Re: [RFC PATCH 5/5] cxl/region: Register device poison before exposing RAM

[Shaikh Kamaluddin <shaikhkamal2012@gmail.com>]

On Thu, Oct 15, 2026 at 03:59:13PM +0800, Richard Cheng wrote:
> On Sat, Oct 10, 2026 at 09:50:13PM +0800, Shaikh Kamaluddin wrote:
> > The trigger_poison_list interface may retrieve poison after CXL memory
> > is already online, which is too late to prevent known-bad frames from
> > entering the page allocator.
> > 
> > During RAM-region probe, retrieve a poison-list snapshot for each
> > participating endpoint decoder. Validate each record against the
> > decoder's assigned DPA range, translate its poisoned DPA units to HPA
> > and PFN, and collect the affected PFNs in an xarray. Coalesce consecutive
> > PFNs and register the resulting ranges with the pre-online hwpoison
> > registry before creating the DAX region.
> > 
> > When Get Poison List is supported, abort region probe if the list is
> > incomplete, malformed, or cannot be translated. If volatile poison-list
> > retrieval is unsupported, preserve the existing behavior and warn that
> > poison cannot be withheld before memory onlining.
> > 
> > Signed-off-by: Shaikh Kamaluddin <shaikhkamal2012@gmail.com>
> > ---
> >  drivers/cxl/core/region.c | 322 ++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 322 insertions(+)
> > 
> > diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c
> > index c72524923b38..e3ffa77927df 100644
> > --- a/drivers/cxl/core/region.c
> > +++ b/drivers/cxl/core/region.c
> > @@ -14,6 +14,7 @@
> >  #include <linux/string_choices.h>
> >  #include <cxlmem.h>
> >  #include <cxl.h>
> > +#include <linux/memory-failure.h>
> >  #include "core.h"
> >  #include "mce.h"
> >  
> > @@ -3038,6 +3039,323 @@ int cxl_get_poison_by_endpoint(struct cxl_port *port)
> >  	return rc;
> >  }
> >  
> > +/*
> > + * Device-reported poison is gathered at RAM-region probe, before the range
> > + * reaches dax/kmem, and registered with the pre-online registry so the
> > + * frames it covers are withheld when the memory is onlined.
> > + */
> > +static int cxl_poison_store_pfn(struct xarray *pfns, u64 hpa)
> > +{
> > +	return xa_err(xa_store(pfns, PHYS_PFN(hpa), xa_mk_value(1),
> > +			GFP_KERNEL));
> > +}
> > +
> > +static int cxl_poison_store_hpa(struct cxl_region *cxlr, struct xarray *pfns,
> > +				u64 hpa)
> > +{
> > +	struct cxl_region_params *p = &cxlr->params;
> > +	int rc;
> > +
> > +	/* Outside this region: not ours to withhold. */
> > +	if (!cxl_resource_contains_addr(p->res, hpa))
> > +		return -ERANGE;
> > +
> > +	rc = cxl_poison_store_pfn(pfns, hpa);
> > +	if (rc || !p->cache_size)
> > +		return rc;
> > +
> > +	/*
> > +	 * An extended linear cache makes the same DPA visible a second time
> > +	 * at hpa - cache_size. Match the translation used by the
> > +	 * cxl_general_media and cxl_dram trace events.
> > +	 */
> > +	if (WARN_ON_ONCE(hpa < p->res->start + p->cache_size))
> > +		return -ERANGE;
> > +
> > +	return cxl_poison_store_pfn(pfns, hpa - p->cache_size);
> > +}
> > +
> > +static int cxl_poison_record_to_pfns(struct cxl_region *cxlr,
> > +				     struct cxl_memdev *cxlmd,
> > +				     const struct cxl_poison_record *record,
> > +				     struct xarray *pfns)
> > +{
> > +	struct cxl_region_params *p = &cxlr->params;
> > +	u64 raw = le64_to_cpu(record->address);
> > +	u64 units = le32_to_cpu(record->length);
> > +	u64 dpa = raw & CXL_POISON_START_MASK;
> > +	u8 source = raw & CXL_POISON_SOURCE_MASK;
> > +	u64 length, end, step, cur;
> > +
> > +	if (!units ||
> > +	    check_mul_overflow(units, (u64)CXL_POISON_LEN_MULT, &length) ||
> > +	    check_add_overflow(dpa, length, &end))
> > +		return -EPROTO;
> > +
> > +	/*
> > +	 * A corrected error never reaches the poison list, so every record
