From: "Marc-André Lureau" <marcandre.lureau@redhat.com>
To: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
Sergio Lopez Pascual <slp@redhat.com>,
Baoquan He <bhe@redhat.com>, "Somlo, Gabriel" <somlo@cmu.edu>,
xiaolong.ye@intel.com
Subject: Re: [PATCH v15 11/11] RFC: fw_cfg: do DMA read operation
Date: Wed, 28 Feb 2018 13:27:02 +0100 [thread overview]
Message-ID: <CAMxuvay=4qSLZBVHoPPX+6ZNEmyS2G-69hHqW3HVdxFoS90h1g@mail.gmail.com> (raw)
In-Reply-To: <20180227020104-mutt-send-email-mst@kernel.org>
Hi
On Tue, Feb 27, 2018 at 1:04 AM, Michael S. Tsirkin <mst@redhat.com> wrote:
> On Thu, Feb 15, 2018 at 10:33:12PM +0100, Marc-André Lureau wrote:
>> Modify fw_cfg_read_blob() to use DMA if the device supports it.
>> Return errors, because the operation may fail.
>>
>> So far, only one call in fw_cfg_register_dir_entries() is using
>> kmalloc'ed buf and is thus clearly eligible to DMA read.
>>
>> Initially, I didn't implement DMA read to speed up boot time, but as a
>> first step before introducing DMA write (since read operations were
>> already presents). Even more, I didn't realize fw-cfg entries were
>> being read by the kernel during boot by default. But actally fw-cfg
>> entries are being populated during module probe. I knew DMA improved a
>> lot bios boot time (the main reason the DMA interface was added
>> afaik). Let see the time it would take to read the whole ACPI
>> tables (128kb allocated)
>>
>> # time cat /sys/firmware/qemu_fw_cfg/by_name/etc/acpi/tables/raw
>> - with DMA: sys 0m0.003s
>> - without DMA (-global fw_cfg.dma_enabled=off): sys 0m7.674s
>>
>> FW_CFG_FILE_DIR (0x19) is the only "file" that is read during kernel
>> boot to populate sysfs qemu_fw_cfg directory, and it is quite
>> small (1-2kb). Since it does not expose itself, in order to measure
>> the time it takes to read such small file, I took a comparable sized
>> file of 2048 bytes and exposed it (-fw_cfg test,file=file with a
>> modified read_raw enabling DMA)
>>
>> # perf stat -r 100 cat /sys/firmware/qemu_fw_cfg/by_name/test/raw >/dev/null
>> - with DMA:
>> 0.636037 task-clock (msec) # 0.141 CPUs utilized ( +- 1.19% )
>> - without DMA:
>> 6.430128 task-clock (msec) # 0.622 CPUs utilized ( +- 0.22% )
>>
>> That's a few msec saved during boot by enabling DMA read (the gain
>> would be more substantial if other & bigger fw-cfg entries are read by
>> others from sysfs, unfortunately, it's not clear if we can always
>> enable DMA there)
>>
>> Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
>> ---
>> drivers/firmware/qemu_fw_cfg.c | 61 ++++++++++++++++++++++++++++++++++--------
>> 1 file changed, 50 insertions(+), 11 deletions(-)
>>
>> diff --git a/drivers/firmware/qemu_fw_cfg.c b/drivers/firmware/qemu_fw_cfg.c
>> index 3015e77aebca..94df57e9be66 100644
>> --- a/drivers/firmware/qemu_fw_cfg.c
>> +++ b/drivers/firmware/qemu_fw_cfg.c
>> @@ -124,12 +124,47 @@ static ssize_t fw_cfg_dma_transfer(void *address, u32 length, u32 control)
>> return ret;
>> }
>>
>> +/* with acpi & dev locks taken */
>> +static ssize_t fw_cfg_read_blob_dma(u16 key,
>> + void *buf, loff_t pos, size_t count)
>> +{
>> + ssize_t ret;
>> +
>> + if (pos == 0) {
>> + ret = fw_cfg_dma_transfer(buf, count, key << 16
>> + | FW_CFG_DMA_CTL_SELECT
>> + | FW_CFG_DMA_CTL_READ);
>> + } else {
>> + fw_cfg_sel_endianness(key);
