Hi Bart,
On 2020-08-06 02:11, Daejun Park wrote:
> > This is a patch for the HPB feature.
> > This patch adds HPB function calls to UFS core driver.
> >
> > The mininum size of the memory pool used in the HPB is implemented as a
> ^^^^^^^
> minimum?
I will fix it.
> > Kconfig parameter (SCSI_UFS_HPB_HOST_MEM), so that it can be configurable.
>
> > +config SCSI_UFS_HPB
> > + bool "Support UFS Host Performance Booster"
> > + depends on SCSI_UFSHCD
> > + help
> > + A UFS HPB Feature improves random read performance. It caches
> ^ ^^^^^^^
> The? feature?
I will fix it.
> > + L2P map of UFS to host DRAM. The driver uses HPB read command
> > + by piggybacking physical page number for bypassing FTL's L2P address
> > + translation.
>
> Please explain what L2P and FTL mean. Not everyone is familiar with SSD
> internals.
I added full name of the abbreviation.
L2P (logical to physical) map of UFS to host DRAM. The driver uses HPB read command
^^^^^^^^^^^^^^^^^^^
by piggybacking physical page number for bypassing FTL (flash translation layer)
^^^^^^^^^^^^^^^^^^^^^^^^
> > +config SCSI_UFS_HPB_HOST_MEM
> > + int "Host-side cached memory size (KB) for HPB support"
> > + default 32
> > + depends on SCSI_UFS_HPB
> > + help
> > + The mininum size of the memory pool used in the HPB module. It can
> > + be configurable by the user. If this value is larger than required
> > + memory size, kernel resizes cached memory size.
> ^^^^^^^ ^^^^^^^^^^^^^^^^^^
> reduces? cache size?
>
> Please make this a kernel module parameter instead of a compile-time constant.
OK, I will change it.
> > +#ifndef CONFIG_SCSI_UFS_HPB
> > +static void ufshpb_resume(struct ufs_hba *hba) {}
> > +static void ufshpb_suspend(struct ufs_hba *hba) {}
> > +static void ufshpb_reset(struct ufs_hba *hba) {}
> > +static void ufshpb_reset_host(struct ufs_hba *hba) {}
> > +static void ufshpb_rsp_upiu(struct ufs_hba *hba, struct ufshcd_lrb *lrbp) {}
> > +static void ufshpb_prep(struct ufs_hba *hba, struct ufshcd_lrb *lrbp) {}
> > +static void ufshpb_remove(struct ufs_hba *hba) {}
> > +static void ufshpb_scan_feature(struct ufs_hba *hba) {}
> > +#endif
>
> Please move these definitions into ufshpb.h since that is the
> recommended Linux kernel coding style.
OK, I will move them.
> > diff --git a/drivers/scsi/ufs/ufshcd.h b/drivers/scsi/ufs/ufshcd.h
> > index b2ef18f1b746..904c19796e09 100644
> > --- a/drivers/scsi/ufs/ufshcd.h
> > +++ b/drivers/scsi/ufs/ufshcd.h
> > @@ -47,6 +47,9 @@
> > #include "ufs.h"
> > #include "ufs_quirks.h"
> > #include "ufshci.h"
> > +#ifdef CONFIG_SCSI_UFS_HPB
> > +#include "ufshpb.h"
> > +#endif
>
> Please move #ifdef CONFIG_SCSI_UFS_HPB / #endif into ufshpb.h. From
> Documentation/process/4.Coding.rst: "As a general rule, #ifdef use
> should be confined to header files whenever possible."
OK, I will fix it.
> > +struct ufsf_feature_info {
> > + atomic_t slave_conf_cnt;
> > + wait_queue_head_t sdev_wait;
> > +};
>
> Please add a comment above this data structure that explains the role
> of this data structure and also what "ufsf" stands for.
"ufsf" is stands for ufs feature. I wiil add comments for the data structure.
> > +static int ufshpb_create_sysfs(struct ufs_hba *hba, struct ufshpb_lu *hpb);
>
> I don't think that this forward declaration is necessary so please leave it
> out.
OK, I will remove it.
