Shrink time spent in copy_data_pages()

copy_data_pages() can take a long (multi-second) time to finish, and currently the only indication of that is the timestamp difference between print messages right before and right after. The timestamp is also immediately reset afterwards to the time before image creation, making it even harder to spot this delay. And this function runs in a critical section with a single CPU online and syscore suspended, so it should be kept as quick as possible to keep the system responsive.

Call into swsusp_show_speed() to report the amount of data, time taken, and copy speed of copy_data_pages() to make it easier to spot delays and verify performance improvements. The current time is obtained through sched_clock() instead of ktime_get() since timekeeping is suspended in this region.

Signed-off-by Nícolas F. R. A. Prado <[email protected]>

---
 kernel/power/snapshot.c | 5 +++++
 1 file changed, 5 insertions(+)

kernel/power/snapshot.c +5 -1

index e249e5786fbc..4f452baf31dc 100644
@@ -13,6 +13,7 @@
 #include <linux/version.h>
 #include <linux/module.h>
 #include <linux/mm.h>
+#include <linux/sched/clock.h>
 #include <linux/suspend.h>
 #include <linux/delay.h>
 #include <linux/bitops.h>

@@ -2109,6 +2110,7 @@ static int swsusp_alloc(struct memory_bitmap *copy_bm,
 asmlinkage __visible int swsusp_save(void)
 {
        unsigned int nr_pages, nr_highmem;
+       ktime_t start, stop;
 
        pr_info("Creating image:\n");
 

@@ -2132,7 +2134,10 @@ asmlinkage __visible int swsusp_save(void)
         * Kill them.
         */
        drain_local_pages(NULL);
+       start = ns_to_ktime(sched_clock());
        nr_copy_pages = copy_data_pages(&copy_bm, &orig_bm, &zero_bm);
+       stop = ns_to_ktime(sched_clock());
+       swsusp_show_speed(start, stop, nr_copy_pages, "Copied");
 
        /*
         * End of critical section. From now on, we can write to memory,
 
--

2.53.0

Quoted 14 lines · Nícolas F. R. A. Prado

copy_data_pages() can take a long (multi-second) time to finish, and
currently the only indication of that is the timestamp difference between print messages right before and right after. The timestamp is
also immediately reset afterwards to the time before image creation,
making it even harder to spot this delay. And this function runs in a
critical section with a single CPU online and syscore suspended, so it
should be kept as quick as possible to keep the system responsive.

Quoted 9 lines

Call into swsusp_show_speed() to report the amount of data, time taken,
and copy speed of copy_data_pages() to make it easier to spot delays and
verify performance improvements. The current time is obtained through
sched_clock() instead of ktime_get() since timekeeping is suspended in
this region.

Why not use local_clock() for this?

Add a 'platform_enter' hibernation test level to allow testing the full hibernation path up until the last point of return in hibernation_platform_enter().

In particular this allows seeing print messages from the code run after the hibernation image is created on platforms that don't have an easily accessible serial console, since these messages are not present upon hibernation resume.

Signed-off-by Nícolas F. R. A. Prado <[email protected]>

---
 kernel/power/hibernate.c | 13 +++++++++++--
 kernel/power/main.c      |  1 +
 kernel/power/power.h     |  1 +
 3 files changed, 13 insertions(+), 2 deletions(-)

kernel/power/hibernate.c +11 -2

index 001e4556b137..15cb7a090770 100644
@@ -75,6 +75,7 @@ enum {
 static int hibernation_mode = HIBERNATION_SHUTDOWN;
 
 bool freezer_test_done;
+bool platform_enter_test_done;
 
 static const struct platform_hibernation_ops *hibernation_ops;
 

@@ -636,6 +637,11 @@ int hibernation_platform_enter(void)
                goto Power_up;
        }
 
+       if (hibernation_test(TEST_PLATFORM_ENTER)) {
+               platform_enter_test_done = true;
+               goto Power_up;
+       }
+
        hibernation_ops->enter();
        /* We should never get here */
        while (1);