> > +	 * is uncorrectable. Only a failure of the device's own media is
> > +	 * withheld: External poison was written by the host into healthy
> > +	 * media and a rewrite or Clear Poison recovers it, and Injected
> > +	 * poison is a test artifact. Withholding either would make a
> > +	 * recoverable state permanent.
> > +	 */
> > +	if (source != CXL_POISON_SOURCE_INTERNAL) {
> > +		dev_dbg(&cxlr->dev, "%s: dpa %#llx source %u not withheld\n",
> > +			dev_name(&cxlmd->dev), dpa, source);
> > +		return 0;
> > +	}
> > +
> > +	/*
> > +	 * Translate once per unit of HPA contiguity rather than per 64
> > +	 * bytes. Consecutive DPAs within an interleave granule map to
> > +	 * consecutive HPAs, and a granule no larger than a page lies within
> > +	 * one page, so stepping by min(granule, PAGE_SIZE) from an aligned
> > +	 * start reaches every page the record covers.
> > +	 */
> > +	step = clamp_t(u64, p->interleave_granularity, CXL_POISON_LEN_MULT,
> > +		       PAGE_SIZE);
> > +
> > +	for (cur = ALIGN_DOWN(dpa, step); cur < end; cur += step) {
> > +		u64 hpa = cxl_dpa_to_hpa(cxlr, cxlmd, max(cur, dpa));
> > +		int rc;
> > +
> > +		/* Not mapped by this region; nothing here to withhold. */
> > +		if (hpa == ULLONG_MAX)
> > +			continue;
> > +
> > +		rc = cxl_poison_store_hpa(cxlr, pfns, hpa);
> > +		if (rc)
> > +			return rc;
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +static int cxl_region_collect_decoder_poison(struct cxl_region *cxlr,
> > +					     struct cxl_endpoint_decoder *cxled,
> > +					     struct xarray *pfns)
> > +{
> > +	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
> > +	struct cxl_poison_snapshot snapshot;
> > +	int rc;
> > +
> > +	rc = cxl_poison_snapshot_init(cxlmd, &snapshot);
> > +	if (rc == -EOPNOTSUPP) {
> > +		dev_info(&cxlr->dev,
> > +			 "%s: no Get Poison List support, nothing withheld\n",
> > +			 dev_name(&cxlmd->dev));
> > +		return 0;
> > +	}
> > +	if (rc)
> > +		return rc;
> > +
> > +	rc = cxl_get_poison_by_decoder(cxled, &snapshot,
> > +				       CXL_POISON_QUERY_PREONLINE);
> > +	if (rc == -EFAULT) {
> > +		/*
> > +		 * The device keeps no poison list for volatile ranges, which
> > +		 * CXL r3.2 8.2.9.8.4.1 permits. Online the region as today.
> > +		 */
> > +		dev_info(&cxlr->dev,
> > +			 "%s: no volatile poison list, nothing withheld\n",
> > +			 dev_name(&cxlmd->dev));
> > +		rc = 0;
> > +		goto out;
> > +	}
> > +	if (rc)
> > +		goto out;
> > +
> > +	/*
> > +	 * Withholding from a partial list would online the region on the
> > +	 * belief that every bad frame had been named.
> > +	 */
> > +	if (snapshot.truncated ||
> > +	    (snapshot.device_flags & CXL_POISON_FLAG_OVERFLOW)) {
> > +		dev_err(&cxlr->dev, "%s: poison list incomplete\n",
> > +			dev_name(&cxlmd->dev));
> > +		rc = -EOVERFLOW;
> > +		goto out;
> > +	}
> > +	if (snapshot.device_flags & CXL_POISON_FLAG_SCANNING) {
> > +		dev_err(&cxlr->dev, "%s: media scan in progress\n",
> > +			dev_name(&cxlmd->dev));
> > +		rc = -EBUSY;
> > +		goto out;
> > +	}
> > +
> > +	for (u32 i = 0; i < snapshot.nr_records; i++) {
> > +		rc = cxl_poison_record_to_pfns(cxlr, cxlmd,
> > +					       &snapshot.records[i], pfns);
> > +		if (rc)
> > +			break;
> > +	}
> > +out:
> > +	cxl_poison_snapshot_destroy(&snapshot);
> > +	return rc;
> > +}
> > +
> > +static unsigned int cxl_count_poison_ranges(struct xarray *pfns)
> > +{
> > +	unsigned long index, previous = 0;
> > +	unsigned int nr = 0;
> > +	void *entry;
> > +	bool first = true;
> > +
> > +	xa_for_each(pfns, index, entry) {
> > +		if (first || index != previous + 1)
> > +			nr++;
> > +		first = false;
> > +		previous = index;
> > +	}
> > +	return nr;
> > +}
> > +
> > +/*
> > + * The xarray holds each PFN once and xa_for_each() walks them in ascending
