Peter Zijlstra wrote: > On Thu, 2006-09-14 at 14:22 -0500, Mike Christie wrote: >> Peter Zijlstra wrote: >>> On Wed, 2006-09-13 at 15:50 -0500, Mike Christie wrote: >>>> Peter Zijlstra wrote: >>>>> Implement sht->swapdev() for iSCSI. This method takes care of reserving >>>>> the extra memory needed and marking all relevant sockets with SOCK_VMIO. >>>>> >>>>> When used for swapping, TCP socket creation is done under GFP_MEMALLOC and >>>>> the TCP connect is done with SOCK_VMIO to ensure their success. Also the >>>>> netlink userspace interface is marked SOCK_VMIO, this will ensure that >>>>> even >>>>> under pressure we can still communicate with the daemon (which runs as >>>>> mlockall() and needs no additional memory to operate). >>>>> >>>>> Netlink requests are handled under the new PF_MEM_NOWAIT when a swapper is >>>>> present. This ensures that the netlink socket will not block. User-space >>>>> will >>>>> need to retry failed requests. >>>>> >>>>> The TCP receive path is handled under PF_MEMALLOC for SOCK_VMIO sockets. >>>>> This makes sure we do not block the critical socket, and that we do not >>>>> fail to process incomming data. >>>>> >>>>> Signed-off-by: Peter Zijlstra <[EMAIL PROTECTED]> >>>>> CC: Mike Christie <[EMAIL PROTECTED]> >>>>> --- >>>>> drivers/scsi/iscsi_tcp.c | 103 >>>>> +++++++++++++++++++++++++++++++----- >>>>> drivers/scsi/scsi_transport_iscsi.c | 23 +++++++- >>>>> include/scsi/libiscsi.h | 1 >>>>> include/scsi/scsi_transport_iscsi.h | 2 >>>>> 4 files changed, 113 insertions(+), 16 deletions(-) >>>>> >>>>> Index: linux-2.6/drivers/scsi/iscsi_tcp.c >>>>> =================================================================== >>>>> --- linux-2.6.orig/drivers/scsi/iscsi_tcp.c >>>>> +++ linux-2.6/drivers/scsi/iscsi_tcp.c >>>>> @@ -42,6 +42,7 @@ >>>>> #include <scsi/scsi_host.h> >>>>> #include <scsi/scsi.h> >>>>> #include <scsi/scsi_transport_iscsi.h> >>>>> +#include <scsi/scsi_device.h> >>>>> >>>>> #include "iscsi_tcp.h" >>>>> >>>>> @@ -845,9 +846,13 @@ iscsi_tcp_data_recv(read_descriptor_t *r >>>>> int rc; >>>>> struct iscsi_conn *conn = rd_desc->arg.data; >>>>> struct iscsi_tcp_conn *tcp_conn = conn->dd_data; >>>>> - int processed; >>>>> + int processed = 0; >>>>> char pad[ISCSI_PAD_LEN]; >>>>> struct scatterlist sg; >>>>> + unsigned long pflags = current->flags; >>>>> + >>>>> + if (sk_has_vmio(tcp_conn->sock->sk)) >>>>> + current->flags |= PF_MEMALLOC; >>>>> >>>> Is this too late or not needed or what is it for? This function gets run >>>> from the network layer's softirq and at this point we have a skbuff with >>>> data that we want to process. The iscsi layer also does not allocate >>>> memory for read or write IO in this path. >>> I thought I found allocations in that path, lemme search... >>> found this: >>> >>> iscsi_tcp_data_recv() >>> iscsi_data_rescv() >>> iscsi_complete_pdu() >>> __iscsi_complete_pdu() >>> iscsi_recv_pdu() >>> alloc_skb( GFP_ATOMIC); >>> >> You are right that is for the netlink interface. Could we move the >> PF_MEMALLOC setting and clearing to iscsi_recv_pdu and and add it to >> iscsi_conn_error in scsi_transport_iscsi.c so that iscsi_iser and >> qla4xxx will have it set when they need it. I will send a patch for this >> along with a way to have the netlink sock vmio set for all iscsi drivers >> that need it. > > I already have such a patch, look at: > http://programming.kicks-ass.net/kernel-patches/vm_deadlock/current/iscsi_vmio.patch >
You are drowning me in patches :) I did not see that one. I was still commenting on this patch :) The new patch looks ok. > but what conditional do you want to use for PF_MEMALLOC, an > unconditional setting will be highly unpopular. > > Hmm, perhaps you could key it of sk_has_vmio(nls)... Yes. - To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