@@ -700,6 +706,8 @@ static int power_down(void)
                break;
        case HIBERNATION_PLATFORM:
                error = hibernation_platform_enter();
+               if (platform_enter_test_done)
+                       goto exit;
                if (error == -EAGAIN || error == -EBUSY) {
                        events_check_enabled = false;
                        pr_info("Wakeup event detected during hibernation, rolling back.\n");

@@ -848,7 +856,7 @@ int hibernate(void)
                        else
                                error = power_down();
                }
-               if (error) {
+               if (error || platform_enter_test_done) {
                        /* recover any devices that refused to thaw */
                        dpm_resume_suspended_devices(PMSG_RECOVER);
                }

@@ -870,8 +878,9 @@ int hibernate(void)
        }
        thaw_processes();
 
-       /* Don't bother checking whether freezer_test_done is true */
+       /* Don't bother checking whether any of the test_done flags are true */
        freezer_test_done = false;
+       platform_enter_test_done = false;
  Exit:
        filesystems_thaw();
        pm_notifier_call_chain(PM_POST_HIBERNATION);

kernel/power/main.c +1 -0

index 7f53ef65b789..04f0c429519c 100644
@@ -273,6 +273,7 @@ static const char * const pm_tests[__TEST_AFTER_LAST] = {
        [TEST_PLATFORM] = "platform",
        [TEST_DEVICES] = "devices",
        [TEST_FREEZER] = "freezer",
+       [TEST_PLATFORM_ENTER] = "platform_enter",
 };
 
 static ssize_t pm_test_show(struct kobject *kobj, struct kobj_attribute *attr,

kernel/power/power.h +0 -0

index 7ccd709af93f..0d33d0e46ed0 100644

Quoted 101 lines · Nícolas F. R. A. Prado

Add a 'platform_enter' hibernation test level to allow testing the full
hibernation path up until the last point of return in hibernation_platform_enter().

In particular this allows seeing print messages from the code run after
the hibernation image is created on platforms that don't have an easily
accessible serial console, since these messages are not present upon
hibernation resume.

Signed-off-by Nícolas F. R. A. Prado <[email protected]>

---
 kernel/power/hibernate.c | 13 +++++++++++--
 kernel/power/main.c      |  1 +
 kernel/power/power.h     |  1 +

 3 files changed, 13 insertions(+), 2 deletions(-)

kernel/power/hibernate.c +6 -0

index 001e4556b137..15cb7a090770 100644
@@ -75,6 +75,7 @@ enum {
  static int hibernation_mode = HIBERNATION_SHUTDOWN;
 
  bool freezer_test_done;
+bool platform_enter_test_done;
 
  static const struct platform_hibernation_ops *hibernation_ops;
 

@@ -636,6 +637,11 @@ int hibernation_platform_enter(void)
                 goto Power_up;
         }
 
+       if (hibernation_test(TEST_PLATFORM_ENTER)) {
+               platform_enter_test_done = true;
+               goto Power_up;
+       }
+
         hibernation_ops->enter();
         /* We should never get here */
         while (1);

@@ -700,6 +706,8 @@ static int power_down(void)
                 break;

        case HIBERNATION_PLATFORM:
                error = hibernation_platform_enter(); +               if (platform_enter_test_done)
+                       goto exit;
                if (error == -EAGAIN || error == -EBUSY) { events_check_enabled = false; pr_info("Wakeup event detected during hibernation, rolling back.\n");

kernel/power/hibernate.c +0 -0

@@ -848,7 +856,7 @@ int hibernate(void)

                        else
                                error = power_down();
                }
-               if (error) {
+               if (error || platform_enter_test_done) { /* recover any devices that refused to thaw */
                        dpm_resume_suspended_devices(PMSG_RECOVER); }

kernel/power/hibernate.c +1 -1

@@ -870,8 +878,9 @@ int hibernate(void)
         }
         thaw_processes();
 
