63.898861] pcieport 0000:00:1c.0: pciehp: Slot(8): Link Down
> [   63.912118] BUG: unable to handle page fault for address: ffffb24d403e5010
> [   63.912122] #PF: supervisor read access in kernel mode
> [   63.912125] #PF: error_code(0x0000) - not-present page
> [   63.912126] PGD 100000067 P4D 100000067 PUD 1001fe067 PMD 100d97067 PTE 0
> [   63.912131] Oops: 0000 [#1] PREEMPT SMP PTI
> [   63.912134] CPU: 3 PID: 534 Comm: kworker/3:10 Not tainted 6.4.0 #6
> [   63.912137] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./H370M Pro4, BIOS P3.40 10/25/2018
> [   63.912138] Workqueue: pm pm_runtime_work
> [   63.912144] RIP: 0010:ioread32+0x2e/0x70
> [   63.912148] Code: ff 03 00 77 25 48 81 ff 00 00 01 00 77 14 8b 15 08 d9 54 01 b8 ff ff ff ff 85 d2 75 14 c3 cc cc cc cc 89 fa ed c3 cc cc cc cc <8b> 07 c3 cc cc cc cc 55 83 ea 01 48 89 fe 48 c7 c7 98 6f 15 99 48
> [   63.912150] RSP: 0018:ffffb24d40a5bd78 EFLAGS: 00010296
> [   63.912152] RAX: ffffb24d403e5000 RBX: 0000000000000152 RCX: 000000000000007f
> [   63.912153] RDX: 000000000000ff00 RSI: ffffb24d403e5010 RDI: ffffb24d403e5010
> [   63.912155] RBP: ffffb24d40a5bd98 R08: ffffb24d403e5010 R09: 0000000000000000
> [   63.912156] R10: ffff9074cd95e7f4 R11: 0000000000000003 R12: 000000000000007f
> [   63.912158] R13: ffff9074e1a68c00 R14: ffff9074e1a68d00 R15: 0000000000009003
> [   63.912159] FS:  0000000000000000(0000) GS:ffff90752a180000(0000) knlGS:0000000000000000
> [   63.912161] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [   63.912162] CR2: ffffb24d403e5010 CR3: 0000000152832006 CR4: 00000000003706e0
> [   63.912164] Call Trace:
> [   63.912165]  <TASK>
> [   63.912167]  ? show_regs+0x68/0x70
> [   63.912171]  ? __die_body+0x20/0x70
> [   63.912173]  ? __die+0x2b/0x40
> [   63.912175]  ? page_fault_oops+0x160/0x480
> [   63.912177]  ? search_bpf_extables+0x63/0x90
> [   63.912180]  ? ioread32+0x2e/0x70
> [   63.912183]  ? search_exception_tables+0x5f/0x70
> [   63.912186]  ? kernelmode_fixup_or_oops+0xa2/0x120
> [   63.912189]  ? __bad_area_nosemaphore+0x179/0x230
> [   63.912191]  ? bad_area_nosemaphore+0x16/0x20
> [   63.912193]  ? do_kern_addr_fault+0x8b/0xa0
> [   63.912195]  ? exc_page_fault+0xe5/0x180
> [   63.912198]  ? asm_exc_page_fault+0x27/0x30
> [   63.912203]  ? ioread32+0x2e/0x70
> [   63.912206]  ? rtsx_pci_write_register+0x5b/0x90 [rtsx_pci]
> [   63.912217]  rtsx_set_l1off_sub+0x1c/0x30 [rtsx_pci]
> [   63.912226]  rts5261_set_l1off_cfg_sub_d0+0x36/0x40 [rtsx_pci]
> [   63.912234]  rtsx_pci_runtime_idle+0xc7/0x160 [rtsx_pci]
> [   63.912243]  ? __pfx_pci_pm_runtime_idle+0x10/0x10
> [   63.912246]  pci_pm_runtime_idle+0x34/0x70
> [   63.912248]  rpm_idle+0xc4/0x2b0
> [   63.912251]  pm_runtime_work+0x93/0xc0
> [   63.912254]  process_one_work+0x21a/0x430
> [   63.912258]  worker_thread+0x4a/0x3c0
> [   63.912261]  ? __pfx_worker_thread+0x10/0x10
> [   63.912263]  kthread+0x106/0x140
> [   63.912266]  ? __pfx_kthread+0x10/0x10
> [   63.912268]  ret_from_fork+0x29/0x50
> [   63.912273]  </TASK>

