[PATCH 11/14] fsi: Improve master indexing

Eddie James eajames at linux.ibm.com
Thu Aug 10 02:21:05 AEST 2023


On 8/9/23 02:08, Joel Stanley wrote:
> On Mon, 12 Jun 2023 at 19:57, Eddie James <eajames at linux.ibm.com> wrote:
>> Master indexing is problematic if a hub is rescanned while the
>> root master is being rescanned. Move the IDA free below the device
>> unregistration, lock the scan mutex in the probe function, and
>> request a specific idx in the hub driver.
> I've applied this series, but taking a closer look at this patch I
> think it can be improved. If you resend, just send this patch.
>
>> Signed-off-by: Eddie James <eajames at linux.ibm.com>
>> ---
>>   drivers/fsi/fsi-core.c       | 41 ++++++++++++++++++++++--------------
>>   drivers/fsi/fsi-master-hub.c |  2 ++
>>   2 files changed, 27 insertions(+), 16 deletions(-)
>>
>> diff --git a/drivers/fsi/fsi-core.c b/drivers/fsi/fsi-core.c
>> index ec4d02264391..503061a6740b 100644
>> --- a/drivers/fsi/fsi-core.c
>> +++ b/drivers/fsi/fsi-core.c
>> @@ -1327,46 +1327,55 @@ static struct class fsi_master_class = {
>>   int fsi_master_register(struct fsi_master *master)
>>   {
>>          int rc;
>> -       struct device_node *np;
>>
>>          mutex_init(&master->scan_lock);
>> -       master->idx = ida_alloc(&master_ida, GFP_KERNEL);
>> +
>> +       if (master->idx) {
> Why do we allocate a new idx if there's already one?


At this point, the master driver (aspeed, hub, i2cr, whatever) might 
just be requesting a certain index. It's not allocated yet, so it needs 
to be allocated so that we don't get overlap.


>
>> +               master->idx = ida_alloc_range(&master_ida, master->idx,
>> +                                             master->idx, GFP_KERNEL);
> If we can't get one in the range we want, we ask for any? Should this
> print a warning?


Perhaps, we could also return error here if we decide that the requested 
index is needed and not just wanted.


>
>> +               if (master->idx < 0)
>> +                       master->idx = ida_alloc(&master_ida, GFP_KERNEL);
>> +       } else {
> If ixd was zero, we create one. This is the "normal" case?


Yes, the assumption is: zero is the default due to zero-alloc'd master 
structures.


>
>> +               master->idx = ida_alloc(&master_ida, GFP_KERNEL);
>> +       }
>> +
> We check the same error condition again.


Yes I might be able to make this cleaner...


>
>>          if (master->idx < 0)
>>                  return master->idx;
>> -       dev_set_name(&master->dev, "fsi%d", master->idx);
>> +       if (!dev_name(&master->dev))
>> +               dev_set_name(&master->dev, "fsi%d", master->idx);
>> +
>>          master->dev.class = &fsi_master_class;
>>
>> +       mutex_lock(&master->scan_lock);
>>          rc = device_register(&master->dev);
>>          if (rc) {
>>                  ida_free(&master_ida, master->idx);
>> -               return rc;
>> -       }
>> +       } else {
>> +               struct device_node *np = dev_of_node(&master->dev);
> This change looks a bit different to the idx changes. What's happening here?


This is just restructuring to get the lock before the scan. It could be 
a separate commit, yea.

Thanks for the review!

Eddie


>> -       np = dev_of_node(&master->dev);
>> -       if (!of_property_read_bool(np, "no-scan-on-init")) {
>> -               mutex_lock(&master->scan_lock);
>> -               fsi_master_scan(master);
>> -               mutex_unlock(&master->scan_lock);
>> +               if (!of_property_read_bool(np, "no-scan-on-init"))
>> +                       fsi_master_scan(master);
>>          }
>>
>> -       return 0;
>> +       mutex_unlock(&master->scan_lock);
>> +       return rc;
>>   }
>>   EXPORT_SYMBOL_GPL(fsi_master_register);
>>
>>   void fsi_master_unregister(struct fsi_master *master)
>>   {
>> -       trace_fsi_master_unregister(master);
>> +       int idx = master->idx;
>>
>> -       if (master->idx >= 0) {
>> -               ida_free(&master_ida, master->idx);
>> -               master->idx = -1;
>> -       }
>> +       trace_fsi_master_unregister(master);
>>
>>          mutex_lock(&master->scan_lock);
>>          fsi_master_unscan(master);
>> +       master->n_links = 0;
>>          mutex_unlock(&master->scan_lock);
>> +
>>          device_unregister(&master->dev);
>> +       ida_free(&master_ida, idx);
>>   }
>>   EXPORT_SYMBOL_GPL(fsi_master_unregister);
>>
>> diff --git a/drivers/fsi/fsi-master-hub.c b/drivers/fsi/fsi-master-hub.c
>> index 6d8b6e8854e5..36da643b3201 100644
>> --- a/drivers/fsi/fsi-master-hub.c
>> +++ b/drivers/fsi/fsi-master-hub.c
>> @@ -12,6 +12,7 @@
>>   #include <linux/slab.h>
>>
>>   #include "fsi-master.h"
>> +#include "fsi-slave.h"
>>
>>   #define FSI_ENGID_HUB_MASTER           0x1c
>>
>> @@ -229,6 +230,7 @@ static int hub_master_probe(struct device *dev)
>>          hub->master.dev.release = hub_master_release;
>>          hub->master.dev.of_node = of_node_get(dev_of_node(dev));
>>
>> +       hub->master.idx = fsi_dev->slave->link + 1;
>>          hub->master.n_links = links;
>>          hub->master.read = hub_master_read;
>>          hub->master.write = hub_master_write;
>> --
>> 2.31.1
>>


More information about the linux-fsi mailing list