* [PATCH RESEND] PCI/VPD: Add simple VPD read cache
@ 2021-11-18 21:10 Heiner Kallweit
2021-12-15 20:53 ` Bjorn Helgaas
0 siblings, 1 reply; 2+ messages in thread
From: Heiner Kallweit @ 2021-11-18 21:10 UTC (permalink / raw)
To: Bjorn Helgaas; +Cc: linux-pci@vger.kernel.org
VPD reads can be time-consuming operations, and when reading bytes
sequentially then worst-case we may read the same VPD dword four times.
Typically it happens in pci_vpd_size() already that the same VPD dword
is read more than once. Therefore let's add a simple read cache that
caches the last read VPD dword.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
---
drivers/pci/vpd.c | 13 +++++++++++++
include/linux/pci.h | 2 ++
2 files changed, 15 insertions(+)
diff --git a/drivers/pci/vpd.c b/drivers/pci/vpd.c
index 75e48df2e..2315e45f6 100644
--- a/drivers/pci/vpd.c
+++ b/drivers/pci/vpd.c
@@ -194,6 +194,11 @@ static ssize_t pci_vpd_read(struct pci_dev *dev, loff_t pos, size_t count,
break;
}
+ if (vpd->cached_offset == (pos & ~3)) {
+ val = vpd->cached_val;
+ goto process;
+ }
+
ret = pci_user_write_config_word(dev, vpd->cap + PCI_VPD_ADDR,
pos & ~3);
if (ret < 0)
@@ -206,6 +211,9 @@ static ssize_t pci_vpd_read(struct pci_dev *dev, loff_t pos, size_t count,
if (ret < 0)
break;
+ vpd->cached_val = val;
+ vpd->cached_offset = pos & ~3;
+process:
skip = pos & 3;
for (i = 0; i < sizeof(u32); i++) {
if (i >= skip) {
@@ -242,6 +250,10 @@ static ssize_t pci_vpd_write(struct pci_dev *dev, loff_t pos, size_t count,
return -EINTR;
while (pos < end) {
+ /* invalidate read cache */
+ if (vpd->cached_offset == pos)
+ vpd->cached_offset = -1;
+
ret = pci_user_write_config_dword(dev, vpd->cap + PCI_VPD_DATA,
get_unaligned_le32(buf));
if (ret < 0)
@@ -270,6 +282,7 @@ void pci_vpd_init(struct pci_dev *dev)
dev->vpd.cap = pci_find_capability(dev, PCI_CAP_ID_VPD);
mutex_init(&dev->vpd.lock);
+ dev->vpd.cached_offset = -1;
}
static ssize_t vpd_read(struct file *filp, struct kobject *kobj,
diff --git a/include/linux/pci.h b/include/linux/pci.h
index cd8aa6fce..fb123c248 100644
--- a/include/linux/pci.h
+++ b/include/linux/pci.h
@@ -297,6 +297,8 @@ enum pcie_link_width pcie_get_width_cap(struct pci_dev *dev);
struct pci_vpd {
struct mutex lock;
unsigned int len;
+ u32 cached_val;
+ int cached_offset;
u8 cap;
};
--
2.33.0
^ permalink raw reply related [flat|nested] 2+ messages in thread* Re: [PATCH RESEND] PCI/VPD: Add simple VPD read cache
2021-11-18 21:10 [PATCH RESEND] PCI/VPD: Add simple VPD read cache Heiner Kallweit
@ 2021-12-15 20:53 ` Bjorn Helgaas
0 siblings, 0 replies; 2+ messages in thread
From: Bjorn Helgaas @ 2021-12-15 20:53 UTC (permalink / raw)
To: Heiner Kallweit; +Cc: Bjorn Helgaas, linux-pci@vger.kernel.org
On Thu, Nov 18, 2021 at 10:10:20PM +0100, Heiner Kallweit wrote:
> VPD reads can be time-consuming operations, and when reading bytes
> sequentially then worst-case we may read the same VPD dword four times.
> Typically it happens in pci_vpd_size() already that the same VPD dword
> is read more than once. Therefore let's add a simple read cache that
> caches the last read VPD dword.
This adds 8 more bytes to pci_dev for everybody. Can you quantify the
benefit? It seems like well-formed VPD typically has 3-4 tags, so
we'd be talking about maybe optimizing 20ish reads (4 bytes per tag,
reading the same dword 4 times per tag) down to 3-4 dword reads. Not
sure the cost/benefit really pays off here.
For crappy devices with junk in their VPD, I certainly see how we
could do a huge number of reads. But that doesn't seem like it would
be worth optimizing.
> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
> ---
> drivers/pci/vpd.c | 13 +++++++++++++
> include/linux/pci.h | 2 ++
> 2 files changed, 15 insertions(+)
>
> diff --git a/drivers/pci/vpd.c b/drivers/pci/vpd.c
> index 75e48df2e..2315e45f6 100644
> --- a/drivers/pci/vpd.c
> +++ b/drivers/pci/vpd.c
> @@ -194,6 +194,11 @@ static ssize_t pci_vpd_read(struct pci_dev *dev, loff_t pos, size_t count,
> break;
> }
>
> + if (vpd->cached_offset == (pos & ~3)) {
> + val = vpd->cached_val;
> + goto process;
> + }
> +
> ret = pci_user_write_config_word(dev, vpd->cap + PCI_VPD_ADDR,
> pos & ~3);
> if (ret < 0)
> @@ -206,6 +211,9 @@ static ssize_t pci_vpd_read(struct pci_dev *dev, loff_t pos, size_t count,
> if (ret < 0)
> break;
>
> + vpd->cached_val = val;
> + vpd->cached_offset = pos & ~3;
> +process:
> skip = pos & 3;
> for (i = 0; i < sizeof(u32); i++) {
> if (i >= skip) {
> @@ -242,6 +250,10 @@ static ssize_t pci_vpd_write(struct pci_dev *dev, loff_t pos, size_t count,
> return -EINTR;
>
> while (pos < end) {
> + /* invalidate read cache */
> + if (vpd->cached_offset == pos)
> + vpd->cached_offset = -1;
> +
> ret = pci_user_write_config_dword(dev, vpd->cap + PCI_VPD_DATA,
> get_unaligned_le32(buf));
> if (ret < 0)
> @@ -270,6 +282,7 @@ void pci_vpd_init(struct pci_dev *dev)
>
> dev->vpd.cap = pci_find_capability(dev, PCI_CAP_ID_VPD);
> mutex_init(&dev->vpd.lock);
> + dev->vpd.cached_offset = -1;
> }
>
> static ssize_t vpd_read(struct file *filp, struct kobject *kobj,
> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index cd8aa6fce..fb123c248 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -297,6 +297,8 @@ enum pcie_link_width pcie_get_width_cap(struct pci_dev *dev);
> struct pci_vpd {
> struct mutex lock;
> unsigned int len;
> + u32 cached_val;
> + int cached_offset;
> u8 cap;
> };
>
> --
> 2.33.0
>
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2021-12-15 20:53 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2021-11-18 21:10 [PATCH RESEND] PCI/VPD: Add simple VPD read cache Heiner Kallweit
2021-12-15 20:53 ` Bjorn Helgaas
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox