[Drbd-dev] [PATCH] drbd: when change susp_uuid[NEW] to true, make sure susp_uuid[OLD] is false

Philipp Reisner philipp.reisner at linbit.com
Thu Dec 7 06:35:28 CET 2023


Hello Zhengbing,

no, I disagree. The code is good as it is since commit e31d63966fac73fc000b.

Here is my explanation:

IO is frozen because the resource lost quorum. The user changes the config
and reduces the quorum setting to 1.

The regular sequence is:

1. in sanitize_state()
   susp_uuid = { false, true }
   DRBD keeps the resource suspended longer.

2. in finish_state_change()
   it sets __NEW_CUR_UUIC and that moves to NEW_CUR_UUID later.

   in w_after_state_change()
3a when susp_uuid goes from false -> true then drbd_uuid_new_current()
3b Trigger another state change that sets susp_uuuid[NEW] back to false.

1. sanitize_sate()    susp_uuid = { true, false }
2. in finish_state_change() nothing relevant happens
3a no positive edge on susp_uuid no NEW_CUR_UUID bit nothing happens
3b susp_uuid[NEW] == false -> nothing happens


Now let us assume between 3a and 3b, another CPU does an unrelated
state change that usually triggers a new current uuid:

CPU0:
1. in sanitize_state()
   susp_uuid = { false, true }
   DRBD keeps the resource suspended longer.

2. in finish_state_change()
   it sets __NEW_CUR_UUIC and that moves to NEW_CUR_UUID later.

   in w_after_state_change()
3a when susp_uuid goes from false -> true
   drbd_uuid_new_current()

    CPU1:
    Here, another thread starts another unrelated state change
    that again wants to set NEW_CUR_UUID. But let us go into the details.

    1. in sanitize_state() susp_uuid = { true, true }
    2. in finish_state_chanage() it does _NOT_ set __NEW_CUR_UUID
       because the condition only allows it, when we have a
       positive edge on susp_uuid (= susp_uuid = { false, true })
       But this time, it is susp_uuid = { true, true }. No edge.
       so __NEW_CUR_UUID stays 0 -> NEW_CUR_UUID stays 0
    3a nothing, since no NEW_CUR_UUID
    3b creates a state change to set susp_uuid[NEW] to false

CPU: continues with w_after_state()
3b Trigger another state change that sets susp_uuuid[NEW] back to false.
   This is an "empty state change" as it is false and does nothing.


I think you have a mistake in your reasoning at T3.

