From: David Matlack <dmatlack@google.com>
To: Alex Williamson <alex@shazbot.org>
Cc: kexec@lists.infradead.org, linux-doc@vger.kernel.org,
linux-kernel@vger.kernel.org, linux-mm@kvack.org,
linux-pci@vger.kernel.org,
Adithya Jayachandran <ajayachandra@nvidia.com>,
Alexander Graf <graf@amazon.com>,
Bjorn Helgaas <bhelgaas@google.com>, Chris Li <chrisl@kernel.org>,
David Rientjes <rientjes@google.com>,
Jacob Pan <jacob.pan@linux.microsoft.com>,
Jason Gunthorpe <jgg@nvidia.com>,
Jonathan Corbet <corbet@lwn.net>, Josh Hilke <jrhilke@google.com>,
Leon Romanovsky <leonro@nvidia.com>,
Lukas Wunner <lukas@wunner.de>, Mike Rapoport <rppt@kernel.org>,
Parav Pandit <parav@nvidia.com>,
Pasha Tatashin <pasha.tatashin@soleen.com>,
Pranjal Shrivastava <praan@google.com>,
Pratyush Yadav <pratyush@kernel.org>,
Randy Dunlap <rdunlap@infradead.org>,
Saeed Mahameed <saeedm@nvidia.com>,
Samiullah Khawaja <skhawaja@google.com>,
Shuah Khan <skhan@linuxfoundation.org>,
Vipin Sharma <vipinsh@google.com>, William Tu <witu@nvidia.com>,
Yi Liu <yi.l.liu@intel.com>
Subject: Re: [PATCH v9 08/13] PCI: Save and restore the ACS Control register
Date: Mon, 21 Sep 2026 20:36:01 +0000 [thread overview]
Message-ID: <arGVMaxfwEw_PkXy@google.com> (raw)
In-Reply-To: <20260918191846.2f68b23b@shazbot.org>
On 2026-09-18 07:18 PM, Alex Williamson wrote:
> On Fri, 18 Sep 2026 20:06:34 +0000
> David Matlack <dmatlack@google.com> wrote:
>
> > Save the ACS Control register in pci_save_state() and write it back
> > in pci_restore_state(), instead of recomputing the ACS controls from
> > scratch with pci_enable_acs().
> >
> > This makes ACS symmetric with the rest of a device's saved state. Today
> > pci_save_state() ignores ACS entirely and pci_restore_state() re-enables
> > the ACS controls from the kernel's current ACS policy. As a result, a
> > device can come out of a reset with different ACS controls than it went
> > in with, e.g. any controls programmed outside of pci_enable_acs() are
> > silently dropped.
>
> Controls, yes, but if we're restoring controls that might have EC set
> now, should the Egress Control Vector also be part of the save state?
>
> Currently EC always gets cleared on reset and won't be restored by
> pci_enable_acs(), so we can lose both EC and the EC vector. With this,
> I think we restore EC but still lose the EC vector. Thanks,
Hi Alex,
The kernel does not enable Egress Control today or program the vector.
Would this be to cover the case where firmware or userspace enabled it?
Here is an updated patch to save/restore the vector, but I don't have
any devices that support Egress Control on my normal testing system so I
haven't been able to really test it yet.
From: David Matlack <dmatlack@google.com>
Date: Fri, 11 Sep 2026 20:05:04 +0000
Subject: [PATCH] PCI: Save and restore the ACS Control register and Egress
Control Vector
Save the ACS Control register and the ACS Egress Control Vector in
pci_save_state() and write them back in pci_restore_state(), instead of
recomputing the ACS controls from scratch with pci_enable_acs().
This makes ACS symmetric with the rest of a device's saved state. Today
pci_save_state() ignores ACS entirely and pci_restore_state() re-enables
the ACS controls from the kernel's current ACS policy. As a result, a
device can come out of a reset with different ACS controls than it went
in with, e.g. any controls programmed outside of pci_enable_acs() are
silently dropped.