-       /* Don't bother checking whether freezer_test_done is true
 */
+       /* Don't bother checking whether any of the test_done flags

are true */
        freezer_test_done = false;
+       platform_enter_test_done = false; Exit:
        filesystems_thaw(); pm_notifier_call_chain(PM_POST_HIBERNATION);

kernel/power/main.c +1 -0

index 7f53ef65b789..04f0c429519c 100644
@@ -273,6 +273,7 @@ static const char * const
 pm_tests[__TEST_AFTER_LAST] = {
         [TEST_PLATFORM] = "platform",
         [TEST_DEVICES] = "devices",
         [TEST_FREEZER] = "freezer",
+       [TEST_PLATFORM_ENTER] = "platform_enter",
  };
 

 static ssize_t pm_test_show(struct kobject *kobj, struct kobj_attribute *attr,

kernel/power/power.h +1 -0

index 7ccd709af93f..0d33d0e46ed0 100644
@@ -259,6 +259,7 @@ enum {
         TEST_PLATFORM,
         TEST_DEVICES,
         TEST_FREEZER,
+       TEST_PLATFORM_ENTER,

Quoted 4 lines

kernel/power/power.h +0 -0

         /* keep last */

        __TEST_AFTER_LAST
 };

In the first place, test levels are designed to be applicable to both system suspend and hibernation and I don't see how this test can be applied to system suspend.

Second, the last device power off transition during hibernation may not be reversible and adding a test that goes past a point of no return is not particularly useful.

The current implementation of copy_data_pages() has at its center do_copy_page(), which copies data of all saveable pages in a loop one long at a time, while checking whether the page is fully zero to allow it to not be saved in the hibernation image.

As the comment above do_copy_page() mentions, this loop was used instead of the more optimized copy_page() and memcpy() due to them depending on fpu_begin()/fpu_end() (see [1] and commit 95018f7c94cb ("[PATCH] swsusp:
do not use memcpy for snapshotting memory")). However, since commit c6dbd3e5e69c ("x86/mmx_32: Remove X86_USE_3DNOW"), this limitation no longer holds, and it only affected x86-32 in the first place.

From testing it is clear that removing the zero page check and using copy_page() makes it almost 3 times quicker:
- Current:

    PM: hibernation: Copied 4940240 kbytes in 11.33 seconds (436.03 MB/s)
- With just the zero check removed:
    PM: hibernation: Copied 4974664 kbytes in 9.03 seconds (550.90 MB/s)
- With the zero check removed and using copy_page() instead:
    PM: hibernation: Copied 6275216 kbytes in 3.96 seconds (1584.65 MB/s)

Given that copy_data_pages() runs inside a critical section where only a single CPU is online and syscore is suspended, it should be kept as short as possible to keep the system responsive.

While switching from the loop to copy_page() brings big speed improvements, it also makes the zero page check much costlier since it can no longer be done in the middle of the copy. In fact upon testing adding the zero check alongside copy_page() made it slower than the current code.

The following shows a rough comparison of a few more metrics between the current code and when both copy_page() is used and the zero page check is disabled:
- Total time to hibernate:
- before: 13.77s
- after: 14.13s

- Total time to resume:
- before: 5.85s
- after: 5.85s
- Total time in syscore_suspend():
- before: 11.3s
- after: 4.08s

- Total image size written to disk:
- before: 4956296kB => 2606402kB compressed = 2.60MB
- after: 6274608 => 2616624kB compressed = 2.61MB

As can be seen the total time to hibernate is roughly the same, suggesting that the time saved with the more efficient copy_page() is payed by having to compress more data (the zero pages). The time to resume remains the same. And the hibernation image size is basically the same, as the zero pages compress well (the case would be quite different if compression is disabled). The big win here is that the time between syscore_suspend() and syscore_resume() is much lower, meaning the system will be more responsive throughout hibernation.

Expose a 'nozerocheck' module parameter

to allow the zero page check to
be disabled and the faster copy_page() to be used, giving the option for userspace to choose between a more responsive system and reducing the disk usage particularly when compression is disabled.

