Re: [PATCH v3 2/6] cxl/mbox: Add GET_POISON_LIST mailbox command

[Alison Schofield]

On Tue, Dec 06, 2022 at 06:41:34PM -0800, Dan Williams wrote:
> alison.schofield@ wrote:
> > From: Alison Schofield <alison.schofield@xxxxxxxxx>
> > 
> > CXL devices maintain a list of locations that are poisoned or result
> > in poison if the addresses are accessed by the host.
> > 
> > Per the spec (CXL 3.0 8.2.9.8.4.1), the device returns this Poison
> > list as a set of  Media Error Records that include the source of the
> > error, the starting device physical address and length. The length is
> > the number of adjacent DPAs in the record and is in units of 64 bytes.
> > 
> > Retrieve the list and log each Media Error Record as a trace event of
> > type 'cxl_poison'.
> > 
> > When the poison list is requested by region, include the region name
> > and uuid in the trace event.
> > 
> > Signed-off-by: Alison Schofield <alison.schofield@xxxxxxxxx>
> > ---
> >  drivers/cxl/core/mbox.c | 81 +++++++++++++++++++++++++++++++++++++++++
> >  drivers/cxl/cxlmem.h    | 37 +++++++++++++++++++
> >  2 files changed, 118 insertions(+)
> > 
> > diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> > index 0c90f13870a4..88f034e97812 100644
> > --- a/drivers/cxl/core/mbox.c
> > +++ b/drivers/cxl/core/mbox.c
> > @@ -9,6 +9,9 @@
> >  
> >  #include "core.h"
> >  
> > +#define CREATE_TRACE_POINTS
> > +#include <trace/events/cxl.h>
> > +
> >  static bool cxl_raw_allow_all;
> >  
> >  /**
> > @@ -752,6 +755,7 @@ int cxl_dev_state_identify(struct cxl_dev_state *cxlds)
> >  {
> >  	/* See CXL 2.0 Table 175 Identify Memory Device Output Payload */
> >  	struct cxl_mbox_identify id;
> > +	__le32 val = 0;
> >  	int rc;
> >  
> >  	rc = cxl_mbox_send_cmd(cxlds, CXL_MBOX_OP_IDENTIFY, NULL, 0, &id,
> > @@ -771,6 +775,9 @@ int cxl_dev_state_identify(struct cxl_dev_state *cxlds)
> >  	cxlds->lsa_size = le32_to_cpu(id.lsa_size);
> >  	memcpy(cxlds->firmware_version, id.fw_revision, sizeof(id.fw_revision));
> >  
> > +	memcpy(&val, id.poison_list_max_mer, 3);
> 
> I see Jonathan already commented on the need for get_unaligned_le24()
> here, seconded.
> 
Got it.

> 
> > +	cxlds->poison_max = min_t(u32, le32_to_cpu(val), CXL_POISON_LIST_MAX);
> > +
> >  	return 0;
> >  }
> >  EXPORT_SYMBOL_NS_GPL(cxl_dev_state_identify, CXL);
> > @@ -835,6 +842,79 @@ int cxl_mem_create_range_info(struct cxl_dev_state *cxlds)
> >  }
> >  EXPORT_SYMBOL_NS_GPL(cxl_mem_create_range_info, CXL);
> >  
> > +static void cxl_trace_poison(struct cxl_mbox_poison_payload_out *po,
> > +			     struct cxl_region *cxlr,
> > +			     const char *memdev_name,
> > +			     const char *pcidev_name)
> 
> Type-safety please. Pass a 'struct cxl_memdev *' and 'struct pci_dev *'.
> Might need to be 'struct device *' instead of 'struct pci_dev *'
> depending on if this needs to be called from cxl_test, but anything is
> better than a non-specific string.
> 
Got it.

> > +{
> > +	const char *region_name = cxlr ? dev_name(&cxlr->dev) : NULL;
> 
> ...and push this conversion into the trace point.
> 
Got that, and also pushed the decode of struct cxl_poison_record to
the trace point like this:

TP_PROTO(const struct device *memdev, const struct pci_dev *pcidev,
                     const struct cxl_region *region,
                     const struct cxl_poison_record *record,
                     u8 flags, __le64 overflow_t),


> > +	struct cxl_region_params *p = cxlr ? &cxlr->params : NULL;
> > +	uuid_t *uuid = p ? &p->uuid : NULL;
> > +	u64 addr, dpa, overflow_t = 0;
> > +	u8 source;
> > +	u32 len;
> > +
> > +	if (po->flags & CXL_POISON_FLAG_OVERFLOW)
> > +		overflow_t = le64_to_cpu(po->overflow_timestamp);
> > +
> > +	for (int i = 0; i < le16_to_cpu(po->count); i++) {
> > +		len = le32_to_cpu(po->record[i].length) * CXL_POISON_LEN_MULT;
> > +		addr = le64_to_cpu(po->record[i].address);
> > +		source = addr & CXL_POISON_SOURCE_MASK;
> > +		dpa = addr & CXL_POISON_START_MASK;
> > +
> > +		trace_cxl_poison(memdev_name, pcidev_name, region_name, uuid,
> > +				 dpa, len, source, po->flags, overflow_t);
> > +	}
> > +}
> > +
> > +int cxl_mem_get_poison(struct cxl_memdev *cxlmd, u64 offset, u64 len,
> > +		       struct cxl_region *cxlr)
> > +{
> > +	struct cxl_dev_state *cxlds = cxlmd->cxlds;
> > +	const char *memdev_name = dev_name(&cxlmd->dev);
> > +	const char *pcidev_name = dev_name(cxlds->dev);
> > +	struct cxl_mbox_poison_payload_out *po;
> > +	struct cxl_mbox_poison_payload_in pi;
> > +	int nr_records = 0;
> > +	int rc;
> > +
> > +	po = kvmalloc(cxlds->payload_size, GFP_KERNEL);
> > +	if (!po)
> > +		return -ENOMEM;
> > +
> > +	pi.offset = cpu_to_le64(offset);
> > +	pi.length = cpu_to_le64(len);
> > +
> > +	rc = mutex_lock_interruptible(&cxlds->poison_list_mutex);
> 
> So I do not know what this mutex is protecting if there is an allocation
> per cxl_mem_get_poison() invocation. Although I suspect that's somewhat
> wasteful. Just allocate one buffer at the beginning of time and then use
> the lock to protect that buffer.

Intent was a single lock on the device to protect the state of the
poison list retrieval - do not allow > 1 reader. With > 1 reader
software may not know if it retrieved the complete list.

I'm not understanding about protecting a buffer, instead of protecting
the state. Here I try to protect the state.

> 
> Although, I wonder if this and Event handling should share locks and one
> preallocated buffer? Otherwise I do think it is important for Events and
> Poison handling to be able to make forward progress without needing to
> allocate up to a megabyte of memory at runtime. The other payload_size
> allocations are for one-off things that run at the beginning of time,
> but Poison and Events run repeatedly.
> 
> > 
snip