linux-kernel - Re: Commit 4257f7e0 ("PCI/ASPM: Save/restore L1SS Capability for suspend/resume") causing hibernate resume failures

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

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Date:   Sun, 27 Dec 2020 22:05:13 -0600
From:   Bjorn Helgaas <helgaas@...nel.org>
To:     "Kenneth R. Crudup" <kenny@...ix.com>
Cc:     Vidya Sagar <vidyas@...dia.com>, linux-pci@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: Commit 4257f7e0 ("PCI/ASPM: Save/restore L1SS Capability for
 suspend/resume") causing hibernate resume failures

> From: Kenneth R. Crudup <kenny@...ix.com>
> 
> I've been running Linus' master branch on my laptop (Dell XPS 13
> 2-in-1).  With this commit in place, after resuming from hibernate
> my machine is essentially useless, with a torrent of disk I/O errors
> on my NVMe device (at least, and possibly other devices affected)
> until a reboot.
> 
> I do use tlp to set the PCIe ASPM to "performance" on AC and
> "powersupersave" on battery.

Thanks a lot for the report, and sorry for the breakage.

4257f7e008ea restores PCI_L1SS_CTL1, then PCI_L1SS_CTL2.  I think it
should do those in the reverse order, since the Enable bits are in
PCI_L1SS_CTL1.  It also restores L1SS state (potentially enabling
L1.x) before we restore the PCIe Capability (potentially enabling ASPM
as a whole).  Those probably should also be in the other order.

If it's convenient, can you try the patch below?  If the debug patch
doesn't help:

  - Are you seeing the hibernate/resume problem when on AC or on
    battery?

  - What if you don't use tlp?  Does hibernate/resume work fine then?
    If tlp makes a difference, can you collect "sudo lspci -vv" output
    with and without using tlp (before hibernate)?

  - If you revert 4257f7e008ea, does hibernate/resume work fine?  Both
    with the tlp tweak and without?

  - Collect "sudo lspci -vv" output before hibernate and (if possible)
    after resume when you see the problem.

I guess tlp only uses /sys/module/pcie_aspm/parameters/policy, so it
sets the same ASPM policy system-wide.

Bjorn

diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index b9fecc25d213..6598b5cd3154 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -1665,9 +1665,8 @@ void pci_restore_state(struct pci_dev *dev)
 	 * LTR itself (in the PCIe capability).
 	 */
 	pci_restore_ltr_state(dev);
-	pci_restore_aspm_l1ss_state(dev);
-
 	pci_restore_pcie_state(dev);
+	pci_restore_aspm_l1ss_state(dev);
 	pci_restore_pasid_state(dev);
 	pci_restore_pri_state(dev);
 	pci_restore_ats_state(dev);
diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index a08e7d6dc248..c4a99274b4bb 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -752,8 +752,8 @@ void pci_save_aspm_l1ss_state(struct pci_dev *dev)
 		return;
 
 	cap = (u32 *)&save_state->cap.data[0];
-	pci_read_config_dword(dev, aspm_l1ss + PCI_L1SS_CTL1, cap++);
 	pci_read_config_dword(dev, aspm_l1ss + PCI_L1SS_CTL2, cap++);
+	pci_read_config_dword(dev, aspm_l1ss + PCI_L1SS_CTL1, cap++);
 }
 
 void pci_restore_aspm_l1ss_state(struct pci_dev *dev)
@@ -774,8 +774,8 @@ void pci_restore_aspm_l1ss_state(struct pci_dev *dev)
 		return;
 
 	cap = (u32 *)&save_state->cap.data[0];
-	pci_write_config_dword(dev, aspm_l1ss + PCI_L1SS_CTL1, *cap++);
 	pci_write_config_dword(dev, aspm_l1ss + PCI_L1SS_CTL2, *cap++);
+	pci_write_config_dword(dev, aspm_l1ss + PCI_L1SS_CTL1, *cap++);
 }
 
 static void pcie_config_aspm_dev(struct pci_dev *pdev, u32 val)