On Thu, 23 Jun 2022 19:46:44 -0700
Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> In preparation for provisioining CXL regions, add accounting for the DPA
> space consumed by existing regions / decoders. Recall, a CXL region is a
> memory range comrpised from one or more endpoint devices contributing a
> mapping of their DPA into HPA space through a decoder.
>
> Record the DPA ranges covered by committed decoders at initial probe of
> endpoint ports relative to a per-device resource tree of the DPA type
> (pmem or volaltile-ram).
>
> The cxl_dpa_rwsem semaphore is introduced to globally synchronize DPA
> state across all endpoints and their decoders at once. The vast majority
> of DPA operations are reads as region creation is expected to be as rare
> as disk partitioning and volume creation. The device_lock() for this
> synchronization is specifically avoided for concern of entangling with
> sysfs attribute removal.
>
> Co-developed-by: Ben Widawsky <bwidawsk@xxxxxxxxxx>
> Signed-off-by: Ben Widawsky <bwidawsk@xxxxxxxxxx>
> Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> ---
> drivers/cxl/core/hdm.c | 148 ++++++++++++++++++++++++++++++++++++++++++++----
> drivers/cxl/cxl.h | 2 +
> drivers/cxl/cxlmem.h | 13 ++++
> 3 files changed, 152 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/cxl/core/hdm.c b/drivers/cxl/core/hdm.c
> index c940a4911fee..daae6e533146 100644
> --- a/drivers/cxl/core/hdm.c
> +++ b/drivers/cxl/core/hdm.c
> @@ -7,6 +7,8 @@
> #include "cxlmem.h"
> #include "core.h"
>
> +static DECLARE_RWSEM(cxl_dpa_rwsem);
I've not checked many files, but pci.c has equivalent static defines after
the DOC: entry so for consistency move this below that?
> +
> /**
> * DOC: cxl core hdm
> *
> @@ -128,10 +130,108 @@ struct cxl_hdm *devm_cxl_setup_hdm(struct cxl_port *port)
> }
> EXPORT_SYMBOL_NS_GPL(devm_cxl_setup_hdm, CXL);
>
> +/*
> + * Must be called in a context that synchronizes against this decoder's
> + * port ->remove() callback (like an endpoint decoder sysfs attribute)
> + */
> +static void cxl_dpa_release(void *cxled);
> +static void __cxl_dpa_release(struct cxl_endpoint_decoder *cxled, bool remove_action)
> +{
> + struct cxl_port *port = cxled_to_port(cxled);
> + struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
> + struct cxl_dev_state *cxlds = cxlmd->cxlds;
> + struct resource *res = cxled->dpa_res;
> +
> + lockdep_assert_held_write(&cxl_dpa_rwsem);
> +
> + if (remove_action)
> + devm_remove_action(&port->dev, cxl_dpa_release, cxled);
This code organization is more surprising than I'd like. Why not move this to
a wrapper that is like devm_kfree() and similar which do the free now and
remove from the devm list?
static void __cxl_dpa_release(struct cxl_endpoint_decoder *cxled)
{
struct cxl_port *port = cxled_to_port(cxled);
struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
struct cxl_dev_state *cxlds = cxlmd->cxlds;
struct resource *res = cxled->dpa_res;
if (cxled->skip)
__release_region(&cxlds->dpa_res, res->start - cxled->skip,
cxled->skip);
cxled->skip = 0;
__release_region(&cxlds->dpa_res, res->start, resource_size(res));
cxled->dpa_res = NULL;
}
/* possibly add some underscores to this name to indicate it's special
in when you can safely call it */
static void devm_cxl_dpa_release(struct cxl_endpoint_decoder *cxled)
{
struct cxl_port *port = cxled_to_port(cxled);
lockdep_assert_held_write(&cxl_dpa_rwsem);
devm_remove_action(&port->dev, cxl_dpa_release, cxled);
__cxl_dpa_release(cxled);
}
static void cxl_dpa_release(void *cxled)
{
down_write(&cxl_dpa_rwsem);
__cxl_dpa_release(cxled, false);
up_write(&cxl_dpa_rwsem);
}
> +
> + if (cxled->skip)
> + __release_region(&cxlds->dpa_res, res->start - cxled->skip,
> + cxled->skip);
> + cxled->skip = 0;
> + __release_region(&cxlds->dpa_res, res->start, resource_size(res));
> + cxled->dpa_res = NULL;
> +}
> +
> +static void cxl_dpa_release(void *cxled)
> +{
> + down_write(&cxl_dpa_rwsem);
> + __cxl_dpa_release(cxled, false);
> + up_write(&cxl_dpa_rwsem);
> +}
> +
> +static int __cxl_dpa_reserve(struct cxl_endpoint_decoder *cxled,
> + resource_size_t base, resource_size_t len,
> + resource_size_t skip)
> +{
> + struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
> + struct cxl_port *port = cxled_to_port(cxled);
> + struct cxl_dev_state *cxlds = cxlmd->cxlds;
> + struct device *dev = &port->dev;
> + struct resource *res;
> +
> + lockdep_assert_held_write(&cxl_dpa_rwsem);
> +
> + if (!len)
> + return 0;
> +
> + if (cxled->dpa_res) {
> + dev_dbg(dev, "decoder%d.%d: existing allocation %pr assigned\n",
> + port->id, cxled->cxld.id, cxled->dpa_res);
> + return -EBUSY;
> + }
> +
> + if (skip) {
> + res = __request_region(&cxlds->dpa_res, base - skip, skip,
> + dev_name(dev), 0);
Interface that uses a backwards definition of skip as what to skip before
the base parameter is a little odd can we rename base parameter to something
like 'current_top' then have base = current_top + skip? current_top naming
not great though...
> + if (!res) {
> + dev_dbg(dev,
> + "decoder%d.%d: failed to reserve skip space\n",
> + port->id, cxled->cxld.id);
> + return -EBUSY;
> + }
> + }
> + res = __request_region(&cxlds->dpa_res, base, len, dev_name(dev), 0);
> + if (!res) {
> + dev_dbg(dev, "decoder%d.%d: failed to reserve allocation\n",
> + port->id, cxled->cxld.id);
> + if (skip)
> + __release_region(&cxlds->dpa_res, base - skip, skip);
> + return -EBUSY;
> + }
> + cxled->dpa_res = res;
> + cxled->skip = skip;
> +
> + return 0;
> +}
> +
...
