RE: [PATCHv2 4/4] PCI: dwc: add prefetchable memory range support

From: Z.q. Hou
Date: Tue Dec 11 2018 - 05:21:17 EST

Next message: Ulf Hansson: "Re: [PATCH v6 10/10] soc: qcom: rpmhpd: Mark mx as a parent for cx"
Previous message: Christoffer Dall: "Re: [PATCH v2 1/4] KVM: arm/arm64: vgic: Do not cond_resched_lock() with IRQs disabled"
Next in thread: Lorenzo Pieralisi: "Re: [PATCHv2 4/4] PCI: dwc: add prefetchable memory range support"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi Lorenzo,

Thanks a lot for your comments!

> -----Original Message-----
> From: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>
> Sent: 2018年11月29日 2:00
> To: Z.q. Hou <zhiqiang.hou@xxxxxxx>
> Cc: linux-pci@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> bhelgaas@xxxxxxxxxx; jingoohan1@xxxxxxxxx;
> gustavo.pimentel@xxxxxxxxxxxx; Roy Zang <roy.zang@xxxxxxx>; Mingkai Hu
> <mingkai.hu@xxxxxxx>; M.h. Lian <minghuan.lian@xxxxxxx>
> Subject: Re: [PATCHv2 4/4] PCI: dwc: add prefetchable memory range support
>
> On Wed, Nov 07, 2018 at 10:09:21AM +0000, Z.q. Hou wrote:
> > From: Hou Zhiqiang <Zhiqiang.Hou@xxxxxxx>
> >
> > The current code only support non-prefetchable memory range, as the
> > non-prefetchable memory range must not be greater than 4GiB, one
> > viewport can cover it, which supports upto 4GiB.
> >
> > To support prefetchable memory range, which is upto 64-bit memory
> > space and can be greater than 4GiB, so we need multiple viewports. And
> > added separate vars to store prefetchable memory range info to prevent
> > overriding the non-prefetchable memory range info.
> >
> > And this patch explicitly assigned the last (if there are only
> > 2 viewports) or last 2 viewports for CFG and I/O windows and the rests
> > for MEM windows.
> >
> > Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@xxxxxxx>
> > ---
> > V2:
> > - Reworded the subject and commit description.
> > - Fix the prefetchable memory range overriding non-perfetchable
> > memory range issue by adding vars to store prefetchable memory
> > range info.
> >
> > .../pci/controller/dwc/pcie-designware-host.c | 107 ++++++++++++++----
> > drivers/pci/controller/dwc/pcie-designware.h | 7 ++
> > 2 files changed, 95 insertions(+), 19 deletions(-)
> >
> > diff --git a/drivers/pci/controller/dwc/pcie-designware-host.c
> > b/drivers/pci/controller/dwc/pcie-designware-host.c
> > index ecacce016489..328aa40a6609 100644
> > --- a/drivers/pci/controller/dwc/pcie-designware-host.c
> > +++ b/drivers/pci/controller/dwc/pcie-designware-host.c
> > @@ -346,6 +346,35 @@ int dw_pcie_host_init(struct pcie_port *pp)
> > dev_err(dev, "Missing *config* reg space\n");
> > }
> >
> > + /*
> > + * If vendor's platform driver has set the num_viewport and it is
> > + * not less than 2, skip getting the num_viewport from DT here.
>
> That's not what the device tree bindings specify. If DT does contain the
> property you *must* read it.

Yes, will change back to *must* read.

>
> > + */
> > + if (pci->num_viewport < 2) {
> > + ret = of_property_read_u32(np, "num-viewport",
> > + &pci->num_viewport);
> > + if (ret || pci->num_viewport < 2)
> > + pci->num_viewport = 2;
> > + }
> > +
> > + /*
> > + * if there are only 2 viewports, assign the last viewport for
> > + * both CFG and IO window, otherwise assign the last 2 viewport
>
> Gah. Can anyone explain to me how this driver works if only two viewports
> are available ? What happens if an IO access happens at the same time of a
> config access (that fiddles with the outbound memory windows) ?

You are right it has potential IO transaction drop issue, but DWC PCIe databook does give a default value 2 of viewport number.