Testing was done on the SteamDeck OLED.

[1] https://lore.kernel.org/all/[email protected]/

Signed-off-by Nícolas F. R. A. Prado <[email protected]>

---
 kernel/power/snapshot.c | 37 ++++++++++++++++++++++++++-----------
 1 file changed, 26 insertions(+), 11 deletions(-)

kernel/power/snapshot.c +26 -12

index 4f452baf31dc..00feac18faae 100644
@@ -40,6 +40,9 @@
 
 #include "power.h"
 
+static bool nozerocheck;
+module_param(nozerocheck, bool, 0644);
+
 #if defined(CONFIG_STRICT_KERNEL_RWX) && defined(CONFIG_ARCH_HAS_SET_MEMORY)
 static bool hibernate_restore_protection;
 static bool hibernate_restore_protection_active;

@@ -1425,11 +1428,9 @@ static unsigned int count_data_pages(void)
 }
 
 /*
- * This is needed, because copy_page and memcpy are not usable for copying
- * task structs. Returns true if the page was filled with only zeros,
- * otherwise false.
+ * Returns true if the page was filled with only zeros, otherwise false.
  */
-static inline bool do_copy_page(long *dst, long *src)
+static inline bool do_copy_page_zerocheck(long *dst, long *src)
 {
        long z = 0;
        int n;

@@ -1441,6 +1442,12 @@ static inline bool do_copy_page(long *dst, long *src)
        return !z;
 }
 
+static inline bool do_copy_page_nozerocheck(long *dst, long *src)
+{
+       copy_page(dst, src);
+       return false;
+}
+
 /**
  * safe_copy_page - Copy a page in a safe way.
  *

@@ -1450,7 +1457,8 @@ static inline bool do_copy_page(long *dst, long *src)
  * always returns 'true'. Returns true if the page was entirely composed of
  * zeros, otherwise it will return false.
  */
-static bool safe_copy_page(void *dst, struct page *s_page)
+static bool safe_copy_page(void *dst, struct page *s_page,
+                          bool (*do_copy_page)(long *dst, long *src))
 {
        bool zeros_only;
 

@@ -1471,7 +1479,8 @@ static inline struct page *page_is_saveable(struct zone *zone, unsigned long pfn
                saveable_highmem_page(zone, pfn) : saveable_page(zone, pfn);
 }
 
-static bool copy_data_page(unsigned long dst_pfn, unsigned long src_pfn)
+static bool copy_data_page(unsigned long dst_pfn, unsigned long src_pfn,
+                          bool (*do_copy_page)(long *dst, long *src))
 {
        struct page *s_page, *d_page;
        void *src, *dst;

@@ -1491,12 +1500,13 @@ static bool copy_data_page(unsigned long dst_pfn, unsigned long src_pfn)
                         * The page pointed to by src may contain some kernel
                         * data modified by kmap_atomic()
                         */
-                       zeros_only = safe_copy_page(buffer, s_page);
+                       zeros_only = safe_copy_page(buffer, s_page, do_copy_page);
                        dst = kmap_local_page(d_page);
                        copy_page(dst, buffer);
                        kunmap_local(dst);
                } else {
-                       zeros_only = safe_copy_page(page_address(d_page), s_page);
+                       zeros_only = safe_copy_page(page_address(d_page),
+                                                   s_page, do_copy_page);
                }
        }
        return zeros_only;

@@ -1504,10 +1514,12 @@ static bool copy_data_page(unsigned long dst_pfn, unsigned long src_pfn)
 #else
 #define page_is_saveable(zone, pfn)    saveable_page(zone, pfn)
 
-static inline int copy_data_page(unsigned long dst_pfn, unsigned long src_pfn)
+static inline int
+copy_data_page(unsigned long dst_pfn, unsigned long src_pfn,
+              bool (*do_copy_page)(long *dst, long *src))
 {
        return safe_copy_page(page_address(pfn_to_page(dst_pfn)),
-                               pfn_to_page(src_pfn));
+                               pfn_to_page(src_pfn), do_copy_page);
 }
 #endif /* CONFIG_HIGHMEM */
 