>> + ret = fw_cfg_dma_transfer(NULL, pos, FW_CFG_DMA_CTL_SKIP);
>> + if (ret < 0)
>> + return ret;
>> + ret = fw_cfg_dma_transfer(buf, count,
>> + FW_CFG_DMA_CTL_READ);
>> + }
>> +
>> + return ret;
>> +}
>> +
>> +/* with acpi & dev locks taken */
>> +static ssize_t fw_cfg_read_blob_io(u16 key,
>> + void *buf, loff_t pos, size_t count)
>> +{
>> + fw_cfg_sel_endianness(key);
>> + while (pos-- > 0)
>> + ioread8(fw_cfg_reg_data);
>> + ioread8_rep(fw_cfg_reg_data, buf, count);
>> + return count;
>> +}
>> +
>> /* read chunk of given fw_cfg blob (caller responsible for sanity-check) */
>> static ssize_t fw_cfg_read_blob(u16 key,
>> - void *buf, loff_t pos, size_t count)
>> + void *buf, loff_t pos, size_t count,
>> + bool dma)
>> {
>> u32 glk = -1U;
>> acpi_status status;
>> + ssize_t ret;
>>
>> /* If we have ACPI, ensure mutual exclusion against any potential
>> * device access by the firmware, e.g. via AML methods:
>
> so this adds a dma flag to fw_cfg_read_blob.
>
>
>
>> @@ -143,14 +178,17 @@ static ssize_t fw_cfg_read_blob(u16 key,
>> }
>>
>> mutex_lock(&fw_cfg_dev_lock);
>> - fw_cfg_sel_endianness(key);
>> - while (pos-- > 0)
>> - ioread8(fw_cfg_reg_data);
>> - ioread8_rep(fw_cfg_reg_data, buf, count);
>> + if (dma && fw_cfg_dma_enabled()) {
>> + ret = fw_cfg_read_blob_dma(key, buf, pos, count);
>> + } else {
>> + ret = fw_cfg_read_blob_io(key, buf, pos, count);
>> + }
>> +
>> mutex_unlock(&fw_cfg_dev_lock);
>>
>> acpi_release_global_lock(glk);
>> - return count;
>> +
>> + return ret;
>> }
>>
>> #ifdef CONFIG_CRASH_CORE
>
> If set to false it does io, if set to true it does dma.
>
> I would prefer passing an accessor function pointer
> since that's clearer than true/false.
ok
>
>> @@ -284,7 +322,7 @@ static int fw_cfg_do_platform_probe(struct platform_device *pdev)
>>
>> /* verify fw_cfg device signature */
>> if (fw_cfg_read_blob(FW_CFG_SIGNATURE, sig,
>> - 0, FW_CFG_SIG_SIZE) < 0 ||
>> + 0, FW_CFG_SIG_SIZE, false) < 0 ||
>> memcmp(sig, "QEMU", FW_CFG_SIG_SIZE) != 0) {
>> fw_cfg_io_cleanup();
>> return -ENODEV;
>> @@ -468,7 +506,8 @@ static ssize_t fw_cfg_sysfs_read_raw(struct file *filp, struct kobject *kobj,
>> if (count > entry->size - pos)
>> count = entry->size - pos;
>>
>> - return fw_cfg_read_blob(entry->select, buf, pos, count);
>> + /* do not use DMA, virt_to_phys(buf) might not be ok */
>> + return fw_cfg_read_blob(entry->select, buf, pos, count, false);
>> }
>>
>> static struct bin_attribute fw_cfg_sysfs_attr_raw = {
>> @@ -634,7 +673,7 @@ static int fw_cfg_register_dir_entries(void)
>> size_t dir_size;
>>
>> ret = fw_cfg_read_blob(FW_CFG_FILE_DIR, &files_count,
>> - 0, sizeof(files_count));
>> + 0, sizeof(files_count), false);
>> if (ret < 0)
>> return ret;
>>
>> @@ -646,7 +685,7 @@ static int fw_cfg_register_dir_entries(void)
>> return -ENOMEM;
>>
>> ret = fw_cfg_read_blob(FW_CFG_FILE_DIR, dir,
>> - sizeof(files_count), dir_size);
>> + sizeof(files_count), dir_size, false);
>> if (ret < 0)
>> goto end;
>>
>> @@ -697,7 +736,7 @@ static int fw_cfg_sysfs_probe(struct platform_device *pdev)
>> goto err_probe;
>>
>> /* get revision number, add matching top-level attribute */
>> - err = fw_cfg_read_blob(FW_CFG_ID, &rev, 0, sizeof(rev));
>> + err = fw_cfg_read_blob(FW_CFG_ID, &rev, 0, sizeof(rev), false);
>> if (err < 0)
>> goto err_probe;
>
>
> Looks like all callers pass in false as parameter.
