Message ID | 20241022074319.512127-1-avri.altman@wdc.com |
---|---|
Headers | show |
Series | Untie the host lock entanglement - part 1 | expand |
> On 10/22/24 12:43 AM, Avri Altman wrote: > > @@ -6877,13 +6874,13 @@ static irqreturn_t ufshcd_check_errors(struct > ufs_hba *hba, u32 intr_status) > > */ > > static irqreturn_t ufshcd_tmc_handler(struct ufs_hba *hba) > > { > > - unsigned long flags, pending, issued; > > + unsigned long flags; > > + unsigned long pending = ufshcd_readl(hba, > REG_UTP_TASK_REQ_DOOR_BELL); > > + unsigned long issued = hba->outstanding_tasks & ~pending; > > irqreturn_t ret = IRQ_NONE; > > int tag; > > > > spin_lock_irqsave(hba->host->host_lock, flags); > > - pending = ufshcd_readl(hba, REG_UTP_TASK_REQ_DOOR_BELL); > > - issued = hba->outstanding_tasks & ~pending; > > Please keep the 'pending' and 'issued' assignments in the function body. > Initializing variables in the declaration block is fine but adding code in the > declaration block that has side effects is a bit controversial. Done. > > > for_each_set_bit(tag, &issued, hba->nutmrs) { > > struct request *req = hba->tmf_rqs[tag]; > > struct completion *c = req->end_io_data; > > Would it be sufficient to hold the SCSI host lock around the > hba->outstanding_tasks read only? I don't think that the > for_each_set_bit() loop needs to be protected with the SCSI host lock. That may cause concurrent access to tmf_rqs? So better withdraw from changing ufshcd_tmc_handler() and just leave the whole function as it is? Thanks, Avri > > Otherwise this patch looks good to me. > > Thanks, > > Bart.
On 10/22/24 11:47 PM, Avri Altman wrote: >> On 10/22/24 12:43 AM, Avri Altman wrote: >>> for_each_set_bit(tag, &issued, hba->nutmrs) { >>> struct request *req = hba->tmf_rqs[tag]; >>> struct completion *c = req->end_io_data; >> >> Would it be sufficient to hold the SCSI host lock around the >> hba->outstanding_tasks read only? I don't think that the >> for_each_set_bit() loop needs to be protected with the SCSI host lock. > > That may cause concurrent access to tmf_rqs? Right, the host_lock serializes hba->tmf_rqs[] accesses. Without having analyzed whether or not removing locking from around the hba->tmf_rqs[] accesses, let's keep this locking. > So better withdraw from changing ufshcd_tmc_handler() and just leave > the whole function as it is? That sounds good to me. Thanks, Bart.