@@ -1524,6 +1536,9 @@ static unsigned long copy_data_pages(struct memory_bitmap *copy_bm,
        unsigned long copied_pages = 0;
        struct zone *zone;
        unsigned long pfn, copy_pfn;
+       bool (*do_copy_page)(long *dst, long *src);
+
+       do_copy_page = nozerocheck ? do_copy_page_nozerocheck : do_copy_page_zerocheck;
 
        for_each_populated_zone(zone) {
                unsigned long max_zone_pfn;

@@ -1541,7 +1556,7 @@ static unsigned long copy_data_pages(struct memory_bitmap *copy_bm,
                pfn = memory_bm_next_pfn(orig_bm);
                if (unlikely(pfn == BM_END_OF_MAP))
                        break;
-               if (copy_data_page(copy_pfn, pfn)) {
+               if (copy_data_page(copy_pfn, pfn, do_copy_page)) {
                        memory_bm_set_bit(zero_bm, pfn);
                        /* Use this copy_pfn for a page that is not full of zeros */
                        continue;
 
--

2.53.0

Quoted 32 lines · Nícolas F. R. A. Prado

The current implementation of copy_data_pages() has at its center do_copy_page(), which copies data of all saveable pages in a loop one
long at a time, while checking whether the page is fully zero to allow
it to not be saved in the hibernation image.

As the comment above do_copy_page() mentions, this loop was used instead
of the more optimized copy_page() and memcpy() due to them depending on
fpu_begin()/fpu_end() (see [1] and commit 95018f7c94cb ("[PATCH] swsusp:
do not use memcpy for snapshotting memory")). However, since commit c6dbd3e5e69c ("x86/mmx_32: Remove X86_USE_3DNOW"), this limitation no
longer holds, and it only affected x86-32 in the first place.

From testing it is clear that removing the zero page check and using
copy_page() makes it almost 3 times quicker:
- Current:
    PM: hibernation: Copied 4940240 kbytes in 11.33 seconds (436.03 MB/s)
- With just the zero check removed:
    PM: hibernation: Copied 4974664 kbytes in 9.03 seconds (550.90 MB/s)
- With the zero check removed and using copy_page() instead:
    PM: hibernation: Copied 6275216 kbytes in 3.96 seconds (1584.65 MB/s)

Quoted 5 lines

Given that copy_data_pages() runs inside a critical section where only
a single CPU is online and syscore is suspended, it should be kept as
short as possible to keep the system responsive.

I'm not sure if being responsive during hibernation is all that important.

Also, the change is about using copy_page() to reduce the time spent in syscore_suspend(), so the subject is a bit confusing.

Quoted 7 lines

While switching from the loop to copy_page() brings big speed improvements, it also makes the zero page check much costlier since it
can no longer be done in the middle of the copy. In fact upon testing
adding the zero check alongside copy_page() made it slower than the current code.

Guess what, CPUs have caches.

Quoted 8 lines

The following shows a rough comparison of a few more metrics between the
current code and when both copy_page() is used and the zero page check
is disabled:
- Total time to hibernate:
  - before: 13.77s
  - after: 14.13s

Score for the current code.

Quoted 6 lines

- Total time to resume:
  - before: 5.85s
  - after: 5.85s
- Total time in syscore_suspend():
  - before: 11.3s
  - after: 4.08s

Why does this matter?

Quoted 4 lines

- Total image size written to disk:
  - before: 4956296kB => 2606402kB compressed = 2.60MB
  - after: 6274608 => 2616624kB compressed = 2.61MB

"kB" seems to be missing in the bottom row.

Which means that this is a score for the current code again because it
needs to copy and compress fewer pages which means that less energy needs to be used (which may be kind of relevant in low-battery situations).

