Re: [PATCH v2 01/16] nvme: fix pci doorbell size calculation

[Maxim Levitsky <mlevitsk@redhat.com>]

On Wed, 2020-04-15 at 15:13 +0200, Philippe Mathieu-Daudé wrote:
> On 4/15/20 3:01 PM, Klaus Jensen wrote:
> > From: Klaus Jensen <k.jensen@samsung.com>
> > 
> > The size of the BAR is 0x1000 (main registers) + 8 bytes for each
> > queue. Currently, the size of the BAR is calculated like so:
> > 
> >      n->reg_size = pow2ceil(0x1004 + 2 * (n->num_queues + 1) * 4);
> > 
> > Since the 'num_queues' parameter already accounts for the admin queue,
> > this should in any case not need to be incremented by one. Also, the
> > size should be initialized to (0x1000).
> > 
> >      n->reg_size = pow2ceil(0x1000 + 2 * n->num_queues * 4);
> > 
> > This, with the default value of num_queues (64), we will set aside room
> > for 1 admin queue and 63 I/O queues (4 bytes per doorbell, 2 doorbells
> > per queue).
> > 
> > Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
> > Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
> > ---
> >   hw/block/nvme.c | 7 ++++++-
> >   1 file changed, 6 insertions(+), 1 deletion(-)
> > 
> > diff --git a/hw/block/nvme.c b/hw/block/nvme.c
> > index d28335cbf377..5b5f75c9d29e 100644
> > --- a/hw/block/nvme.c
> > +++ b/hw/block/nvme.c
> > @@ -43,6 +43,9 @@
> >   #include "trace.h"
> >   #include "nvme.h"
> >   
> > +#define NVME_REG_SIZE 0x1000
> > +#define NVME_DB_SIZE  4
> > +
> >   #define NVME_GUEST_ERR(trace, fmt, ...) \
> >       do { \
> >           (trace_##trace)(__VA_ARGS__); \
> > @@ -1345,7 +1348,9 @@ static void nvme_realize(PCIDevice *pci_dev, Error **errp)
> >       pcie_endpoint_cap_init(pci_dev, 0x80);
> >   
> >       n->num_namespaces = 1;
> > -    n->reg_size = pow2ceil(0x1004 + 2 * (n->num_queues + 1) * 4);
> > +
> > +    /* num_queues is really number of pairs, so each has two doorbells */
> > +    n->reg_size = pow2ceil(NVME_REG_SIZE + 2 * n->num_queues * NVME_DB_SIZE);
> 
> Unrelated to this change, but it would be cleaner to initialize reg_size 
> using MAX_NUM_QUEUES, then in the I/O handler log GUEST_ERROR when 
> registers > n->num_queues accessed. This would model closer to the hardware.
Agree.

Also keep in mind that NVME_DB_SIZE is configurable by setting the doorbell stride.
(but this is optional, so currently this code is OK)

Other than that,
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>

> 
> >       n->ns_size = bs_size / (uint64_t)n->num_namespaces;
> >   
> >       n->namespaces = g_new0(NvmeNamespace, n->num_namespaces);
> > 
> 
> 


Best regards,
	Maxim Levitsky