> Given this, how can this speed up any operations?
>
> Are you sure you tested this properly?
I did modify read_raw to conduct testing ( the part "with a
modified read_raw enabling DMA" should be before, updating commit message).
>
>> --
>> 2.16.1.73.g5832b7e9f2
next prev parent reply other threads:[~2018-02-28 12:27 UTC|newest]
Thread overview: 36+ messages / expand[flat|nested] mbox.gz Atom feed top
2018-02-15 21:33 [PATCH v15 00/11] fw_cfg: add DMA operations & etc/vmcoreinfo support Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 01/11] crash: export paddr_vmcoreinfo_note() Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 02/11] fw_cfg: add a public uapi header Marc-André Lureau
2018-02-27 0:06 ` Michael S. Tsirkin
2018-02-28 11:50 ` Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 03/11] fw_cfg: fix sparse warnings in fw_cfg_sel_endianness() Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 04/11] fw_cfg: fix sparse warnings with fw_cfg_file Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 05/11] fw_cfg: fix sparse warning reading FW_CFG_ID Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 06/11] fw_cfg: fix sparse warnings around FW_CFG_FILE_DIR read Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 07/11] fw_cfg: remove inline from fw_cfg_read_blob() Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 08/11] fw_cfg: handle fw_cfg_read_blob() error Marc-André Lureau
2018-02-27 0:20 ` Michael S. Tsirkin
2018-02-28 11:49 ` Marc-André Lureau
2018-02-28 13:01 ` Gabriel Somlo
2018-02-28 15:32 ` Michael S. Tsirkin
2018-02-28 23:25 ` Gabriel Somlo
2018-02-28 23:58 ` Michael S. Tsirkin
2018-02-28 23:58 ` Michael S. Tsirkin
2018-03-01 0:49 ` Gabriel Somlo
2018-03-01 0:35 ` Gabriel Somlo
2018-02-15 21:33 ` [PATCH v15 09/11] fw_cfg: add DMA register Marc-André Lureau
2018-02-15 21:33 ` [PATCH v15 10/11] fw_cfg: write vmcoreinfo details Marc-André Lureau
2018-02-27 0:28 ` Michael S. Tsirkin
2018-02-28 12:22 ` Marc-André Lureau
2018-02-28 15:34 ` Michael S. Tsirkin
2018-02-15 21:33 ` [PATCH v15 11/11] RFC: fw_cfg: do DMA read operation Marc-André Lureau
2018-02-27 0:04 ` Michael S. Tsirkin
2018-02-28 12:27 ` Marc-André Lureau [this message]
2018-02-28 15:35 ` Michael S. Tsirkin
2018-02-28 15:41 ` Marc-André Lureau
2018-02-28 15:48 ` Michael S. Tsirkin
2018-02-28 16:00 ` Marc-André Lureau
2018-02-28 17:16 ` Michael S. Tsirkin
2018-02-28 17:17 ` Michael S. Tsirkin
2018-02-28 17:22 ` Marc-André Lureau
2018-02-27 0:29 ` [PATCH v15 00/11] fw_cfg: add DMA operations & etc/vmcoreinfo support Michael S. Tsirkin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAMxuvay=4qSLZBVHoPPX+6ZNEmyS2G-69hHqW3HVdxFoS90h1g@mail.gmail.com' \
--to=marcandre.lureau@redhat.com \
--cc=bhe@redhat.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mst@redhat.com \
--cc=slp@redhat.com \
--cc=somlo@cmu.edu \
--cc=xiaolong.ye@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).