Can you strip out the call trace stuff that's not relevant so the call
path is clear?  I'm guessing we're in ioread32(), and nothing above
do_kern_addr_fault() or below worker_thread() is relevant.

> [   63.912274] Modules linked in: nvme nvme_core snd_hda_codec_hdmi snd_sof_pci_intel_cnl snd_sof_intel_hda_common snd_hda_codec_realtek snd_hda_codec_generic snd_soc_hdac_hda soundwire_intel ledtrig_audio nls_iso8859_1 soundwire_generic_allocation soundwire_cadence snd_sof_intel_hda_mlink snd_sof_intel_hda snd_sof_pci snd_sof_xtensa_dsp snd_sof snd_sof_utils snd_hda_ext_core snd_soc_acpi_intel_match snd_soc_acpi soundwire_bus snd_soc_core snd_compress ac97_bus snd_pcm_dmaengine snd_hda_intel i915 snd_intel_dspcfg snd_intel_sdw_acpi intel_rapl_msr snd_hda_codec intel_rapl_common snd_hda_core x86_pkg_temp_thermal intel_powerclamp snd_hwdep coretemp snd_pcm kvm_intel drm_buddy ttm mei_hdcp kvm drm_display_helper snd_seq_midi snd_seq_midi_event cec crct10dif_pclmul ghash_clmulni_intel sha512_ssse3 aesni_intel crypto_simd rc_core cryptd rapl snd_rawmidi drm_kms_helper binfmt_misc intel_cstate i2c_algo_bit joydev snd_seq snd_seq_device syscopyarea wmi_bmof snd_timer sysfillrect input_leds snd ee1004 sysimgblt mei_me soundcore
> [   63.912324]  mei intel_pch_thermal mac_hid acpi_tad acpi_pad sch_fq_codel msr parport_pc ppdev lp ramoops drm parport reed_solomon efi_pstore ip_tables x_tables autofs4 hid_generic usbhid hid rtsx_pci_sdmmc crc32_pclmul ahci e1000e i2c_i801 i2c_smbus rtsx_pci xhci_pci libahci xhci_pci_renesas video wmi

The module list doesn't look relevant here.  Nor the timestamps.

> [   63.912346] CR2: ffffb24d403e5010
> [   63.912348] ---[ end trace 0000000000000000 ]---
> 
> This happens because scheduled pm_runtime_idle() is not cancelled.
> 
> So before releasing the device, stop all runtime power managements by
> using pm_runtime_barrier() to fix the issue.
>
> Link: https://lore.kernel.org/all/2ce258f371234b1f8a1a470d5488d00e@realtek.com/
> Tested-by: Ricky Wu <ricky_wu@realtek.com>
> ---
> v2:
>   Cover more cases than just pciehp.
>  
>  drivers/pci/remove.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c
> index d749ea8250d6..c69b4ce5dbfd 100644
> --- a/drivers/pci/remove.c
> +++ b/drivers/pci/remove.c
> @@ -1,6 +1,7 @@
>  // SPDX-License-Identifier: GPL-2.0
>  #include <linux/pci.h>
>  #include <linux/module.h>
> +#include <linux/pm_runtime.h>
>  #include "pci.h"
>  
>  static void pci_free_resources(struct pci_dev *dev)
> @@ -18,6 +19,7 @@ static void pci_stop_dev(struct pci_dev *dev)
>  	pci_pme_active(dev, false);
>  
>  	if (pci_dev_is_added(dev)) {
> +		pm_runtime_barrier(&dev->dev);

If pm_runtime_barrier() is really the solution, it seems like this
should go somewhere in pci-driver.c where we call the driver PM
callbacks.

