Check may_checkpoint() early
RFC PATCH 1/1 Check may_checkpoint() early
From: Sukadev Bhattiprolu <[email protected]>
Date: Mon, 16 Feb 2009 12:05:50 -0800
Subject: [PATCH] Check 'may_checkpoint()' early
We currently check if a task is checkpointable when writing the task information to checkpoint file. The small downside of doing this check late is that we may have processed several processes in the tree before hitting one that cannot be checkpointed.
We anyway walk the process tree we are checkpointing earlier, while counting the tasks. We could check if all processes in the tree are checkpointable at that time and fail early if any of them are not. Since the process tree should be frozen, checking earlier should not matter ?
For now, the patch leaves the existing check in cr_write_pids(). We can remove that later.
--- checkpoint/checkpoint.c | 9 +++++++++ 1 files changed, 9 insertions(+), 0 deletions(-)
checkpoint/checkpoint.c +9 -0
@@ -347,11 +347,18 @@ static int cr_tree_count_tasks(struct cr_ctx *ctx)struct task_struct **tasks_arr = ctx->tasks_arr;int tasks_nr = ctx->tasks_nr;int nr =0 ; +int ret;read_lock (&tasklist_lock);/* count tasks via DFS scan of the tree */ while (1 ) { + ret =cr_may_checkpoint_task (task, ctx); +if (ret <0 ) { + nr = ret; +break ; + } +if (tasks_arr) {/* unlikely... but if so then try again later */ if (nr == tasks_nr) {
@@ -409,6 +416,8 @@ static int cr_build_tree(struct cr_ctx *ctx)/* count tasks (no side effects) */ n =cr_tree_count_tasks (ctx); +if (n <0 ) +return n; ctx->tasks_nr = n; ctx->tasks_arr =kzalloc (n *sizeof (*ctx->tasks_arr),GFP_KERNEL );
Signature · 1 line
1.5.2.5
A message that is not in the archive (sent off-list or not yet received): 20090221201221.GA13532-r/[email protected]
Quoted 10 lines · Sukadev Bhattiprolu
From: Sukadev Bhattiprolu <[email protected]>
Date: Mon, 16 Feb 2009 12:05:50 -0800
Subject: [PATCH] Check 'may_checkpoint()' earlyWe currently check if a task is checkpointable when writing the task information to checkpoint file. The small downside of doing this check late is that we may have processed several processes in the tree before hitting one that cannot be checkpointed.
Quoted 7 lines
We anyway walk the process tree we are checkpointing earlier, while counting the tasks. We could check if all processes in the tree are checkpointable at that time and fail early if any of them are not. Since the process tree should be frozen, checking earlier should not matter ?
For now, the patch leaves the existing check in
cr_write_pids(). We
Yes we'll at least need to add checks that there is no running task sharing an mm or files_struct with a checkpointed task, else some of our checks will definately be racy (and are even now).
Quoted 41 lines
can remove that later.
Signed-off-by Sukadev Bhattiprolu <[email protected]>--- checkpoint/checkpoint.c | 9 +++++++++ 1 files changed, 9 insertions(+), 0 deletions(-)checkpoint/checkpoint.c +9
-0@@ -347,11 +347,18 @@ static int cr_tree_count_tasks(struct cr_ctx *ctx)struct task_struct **tasks_arr = ctx->tasks_arr;int tasks_nr = ctx->tasks_nr;int nr =0 ; +int ret;read_lock (&tasklist_lock);/* count tasks via DFS scan of the tree */ while (1 ) { + ret =cr_may_checkpoint_task (task, ctx); +if (ret <0 ) { + nr = ret; +break ; + } +if (tasks_arr) {/* unlikely... but if so then try again later */ if (nr == tasks_nr) {@@ -409,6 +416,8 @@ static int cr_build_tree(struct cr_ctx *ctx)/* count tasks (no side effects) */ n =cr_tree_count_tasks (ctx); +if (n <0 ) +return n; ctx->tasks_nr = n; ctx->tasks_arr =kzalloc (n *sizeof (*ctx->tasks_arr),GFP_KERNEL );Signature · 1 line
1.5.2.5