best regards,
 Philipp

On Mon, Nov 27, 2023 at 3:48 PM 黄正兵 <zhengbing.huang at easystack.cn> wrote:
>
> Hi Philipp,
>
>
> The third point about modification is that it may not cover all scenarios.
>
> In w_after_state_change(), between generating the new current_uuid and
> clearing the susp_uuid, it is possible that another thread set the __NEW_CUR_UUID flag.
>
> The problem scenario is as follows:
>
> T1:
> begin_state_change()
> end_state_change()
>   __end_state_change()
>     ___end_state_change()
>     # susp_uuid value is {True, True}
>       T2: w_after_state_change()
>             drbd_uuid_new_current()
>
>                                                                               T3:
>                                                                                     begin_state_change()
>                                                                                     # set __NEW_CUR_UUID flag
>                                                                                     end_state_change()
>
>             begin_state_change()
>             # set susp_uuid[NEW] = False
>             end_state_change()
>
> In the end, the NEW_CUR_UUID flag bit is generated and not cleaned up.
>
> Therefore, the third point about modification can be change to this:
>
> diff --git a/drbd/drbd_state.c b/drbd/drbd_state.c
> index e35150340..2101c0ebc 100644
> --- a/drbd/drbd_state.c
> +++ b/drbd/drbd_state.c
> @@ -2553,6 +2553,7 @@ static bool should_try_become_up_to_date(struct drbd_device *device, enum drbd_d
>  static void finish_state_change(struct drbd_resource *resource, const char *tag)
>  {
>         enum drbd_role *role = resource->role;
> +       bool *susp_uuid = resource->susp_uuid;
>         struct drbd_device *device;
>         struct drbd_connection *connection;
>         bool starting_resync = false;
> @@ -2828,7 +2829,7 @@ static void finish_state_change(struct drbd_resource *resource, const char *tag)
>                 if (role[OLD] == R_SECONDARY && role[NEW] == R_PRIMARY)
>                         create_new_uuid = true;
>
> -               if (create_new_uuid)
> +               if (create_new_uuid && !susp_uuid[OLD])
>                         set_bit(__NEW_CUR_UUID, &device->flags);
>
>                 if (disk_state[NEW] != D_NEGOTIATING && get_ldev_if_state(device, D_DETACHING)) {
>
> best regards,
> Zhengbing
>
>
>
>
>
> From: Philipp Reisner <philipp.reisner at linbit.com>
> Date: 2023-11-25 15:47:16
> To:  "黄正兵" <zhengbing.huang at easystack.cn>
> Cc:  dongsheng.yang at easystack.cn,drbd-dev at lists.linbit.com
> Subject: Re: Re: [PATCH] drbd: when change susp_uuid[NEW] to true, make sure susp_uuid[OLD] is false>Hi Zhengbing,
> >
> >Lars and I came up with this:
> >
> >commit e31d63966fac73fc000b584ccb2580016eccda17
> >Author: Philipp Reisner <philipp.reisner at linbit.com>
> >Date:   Thu Nov 23 06:31:28 2023 +0100
> >
> >   drbd: fix susp_uuid clearing
> >
> >   I introduced susp_uuid for keeping a resource longer suspended for
> >   writing a new current-uuid.
> >   But in the following scenario, it failed to clear the susp_uuid bit:
> >
> >    1. The network connection between a primary and secondary node
> >       fails. The primary has the setting quorum=2. IO freezes on the
> >       primary node.
> >
> >    2. The user changes the quorum setting from 2 to one. The primary
> >       node generates a new current-uuid and clears the susp_uuid bit in
> >       a second state change.
> >
> >    3. If, during the second state change, another condition for creating
> >       a new current UUID evaluates to true, the code fails to clear the
> >       suspen_uuid bit.
> >
> >   Fixing this in three places:
> >    1. in sanitize_state(): Only set susp_uuid[NEW] if susp_uuid[OLD]
> >       was not set. That avoids it happening a second time after
> >       one iteration.
> >
> >    2. in w_after_state_chage(): Always end the cycle by setting clearing
> >       susp_uuid[NEW] if it is set. (Independently of susp_uuid[OLD].
> >
> >    3. in finish_sate_change(): Do not set NEW_CUR_UUID again if we are
> >       in this operation.
> >
> >   Strictly speaking, (2) is unnecessary after (1), but I also do it
> >   since it shortens the code. Also, (3) alone would fix the issue.
> >