> > + * for CFG and IO window specific. And the rest viewports are
> > + * assigned to MEM windows.
> > + */
> > + if (pci->num_viewport == 2) {
> > + pp->cfg_idx = pp->io_idx = PCIE_ATU_REGION_INDEX1;
> > + pp->mem_wins = 1;
> > + } else {
> > + pp->cfg_idx = pci->num_viewport - 1;
> > + pp->io_idx = pci->num_viewport - 2;
> > + pp->mem_wins = pci->num_viewport - 2;
> > + }
> > +
> > + dev_dbg(dev, "CFG win id: %d, I/O win id: %d, Total MEM win: %d\n",
> > + pp->cfg_idx, pp->io_idx, pp->mem_wins);
> > +
> > bridge = pci_alloc_host_bridge(0);
> > if (!bridge)
> > return -ENOMEM;
> > @@ -377,10 +406,20 @@ int dw_pcie_host_init(struct pcie_port *pp)
> > }
> > break;
> > case IORESOURCE_MEM:
> > - pp->mem = win->res;
> > - pp->mem->name = "MEM";
> > - pp->mem_size = resource_size(pp->mem);
> > - pp->mem_bus_addr = pp->mem->start - win->offset;
> > + if (win->res->flags & IORESOURCE_PREFETCH) {
> > + pp->mem_perf = win->res;
> > + pp->mem_perf->name = "MEM perf";
>
> Nit: Why "perf" and not "pref" ?
>
> It is confusing but that's the least of this patch problems.

My bad, it is typo, will fix in next version.

>
> > + pp->mem_perf_size = resource_size(pp->mem_perf);
> > + pp->mem_perf_bus_addr = pp->mem_perf->start -
> > + win->offset;
> > + pp->mem_perf_base = pp->mem_perf->start;
> > + } else {
> > + pp->mem = win->res;
> > + pp->mem->name = "MEM";
> > + pp->mem_size = resource_size(pp->mem);
> > + pp->mem_bus_addr = pp->mem->start - win->offset;
> > + pp->mem_base = pp->mem->start;
> > + }
> > break;
> > case 0:
> > pp->cfg = win->res;
> > @@ -406,8 +445,6 @@ int dw_pcie_host_init(struct pcie_port *pp)
> > }
> > }
> >
> > - pp->mem_base = pp->mem->start;
> > -
> > if (!pp->va_cfg0_base) {
> > pp->va_cfg0_base = devm_pci_remap_cfgspace(dev,
> > pp->cfg0_base, pp->cfg0_size);
> > @@ -534,12 +571,12 @@ static int dw_pcie_rd_other_conf(struct
> pcie_port *pp, struct pci_bus *bus,
> > va_cfg_base = pp->va_cfg1_base;
> > }
> >
> > - dw_pcie_prog_outbound_atu(pci, PCIE_ATU_REGION_INDEX1,
> > + dw_pcie_prog_outbound_atu(pci, pp->cfg_idx,
> > type, cpu_addr,
> > busdev, cfg_size);
> > ret = dw_pcie_read(va_cfg_base + where, size, val);
> > - if (pci->num_viewport <= 2)
> > - dw_pcie_prog_outbound_atu(pci, PCIE_ATU_REGION_INDEX1,
> > + if (pp->cfg_idx == pp->io_idx)
> > + dw_pcie_prog_outbound_atu(pci, pp->io_idx,
> > PCIE_ATU_TYPE_IO, pp->io_base,
> > pp->io_bus_addr, pp->io_size);
>
> See above, even though this is not related to this patch.

I think it is more reasonable to check if the CFG and IO shared viewport than the viewport number.

>
> >
> > @@ -573,12 +610,12 @@ static int dw_pcie_wr_other_conf(struct
> pcie_port *pp, struct pci_bus *bus,
> > va_cfg_base = pp->va_cfg1_base;
> > }
> >
> > - dw_pcie_prog_outbound_atu(pci, PCIE_ATU_REGION_INDEX1,
> > + dw_pcie_prog_outbound_atu(pci, pp->cfg_idx,
> > type, cpu_addr,
> > busdev, cfg_size);
> > ret = dw_pcie_write(va_cfg_base + where, size, val);
> > - if (pci->num_viewport <= 2)
> > - dw_pcie_prog_outbound_atu(pci, PCIE_ATU_REGION_INDEX1,
> > + if (pp->cfg_idx == pp->io_idx)
> > + dw_pcie_prog_outbound_atu(pci, pp->io_idx,
> > PCIE_ATU_TYPE_IO, pp->io_base,
> > pp->io_bus_addr, pp->io_size);
> >
> > @@ -652,6 +689,9 @@ static u8 dw_pcie_iatu_unroll_enabled(struct
> > dw_pcie *pci) void dw_pcie_setup_rc(struct pcie_port *pp) {
> > u32 val, ctrl, num_ctrls;
> > + u64 remain_size, base, win_size;
> > + phys_addr_t bus_addr;
> > + int i;
> > struct dw_pcie *pci = to_dw_pcie_from_pp(pp);
> >
> > dw_pcie_setup(pci);
> > @@ -700,13 +740,42 @@ void dw_pcie_setup_rc(struct pcie_port *pp)
> > dev_dbg(pci->dev, "iATU unroll: %s\n",
> > pci->iatu_unroll_enabled ? "enabled" : "disabled");
> >
> > - dw_pcie_prog_outbound_atu(pci, PCIE_ATU_REGION_INDEX0,
> > - PCIE_ATU_TYPE_MEM, pp->mem_base,
> > - pp->mem_bus_addr, pp->mem_size);
> > - if (pci->num_viewport > 2)
> > - dw_pcie_prog_outbound_atu(pci, PCIE_ATU_REGION_INDEX2,
> > - PCIE_ATU_TYPE_IO, pp->io_base,
> > - pp->io_bus_addr, pp->io_size);
> > + /*
> > + * The maximum region size is 4 GB, and a region
> > + * must not cross a 4 GB boundary.
> > + */
> > + win_size = SZ_4G - (pp->mem_base & (SZ_4G - 1));
> > + win_size = min(win_size, pp->mem_size);
> > + dw_pcie_prog_outbound_atu(pci, 0, PCIE_ATU_TYPE_MEM,
> > + pp->mem_base, pp->mem_bus_addr,
> > + win_size);
> > + dev_dbg(pci->dev,
> > + "iATU: non-pref MEM: win = %d: base = %llx, bus_addr = %pa,
> size = %llx\n",
> > + 0, pp->mem_base, &pp->mem_bus_addr, win_size);
> > +
> > + /* Prefetchable range can be 64bit space */
> > + remain_size = pp->mem_perf_size;
>
> unallocated/free_size ?