Notably, this prepares the kernel to be able to adopt the ACS controls
established by a previous kernel across a Live Update rather than
assigning new ones through pci_enable_acs().
Save and restore the Egress Control Vector as well to keep it in sync
with the now-properly-restored Egress Control Enable bit in the ACS
control register. The kernel never enables Egress Controls or programs
the vector, but the firmware could have and they need to be kept in sync
to avoid changing how P2P traffic is rounted.
The size of the Egress Control Vector is not known when
pci_allocate_cap_save_buffers() runs, as the ACS Capability register is
only read later, in pci_acs_init(). Rather than move the allocation,
size the save buffer for the largest vector a device can implement,
which costs at most 32 bytes per ACS-capable device.
pci_enable_acs() runs when a driver binds to a device
(pci_dma_configure()), i.e. after pci_bus_add_device() has already saved
the device's state. Refresh the saved ACS state there as well, otherwise
a subsequent reset would revert ACS back to the configuration left
behind by firmware.
Devices that rely on device-specific quirks to enable an ACS equivalent
keep that configuration outside of the ACS Control register, so keep
configuring ACS from scratch for them. Do the same for devices that have
no saved ACS state at all.
Assisted-by: Claude:claude-opus-5
Signed-off-by: David Matlack <dmatlack@google.com>
---
drivers/pci/pci.c | 117 +++++++++++++++++++++++++++++++++-
drivers/pci/pci.h | 5 ++
drivers/pci/quirks.c | 7 ++
include/uapi/linux/pci_regs.h | 1 +
4 files changed, 129 insertions(+), 1 deletion(-)
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index b2879a6be5f8..e1c05866e2ee 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -1021,6 +1021,106 @@ static void pci_std_enable_acs(struct pci_dev *dev, struct pci_acs *caps)
caps->ctrl |= (dev->acs_capabilities & PCI_ACS_TB);
}
+/*
+ * Layout of the ACS save buffer. @ecv holds the ACS Egress Control Vector and
+ * is sized for the largest vector a device can implement.
+ */
+struct pci_acs_saved_state {
+ u16 ctrl;
+ u32 ecv[8];
+};
+
+/*
+ * Return the size in bytes of the ACS Egress Control Vector, or 0 if the
+ * device does not implement P2P Egress Control.
+ */
+static unsigned int pci_acs_ecv_size(struct pci_dev *dev)
+{
+ unsigned int bits;
+
+ if (!dev->acs_cap || !(dev->acs_capabilities & PCI_ACS_EC))
+ return 0;
+
+ /* An Egress Control Vector Size of 0 means 256 bits */
+ bits = FIELD_GET(PCI_ACS_EGRESS_BITS_MASK, dev->acs_capabilities);
+ if (!bits)
+ bits = 256;
+
+ /* The vector is implemented as a series of DWORD registers */
+ return round_up(bits, 32) / 8;
+}
+
+/**
+ * pci_save_acs_state - save the ACS Control register and Egress Control Vector
+ * @dev: the PCI device
+ *
+ * Record the ACS configuration currently programmed in hardware so that
+ * pci_restore_acs_state() can reapply it after a reset.
