Re: [PATCH v2] net/core: Increase default optmem_max limit

[Stephan Mueller]

Am Dienstag, 16. Januar 2018, 18:16:43 CET schrieb Björn 'besser82' Esser:

Hi Björn,

> With the new Linux Kernel Crypto API User Space Interface
> and its underlying AF_ALG socket, the current default value
> for `net.core.optmem_max` can be exhausted pretty quick when
> using asynchronous IO; on 32 bit systems it is not even enough
> for sending about 10 IOVECs at once to the socket interface.
> 
> To provide consumers of this new user space interface a well
> sufficient and reasonable maximum ancillary buffer size per
> socket by default, the limit is increased to four times of
> the previous setting:
> 
>   * 32 bit systems:  from 10240 bytes to 40960 bytes
>   * 64 bit systems:  from 20480 bytes to 81920 bytes
> 
> This allows for sending 32/64 (32/64 bit) parallel IOVECs at
> once to the socket interface, which should be enough for use
> in real world applications.
> 
> Signed-off-by: Björn Esser 

Considering NR_FILE defining the default maximum number of file descriptors, 
at max 335 MB of RAM (32 bit) or 670 MB (64 bit) could be allocated which I 
would assume to be ok in current systems.

Reviewed-by: Stephan Mueller 

Ciao
Stephan

Re: [PATCH v2] net/core: Increase default optmem_max limit

[Stephan Mueller]

Am Dienstag, 16. Januar 2018, 18:16:43 CET schrieb Björn 'besser82' Esser:

Hi Björn,

> With the new Linux Kernel Crypto API User Space Interface
> and its underlying AF_ALG socket, the current default value
> for `net.core.optmem_max` can be exhausted pretty quick when
> using asynchronous IO; on 32 bit systems it is not even enough
> for sending about 10 IOVECs at once to the socket interface.
> 
> To provide consumers of this new user space interface a well
> sufficient and reasonable maximum ancillary buffer size per
> socket by default, the limit is increased to four times of
> the previous setting:
> 
>   * 32 bit systems:  from 10240 bytes to 40960 bytes
>   * 64 bit systems:  from 20480 bytes to 81920 bytes
> 
> This allows for sending 32/64 (32/64 bit) parallel IOVECs at
> once to the socket interface, which should be enough for use
> in real world applications.
> 
> Signed-off-by: Björn Esser 

Considering NR_FILE defining the default maximum number of file descriptors, 
at max 335 MB of RAM (32 bit) or 670 MB (64 bit) could be allocated which I 
would assume to be ok in current systems.

Reviewed-by: Stephan Mueller 

Ciao
Stephan

Re: [PATCH v2] net/core: Increase default optmem_max limit

[Björn 'besser82' Esser]

With the new Linux Kernel Crypto API User Space Interface
and its underlying AF_ALG socket, the current default value
for `net.core.optmem_max` can be exhausted pretty quick when
using asynchronous IO; on 32 bit systems it is not even enough
for sending about 10 IOVECs at once to the socket interface.

To provide consumers of this new user space interface a well
sufficient and reasonable maximum ancillary buffer size per
socket by default, the limit is increased to four times of
the previous setting:

  * 32 bit systems:  from 10240 bytes to 40960 bytes
  * 64 bit systems:  from 20480 bytes to 81920 bytes

This allows for sending 32/64 (32/64 bit) parallel IOVECs at
once to the socket interface, which should be enough for use
in real world applications.

Signed-off-by: Björn Esser 
---
 net/core/sock.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/core/sock.c b/net/core/sock.c
index c0b5b2f17412..de00c571c933 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -316,7 +316,7 @@ __u32 sysctl_wmem_default __read_mostly = SK_WMEM_MAX;
 __u32 sysctl_rmem_default __read_mostly = SK_RMEM_MAX;
 
 /* Maximal space eaten by iovec or ancillary data plus some space */
-int sysctl_optmem_max __read_mostly = sizeof(unsigned long)*(2*UIO_MAXIOV+512);
+int sysctl_optmem_max __read_mostly = sizeof(unsigned 
long)*4*(2*UIO_MAXIOV+512);
 EXPORT_SYMBOL(sysctl_optmem_max);
 
 int sysctl_tstamp_allow_data __read_mostly = 1;

Re: [PATCH v2] net/core: Increase default optmem_max limit

[Björn 'besser82' Esser]

With the new Linux Kernel Crypto API User Space Interface
and its underlying AF_ALG socket, the current default value
for `net.core.optmem_max` can be exhausted pretty quick when
using asynchronous IO; on 32 bit systems it is not even enough
for sending about 10 IOVECs at once to the socket interface.

To provide consumers of this new user space interface a well
sufficient and reasonable maximum ancillary buffer size per
socket by default, the limit is increased to four times of
the previous setting:

  * 32 bit systems:  from 10240 bytes to 40960 bytes
  * 64 bit systems:  from 20480 bytes to 81920 bytes

This allows for sending 32/64 (32/64 bit) parallel IOVECs at
once to the socket interface, which should be enough for use
in real world applications.

Signed-off-by: Björn Esser 
---
 net/core/sock.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/core/sock.c b/net/core/sock.c
index c0b5b2f17412..de00c571c933 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -316,7 +316,7 @@ __u32 sysctl_wmem_default __read_mostly = SK_WMEM_MAX;
 __u32 sysctl_rmem_default __read_mostly = SK_RMEM_MAX;
 
 /* Maximal space eaten by iovec or ancillary data plus some space */
-int sysctl_optmem_max __read_mostly = sizeof(unsigned long)*(2*UIO_MAXIOV+512);
+int sysctl_optmem_max __read_mostly = sizeof(unsigned 
long)*4*(2*UIO_MAXIOV+512);
 EXPORT_SYMBOL(sysctl_optmem_max);
 
 int sysctl_tstamp_allow_data __read_mostly = 1;

[PATCH v2] net/core: Increase default optmem_max limit

[Björn 'besser82' Esser]

[PATCH v2] net/core: Increase default optmem_max limit

[Björn 'besser82' Esser]