Sorry, I don't understand your question.

>
> > + base = pp->mem_perf_base;
> > + bus_addr = pp->mem_perf_bus_addr;
> > + for (i = 1; remain_size > 0 && i < pp->mem_wins; i++) {
> > + win_size = SZ_4G - (base & (SZ_4G - 1));
> > + win_size = min(win_size, remain_size);
> > + dw_pcie_prog_outbound_atu(pci, i, PCIE_ATU_TYPE_MEM,
> > + base, bus_addr, win_size);
> > + dev_dbg(pci->dev, "iATU: pref MEM: win = %d: base = %llx,
> bus_addr = %pa, size = %llx\n",
> > + i, base, &bus_addr, win_size);
> > +
> > + base += win_size;
> > + bus_addr += win_size;
> > + remain_size -= win_size;
> > + }
> > +
> > + if (remain_size > 0)
> > + dev_info(pci->dev, "iATU: MEM window isn't enough\n");
> > +
> > + dw_pcie_prog_outbound_atu(pci, pp->io_idx, PCIE_ATU_TYPE_IO,
> > + pp->io_base, pp->io_bus_addr,
> > + pp->io_size);
> > }
> >
> > dw_pcie_wr_own_conf(pp, PCI_BASE_ADDRESS_0, 4, 0); diff --git
> > a/drivers/pci/controller/dwc/pcie-designware.h
> > b/drivers/pci/controller/dwc/pcie-designware.h
> > index a438c3879aa9..0197f67f82b7 100644
> > --- a/drivers/pci/controller/dwc/pcie-designware.h
> > +++ b/drivers/pci/controller/dwc/pcie-designware.h
> > @@ -148,15 +148,22 @@ struct pcie_port {
> > u64 cfg1_base;
> > void __iomem *va_cfg1_base;
> > u32 cfg1_size;
> > + u32 cfg_idx;
> > resource_size_t io_base;
> > phys_addr_t io_bus_addr;
> > u32 io_size;
> > + u32 io_idx;
> > u64 mem_base;
> > phys_addr_t mem_bus_addr;
> > u64 mem_size;
> > + u64 mem_perf_base;
>
> This is a phys_addr_t
>
> > + phys_addr_t mem_perf_bus_addr;
>
> pci_bus_addr_t ?

Will fix in next version.

>
> Lorenzo
>
> > + u64 mem_perf_size;
> > + u32 mem_wins;
> > struct resource *cfg;
> > struct resource *io;
> > struct resource *mem;
> > + struct resource *mem_perf;
> > struct resource *busn;
> > int irq;
> > const struct dw_pcie_host_ops *ops;
> > --
> > 2.17.1
> >

Thanks,
Zhiqiang

Next message: Ulf Hansson: "Re: [PATCH v6 10/10] soc: qcom: rpmhpd: Mark mx as a parent for cx"
Previous message: Christoffer Dall: "Re: [PATCH v2 1/4] KVM: arm/arm64: vgic: Do not cond_resched_lock() with IRQs disabled"
Next in thread: Lorenzo Pieralisi: "Re: [PATCHv2 4/4] PCI: dwc: add prefetchable memory range support"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]