>  		device_release_driver(&dev->dev);
>  		pci_proc_detach_device(dev);
> -- 
> 2.34.1
>
Kai-Heng Feng Dec. 22, 2023, 3:53 a.m. UTC | #2
Hi Bjorn,

On Fri, Dec 15, 2023 at 1:16 AM Bjorn Helgaas <helgaas@kernel.org> wrote:
>
> [+cc Rafael, runtime PM expert :)]
>
> On Tue, Dec 12, 2023 at 12:38:07PM +0800, Kai-Heng Feng wrote:
> > When inserting an SD7.0 card to Realtek card reader, the card reader
> > unplugs itself and morph into a NVMe device. The slot Link down on hot
> > unplugged can cause the following error:
>
> A page fault in a driver following a link down event sounds like
> either a driver defect or a PCI core defect that could affect any
> driver.  The rtsx power and ASPM management is very unusual, so I
> don't feel super confident in it.
>
> I guess the theory here is that while we're running
> rtsx_pci_runtime_idle(), the link down event happens and we run
> rtsx_pci_remove(), which unmaps the pcr->remap_addr page, and then
> rtsx_pci_readl(RTSX_HAIMR) in the rtsx_pci_runtime_idle() path
> references that unmapped page?
>
> I looked through other drivers that use runtime PM.  The typical
> pattern is:
>
>   *_probe()
>     pm_runtime_put
>     pm_runtime_allow
>
>   *_remove()
>     pm_runtime_forbid
>     pm_runtime_get
>
> rtsx does the put/allow and forbid/get in the reverse order:
>
>   rtsx_pci_probe()
>     pm_runtime_allow
>     pm_runtime_put
>
>   rtsx_pci_remove()
>     pm_runtime_get_sync
>     pm_runtime_forbid
>     iounmap(pcr->remap_addr)               # <-- unmap the page
>
>   rtsx_pci_runtime_idle()
>     ...
>       ioread32(pcr->remap_addr + reg)      # <-- read from unmapped page
>
> I don't know whether this is an issue, and isp_probe() and nhi_probe()
> also use this reverse order, so maybe it's all fine.  But I do wonder
> whether there's a reason to do it differently.
>
> > [   63.898861] pcieport 0000:00:1c.0: pciehp: Slot(8): Link Down
> > [   63.912118] BUG: unable to handle page fault for address: ffffb24d403e5010
> > [   63.912122] #PF: supervisor read access in kernel mode
> > [   63.912125] #PF: error_code(0x0000) - not-present page
> > [   63.912126] PGD 100000067 P4D 100000067 PUD 1001fe067 PMD 100d97067 PTE 0
> > [   63.912131] Oops: 0000 [#1] PREEMPT SMP PTI
> > [   63.912134] CPU: 3 PID: 534 Comm: kworker/3:10 Not tainted 6.4.0 #6
> > [   63.912137] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./H370M Pro4, BIOS P3.40 10/25/2018
> > [   63.912138] Workqueue: pm pm_runtime_work
> > [   63.912144] RIP: 0010:ioread32+0x2e/0x70