> > +static inline int ufshpb_is_valid_srgn(struct ufshpb_region *rgn,
> > + struct ufshpb_subregion *srgn)
> > +{
> > + return rgn->rgn_state != HPB_RGN_INACTIVE &&
> > + srgn->srgn_state == HPB_SRGN_VALID;
> > +}
>
> Please do not declare functions inside .c files inline but instead let
> the compiler decide which functions to inline. Modern compilers are really
> good at this.
I didn't know about it. Thanks.
> > +static struct kobj_type ufshpb_ktype = {
> > + .sysfs_ops = &ufshpb_sysfs_ops,
> > + .release = NULL,
> > +};
>
> If the release method of a kobj_type is NULL that is a strong sign that
> there is something wrong ...
>
> > +static int ufshpb_create_sysfs(struct ufs_hba *hba, struct ufshpb_lu *hpb)
> > +{
> > + int ret;
> > +
> > + ufshpb_stat_init(hpb);
> > +
> > + kobject_init(&hpb->kobj, &ufshpb_ktype);
> > + mutex_init(&hpb->sysfs_lock);
> > +
> > + ret = kobject_add(&hpb->kobj, kobject_get(&hba->dev->kobj),
> > + "ufshpb_lu%d", hpb->lun);
> > +
> > + if (ret)
> > + return ret;
> > +
> > + ret = sysfs_create_group(&hpb->kobj, &ufshpb_sysfs_group);
> > +
> > + if (ret) {
> > + dev_err(hba->dev, "ufshpb_lu%d create file error\n", hpb->lun);
> > + return ret;
> > + }
> > +
> > + dev_info(hba->dev, "ufshpb_lu%d sysfs adds uevent", hpb->lun);
> > + kobject_uevent(&hpb->kobj, KOBJ_ADD);
> > +
> > + return 0;
> > +}
>
> Please attach these sysfs attributes to /sys/class/scsi_device/*/device
> instead of creating a new kobject. Consider using the following
> scsi_host_template member to define LUN sysfs attributes:
I am not rejecting your comment. But I added kobject for distinguishing
between other attributes and attributes related to HPB feature.
If you think it's pointless, I'll fix it.
> /*
> * Pointer to the SCSI device attribute groups for this host,
> * NULL terminated.
> */
> const struct attribute_group **sdev_groups;
>
> > +static void ufshpb_scan_hpb_lu(struct ufs_hba *hba,
> > + struct ufshpb_dev_info *hpb_dev_info,
> > + u8 *desc_buf)
> > +{
> > + struct scsi_device *sdev;
> > + struct ufshpb_lu *hpb;
> > + int find_hpb_lu = 0;
> > + int ret;
> > +
> > + shost_for_each_device(sdev, hba->host) {
> > + struct ufshpb_lu_info hpb_lu_info = { 0 };
> > + int lun = sdev->lun;
> > +
> > + if (lun >= hba->dev_info.max_lu_supported)
> > + continue;
> > +
> > + ret = ufshpb_get_lu_info(hba, lun, &hpb_lu_info, desc_buf);
> > + if (ret)
> > + continue;
> > +
> > + hpb = ufshpb_alloc_hpb_lu(hba, lun, hpb_dev_info,
> > + &hpb_lu_info);
> > + if (!hpb)
> > + continue;
> > +
> > + hpb->sdev_ufs_lu = sdev;
> > + sdev->hostdata = hpb;
> > +
> > + list_add_tail(&hpb->list_hpb_lu, &lh_hpb_lu);
> > + find_hpb_lu++;
> > + }
> > +
> > + if (!find_hpb_lu)
> > + return;
> > +
> > + ufshpb_check_hpb_reset_query(hba);
> > +
> > + list_for_each_entry(hpb, &lh_hpb_lu, list_hpb_lu) {
> > + dev_info(hba->dev, "set state to present\n");
> > + ufshpb_set_state(hpb, HPB_PRESENT);
> > + }
> > +}
>
> Please remove the loop from the above function, make this function accept a
> SCSI device pointer as argument and call this function from
> ufshcd_slave_configure() or ufshcd_hpb_configure().
I will move the loop to ufshcd_hpb_configure().
