From: Bjorn Helgaas <helgaas@kernel.org>
To: "Christian König" <christian.koenig@amd.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>,
linux-pci@vger.kernel.org,
Alex Deucher <alexander.deucher@amd.com>
Subject: Re: Resizeable PCI BAR support
Date: Mon, 7 Dec 2015 10:00:54 -0600 [thread overview]
Message-ID: <20151207160054.GF7994@localhost> (raw)
In-Reply-To: <5662C61F.50303@amd.com>
On Sat, Dec 05, 2015 at 12:10:23PM +0100, Christian König wrote:
> Hi Bjorn,
>
> reading up on the mailing list this is probably not the first time
> this topic comes up, but this time it is not just for FPGAs but for
> really end user hardware.
>
> Alex and I are the maintainers of the Radeon and Amdgpu kernel
> drivers for the AMD graphics hardware. Now the newest hardware
> generation of those devices started to support resizeable PCI BARs
> as specified here https://pcisig.com/sites/default/files/specification_documents/ECN_Resizable-BAR_24Apr2008.pdf.
>
> That feature is rather important for graphics hardware, because the
> PCI BARs are usually limited to 256MB while on modern cards you can
> easily find 4GB or more VRAM. The end result is that only a fraction
> of that VRAM is CPU accessible, causing a whole bunch of workarounds
> in the driver stack for that hardware.
>
> We of course want to avoid that hassle and so I've attached two
> patches created to test the functionality. The first one is just
> infrastructure for handling the extension and the second is a hack
> to resize all BARs to their maximum before probing them by the PCI
> subsystem.
>
> That resizing works astonishing fine, the only problem is that on
> practical no system I've tested this feature actually works. The
> reason is that the PCI root bridge doesn't get enough address space
> assigned by the BIOS for this and the PCI subsystem in the kernel
> isn't able to reprogram it.
>
> Any clever idea how to get this working without waiting for system
> BIOS providers to pick up that extension? Would adding support for
> reprogramming the root bridges be possible?
Many host bridges support programmable windows in the hardware, so in
principle we could reprogram them at run-time. But the CPU side of
the bridge is out of the PCI domain, so there's no single way to do it
-- it's all implementation-specific. ACPI has some pieces that would
make it more generic (_PRS tells us whether the bridge supports
reprogramming, the Linux resource map tells us what address space is
available, _SRS lets us actually reprogram the bridge), but I don't
know whether any BIOSes actually support _PRS and _SRS on host
bridges. I don't know of any OS support that would use it.
Bottom line, I don't have any clever ideas for how to do this.
> From ea99917e1760b2a1ecd6d353d7a8a060ea24dd03 Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig@amd.com>
> Date: Wed, 25 Nov 2015 12:46:29 +0100
> Subject: [PATCH 1/2] PCI: add resizeable BAR infrastructure
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Just the defines and helper functions to read the possible sizes of a BAR and
> update it's size.
>
> See https://pcisig.com/sites/default/files/specification_documents/ECN_Resizable-BAR_24Apr2008.pdf.
>
> Signed-off-by: Christian K??nig <christian.koenig@amd.com>
> ---
> drivers/pci/pci.c | 78 +++++++++++++++++++++++++++++++++++++++++++
> include/linux/pci.h | 2 ++
> include/uapi/linux/pci_regs.h | 7 ++++
> 3 files changed, 87 insertions(+)
>
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 6a9a111..862247f 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -2453,6 +2453,84 @@ bool pci_acs_path_enabled(struct pci_dev *start,
> }
>
> /**
> + * pci_rbar_get_sizes - get possible sizes for BAR
> + * @dev: PCI device
> + * @bar: BAR to query
> + *
> + * Get the possible sizes of a resizeable BAR as bitmask defined in the spec
> + * (bit 0=1MB, bit 19=512GB). Returns 0 if BAR isn't resizeable.
> + */
> +u32 pci_rbar_get_sizes(struct pci_dev *pdev, int bar)
> +{
> + int pos, nbars;
> + u32 ctrl, cap;
> + int i;
> +
> + pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_REBAR);
> + if (!pos)
> + return 0x0;
> +
> + pci_read_config_dword(pdev, pos + PCI_REBAR_CTRL, &ctrl);
> + nbars = (ctrl & PCI_REBAR_CTRL_NBAR_MASK) >> PCI_REBAR_CTRL_NBAR_SHIFT;
> +
> + for (i = 0; i < nbars; ++i, pos += 8) {
> + int bar_idx;
> +
> + pci_read_config_dword(pdev, pos + PCI_REBAR_CTRL, &ctrl);
> + bar_idx = (ctrl & PCI_REBAR_CTRL_BAR_IDX_MASK) >>
> + PCI_REBAR_CTRL_BAR_IDX_SHIFT;
> + if (bar_idx != bar)
> + continue;
> +
> + pci_read_config_dword(pdev, pos + PCI_REBAR_CAP, &cap);
> + return (cap & PCI_REBAR_CTRL_SIZES_MASK) >>
> + PCI_REBAR_CTRL_SIZES_SHIFT;
> + }
> +
> + return 0x0;
> +}
> +
> +/**
> + * pci_rbar_set_size - set a new size for a BAR
> + * @dev: PCI device
> + * @bar: BAR to set size to
> + * @size: new size as defined in the spec.
> + *
> + * Set the new size of a BAR as defined in the spec (0=1MB, 19=512GB).
> + * Returns true if resizing was successful, false otherwise.
