On 21-09-21 16:39:30, Dan Williams wrote:
> On Tue, Sep 21, 2021 at 3:05 PM Ben Widawsky <ben.widawsky@xxxxxxxxx> wrote:
> >
> > In preparation for moving parts of register mapping to cxl_core, the
> > cxl_pci driver is refactored to utilize a new helper to find register
> > blocks by type.
> >
> > cxl_pci scanned through all register blocks and mapping the ones that
> > the driver will use. This logic is inverted so that the driver
> > specifically requests the register blocks from a new helper. Under the
> > hood, the same implementation of scanning through all register locator
> > DVSEC entries exists.
> >
> > There are 2 behavioral changes (#2 is arguable):
> > 1. A dev_err is introduced if cxl_map_regs fails.
> > 2. The previous logic would try to map component registers and device
> > registers multiple times if there were present and keep the mapping
> > of the last one found (furthest offset in the register locator).
> > While this is disallowed in the spec, CXL 2.0 8.1.9: "Each register
> > block identifier shall only occur once in the Register Locator DVSEC
> > structure" it was how the driver would respond to the spec violation.
> > The new logic will take the first found register block by type and
> > move on.
>
> Yeah, I think it's silly to try to predict how hardware might violate
> the specification. Just wait until there is a known shipping device
> with a problem and then add a quirk to the driver.
>
> >
> > Signed-off-by: Ben Widawsky <ben.widawsky@xxxxxxxxx>
> > ---
> > drivers/cxl/pci.c | 113 ++++++++++++++++++++++++++--------------------
> > 1 file changed, 65 insertions(+), 48 deletions(-)
> >
> > diff --git a/drivers/cxl/pci.c b/drivers/cxl/pci.c
> > index ccc7c2573ddc..6e5c026f5262 100644
> > --- a/drivers/cxl/pci.c
> > +++ b/drivers/cxl/pci.c
> > @@ -428,46 +428,28 @@ static void cxl_decode_register_block(u32 reg_lo, u32 reg_hi,
> > *reg_type = FIELD_GET(CXL_REGLOC_RBI_MASK, reg_lo);
> > }
> >
> > -/**
> > - * cxl_pci_setup_regs() - Setup necessary MMIO.
> > - * @cxlm: The CXL memory device to communicate with.
> > - *
> > - * Return: 0 if all necessary registers mapped.
> > - *
> > - * A memory device is required by spec to implement a certain set of MMIO
> > - * regions. The purpose of this function is to enumerate and map those
> > - * registers.
> > - */
> > -static int cxl_pci_setup_regs(struct cxl_mem *cxlm)
> > +static int find_register_block(struct pci_dev *pdev, enum cxl_regloc_type type,
> > + struct cxl_register_map *map)
> > {
> > - void __iomem *base;
> > + int regloc, i, rc = -ENODEV;
> > u32 regloc_size, regblocks;
> > - int regloc, i, n_maps, ret = 0;
> > - struct device *dev = cxlm->dev;
> > - struct pci_dev *pdev = to_pci_dev(dev);
> > - struct cxl_register_map *map, maps[CXL_REGLOC_RBI_TYPES];
> > +
> > + memset(map, 0, sizeof(*map));
>
> Is this necessary? It seems this fills in all fields on success, why
> does it need to zero-init?
>
Am earlier version of this patch attempted to determine success based on the
values in @map. It is no longer necessary.