> >   Fixes: d47f7456ab ("drbd: create new UUID before resuming IO upon
> >   regaining quorum")
> >
> >   Co-developed-by: zhengbing.huang <zhengbing.huang at easystack.cn>
> >   Co-developed-by: Lars Ellenberg <lars.ellenberg at linbit.com>
> >
> >diff --git a/drbd/drbd_state.c b/drbd/drbd_state.c
> >index e35150340..b884bb622 100644
> >--- a/drbd/drbd_state.c
> >+++ b/drbd/drbd_state.c
> >@@ -2353,7 +2353,8 @@ static void sanitize_state(struct drbd_resource *resource)
> >       resource->susp_quorum[NEW] =
> >               resource->res_opts.on_no_quorum == ONQ_SUSPEND_IO ?
> >!resource_has_quorum : false;
> >
> >-       if (resource_is_suspended(resource, OLD) &&
> >!resource_is_suspended(resource, NEW)) {
> >+       if (!resource->susp_uuid[OLD] &&
> >+           resource_is_suspended(resource, OLD) &&
> >!resource_is_suspended(resource, NEW)) {
> >               idr_for_each_entry(&resource->devices, device, vnr) {
> >                       if (test_bit(NEW_CUR_UUID, &device->flags)) {
> >                               resource->susp_uuid[NEW] = true;
> >@@ -2553,6 +2554,7 @@ static bool should_try_become_up_to_date(struct
> >drbd_device *device, enum drbd_d
> >static void finish_state_change(struct drbd_resource *resource, const char *tag)
> >{
> >       enum drbd_role *role = resource->role;
> >+       bool *susp_uuid = resource->susp_uuid;
> >       struct drbd_device *device;
> >       struct drbd_connection *connection;
> >       bool starting_resync = false;
> >@@ -2828,7 +2830,10 @@ static void finish_state_change(struct
> >drbd_resource *resource, const char *tag)
> >               if (role[OLD] == R_SECONDARY && role[NEW] == R_PRIMARY)
> >                       create_new_uuid = true;
> >
> >-               if (create_new_uuid)
> >+               /* When susp_uuid goes from true to false, we just created a new
> >+                * current-uuid, it is pointless to do this one more time
> >+                */
> >+               if (create_new_uuid && !(susp_uuid[OLD] && !susp_uuid[NEW]))
> >                       set_bit(__NEW_CUR_UUID, &device->flags);
> >
> >               if (disk_state[NEW] != D_NEGOTIATING &&
> >get_ldev_if_state(device, D_DETACHING)) {
> >@@ -4189,7 +4194,7 @@ static int w_after_state_change(struct drbd_work
> >*w, int unused)
> >                       still_connected = true;
> >       }
> >
> >-       if (!susp_uuid[OLD] && susp_uuid[NEW]) {
> >+       if (susp_uuid[NEW]) {
> >               unsigned long irq_flags;
> >
> >               begin_state_change(resource, &irq_flags, CS_VERBOSE);
> >
> >On Thu, Nov 23, 2023 at 9:48 AM 黄正兵 <zhengbing.huang at easystack.cn> wrote:
> >>
> >> Hi Philipp,
> >>
> >> Such modifications may cause new problems.
> >>
> >>  If, during the second state change,  the NEW_CUR_UUID flag was not cleared.
> >>  clears the susp_uuid bit in a  thirdly state change.
> >>
> >> Eventually, it becomes an endless cycle.
> >>
> >> So, is it possible to add the following modifications:
> >>
> >> diff --git a/drbd/drbd_state.c b/drbd/drbd_state.c
> >> index 2fbd0b1..6bd6431 100644
> >> --- a/drbd/drbd_state.c
> >> +++ b/drbd/drbd_state.c
> >> @@ -3938,9 +3938,13 @@ static int w_after_state_change(struct drbd_work *w, int unused)
> >>                                 drbd_maybe_khelper(device, connection, "disconnected");
> >>                 }
> >>
> >> -               if (!susp_uuid[OLD] && susp_uuid[NEW] &&
> >> -                   test_and_clear_bit(NEW_CUR_UUID, &device->flags))
> >> -                       new_current_uuid = true;
> >> +               if (susp_uuid[NEW]) {
> >> +                       if (susp_uuid[OLD])
> >> +                               clear_bit(NEW_CUR_UUID, &device->flags);
> >> +                       else if (test_and_clear_bit(NEW_CUR_UUID, &device->flags))) {
> >> +                               new_curr_uuid = true;
> >> +                       }
> >> +               }
> >>
> >>                 if (new_current_uuid)
> >>                         drbd_uuid_new_current(device, false);
> >>
> >> best regards,
> >> Zhengbing
> >>
> >>
> >>
> >>
> >>
> >> From: Philipp Reisner <philipp.reisner at linbit.com>
> >> Date: 2023-11-23 14:06:30
> >> To:  "zhengbing.huang" <zhengbing.huang at easystack.cn>
> >> Cc:  drbd-dev at lists.linbit.com,joel.colledge at linbit.com