> + */
> +bool pci_rbar_set_size(struct pci_dev *pdev, int bar, int size)
> +{
> + int pos, nbars;
> + u32 ctrl;
> + int i;
> +
> + pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_REBAR);
> + if (!pos)
> + return false;
> +
> + pci_read_config_dword(pdev, pos + PCI_REBAR_CTRL, &ctrl);
> + nbars = (ctrl & PCI_REBAR_CTRL_NBAR_MASK) >> PCI_REBAR_CTRL_NBAR_SHIFT;
> +
> + for (i = 0; i < nbars; ++i, pos += 8) {
> + int bar_idx;
> +
> + pci_read_config_dword(pdev, pos + PCI_REBAR_CTRL, &ctrl);
> + bar_idx = (ctrl & PCI_REBAR_CTRL_BAR_IDX_MASK) >>
> + PCI_REBAR_CTRL_BAR_IDX_SHIFT;
> + if (bar_idx != bar)
> + continue;
> +
> + ctrl &= ~PCI_REBAR_CTRL_BAR_SIZE_MASK;
> + ctrl |= size << PCI_REBAR_CTRL_BAR_SIZE_SHIFT;
> + pci_write_config_dword(pdev, pos + PCI_REBAR_CTRL, ctrl);
> + return true;
> + }
> +
> + return false;
> +}
> +
> +/**
> * pci_swizzle_interrupt_pin - swizzle INTx for device behind bridge
> * @dev: the PCI device
> * @pin: the INTx pin (1=INTA, 2=INTB, 3=INTC, 4=INTD)
> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index e90eb22..fafb974 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -1801,6 +1801,8 @@ void pci_request_acs(void);
> bool pci_acs_enabled(struct pci_dev *pdev, u16 acs_flags);
> bool pci_acs_path_enabled(struct pci_dev *start,
> struct pci_dev *end, u16 acs_flags);
> +u32 pci_rbar_get_sizes(struct pci_dev *pdev, int bar);
> +bool pci_rbar_set_size(struct pci_dev *pdev, int bar, int size);
>
> #define PCI_VPD_LRDT 0x80 /* Large Resource Data Type */
> #define PCI_VPD_LRDT_ID(x) ((x) | PCI_VPD_LRDT)
> diff --git a/include/uapi/linux/pci_regs.h b/include/uapi/linux/pci_regs.h
> index 413417f..b746cbb 100644
> --- a/include/uapi/linux/pci_regs.h
> +++ b/include/uapi/linux/pci_regs.h
> @@ -886,9 +886,16 @@
> #define PCI_SATA_SIZEOF_LONG 16
>
> /* Resizable BARs */
> +#define PCI_REBAR_CAP 4 /* capability register */
> +#define PCI_REBAR_CTRL_SIZES_MASK (0xFFFFF << 4) /* mask for sizes */
> +#define PCI_REBAR_CTRL_SIZES_SHIFT 4 /* shift for sizes */
> #define PCI_REBAR_CTRL 8 /* control register */
> +#define PCI_REBAR_CTRL_BAR_IDX_MASK (7 << 0) /* mask for bar index */
> +#define PCI_REBAR_CTRL_BAR_IDX_SHIFT 0 /* shift for bar index */
> #define PCI_REBAR_CTRL_NBAR_MASK (7 << 5) /* mask for # bars */
> #define PCI_REBAR_CTRL_NBAR_SHIFT 5 /* shift for # bars */
> +#define PCI_REBAR_CTRL_BAR_SIZE_MASK (0x1F << 8) /* mask for bar size */
> +#define PCI_REBAR_CTRL_BAR_SIZE_SHIFT 8 /* shift for bar size */
>
> /* Dynamic Power Allocation */
> #define PCI_DPA_CAP 4 /* capability register */
> --
> 2.5.0
>
> From b7201b6af1b3f5452549351b018a27806ee5a723 Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig@amd.com>
> Date: Wed, 25 Nov 2015 12:51:24 +0100
> Subject: [PATCH 2/2] PCI: WIP hack to resize all BARs to their maximum
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig@amd.com>
> ---
> drivers/pci/probe.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index 8361d27..1d8fc20 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -318,6 +318,17 @@ static void pci_read_bases(struct pci_dev *dev, unsigned int howmany, int rom)
>
> for (pos = 0; pos < howmany; pos++) {
> struct resource *res = &dev->resource[pos];
> + u32 sizes = pci_rbar_get_sizes(dev, pos);
> +
> + if (sizes) {
> + /* Just resize to the maximum for now */
> + int size = fls(sizes) - 1;
> +
> + if (pci_rbar_set_size(dev, pos, size))
> + dev_info(&dev->dev, "Resized BAR %d to %dMB\n",
> + pos, 1 << size);
> + }
> +
> reg = PCI_BASE_ADDRESS_0 + (pos << 2);
> pos += __pci_read_base(dev, pci_bar_unknown, res, reg);
> }
> --
> 2.5.0
>
prev parent reply other threads:[~2015-12-07 16:01 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2015-12-05 11:10 Resizeable PCI BAR support Christian König
2015-12-05 13:27 ` kbuild test robot
2015-12-07 16:00 ` Bjorn Helgaas [this message]
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=20151207160054.GF7994@localhost \
--to=helgaas@kernel.org \
--cc=alexander.deucher@amd.com \
--cc=bhelgaas@google.com \
--cc=christian.koenig@amd.com \
--cc=linux-pci@vger.kernel.org \
/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;
as well as URLs for NNTP newsgroup(s).