> >
> > regloc = cxl_pci_dvsec(pdev, PCI_DVSEC_ID_CXL_REGLOC_DVSEC_ID);
> > - if (!regloc) {
> > - dev_err(dev, "register location dvsec not found\n");
> > + if (!regloc)
> > return -ENXIO;
> > - }
> > -
> > - if (pci_request_mem_regions(pdev, pci_name(pdev)))
> > - return -ENODEV;
> >
> > - /* Get the size of the Register Locator DVSEC */
> > pci_read_config_dword(pdev, regloc + PCI_DVSEC_HEADER1, ®loc_size);
> > regloc_size = FIELD_GET(PCI_DVSEC_HEADER1_LENGTH_MASK, regloc_size);
> >
> > regloc += PCI_DVSEC_ID_CXL_REGLOC_BLOCK1_OFFSET;
> > regblocks = (regloc_size - PCI_DVSEC_ID_CXL_REGLOC_BLOCK1_OFFSET) / 8;
> >
> > - for (i = 0, n_maps = 0; i < regblocks; i++, regloc += 8) {
> > + for (i = 0; i < regblocks; i++, regloc += 8) {
> > u32 reg_lo, reg_hi;
> > - u8 reg_type;
> > + u8 reg_type, bar;
> > u64 offset;
> > - u8 bar;
> >
> > pci_read_config_dword(pdev, regloc, ®_lo);
> > pci_read_config_dword(pdev, regloc + 4, ®_hi);
> > @@ -475,39 +457,74 @@ static int cxl_pci_setup_regs(struct cxl_mem *cxlm)
> > cxl_decode_register_block(reg_lo, reg_hi, &bar, &offset,
> > ®_type);
> >
> > - /* Ignore unknown register block types */
> > - if (reg_type > CXL_REGLOC_RBI_MEMDEV)
> > - continue;
> > + if (reg_type == type) {
> > + map->barno = bar;
> > + map->block_offset = offset;
> > + map->reg_type = reg_type;
> > + rc = 0;
> > + break;
>
> As this patch is already adding helpers, perhaps rather than a loop
> break, make the loop a helper so it can just "return 0;" directly:
>
> Something like:
>
> pci_request_mem_regions(...);
> rc = __find_register_block(...);
> pci_release_mem_regions(...);
>
> ...although, now that I see it written that way I think the request +
> release regions should probably just be dropped. It's not like any of
> the register enumeration would collide with someone else who already
> has the registers mapped. The collision only comes when the registers
> are mapped for their final usage, and that will have more precision in
> the request.
>
I think the origin of this was copy-pasta on my part. However, what's the
idiomatic way to do this? I guess I don't fully understand what kind of
collisions exist today and how to prevent them. Removing this entirely would
make easier code and if you assert we can do that, I'm all for it.
> > + }
> > + }
> >
> > - base = cxl_pci_map_regblock(cxlm, bar, offset);
> > - if (!base)
> > - return -ENOMEM;
> > + pci_release_mem_regions(pdev);
> >
> > - map = &maps[n_maps];
> > - map->barno = bar;
> > - map->block_offset = offset;
> > - map->reg_type = reg_type;
> > + return rc;
> > +}
> >
> > - ret = cxl_probe_regs(cxlm, base + offset, map);
> > +/**
> > + * cxl_pci_setup_regs() - Setup necessary MMIO.
> > + * @cxlm: The CXL memory device to communicate with.
> > + *
> > + * Return: 0 if all necessary registers mapped.
> > + *
> > + * A memory device is required by spec to implement a certain set of MMIO
> > + * regions. The purpose of this function is to enumerate and map those
> > + * registers.
> > + */
> > +static int cxl_pci_setup_regs(struct cxl_mem *cxlm)
> > +{
> > + int rc, i;
> > + struct device *dev = cxlm->dev;
> > + struct pci_dev *pdev = to_pci_dev(dev);
> > + const enum cxl_regloc_type types[] = { CXL_REGLOC_RBI_MEMDEV,
> > + CXL_REGLOC_RBI_COMPONENT };
> >
> > - /* Always unmap the regblock regardless of probe success */
> > - cxl_pci_unmap_regblock(cxlm, base);
> > + if (pci_request_mem_regions(pdev, pci_name(pdev)))
> > + return -ENODEV;
> >
> > - if (ret)
> > - return ret;
> > + for (i = 0; i < ARRAY_SIZE(types); i++) {
> > + struct cxl_register_map map;
> > + void __iomem *base;
> >
> > - n_maps++;
> > - }
> > + rc = find_register_block(pdev, types[i], &map);
> > + if (rc) {
> > + dev_err(dev, "Couldn't find %s register block\n",
> > + types[i] == CXL_REGLOC_RBI_MEMDEV ?
> > + "device" :
> > + "component");
> > + break;
> > + }
> >
> > - pci_release_mem_regions(pdev);
> > + base = cxl_pci_map_regblock(cxlm, map.barno, map.block_offset);
> > + if (!base) {
> > + rc = -ENOMEM;
> > + break;
> > + }
> >
> > - for (i = 0; i < n_maps; i++) {
> > - ret = cxl_map_regs(cxlm, &maps[i]);
> > - if (ret)
> > + rc = cxl_probe_regs(cxlm, base + map.block_offset, &map);
>
> It strikes me as odd @map has everything except a copy of @base. I
> wonder if this patch becomes easier to read if patch4 comes before
> this one and all the map_offset usage is dropped because @map can
> carry the required information directly. I'm not sure this suggestion
> is a win. I'm struggling to make sense of diff in isolation so will
> need to circle back when I can apply this and look at the result, for
> now it's just these edge comments.
It's a good point that I've also thought about. I think Ira's original goal was
to try to keep the iomem base in the API, but if we consider this series a win,
I think everything might benefit moving into @map. I'm certainly willing to try
it.