> > [   63.912148] Code: ff 03 00 77 25 48 81 ff 00 00 01 00 77 14 8b 15 08 d9 54 01 b8 ff ff ff ff 85 d2 75 14 c3 cc cc cc cc 89 fa ed c3 cc cc cc cc <8b> 07 c3 cc cc cc cc 55 83 ea 01 48 89 fe 48 c7 c7 98 6f 15 99 48
> > [   63.912150] RSP: 0018:ffffb24d40a5bd78 EFLAGS: 00010296
> > [   63.912152] RAX: ffffb24d403e5000 RBX: 0000000000000152 RCX: 000000000000007f
> > [   63.912153] RDX: 000000000000ff00 RSI: ffffb24d403e5010 RDI: ffffb24d403e5010
> > [   63.912155] RBP: ffffb24d40a5bd98 R08: ffffb24d403e5010 R09: 0000000000000000
> > [   63.912156] R10: ffff9074cd95e7f4 R11: 0000000000000003 R12: 000000000000007f
> > [   63.912158] R13: ffff9074e1a68c00 R14: ffff9074e1a68d00 R15: 0000000000009003
> > [   63.912159] FS:  0000000000000000(0000) GS:ffff90752a180000(0000) knlGS:0000000000000000
> > [   63.912161] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > [   63.912162] CR2: ffffb24d403e5010 CR3: 0000000152832006 CR4: 00000000003706e0
> > [   63.912164] Call Trace:
> > [   63.912165]  <TASK>
> > [   63.912167]  ? show_regs+0x68/0x70
> > [   63.912171]  ? __die_body+0x20/0x70
> > [   63.912173]  ? __die+0x2b/0x40
> > [   63.912175]  ? page_fault_oops+0x160/0x480
> > [   63.912177]  ? search_bpf_extables+0x63/0x90
> > [   63.912180]  ? ioread32+0x2e/0x70
> > [   63.912183]  ? search_exception_tables+0x5f/0x70
> > [   63.912186]  ? kernelmode_fixup_or_oops+0xa2/0x120
> > [   63.912189]  ? __bad_area_nosemaphore+0x179/0x230
> > [   63.912191]  ? bad_area_nosemaphore+0x16/0x20
> > [   63.912193]  ? do_kern_addr_fault+0x8b/0xa0
> > [   63.912195]  ? exc_page_fault+0xe5/0x180
> > [   63.912198]  ? asm_exc_page_fault+0x27/0x30
> > [   63.912203]  ? ioread32+0x2e/0x70
> > [   63.912206]  ? rtsx_pci_write_register+0x5b/0x90 [rtsx_pci]
> > [   63.912217]  rtsx_set_l1off_sub+0x1c/0x30 [rtsx_pci]
> > [   63.912226]  rts5261_set_l1off_cfg_sub_d0+0x36/0x40 [rtsx_pci]
> > [   63.912234]  rtsx_pci_runtime_idle+0xc7/0x160 [rtsx_pci]
> > [   63.912243]  ? __pfx_pci_pm_runtime_idle+0x10/0x10
> > [   63.912246]  pci_pm_runtime_idle+0x34/0x70
> > [   63.912248]  rpm_idle+0xc4/0x2b0
> > [   63.912251]  pm_runtime_work+0x93/0xc0
> > [   63.912254]  process_one_work+0x21a/0x430
> > [   63.912258]  worker_thread+0x4a/0x3c0
> > [   63.912261]  ? __pfx_worker_thread+0x10/0x10
> > [   63.912263]  kthread+0x106/0x140
> > [   63.912266]  ? __pfx_kthread+0x10/0x10
> > [   63.912268]  ret_from_fork+0x29/0x50
> > [   63.912273]  </TASK>
>
> Can you strip out the call trace stuff that's not relevant so the call
> path is clear?  I'm guessing we're in ioread32(), and nothing above
> do_kern_addr_fault() or below worker_thread() is relevant.