> > + * order, so runs of consecutive indices are exactly the sorted,
> > + * nonoverlapping ranges preonline_hwpoison_register() requires.
> > + */
> > +static int cxl_pfns_to_ranges(struct xarray *pfns,
> > +			      struct preonline_hwpoison_range **ranges_out,
> > +			      unsigned int *nr_ranges_out)
> > +{
> > +	struct preonline_hwpoison_range *ranges;
> > +	unsigned long index, previous = 0, run_start = 0;
> > +	unsigned int nr_ranges, nr = 0;
> > +	void *entry;
> > +	bool first = true;
> > +
> > +	nr_ranges = cxl_count_poison_ranges(pfns);
> > +	if (!nr_ranges) {
> > +		*ranges_out = NULL;
> > +		*nr_ranges_out = 0;
> > +		return 0;
> > +	}
> > +
> > +	ranges = kvmalloc_array(nr_ranges, sizeof(*ranges), GFP_KERNEL);
> > +	if (!ranges)
> > +		return -ENOMEM;
> > +
> > +	xa_for_each(pfns, index, entry) {
> > +		if (first) {
> > +			run_start = index;
> > +			previous = index;
> > +			first = false;
> > +			continue;
> > +		}
> > +		if (index == previous + 1) {
> > +			previous = index;
> > +			continue;
> > +		}
> > +		ranges[nr++] = (struct preonline_hwpoison_range) {
> > +			.start_pfn = run_start,
> > +			.nr_pages = previous - run_start + 1,
> > +		};
> > +		run_start = index;
> > +		previous = index;
> > +	}
> > +	ranges[nr++] = (struct preonline_hwpoison_range){
> > +		.start_pfn = run_start,
> > +		.nr_pages = previous - run_start + 1,
> > +	};
> > +
> > +	WARN_ON_ONCE(nr != nr_ranges);
> > +	*ranges_out = ranges;
> > +	*nr_ranges_out = nr;
> > +	return 0;
> > +}
> > +
> > +static void cxl_unregister_device_poison(void *data)
> > +{
> > +	preonline_hwpoison_unregister(data);
> > +}
> > +
> > +static int cxl_region_register_device_poison(struct cxl_region *cxlr)
> > +{
> > +	struct cxl_region_params *p = &cxlr->params;
> > +	struct preonline_hwpoison_range *ranges = NULL;
> > +	struct preonline_hwpoison *handle;
> > +	unsigned int nr_ranges = 0;
> > +	unsigned long nr_frames = 0;
> > +	struct xarray pfns;
> > +	int rc = 0;
> > +
> > +	lockdep_assert_held(&cxl_rwsem.region);
> > +	lockdep_assert_held(&cxl_rwsem.dpa);
> > +
> > +	/* The region may have been decommitted while the lock was dropped. */
> > +	if (p->state < CXL_CONFIG_COMMIT)
> > +		return -ENXIO;
> > +
> > +	if (test_bit(CXL_REGION_F_NORMALIZED_ADDRESSING, &cxlr->flags)) {
> > +		dev_info(&cxlr->dev,
> > +			 "normalized addressing, poison not translated\n");
> > +		return 0;
> > +	}
> > +
> > +	xa_init(&pfns);
> > +
> > +	for (int i = 0; i < p->nr_targets; i++) {
> > +		struct cxl_endpoint_decoder *cxled = p->targets[i];
> > +
> > +		if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
> > +			continue;
> > +		rc = cxl_region_collect_decoder_poison(cxlr, cxled, &pfns);
> > +		if (rc)
> > +			goto out;
> > +	}
> > +
> > +	if (xa_empty(&pfns))
> > +		goto out;
> > +
> > +	rc = cxl_pfns_to_ranges(&pfns, &ranges, &nr_ranges);
> > +	if (rc)
> > +		goto out;
> > +
> > +	rc = preonline_hwpoison_register(ranges, nr_ranges, &handle);
> > +	if (rc == -EOPNOTSUPP) {
> > +		/* No poison tracking in this kernel; nothing to withhold. */
> > +		rc = 0;
> > +		goto out;
> > +	}
> > +	if (rc)
> > +		goto out;
> > +
> > +	for (unsigned int i = 0; i < nr_ranges; i++)
> > +		nr_frames += ranges[i].nr_pages;
> > +	dev_warn(&cxlr->dev,
> > +		 "withholding %lu frame(s) reported poisoned by the device\n",
> > +		 nr_frames);
> > +
> > +	rc = devm_add_action_or_reset(&cxlr->dev, cxl_unregister_device_poison,
> > +				      handle);
> 
> Hi Shaikh,
> 
> I think you are trying to propose the architecture as the following ?
> - CXL region driver collects the device's poison info and owns the registration
> - MM use the registration to keep poisoned pages out of the page allocator
>   when memory is onlined
> - DAX/kmem adds and removes the corresponding system RAM
> 