>
> > +static void ufshpb_init(void *data, async_cookie_t cookie)
> > +{
> > + struct ufsf_feature_info *ufsf = (struct ufsf_feature_info *)data;
> > + struct ufs_hba *hba;
> > + struct ufshpb_dev_info hpb_dev_info = { 0 };
> > + char *desc_buf;
> > + int ret;
> > +
> > + hba = container_of(ufsf, struct ufs_hba, ufsf);
> > +
> > + desc_buf = kzalloc(QUERY_DESC_MAX_SIZE, GFP_KERNEL);
> > + if (!desc_buf)
> > + goto release_desc_buf;
> > +
> > + ret = ufshpb_get_dev_info(hba, &hpb_dev_info, desc_buf);
> > + if (ret)
> > + goto release_desc_buf;
> > +
> > + /*
> > + * Because HPB driver uses scsi_device data structure,
> > + * we should wait at this point until finishing initialization of all
> > + * scsi devices. Even if timeout occurs, HPB driver will search
> > + * the scsi_device list on struct scsi_host (shost->__host list_head)
> > + * and can find out HPB logical units in all scsi_devices
> > + */
> > + wait_event_timeout(hba->ufsf.sdev_wait,
> > + (atomic_read(&hba->ufsf.slave_conf_cnt)
> > + == hpb_dev_info.num_lu),
> > + SDEV_WAIT_TIMEOUT);
> > +
> > + ufshpb_issue_hpb_reset_query(hba);
> > +
> > + dev_dbg(hba->dev, "ufshpb: slave count %d, lu count %d\n",
> > + atomic_read(&hba->ufsf.slave_conf_cnt), hpb_dev_info.num_lu);
> > +
> > + ufshpb_scan_hpb_lu(hba, &hpb_dev_info, desc_buf);
> > +
> > +release_desc_buf:
> > + kfree(desc_buf);
> > +}
>
> Since the UFS driver calls scsi_scan_host() from ufshcd_add_lus(), do you
> agree that the above wait_event_timeout() call can be eliminated by splitting
> ufshpb_init() into two functions and by calling the part below
> wait_event_timeout() after scsi_scan_host() has finished?
Yes, I agree the above wait_event_timeout() call can be eliminated.
I will change these functions.
> > +void ufshpb_remove(struct ufs_hba *hba)
> > +{
> > + struct ufshpb_lu *hpb, *n_hpb;
> > + struct ufsf_feature_info *ufsf;
> > + struct scsi_device *sdev;
> > +
> > + ufsf = &hba->ufsf;
> > +
> > + list_for_each_entry_safe(hpb, n_hpb, &lh_hpb_lu, list_hpb_lu) {
> > + ufshpb_set_state(hpb, HPB_FAILED);
> > +
> > + sdev = hpb->sdev_ufs_lu;
> > + sdev->hostdata = NULL;
> > +
> > + ufshpb_destroy_region_tbl(hpb);
> > +
> > + list_del_init(&hpb->list_hpb_lu);
> > + ufshpb_remove_sysfs(hpb);
> > +
> > + kfree(hpb);
> > + }
> > +
> > + dev_info(hba->dev, "ufshpb: remove success\n");
> > +}
>
> Should the code in the body of the above loop perhaps be called from inside
> ufshcd_slave_destroy()?
Moving other stuffs in the loop is good idea, but removing attributes is problem.
To avoid adding new kobject, I will try to use sysfs_merge_group()
for adding attributes. To delete merged attributes, sysfs_unmerge_group()
should be called. But sysfs_remove_groups() is called before calling ufshcd_slave_destroy().
> > +void ufshpb_scan_feature(struct ufs_hba *hba)
> > +{
> > + init_waitqueue_head(&hba->ufsf.sdev_wait);
> > + atomic_set(&hba->ufsf.slave_conf_cnt, 0);
> > +
> > + if (hba->dev_info.wspecversion >= HPB_SUPPORT_VERSION &&
> > + (hba->dev_info.b_ufs_feature_sup & UFS_DEV_HPB_SUPPORT))
> > + async_schedule(ufshpb_init, &hba->ufsf);
> > +}
>
> Why does this function use async_schedule()?
The wait_event_timeout() call will be removed, so it will be changed.
Thanks,
Daejun