Quoted 14 lines

As can be seen the total time to hibernate is roughly the same, suggesting that the time saved with the more efficient copy_page() is
payed by having to compress more data (the zero pages). The time to resume remains the same. And the hibernation image size is basically the
same, as the zero pages compress well (the case would be quite different
if compression is disabled). The big win here is that the time between
syscore_suspend() and syscore_resume() is much lower, meaning the system
will be more responsive throughout hibernation.

Quoted 1 line

Expose a 'nozerocheck' module parameter

I'd rather not.

Either the switch-over is compelling enough to do it unconditionally, or it is not, in which case the existing code can be used just fine.

As it stands, I'm not convinced.

Quoted 159 lines

to allow the zero page check to
be disabled and the faster copy_page() to be used, giving the option for
userspace to choose between a more responsive system and reducing the
disk usage particularly when compression is disabled.

Testing was done on the SteamDeck OLED.

[1]
https://lore.kernel.org/all/[email protected]/

Signed-off-by Nícolas F. R. A. Prado <[email protected]>

---
 kernel/power/snapshot.c | 37 ++++++++++++++++++++++++++-----------

 1 file changed, 26 insertions(+), 11 deletions(-)

kernel/power/snapshot.c +3 -0

index 4f452baf31dc..00feac18faae 100644
@@ -40,6 +40,9 @@
 
  #include "power.h"
 
+static bool nozerocheck;
+module_param(nozerocheck, bool, 0644);
+
  #if defined(CONFIG_STRICT_KERNEL_RWX) &&
 defined(CONFIG_ARCH_HAS_SET_MEMORY)
  static bool hibernate_restore_protection;

 static bool hibernate_restore_protection_active;

kernel/power/snapshot.c +0 -1

@@ -1425,11 +1428,9 @@ static unsigned int count_data_pages(void)
  }
 
  /*
- * This is needed, because copy_page and memcpy are not usable for

copying
- * task structs. Returns true if the page was filled with only zeros,
- * otherwise false.
+ * Returns true if the page was filled with only zeros, otherwise false.
  */

kernel/power/snapshot.c +1 -1

-static inline bool do_copy_page(long *dst, long *src)
+static inline bool do_copy_page_zerocheck(long *dst, long *src)

 {
        long z = 0;
        int n;

kernel/power/snapshot.c +6 -0

@@ -1441,6 +1442,12 @@ static inline bool do_copy_page(long *dst,
 long *src)
         return !z;
  }
 
+static inline bool do_copy_page_nozerocheck(long *dst, long *src)
+{
+       copy_page(dst, src);
+       return false;
+}
+
  /**
   * safe_copy_page - Copy a page in a safe way.

  *

kernel/power/snapshot.c +0 -0

@@ -1450,7 +1457,8 @@ static inline bool do_copy_page(long *dst,
 long *src)
   * always returns 'true'. Returns true if the page was entirely

composed of
  * zeros, otherwise it will return false.
  */

kernel/power/snapshot.c +2 -1

-static bool safe_copy_page(void *dst, struct page *s_page)
+static bool safe_copy_page(void *dst, struct page *s_page,
+                          bool (*do_copy_page)(long *dst, long

*src))
 {
        bool zeros_only;

kernel/power/snapshot.c +0 -0

@@ -1471,7 +1479,8 @@ static inline struct page
 *page_is_saveable(struct zone *zone, unsigned long pfn

                saveable_highmem_page(zone, pfn) :
saveable_page(zone, pfn);
 }

-static bool copy_data_page(unsigned long dst_pfn, unsigned long src_pfn)
+static bool copy_data_page(unsigned long dst_pfn, unsigned long src_pfn,
+                          bool (*do_copy_page)(long *dst, long *src))
 {
        struct page *s_page, *d_page;
        void *src, *dst;

kernel/power/snapshot.c +0 -0

@@ -1491,12 +1500,13 @@ static bool copy_data_page(unsigned long
 dst_pfn, unsigned long src_pfn)
                          * The page pointed to by src may contain