Yes, that's how the RFC works today. If we go with Dave's suggestion below,
only the first point changes: CXL only provides the entries, and MM owns
them.

> The registration belongs to the driver, but the memory that needs it can
> outlive the driver binding.
> 
> Unbinding the driver doesn't necessarily remove the memory. If some mem blocks
> are still online, removal of the range can fail.
> When driver cleanup removes the registration, MM loses poison info of those
> memory blocks as well.
>

You're right. If any block is still online, dev_dax_kmem_remove() can't
remove the range and leaves all its blocks in place, but the devm action
still drops the entries. Blocks that were online are fine, their bad
frames are already marked. The problem is a block that was offline: it
can be onlined later, and its bad frame then goes to the buddy.

> I don't have a good idea how we should re-design the ownership here, maybe
> we can discuss about it.
>

Hand the poison info to MM when the memory is added. The driver's
registration only has to live until add_memory_driver_managed().
At that point MM copies the registered ranges inside the added
range and keeps them with the memory blocks (or the memory group
dax/kmem already registers), and frees them when the memory is
removed. After that the driver binding doesn't matter. Hook A
already makes its decision at add time, so both hooks would read
MM's copy.

Does that sound right? If so I'll do it this way in v2.

> Best regards,
> Richard Cheng.
> 
> 
> 
> 
> > +out:
> > +	kvfree(ranges);
> > +	xa_destroy(&pfns);
> > +	return rc;
> > +}
> > +
> > +static int cxl_region_scan_device_poison(struct cxl_region *cxlr)
> > +{
> > +	int rc;
> > +
> > +	/* Nothing would consume the result. */
> > +	if (!IS_ENABLED(CONFIG_MEMORY_FAILURE))
> > +		return 0;
> > +
> > +	ACQUIRE(rwsem_read_intr, region_rwsem)(&cxl_rwsem.region);
> > +	rc = ACQUIRE_ERR(rwsem_read_intr, &region_rwsem);
> > +	if (rc)
> > +		return rc;
> > +	ACQUIRE(rwsem_read_intr, dpa_rwsem)(&cxl_rwsem.dpa);
> > +	rc = ACQUIRE_ERR(rwsem_read_intr, &dpa_rwsem);
> > +	if (rc)
> > +		return rc;
> > +
> > +	return cxl_region_register_device_poison(cxlr);
> > +}
> > +
> >  struct cxl_dpa_to_region_context {
> >  	struct cxl_region *cxlr;
> >  	u64 dpa;
> > @@ -4262,6 +4580,10 @@ static int cxl_region_probe(struct device *dev)
> >  					p->res->start, p->res->end, cxlr,
> >  					is_system_ram) > 0)
> >  			return 0;
> > +		rc = cxl_region_scan_device_poison(cxlr);
> > +		if (rc)
> > +			return rc;
> > +
> >  		return devm_cxl_add_dax_region(cxlr);
> >  	default:
> >  		dev_dbg(&cxlr->dev, "unsupported region mode: %d\n",
> > -- 
> > 2.43.0
> >