Commit "aio: percpu reqs_available" added some math to the nr_requests
calculation, but didn't correct the overflow calculations to handle that.
This means that this:
#include <linux/aio_abi.h>
void main(void)
{
aio_context_t ctx_idp;
io_setup(0x80000001, &ctx_idp);
}
Would trigger the newly added BUG() couple of lines after the overflow
checks.
Signed-off-by: Sasha Levin <[email protected]>
---
fs/aio.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/fs/aio.c b/fs/aio.c
index 5b7ed78..0ae450a 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -411,7 +411,8 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
/* Prevent overflows */
if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
- (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
+ (nr_events > (0x10000000U / sizeof(struct kiocb))) ||
+ (nr_events < num_possible_cpus() * 4)) {
pr_debug("ENOMEM: nr_events too high\n");
return ERR_PTR(-EINVAL);
}
--
1.8.2.1
On Fri, May 17, 2013 at 02:23:54PM -0400, Sasha Levin wrote:
> Commit "aio: percpu reqs_available" added some math to the nr_requests
> calculation, but didn't correct the overflow calculations to handle that.
>
> This means that this:
>
> #include <linux/aio_abi.h>
> void main(void)
> {
> aio_context_t ctx_idp;
> io_setup(0x80000001, &ctx_idp);
> }
>
> Would trigger the newly added BUG() couple of lines after the overflow
> checks.
This BUG() isn't in Linus' tree, and probably should be removed before
it gets there.
> Signed-off-by: Sasha Levin <[email protected]>
> ---
> fs/aio.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/fs/aio.c b/fs/aio.c
> index 5b7ed78..0ae450a 100644
> --- a/fs/aio.c
> +++ b/fs/aio.c
> @@ -411,7 +411,8 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
>
> /* Prevent overflows */
> if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
> - (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
> + (nr_events > (0x10000000U / sizeof(struct kiocb))) ||
> + (nr_events < num_possible_cpus() * 4)) {
> pr_debug("ENOMEM: nr_events too high\n");
> return ERR_PTR(-EINVAL);
This is completely wrong. Enforcing a minimum needs to be done in a way
that doesn't fail for existing users that potentially use a minimum
smaller than what is newly required. That is: an existing userland program
that only requests 16 events must not fail because of changes to the kernel
that increase the minimum number of requests. So I have to NACK this patch
as it stands.
-ben
> }
> --
> 1.8.2.1
--
"Thought is the essence of where you are now."
On 05/17/2013 02:53 PM, Benjamin LaHaise wrote:
> On Fri, May 17, 2013 at 02:23:54PM -0400, Sasha Levin wrote:
>> Commit "aio: percpu reqs_available" added some math to the nr_requests
>> calculation, but didn't correct the overflow calculations to handle that.
>>
>> This means that this:
>>
>> #include <linux/aio_abi.h>
>> void main(void)
>> {
>> aio_context_t ctx_idp;
>> io_setup(0x80000001, &ctx_idp);
>> }
>>
>> Would trigger the newly added BUG() couple of lines after the overflow
>> checks.
>
> This BUG() isn't in Linus' tree, and probably should be removed before
> it gets there.
It's not, it's in -next though.
>> Signed-off-by: Sasha Levin <[email protected]>
>> ---
>> fs/aio.c | 3 ++-
>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/aio.c b/fs/aio.c
>> index 5b7ed78..0ae450a 100644
>> --- a/fs/aio.c
>> +++ b/fs/aio.c
>> @@ -411,7 +411,8 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
>>
>> /* Prevent overflows */
>> if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
>> - (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
>> + (nr_events > (0x10000000U / sizeof(struct kiocb))) ||
>> + (nr_events < num_possible_cpus() * 4)) {
>> pr_debug("ENOMEM: nr_events too high\n");
>> return ERR_PTR(-EINVAL);
>
> This is completely wrong. Enforcing a minimum needs to be done in a way
> that doesn't fail for existing users that potentially use a minimum
> smaller than what is newly required. That is: an existing userland program
> that only requests 16 events must not fail because of changes to the kernel
> that increase the minimum number of requests. So I have to NACK this patch
> as it stands.
You didn't look around the context of the patch.
Couple of lines before that check, this happens:
nr_events = max(nr_events, num_possible_cpus() * 4);
nr_events *= 2;
The check I've added would only make sense if nr_events wrapped around, not if
nr_events was originally smaller than (num_possible_cpus() * 4).
Thanks,
Sasha