+ */
+static void pci_save_acs_state(struct pci_dev *dev)
+{
+ struct pci_cap_saved_state *save_state;
+ struct pci_acs_saved_state *acs;
+ unsigned int i, dwords;
+
+ if (!dev->acs_cap)
+ return;
+
+ save_state = pci_find_saved_ext_cap(dev, PCI_EXT_CAP_ID_ACS);
+ if (!save_state)
+ return;
+
+ acs = (struct pci_acs_saved_state *)save_state->cap.data;
+
+ pci_read_config_word(dev, dev->acs_cap + PCI_ACS_CTRL, &acs->ctrl);
+
+ dwords = pci_acs_ecv_size(dev) / sizeof(u32);
+ for (i = 0; i < dwords; i++)
+ pci_read_config_dword(dev, dev->acs_cap + PCI_ACS_EGRESS_CTL_V +
+ i * sizeof(u32), &acs->ecv[i]);
+}
+
+/**
+ * pci_restore_acs_state - restore the ACS Control register and Egress Control
+ * Vector
+ * @dev: the PCI device
+ */
+static void pci_restore_acs_state(struct pci_dev *dev)
+{
+ struct pci_cap_saved_state *save_state = NULL;
+ struct pci_acs_saved_state *acs;
+ unsigned int i, dwords;
+
+ if (dev->acs_cap && !pci_need_dev_specific_enable_acs(dev))
+ save_state = pci_find_saved_ext_cap(dev, PCI_EXT_CAP_ID_ACS);
+
+ /*
+ * Devices that rely on device-specific quirks to enable an ACS
+ * equivalent keep that configuration outside of the ACS Control
+ * register, so there is nothing useful to restore for them. Configure
+ * ACS from scratch instead, which also covers devices that have no
+ * saved ACS state at all.
+ */
+ if (!save_state) {
+ pci_enable_acs(dev);
+ return;
+ }
+
+ acs = (struct pci_acs_saved_state *)save_state->cap.data;
+
+ /*
+ * Restore the Egress Control Vector before the ACS Control register.
+ * The vector resets to zero, so enabling P2P Egress Control first
+ * would briefly apply the reset vector instead of the saved one.
+ */
+ dwords = pci_acs_ecv_size(dev) / sizeof(u32);
+ for (i = 0; i < dwords; i++)
+ pci_write_config_dword(dev, dev->acs_cap + PCI_ACS_EGRESS_CTL_V +
+ i * sizeof(u32), acs->ecv[i]);
+
+ pci_write_config_word(dev, dev->acs_cap + PCI_ACS_CTRL, acs->ctrl);
+}
+
/**
* pci_enable_acs - enable ACS if hardware support it
* @dev: the PCI device
@@ -1057,6 +1157,15 @@ void pci_enable_acs(struct pci_dev *dev)
__pci_config_acs(dev, &caps, config_acs_param, 0, 0);
pci_write_config_word(dev, pos + PCI_ACS_CTRL, caps.ctrl);
+
+ /*
+ * pci_enable_acs() runs when a driver binds to the device, i.e. after
+ * pci_bus_add_device() has already saved the device's state. Refresh
+ * the saved ACS state so that a subsequent reset restores the
+ * configuration programmed here rather than the one left behind by
+ * firmware.
+ */
+ pci_save_acs_state(dev);
}
/**
@@ -1800,6 +1909,7 @@ int pci_save_state(struct pci_dev *dev)
pci_save_aer_state(dev);
pci_save_ptm_state(dev);
pci_save_tph_state(dev);
+ pci_save_acs_state(dev);
return pci_save_vc_state(dev);
}
EXPORT_SYMBOL(pci_save_state);
@@ -1877,7 +1987,7 @@ void pci_restore_state(struct pci_dev *dev)
pci_restore_msi_state(dev);
/* Restore ACS and IOV configuration state */
- pci_enable_acs(dev);
+ pci_restore_acs_state(dev);
pci_restore_iov_state(dev);
dev->state_saved = false;