> >> Subject: Re: [PATCH] drbd: when change susp_uuid[NEW] to true, make sure susp_uuid[OLD] is false>Hello Zhengbing,
> >> >
> >> >Thank you for pointing this out. I created a little bit different patch.
> >> >First, the sanitize_state() function should clean up the new state,
> >> >but I think it should never change the old state.
> >> >It can not rewrite history.
> >> >Second, we do not want to create two new current-uuids while I/O is frozen.
> >> >
> >> >    drbd: fix susp_uuid clearing
> >> >
> >> >   I introduced susp_uuid for keeping a resource longer suspended for
> >> >   writing a new current-uuid.
> >> >   But in the following scenario, it failed to clear the susp_uuid bit:
> >> >
> >> >    1. The network connection between a primary and secondary node
> >> >    fails. The primary has the setting quorum=2. IO freezes on the
> >> >    primary node.
> >> >
> >> >    2. The user changes the quorum setting from 2 to one. The primary
> >> >    node generates a new current-uuid and clears the susp_uuid bit in a
> >> >    second state change.
> >> >
> >> >    3. If, during the second state change, another condition for creating
> >> >    a new current UUID evaluates to true, the code fails to clear the
> >> >    suspen_uuid bit.
> >> >
> >> >   The second attempt to create a new current-uuid gets ignored since it
> >> >   is pointless to create two current-uuids while I/O is frozen.
> >> >
> >> >   Fixes: d47f7456ab ("drbd: create new UUID before resuming IO upon
> >> >   regaining quorum")
> >> >
> >> >   Co-developed-by: zhengbing.huang <zhengbing.huang at easystack.cn>
> >> >
> >> >diff --git a/drbd/drbd_state.c b/drbd/drbd_state.c
> >> >index e35150340..316d3fd39 100644
> >> >--- a/drbd/drbd_state.c
> >> >+++ b/drbd/drbd_state.c
> >> >@@ -4189,7 +4189,7 @@ static int w_after_state_change(struct drbd_work
> >> >*w, int unused)
> >> >                       still_connected = true;
> >> >       }
> >> >
> >> >-       if (!susp_uuid[OLD] && susp_uuid[NEW]) {
> >> >+       if (susp_uuid[NEW]) {
> >> >               unsigned long irq_flags;
> >> >
> >> >               begin_state_change(resource, &irq_flags, CS_VERBOSE);
> >> >
> >> >
> >> >I give this to internal review for merging.
> >> >
> >> >best regards,
> >> > Philipp
> >> >
> >> >
> >> >On Wed, Nov 22, 2023 at 4:25 AM zhengbing.huang
> >> ><zhengbing.huang at easystack.cn> wrote:
> >> >>
> >> >> The problem scenario is as follows:
> >> >>
> >> >> 1. drbd is built on two nodes, role is primary and secondary, quorum is 2.
> >> >>    then drbd's network fails. IO will be suspended.
> >> >> 2. primary modify quorum to 1, during this state change,
> >> >>    drbd will set susp_uuid[NEW] to true and generate a new UUID.
> >> >> 3. then in w_after_state_change, start the second state change,
> >> >>    set susp_uuid[NEW] to false. but during the second state change,
> >> >>    it's possible to find NEW_CUR_UUID flag was set by others.
> >> >>    then sanitize_state() will set susp_uuid[NEW] to true.
> >> >>
> >> >> Finally susp_uuid value is {true, true}, IO is frozen.
> >> >> And there is no way to set susp_uuid to false after that.
> >> >>
> >> >> So, while susp_uuid[NEW] is set to true, we want susp_uuid[OLD] to be false.
> >> >>
> >> >> Fixes: d47f7456ab ("drbd: create new UUID before resuming IO upon regaining quorum")
> >> >> Signed-off-by: zhengbing.huang <zhengbing.huang at easystack.cn>
> >> >> ---
> >> >>  drbd/drbd_state.c | 1 +
> >> >>  1 file changed, 1 insertion(+)
> >> >>
> >> >> diff --git a/drbd/drbd_state.c b/drbd/drbd_state.c
> >> >> index e35150340..0dedd2dae 100644
> >> >> --- a/drbd/drbd_state.c
> >> >> +++ b/drbd/drbd_state.c
> >> >> @@ -2356,6 +2356,7 @@ static void sanitize_state(struct drbd_resource *resource)
> >> >>         if (resource_is_suspended(resource, OLD) && !resource_is_suspended(resource, NEW)) {
> >> >>                 idr_for_each_entry(&resource->devices, device, vnr) {
> >> >>                         if (test_bit(NEW_CUR_UUID, &device->flags)) {
> >> >> +                               resource->susp_uuid[OLD] = false;
> >> >>                                 resource->susp_uuid[NEW] = true;
> >> >>                                 break;
> >> >>                         }
> >> >> --
> >> >> 2.17.1
> >> >>
> >>
> >>
>
>


More information about the drbd-dev mailing list