some kernel
                         * data modified by kmap_atomic() */
-                       zeros_only = safe_copy_page(buffer, s_page);
+                       zeros_only = safe_copy_page(buffer, s_page, do_copy_page);
                        dst = kmap_local_page(d_page); copy_page(dst, buffer); kunmap_local(dst);
                } else {
-                       zeros_only = safe_copy_page(page_address(d_page), s_page);
+                       zeros_only = safe_copy_page(page_address(d_page),
+                                                   s_page, do_copy_page);
                }
        }
        return zeros_only;

kernel/power/snapshot.c +4 -2

@@ -1504,10 +1514,12 @@ static bool copy_data_page(unsigned long
 dst_pfn, unsigned long src_pfn)
  #else
  #define page_is_saveable(zone, pfn)    saveable_page(zone, pfn)
 
-static inline int copy_data_page(unsigned long dst_pfn, unsigned
 long src_pfn)
+static inline int
+copy_data_page(unsigned long dst_pfn, unsigned long src_pfn,
+              bool (*do_copy_page)(long *dst, long *src))
  {
         return safe_copy_page(page_address(pfn_to_page(dst_pfn)),
-                               pfn_to_page(src_pfn));
+                               pfn_to_page(src_pfn),
 do_copy_page);

 }
 #endif /* CONFIG_HIGHMEM */

kernel/power/snapshot.c +3 -0

@@ -1524,6 +1536,9 @@ static unsigned long copy_data_pages(struct
 memory_bitmap *copy_bm,
         unsigned long copied_pages = 0;
         struct zone *zone;
         unsigned long pfn, copy_pfn;
+       bool (*do_copy_page)(long *dst, long *src);
+
+       do_copy_page = nozerocheck ? do_copy_page_nozerocheck :
 do_copy_page_zerocheck;
 

        for_each_populated_zone(zone) {
                unsigned long max_zone_pfn;

kernel/power/snapshot.c +1 -1

@@ -1541,7 +1556,7 @@ static unsigned long copy_data_pages(struct
 memory_bitmap *copy_bm,
                 pfn = memory_bm_next_pfn(orig_bm);
                 if (unlikely(pfn == BM_END_OF_MAP))
                         break;
-               if (copy_data_page(copy_pfn, pfn)) {
+               if (copy_data_page(copy_pfn, pfn, do_copy_page)) {
                         memory_bm_set_bit(zero_bm, pfn);
                         /* Use this copy_pfn for a page that is not

full of zeros */
                        continue;

--
2.53.0

Quoted 74 lines · Rafael J. Wysocki

Quoted 30 lines · Nícolas F. R. A. Prado

The current implementation of copy_data_pages() has at its center do_copy_page(), which copies data of all saveable pages in a loop one
long at a time, while checking whether the page is fully zero to allow
it to not be saved in the hibernation image.

As the comment above do_copy_page() mentions, this loop was used instead
of the more optimized copy_page() and memcpy() due to them depending on
fpu_begin()/fpu_end() (see [1] and commit 95018f7c94cb ("[PATCH] swsusp:
do not use memcpy for snapshotting memory")). However, since commit c6dbd3e5e69c ("x86/mmx_32: Remove X86_USE_3DNOW"), this limitation no
longer holds, and it only affected x86-32 in the first place.

From testing it is clear that removing the zero page check and using
copy_page() makes it almost 3 times quicker:
- Current:

    PM: hibernation: Copied 4940240 kbytes in 11.33 seconds (436.03
MB/s)
- With just the zero check removed:
    PM: hibernation: Copied 4974664 kbytes in 9.03 seconds (550.90
MB/s)
- With the zero check removed and using copy_page() instead:
    PM: hibernation: Copied 6275216 kbytes in 3.96 seconds (1584.65

Quoted 7 lines

MB/s)

Given that copy_data_pages() runs inside a critical section where only
a single CPU is online and syscore is suspended, it should be kept as
short as possible to keep the system responsive.