@@ -3532,6 +3642,11 @@ void pci_allocate_cap_save_buffers(struct pci_dev *dev)
if (error)
pci_err(dev, "unable to allocate suspend buffer for LTR\n");
+ error = pci_add_ext_cap_save_buffer(dev, PCI_EXT_CAP_ID_ACS,
+ sizeof(struct pci_acs_saved_state));
+ if (error)
+ pci_err(dev, "unable to allocate suspend buffer for ACS\n");
+
pci_allocate_vc_save_buffers(dev);
}
diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h
index ba3c3fddddc2..037c1674f164 100644
--- a/drivers/pci/pci.h
+++ b/drivers/pci/pci.h
@@ -1095,6 +1095,7 @@ void pci_acs_init(struct pci_dev *dev);
void pci_enable_acs(struct pci_dev *dev);
#ifdef CONFIG_PCI_QUIRKS
int pci_dev_specific_acs_enabled(struct pci_dev *dev, u16 acs_flags);
+bool pci_need_dev_specific_enable_acs(struct pci_dev *dev);
int pci_dev_specific_enable_acs(struct pci_dev *dev);
int pci_dev_specific_disable_acs_redir(struct pci_dev *dev);
void pci_disable_broken_acs_cap(struct pci_dev *pdev);
@@ -1105,6 +1106,10 @@ static inline int pci_dev_specific_acs_enabled(struct pci_dev *dev,
{
return -ENOTTY;
}
+static inline bool pci_need_dev_specific_enable_acs(struct pci_dev *dev)
+{
+ return false;
+}
static inline int pci_dev_specific_enable_acs(struct pci_dev *dev)
{
return -ENOTTY;
diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
index 7aee30734303..e500c202d2ec 100644
--- a/drivers/pci/quirks.c
+++ b/drivers/pci/quirks.c
@@ -5476,6 +5476,13 @@ static const struct pci_dev_acs_ops *pci_dev_acs_ops_get(struct pci_dev *dev)
return NULL;
}
+bool pci_need_dev_specific_enable_acs(struct pci_dev *dev)
+{
+ const struct pci_dev_acs_ops *p = pci_dev_acs_ops_get(dev);
+
+ return p && p->enable_acs;
+}
+
int pci_dev_specific_enable_acs(struct pci_dev *dev)
{
const struct pci_dev_acs_ops *p = pci_dev_acs_ops_get(dev);
diff --git a/include/uapi/linux/pci_regs.h b/include/uapi/linux/pci_regs.h
index facaa324bd86..66359cb94f0d 100644
--- a/include/uapi/linux/pci_regs.h
+++ b/include/uapi/linux/pci_regs.h
@@ -1023,6 +1023,7 @@
#define PCI_ACS_UF 0x0010 /* Upstream Forwarding */
#define PCI_ACS_EC 0x0020 /* P2P Egress Control */
#define PCI_ACS_DT 0x0040 /* Direct Translated P2P */
+#define PCI_ACS_EGRESS_BITS_MASK 0xff00 /* Egress Control Vector Size */
#define PCI_ACS_EGRESS_BITS 0x05 /* ACS Egress Control Vector Size */
#define PCI_ACS_CTRL 0x06 /* ACS Control Register */
#define PCI_ACS_EGRESS_CTL_V 0x08 /* ACS Egress Control Vector */
next prev parent reply other threads:[~2026-09-21 20:36 UTC|newest]
Thread overview: 45+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-09-18 20:06 [PATCH v9 00/13] PCI: liveupdate: PCI core support for Live Update David Matlack
2026-09-18 20:06 ` [PATCH v9 01/13] PCI: liveupdate: Set up FLB handler for the PCI core David Matlack
2026-09-18 20:15 ` sashiko-bot
2026-09-25 19:55 ` Bjorn Helgaas
2026-09-18 20:06 ` [PATCH v9 02/13] PCI: liveupdate: Track outgoing preserved PCI devices David Matlack
2026-09-18 20:23 ` sashiko-bot
2026-09-18 20:06 ` [PATCH v9 03/13] PCI: liveupdate: Track incoming " David Matlack
2026-09-18 20:20 ` sashiko-bot
