Re: [Qemu-devel] [PATCH v12 17/21] migration: Create ram_multifd_page

[Juan Quintela]

Peter Xu <address@hidden> wrote:
> On Wed, Apr 25, 2018 at 01:27:19PM +0200, Juan Quintela wrote:
>> The function still don't use multifd, but we have simplified
>> ram_save_page, xbzrle and RDMA stuff is gone.  We have added a new
>> counter.
>> 
>> Signed-off-by: Juan Quintela <address@hidden>
>> 
>> --
>> Add last_page parameter
>> Add commets for done and address
>> Remove multifd field, it is the same than normal pages
>> Merge next patch, now we send multiple pages at a time
>> Remove counter for multifd pages, it is identical to normal pages
>> Use iovec's instead of creating the equivalent.
>> Clear memory used by pages (dave)
>> Use g_new0(danp)
>> define MULTIFD_CONTINUE
>> now pages member is a pointer
>> Fix off-by-one in number of pages in one packet
>> Remove RAM_SAVE_FLAG_MULTIFD_PAGE
>> s/multifd_pages_t/MultiFDPages_t/
>> ---
>>  migration/ram.c | 93 +++++++++++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 93 insertions(+)
>> 
>> diff --git a/migration/ram.c b/migration/ram.c
>> index 398cb0af3b..862ec53d32 100644
>> --- a/migration/ram.c
>> +++ b/migration/ram.c
>> @@ -54,6 +54,7 @@
>>  #include "migration/block.h"
>>  #include "sysemu/sysemu.h"
>>  #include "qemu/uuid.h"
>> +#include "qemu/iov.h"
>>  
>>  /***********************************************************/
>>  /* ram save/restore */
>> @@ -692,8 +693,65 @@ struct {
>>      QemuSemaphore sem_sync;
>>      /* global number of generated multifd packets */
>>      uint32_t seq;
>> +    /* send channels ready */
>> +    QemuSemaphore channels_ready;
>>  } *multifd_send_state;
>>  
>> +static void multifd_send_pages(void)
>> +{
>> +    int i;
>> +    static int next_channel;
>> +    MultiFDSendParams *p = NULL; /* make happy gcc */
>> +    MultiFDPages_t *pages = multifd_send_state->pages;
>> +
>> +    qemu_sem_wait(&multifd_send_state->channels_ready);
>
> This sem is posted when a thread has finished its work.  However this
> is called in the main migration thread.  If with this line, are the
> threads really sending things in parallel?  Since it looks to me that
> this function (and the main thread) won't send the 2nd page array if
> the 1st hasn't finished, and won't send the 3rd if the 2nd hasn't,
> vice versa...
>
> Maybe I misunderstood something.  Please feel free to correct me.

@@ -824,6 +888,7 @@ static void *multifd_send_thread(void *opaque)
             if (flags & MULTIFD_FLAG_SYNC) {
                 qemu_sem_post(&multifd_send_state->sem_sync);
             }
+            qemu_sem_post(&multifd_send_state->channels_ready);
         } else if (p->quit) {
             qemu_mutex_unlock(&p->mutex);
             break;


Notice this bit on multifd_send_thread.  We are adding one to channels_ready.
But we can enter there for two reasons:
- we need to send a new packet full of pages
- we need to send a syncronization packet.
  this is what happens when we start.

Before the main thread start, all the other channels have to be created,
so we have that semaphore initialized to the right number of channels.

How do you preffer this to be documented?

Later, Juan.



>> +    for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) {
>> +        p = &multifd_send_state->params[i];
>> +
>> +        qemu_mutex_lock(&p->mutex);
>> +        if (!p->pending_job) {
>> +            p->pending_job++;
>> +            next_channel = (i + 1) % migrate_multifd_channels();
>> +            break;
>> +        }
>> +        qemu_mutex_unlock(&p->mutex);
>> +    }
>> +    p->pages->used = 0;
>> +    multifd_send_state->seq++;
>> +    p->seq = multifd_send_state->seq;
>> +    p->pages->block = NULL;
>> +    multifd_send_state->pages = p->pages;
>> +    p->pages = pages;

[1]

>
> Here we directly replaced MultiFDSendParams.pages with
> multifd_send_state->pages.  Then are we always using a single
> MultiFDPages_t struct?  And if so, will all the initial
> MultiFDSendParams.pages memory leaked without freed?

Multifdsend_state_pages is stored in pages variable.
We stored it at [1].

What we have (sending side) is:

- 1 multifd_pages by channel
- 1 multifd_pages by main thread

What we do here is:

  pages = multifd_send_state->pages;
  multifd_send_state->pages = channel[i]->pages;
  channel[i]->pages = pages;

So we are just doing a swap.  We do that through the whole loop to have
smaller names (ii.e. basically we do everything over pages->foo), buht
the idea is what I put there. (Ok, what I called channel[i] is "p").

But everywhere on that file (compression threads and multifd ones) use p
to mean the parameters of a thread.

Later, Juan.