I'm not sure if being responsive during hibernation is all that important.

Also, the change is about using copy_page() to reduce the time spent in syscore_suspend(), so the subject is a bit confusing.

Quoted 7 lines

While switching from the loop to copy_page() brings big speed improvements, it also makes the zero page check much costlier since it
can no longer be done in the middle of the copy. In fact upon testing
adding the zero check alongside copy_page() made it slower than the current code.

Guess what, CPUs have caches.

Quoted 8 lines

The following shows a rough comparison of a few more metrics between the
current code and when both copy_page() is used and the zero page check
is disabled:
- Total time to hibernate:
- before: 13.77s
- after: 14.13s

Score for the current code.

Quoted 6 lines

- Total time to resume:
- before: 5.85s
- after: 5.85s
- Total time in syscore_suspend():
- before: 11.3s
- after: 4.08s

Quoted 1 line

Why does this matter?

Reducing the time spent here means reduced time between pm_wakeup_pending() checks, which ultimately means the kernel is quicker to abort hibernation in reaction to a pending wakeup.

Quoted 92 lines · Nícolas F. R. A. Prado

Quoted 90 lines · Rafael J. Wysocki

Quoted 42 lines · Nícolas F. R. A. Prado

The current implementation of copy_data_pages() has at its center
do_copy_page(), which copies data of all saveable pages in a loop
one
long at a time, while checking whether the page is fully zero to
allow
it to not be saved in the hibernation image.

As the comment above do_copy_page() mentions, this loop was used
instead
of the more optimized copy_page() and memcpy() due to them depending on
fpu_begin()/fpu_end() (see [1] and commit 95018f7c94cb ("[PATCH]
swsusp:
do not use memcpy for snapshotting memory")). However, since commit
c6dbd3e5e69c ("x86/mmx_32: Remove X86_USE_3DNOW"), this limitation
no
longer holds, and it only affected x86-32 in the first place.

From testing it is clear that removing the zero page check and using
copy_page() makes it almost 3 times quicker:
- Current:
    PM: hibernation: Copied 4940240 kbytes in 11.33 seconds (436.03
MB/s)
- With just the zero check removed:
    PM: hibernation: Copied 4974664 kbytes in 9.03 seconds (550.90
MB/s)
- With the zero check removed and using copy_page() instead:
    PM: hibernation: Copied 6275216 kbytes in 3.96 seconds (1584.65
MB/s)

Quoted 7 lines

Given that copy_data_pages() runs inside a critical section where
only
a single CPU is online and syscore is suspended, it should be kept
as
short as possible to keep the system responsive.

I'm not sure if being responsive during hibernation is all that important.

Also, the change is about using copy_page() to reduce the time spent
in syscore_suspend(), so the subject is a bit confusing.

Quoted 9 lines

While switching from the loop to copy_page() brings big speed improvements, it also makes the zero page check much costlier since
it
can no longer be done in the middle of the copy. In fact upon testing
adding the zero check alongside copy_page() made it slower than the
current code.

Guess what, CPUs have caches.

Quoted 9 lines

The following shows a rough comparison of a few more metrics between the
current code and when both copy_page() is used and the zero page
check
is disabled:
- Total time to hibernate:
  - before: 13.77s
  - after: 14.13s

Score for the current code.

Quoted 6 lines

- Total time to resume:
  - before: 5.85s
  - after: 5.85s
- Total time in syscore_suspend():
  - before: 11.3s
  - after: 4.08s

Quoted 5 lines

Quoted 1 line

Why does this matter?

Reducing the time spent here means reduced time between pm_wakeup_pending() checks, which ultimately means the kernel is quicker to abort hibernation in reaction to a pending wakeup.

That could be spelled out more directly in the patch changelog.

I'm still unsure why it matters though.  Hibernation is slow pretty much by definition, so why would anyone expect a fast response to a wakeup event during hibernation?  If they have latency constraints, shouldn't they use suspend instead?