2026-09-25 19:59 ` Bjorn Helgaas
2026-09-18 20:06 ` [PATCH v9 04/13] PCI: liveupdate: Document driver binding responsibilities David Matlack
2026-09-18 20:09 ` sashiko-bot
2026-09-18 20:06 ` [PATCH v9 05/13] PCI: liveupdate: Auto-preserve upstream bridges across Live Update David Matlack
2026-09-18 20:14 ` sashiko-bot
2026-09-25 20:02 ` Bjorn Helgaas
2026-09-18 20:06 ` [PATCH v9 06/13] PCI: liveupdate: Preserve bus numbers during " David Matlack
2026-09-18 20:13 ` sashiko-bot
2026-09-25 20:04 ` Bjorn Helgaas
2026-09-28 17:16 ` Pranjal Shrivastava
2026-09-18 20:06 ` [PATCH v9 07/13] PCI: Refactor matching logic for pci_dev_acs_ops David Matlack
2026-09-18 20:13 ` sashiko-bot
2026-09-18 20:06 ` [PATCH v9 08/13] PCI: Save and restore the ACS Control register David Matlack
2026-09-18 20:15 ` sashiko-bot
2026-09-19 1:18 ` Alex Williamson
2026-09-21 20:36 ` David Matlack [this message]
2026-10-07 9:20 ` David Matlack
2026-09-28 17:18 ` Pranjal Shrivastava
2026-09-18 20:06 ` [PATCH v9 09/13] PCI: liveupdate: Adopt ACS controls in incoming preserved devices David Matlack
2026-09-18 20:13 ` sashiko-bot
2026-09-28 17:19 ` Pranjal Shrivastava
2026-09-18 20:06 ` [PATCH v9 10/13] PCI: liveupdate: Adopt ARI Forwarding Enable on preserved bridges David Matlack
2026-09-18 20:19 ` sashiko-bot
2026-09-18 20:06 ` [PATCH v9 11/13] PCI: liveupdate: Freeze preservation status during shutdown David Matlack
2026-09-18 20:19 ` sashiko-bot
2026-09-18 20:06 ` [PATCH v9 12/13] PCI: liveupdate: Do not disable bus mastering on preserved devices during kexec David Matlack
2026-09-18 20:17 ` sashiko-bot
2026-09-18 20:06 ` [PATCH v9 13/13] Documentation: PCI: Add documentation for Live Update David Matlack
2026-09-18 20:14 ` sashiko-bot
2026-09-22 18:36 ` [PATCH v9 00/13] PCI: liveupdate: PCI core support " Zhu Yanjun
2026-09-22 18:53 ` David Matlack
[not found] ` <d4547411-eeae-46c5-a30b-12d8cdde9329@linux.dev>
2026-09-24 21:59 ` David Matlack
2026-09-25 4:47 ` Zhu Yanjun
2026-09-25 18:59 ` Samiullah Khawaja
2026-09-22 18:54 ` Zhu Yanjun
2026-09-22 21:47 ` Zhu Yanjun
2026-09-28 15:46 ` Pasha Tatashin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=arGVMaxfwEw_PkXy@google.com \
--to=dmatlack@google.com \
--cc=ajayachandra@nvidia.com \
--cc=alex@shazbot.org \
--cc=bhelgaas@google.com \
--cc=chrisl@kernel.org \
--cc=corbet@lwn.net \
--cc=graf@amazon.com \
--cc=jacob.pan@linux.microsoft.com \
--cc=jgg@nvidia.com \
--cc=jrhilke@google.com \
--cc=kexec@lists.infradead.org \
--cc=leonro@nvidia.com \
--cc=linux-doc@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=linux-pci@vger.kernel.org \
--cc=lukas@wunner.de \
--cc=parav@nvidia.com \
--cc=pasha.tatashin@soleen.com \
--cc=praan@google.com \
--cc=pratyush@kernel.org \
--cc=rdunlap@infradead.org \
--cc=rientjes@google.com \
--cc=rppt@kernel.org \
--cc=saeedm@nvidia.com \
--cc=skhan@linuxfoundation.org \
--cc=skhawaja@google.com \
--cc=vipinsh@google.com \
--cc=witu@nvidia.com \
--cc=yi.l.liu@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®