While developing a 9P server (https://github.com/Barre/ZeroFS) and
testing it under high-load, I was running into allocation failures.
The failures occur even with plenty of free memory available because
kmalloc requires contiguous physical memory.
This results in errors like:
ls: page allocation failure: order:7, mode:0x40c40(GFP_NOFS|__GFP_COMP)
This patch introduces a transport capability flag (supports_vmalloc)
that indicates whether a transport can work with vmalloc'd buffers
(non-physically contiguous memory). Transports requiring DMA should
leave this flag as false.
The fd-based transports (tcp, unix, fd) set this flag to true, and
p9_fcall_init will use kvmalloc instead of kmalloc for these
transports. This allows the allocator to fall back to vmalloc when
contiguous physical memory is not available.
Additionally, if kmem_cache_alloc fails, the code falls back to
kvmalloc for transports that support it.
Signed-off-by: Pierre Barre <pierre@barre.sh>
Reviewed-by: Christian Schoenebeck <linux_oss@crudebyte.com>
Message-ID: <
d2017c29-11fb-44a5-bd0f-
4204329bbefb@app.fastmail.com>
Signed-off-by: Dominique Martinet <asmadeus@codewreck.org>
* we're less flexible when choosing the response message
* size in this case
* @def: set if this transport should be considered the default
+ * @supports_vmalloc: set if this transport can work with vmalloc'd buffers
+ * (non-physically contiguous memory). Transports requiring
+ * DMA should leave this as false.
* @create: member function to create a new connection on this transport
* @close: member function to discard a connection on this transport
* @request: member function to issue a request to the transport
int maxsize; /* max message size of transport */
bool pooled_rbuffers;
int def; /* this transport should be default */
+ bool supports_vmalloc; /* can work with vmalloc'd buffers */
struct module *owner;
int (*create)(struct p9_client *client,
const char *devname, char *args);
if (likely(c->fcall_cache) && alloc_msize == c->msize) {
fc->sdata = kmem_cache_alloc(c->fcall_cache, GFP_NOFS);
fc->cache = c->fcall_cache;
+ if (!fc->sdata && c->trans_mod->supports_vmalloc) {
+ fc->sdata = kvmalloc(alloc_msize, GFP_NOFS);
+ fc->cache = NULL;
+ }
} else {
- fc->sdata = kmalloc(alloc_msize, GFP_NOFS);
+ if (c->trans_mod->supports_vmalloc)
+ fc->sdata = kvmalloc(alloc_msize, GFP_NOFS);
+ else
+ fc->sdata = kmalloc(alloc_msize, GFP_NOFS);
fc->cache = NULL;
}
if (!fc->sdata)
if (fc->cache)
kmem_cache_free(fc->cache, fc->sdata);
else
- kfree(fc->sdata);
+ kvfree(fc->sdata);
}
EXPORT_SYMBOL(p9_fcall_fini);
.maxsize = MAX_SOCK_BUF,
.pooled_rbuffers = false,
.def = 0,
+ .supports_vmalloc = true,
.create = p9_fd_create_tcp,
.close = p9_fd_close,
.request = p9_fd_request,
.name = "unix",
.maxsize = MAX_SOCK_BUF,
.def = 0,
+ .supports_vmalloc = true,
.create = p9_fd_create_unix,
.close = p9_fd_close,
.request = p9_fd_request,
.name = "fd",
.maxsize = MAX_SOCK_BUF,
.def = 0,
+ .supports_vmalloc = true,
.create = p9_fd_create,
.close = p9_fd_close,
.request = p9_fd_request,
.maxsize = P9_RDMA_MAXSIZE,
.pooled_rbuffers = true,
.def = 0,
+ .supports_vmalloc = false,
.owner = THIS_MODULE,
.create = rdma_create_trans,
.close = rdma_close,
.close = p9_usbg_close,
.request = p9_usbg_request,
.cancel = p9_usbg_cancel,
+ .supports_vmalloc = false,
.owner = THIS_MODULE,
};
.maxsize = PAGE_SIZE * (VIRTQUEUE_NUM - 3),
.pooled_rbuffers = false,
.def = 1,
+ .supports_vmalloc = false,
.owner = THIS_MODULE,
};
.maxsize = 1 << (XEN_9PFS_RING_ORDER + XEN_PAGE_SHIFT - 2),
.pooled_rbuffers = false,
.def = 1,
+ .supports_vmalloc = false,
.create = p9_xen_create,
.close = p9_xen_close,
.request = p9_xen_request,