Hello!
I''m working on a guest OS port which uses NetBSD drivers, and
I''m
currently working on the netfront support. I''m basing the driver on
the FreeBSD 7.0 driver, and I''ve run into a problem in
network_alloc_rx_buffers.
Parts of the code with line numbers is given below, and I''ll try to
explain where the problem occurs.
The basic problem is that the hypercall on line 112 below fails
because Xen is handed the page number -1.
1 static void
2 network_alloc_rx_buffers(struct netfront_info *sc)
3 {
[...]
48 for (i = 0, m_new = sc->xn_rx_batch; m_new;
49 i++, sc->xn_rx_batchlen--, m_new = next) {
50
[...]
70 rx_pfn_array[i] = vtomach(mtod(m_new,vm_offset_t)) >> PAGE_SHIFT;
The above call fails gives -1 (i.e., an invalid virtual -> machine
translation) time since some mbufs are placed on the same page.
71
72 /* Remove this page from pseudo phys map before passing back to Xen. */
73 xen_phys_machine[((unsigned long)m_new->m_ext.ext_args >>
PAGE_SHIFT)]
74 = INVALID_P2M_ENTRY;
... because this invalidates that mapping. I therefore get every
second entry in rx_pfn_array set to -1.
75
76 rx_mcl[i].op = __HYPERVISOR_update_va_mapping;
77 rx_mcl[i].args[0] = (unsigned long)mtod(m_new,vm_offset_t);
78 rx_mcl[i].args[1] = 0;
79 rx_mcl[i].args[2] = 0;
80
81 }
[...]
111 /* Zap PTEs and give away pages in one big multicall. */
112 (void)HYPERVISOR_multicall(rx_mcl, i+1);
[...]
126 }
... So the multicall fails.
Questions: I don''t know FreeBSD very well, but does FreeBSD always
place the mbufs on separate pages? How could the above code otherwise
work?
I hope I didn''t just make any stupid mistake when porting the code.
// Simon
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel
At Thu, 09 Mar 2006 16:36:13 +0100, Simon Kagstrom wrote:> 48 for (i = 0, m_new = sc->xn_rx_batch; m_new; > 49 i++, sc->xn_rx_batchlen--, m_new = next) { > 50 > [...] > 70 rx_pfn_array[i] = vtomach(mtod(m_new,vm_offset_t)) >> PAGE_SHIFT; > > The above call fails gives -1 (i.e., an invalid virtual -> machine > translation) time since some mbufs are placed on the same page. > > 71 > 72 /* Remove this page from pseudo phys map before passing back to Xen. */ > 73 xen_phys_machine[((unsigned long)m_new->m_ext.ext_args >> PAGE_SHIFT)] > 74 = INVALID_P2M_ENTRY; > > ... because this invalidates that mapping. I therefore get every > [...] > second entry in rx_pfn_array set to -1. > Questions: I don''t know FreeBSD very well, but does FreeBSD always > place the mbufs on separate pages? How could the above code otherwise > work?I think I found the problem myself, in sys/i386-xen/conf/XENCONF, MCLSHIFT is defined to 12, which means that each mbuf cluster will be 4KB large. I had it at 11, which should explain why every second rx_pfn_array was reset. So, then I''m curious: was this the reason for setting MCLSHIFT to 12 in FreeBSD? // Simon _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel