| Message ID | 172964783668.81806.14962699553881333486.stgit@dwillia2-xfh.jf.intel.com |
|---|---|
| State | Accepted |
| Commit | ced502c17c327ddd874cd2edb3eb4b1ff05dceed |
| Headers | show |
| Series | cxl: Initialization and shutdown fixes | expand |
On Tue, 22 Oct 2024 18:43:57 -0700 Dan Williams <dan.j.williams@intel.com> wrote: > With the recent change to allow out-of-order decoder de-commit it > highlights a need to strengthen the in-order decoder commit guarantees. > As it stands match_free_decoder() ensures that if 2 regions are racing > decoder allocations the one that wins the race will get the lower id > decoder, but that still leaves the race to *commit* the decoder. > > Rather than have this complicated case of "reserved in-order, but may > still commit out-of-order", just arrange for the reservation order to > match the commit-order. In other words, prevent subsequent allocations > until the last reservation is committed. > > This precludes overlapping region creation events and requires the > previous regionN to either move forward to the decoder commit stage or > drop its reservation before regionN+1 can move forward. That is, > provided that regionN and regionN+1 decode through the same switch port. > > As a side effect this allows match_free_decoder() to drop its dependency > on needing write access to the device_find_child() @data parameter [1]. > > Reported-by: Zijun Hu <quic_zijuhu@quicinc.com> > Closes: http://lore.kernel.org/20240905-const_dfc_prepare-v4-0-4180e1d5a244@quicinc.com > Cc: Davidlohr Bueso <dave@stgolabs.net> > Cc: Vishal Verma <vishal.l.verma@intel.com> > Cc: Alison Schofield <alison.schofield@intel.com> > Cc: Jonathan Cameron <jonathan.cameron@huawei.com> > Signed-off-by: Dan Williams <dan.j.williams@intel.com> LGTM Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Dan Williams wrote: > With the recent change to allow out-of-order decoder de-commit it > highlights a need to strengthen the in-order decoder commit guarantees. > As it stands match_free_decoder() ensures that if 2 regions are racing > decoder allocations the one that wins the race will get the lower id > decoder, but that still leaves the race to *commit* the decoder. > > Rather than have this complicated case of "reserved in-order, but may > still commit out-of-order", just arrange for the reservation order to > match the commit-order. In other words, prevent subsequent allocations > until the last reservation is committed. > > This precludes overlapping region creation events and requires the > previous regionN to either move forward to the decoder commit stage or > drop its reservation before regionN+1 can move forward. That is, > provided that regionN and regionN+1 decode through the same switch port. > > As a side effect this allows match_free_decoder() to drop its dependency > on needing write access to the device_find_child() @data parameter [1]. > > Reported-by: Zijun Hu <quic_zijuhu@quicinc.com> > Closes: http://lore.kernel.org/20240905-const_dfc_prepare-v4-0-4180e1d5a244@quicinc.com > Cc: Davidlohr Bueso <dave@stgolabs.net> > Cc: Vishal Verma <vishal.l.verma@intel.com> > Cc: Alison Schofield <alison.schofield@intel.com> > Cc: Jonathan Cameron <jonathan.cameron@huawei.com> > Signed-off-by: Dan Williams <dan.j.williams@intel.com> Reviewed-by: Ira Weiny <ira.weiny@intel.com> [snip]
diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c index 3478d2058303..dff618c708dc 100644 --- a/drivers/cxl/core/region.c +++ b/drivers/cxl/core/region.c @@ -778,26 +778,50 @@ static size_t show_targetN(struct cxl_region *cxlr, char *buf, int pos) return rc; } +static int check_commit_order(struct device *dev, const void *data) +{ + struct cxl_decoder *cxld = to_cxl_decoder(dev); + + /* + * if port->commit_end is not the only free decoder, then out of + * order shutdown has occurred, block further allocations until + * that is resolved + */ + if (((cxld->flags & CXL_DECODER_F_ENABLE) == 0)) + return -EBUSY; + return 0; +} + static int match_free_decoder(struct device *dev, void *data) { + struct cxl_port *port = to_cxl_port(dev->parent); struct cxl_decoder *cxld; - int *id = data; + int rc; if (!is_switch_decoder(dev)) return 0; cxld = to_cxl_decoder(dev); - /* enforce ordered allocation */ - if (cxld->id != *id) + if (cxld->id != port->commit_end + 1) return 0; - if (!cxld->region) - return 1; - - (*id)++; + if (cxld->region) { + dev_dbg(dev->parent, + "next decoder to commit (%s) is already reserved (%s)\n", + dev_name(dev), dev_name(&cxld->region->dev)); + return 0; + } - return 0; + rc = device_for_each_child_reverse_from(dev->parent, dev, NULL, + check_commit_order); + if (rc) { + dev_dbg(dev->parent, + "unable to allocate %s due to out of order shutdown\n", + dev_name(dev)); + return 0; + } + return 1; } static int match_auto_decoder(struct device *dev, void *data) @@ -824,7 +848,6 @@ cxl_region_find_decoder(struct cxl_port *port, struct cxl_region *cxlr) { struct device *dev; - int id = 0; if (port == cxled_to_port(cxled)) return &cxled->cxld; @@ -833,7 +856,7 @@ cxl_region_find_decoder(struct cxl_port *port, dev = device_find_child(&port->dev, &cxlr->params, match_auto_decoder); else - dev = device_find_child(&port->dev, &id, match_free_decoder); + dev = device_find_child(&port->dev, NULL, match_free_decoder); if (!dev) return NULL; /*
With the recent change to allow out-of-order decoder de-commit it highlights a need to strengthen the in-order decoder commit guarantees. As it stands match_free_decoder() ensures that if 2 regions are racing decoder allocations the one that wins the race will get the lower id decoder, but that still leaves the race to *commit* the decoder. Rather than have this complicated case of "reserved in-order, but may still commit out-of-order", just arrange for the reservation order to match the commit-order. In other words, prevent subsequent allocations until the last reservation is committed. This precludes overlapping region creation events and requires the previous regionN to either move forward to the decoder commit stage or drop its reservation before regionN+1 can move forward. That is, provided that regionN and regionN+1 decode through the same switch port. As a side effect this allows match_free_decoder() to drop its dependency on needing write access to the device_find_child() @data parameter [1]. Reported-by: Zijun Hu <quic_zijuhu@quicinc.com> Closes: http://lore.kernel.org/20240905-const_dfc_prepare-v4-0-4180e1d5a244@quicinc.com Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Vishal Verma <vishal.l.verma@intel.com> Cc: Alison Schofield <alison.schofield@intel.com> Cc: Jonathan Cameron <jonathan.cameron@huawei.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> --- drivers/cxl/core/region.c | 43 +++++++++++++++++++++++++++++++++---------- 1 file changed, 33 insertions(+), 10 deletions(-)