Sure.

>
> > [   63.912274] Modules linked in: nvme nvme_core snd_hda_codec_hdmi snd_sof_pci_intel_cnl snd_sof_intel_hda_common snd_hda_codec_realtek snd_hda_codec_generic snd_soc_hdac_hda soundwire_intel ledtrig_audio nls_iso8859_1 soundwire_generic_allocation soundwire_cadence snd_sof_intel_hda_mlink snd_sof_intel_hda snd_sof_pci snd_sof_xtensa_dsp snd_sof snd_sof_utils snd_hda_ext_core snd_soc_acpi_intel_match snd_soc_acpi soundwire_bus snd_soc_core snd_compress ac97_bus snd_pcm_dmaengine snd_hda_intel i915 snd_intel_dspcfg snd_intel_sdw_acpi intel_rapl_msr snd_hda_codec intel_rapl_common snd_hda_core x86_pkg_temp_thermal intel_powerclamp snd_hwdep coretemp snd_pcm kvm_intel drm_buddy ttm mei_hdcp kvm drm_display_helper snd_seq_midi snd_seq_midi_event cec crct10dif_pclmul ghash_clmulni_intel sha512_ssse3 aesni_intel crypto_simd rc_core cryptd rapl snd_rawmidi drm_kms_helper binfmt_misc intel_cstate i2c_algo_bit joydev snd_seq snd_seq_device syscopyarea wmi_bmof snd_timer sysfillrect input_leds snd ee1004 sysimgblt mei_me soundcore
> > [   63.912324]  mei intel_pch_thermal mac_hid acpi_tad acpi_pad sch_fq_codel msr parport_pc ppdev lp ramoops drm parport reed_solomon efi_pstore ip_tables x_tables autofs4 hid_generic usbhid hid rtsx_pci_sdmmc crc32_pclmul ahci e1000e i2c_i801 i2c_smbus rtsx_pci xhci_pci libahci xhci_pci_renesas video wmi
>
> The module list doesn't look relevant here.  Nor the timestamps.
>
> > [   63.912346] CR2: ffffb24d403e5010
> > [   63.912348] ---[ end trace 0000000000000000 ]---
> >
> > This happens because scheduled pm_runtime_idle() is not cancelled.
> >
> > So before releasing the device, stop all runtime power managements by
> > using pm_runtime_barrier() to fix the issue.
> >
> > Link: https://lore.kernel.org/all/2ce258f371234b1f8a1a470d5488d00e@realtek.com/
> > Tested-by: Ricky Wu <ricky_wu@realtek.com>
> > ---
> > v2:
> >   Cover more cases than just pciehp.
> >
> >  drivers/pci/remove.c | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c
> > index d749ea8250d6..c69b4ce5dbfd 100644
> > --- a/drivers/pci/remove.c
> > +++ b/drivers/pci/remove.c
> > @@ -1,6 +1,7 @@
> >  // SPDX-License-Identifier: GPL-2.0
> >  #include <linux/pci.h>
> >  #include <linux/module.h>
> > +#include <linux/pm_runtime.h>
> >  #include "pci.h"
> >
> >  static void pci_free_resources(struct pci_dev *dev)
> > @@ -18,6 +19,7 @@ static void pci_stop_dev(struct pci_dev *dev)
> >       pci_pme_active(dev, false);
> >
> >       if (pci_dev_is_added(dev)) {
> > +             pm_runtime_barrier(&dev->dev);
>
> If pm_runtime_barrier() is really the solution, it seems like this
> should go somewhere in pci-driver.c where we call the driver PM
> callbacks.

What makes pm_runtime_barrier() work is because it calls
'cancel_work_sync(&dev->power.work);' to cancel the pm_runtime_idle()
work, while pm_runtime_forbid() doesn't.
So should pm_runtime_forbid() also use cancel_work_sync()?

Kai-Heng

>
> >               device_release_driver(&dev->dev);
> >               pci_proc_detach_device(dev);
> > --
> > 2.34.1
> >
diff mbox series

Patch

diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c
index d749ea8250d6..c69b4ce5dbfd 100644
--- a/drivers/pci/remove.c
+++ b/drivers/pci/remove.c
@@ -1,6 +1,7 @@ 
 // SPDX-License-Identifier: GPL-2.0
 #include <linux/pci.h>
 #include <linux/module.h>
+#include <linux/pm_runtime.h>
 #include "pci.h"
 
 static void pci_free_resources(struct pci_dev *dev)
@@ -18,6 +19,7 @@  static void pci_stop_dev(struct pci_dev *dev)
 	pci_pme_active(dev, false);
 
 	if (pci_dev_is_added(dev)) {
+		pm_runtime_barrier(&dev->dev);
 
 		device_release_driver(&dev->dev);
 		pci_proc_detach_device(dev);